diff --git a/.gitattributes b/.gitattributes index c1dfee9159fcffd480ab582cf477449d5e4de473..6823478c3e49f0478a526633b72d21464a54743a 100644 --- a/.gitattributes +++ b/.gitattributes @@ -7010,3 +7010,34 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2023/Target-Based[[:space:]]Offensive[[:space:]]Language[[:space:]]Identification/86bf1a0f-f75b-4c45-9ad4-3f47bb9b4c21_origin.pdf filter=lfs diff=lfs merge=lfs -text 2023/Task-Aware[[:space:]]Specialization[[:space:]]for[[:space:]]Efficient[[:space:]]and[[:space:]]Robust[[:space:]]Dense[[:space:]]Retrieval[[:space:]]for[[:space:]]Open-Domain[[:space:]]Question[[:space:]]Answering/7fe9fec5-2530-4d79-b0cb-354977af8b03_origin.pdf filter=lfs diff=lfs merge=lfs -text 2023/TeCS_[[:space:]]A[[:space:]]Dataset[[:space:]]and[[:space:]]Benchmark[[:space:]]for[[:space:]]Tense[[:space:]]Consistency[[:space:]]of[[:space:]]Machine[[:space:]]Translation/0dc2646a-6c7c-4c5a-a2e7-a8b77ccd8cf5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Teaching[[:space:]]Small[[:space:]]Language[[:space:]]Models[[:space:]]to[[:space:]]Reason/88702c47-13fb-46e6-923c-6b020286bf7a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Text-to-SQL[[:space:]]Error[[:space:]]Correction[[:space:]]with[[:space:]]Language[[:space:]]Models[[:space:]]of[[:space:]]Code/63d877fb-bf81-4e0e-9af4-85d702a6c368_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/The[[:space:]]Art[[:space:]]of[[:space:]]Prompting_[[:space:]]Event[[:space:]]Detection[[:space:]]based[[:space:]]on[[:space:]]Type[[:space:]]Specific[[:space:]]Prompts/5e1d8751-6f7a-44c9-99a0-fd4f3669ec72_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/The[[:space:]]Ecological[[:space:]]Fallacy[[:space:]]in[[:space:]]Annotation_[[:space:]]Modeling[[:space:]]Human[[:space:]]Label[[:space:]]Variation[[:space:]]goes[[:space:]]beyond[[:space:]]Sociodemographics/8b624391-c7ce-443d-9caf-875f1d838500_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/The[[:space:]]Inside[[:space:]]Story_[[:space:]]Towards[[:space:]]Better[[:space:]]Understanding[[:space:]]of[[:space:]]Machine[[:space:]]Translation[[:space:]]Neural[[:space:]]Evaluation[[:space:]]Metrics/d25ab395-c57a-4c68-8e6d-00772d415f43_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/The[[:space:]]Mechanical[[:space:]]Bard_[[:space:]]An[[:space:]]Interpretable[[:space:]]Machine[[:space:]]Learning[[:space:]]Approach[[:space:]]to[[:space:]]Shakespearean[[:space:]]Sonnet[[:space:]]Generation/3330f90d-0cb8-4666-aeb2-f9f31ccb9534_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/The[[:space:]]Role[[:space:]]of[[:space:]]Global[[:space:]]and[[:space:]]Local[[:space:]]Context[[:space:]]in[[:space:]]Named[[:space:]]Entity[[:space:]]Recognition/d05b6607-6d83-45a3-9ebc-fc5492b0e085_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/The[[:space:]]Tail[[:space:]]Wagging[[:space:]]the[[:space:]]Dog_[[:space:]]Dataset[[:space:]]Construction[[:space:]]Biases[[:space:]]of[[:space:]]Social[[:space:]]Bias[[:space:]]Benchmarks/fca46b6f-67a0-44f8-a5a2-d1a3682b3c41_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Theory-Grounded[[:space:]]Computational[[:space:]]Text[[:space:]]Analysis/11a89724-ae4c-4578-823c-45401a11d43f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Token-Level[[:space:]]Self-Evolution[[:space:]]Training[[:space:]]for[[:space:]]Sequence-to-Sequence[[:space:]]Learning/8f4944a3-fb6e-4f22-bf90-4b7384f2525a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Toward[[:space:]]Expanding[[:space:]]the[[:space:]]Scope[[:space:]]of[[:space:]]Radiology[[:space:]]Report[[:space:]]Summarization[[:space:]]to[[:space:]]Multiple[[:space:]]Anatomies[[:space:]]and[[:space:]]Modalities/5f10cf4c-c72c-4642-9a79-12d04b80e5f5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Towards[[:space:]]Adaptive[[:space:]]Prefix[[:space:]]Tuning[[:space:]]for[[:space:]]Parameter-Efficient[[:space:]]Language[[:space:]]Model[[:space:]]Fine-tuning/4802d16c-8245-489f-8ead-82b990dc5bd2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Towards[[:space:]]Fewer[[:space:]]Hallucinations[[:space:]]in[[:space:]]Knowledge-Grounded[[:space:]]Dialogue[[:space:]]Generation[[:space:]]via[[:space:]]Augmentative[[:space:]]and[[:space:]]Contrastive[[:space:]]Knowledge-Dialogue/989c66cd-1d3d-4fda-ba4f-e2869fc255a6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Tracing[[:space:]]Linguistic[[:space:]]Markers[[:space:]]of[[:space:]]Influence[[:space:]]in[[:space:]]a[[:space:]]Large[[:space:]]Online[[:space:]]Organisation/e643dbbe-47ab-405d-a70d-46f0528045a3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Trading[[:space:]]Syntax[[:space:]]Trees[[:space:]]for[[:space:]]Wordpieces_[[:space:]]Target-oriented[[:space:]]Opinion[[:space:]]Words[[:space:]]Extraction[[:space:]]with[[:space:]]Wordpieces[[:space:]]and[[:space:]]Aspect[[:space:]]Enhancement/3b3076af-e14d-4a09-b446-5edad95e9345_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Transformed[[:space:]]Protoform[[:space:]]Reconstruction/244a271d-f768-4ade-bdfe-e0da000e90a0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/TwistList_[[:space:]]Resources[[:space:]]and[[:space:]]Baselines[[:space:]]for[[:space:]]Tongue[[:space:]]Twister[[:space:]]Generation/450a8be3-d9ba-4e62-a952-97c7e24dcba6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Typo-Robust[[:space:]]Representation[[:space:]]Learning[[:space:]]for[[:space:]]Dense[[:space:]]Retrieval/8977932e-2af7-46c1-9920-a3e3abaf9a04_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Uncertainty-Aware[[:space:]]Bootstrap[[:space:]]Learning[[:space:]]for[[:space:]]Joint[[:space:]]Extraction[[:space:]]on[[:space:]]Distantly-Supervised[[:space:]]Data/a0f755f8-19b9-4154-a835-8d33c06b518f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Understanding[[:space:]]Demonstration-based[[:space:]]Learning[[:space:]]from[[:space:]]a[[:space:]]Causal[[:space:]]Perspective/58464278-fa19-45b3-a09f-9c2dd2d2dc18_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/UniTRec_[[:space:]]A[[:space:]]Unified[[:space:]]Text-to-Text[[:space:]]Transformer[[:space:]]and[[:space:]]Joint[[:space:]]Contrastive[[:space:]]Learning[[:space:]]Framework[[:space:]]for[[:space:]]Text-based[[:space:]]Recommendation/31a35e47-82fb-45ce-aa23-0bc4dd29fd03_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Unsupervised[[:space:]]Subtitle[[:space:]]Segmentation[[:space:]]with[[:space:]]Masked[[:space:]]Language[[:space:]]Models/d5fa5db2-10b5-46c6-ab9c-a92346dcaabb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Using[[:space:]]contradictions[[:space:]]improves[[:space:]]question[[:space:]]answering[[:space:]]systems/56951eb2-a64c-4213-acb6-6239f47120c0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/When[[:space:]]to[[:space:]]Use[[:space:]]Efficient[[:space:]]Self[[:space:]]Attention_[[:space:]]Profiling[[:space:]]Text,[[:space:]]Speech[[:space:]]and[[:space:]]Image[[:space:]]Transformer[[:space:]]Variants/8b2664b3-8c7e-4af7-bd6b-26db06c033c3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/With[[:space:]]a[[:space:]]Little[[:space:]]Push,[[:space:]]NLI[[:space:]]Models[[:space:]]can[[:space:]]Robustly[[:space:]]and[[:space:]]Efficiently[[:space:]]Predict[[:space:]]Faithfulness/d3da2b49-e96a-43e5-ae6f-e575c826018a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/XL-LEXEME_[[:space:]]WiC[[:space:]]Pretrained[[:space:]]Model[[:space:]]for[[:space:]]Cross-Lingual[[:space:]]LEXical[[:space:]]sEMantic[[:space:]]changE/4754c7cf-ef79-4237-89fc-dec2c7680fce_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Zero-Shot[[:space:]]and[[:space:]]Few-Shot[[:space:]]Stance[[:space:]]Detection[[:space:]]on[[:space:]]Varied[[:space:]]Topics[[:space:]]via[[:space:]]Conditional[[:space:]]Generation/6aade9ed-d045-4bed-80d3-d9ddb4ea3243_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Zero-shot[[:space:]]Cross-lingual[[:space:]]Transfer[[:space:]]With[[:space:]]Learned[[:space:]]Projections[[:space:]]Using[[:space:]]Unlabeled[[:space:]]Target-Language[[:space:]]Data/ce957704-0301-4c9b-84de-73d565462619_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/mOKB6_[[:space:]]A[[:space:]]Multilingual[[:space:]]Open[[:space:]]Knowledge[[:space:]]Base[[:space:]]Completion[[:space:]]Benchmark/70c57a7f-70d1-4b2e-a713-7ef674b2ac69_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/mPMR_[[:space:]]A[[:space:]]Multilingual[[:space:]]Pre-trained[[:space:]]Machine[[:space:]]Reader[[:space:]]at[[:space:]]Scale/24809581-970f-402d-a256-c7d8e2f40ff3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/xSIM++_[[:space:]]An[[:space:]]Improved[[:space:]]Proxy[[:space:]]to[[:space:]]Bitext[[:space:]]Mining[[:space:]]Performance[[:space:]]for[[:space:]]Low-Resource[[:space:]]Languages/632d8eaf-c025-45c6-9916-28d9ca6a54ad_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2023/Teaching Small Language Models to Reason/88702c47-13fb-46e6-923c-6b020286bf7a_content_list.json b/2023/Teaching Small Language Models to Reason/88702c47-13fb-46e6-923c-6b020286bf7a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a908eeaca3006b9b70bb088bb617168f145d0055 --- /dev/null +++ b/2023/Teaching Small Language Models to Reason/88702c47-13fb-46e6-923c-6b020286bf7a_content_list.json @@ -0,0 +1,1444 @@ +[ + { + "type": "text", + "text": "Teaching Small Language Models to Reason", + "text_level": 1, + "bbox": [ + 268, + 84, + 727, + 104 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lucie Charlotte Magister*", + "bbox": [ + 139, + 114, + 366, + 131 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "University of Cambridge", + "bbox": [ + 147, + 131, + 351, + 148 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1cm67@cam.ac.uk", + "bbox": [ + 159, + 149, + 341, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jonathan Mallinson", + "bbox": [ + 410, + 115, + 584, + 130 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Google Research", + "bbox": [ + 426, + 131, + 569, + 147 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jonmall@google.com", + "bbox": [ + 391, + 149, + 608, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jakub Adamek", + "bbox": [ + 682, + 115, + 816, + 130 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Google Research", + "bbox": [ + 678, + 131, + 820, + 147 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "enkait@google.com", + "bbox": [ + 645, + 149, + 853, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Eric Malmi", + "text_level": 1, + "bbox": [ + 278, + 180, + 381, + 195 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Google Research", + "bbox": [ + 260, + 197, + 401, + 212 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "emailmi@google.com", + "bbox": [ + 228, + 215, + 435, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Aliaksei Severyn", + "text_level": 1, + "bbox": [ + 591, + 180, + 739, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Google Research", + "bbox": [ + 596, + 197, + 736, + 212 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "severyn@google.com", + "bbox": [ + 557, + 215, + 776, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 252, + 339, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chain of thought prompting successfully improves the reasoning capabilities of large language models, achieving state of the art results on a range of datasets. However, these reasoning capabilities only appear to emerge in models with at least tens of billions of parameters. In this paper, we explore the transfer of such reasoning capabilities to smaller models via knowledge distillation, also investigating model and dataset size trade-off. Specifically, we finetune a student model on the chain of thought outputs generated by a larger teacher model. Our experiments show that the proposed method improves task performance across arithmetic, commonsense and symbolic reasoning datasets. For example, the accuracy of T5 XXL on GSM8K improves from $8.11\\%$ to $21.99\\%$ and $18.42\\%$ when finetuned on PaLM 540B and GPT-3 175B generated chains of thought, respectively.", + "bbox": [ + 141, + 281, + 460, + 565 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 114, + 580, + 258, + 594 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chain of thought (CoT) prompting encourages language models (LMs) to break down a reasoning task into a series of intermediate steps (Wei et al., 2022). They demonstrate that this prompting significantly increases the task accuracy of large language models (LLMs) across commonsense, symbolic and mathematical reasoning datasets. Here, LLMs are models with at least tens of billions of parameters, such as PaLM 540B (Chowdhery et al., 2022), GPT-3 175B (Brown et al., 2020), or UL2 20B (Tay et al., 2022). However, the reasoning capabilities of smaller LMs do not improve with CoT prompting, mostly producing illogical CoT. Notably, CoT prompting even reduces the accuracy of models with less than 10 billion parameters. Wei et al. (2022) attribute this to abilities, such as semantic understanding and symbolic mapping, only emerging at larger scales. This leads us to our re", + "bbox": [ + 112, + 606, + 489, + 896 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "search question: can the reasoning capabilities of LLMs be transferred to smaller LMs via finetuning?", + "bbox": [ + 507, + 253, + 884, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This work explores CoT knowledge distillation (Hinton et al., 2015) from PaLM 540B (Chowdhery et al., 2022) and GPT-3 175B (Brown et al., 2020) to different sizes of the smaller language model T5 (Raffel et al., 2020), such as T5 XXL, XL and base, which have 11 billion, 3 billion and 220 million parameters, respectively. As a result of our work, we make two recommendations: (1) perform knowledge distillation by finetuning the student model on the CoT generated by a large teacher model; and (2) generate the CoT from an LLM, as proposed by Wei et al. (2022), but crucially provide the solution to the task in the few-shot prompt. We demonstrate that the proposed method improves task performance across arithmetic, commonsense and symbolic reasoning datasets irrespective of the teacher model used. For example, we show an accuracy increase from $8.11\\%$ to $21.99\\%$ and $18.42\\%$ on the GSM8K (Cobbe et al., 2021) dataset when finetuning T5 XXL on PaLM 540B and GPT-3 175B generated CoT data, respectively.", + "bbox": [ + 507, + 285, + 884, + 623 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 509, + 634, + 663, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This work is inspired by the seminal work of Wei et al. (2022) on CoT prompting. They demonstrate that prefixing an input with 2-8 exemplars of CoT reasoning encourages LMs to do the same, reaching state-of-the-art performance on datasets such as GSM8K (Cobbe et al., 2021). Wang et al. (2022) show that task accuracy can be further improved by using self-consistency in CoT prompting. Self-consistency samples CoT reasoning paths from a model's decoder and returns the most consistent path by taking the majority vote. Subsequently, Chung et al. (2022) explore finetuning a FLAN-based (Wei et al., 2021) version of PaLM on manually generated CoT data.", + "bbox": [ + 507, + 661, + 884, + 885 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Concurrent to our work, a small number of other works propose methods focused on CoT student-", + "bbox": [ + 507, + 887, + 884, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Research conducted during an internship at Google.", + "bbox": [ + 141, + 904, + 465, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1773", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics", + "bbox": [ + 226, + 945, + 769, + 958 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Volume 2: Short Papers, pages 1773-1781", + "bbox": [ + 368, + 959, + 628, + 971 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "July 9-14, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 295, + 972, + 700, + 985 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "teacher knowledge distillation. Ho et al. (2022) and Li et al. (2022) also explore knowledge distillation with the difference of proposing diverse sampling and rationalization prompting, respectively. In contrast to their work, our work explores more teacher models and demonstrates both the effects of dataset and model size on accuracy. We also achieve a higher accuracy on common datasets, such as GSM8K, than Ho et al. (2022). In contrast to our work, Shridhar et al. (2022) focus on training two models, one for problem decomposition and one for solving. Yet differently, the focus of Eisenstein et al. (2022) relies on producing markup-and-mask explanations for open-book question answering. Lastly, Huang et al. (2022) present one related experiment, however, we present a more in-depth exploration on more datasets. To the best of our knowledge, our work is the first to extensively explore the improvement of the reasoning ability of small LMs via knowledge distillation across multiple model architectures, and observing the effects of student model size and dataset size on accuracy.", + "bbox": [ + 112, + 84, + 492, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 114, + 464, + 218, + 479 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose a two-step pipeline for CoT knowledge distillation. The first step comprises annotating an existing supervised dataset with CoT reasoning generated by a teacher model. To generate high quality data, we propose using LLMs, such as PaLM 540B or GPT-3 175B, as teachers, based on the finding that CoT reasoning improves with model scale (Wei et al., 2022). Specifically, we perform few-shot prompting with 8 exemplars on these models to generate CoTs. However, we make a key modification to the prompts proposed by Wei et al. (2022). We adapt the few-shot prompts to provide the model with the target after posing the question and before providing example CoT. This is based on the observation that providing this guidance allows LLMs to correct small mistakes in the CoT. Lastly, we remove all incorrect CoT based on the target answer to prevent the student to learn from bad examples. The second step comprises finetuning a student model via teacher forcing (Williams and Zipser, 1989). The student is provided with the question as input, and the CoT and answer as the target. As the model is trained on producing a CoT during finetuning, prompting is not required. Figure 1 provides an overview of the proposed method.", + "bbox": [ + 112, + 500, + 489, + 917 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e48afad38d228e4820b40a7fb5a37dfe0cb0633ef7e7a40866fd7553ce4a3c2a.jpg", + "image_caption": [ + "Figure 1: Overview of the proposed method." + ], + "image_footnote": [], + "bbox": [ + 509, + 85, + 884, + 395 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "4 Experimental Setup", + "text_level": 1, + "bbox": [ + 507, + 445, + 717, + 462 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We follow a similar experimental setup to Wei et al. (2022), focusing on tasks covering arithmetic, commonsense and symbolic reasoning.", + "bbox": [ + 507, + 473, + 885, + 521 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "4.1 Benchmarks and Metrics", + "text_level": 1, + "bbox": [ + 507, + 537, + 754, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "4.1.1 Arithmetic Reasoning", + "text_level": 1, + "bbox": [ + 507, + 560, + 742, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We benchmark the proposed method on the following math word problem datasets: (1) GSM8K (Cobbe et al., 2021), (2) MAWPS (Koncel-Kedziorski et al., 2016) and (3) ASDiv (Miao et al., 2021). We use the official training and testing split for GSM8K, taking the last $10\\%$ of the training split for validation, and the 5-fold cross validation splits available for MAWPS and ASDiv. We evaluate task accuracy by checking for the target answer as the final answer in the CoT. In addition, we compute the task accuracy given an external calculator, to account for arithmetic mistakes made by the model, despite the CoT being correct. The external calculator moves through the generated output, recalculating the left-hand-side of equations. It then replaces the right-hand side with the calculated output, to avoid arithmetic mistakes being carried forward. For example, if a model outputted $^{\\prime}5 + 5 = 11$ . $11*2 = 22'$ , then the external calculator would first calculate $^{\\prime}5 + 5$ and replace the '11' with a '10'. In the subsequent equation, it would", + "bbox": [ + 505, + 581, + 885, + 919 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1774", + "bbox": [ + 480, + 927, + 521, + 940 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "also replace the '11' with a '10' and arrive at the final result of '20'.", + "bbox": [ + 112, + 84, + 485, + 115 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1.2 Commonsense Reasoning", + "text_level": 1, + "bbox": [ + 112, + 127, + 374, + 142 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We benchmark the model's ability to perform commonsense reasoning on the StrategyQA dataset (Geva et al., 2021a). As a testing split is not available, we do not shuffle the dataset to allow reproducing our split of taking the first $80\\%$ as training data, the following $10\\%$ as validation data, and the final $10\\%$ as testing data. We compute task accuracy in the same manner as previously mentioned.", + "bbox": [ + 112, + 147, + 489, + 275 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1.3 Symbolic Reasoning", + "text_level": 1, + "bbox": [ + 112, + 286, + 332, + 300 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Lastly, we benchmark the model on two synthetic tasks for symbolic reasoning: (1) last letter concatenation and (2) coinflip (Wei et al., 2022). Last letter concatenation prompts the model to concatenate the last letter of each word in a string. Coinflip prompts the model to perform state tracking of the coin being flipped. We evaluate task accuracy in the same manner as before. Due to the rigid structure of the datasets, we focus on evaluating the model's generalizability to out-of-distribution (OOD) examples. We finetune the models on examples of length two and evaluate on sequences of length three and four. We initially infer the CoT using PaLM 540B, however, find that the LLM is able to perfectly replicate the desired CoT bar one example due to the rigidity of the template. We therefore decide to use the template generated CoT in our experiments.", + "bbox": [ + 112, + 305, + 489, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.2 Baselines and setup", + "text_level": 1, + "bbox": [ + 112, + 608, + 314, + 623 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We select PaLM 540B (Chowdhery et al., 2022) and GPT-3 175B (Brown et al., 2020) as teacher models. We select PaLM 540B based on the state-of-the-art results on the benchmarking datasets reported by Wei et al. (2022), and confirm the observed trends with GPT-3 175B. The publicly accessible teacher models are prompted as described in Section 3.", + "bbox": [ + 112, + 627, + 489, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We select different sizes of T5 (Raffel et al., 2020) as student models, as T5 is publicly available in many sizes. The student models are trained on the PaLM 540B or GPT-3 175B generated CoT data as described in Section 3. We establish T5 XXL model finetuned on the original target as the baseline. We refrain from shuffling the datasets to allow for reproducibility. For the MAWPS and ASDiv dataset, we perform 5-fold cross validation. For all remaining datasets, we take $10\\%$ of the", + "bbox": [ + 112, + 758, + 489, + 917 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Input: Roger has 5 tennis balls. He buys 2 more cans of tennis balls. Each can has 3 tennis balls. How many tennis balls does he have now?", + "bbox": [ + 539, + 96, + 835, + 142 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Output: Roger started with 5 balls. 2 cans of 3 tennis balls each is 6 tennis balls. $5 + 6 = 11$ . The answer is 11.", + "bbox": [ + 541, + 152, + 810, + 199 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 2: A training example from Wei et al. (2022) demonstrating the input and output provided to T5.", + "bbox": [ + 507, + 228, + 882, + 256 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "training set as a validation set to select the best model checkpoint. Figure 2 showcases an input examples for T5. We refer the reader to Wei et al. (2022) for more training examples, as well as the prompts used for generating the CoT using PaLM 540B and GPT-3 175B.", + "bbox": [ + 507, + 284, + 882, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We refer the reader to Appendix A for an overview of the dataset licenses. We also refer the reader to Appendix B for an overview of the computational resources.", + "bbox": [ + 507, + 380, + 882, + 444 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "5 Results", + "text_level": 1, + "bbox": [ + 509, + 458, + 606, + 472 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "5.1 Arithmetic reasoning", + "text_level": 1, + "bbox": [ + 507, + 485, + 722, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1 details the task accuracy with and without an external calculator for the arithmetic reasoning benchmarks. Our results show that the proposed method improves task accuracy across all datasets. Most notably, the task accuracy of MAwPS is significantly improved. The accuracy achieved given a calculator comes close to the accuracy of 8-shot PaLM 540B, demonstrating that knowledge distillation is effective, but potentially limited by the mathematical abilities of small models.", + "bbox": [ + 507, + 506, + 882, + 665 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/66bf5edd334996a8b4e746bb088737f1f8763fe3752638e51af40a920e3a1e33.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Baseline T5 XXLCoT Finetuned T5 XXLCoT 8-shot PaLM 540B
Acc.Acc.Acc. with Calc.Acc.Acc. with Calc.
GSM8K8.1121.9938.2156.9058.60
Dataset Size672553375337--
MAWPS54.1570.4188.2293.0093.66
Dataset Size159015901590--
ASDiv39.6442.1260.7373.972.6
Dataset Size184415441544--
", + "bbox": [ + 510, + 676, + 877, + 804 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1: Task accuracy across arithmetic reasoning datasets for T5 XXL without finetuning (baseline) and finetuned on PaLM 540B generated chain-of-thought (CoT). We report the accuracy of PaLM 540B on the used datasets for reference. We do not finetune PaLM for this, but employ 8 chain of thought prompts.", + "bbox": [ + 507, + 814, + 880, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1775", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "5.1.1 Ablation study on generating chain-of-thought data", + "text_level": 1, + "bbox": [ + 112, + 84, + 403, + 116 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We perform an ablation study to confirm that providing a LLM with the target during CoT generation is beneficial. We found that for the GSM8K dataset, PaLM 540B only achieves a $59.98\\%$ accuracy if prompted without the target. In comparison, when including the target in the prompt the accuracy is $79.37\\%$ . A superficial explanation would be that when the model is conditioned on the expected answer, it produces the same CoT but copies the answer. However, an analysis of a subset of the differences between CoT produced with and without this conditioning shows that most of the benefits actually come from the model correcting CoT that had a single step missing or was wrong.", + "bbox": [ + 112, + 118, + 489, + 344 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.2 Commonsense reasoning", + "text_level": 1, + "bbox": [ + 112, + 354, + 356, + 370 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the StrategyQA dataset (Table 3), we found that using CoT finetuning improves accuracy from $68.12\\%$ to $71.98\\%$ , using only 1319 of the original 1648 examples. Compared to the arithmetic reasoning datasets, the improvement is not as significant. This can be explained by the model lacking factual knowledge that the dataset requires. The task is heavily focused on the model reasoning on such knowledge, however, a smaller LM is most likely not in possession of this knowledge compared to a larger model with higher memorisation capacity.", + "bbox": [ + 112, + 375, + 489, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.3 Symbolic reasoning", + "text_level": 1, + "bbox": [ + 112, + 562, + 315, + 579 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 2 shows the results obtained for the synthetic symbolic reasoning datasets, focusing on OOD generalization. Focusing on Last Letter Concatenation, it can be stated that both traditional finetuning and the suggested method fail at generalizing to a longer sequence length. In comparison, the proposed method significantly increases accuracy for the Coinflip dataset with regard to generalizing to three coinflips. In contrast, generalisation to four coinflips is slightly weaker than the baseline, which performs very strongly. This may be related to the task length being twice that of the training task.", + "bbox": [ + 112, + 582, + 489, + 776 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.4 Replicating Results using different Teacher Models", + "text_level": 1, + "bbox": [ + 112, + 785, + 431, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We demonstrate the robustness of our method using a different teacher model, namely GPT-3 175B. Table 3 shows the results for GSM8K and StrategyQA when T5 XXL is finetuned on CoT data generated by GPT-3. The results show that the proposed method elicits improvements also with other", + "bbox": [ + 112, + 822, + 489, + 917 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/4aca8cb380aeace2cb83474e36ba8a73711ef1b6ff027eb5f8e8502785f78258.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Baseline T5 XXLCoT Finetuned T5 XXLCoT 8-shot PaLM 540B
Last LetterOOD: 30.000.0094.8
Concat.OOD: 40.000.0063.0
CoinflipOOD: 313.1086.7098.6
OOD: 473.8070.5090.2
", + "bbox": [ + 510, + 80, + 882, + 154 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LLMs as teachers. We also report the accuracy of T5 XXL finetuned on golden CoT provided with the datasets. For the StrategyQA dataset, the model finetuned on the golden CoT performs best, which may be attributed to the dataset being the largest, as both PaLM and GPT-3 get some examples wrong. In contrast, the model finetuned on PaLM generated CoT performs the best for GSM8K.", + "bbox": [ + 507, + 259, + 884, + 388 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/aabf7c9e1ad145c8693e81fe4f461100f7180bbd1ba1b16c72f371abbba0fa42.jpg", + "table_caption": [ + "Table 2: Task accuracy across the symbolic reasoning datasets for T5 XXL finetuned on chain-of-thought (CoT) data. For each dataset, there are 1000 training and testing examples. We report the accuracy of PaLM 540B from (Wei et al., 2022) for reference." + ], + "table_footnote": [], + "table_body": "
Base TaskOriginal CotCoT finetuned T5 XXL using PaLM 540BGPT-3 175BCoT 8-Shot PaLM 540BGPT-3 175B
GSM8K8.1119.9421.9918.4256.946.9
acc. with Calc.-26.9938.2133.0658.649.6
Dataset Size6725672553375298--
StrategyQA68.1271.9867.1563.7777.865.4
Dataset Size1648164813191319--
", + "bbox": [ + 510, + 398, + 882, + 506 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 3: Task accuracy for T5 XXL finetuned on chain-of-thought (CoT) data generated by PaLM 540B and GPT-3 175B. We also finetune on the reasoning steps provided by the datasets. We report the accuracy of PaLM 540B on the used datasets for reference. We do not finetune PaLM for this, but employ 8 chain of thought prompts.", + "bbox": [ + 507, + 514, + 882, + 615 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.5 Ablation study on model size", + "text_level": 1, + "bbox": [ + 507, + 642, + 781, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We investigate the performance gain achieved via finetuning student models of different sizes. Figure 3 shows the performance gain achieved when finetuning T5 of different sizes on the GSM8K dataset. Our results show that T5 base, with 44 times fewer parameters than T5 XXL, matches the performance of the baseline T5 XXL when trained on CoT data. Moreover, given an external calculator, even T5 small outperforms the baseline T5 XXL.", + "bbox": [ + 507, + 662, + 882, + 806 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.6 Ablation study on dataset size", + "text_level": 1, + "bbox": [ + 507, + 818, + 789, + 833 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We also investigate the trade-off between the performance gain from CoT finetuning and dataset size. Table 4 details the test accuracy achieved when finetuning T5 XXL on only $4\\%$ and $20\\%$ of the data, randomly selected. In comparison to the", + "bbox": [ + 507, + 838, + 882, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1776", + "bbox": [ + 480, + 927, + 521, + 940 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cd13943609b671f1f35d570693adefd8c778ebba2260743db1aaa7f486ecbc0e.jpg", + "image_caption": [ + "Figure 3: Effect of student model (T5) size on accuracy on GSM8K." + ], + "image_footnote": [], + "bbox": [ + 122, + 85, + 480, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "baseline accuracy of $8.11\\%$ (Table 3), we see that our method is 6x more data efficient, achieving accuracy of $11.22\\%$ with only $20\\%$ of the examples. However, training on just $20\\%$ of the data still creates a quality gap, and it's possible that with e.g. $200\\%$ larger dataset we could outperform the results in Table 3.", + "bbox": [ + 112, + 354, + 489, + 464 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ce71baa333d2588dca103acec56b9e0fd43d33f996534ca55498bef91f944e0c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Percentage of GSM8K data used to trainCoT finetuned T5 XXL
Acc.Acc. with Calc.
4% (213 examples)6.2912.28
20% (1067 examples)11.2220.47
100% (5337 examples)21.9938.21
", + "bbox": [ + 115, + 474, + 487, + 554 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 4: Task accuracy of T5 XXL finetuned on different amounts of chain-of-thought (CoT) data generated by PaLM 540B.", + "bbox": [ + 112, + 562, + 489, + 607 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6 Discussion", + "text_level": 1, + "bbox": [ + 112, + 636, + 240, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We demonstrate that finetuning larger LMs on the CoT data generated by LLMs of over 100 billion parameters can significantly improve task accuracy. Even a small number of CoT examples appear to suffice for this. However, such improvements appear to be task dependent. For example, the effects are limited for the StrategyQA dataset, which can be attributed to the task requiring specific factual knowledge, which smaller LMs may not have memorised due to their limited capacity. Nevertheless, there is some performance improvement, which may be attributed to the model learning how to approach such tasks. Moreover, the CoT knowledge distillation pipeline presented allows to trade-off model and dataset size with accuracy. Future work could explore improving the reasoning of small", + "bbox": [ + 112, + 661, + 489, + 919 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "models in multi-task settings, as well as the generation of new training data using LLMs, rather than annotating existing datasets.", + "bbox": [ + 507, + 84, + 884, + 134 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "7 Conclusion", + "text_level": 1, + "bbox": [ + 509, + 146, + 642, + 161 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This work explores CoT knowledge distillation from LLMs of over 100 billion parameters to smaller LMs. We propose a knowledge distillation pipeline consisting of two keys steps: (1) generate CoT for existing datasets using LLMs and (2) finetune smaller LMs on the CoT. Our results demonstrate that finetuning on CoT improves task accuracy across a range of benchmarking datasets.", + "bbox": [ + 507, + 172, + 885, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "8 Limitations", + "text_level": 1, + "bbox": [ + 507, + 313, + 645, + 329 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The results we present must be viewed in the context of a few limitations. A limitation is that we only perform experiments in English and on one task at a time. To be more comparable to a LLM few-shot settings, other languages and a multi-task setup could be explored. Furthermore, in order to replicate the results access to none public models is required and inference must be performed on large amounts of data. Another limitation of our work is that it only explores the original CoT prompting approach, but we do not explore subsequent improvements, such a self-consistency (Wang et al., 2022).", + "bbox": [ + 507, + 341, + 885, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "9 Ethical Considerations", + "text_level": 1, + "bbox": [ + 507, + 563, + 741, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The main ethical considerations of our research arise from the text generation performed. The concerns here are that both the teacher and student model may potentially generate non-factual (Ji et al., 2022; Pagnoni et al., 2021; Kreps et al., 2022) or offensive output (Gehman et al., 2020). This is largely influenced by the input data, which is our case are standard, peer-reviewed benchmarking tasks in the NLP domain.", + "bbox": [ + 507, + 590, + 885, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 509, + 763, + 608, + 778 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "BIG-bench collaboration. 2021. Beyond the imitation game: Measuring and extrapolating the capabilities of language models. In preparation.", + "Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901." + ], + "bbox": [ + 509, + 787, + 884, + 917 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1777", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. 2022. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311.", + "Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. 2022. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416.", + "Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. 2021. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168.", + "Jacob Eisenstein, Daniel Andor, Bernd Bohnet, Michael Collins, and David Mimno. 2022. Honest students from untrusted teachers: Learning an interpretable question-answering pipeline from a pretrained language model. arXiv preprint arXiv:2210.02498.", + "Samuel Gehman, Suchin Gururangan, Maarten Sap, Yejin Choi, and Noah A Smith. 2020. Realtoxicityprompts: Evaluating neural toxic degeneration in language models. arXiv preprint arXiv:2009.11462.", + "Mor Geva, Daniel Khashabi, Elad Segal, Tushar Khot, Dan Roth, and Jonathan Berant. 2021a. Did aristotle use a laptop? a question answering benchmark with implicit reasoning strategies. Transactions of the Association for Computational Linguistics, 9:346-361.", + "Mor Geva, Daniel Khashabi, Elad Segal, Tushar Khot, Dan Roth, and Jonathan Berant. 2021b. Did aristotle use a laptop? a question answering benchmark with implicit reasoning strategies. Transactions of the Association for Computational Linguistics, 9:346-361.", + "Geoffrey Hinton, Oriol Vinyals, Jeff Dean, et al. 2015. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2(7).", + "Namgyu Ho, Laura Schmid, and Se-Young Yun. 2022. Large language models are reasoning teachers. arXiv preprint arXiv:2212.10071.", + "Jiaxin Huang, Shixiang Shane Gu, Le Hou, Yuexin Wu, Xuezhi Wang, Hongkun Yu, and Jiawei Han. 2022. Large language models can self-improve. arXiv preprint arXiv:2210.11610.", + "Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Yejin Bang, Andrea Madotto, and Pascale Fung. 2022. Survey of hallucination in natural language generation. ACM Computing Surveys.", + "Rik Koncel-Kedziorski, Subhro Roy, Aida Amini, Nate Kushman, and Hannaneh Hajishirzi. 2016. Mawps: A math word problem repository. In Proceedings of" + ], + "bbox": [ + 115, + 85, + 489, + 917 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1152-1157.", + "Sarah Kreps, R Miles McCain, and Miles Brundage. 2022. All the news that's fit to fabricate: A-generated text as a tool of media misinformation. Journal of Experimental Political Science, 9(1):104-117.", + "Shiyang Li, Jianshu Chen, Yelong Shen, Zhiyu Chen, Xinlu Zhang, Zekun Li, Hong Wang, Jing Qian, Baolin Peng, Yi Mao, et al. 2022. Explanations from large language models make small reasoners better. arXiv preprint arXiv:2210.06726.", + "Shen-Yun Miao, Chao-Chun Liang, and Keh-Yih Su. 2021. A diverse corpus for evaluating and developing english math word problem solvers. arXiv preprint arXiv:2106.15772.", + "Artidoro Pagnoni, Vidhisha Balachandran, and Yulia Tsvetkov. 2021. Understanding factuality in abstractive summarization with frank: A benchmark for factuality metrics. arXiv preprint arXiv:2104.13346.", + "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J Liu, et al. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(140):1-67.", + "Kumar Shridhar, Alessandro Stolfo, and Mrinmaya Sachan. 2022. Distilling multi-step reasoning capabilities of large language models into smaller models via semantic decompositions. arXiv preprint arXiv:2212.00193.", + "Yi Tay, Mostafa Dehghani, Vinh Q Tran, Xavier Garcia, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Neil Houlsby, and Donald Metzler. 2022. Unifying language learning paradigms. arXiv preprint arXiv:2205.05131.", + "Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, and Denny Zhou. 2022. Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171.", + "Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. 2021. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652.", + "Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed Chi, Quoc Le, and Denny Zhou. 2022. Chain of thought prompting elicits reasoning in large language models. arXiv preprint arXiv:2201.11903.", + "Ronald J Williams and David Zipser. 1989. A learning algorithm for continually running fully recurrent neural networks. *Neural computation*, 1(2):270-280." + ], + "bbox": [ + 510, + 85, + 882, + 916 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1778", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A Dataset Usage and Licenses", + "text_level": 1, + "bbox": [ + 114, + 84, + 389, + 99 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we list the licenses for the datasets used and any ethical concerns regarding their usage. We describe the dataset splits used for all datasets in Section 4 of the paper.", + "bbox": [ + 112, + 109, + 489, + 174 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A.1 Arithmetic Reasoning", + "text_level": 1, + "bbox": [ + 114, + 184, + 337, + 200 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The GSM8K dataset (Cobbe et al., 2021) is available under the MIT license. The MAWPS dataset (Koncel-Kedziorski et al., 2016) is available under the CC BY 4.0 and the ASDiv dataset (Miao et al., 2021) is available under the CC BY-NC 4.0 license. We follow the intended usage of the datasets.", + "bbox": [ + 112, + 204, + 490, + 300 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A.2 Commonsense Reasoning", + "text_level": 1, + "bbox": [ + 114, + 312, + 366, + 328 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The StrategyQA dataset (Geva et al., 2021b) is available under the MIT license. Similar to Wei et al. (2022), we use the open-domain setting version available as part of the Big-bench collaboration (BIG-bench collaboration, 2021), available under the Apache License 2.0. We follow the intended usage of the datasets.", + "bbox": [ + 112, + 332, + 489, + 445 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A.3 Symbolic Reasoning", + "text_level": 1, + "bbox": [ + 114, + 456, + 322, + 472 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We generate the symbolic reasoning datasets as described in Wei et al. (2022).", + "bbox": [ + 112, + 476, + 485, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "B Computational Resources", + "text_level": 1, + "bbox": [ + 114, + 520, + 374, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We perform inference and finetuning on different sizes of T5 on TPUs. We perform inference on PaLM 540B also on TPUs. Our results can be replicated via the public API (https://developersgenerativeai.google/products/palm). To make requests to GPT-3 175B, we use the public API (https://beta.openai.com/docs/introduction).", + "bbox": [ + 112, + 546, + 489, + 690 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1779", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A For every submission:", + "bbox": [ + 114, + 107, + 322, + 122 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "A1. Did you describe the limitations of your work? Section 8", + "A2. Did you discuss any potential risks of your work? Section 9", + "A3. Do the abstract and introduction summarize the paper's main claims? Abstract and Section 1", + "A4. Have you used AI writing assistants when working on this paper? Left blank." + ], + "bbox": [ + 129, + 126, + 695, + 285 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "B Did you use or create scientific artifacts?", + "bbox": [ + 114, + 297, + 487, + 313 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Section 4", + "bbox": [ + 132, + 319, + 206, + 332 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "B1. Did you cite the creators of artifacts you used? Section 4", + "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Appendix A", + "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Section 4", + "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? We did not discuss this as the datasets are commonly used NLP benchmarks that do not contain personal data.", + "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? We discuss this in Section 8, the limitations section. We discuss the coverage of domains in Section 4.", + "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. We discuss this in Section 4." + ], + "bbox": [ + 129, + 343, + 880, + 764 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "C Did you run computational experiments?", + "bbox": [ + 114, + 774, + 492, + 791 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sections 4 and 5", + "bbox": [ + 132, + 796, + 258, + 810 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? We report the model specifics in section 4. We describe the computing infrastructure in Appendix 2, but do not estimate the computational budget.", + "bbox": [ + 129, + 820, + 880, + 885 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "ACL 2023 Responsible NLP Checklist", + "bbox": [ + 132, + 84, + 433, + 99 + ], + "page_idx": 7 + }, + { + "type": "footer", + "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.", + "bbox": [ + 112, + 892, + 877, + 916 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1780", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Section 4", + "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Sections 4 and 5", + "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Not applicable. Left blank." + ], + "bbox": [ + 127, + 83, + 880, + 282 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank.", + "bbox": [ + 112, + 293, + 877, + 330 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Not applicable. Left blank.", + "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Not applicable. Left blank.", + "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Not applicable. Left blank.", + "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Not applicable. Left blank.", + "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Not applicable. Left blank." + ], + "bbox": [ + 127, + 340, + 880, + 640 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1781", + "bbox": [ + 482, + 928, + 517, + 940 + ], + "page_idx": 8 + } +] \ No newline at end of file diff --git a/2023/Teaching Small Language Models to Reason/88702c47-13fb-46e6-923c-6b020286bf7a_model.json b/2023/Teaching Small Language Models to Reason/88702c47-13fb-46e6-923c-6b020286bf7a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e1ddb8e27a1fb401cf0b7500ed7f5f0953b85bfb --- /dev/null +++ b/2023/Teaching Small Language Models to Reason/88702c47-13fb-46e6-923c-6b020286bf7a_model.json @@ -0,0 +1,1868 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.27, + 0.085, + 0.728, + 0.105 + ], + "angle": 0, + "content": "Teaching Small Language Models to Reason" + }, + { + "type": "text", + "bbox": [ + 0.14, + 0.115, + 0.368, + 0.132 + ], + "angle": 0, + "content": "Lucie Charlotte Magister*" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.133, + 0.352, + 0.149 + ], + "angle": 0, + "content": "University of Cambridge" + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.15, + 0.342, + 0.163 + ], + "angle": 0, + "content": "1cm67@cam.ac.uk" + }, + { + "type": "text", + "bbox": [ + 0.411, + 0.116, + 0.586, + 0.131 + ], + "angle": 0, + "content": "Jonathan Mallinson" + }, + { + "type": "text", + "bbox": [ + 0.428, + 0.133, + 0.57, + 0.148 + ], + "angle": 0, + "content": "Google Research" + }, + { + "type": "text", + "bbox": [ + 0.392, + 0.15, + 0.609, + 0.164 + ], + "angle": 0, + "content": "jonmall@google.com" + }, + { + "type": "text", + "bbox": [ + 0.683, + 0.116, + 0.817, + 0.131 + ], + "angle": 0, + "content": "Jakub Adamek" + }, + { + "type": "text", + "bbox": [ + 0.68, + 0.133, + 0.821, + 0.148 + ], + "angle": 0, + "content": "Google Research" + }, + { + "type": "text", + "bbox": [ + 0.647, + 0.15, + 0.854, + 0.164 + ], + "angle": 0, + "content": "enkait@google.com" + }, + { + "type": "title", + "bbox": [ + 0.28, + 0.181, + 0.383, + 0.196 + ], + "angle": 0, + "content": "Eric Malmi" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.198, + 0.402, + 0.214 + ], + "angle": 0, + "content": "Google Research" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.216, + 0.436, + 0.23 + ], + "angle": 0, + "content": "emailmi@google.com" + }, + { + "type": "title", + "bbox": [ + 0.593, + 0.181, + 0.74, + 0.197 + ], + "angle": 0, + "content": "Aliaksei Severyn" + }, + { + "type": "text", + "bbox": [ + 0.597, + 0.198, + 0.737, + 0.214 + ], + "angle": 0, + "content": "Google Research" + }, + { + "type": "text", + "bbox": [ + 0.558, + 0.216, + 0.777, + 0.23 + ], + "angle": 0, + "content": "severyn@google.com" + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.253, + 0.341, + 0.268 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.282, + 0.461, + 0.567 + ], + "angle": 0, + "content": "Chain of thought prompting successfully improves the reasoning capabilities of large language models, achieving state of the art results on a range of datasets. However, these reasoning capabilities only appear to emerge in models with at least tens of billions of parameters. In this paper, we explore the transfer of such reasoning capabilities to smaller models via knowledge distillation, also investigating model and dataset size trade-off. Specifically, we finetune a student model on the chain of thought outputs generated by a larger teacher model. Our experiments show that the proposed method improves task performance across arithmetic, commonsense and symbolic reasoning datasets. For example, the accuracy of T5 XXL on GSM8K improves from \\(8.11\\%\\) to \\(21.99\\%\\) and \\(18.42\\%\\) when finetuned on PaLM 540B and GPT-3 175B generated chains of thought, respectively." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.581, + 0.26, + 0.595 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.607, + 0.49, + 0.897 + ], + "angle": 0, + "content": "Chain of thought (CoT) prompting encourages language models (LMs) to break down a reasoning task into a series of intermediate steps (Wei et al., 2022). They demonstrate that this prompting significantly increases the task accuracy of large language models (LLMs) across commonsense, symbolic and mathematical reasoning datasets. Here, LLMs are models with at least tens of billions of parameters, such as PaLM 540B (Chowdhery et al., 2022), GPT-3 175B (Brown et al., 2020), or UL2 20B (Tay et al., 2022). However, the reasoning capabilities of smaller LMs do not improve with CoT prompting, mostly producing illogical CoT. Notably, CoT prompting even reduces the accuracy of models with less than 10 billion parameters. Wei et al. (2022) attribute this to abilities, such as semantic understanding and symbolic mapping, only emerging at larger scales. This leads us to our re" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.254, + 0.885, + 0.285 + ], + "angle": 0, + "content": "search question: can the reasoning capabilities of LLMs be transferred to smaller LMs via finetuning?" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.286, + 0.885, + 0.624 + ], + "angle": 0, + "content": "This work explores CoT knowledge distillation (Hinton et al., 2015) from PaLM 540B (Chowdhery et al., 2022) and GPT-3 175B (Brown et al., 2020) to different sizes of the smaller language model T5 (Raffel et al., 2020), such as T5 XXL, XL and base, which have 11 billion, 3 billion and 220 million parameters, respectively. As a result of our work, we make two recommendations: (1) perform knowledge distillation by finetuning the student model on the CoT generated by a large teacher model; and (2) generate the CoT from an LLM, as proposed by Wei et al. (2022), but crucially provide the solution to the task in the few-shot prompt. We demonstrate that the proposed method improves task performance across arithmetic, commonsense and symbolic reasoning datasets irrespective of the teacher model used. For example, we show an accuracy increase from \\(8.11\\%\\) to \\(21.99\\%\\) and \\(18.42\\%\\) on the GSM8K (Cobbe et al., 2021) dataset when finetuning T5 XXL on PaLM 540B and GPT-3 175B generated CoT data, respectively." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.636, + 0.665, + 0.651 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.662, + 0.885, + 0.886 + ], + "angle": 0, + "content": "This work is inspired by the seminal work of Wei et al. (2022) on CoT prompting. They demonstrate that prefixing an input with 2-8 exemplars of CoT reasoning encourages LMs to do the same, reaching state-of-the-art performance on datasets such as GSM8K (Cobbe et al., 2021). Wang et al. (2022) show that task accuracy can be further improved by using self-consistency in CoT prompting. Self-consistency samples CoT reasoning paths from a model's decoder and returns the most consistent path by taking the majority vote. Subsequently, Chung et al. (2022) explore finetuning a FLAN-based (Wei et al., 2021) version of PaLM on manually generated CoT data." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.888, + 0.885, + 0.919 + ], + "angle": 0, + "content": "Concurrent to our work, a small number of other works propose methods focused on CoT student-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.142, + 0.905, + 0.467, + 0.919 + ], + "angle": 0, + "content": "*Research conducted during an internship at Google." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1773" + }, + { + "type": "footer", + "bbox": [ + 0.227, + 0.946, + 0.77, + 0.959 + ], + "angle": 0, + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + }, + { + "type": "footer", + "bbox": [ + 0.369, + 0.96, + 0.629, + 0.972 + ], + "angle": 0, + "content": "Volume 2: Short Papers, pages 1773-1781" + }, + { + "type": "footer", + "bbox": [ + 0.297, + 0.973, + 0.702, + 0.986 + ], + "angle": 0, + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.493, + 0.441 + ], + "angle": 0, + "content": "teacher knowledge distillation. Ho et al. (2022) and Li et al. (2022) also explore knowledge distillation with the difference of proposing diverse sampling and rationalization prompting, respectively. In contrast to their work, our work explores more teacher models and demonstrates both the effects of dataset and model size on accuracy. We also achieve a higher accuracy on common datasets, such as GSM8K, than Ho et al. (2022). In contrast to our work, Shridhar et al. (2022) focus on training two models, one for problem decomposition and one for solving. Yet differently, the focus of Eisenstein et al. (2022) relies on producing markup-and-mask explanations for open-book question answering. Lastly, Huang et al. (2022) present one related experiment, however, we present a more in-depth exploration on more datasets. To the best of our knowledge, our work is the first to extensively explore the improvement of the reasoning ability of small LMs via knowledge distillation across multiple model architectures, and observing the effects of student model size and dataset size on accuracy." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.465, + 0.219, + 0.48 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.501, + 0.49, + 0.919 + ], + "angle": 0, + "content": "We propose a two-step pipeline for CoT knowledge distillation. The first step comprises annotating an existing supervised dataset with CoT reasoning generated by a teacher model. To generate high quality data, we propose using LLMs, such as PaLM 540B or GPT-3 175B, as teachers, based on the finding that CoT reasoning improves with model scale (Wei et al., 2022). Specifically, we perform few-shot prompting with 8 exemplars on these models to generate CoTs. However, we make a key modification to the prompts proposed by Wei et al. (2022). We adapt the few-shot prompts to provide the model with the target after posing the question and before providing example CoT. This is based on the observation that providing this guidance allows LLMs to correct small mistakes in the CoT. Lastly, we remove all incorrect CoT based on the target answer to prevent the student to learn from bad examples. The second step comprises finetuning a student model via teacher forcing (Williams and Zipser, 1989). The student is provided with the question as input, and the CoT and answer as the target. As the model is trained on producing a CoT during finetuning, prompting is not required. Figure 1 provides an overview of the proposed method." + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.086, + 0.885, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.543, + 0.404, + 0.849, + 0.42 + ], + "angle": 0, + "content": "Figure 1: Overview of the proposed method." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.446, + 0.719, + 0.463 + ], + "angle": 0, + "content": "4 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.474, + 0.886, + 0.523 + ], + "angle": 0, + "content": "We follow a similar experimental setup to Wei et al. (2022), focusing on tasks covering arithmetic, commonsense and symbolic reasoning." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.538, + 0.756, + 0.552 + ], + "angle": 0, + "content": "4.1 Benchmarks and Metrics" + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.561, + 0.744, + 0.576 + ], + "angle": 0, + "content": "4.1.1 Arithmetic Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.582, + 0.886, + 0.92 + ], + "angle": 0, + "content": "We benchmark the proposed method on the following math word problem datasets: (1) GSM8K (Cobbe et al., 2021), (2) MAWPS (Koncel-Kedziorski et al., 2016) and (3) ASDiv (Miao et al., 2021). We use the official training and testing split for GSM8K, taking the last \\(10\\%\\) of the training split for validation, and the 5-fold cross validation splits available for MAWPS and ASDiv. We evaluate task accuracy by checking for the target answer as the final answer in the CoT. In addition, we compute the task accuracy given an external calculator, to account for arithmetic mistakes made by the model, despite the CoT being correct. The external calculator moves through the generated output, recalculating the left-hand-side of equations. It then replaces the right-hand side with the calculated output, to avoid arithmetic mistakes being carried forward. For example, if a model outputted \\(^{\\prime}5 + 5 = 11\\). \\(11*2 = 22'\\), then the external calculator would first calculate \\(^{\\prime}5 + 5\\) and replace the '11' with a '10'. In the subsequent equation, it would" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.928, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1774" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.114, + 0.085, + 0.486, + 0.116 + ], + "angle": 0, + "content": "also replace the '11' with a '10' and arrive at the final result of '20'." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.128, + 0.376, + 0.143 + ], + "angle": 0, + "content": "4.1.2 Commonsense Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.148, + 0.49, + 0.276 + ], + "angle": 0, + "content": "We benchmark the model's ability to perform commonsense reasoning on the StrategyQA dataset (Geva et al., 2021a). As a testing split is not available, we do not shuffle the dataset to allow reproducing our split of taking the first \\(80\\%\\) as training data, the following \\(10\\%\\) as validation data, and the final \\(10\\%\\) as testing data. We compute task accuracy in the same manner as previously mentioned." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.287, + 0.334, + 0.302 + ], + "angle": 0, + "content": "4.1.3 Symbolic Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.306, + 0.49, + 0.595 + ], + "angle": 0, + "content": "Lastly, we benchmark the model on two synthetic tasks for symbolic reasoning: (1) last letter concatenation and (2) coinflip (Wei et al., 2022). Last letter concatenation prompts the model to concatenate the last letter of each word in a string. Coinflip prompts the model to perform state tracking of the coin being flipped. We evaluate task accuracy in the same manner as before. Due to the rigid structure of the datasets, we focus on evaluating the model's generalizability to out-of-distribution (OOD) examples. We finetune the models on examples of length two and evaluate on sequences of length three and four. We initially infer the CoT using PaLM 540B, however, find that the LLM is able to perfectly replicate the desired CoT bar one example due to the rigidity of the template. We therefore decide to use the template generated CoT in our experiments." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.609, + 0.315, + 0.624 + ], + "angle": 0, + "content": "4.2 Baselines and setup" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.629, + 0.49, + 0.756 + ], + "angle": 0, + "content": "We select PaLM 540B (Chowdhery et al., 2022) and GPT-3 175B (Brown et al., 2020) as teacher models. We select PaLM 540B based on the state-of-the-art results on the benchmarking datasets reported by Wei et al. (2022), and confirm the observed trends with GPT-3 175B. The publicly accessible teacher models are prompted as described in Section 3." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.759, + 0.49, + 0.919 + ], + "angle": 0, + "content": "We select different sizes of T5 (Raffel et al., 2020) as student models, as T5 is publicly available in many sizes. The student models are trained on the PaLM 540B or GPT-3 175B generated CoT data as described in Section 3. We establish T5 XXL model finetuned on the original target as the baseline. We refrain from shuffling the datasets to allow for reproducibility. For the MAWPS and ASDiv dataset, we perform 5-fold cross validation. For all remaining datasets, we take \\(10\\%\\) of the" + }, + { + "type": "text", + "bbox": [ + 0.541, + 0.097, + 0.836, + 0.143 + ], + "angle": 0, + "content": "Input: Roger has 5 tennis balls. He buys 2 more cans of tennis balls. Each can has 3 tennis balls. How many tennis balls does he have now?" + }, + { + "type": "text", + "bbox": [ + 0.542, + 0.153, + 0.811, + 0.2 + ], + "angle": 0, + "content": "Output: Roger started with 5 balls. 2 cans of 3 tennis balls each is 6 tennis balls. \\(5 + 6 = 11\\). The answer is 11." + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.229, + 0.883, + 0.258 + ], + "angle": 0, + "content": "Figure 2: A training example from Wei et al. (2022) demonstrating the input and output provided to T5." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.285, + 0.884, + 0.379 + ], + "angle": 0, + "content": "training set as a validation set to select the best model checkpoint. Figure 2 showcases an input examples for T5. We refer the reader to Wei et al. (2022) for more training examples, as well as the prompts used for generating the CoT using PaLM 540B and GPT-3 175B." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.381, + 0.884, + 0.445 + ], + "angle": 0, + "content": "We refer the reader to Appendix A for an overview of the dataset licenses. We also refer the reader to Appendix B for an overview of the computational resources." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.459, + 0.608, + 0.473 + ], + "angle": 0, + "content": "5 Results" + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.486, + 0.724, + 0.501 + ], + "angle": 0, + "content": "5.1 Arithmetic reasoning" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.507, + 0.884, + 0.667 + ], + "angle": 0, + "content": "Table 1 details the task accuracy with and without an external calculator for the arithmetic reasoning benchmarks. Our results show that the proposed method improves task accuracy across all datasets. Most notably, the task accuracy of MAwPS is significantly improved. The accuracy achieved given a calculator comes close to the accuracy of 8-shot PaLM 540B, demonstrating that knowledge distillation is effective, but potentially limited by the mathematical abilities of small models." + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.677, + 0.878, + 0.805 + ], + "angle": 0, + "content": "
Baseline T5 XXLCoT Finetuned T5 XXLCoT 8-shot PaLM 540B
Acc.Acc.Acc. with Calc.Acc.Acc. with Calc.
GSM8K8.1121.9938.2156.9058.60
Dataset Size672553375337--
MAWPS54.1570.4188.2293.0093.66
Dataset Size159015901590--
ASDiv39.6442.1260.7373.972.6
Dataset Size184415441544--
" + }, + { + "type": "table_caption", + "bbox": [ + 0.508, + 0.815, + 0.882, + 0.9 + ], + "angle": 0, + "content": "Table 1: Task accuracy across arithmetic reasoning datasets for T5 XXL without finetuning (baseline) and finetuned on PaLM 540B generated chain-of-thought (CoT). We report the accuracy of PaLM 540B on the used datasets for reference. We do not finetune PaLM for this, but employ 8 chain of thought prompts." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1775" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.114, + 0.085, + 0.404, + 0.117 + ], + "angle": 0, + "content": "5.1.1 Ablation study on generating chain-of-thought data" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.12, + 0.49, + 0.346 + ], + "angle": 0, + "content": "We perform an ablation study to confirm that providing a LLM with the target during CoT generation is beneficial. We found that for the GSM8K dataset, PaLM 540B only achieves a \\(59.98\\%\\) accuracy if prompted without the target. In comparison, when including the target in the prompt the accuracy is \\(79.37\\%\\). A superficial explanation would be that when the model is conditioned on the expected answer, it produces the same CoT but copies the answer. However, an analysis of a subset of the differences between CoT produced with and without this conditioning shows that most of the benefits actually come from the model correcting CoT that had a single step missing or was wrong." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.355, + 0.357, + 0.372 + ], + "angle": 0, + "content": "5.2 Commonsense reasoning" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.376, + 0.49, + 0.552 + ], + "angle": 0, + "content": "For the StrategyQA dataset (Table 3), we found that using CoT finetuning improves accuracy from \\(68.12\\%\\) to \\(71.98\\%\\), using only 1319 of the original 1648 examples. Compared to the arithmetic reasoning datasets, the improvement is not as significant. This can be explained by the model lacking factual knowledge that the dataset requires. The task is heavily focused on the model reasoning on such knowledge, however, a smaller LM is most likely not in possession of this knowledge compared to a larger model with higher memorisation capacity." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.563, + 0.317, + 0.58 + ], + "angle": 0, + "content": "5.3 Symbolic reasoning" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.583, + 0.49, + 0.777 + ], + "angle": 0, + "content": "Table 2 shows the results obtained for the synthetic symbolic reasoning datasets, focusing on OOD generalization. Focusing on Last Letter Concatenation, it can be stated that both traditional finetuning and the suggested method fail at generalizing to a longer sequence length. In comparison, the proposed method significantly increases accuracy for the Coinflip dataset with regard to generalizing to three coinflips. In contrast, generalisation to four coinflips is slightly weaker than the baseline, which performs very strongly. This may be related to the task length being twice that of the training task." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.787, + 0.433, + 0.817 + ], + "angle": 0, + "content": "5.4 Replicating Results using different Teacher Models" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.823, + 0.49, + 0.919 + ], + "angle": 0, + "content": "We demonstrate the robustness of our method using a different teacher model, namely GPT-3 175B. Table 3 shows the results for GSM8K and StrategyQA when T5 XXL is finetuned on CoT data generated by GPT-3. The results show that the proposed method elicits improvements also with other" + }, + { + "type": "table", + "bbox": [ + 0.511, + 0.081, + 0.883, + 0.155 + ], + "angle": 0, + "content": "
Baseline T5 XXLCoT Finetuned T5 XXLCoT 8-shot PaLM 540B
Last LetterOOD: 30.000.0094.8
Concat.OOD: 40.000.0063.0
CoinflipOOD: 313.1086.7098.6
OOD: 473.8070.5090.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.508, + 0.163, + 0.885, + 0.234 + ], + "angle": 0, + "content": "Table 2: Task accuracy across the symbolic reasoning datasets for T5 XXL finetuned on chain-of-thought (CoT) data. For each dataset, there are 1000 training and testing examples. We report the accuracy of PaLM 540B from (Wei et al., 2022) for reference." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.26, + 0.885, + 0.389 + ], + "angle": 0, + "content": "LLMs as teachers. We also report the accuracy of T5 XXL finetuned on golden CoT provided with the datasets. For the StrategyQA dataset, the model finetuned on the golden CoT performs best, which may be attributed to the dataset being the largest, as both PaLM and GPT-3 get some examples wrong. In contrast, the model finetuned on PaLM generated CoT performs the best for GSM8K." + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.399, + 0.883, + 0.507 + ], + "angle": 0, + "content": "
Base TaskOriginal CotCoT finetuned T5 XXL using PaLM 540BGPT-3 175BCoT 8-Shot PaLM 540BGPT-3 175B
GSM8K8.1119.9421.9918.4256.946.9
acc. with Calc.-26.9938.2133.0658.649.6
Dataset Size6725672553375298--
StrategyQA68.1271.9867.1563.7777.865.4
Dataset Size1648164813191319--
" + }, + { + "type": "table_caption", + "bbox": [ + 0.508, + 0.516, + 0.884, + 0.617 + ], + "angle": 0, + "content": "Table 3: Task accuracy for T5 XXL finetuned on chain-of-thought (CoT) data generated by PaLM 540B and GPT-3 175B. We also finetune on the reasoning steps provided by the datasets. We report the accuracy of PaLM 540B on the used datasets for reference. We do not finetune PaLM for this, but employ 8 chain of thought prompts." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.643, + 0.783, + 0.658 + ], + "angle": 0, + "content": "5.5 Ablation study on model size" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.663, + 0.884, + 0.807 + ], + "angle": 0, + "content": "We investigate the performance gain achieved via finetuning student models of different sizes. Figure 3 shows the performance gain achieved when finetuning T5 of different sizes on the GSM8K dataset. Our results show that T5 base, with 44 times fewer parameters than T5 XXL, matches the performance of the baseline T5 XXL when trained on CoT data. Moreover, given an external calculator, even T5 small outperforms the baseline T5 XXL." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.819, + 0.791, + 0.834 + ], + "angle": 0, + "content": "5.6 Ablation study on dataset size" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.839, + 0.884, + 0.919 + ], + "angle": 0, + "content": "We also investigate the trade-off between the performance gain from CoT finetuning and dataset size. Table 4 details the test accuracy achieved when finetuning T5 XXL on only \\(4\\%\\) and \\(20\\%\\) of the data, randomly selected. In comparison to the" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.928, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1776" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.086, + 0.481, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.301, + 0.49, + 0.331 + ], + "angle": 0, + "content": "Figure 3: Effect of student model (T5) size on accuracy on GSM8K." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.355, + 0.49, + 0.466 + ], + "angle": 0, + "content": "baseline accuracy of \\(8.11\\%\\) (Table 3), we see that our method is 6x more data efficient, achieving accuracy of \\(11.22\\%\\) with only \\(20\\%\\) of the examples. However, training on just \\(20\\%\\) of the data still creates a quality gap, and it's possible that with e.g. \\(200\\%\\) larger dataset we could outperform the results in Table 3." + }, + { + "type": "table", + "bbox": [ + 0.117, + 0.475, + 0.488, + 0.555 + ], + "angle": 0, + "content": "
Percentage of GSM8K data used to trainCoT finetuned T5 XXL
Acc.Acc. with Calc.
4% (213 examples)6.2912.28
20% (1067 examples)11.2220.47
100% (5337 examples)21.9938.21
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.563, + 0.49, + 0.608 + ], + "angle": 0, + "content": "Table 4: Task accuracy of T5 XXL finetuned on different amounts of chain-of-thought (CoT) data generated by PaLM 540B." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.637, + 0.242, + 0.652 + ], + "angle": 0, + "content": "6 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.662, + 0.49, + 0.92 + ], + "angle": 0, + "content": "We demonstrate that finetuning larger LMs on the CoT data generated by LLMs of over 100 billion parameters can significantly improve task accuracy. Even a small number of CoT examples appear to suffice for this. However, such improvements appear to be task dependent. For example, the effects are limited for the StrategyQA dataset, which can be attributed to the task requiring specific factual knowledge, which smaller LMs may not have memorised due to their limited capacity. Nevertheless, there is some performance improvement, which may be attributed to the model learning how to approach such tasks. Moreover, the CoT knowledge distillation pipeline presented allows to trade-off model and dataset size with accuracy. Future work could explore improving the reasoning of small" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.135 + ], + "angle": 0, + "content": "models in multi-task settings, as well as the generation of new training data using LLMs, rather than annotating existing datasets." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.147, + 0.643, + 0.162 + ], + "angle": 0, + "content": "7 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.173, + 0.886, + 0.304 + ], + "angle": 0, + "content": "This work explores CoT knowledge distillation from LLMs of over 100 billion parameters to smaller LMs. We propose a knowledge distillation pipeline consisting of two keys steps: (1) generate CoT for existing datasets using LLMs and (2) finetune smaller LMs on the CoT. Our results demonstrate that finetuning on CoT improves task accuracy across a range of benchmarking datasets." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.315, + 0.646, + 0.33 + ], + "angle": 0, + "content": "8 Limitations" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.342, + 0.886, + 0.55 + ], + "angle": 0, + "content": "The results we present must be viewed in the context of a few limitations. A limitation is that we only perform experiments in English and on one task at a time. To be more comparable to a LLM few-shot settings, other languages and a multi-task setup could be explored. Furthermore, in order to replicate the results access to none public models is required and inference must be performed on large amounts of data. Another limitation of our work is that it only explores the original CoT prompting approach, but we do not explore subsequent improvements, such a self-consistency (Wang et al., 2022)." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.564, + 0.742, + 0.58 + ], + "angle": 0, + "content": "9 Ethical Considerations" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.592, + 0.886, + 0.736 + ], + "angle": 0, + "content": "The main ethical considerations of our research arise from the text generation performed. The concerns here are that both the teacher and student model may potentially generate non-factual (Ji et al., 2022; Pagnoni et al., 2021; Kreps et al., 2022) or offensive output (Gehman et al., 2020). This is largely influenced by the input data, which is our case are standard, peer-reviewed benchmarking tasks in the NLP domain." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.764, + 0.61, + 0.779 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.788, + 0.883, + 0.829 + ], + "angle": 0, + "content": "BIG-bench collaboration. 2021. Beyond the imitation game: Measuring and extrapolating the capabilities of language models. In preparation." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.84, + 0.885, + 0.919 + ], + "angle": 0, + "content": "Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901." + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.788, + 0.885, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1777" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.165 + ], + "angle": 0, + "content": "Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. 2022. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.176, + 0.49, + 0.241 + ], + "angle": 0, + "content": "Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. 2022. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.253, + 0.489, + 0.318 + ], + "angle": 0, + "content": "Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. 2021. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.33, + 0.489, + 0.407 + ], + "angle": 0, + "content": "Jacob Eisenstein, Daniel Andor, Bernd Bohnet, Michael Collins, and David Mimno. 2022. Honest students from untrusted teachers: Learning an interpretable question-answering pipeline from a pretrained language model. arXiv preprint arXiv:2210.02498." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.42, + 0.489, + 0.472 + ], + "angle": 0, + "content": "Samuel Gehman, Suchin Gururangan, Maarten Sap, Yejin Choi, and Noah A Smith. 2020. Realtoxicityprompts: Evaluating neural toxic degeneration in language models. arXiv preprint arXiv:2009.11462." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.483, + 0.489, + 0.549 + ], + "angle": 0, + "content": "Mor Geva, Daniel Khashabi, Elad Segal, Tushar Khot, Dan Roth, and Jonathan Berant. 2021a. Did aristotle use a laptop? a question answering benchmark with implicit reasoning strategies. Transactions of the Association for Computational Linguistics, 9:346-361." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.56, + 0.489, + 0.626 + ], + "angle": 0, + "content": "Mor Geva, Daniel Khashabi, Elad Segal, Tushar Khot, Dan Roth, and Jonathan Berant. 2021b. Did aristotle use a laptop? a question answering benchmark with implicit reasoning strategies. Transactions of the Association for Computational Linguistics, 9:346-361." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.637, + 0.489, + 0.677 + ], + "angle": 0, + "content": "Geoffrey Hinton, Oriol Vinyals, Jeff Dean, et al. 2015. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2(7)." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.687, + 0.489, + 0.727 + ], + "angle": 0, + "content": "Namgyu Ho, Laura Schmid, and Se-Young Yun. 2022. Large language models are reasoning teachers. arXiv preprint arXiv:2212.10071." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.738, + 0.489, + 0.791 + ], + "angle": 0, + "content": "Jiaxin Huang, Shixiang Shane Gu, Le Hou, Yuexin Wu, Xuezhi Wang, Hongkun Yu, and Jiawei Han. 2022. Large language models can self-improve. arXiv preprint arXiv:2210.11610." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.802, + 0.489, + 0.868 + ], + "angle": 0, + "content": "Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Yejin Bang, Andrea Madotto, and Pascale Fung. 2022. Survey of hallucination in natural language generation. ACM Computing Surveys." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.879, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Rik Koncel-Kedziorski, Subhro Roy, Aida Amini, Nate Kushman, and Hannaneh Hajishirzi. 2016. Mawps: A math word problem repository. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.529, + 0.086, + 0.884, + 0.127 + ], + "angle": 0, + "content": "the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1152-1157." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.138, + 0.884, + 0.203 + ], + "angle": 0, + "content": "Sarah Kreps, R Miles McCain, and Miles Brundage. 2022. All the news that's fit to fabricate: A-generated text as a tool of media misinformation. Journal of Experimental Political Science, 9(1):104-117." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.216, + 0.884, + 0.282 + ], + "angle": 0, + "content": "Shiyang Li, Jianshu Chen, Yelong Shen, Zhiyu Chen, Xinlu Zhang, Zekun Li, Hong Wang, Jing Qian, Baolin Peng, Yi Mao, et al. 2022. Explanations from large language models make small reasoners better. arXiv preprint arXiv:2210.06726." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.294, + 0.884, + 0.347 + ], + "angle": 0, + "content": "Shen-Yun Miao, Chao-Chun Liang, and Keh-Yih Su. 2021. A diverse corpus for evaluating and developing english math word problem solvers. arXiv preprint arXiv:2106.15772." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.359, + 0.884, + 0.412 + ], + "angle": 0, + "content": "Artidoro Pagnoni, Vidhisha Balachandran, and Yulia Tsvetkov. 2021. Understanding factuality in abstractive summarization with frank: A benchmark for factuality metrics. arXiv preprint arXiv:2104.13346." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.424, + 0.884, + 0.489 + ], + "angle": 0, + "content": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J Liu, et al. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(140):1-67." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.502, + 0.884, + 0.567 + ], + "angle": 0, + "content": "Kumar Shridhar, Alessandro Stolfo, and Mrinmaya Sachan. 2022. Distilling multi-step reasoning capabilities of large language models into smaller models via semantic decompositions. arXiv preprint arXiv:2212.00193." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.58, + 0.884, + 0.645 + ], + "angle": 0, + "content": "Yi Tay, Mostafa Dehghani, Vinh Q Tran, Xavier Garcia, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Neil Houlsby, and Donald Metzler. 2022. Unifying language learning paradigms. arXiv preprint arXiv:2205.05131." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.658, + 0.884, + 0.711 + ], + "angle": 0, + "content": "Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, and Denny Zhou. 2022. Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.723, + 0.884, + 0.788 + ], + "angle": 0, + "content": "Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. 2021. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.801, + 0.884, + 0.854 + ], + "angle": 0, + "content": "Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed Chi, Quoc Le, and Denny Zhou. 2022. Chain of thought prompting elicits reasoning in large language models. arXiv preprint arXiv:2201.11903." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.866, + 0.884, + 0.917 + ], + "angle": 0, + "content": "Ronald J Williams and David Zipser. 1989. A learning algorithm for continually running fully recurrent neural networks. *Neural computation*, 1(2):270-280." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.884, + 0.917 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1778" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.115, + 0.085, + 0.391, + 0.1 + ], + "angle": 0, + "content": "A Dataset Usage and Licenses" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.11, + 0.49, + 0.175 + ], + "angle": 0, + "content": "In this section, we list the licenses for the datasets used and any ethical concerns regarding their usage. We describe the dataset splits used for all datasets in Section 4 of the paper." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.185, + 0.338, + 0.202 + ], + "angle": 0, + "content": "A.1 Arithmetic Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.205, + 0.491, + 0.302 + ], + "angle": 0, + "content": "The GSM8K dataset (Cobbe et al., 2021) is available under the MIT license. The MAWPS dataset (Koncel-Kedziorski et al., 2016) is available under the CC BY 4.0 and the ASDiv dataset (Miao et al., 2021) is available under the CC BY-NC 4.0 license. We follow the intended usage of the datasets." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.313, + 0.367, + 0.329 + ], + "angle": 0, + "content": "A.2 Commonsense Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.334, + 0.49, + 0.446 + ], + "angle": 0, + "content": "The StrategyQA dataset (Geva et al., 2021b) is available under the MIT license. Similar to Wei et al. (2022), we use the open-domain setting version available as part of the Big-bench collaboration (BIG-bench collaboration, 2021), available under the Apache License 2.0. We follow the intended usage of the datasets." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.457, + 0.324, + 0.473 + ], + "angle": 0, + "content": "A.3 Symbolic Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.478, + 0.487, + 0.509 + ], + "angle": 0, + "content": "We generate the symbolic reasoning datasets as described in Wei et al. (2022)." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.521, + 0.376, + 0.538 + ], + "angle": 0, + "content": "B Computational Resources" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.547, + 0.49, + 0.692 + ], + "angle": 0, + "content": "We perform inference and finetuning on different sizes of T5 on TPUs. We perform inference on PaLM 540B also on TPUs. Our results can be replicated via the public API (https://developersgenerativeai.google/products/palm). To make requests to GPT-3 175B, we use the public API (https://beta.openai.com/docs/introduction)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1779" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.133, + 0.085, + 0.435, + 0.1 + ], + "angle": 0, + "content": "ACL 2023 Responsible NLP Checklist" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.108, + 0.323, + 0.123 + ], + "angle": 0, + "content": "A For every submission:" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.127, + 0.533, + 0.158 + ], + "angle": 0, + "content": "A1. Did you describe the limitations of your work? Section 8" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.169, + 0.553, + 0.2 + ], + "angle": 0, + "content": "A2. Did you discuss any potential risks of your work? Section 9" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.212, + 0.696, + 0.243 + ], + "angle": 0, + "content": "A3. Do the abstract and introduction summarize the paper's main claims? Abstract and Section 1" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.255, + 0.669, + 0.286 + ], + "angle": 0, + "content": "A4. Have you used AI writing assistants when working on this paper? Left blank." + }, + { + "type": "list", + "bbox": [ + 0.13, + 0.127, + 0.696, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.298, + 0.489, + 0.314 + ], + "angle": 0, + "content": "B Did you use or create scientific artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.32, + 0.207, + 0.333 + ], + "angle": 0, + "content": "Section 4" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.344, + 0.53, + 0.375 + ], + "angle": 0, + "content": "B1. Did you cite the creators of artifacts you used? Section 4" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.386, + 0.779, + 0.419 + ], + "angle": 0, + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Appendix A" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.429, + 0.881, + 0.507 + ], + "angle": 0, + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Section 4" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.521, + 0.881, + 0.601 + ], + "angle": 0, + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? We did not discuss this as the datasets are commonly used NLP benchmarks that do not contain personal data." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.611, + 0.881, + 0.659 + ], + "angle": 0, + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? We discuss this in Section 8, the limitations section. We discuss the coverage of domains in Section 4." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.669, + 0.881, + 0.765 + ], + "angle": 0, + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. We discuss this in Section 4." + }, + { + "type": "list", + "bbox": [ + 0.13, + 0.344, + 0.881, + 0.765 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.775, + 0.494, + 0.792 + ], + "angle": 0, + "content": "C Did you run computational experiments?" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.797, + 0.259, + 0.811 + ], + "angle": 0, + "content": "Sections 4 and 5" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.821, + 0.881, + 0.887 + ], + "angle": 0, + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? We report the model specifics in section 4. We describe the computing infrastructure in Appendix 2, but do not estimate the computational budget." + }, + { + "type": "footer", + "bbox": [ + 0.114, + 0.894, + 0.878, + 0.917 + ], + "angle": 0, + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1780" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.13, + 0.084, + 0.88, + 0.131 + ], + "angle": 0, + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Section 4" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.143, + 0.881, + 0.207 + ], + "angle": 0, + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Sections 4 and 5" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.219, + 0.881, + 0.283 + ], + "angle": 0, + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Not applicable. Left blank." + }, + { + "type": "list", + "bbox": [ + 0.129, + 0.084, + 0.881, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.294, + 0.878, + 0.331 + ], + "angle": 0, + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.341, + 0.881, + 0.389 + ], + "angle": 0, + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.4, + 0.881, + 0.464 + ], + "angle": 0, + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.476, + 0.881, + 0.54 + ], + "angle": 0, + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.55, + 0.873, + 0.582 + ], + "angle": 0, + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.593, + 0.881, + 0.642 + ], + "angle": 0, + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Not applicable. Left blank." + }, + { + "type": "list", + "bbox": [ + 0.128, + 0.341, + 0.881, + 0.642 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.519, + 0.941 + ], + "angle": 0, + "content": "1781" + } + ] +] \ No newline at end of file diff --git a/2023/Teaching Small Language Models to Reason/88702c47-13fb-46e6-923c-6b020286bf7a_origin.pdf b/2023/Teaching Small Language Models to Reason/88702c47-13fb-46e6-923c-6b020286bf7a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..48ca9424a406add6e8002d573e46402e5c6ffbf2 --- /dev/null +++ b/2023/Teaching Small Language Models to Reason/88702c47-13fb-46e6-923c-6b020286bf7a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:49f059e501c1cd8cf10ede3e65df1918d4125a217b429d998cc501402fcb29af +size 241686 diff --git a/2023/Teaching Small Language Models to Reason/full.md b/2023/Teaching Small Language Models to Reason/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a71ce07b7ef73da68bfb5a40824962c0ceb53217 --- /dev/null +++ b/2023/Teaching Small Language Models to Reason/full.md @@ -0,0 +1,252 @@ +# Teaching Small Language Models to Reason + +Lucie Charlotte Magister* + +University of Cambridge + +1cm67@cam.ac.uk + +Jonathan Mallinson + +Google Research + +jonmall@google.com + +Jakub Adamek + +Google Research + +enkait@google.com + +# Eric Malmi + +Google Research + +emailmi@google.com + +# Aliaksei Severyn + +Google Research + +severyn@google.com + +# Abstract + +Chain of thought prompting successfully improves the reasoning capabilities of large language models, achieving state of the art results on a range of datasets. However, these reasoning capabilities only appear to emerge in models with at least tens of billions of parameters. In this paper, we explore the transfer of such reasoning capabilities to smaller models via knowledge distillation, also investigating model and dataset size trade-off. Specifically, we finetune a student model on the chain of thought outputs generated by a larger teacher model. Our experiments show that the proposed method improves task performance across arithmetic, commonsense and symbolic reasoning datasets. For example, the accuracy of T5 XXL on GSM8K improves from $8.11\%$ to $21.99\%$ and $18.42\%$ when finetuned on PaLM 540B and GPT-3 175B generated chains of thought, respectively. + +# 1 Introduction + +Chain of thought (CoT) prompting encourages language models (LMs) to break down a reasoning task into a series of intermediate steps (Wei et al., 2022). They demonstrate that this prompting significantly increases the task accuracy of large language models (LLMs) across commonsense, symbolic and mathematical reasoning datasets. Here, LLMs are models with at least tens of billions of parameters, such as PaLM 540B (Chowdhery et al., 2022), GPT-3 175B (Brown et al., 2020), or UL2 20B (Tay et al., 2022). However, the reasoning capabilities of smaller LMs do not improve with CoT prompting, mostly producing illogical CoT. Notably, CoT prompting even reduces the accuracy of models with less than 10 billion parameters. Wei et al. (2022) attribute this to abilities, such as semantic understanding and symbolic mapping, only emerging at larger scales. This leads us to our re + +search question: can the reasoning capabilities of LLMs be transferred to smaller LMs via finetuning? + +This work explores CoT knowledge distillation (Hinton et al., 2015) from PaLM 540B (Chowdhery et al., 2022) and GPT-3 175B (Brown et al., 2020) to different sizes of the smaller language model T5 (Raffel et al., 2020), such as T5 XXL, XL and base, which have 11 billion, 3 billion and 220 million parameters, respectively. As a result of our work, we make two recommendations: (1) perform knowledge distillation by finetuning the student model on the CoT generated by a large teacher model; and (2) generate the CoT from an LLM, as proposed by Wei et al. (2022), but crucially provide the solution to the task in the few-shot prompt. We demonstrate that the proposed method improves task performance across arithmetic, commonsense and symbolic reasoning datasets irrespective of the teacher model used. For example, we show an accuracy increase from $8.11\%$ to $21.99\%$ and $18.42\%$ on the GSM8K (Cobbe et al., 2021) dataset when finetuning T5 XXL on PaLM 540B and GPT-3 175B generated CoT data, respectively. + +# 2 Related Work + +This work is inspired by the seminal work of Wei et al. (2022) on CoT prompting. They demonstrate that prefixing an input with 2-8 exemplars of CoT reasoning encourages LMs to do the same, reaching state-of-the-art performance on datasets such as GSM8K (Cobbe et al., 2021). Wang et al. (2022) show that task accuracy can be further improved by using self-consistency in CoT prompting. Self-consistency samples CoT reasoning paths from a model's decoder and returns the most consistent path by taking the majority vote. Subsequently, Chung et al. (2022) explore finetuning a FLAN-based (Wei et al., 2021) version of PaLM on manually generated CoT data. + +Concurrent to our work, a small number of other works propose methods focused on CoT student- + +teacher knowledge distillation. Ho et al. (2022) and Li et al. (2022) also explore knowledge distillation with the difference of proposing diverse sampling and rationalization prompting, respectively. In contrast to their work, our work explores more teacher models and demonstrates both the effects of dataset and model size on accuracy. We also achieve a higher accuracy on common datasets, such as GSM8K, than Ho et al. (2022). In contrast to our work, Shridhar et al. (2022) focus on training two models, one for problem decomposition and one for solving. Yet differently, the focus of Eisenstein et al. (2022) relies on producing markup-and-mask explanations for open-book question answering. Lastly, Huang et al. (2022) present one related experiment, however, we present a more in-depth exploration on more datasets. To the best of our knowledge, our work is the first to extensively explore the improvement of the reasoning ability of small LMs via knowledge distillation across multiple model architectures, and observing the effects of student model size and dataset size on accuracy. + +# 3 Method + +We propose a two-step pipeline for CoT knowledge distillation. The first step comprises annotating an existing supervised dataset with CoT reasoning generated by a teacher model. To generate high quality data, we propose using LLMs, such as PaLM 540B or GPT-3 175B, as teachers, based on the finding that CoT reasoning improves with model scale (Wei et al., 2022). Specifically, we perform few-shot prompting with 8 exemplars on these models to generate CoTs. However, we make a key modification to the prompts proposed by Wei et al. (2022). We adapt the few-shot prompts to provide the model with the target after posing the question and before providing example CoT. This is based on the observation that providing this guidance allows LLMs to correct small mistakes in the CoT. Lastly, we remove all incorrect CoT based on the target answer to prevent the student to learn from bad examples. The second step comprises finetuning a student model via teacher forcing (Williams and Zipser, 1989). The student is provided with the question as input, and the CoT and answer as the target. As the model is trained on producing a CoT during finetuning, prompting is not required. Figure 1 provides an overview of the proposed method. + +![](images/e48afad38d228e4820b40a7fb5a37dfe0cb0633ef7e7a40866fd7553ce4a3c2a.jpg) +Figure 1: Overview of the proposed method. + +# 4 Experimental Setup + +We follow a similar experimental setup to Wei et al. (2022), focusing on tasks covering arithmetic, commonsense and symbolic reasoning. + +# 4.1 Benchmarks and Metrics + +# 4.1.1 Arithmetic Reasoning + +We benchmark the proposed method on the following math word problem datasets: (1) GSM8K (Cobbe et al., 2021), (2) MAWPS (Koncel-Kedziorski et al., 2016) and (3) ASDiv (Miao et al., 2021). We use the official training and testing split for GSM8K, taking the last $10\%$ of the training split for validation, and the 5-fold cross validation splits available for MAWPS and ASDiv. We evaluate task accuracy by checking for the target answer as the final answer in the CoT. In addition, we compute the task accuracy given an external calculator, to account for arithmetic mistakes made by the model, despite the CoT being correct. The external calculator moves through the generated output, recalculating the left-hand-side of equations. It then replaces the right-hand side with the calculated output, to avoid arithmetic mistakes being carried forward. For example, if a model outputted $^{\prime}5 + 5 = 11$ . $11*2 = 22'$ , then the external calculator would first calculate $^{\prime}5 + 5$ and replace the '11' with a '10'. In the subsequent equation, it would + +also replace the '11' with a '10' and arrive at the final result of '20'. + +# 4.1.2 Commonsense Reasoning + +We benchmark the model's ability to perform commonsense reasoning on the StrategyQA dataset (Geva et al., 2021a). As a testing split is not available, we do not shuffle the dataset to allow reproducing our split of taking the first $80\%$ as training data, the following $10\%$ as validation data, and the final $10\%$ as testing data. We compute task accuracy in the same manner as previously mentioned. + +# 4.1.3 Symbolic Reasoning + +Lastly, we benchmark the model on two synthetic tasks for symbolic reasoning: (1) last letter concatenation and (2) coinflip (Wei et al., 2022). Last letter concatenation prompts the model to concatenate the last letter of each word in a string. Coinflip prompts the model to perform state tracking of the coin being flipped. We evaluate task accuracy in the same manner as before. Due to the rigid structure of the datasets, we focus on evaluating the model's generalizability to out-of-distribution (OOD) examples. We finetune the models on examples of length two and evaluate on sequences of length three and four. We initially infer the CoT using PaLM 540B, however, find that the LLM is able to perfectly replicate the desired CoT bar one example due to the rigidity of the template. We therefore decide to use the template generated CoT in our experiments. + +# 4.2 Baselines and setup + +We select PaLM 540B (Chowdhery et al., 2022) and GPT-3 175B (Brown et al., 2020) as teacher models. We select PaLM 540B based on the state-of-the-art results on the benchmarking datasets reported by Wei et al. (2022), and confirm the observed trends with GPT-3 175B. The publicly accessible teacher models are prompted as described in Section 3. + +We select different sizes of T5 (Raffel et al., 2020) as student models, as T5 is publicly available in many sizes. The student models are trained on the PaLM 540B or GPT-3 175B generated CoT data as described in Section 3. We establish T5 XXL model finetuned on the original target as the baseline. We refrain from shuffling the datasets to allow for reproducibility. For the MAWPS and ASDiv dataset, we perform 5-fold cross validation. For all remaining datasets, we take $10\%$ of the + +Input: Roger has 5 tennis balls. He buys 2 more cans of tennis balls. Each can has 3 tennis balls. How many tennis balls does he have now? + +Output: Roger started with 5 balls. 2 cans of 3 tennis balls each is 6 tennis balls. $5 + 6 = 11$ . The answer is 11. + +Figure 2: A training example from Wei et al. (2022) demonstrating the input and output provided to T5. + +training set as a validation set to select the best model checkpoint. Figure 2 showcases an input examples for T5. We refer the reader to Wei et al. (2022) for more training examples, as well as the prompts used for generating the CoT using PaLM 540B and GPT-3 175B. + +We refer the reader to Appendix A for an overview of the dataset licenses. We also refer the reader to Appendix B for an overview of the computational resources. + +# 5 Results + +# 5.1 Arithmetic reasoning + +Table 1 details the task accuracy with and without an external calculator for the arithmetic reasoning benchmarks. Our results show that the proposed method improves task accuracy across all datasets. Most notably, the task accuracy of MAwPS is significantly improved. The accuracy achieved given a calculator comes close to the accuracy of 8-shot PaLM 540B, demonstrating that knowledge distillation is effective, but potentially limited by the mathematical abilities of small models. + +
Baseline T5 XXLCoT Finetuned T5 XXLCoT 8-shot PaLM 540B
Acc.Acc.Acc. with Calc.Acc.Acc. with Calc.
GSM8K8.1121.9938.2156.9058.60
Dataset Size672553375337--
MAWPS54.1570.4188.2293.0093.66
Dataset Size159015901590--
ASDiv39.6442.1260.7373.972.6
Dataset Size184415441544--
+ +Table 1: Task accuracy across arithmetic reasoning datasets for T5 XXL without finetuning (baseline) and finetuned on PaLM 540B generated chain-of-thought (CoT). We report the accuracy of PaLM 540B on the used datasets for reference. We do not finetune PaLM for this, but employ 8 chain of thought prompts. + +# 5.1.1 Ablation study on generating chain-of-thought data + +We perform an ablation study to confirm that providing a LLM with the target during CoT generation is beneficial. We found that for the GSM8K dataset, PaLM 540B only achieves a $59.98\%$ accuracy if prompted without the target. In comparison, when including the target in the prompt the accuracy is $79.37\%$ . A superficial explanation would be that when the model is conditioned on the expected answer, it produces the same CoT but copies the answer. However, an analysis of a subset of the differences between CoT produced with and without this conditioning shows that most of the benefits actually come from the model correcting CoT that had a single step missing or was wrong. + +# 5.2 Commonsense reasoning + +For the StrategyQA dataset (Table 3), we found that using CoT finetuning improves accuracy from $68.12\%$ to $71.98\%$ , using only 1319 of the original 1648 examples. Compared to the arithmetic reasoning datasets, the improvement is not as significant. This can be explained by the model lacking factual knowledge that the dataset requires. The task is heavily focused on the model reasoning on such knowledge, however, a smaller LM is most likely not in possession of this knowledge compared to a larger model with higher memorisation capacity. + +# 5.3 Symbolic reasoning + +Table 2 shows the results obtained for the synthetic symbolic reasoning datasets, focusing on OOD generalization. Focusing on Last Letter Concatenation, it can be stated that both traditional finetuning and the suggested method fail at generalizing to a longer sequence length. In comparison, the proposed method significantly increases accuracy for the Coinflip dataset with regard to generalizing to three coinflips. In contrast, generalisation to four coinflips is slightly weaker than the baseline, which performs very strongly. This may be related to the task length being twice that of the training task. + +# 5.4 Replicating Results using different Teacher Models + +We demonstrate the robustness of our method using a different teacher model, namely GPT-3 175B. Table 3 shows the results for GSM8K and StrategyQA when T5 XXL is finetuned on CoT data generated by GPT-3. The results show that the proposed method elicits improvements also with other + +
Baseline T5 XXLCoT Finetuned T5 XXLCoT 8-shot PaLM 540B
Last LetterOOD: 30.000.0094.8
Concat.OOD: 40.000.0063.0
CoinflipOOD: 313.1086.7098.6
OOD: 473.8070.5090.2
+ +LLMs as teachers. We also report the accuracy of T5 XXL finetuned on golden CoT provided with the datasets. For the StrategyQA dataset, the model finetuned on the golden CoT performs best, which may be attributed to the dataset being the largest, as both PaLM and GPT-3 get some examples wrong. In contrast, the model finetuned on PaLM generated CoT performs the best for GSM8K. + +Table 2: Task accuracy across the symbolic reasoning datasets for T5 XXL finetuned on chain-of-thought (CoT) data. For each dataset, there are 1000 training and testing examples. We report the accuracy of PaLM 540B from (Wei et al., 2022) for reference. + +
Base TaskOriginal CotCoT finetuned T5 XXL using PaLM 540BGPT-3 175BCoT 8-Shot PaLM 540BGPT-3 175B
GSM8K8.1119.9421.9918.4256.946.9
acc. with Calc.-26.9938.2133.0658.649.6
Dataset Size6725672553375298--
StrategyQA68.1271.9867.1563.7777.865.4
Dataset Size1648164813191319--
+ +Table 3: Task accuracy for T5 XXL finetuned on chain-of-thought (CoT) data generated by PaLM 540B and GPT-3 175B. We also finetune on the reasoning steps provided by the datasets. We report the accuracy of PaLM 540B on the used datasets for reference. We do not finetune PaLM for this, but employ 8 chain of thought prompts. + +# 5.5 Ablation study on model size + +We investigate the performance gain achieved via finetuning student models of different sizes. Figure 3 shows the performance gain achieved when finetuning T5 of different sizes on the GSM8K dataset. Our results show that T5 base, with 44 times fewer parameters than T5 XXL, matches the performance of the baseline T5 XXL when trained on CoT data. Moreover, given an external calculator, even T5 small outperforms the baseline T5 XXL. + +# 5.6 Ablation study on dataset size + +We also investigate the trade-off between the performance gain from CoT finetuning and dataset size. Table 4 details the test accuracy achieved when finetuning T5 XXL on only $4\%$ and $20\%$ of the data, randomly selected. In comparison to the + +![](images/cd13943609b671f1f35d570693adefd8c778ebba2260743db1aaa7f486ecbc0e.jpg) +Figure 3: Effect of student model (T5) size on accuracy on GSM8K. + +baseline accuracy of $8.11\%$ (Table 3), we see that our method is 6x more data efficient, achieving accuracy of $11.22\%$ with only $20\%$ of the examples. However, training on just $20\%$ of the data still creates a quality gap, and it's possible that with e.g. $200\%$ larger dataset we could outperform the results in Table 3. + +
Percentage of GSM8K data used to trainCoT finetuned T5 XXL
Acc.Acc. with Calc.
4% (213 examples)6.2912.28
20% (1067 examples)11.2220.47
100% (5337 examples)21.9938.21
+ +Table 4: Task accuracy of T5 XXL finetuned on different amounts of chain-of-thought (CoT) data generated by PaLM 540B. + +# 6 Discussion + +We demonstrate that finetuning larger LMs on the CoT data generated by LLMs of over 100 billion parameters can significantly improve task accuracy. Even a small number of CoT examples appear to suffice for this. However, such improvements appear to be task dependent. For example, the effects are limited for the StrategyQA dataset, which can be attributed to the task requiring specific factual knowledge, which smaller LMs may not have memorised due to their limited capacity. Nevertheless, there is some performance improvement, which may be attributed to the model learning how to approach such tasks. Moreover, the CoT knowledge distillation pipeline presented allows to trade-off model and dataset size with accuracy. Future work could explore improving the reasoning of small + +models in multi-task settings, as well as the generation of new training data using LLMs, rather than annotating existing datasets. + +# 7 Conclusion + +This work explores CoT knowledge distillation from LLMs of over 100 billion parameters to smaller LMs. We propose a knowledge distillation pipeline consisting of two keys steps: (1) generate CoT for existing datasets using LLMs and (2) finetune smaller LMs on the CoT. Our results demonstrate that finetuning on CoT improves task accuracy across a range of benchmarking datasets. + +# 8 Limitations + +The results we present must be viewed in the context of a few limitations. A limitation is that we only perform experiments in English and on one task at a time. To be more comparable to a LLM few-shot settings, other languages and a multi-task setup could be explored. Furthermore, in order to replicate the results access to none public models is required and inference must be performed on large amounts of data. Another limitation of our work is that it only explores the original CoT prompting approach, but we do not explore subsequent improvements, such a self-consistency (Wang et al., 2022). + +# 9 Ethical Considerations + +The main ethical considerations of our research arise from the text generation performed. The concerns here are that both the teacher and student model may potentially generate non-factual (Ji et al., 2022; Pagnoni et al., 2021; Kreps et al., 2022) or offensive output (Gehman et al., 2020). This is largely influenced by the input data, which is our case are standard, peer-reviewed benchmarking tasks in the NLP domain. + +# References + +BIG-bench collaboration. 2021. Beyond the imitation game: Measuring and extrapolating the capabilities of language models. In preparation. +Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901. + +Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. 2022. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311. +Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. 2022. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416. +Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. 2021. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168. +Jacob Eisenstein, Daniel Andor, Bernd Bohnet, Michael Collins, and David Mimno. 2022. Honest students from untrusted teachers: Learning an interpretable question-answering pipeline from a pretrained language model. arXiv preprint arXiv:2210.02498. +Samuel Gehman, Suchin Gururangan, Maarten Sap, Yejin Choi, and Noah A Smith. 2020. Realtoxicityprompts: Evaluating neural toxic degeneration in language models. arXiv preprint arXiv:2009.11462. +Mor Geva, Daniel Khashabi, Elad Segal, Tushar Khot, Dan Roth, and Jonathan Berant. 2021a. Did aristotle use a laptop? a question answering benchmark with implicit reasoning strategies. Transactions of the Association for Computational Linguistics, 9:346-361. +Mor Geva, Daniel Khashabi, Elad Segal, Tushar Khot, Dan Roth, and Jonathan Berant. 2021b. Did aristotle use a laptop? a question answering benchmark with implicit reasoning strategies. Transactions of the Association for Computational Linguistics, 9:346-361. +Geoffrey Hinton, Oriol Vinyals, Jeff Dean, et al. 2015. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2(7). +Namgyu Ho, Laura Schmid, and Se-Young Yun. 2022. Large language models are reasoning teachers. arXiv preprint arXiv:2212.10071. +Jiaxin Huang, Shixiang Shane Gu, Le Hou, Yuexin Wu, Xuezhi Wang, Hongkun Yu, and Jiawei Han. 2022. Large language models can self-improve. arXiv preprint arXiv:2210.11610. +Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Yejin Bang, Andrea Madotto, and Pascale Fung. 2022. Survey of hallucination in natural language generation. ACM Computing Surveys. +Rik Koncel-Kedziorski, Subhro Roy, Aida Amini, Nate Kushman, and Hannaneh Hajishirzi. 2016. Mawps: A math word problem repository. In Proceedings of + +the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1152-1157. +Sarah Kreps, R Miles McCain, and Miles Brundage. 2022. All the news that's fit to fabricate: A-generated text as a tool of media misinformation. Journal of Experimental Political Science, 9(1):104-117. +Shiyang Li, Jianshu Chen, Yelong Shen, Zhiyu Chen, Xinlu Zhang, Zekun Li, Hong Wang, Jing Qian, Baolin Peng, Yi Mao, et al. 2022. Explanations from large language models make small reasoners better. arXiv preprint arXiv:2210.06726. +Shen-Yun Miao, Chao-Chun Liang, and Keh-Yih Su. 2021. A diverse corpus for evaluating and developing english math word problem solvers. arXiv preprint arXiv:2106.15772. +Artidoro Pagnoni, Vidhisha Balachandran, and Yulia Tsvetkov. 2021. Understanding factuality in abstractive summarization with frank: A benchmark for factuality metrics. arXiv preprint arXiv:2104.13346. +Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J Liu, et al. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(140):1-67. +Kumar Shridhar, Alessandro Stolfo, and Mrinmaya Sachan. 2022. Distilling multi-step reasoning capabilities of large language models into smaller models via semantic decompositions. arXiv preprint arXiv:2212.00193. +Yi Tay, Mostafa Dehghani, Vinh Q Tran, Xavier Garcia, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Neil Houlsby, and Donald Metzler. 2022. Unifying language learning paradigms. arXiv preprint arXiv:2205.05131. +Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, and Denny Zhou. 2022. Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171. +Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. 2021. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652. +Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed Chi, Quoc Le, and Denny Zhou. 2022. Chain of thought prompting elicits reasoning in large language models. arXiv preprint arXiv:2201.11903. +Ronald J Williams and David Zipser. 1989. A learning algorithm for continually running fully recurrent neural networks. *Neural computation*, 1(2):270-280. + +# A Dataset Usage and Licenses + +In this section, we list the licenses for the datasets used and any ethical concerns regarding their usage. We describe the dataset splits used for all datasets in Section 4 of the paper. + +# A.1 Arithmetic Reasoning + +The GSM8K dataset (Cobbe et al., 2021) is available under the MIT license. The MAWPS dataset (Koncel-Kedziorski et al., 2016) is available under the CC BY 4.0 and the ASDiv dataset (Miao et al., 2021) is available under the CC BY-NC 4.0 license. We follow the intended usage of the datasets. + +# A.2 Commonsense Reasoning + +The StrategyQA dataset (Geva et al., 2021b) is available under the MIT license. Similar to Wei et al. (2022), we use the open-domain setting version available as part of the Big-bench collaboration (BIG-bench collaboration, 2021), available under the Apache License 2.0. We follow the intended usage of the datasets. + +# A.3 Symbolic Reasoning + +We generate the symbolic reasoning datasets as described in Wei et al. (2022). + +# B Computational Resources + +We perform inference and finetuning on different sizes of T5 on TPUs. We perform inference on PaLM 540B also on TPUs. Our results can be replicated via the public API (https://developersgenerativeai.google/products/palm). To make requests to GPT-3 175B, we use the public API (https://beta.openai.com/docs/introduction). + +A For every submission: + +A1. Did you describe the limitations of your work? Section 8 +A2. Did you discuss any potential risks of your work? Section 9 +A3. Do the abstract and introduction summarize the paper's main claims? Abstract and Section 1 +A4. Have you used AI writing assistants when working on this paper? Left blank. + +B Did you use or create scientific artifacts? + +Section 4 + +B1. Did you cite the creators of artifacts you used? Section 4 +B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Appendix A +B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Section 4 +B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? We did not discuss this as the datasets are commonly used NLP benchmarks that do not contain personal data. +B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? We discuss this in Section 8, the limitations section. We discuss the coverage of domains in Section 4. +B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. We discuss this in Section 4. + +C Did you run computational experiments? + +Sections 4 and 5 + +C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? We report the model specifics in section 4. We describe the computing infrastructure in Appendix 2, but do not estimate the computational budget. + +C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Section 4 +C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Sections 4 and 5 +C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Not applicable. Left blank. + +D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank. + +D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Not applicable. Left blank. +D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Not applicable. Left blank. +D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Not applicable. Left blank. +D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Not applicable. Left blank. +D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Not applicable. Left blank. \ No newline at end of file diff --git a/2023/Teaching Small Language Models to Reason/images.zip b/2023/Teaching Small Language Models to Reason/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2bfb582407dd8e4679f749b8992b6abb8e15a335 --- /dev/null +++ b/2023/Teaching Small Language Models to Reason/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:03f6d928ff2b5fab66e1a92b01d2efadb6a780cef4bf210f411c7db7a8a095cd +size 158076 diff --git a/2023/Teaching Small Language Models to Reason/layout.json b/2023/Teaching Small Language Models to Reason/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c508687144151673155848537e70848b6124ba17 --- /dev/null +++ b/2023/Teaching Small Language Models to Reason/layout.json @@ -0,0 +1,6112 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 160, + 71, + 433, + 88 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 71, + 433, + 88 + ], + "spans": [ + { + "bbox": [ + 160, + 71, + 433, + 88 + ], + "type": "text", + "content": "Teaching Small Language Models to Reason" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 83, + 96, + 218, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 96, + 218, + 111 + ], + "spans": [ + { + "bbox": [ + 83, + 96, + 218, + 111 + ], + "type": "text", + "content": "Lucie Charlotte Magister*" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 88, + 111, + 209, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 111, + 209, + 125 + ], + "spans": [ + { + "bbox": [ + 88, + 111, + 209, + 125 + ], + "type": "text", + "content": "University of Cambridge" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 95, + 126, + 203, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 126, + 203, + 137 + ], + "spans": [ + { + "bbox": [ + 95, + 126, + 203, + 137 + ], + "type": "text", + "content": "1cm67@cam.ac.uk" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 244, + 97, + 348, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 97, + 348, + 110 + ], + "spans": [ + { + "bbox": [ + 244, + 97, + 348, + 110 + ], + "type": "text", + "content": "Jonathan Mallinson" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 254, + 111, + 339, + 124 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 111, + 339, + 124 + ], + "spans": [ + { + "bbox": [ + 254, + 111, + 339, + 124 + ], + "type": "text", + "content": "Google Research" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 233, + 126, + 362, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 126, + 362, + 137 + ], + "spans": [ + { + "bbox": [ + 233, + 126, + 362, + 137 + ], + "type": "text", + "content": "jonmall@google.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 406, + 97, + 486, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 97, + 486, + 110 + ], + "spans": [ + { + "bbox": [ + 406, + 97, + 486, + 110 + ], + "type": "text", + "content": "Jakub Adamek" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 404, + 111, + 488, + 124 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 404, + 111, + 488, + 124 + ], + "spans": [ + { + "bbox": [ + 404, + 111, + 488, + 124 + ], + "type": "text", + "content": "Google Research" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 384, + 126, + 508, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 126, + 508, + 137 + ], + "spans": [ + { + "bbox": [ + 384, + 126, + 508, + 137 + ], + "type": "text", + "content": "enkait@google.com" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 166, + 152, + 227, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 152, + 227, + 164 + ], + "spans": [ + { + "bbox": [ + 166, + 152, + 227, + 164 + ], + "type": "text", + "content": "Eric Malmi" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 155, + 166, + 239, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 166, + 239, + 179 + ], + "spans": [ + { + "bbox": [ + 155, + 166, + 239, + 179 + ], + "type": "text", + "content": "Google Research" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 136, + 181, + 259, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 181, + 259, + 193 + ], + "spans": [ + { + "bbox": [ + 136, + 181, + 259, + 193 + ], + "type": "text", + "content": "emailmi@google.com" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 352, + 152, + 440, + 165 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 152, + 440, + 165 + ], + "spans": [ + { + "bbox": [ + 352, + 152, + 440, + 165 + ], + "type": "text", + "content": "Aliaksei Severyn" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 355, + 166, + 438, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 166, + 438, + 179 + ], + "spans": [ + { + "bbox": [ + 355, + 166, + 438, + 179 + ], + "type": "text", + "content": "Google Research" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 332, + 181, + 462, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 181, + 462, + 193 + ], + "spans": [ + { + "bbox": [ + 332, + 181, + 462, + 193 + ], + "type": "text", + "content": "severyn@google.com" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 155, + 212, + 202, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 212, + 202, + 225 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 202, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 84, + 237, + 274, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 237, + 274, + 476 + ], + "spans": [ + { + "bbox": [ + 84, + 237, + 274, + 476 + ], + "type": "text", + "content": "Chain of thought prompting successfully improves the reasoning capabilities of large language models, achieving state of the art results on a range of datasets. However, these reasoning capabilities only appear to emerge in models with at least tens of billions of parameters. In this paper, we explore the transfer of such reasoning capabilities to smaller models via knowledge distillation, also investigating model and dataset size trade-off. Specifically, we finetune a student model on the chain of thought outputs generated by a larger teacher model. Our experiments show that the proposed method improves task performance across arithmetic, commonsense and symbolic reasoning datasets. For example, the accuracy of T5 XXL on GSM8K improves from " + }, + { + "bbox": [ + 84, + 237, + 274, + 476 + ], + "type": "inline_equation", + "content": "8.11\\%" + }, + { + "bbox": [ + 84, + 237, + 274, + 476 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 84, + 237, + 274, + 476 + ], + "type": "inline_equation", + "content": "21.99\\%" + }, + { + "bbox": [ + 84, + 237, + 274, + 476 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 84, + 237, + 274, + 476 + ], + "type": "inline_equation", + "content": "18.42\\%" + }, + { + "bbox": [ + 84, + 237, + 274, + 476 + ], + "type": "text", + "content": " when finetuned on PaLM 540B and GPT-3 175B generated chains of thought, respectively." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 68, + 488, + 154, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 488, + 154, + 500 + ], + "spans": [ + { + "bbox": [ + 68, + 488, + 154, + 500 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 67, + 510, + 291, + 754 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 510, + 291, + 754 + ], + "spans": [ + { + "bbox": [ + 67, + 510, + 291, + 754 + ], + "type": "text", + "content": "Chain of thought (CoT) prompting encourages language models (LMs) to break down a reasoning task into a series of intermediate steps (Wei et al., 2022). They demonstrate that this prompting significantly increases the task accuracy of large language models (LLMs) across commonsense, symbolic and mathematical reasoning datasets. Here, LLMs are models with at least tens of billions of parameters, such as PaLM 540B (Chowdhery et al., 2022), GPT-3 175B (Brown et al., 2020), or UL2 20B (Tay et al., 2022). However, the reasoning capabilities of smaller LMs do not improve with CoT prompting, mostly producing illogical CoT. Notably, CoT prompting even reduces the accuracy of models with less than 10 billion parameters. Wei et al. (2022) attribute this to abilities, such as semantic understanding and symbolic mapping, only emerging at larger scales. This leads us to our re" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 302, + 213, + 526, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 213, + 526, + 239 + ], + "spans": [ + { + "bbox": [ + 302, + 213, + 526, + 239 + ], + "type": "text", + "content": "search question: can the reasoning capabilities of LLMs be transferred to smaller LMs via finetuning?" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 302, + 240, + 526, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 240, + 526, + 524 + ], + "spans": [ + { + "bbox": [ + 302, + 240, + 526, + 524 + ], + "type": "text", + "content": "This work explores CoT knowledge distillation (Hinton et al., 2015) from PaLM 540B (Chowdhery et al., 2022) and GPT-3 175B (Brown et al., 2020) to different sizes of the smaller language model T5 (Raffel et al., 2020), such as T5 XXL, XL and base, which have 11 billion, 3 billion and 220 million parameters, respectively. As a result of our work, we make two recommendations: (1) perform knowledge distillation by finetuning the student model on the CoT generated by a large teacher model; and (2) generate the CoT from an LLM, as proposed by Wei et al. (2022), but crucially provide the solution to the task in the few-shot prompt. We demonstrate that the proposed method improves task performance across arithmetic, commonsense and symbolic reasoning datasets irrespective of the teacher model used. For example, we show an accuracy increase from " + }, + { + "bbox": [ + 302, + 240, + 526, + 524 + ], + "type": "inline_equation", + "content": "8.11\\%" + }, + { + "bbox": [ + 302, + 240, + 526, + 524 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 302, + 240, + 526, + 524 + ], + "type": "inline_equation", + "content": "21.99\\%" + }, + { + "bbox": [ + 302, + 240, + 526, + 524 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 302, + 240, + 526, + 524 + ], + "type": "inline_equation", + "content": "18.42\\%" + }, + { + "bbox": [ + 302, + 240, + 526, + 524 + ], + "type": "text", + "content": " on the GSM8K (Cobbe et al., 2021) dataset when finetuning T5 XXL on PaLM 540B and GPT-3 175B generated CoT data, respectively." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 303, + 534, + 395, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 534, + 395, + 547 + ], + "spans": [ + { + "bbox": [ + 303, + 534, + 395, + 547 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 302, + 556, + 526, + 745 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 556, + 526, + 745 + ], + "spans": [ + { + "bbox": [ + 302, + 556, + 526, + 745 + ], + "type": "text", + "content": "This work is inspired by the seminal work of Wei et al. (2022) on CoT prompting. They demonstrate that prefixing an input with 2-8 exemplars of CoT reasoning encourages LMs to do the same, reaching state-of-the-art performance on datasets such as GSM8K (Cobbe et al., 2021). Wang et al. (2022) show that task accuracy can be further improved by using self-consistency in CoT prompting. Self-consistency samples CoT reasoning paths from a model's decoder and returns the most consistent path by taking the majority vote. Subsequently, Chung et al. (2022) explore finetuning a FLAN-based (Wei et al., 2021) version of PaLM on manually generated CoT data." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 302, + 746, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 746, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 746, + 526, + 772 + ], + "type": "text", + "content": "Concurrent to our work, a small number of other works propose methods focused on CoT student-" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 84, + 761, + 277, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 761, + 277, + 772 + ], + "spans": [ + { + "bbox": [ + 84, + 761, + 277, + 772 + ], + "type": "text", + "content": "*Research conducted during an internship at Google." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "1773" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 135, + 795, + 458, + 806 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 795, + 458, + 806 + ], + "spans": [ + { + "bbox": [ + 135, + 795, + 458, + 806 + ], + "type": "text", + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 219, + 807, + 374, + 817 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 807, + 374, + 817 + ], + "spans": [ + { + "bbox": [ + 219, + 807, + 374, + 817 + ], + "type": "text", + "content": "Volume 2: Short Papers, pages 1773-1781" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "spans": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "text", + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 293, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 293, + 370 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 293, + 370 + ], + "type": "text", + "content": "teacher knowledge distillation. Ho et al. (2022) and Li et al. (2022) also explore knowledge distillation with the difference of proposing diverse sampling and rationalization prompting, respectively. In contrast to their work, our work explores more teacher models and demonstrates both the effects of dataset and model size on accuracy. We also achieve a higher accuracy on common datasets, such as GSM8K, than Ho et al. (2022). In contrast to our work, Shridhar et al. (2022) focus on training two models, one for problem decomposition and one for solving. Yet differently, the focus of Eisenstein et al. (2022) relies on producing markup-and-mask explanations for open-book question answering. Lastly, Huang et al. (2022) present one related experiment, however, we present a more in-depth exploration on more datasets. To the best of our knowledge, our work is the first to extensively explore the improvement of the reasoning ability of small LMs via knowledge distillation across multiple model architectures, and observing the effects of student model size and dataset size on accuracy." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 391, + 130, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 391, + 130, + 403 + ], + "spans": [ + { + "bbox": [ + 68, + 391, + 130, + 403 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 421, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 421, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 421, + 291, + 772 + ], + "type": "text", + "content": "We propose a two-step pipeline for CoT knowledge distillation. The first step comprises annotating an existing supervised dataset with CoT reasoning generated by a teacher model. To generate high quality data, we propose using LLMs, such as PaLM 540B or GPT-3 175B, as teachers, based on the finding that CoT reasoning improves with model scale (Wei et al., 2022). Specifically, we perform few-shot prompting with 8 exemplars on these models to generate CoTs. However, we make a key modification to the prompts proposed by Wei et al. (2022). We adapt the few-shot prompts to provide the model with the target after posing the question and before providing example CoT. This is based on the observation that providing this guidance allows LLMs to correct small mistakes in the CoT. Lastly, we remove all incorrect CoT based on the target answer to prevent the student to learn from bad examples. The second step comprises finetuning a student model via teacher forcing (Williams and Zipser, 1989). The student is provided with the question as input, and the CoT and answer as the target. As the model is trained on producing a CoT during finetuning, prompting is not required. Figure 1 provides an overview of the proposed method." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 303, + 72, + 526, + 333 + ], + "blocks": [ + { + "bbox": [ + 303, + 72, + 526, + 333 + ], + "lines": [ + { + "bbox": [ + 303, + 72, + 526, + 333 + ], + "spans": [ + { + "bbox": [ + 303, + 72, + 526, + 333 + ], + "type": "image", + "image_path": "e48afad38d228e4820b40a7fb5a37dfe0cb0633ef7e7a40866fd7553ce4a3c2a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 339, + 505, + 353 + ], + "lines": [ + { + "bbox": [ + 323, + 339, + 505, + 353 + ], + "spans": [ + { + "bbox": [ + 323, + 339, + 505, + 353 + ], + "type": "text", + "content": "Figure 1: Overview of the proposed method." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 302, + 375, + 427, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 375, + 427, + 389 + ], + "spans": [ + { + "bbox": [ + 302, + 375, + 427, + 389 + ], + "type": "text", + "content": "4 Experimental Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 398, + 527, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 398, + 527, + 439 + ], + "spans": [ + { + "bbox": [ + 302, + 398, + 527, + 439 + ], + "type": "text", + "content": "We follow a similar experimental setup to Wei et al. (2022), focusing on tasks covering arithmetic, commonsense and symbolic reasoning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 452, + 449, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 452, + 449, + 464 + ], + "spans": [ + { + "bbox": [ + 302, + 452, + 449, + 464 + ], + "type": "text", + "content": "4.1 Benchmarks and Metrics" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 471, + 442, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 471, + 442, + 484 + ], + "spans": [ + { + "bbox": [ + 302, + 471, + 442, + 484 + ], + "type": "text", + "content": "4.1.1 Arithmetic Reasoning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 301, + 489, + 527, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 489, + 527, + 773 + ], + "spans": [ + { + "bbox": [ + 301, + 489, + 527, + 773 + ], + "type": "text", + "content": "We benchmark the proposed method on the following math word problem datasets: (1) GSM8K (Cobbe et al., 2021), (2) MAWPS (Koncel-Kedziorski et al., 2016) and (3) ASDiv (Miao et al., 2021). We use the official training and testing split for GSM8K, taking the last " + }, + { + "bbox": [ + 301, + 489, + 527, + 773 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 301, + 489, + 527, + 773 + ], + "type": "text", + "content": " of the training split for validation, and the 5-fold cross validation splits available for MAWPS and ASDiv. We evaluate task accuracy by checking for the target answer as the final answer in the CoT. In addition, we compute the task accuracy given an external calculator, to account for arithmetic mistakes made by the model, despite the CoT being correct. The external calculator moves through the generated output, recalculating the left-hand-side of equations. It then replaces the right-hand side with the calculated output, to avoid arithmetic mistakes being carried forward. For example, if a model outputted " + }, + { + "bbox": [ + 301, + 489, + 527, + 773 + ], + "type": "inline_equation", + "content": "^{\\prime}5 + 5 = 11" + }, + { + "bbox": [ + 301, + 489, + 527, + 773 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 301, + 489, + 527, + 773 + ], + "type": "inline_equation", + "content": "11*2 = 22'" + }, + { + "bbox": [ + 301, + 489, + 527, + 773 + ], + "type": "text", + "content": ", then the external calculator would first calculate " + }, + { + "bbox": [ + 301, + 489, + 527, + 773 + ], + "type": "inline_equation", + "content": "^{\\prime}5 + 5" + }, + { + "bbox": [ + 301, + 489, + 527, + 773 + ], + "type": "text", + "content": " and replace the '11' with a '10'. In the subsequent equation, it would" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 780, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 310, + 791 + ], + "type": "text", + "content": "1774" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 289, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 289, + 97 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 289, + 97 + ], + "type": "text", + "content": "also replace the '11' with a '10' and arrive at the final result of '20'." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 107, + 223, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 107, + 223, + 120 + ], + "spans": [ + { + "bbox": [ + 67, + 107, + 223, + 120 + ], + "type": "text", + "content": "4.1.2 Commonsense Reasoning" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 124, + 291, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 124, + 291, + 232 + ], + "spans": [ + { + "bbox": [ + 67, + 124, + 291, + 232 + ], + "type": "text", + "content": "We benchmark the model's ability to perform commonsense reasoning on the StrategyQA dataset (Geva et al., 2021a). As a testing split is not available, we do not shuffle the dataset to allow reproducing our split of taking the first " + }, + { + "bbox": [ + 67, + 124, + 291, + 232 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 67, + 124, + 291, + 232 + ], + "type": "text", + "content": " as training data, the following " + }, + { + "bbox": [ + 67, + 124, + 291, + 232 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 67, + 124, + 291, + 232 + ], + "type": "text", + "content": " as validation data, and the final " + }, + { + "bbox": [ + 67, + 124, + 291, + 232 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 67, + 124, + 291, + 232 + ], + "type": "text", + "content": " as testing data. We compute task accuracy in the same manner as previously mentioned." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 241, + 198, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 241, + 198, + 253 + ], + "spans": [ + { + "bbox": [ + 67, + 241, + 198, + 253 + ], + "type": "text", + "content": "4.1.3 Symbolic Reasoning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 257, + 291, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 257, + 291, + 500 + ], + "spans": [ + { + "bbox": [ + 67, + 257, + 291, + 500 + ], + "type": "text", + "content": "Lastly, we benchmark the model on two synthetic tasks for symbolic reasoning: (1) last letter concatenation and (2) coinflip (Wei et al., 2022). Last letter concatenation prompts the model to concatenate the last letter of each word in a string. Coinflip prompts the model to perform state tracking of the coin being flipped. We evaluate task accuracy in the same manner as before. Due to the rigid structure of the datasets, we focus on evaluating the model's generalizability to out-of-distribution (OOD) examples. We finetune the models on examples of length two and evaluate on sequences of length three and four. We initially infer the CoT using PaLM 540B, however, find that the LLM is able to perfectly replicate the desired CoT bar one example due to the rigidity of the template. We therefore decide to use the template generated CoT in our experiments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 512, + 187, + 524 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 512, + 187, + 524 + ], + "spans": [ + { + "bbox": [ + 67, + 512, + 187, + 524 + ], + "type": "text", + "content": "4.2 Baselines and setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 528, + 291, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 528, + 291, + 635 + ], + "spans": [ + { + "bbox": [ + 67, + 528, + 291, + 635 + ], + "type": "text", + "content": "We select PaLM 540B (Chowdhery et al., 2022) and GPT-3 175B (Brown et al., 2020) as teacher models. We select PaLM 540B based on the state-of-the-art results on the benchmarking datasets reported by Wei et al. (2022), and confirm the observed trends with GPT-3 175B. The publicly accessible teacher models are prompted as described in Section 3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 638, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 638, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 638, + 291, + 772 + ], + "type": "text", + "content": "We select different sizes of T5 (Raffel et al., 2020) as student models, as T5 is publicly available in many sizes. The student models are trained on the PaLM 540B or GPT-3 175B generated CoT data as described in Section 3. We establish T5 XXL model finetuned on the original target as the baseline. We refrain from shuffling the datasets to allow for reproducibility. For the MAWPS and ASDiv dataset, we perform 5-fold cross validation. For all remaining datasets, we take " + }, + { + "bbox": [ + 67, + 638, + 291, + 772 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 67, + 638, + 291, + 772 + ], + "type": "text", + "content": " of the" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 321, + 81, + 497, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 81, + 497, + 120 + ], + "spans": [ + { + "bbox": [ + 321, + 81, + 497, + 120 + ], + "type": "text", + "content": "Input: Roger has 5 tennis balls. He buys 2 more cans of tennis balls. Each can has 3 tennis balls. How many tennis balls does he have now?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 322, + 128, + 482, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 128, + 482, + 168 + ], + "spans": [ + { + "bbox": [ + 322, + 128, + 482, + 168 + ], + "type": "text", + "content": "Output: Roger started with 5 balls. 2 cans of 3 tennis balls each is 6 tennis balls. " + }, + { + "bbox": [ + 322, + 128, + 482, + 168 + ], + "type": "inline_equation", + "content": "5 + 6 = 11" + }, + { + "bbox": [ + 322, + 128, + 482, + 168 + ], + "type": "text", + "content": ". The answer is 11." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 192, + 525, + 216 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 192, + 525, + 216 + ], + "spans": [ + { + "bbox": [ + 302, + 192, + 525, + 216 + ], + "type": "text", + "content": "Figure 2: A training example from Wei et al. (2022) demonstrating the input and output provided to T5." + } + ] + } + ], + "index": 10, + "type": "text" + }, + { + "bbox": [ + 302, + 239, + 525, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 239, + 525, + 318 + ], + "spans": [ + { + "bbox": [ + 302, + 239, + 525, + 318 + ], + "type": "text", + "content": "training set as a validation set to select the best model checkpoint. Figure 2 showcases an input examples for T5. We refer the reader to Wei et al. (2022) for more training examples, as well as the prompts used for generating the CoT using PaLM 540B and GPT-3 175B." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 320, + 525, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 320, + 525, + 374 + ], + "spans": [ + { + "bbox": [ + 302, + 320, + 525, + 374 + ], + "type": "text", + "content": "We refer the reader to Appendix A for an overview of the dataset licenses. We also refer the reader to Appendix B for an overview of the computational resources." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 303, + 386, + 361, + 397 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 386, + 361, + 397 + ], + "spans": [ + { + "bbox": [ + 303, + 386, + 361, + 397 + ], + "type": "text", + "content": "5 Results" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 408, + 430, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 408, + 430, + 421 + ], + "spans": [ + { + "bbox": [ + 302, + 408, + 430, + 421 + ], + "type": "text", + "content": "5.1 Arithmetic reasoning" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 426, + 525, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 426, + 525, + 560 + ], + "spans": [ + { + "bbox": [ + 302, + 426, + 525, + 560 + ], + "type": "text", + "content": "Table 1 details the task accuracy with and without an external calculator for the arithmetic reasoning benchmarks. Our results show that the proposed method improves task accuracy across all datasets. Most notably, the task accuracy of MAwPS is significantly improved. The accuracy achieved given a calculator comes close to the accuracy of 8-shot PaLM 540B, demonstrating that knowledge distillation is effective, but potentially limited by the mathematical abilities of small models." + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 304, + 569, + 522, + 677 + ], + "blocks": [ + { + "bbox": [ + 304, + 569, + 522, + 677 + ], + "lines": [ + { + "bbox": [ + 304, + 569, + 522, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 522, + 677 + ], + "type": "table", + "html": "
Baseline T5 XXLCoT Finetuned T5 XXLCoT 8-shot PaLM 540B
Acc.Acc.Acc. with Calc.Acc.Acc. with Calc.
GSM8K8.1121.9938.2156.9058.60
Dataset Size672553375337--
MAWPS54.1570.4188.2293.0093.66
Dataset Size159015901590--
ASDiv39.6442.1260.7373.972.6
Dataset Size184415441544--
", + "image_path": "66bf5edd334996a8b4e746bb088737f1f8763fe3752638e51af40a920e3a1e33.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 302, + 685, + 524, + 756 + ], + "lines": [ + { + "bbox": [ + 302, + 685, + 524, + 756 + ], + "spans": [ + { + "bbox": [ + 302, + 685, + 524, + 756 + ], + "type": "text", + "content": "Table 1: Task accuracy across arithmetic reasoning datasets for T5 XXL without finetuning (baseline) and finetuned on PaLM 540B generated chain-of-thought (CoT). We report the accuracy of PaLM 540B on the used datasets for reference. We do not finetune PaLM for this, but employ 8 chain of thought prompts." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "1775" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 240, + 98 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 240, + 98 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 240, + 98 + ], + "type": "text", + "content": "5.1.1 Ablation study on generating chain-of-thought data" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 100, + 291, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 100, + 291, + 290 + ], + "spans": [ + { + "bbox": [ + 67, + 100, + 291, + 290 + ], + "type": "text", + "content": "We perform an ablation study to confirm that providing a LLM with the target during CoT generation is beneficial. We found that for the GSM8K dataset, PaLM 540B only achieves a " + }, + { + "bbox": [ + 67, + 100, + 291, + 290 + ], + "type": "inline_equation", + "content": "59.98\\%" + }, + { + "bbox": [ + 67, + 100, + 291, + 290 + ], + "type": "text", + "content": " accuracy if prompted without the target. In comparison, when including the target in the prompt the accuracy is " + }, + { + "bbox": [ + 67, + 100, + 291, + 290 + ], + "type": "inline_equation", + "content": "79.37\\%" + }, + { + "bbox": [ + 67, + 100, + 291, + 290 + ], + "type": "text", + "content": ". A superficial explanation would be that when the model is conditioned on the expected answer, it produces the same CoT but copies the answer. However, an analysis of a subset of the differences between CoT produced with and without this conditioning shows that most of the benefits actually come from the model correcting CoT that had a single step missing or was wrong." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 298, + 212, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 298, + 212, + 312 + ], + "spans": [ + { + "bbox": [ + 67, + 298, + 212, + 312 + ], + "type": "text", + "content": "5.2 Commonsense reasoning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 316, + 291, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 316, + 291, + 464 + ], + "spans": [ + { + "bbox": [ + 67, + 316, + 291, + 464 + ], + "type": "text", + "content": "For the StrategyQA dataset (Table 3), we found that using CoT finetuning improves accuracy from " + }, + { + "bbox": [ + 67, + 316, + 291, + 464 + ], + "type": "inline_equation", + "content": "68.12\\%" + }, + { + "bbox": [ + 67, + 316, + 291, + 464 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 67, + 316, + 291, + 464 + ], + "type": "inline_equation", + "content": "71.98\\%" + }, + { + "bbox": [ + 67, + 316, + 291, + 464 + ], + "type": "text", + "content": ", using only 1319 of the original 1648 examples. Compared to the arithmetic reasoning datasets, the improvement is not as significant. This can be explained by the model lacking factual knowledge that the dataset requires. The task is heavily focused on the model reasoning on such knowledge, however, a smaller LM is most likely not in possession of this knowledge compared to a larger model with higher memorisation capacity." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 473, + 188, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 473, + 188, + 487 + ], + "spans": [ + { + "bbox": [ + 67, + 473, + 188, + 487 + ], + "type": "text", + "content": "5.3 Symbolic reasoning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 490, + 291, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 490, + 291, + 653 + ], + "spans": [ + { + "bbox": [ + 67, + 490, + 291, + 653 + ], + "type": "text", + "content": "Table 2 shows the results obtained for the synthetic symbolic reasoning datasets, focusing on OOD generalization. Focusing on Last Letter Concatenation, it can be stated that both traditional finetuning and the suggested method fail at generalizing to a longer sequence length. In comparison, the proposed method significantly increases accuracy for the Coinflip dataset with regard to generalizing to three coinflips. In contrast, generalisation to four coinflips is slightly weaker than the baseline, which performs very strongly. This may be related to the task length being twice that of the training task." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 661, + 257, + 687 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 661, + 257, + 687 + ], + "spans": [ + { + "bbox": [ + 67, + 661, + 257, + 687 + ], + "type": "text", + "content": "5.4 Replicating Results using different Teacher Models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 692, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 692, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 692, + 291, + 772 + ], + "type": "text", + "content": "We demonstrate the robustness of our method using a different teacher model, namely GPT-3 175B. Table 3 shows the results for GSM8K and StrategyQA when T5 XXL is finetuned on CoT data generated by GPT-3. The results show that the proposed method elicits improvements also with other" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 304, + 68, + 525, + 130 + ], + "blocks": [ + { + "bbox": [ + 304, + 68, + 525, + 130 + ], + "lines": [ + { + "bbox": [ + 304, + 68, + 525, + 130 + ], + "spans": [ + { + "bbox": [ + 304, + 68, + 525, + 130 + ], + "type": "table", + "html": "
Baseline T5 XXLCoT Finetuned T5 XXLCoT 8-shot PaLM 540B
Last LetterOOD: 30.000.0094.8
Concat.OOD: 40.000.0063.0
CoinflipOOD: 313.1086.7098.6
OOD: 473.8070.5090.2
", + "image_path": "4aca8cb380aeace2cb83474e36ba8a73711ef1b6ff027eb5f8e8502785f78258.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 218, + 526, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 218, + 526, + 327 + ], + "spans": [ + { + "bbox": [ + 302, + 218, + 526, + 327 + ], + "type": "text", + "content": "LLMs as teachers. We also report the accuracy of T5 XXL finetuned on golden CoT provided with the datasets. For the StrategyQA dataset, the model finetuned on the golden CoT performs best, which may be attributed to the dataset being the largest, as both PaLM and GPT-3 get some examples wrong. In contrast, the model finetuned on PaLM generated CoT performs the best for GSM8K." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 304, + 335, + 525, + 426 + ], + "blocks": [ + { + "bbox": [ + 302, + 137, + 526, + 196 + ], + "lines": [ + { + "bbox": [ + 302, + 137, + 526, + 196 + ], + "spans": [ + { + "bbox": [ + 302, + 137, + 526, + 196 + ], + "type": "text", + "content": "Table 2: Task accuracy across the symbolic reasoning datasets for T5 XXL finetuned on chain-of-thought (CoT) data. For each dataset, there are 1000 training and testing examples. We report the accuracy of PaLM 540B from (Wei et al., 2022) for reference." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 304, + 335, + 525, + 426 + ], + "lines": [ + { + "bbox": [ + 304, + 335, + 525, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 525, + 426 + ], + "type": "table", + "html": "
Base TaskOriginal CotCoT finetuned T5 XXL using PaLM 540BGPT-3 175BCoT 8-Shot PaLM 540BGPT-3 175B
GSM8K8.1119.9421.9918.4256.946.9
acc. with Calc.-26.9938.2133.0658.649.6
Dataset Size6725672553375298--
StrategyQA68.1271.9867.1563.7777.865.4
Dataset Size1648164813191319--
", + "image_path": "aabf7c9e1ad145c8693e81fe4f461100f7180bbd1ba1b16c72f371abbba0fa42.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 433, + 525, + 518 + ], + "lines": [ + { + "bbox": [ + 302, + 433, + 525, + 518 + ], + "spans": [ + { + "bbox": [ + 302, + 433, + 525, + 518 + ], + "type": "text", + "content": "Table 3: Task accuracy for T5 XXL finetuned on chain-of-thought (CoT) data generated by PaLM 540B and GPT-3 175B. We also finetune on the reasoning steps provided by the datasets. We report the accuracy of PaLM 540B on the used datasets for reference. We do not finetune PaLM for this, but employ 8 chain of thought prompts." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 302, + 540, + 465, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 540, + 465, + 553 + ], + "spans": [ + { + "bbox": [ + 302, + 540, + 465, + 553 + ], + "type": "text", + "content": "5.5 Ablation study on model size" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 557, + 525, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 557, + 525, + 678 + ], + "spans": [ + { + "bbox": [ + 302, + 557, + 525, + 678 + ], + "type": "text", + "content": "We investigate the performance gain achieved via finetuning student models of different sizes. Figure 3 shows the performance gain achieved when finetuning T5 of different sizes on the GSM8K dataset. Our results show that T5 base, with 44 times fewer parameters than T5 XXL, matches the performance of the baseline T5 XXL when trained on CoT data. Moreover, given an external calculator, even T5 small outperforms the baseline T5 XXL." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 688, + 470, + 701 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 688, + 470, + 701 + ], + "spans": [ + { + "bbox": [ + 302, + 688, + 470, + 701 + ], + "type": "text", + "content": "5.6 Ablation study on dataset size" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 705, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 705, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 705, + 525, + 772 + ], + "type": "text", + "content": "We also investigate the trade-off between the performance gain from CoT finetuning and dataset size. Table 4 details the test accuracy achieved when finetuning T5 XXL on only " + }, + { + "bbox": [ + 302, + 705, + 525, + 772 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 302, + 705, + 525, + 772 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 302, + 705, + 525, + 772 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 302, + 705, + 525, + 772 + ], + "type": "text", + "content": " of the data, randomly selected. In comparison to the" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 780, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 310, + 791 + ], + "type": "text", + "content": "1776" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 72, + 286, + 242 + ], + "blocks": [ + { + "bbox": [ + 73, + 72, + 286, + 242 + ], + "lines": [ + { + "bbox": [ + 73, + 72, + 286, + 242 + ], + "spans": [ + { + "bbox": [ + 73, + 72, + 286, + 242 + ], + "type": "image", + "image_path": "cd13943609b671f1f35d570693adefd8c778ebba2260743db1aaa7f486ecbc0e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 253, + 291, + 278 + ], + "lines": [ + { + "bbox": [ + 67, + 253, + 291, + 278 + ], + "spans": [ + { + "bbox": [ + 67, + 253, + 291, + 278 + ], + "type": "text", + "content": "Figure 3: Effect of student model (T5) size on accuracy on GSM8K." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 298, + 291, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 298, + 291, + 391 + ], + "spans": [ + { + "bbox": [ + 67, + 298, + 291, + 391 + ], + "type": "text", + "content": "baseline accuracy of " + }, + { + "bbox": [ + 67, + 298, + 291, + 391 + ], + "type": "inline_equation", + "content": "8.11\\%" + }, + { + "bbox": [ + 67, + 298, + 291, + 391 + ], + "type": "text", + "content": " (Table 3), we see that our method is 6x more data efficient, achieving accuracy of " + }, + { + "bbox": [ + 67, + 298, + 291, + 391 + ], + "type": "inline_equation", + "content": "11.22\\%" + }, + { + "bbox": [ + 67, + 298, + 291, + 391 + ], + "type": "text", + "content": " with only " + }, + { + "bbox": [ + 67, + 298, + 291, + 391 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 67, + 298, + 291, + 391 + ], + "type": "text", + "content": " of the examples. However, training on just " + }, + { + "bbox": [ + 67, + 298, + 291, + 391 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 67, + 298, + 291, + 391 + ], + "type": "text", + "content": " of the data still creates a quality gap, and it's possible that with e.g. " + }, + { + "bbox": [ + 67, + 298, + 291, + 391 + ], + "type": "inline_equation", + "content": "200\\%" + }, + { + "bbox": [ + 67, + 298, + 291, + 391 + ], + "type": "text", + "content": " larger dataset we could outperform the results in Table 3." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 69, + 399, + 290, + 466 + ], + "blocks": [ + { + "bbox": [ + 69, + 399, + 290, + 466 + ], + "lines": [ + { + "bbox": [ + 69, + 399, + 290, + 466 + ], + "spans": [ + { + "bbox": [ + 69, + 399, + 290, + 466 + ], + "type": "table", + "html": "
Percentage of GSM8K data used to trainCoT finetuned T5 XXL
Acc.Acc. with Calc.
4% (213 examples)6.2912.28
20% (1067 examples)11.2220.47
100% (5337 examples)21.9938.21
", + "image_path": "ce71baa333d2588dca103acec56b9e0fd43d33f996534ca55498bef91f944e0c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 473, + 291, + 511 + ], + "lines": [ + { + "bbox": [ + 67, + 473, + 291, + 511 + ], + "spans": [ + { + "bbox": [ + 67, + 473, + 291, + 511 + ], + "type": "text", + "content": "Table 4: Task accuracy of T5 XXL finetuned on different amounts of chain-of-thought (CoT) data generated by PaLM 540B." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 535, + 143, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 535, + 143, + 548 + ], + "spans": [ + { + "bbox": [ + 67, + 535, + 143, + 548 + ], + "type": "text", + "content": "6 Discussion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 556, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 556, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 556, + 291, + 773 + ], + "type": "text", + "content": "We demonstrate that finetuning larger LMs on the CoT data generated by LLMs of over 100 billion parameters can significantly improve task accuracy. Even a small number of CoT examples appear to suffice for this. However, such improvements appear to be task dependent. For example, the effects are limited for the StrategyQA dataset, which can be attributed to the task requiring specific factual knowledge, which smaller LMs may not have memorised due to their limited capacity. Nevertheless, there is some performance improvement, which may be attributed to the model learning how to approach such tasks. Moreover, the CoT knowledge distillation pipeline presented allows to trade-off model and dataset size with accuracy. Future work could explore improving the reasoning of small" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 71, + 526, + 113 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 113 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 113 + ], + "type": "text", + "content": "models in multi-task settings, as well as the generation of new training data using LLMs, rather than annotating existing datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 303, + 123, + 382, + 136 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 123, + 382, + 136 + ], + "spans": [ + { + "bbox": [ + 303, + 123, + 382, + 136 + ], + "type": "text", + "content": "7 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 145, + 527, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 145, + 527, + 255 + ], + "spans": [ + { + "bbox": [ + 302, + 145, + 527, + 255 + ], + "type": "text", + "content": "This work explores CoT knowledge distillation from LLMs of over 100 billion parameters to smaller LMs. We propose a knowledge distillation pipeline consisting of two keys steps: (1) generate CoT for existing datasets using LLMs and (2) finetune smaller LMs on the CoT. Our results demonstrate that finetuning on CoT improves task accuracy across a range of benchmarking datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 264, + 384, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 264, + 384, + 277 + ], + "spans": [ + { + "bbox": [ + 302, + 264, + 384, + 277 + ], + "type": "text", + "content": "8 Limitations" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 287, + 527, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 287, + 527, + 462 + ], + "spans": [ + { + "bbox": [ + 302, + 287, + 527, + 462 + ], + "type": "text", + "content": "The results we present must be viewed in the context of a few limitations. A limitation is that we only perform experiments in English and on one task at a time. To be more comparable to a LLM few-shot settings, other languages and a multi-task setup could be explored. Furthermore, in order to replicate the results access to none public models is required and inference must be performed on large amounts of data. Another limitation of our work is that it only explores the original CoT prompting approach, but we do not explore subsequent improvements, such a self-consistency (Wang et al., 2022)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 474, + 441, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 474, + 441, + 487 + ], + "spans": [ + { + "bbox": [ + 302, + 474, + 441, + 487 + ], + "type": "text", + "content": "9 Ethical Considerations" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 497, + 527, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 497, + 527, + 618 + ], + "spans": [ + { + "bbox": [ + 302, + 497, + 527, + 618 + ], + "type": "text", + "content": "The main ethical considerations of our research arise from the text generation performed. The concerns here are that both the teacher and student model may potentially generate non-factual (Ji et al., 2022; Pagnoni et al., 2021; Kreps et al., 2022) or offensive output (Gehman et al., 2020). This is largely influenced by the input data, which is our case are standard, peer-reviewed benchmarking tasks in the NLP domain." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 303, + 642, + 362, + 655 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 642, + 362, + 655 + ], + "spans": [ + { + "bbox": [ + 303, + 642, + 362, + 655 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 303, + 662, + 526, + 772 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 303, + 662, + 525, + 697 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 662, + 525, + 697 + ], + "spans": [ + { + "bbox": [ + 303, + 662, + 525, + 697 + ], + "type": "text", + "content": "BIG-bench collaboration. 2021. Beyond the imitation game: Measuring and extrapolating the capabilities of language models. In preparation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 303, + 706, + 526, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 706, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 303, + 706, + 526, + 772 + ], + "type": "text", + "content": "Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "1777" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 291, + 138 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 291, + 138 + ], + "type": "text", + "content": "Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. 2022. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 148, + 291, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 148, + 291, + 202 + ], + "spans": [ + { + "bbox": [ + 69, + 148, + 291, + 202 + ], + "type": "text", + "content": "Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. 2022. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 212, + 290, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 212, + 290, + 267 + ], + "spans": [ + { + "bbox": [ + 69, + 212, + 290, + 267 + ], + "type": "text", + "content": "Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. 2021. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 277, + 290, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 277, + 290, + 342 + ], + "spans": [ + { + "bbox": [ + 69, + 277, + 290, + 342 + ], + "type": "text", + "content": "Jacob Eisenstein, Daniel Andor, Bernd Bohnet, Michael Collins, and David Mimno. 2022. Honest students from untrusted teachers: Learning an interpretable question-answering pipeline from a pretrained language model. arXiv preprint arXiv:2210.02498." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 353, + 290, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 353, + 290, + 396 + ], + "spans": [ + { + "bbox": [ + 69, + 353, + 290, + 396 + ], + "type": "text", + "content": "Samuel Gehman, Suchin Gururangan, Maarten Sap, Yejin Choi, and Noah A Smith. 2020. Realtoxicityprompts: Evaluating neural toxic degeneration in language models. arXiv preprint arXiv:2009.11462." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 406, + 290, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 406, + 290, + 461 + ], + "spans": [ + { + "bbox": [ + 69, + 406, + 290, + 461 + ], + "type": "text", + "content": "Mor Geva, Daniel Khashabi, Elad Segal, Tushar Khot, Dan Roth, and Jonathan Berant. 2021a. Did aristotle use a laptop? a question answering benchmark with implicit reasoning strategies. Transactions of the Association for Computational Linguistics, 9:346-361." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 470, + 290, + 526 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 470, + 290, + 526 + ], + "spans": [ + { + "bbox": [ + 69, + 470, + 290, + 526 + ], + "type": "text", + "content": "Mor Geva, Daniel Khashabi, Elad Segal, Tushar Khot, Dan Roth, and Jonathan Berant. 2021b. Did aristotle use a laptop? a question answering benchmark with implicit reasoning strategies. Transactions of the Association for Computational Linguistics, 9:346-361." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 535, + 290, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 535, + 290, + 569 + ], + "spans": [ + { + "bbox": [ + 69, + 535, + 290, + 569 + ], + "type": "text", + "content": "Geoffrey Hinton, Oriol Vinyals, Jeff Dean, et al. 2015. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2(7)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 577, + 290, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 577, + 290, + 611 + ], + "spans": [ + { + "bbox": [ + 69, + 577, + 290, + 611 + ], + "type": "text", + "content": "Namgyu Ho, Laura Schmid, and Se-Young Yun. 2022. Large language models are reasoning teachers. arXiv preprint arXiv:2212.10071." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 620, + 290, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 620, + 290, + 665 + ], + "spans": [ + { + "bbox": [ + 69, + 620, + 290, + 665 + ], + "type": "text", + "content": "Jiaxin Huang, Shixiang Shane Gu, Le Hou, Yuexin Wu, Xuezhi Wang, Hongkun Yu, and Jiawei Han. 2022. Large language models can self-improve. arXiv preprint arXiv:2210.11610." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 674, + 290, + 729 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 674, + 290, + 729 + ], + "spans": [ + { + "bbox": [ + 69, + 674, + 290, + 729 + ], + "type": "text", + "content": "Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Yejin Bang, Andrea Madotto, and Pascale Fung. 2022. Survey of hallucination in natural language generation. ACM Computing Surveys." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 69, + 739, + 291, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 739, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 739, + 291, + 772 + ], + "type": "text", + "content": "Rik Koncel-Kedziorski, Subhro Roy, Aida Amini, Nate Kushman, and Hannaneh Hajishirzi. 2016. Mawps: A math word problem repository. In Proceedings of" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 525, + 771 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 314, + 72, + 525, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 525, + 106 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 525, + 106 + ], + "type": "text", + "content": "the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1152-1157." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 116, + 525, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 116, + 525, + 170 + ], + "spans": [ + { + "bbox": [ + 304, + 116, + 525, + 170 + ], + "type": "text", + "content": "Sarah Kreps, R Miles McCain, and Miles Brundage. 2022. All the news that's fit to fabricate: A-generated text as a tool of media misinformation. Journal of Experimental Political Science, 9(1):104-117." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 181, + 525, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 181, + 525, + 237 + ], + "spans": [ + { + "bbox": [ + 304, + 181, + 525, + 237 + ], + "type": "text", + "content": "Shiyang Li, Jianshu Chen, Yelong Shen, Zhiyu Chen, Xinlu Zhang, Zekun Li, Hong Wang, Jing Qian, Baolin Peng, Yi Mao, et al. 2022. Explanations from large language models make small reasoners better. arXiv preprint arXiv:2210.06726." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 247, + 525, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 525, + 291 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 525, + 291 + ], + "type": "text", + "content": "Shen-Yun Miao, Chao-Chun Liang, and Keh-Yih Su. 2021. A diverse corpus for evaluating and developing english math word problem solvers. arXiv preprint arXiv:2106.15772." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 301, + 525, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 301, + 525, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 301, + 525, + 346 + ], + "type": "text", + "content": "Artidoro Pagnoni, Vidhisha Balachandran, and Yulia Tsvetkov. 2021. Understanding factuality in abstractive summarization with frank: A benchmark for factuality metrics. arXiv preprint arXiv:2104.13346." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 356, + 525, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 356, + 525, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 356, + 525, + 411 + ], + "type": "text", + "content": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J Liu, et al. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(140):1-67." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 422, + 525, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 525, + 476 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 525, + 476 + ], + "type": "text", + "content": "Kumar Shridhar, Alessandro Stolfo, and Mrinmaya Sachan. 2022. Distilling multi-step reasoning capabilities of large language models into smaller models via semantic decompositions. arXiv preprint arXiv:2212.00193." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 487, + 525, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 525, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 525, + 542 + ], + "type": "text", + "content": "Yi Tay, Mostafa Dehghani, Vinh Q Tran, Xavier Garcia, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Neil Houlsby, and Donald Metzler. 2022. Unifying language learning paradigms. arXiv preprint arXiv:2205.05131." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 553, + 525, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 553, + 525, + 597 + ], + "spans": [ + { + "bbox": [ + 304, + 553, + 525, + 597 + ], + "type": "text", + "content": "Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, and Denny Zhou. 2022. Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 608, + 525, + 662 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 608, + 525, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 608, + 525, + 662 + ], + "type": "text", + "content": "Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. 2021. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 673, + 525, + 718 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 673, + 525, + 718 + ], + "spans": [ + { + "bbox": [ + 304, + 673, + 525, + 718 + ], + "type": "text", + "content": "Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed Chi, Quoc Le, and Denny Zhou. 2022. Chain of thought prompting elicits reasoning in large language models. arXiv preprint arXiv:2201.11903." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 728, + 525, + 771 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 728, + 525, + 771 + ], + "spans": [ + { + "bbox": [ + 304, + 728, + 525, + 771 + ], + "type": "text", + "content": "Ronald J Williams and David Zipser. 1989. A learning algorithm for continually running fully recurrent neural networks. *Neural computation*, 1(2):270-280." + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1778" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 71, + 232, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 71, + 232, + 84 + ], + "spans": [ + { + "bbox": [ + 68, + 71, + 232, + 84 + ], + "type": "text", + "content": "A Dataset Usage and Licenses" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 92, + 291, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 92, + 291, + 147 + ], + "spans": [ + { + "bbox": [ + 67, + 92, + 291, + 147 + ], + "type": "text", + "content": "In this section, we list the licenses for the datasets used and any ethical concerns regarding their usage. We describe the dataset splits used for all datasets in Section 4 of the paper." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 68, + 155, + 201, + 169 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 155, + 201, + 169 + ], + "spans": [ + { + "bbox": [ + 68, + 155, + 201, + 169 + ], + "type": "text", + "content": "A.1 Arithmetic Reasoning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 172, + 292, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 172, + 292, + 253 + ], + "spans": [ + { + "bbox": [ + 67, + 172, + 292, + 253 + ], + "type": "text", + "content": "The GSM8K dataset (Cobbe et al., 2021) is available under the MIT license. The MAWPS dataset (Koncel-Kedziorski et al., 2016) is available under the CC BY 4.0 and the ASDiv dataset (Miao et al., 2021) is available under the CC BY-NC 4.0 license. We follow the intended usage of the datasets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 263, + 218, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 263, + 218, + 276 + ], + "spans": [ + { + "bbox": [ + 68, + 263, + 218, + 276 + ], + "type": "text", + "content": "A.2 Commonsense Reasoning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 280, + 291, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 280, + 291, + 375 + ], + "spans": [ + { + "bbox": [ + 67, + 280, + 291, + 375 + ], + "type": "text", + "content": "The StrategyQA dataset (Geva et al., 2021b) is available under the MIT license. Similar to Wei et al. (2022), we use the open-domain setting version available as part of the Big-bench collaboration (BIG-bench collaboration, 2021), available under the Apache License 2.0. We follow the intended usage of the datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 384, + 192, + 397 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 384, + 192, + 397 + ], + "spans": [ + { + "bbox": [ + 68, + 384, + 192, + 397 + ], + "type": "text", + "content": "A.3 Symbolic Reasoning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 401, + 289, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 401, + 289, + 428 + ], + "spans": [ + { + "bbox": [ + 67, + 401, + 289, + 428 + ], + "type": "text", + "content": "We generate the symbolic reasoning datasets as described in Wei et al. (2022)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 68, + 438, + 223, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 438, + 223, + 452 + ], + "spans": [ + { + "bbox": [ + 68, + 438, + 223, + 452 + ], + "type": "text", + "content": "B Computational Resources" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 460, + 291, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 460, + 291, + 581 + ], + "spans": [ + { + "bbox": [ + 67, + 460, + 291, + 581 + ], + "type": "text", + "content": "We perform inference and finetuning on different sizes of T5 on TPUs. We perform inference on PaLM 540B also on TPUs. Our results can be replicated via the public API (https://developersgenerativeai.google/products/palm). To make requests to GPT-3 175B, we use the public API (https://beta.openai.com/docs/introduction)." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1779" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 90, + 192, + 103 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 90, + 192, + 103 + ], + "spans": [ + { + "bbox": [ + 68, + 90, + 192, + 103 + ], + "type": "text", + "content": "A For every submission:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 106, + 414, + 240 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "spans": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "type": "text", + "content": "A1. Did you describe the limitations of your work? Section 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 142, + 329, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 142, + 329, + 168 + ], + "spans": [ + { + "bbox": [ + 77, + 142, + 329, + 168 + ], + "type": "text", + "content": "A2. Did you discuss any potential risks of your work? Section 9" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 178, + 414, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 178, + 414, + 204 + ], + "spans": [ + { + "bbox": [ + 77, + 178, + 414, + 204 + ], + "type": "text", + "content": "A3. Do the abstract and introduction summarize the paper's main claims? Abstract and Section 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 77, + 214, + 398, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 214, + 398, + 240 + ], + "spans": [ + { + "bbox": [ + 77, + 214, + 398, + 240 + ], + "type": "text", + "content": "A4. Have you used AI writing assistants when working on this paper? Left blank." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 68, + 250, + 290, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 250, + 290, + 264 + ], + "spans": [ + { + "bbox": [ + 68, + 250, + 290, + 264 + ], + "type": "text", + "content": "B Did you use or create scientific artifacts?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 79, + 269, + 123, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 269, + 123, + 280 + ], + "spans": [ + { + "bbox": [ + 79, + 269, + 123, + 280 + ], + "type": "text", + "content": "Section 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 77, + 289, + 524, + 643 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 77, + 289, + 315, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 289, + 315, + 315 + ], + "spans": [ + { + "bbox": [ + 77, + 289, + 315, + 315 + ], + "type": "text", + "content": "B1. Did you cite the creators of artifacts you used? Section 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 77, + 324, + 463, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 324, + 463, + 352 + ], + "spans": [ + { + "bbox": [ + 77, + 324, + 463, + 352 + ], + "type": "text", + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Appendix A" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 77, + 360, + 524, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 360, + 524, + 426 + ], + "spans": [ + { + "bbox": [ + 77, + 360, + 524, + 426 + ], + "type": "text", + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Section 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 77, + 438, + 524, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 438, + 524, + 505 + ], + "spans": [ + { + "bbox": [ + 77, + 438, + 524, + 505 + ], + "type": "text", + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? We did not discuss this as the datasets are commonly used NLP benchmarks that do not contain personal data." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 77, + 513, + 524, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 513, + 524, + 554 + ], + "spans": [ + { + "bbox": [ + 77, + 513, + 524, + 554 + ], + "type": "text", + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? We discuss this in Section 8, the limitations section. We discuss the coverage of domains in Section 4." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 77, + 562, + 524, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 562, + 524, + 643 + ], + "spans": [ + { + "bbox": [ + 77, + 562, + 524, + 643 + ], + "type": "text", + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. We discuss this in Section 4." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 68, + 651, + 293, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 651, + 293, + 666 + ], + "spans": [ + { + "bbox": [ + 68, + 651, + 293, + 666 + ], + "type": "text", + "content": "C Did you run computational experiments?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 79, + 670, + 154, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 670, + 154, + 682 + ], + "spans": [ + { + "bbox": [ + 79, + 670, + 154, + 682 + ], + "type": "text", + "content": "Sections 4 and 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 77, + 690, + 524, + 745 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 690, + 524, + 745 + ], + "spans": [ + { + "bbox": [ + 77, + 690, + 524, + 745 + ], + "type": "text", + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? We report the model specifics in section 4. We describe the computing infrastructure in Appendix 2, but do not estimate the computational budget." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "spans": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "text", + "content": "ACL 2023 Responsible NLP Checklist" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 751, + 522, + 771 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 751, + 522, + 771 + ], + "spans": [ + { + "bbox": [ + 67, + 751, + 522, + 771 + ], + "type": "text", + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1780" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 76, + 70, + 524, + 238 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 77, + 70, + 523, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 70, + 523, + 110 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 523, + 110 + ], + "type": "text", + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Section 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 77, + 120, + 524, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 120, + 524, + 174 + ], + "spans": [ + { + "bbox": [ + 77, + 120, + 524, + 174 + ], + "type": "text", + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Sections 4 and 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 76, + 184, + 524, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 184, + 524, + 238 + ], + "spans": [ + { + "bbox": [ + 76, + 184, + 524, + 238 + ], + "type": "text", + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Not applicable. Left blank." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 67, + 247, + 522, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 247, + 522, + 278 + ], + "spans": [ + { + "bbox": [ + 67, + 247, + 522, + 278 + ], + "type": "text", + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 76, + 286, + 524, + 539 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 76, + 286, + 524, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 286, + 524, + 327 + ], + "spans": [ + { + "bbox": [ + 76, + 286, + 524, + 327 + ], + "type": "text", + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Not applicable. Left blank." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 76, + 336, + 524, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 336, + 524, + 390 + ], + "spans": [ + { + "bbox": [ + 76, + 336, + 524, + 390 + ], + "type": "text", + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Not applicable. Left blank." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 76, + 400, + 524, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 400, + 524, + 454 + ], + "spans": [ + { + "bbox": [ + 76, + 400, + 524, + 454 + ], + "type": "text", + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Not applicable. Left blank." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 76, + 462, + 519, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 462, + 519, + 489 + ], + "spans": [ + { + "bbox": [ + 76, + 462, + 519, + 489 + ], + "type": "text", + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Not applicable. Left blank." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 76, + 498, + 524, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 498, + 524, + 539 + ], + "spans": [ + { + "bbox": [ + 76, + 498, + 524, + 539 + ], + "type": "text", + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Not applicable. Left blank." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 308, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 308, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 308, + 791 + ], + "type": "text", + "content": "1781" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Text-to-SQL Error Correction with Language Models of Code/63d877fb-bf81-4e0e-9af4-85d702a6c368_content_list.json b/2023/Text-to-SQL Error Correction with Language Models of Code/63d877fb-bf81-4e0e-9af4-85d702a6c368_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f7697b13242c0e67febebb845f0f9cbdeb887cd3 --- /dev/null +++ b/2023/Text-to-SQL Error Correction with Language Models of Code/63d877fb-bf81-4e0e-9af4-85d702a6c368_content_list.json @@ -0,0 +1,1797 @@ +[ + { + "type": "text", + "text": "Text-to-SQL Error Correction with Language Models of Code", + "text_level": 1, + "bbox": [ + 176, + 89, + 821, + 111 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ziru Chen $^{1}$ , Shijie Chen $^{1}$ , Michael White $^{1}$ , Raymond Mooney $^{2}$ , Ali Payani $^{3}$ , Jayanth Srinivasa $^{3}$ , Yu Su $^{1}$ , Huan Sun $^{1}$", + "bbox": [ + 228, + 124, + 771, + 158 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1The Ohio State University", + "bbox": [ + 391, + 159, + 611, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ The University of Texas at Austin $^{3}$ Cisco Research", + "bbox": [ + 278, + 175, + 719, + 191 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{chen.8336, chen.10216, white.1240, su.809, sun.397}@osu.edu", + "bbox": [ + 196, + 192, + 805, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mooney@cs.utexas.edu {apayani, jasriniv}@cisco", + "bbox": [ + 258, + 210, + 741, + 225 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 252, + 339, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite recent progress in text-to-SQL parsing, current semantic parsers are still not accurate enough for practical use. In this paper, we investigate how to build automatic text-to-SQL error correction models. Noticing that token-level edits are out of context and sometimes ambiguous, we propose building clause-level edit models instead. Besides, while most language models of code are not specifically pre-trained for SQL, they know common data structures and their operations in programming languages such as Python. Thus, we propose a novel representation for SQL queries and their edits that adheres more closely to the pre-training corpora of language models of code. Our error correction model improves the exact set match accuracy of different parsers by 2.4-6.5 and obtains up to 4.3 point absolute improvement over two strong baselines. $^1$", + "bbox": [ + 141, + 275, + 460, + 546 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 114, + 555, + 258, + 569 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-SQL parsing is a classic semantic parsing task that finds wide applications (Zelle and Mooney, 1996; Tang and Mooney, 2000). Since the release of Spider (Yu et al., 2018), a cross-database text-to-SQL benchmark, many semantic parsers with decent performance have been developed (Lin et al., 2020; Wang et al., 2020; Deng et al., 2021; Rubin and Berant, 2021; Scholak et al., 2021). Nonetheless, state-of-the-art semantic parsers are still not accurate enough. As a result, their users need to constantly correct wrongly predicted SQL queries, which can be as time-consuming and error-prone as writing a SQL query from scratch (Jorgensen and Shepperd, 2007; Weiss et al., 2007). Therefore, in this paper, we study the problem of automatic text-to-SQL error correction to better assist users in querying complex databases.", + "bbox": [ + 112, + 580, + 489, + 851 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We first highlight that it is essential to factor in the compositional substructures within SQL", + "bbox": [ + 112, + 854, + 487, + 885 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "queries, such as abstract syntax trees (Yin and Neubig, 2017; Guo et al., 2022) and data-flow graphs (Guo et al., 2021), instead of treating code snippets as string sequences. Compared to individual tokens, substructures (e.g. SQL clauses) include more context of the entire program and are more semantically meaningful. Consequently, edit patterns of such substructures are more intuitive for humans to understand and easier for language models to learn. Moreover, while the pre-training corpora for language models of code, such as CodeT5 (Wang et al., 2021), do not include many SQL queries based on their documentation, they naturally contain abundant examples of common data structures like dictionaries. Therefore, we hypothesize that transforming unfamiliar SQL queries into familiar data structures can help language models of code better perform structural editing of SQL queries.", + "bbox": [ + 507, + 253, + 884, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Based on these observations, we develop our error correction model and make two contributions. First, we propose considering SQL clauses instead of tokens as basic semantic units for editing. Using a context-free grammar, we can decompose a SQL query and identify its clauses by traversing its abstract syntax tree. Second, we propose a new representation of SQL queries and their edits that adheres more closely to common code pre-training corpora, including CodeSearchNet (Husain et al., 2020), and makes the structures of a SQL query more explicit. With a decomposed SQL query, we pair each clause with its SQL keyword and represent the entire query as a Python dictionary. Then, we format edits on a wrong SQL query as a program that modifies data of the query's corresponding dictionary. Unlike token-level edits in existing work (Zhang et al., 2023), such dictionary operations define all edits unambiguously and can be directly executed with a Python interpreter.", + "bbox": [ + 507, + 546, + 884, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Through comprehensive experiments with different representations, we show that: (1) our proposed representation has the lowest zero-shot perplexity", + "bbox": [ + 507, + 870, + 882, + 919 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1Our code and data are available at https://github. com/OSU-NLP-Group/Auto-SQL-Correction.", + "bbox": [ + 112, + 891, + 489, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1359", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics", + "bbox": [ + 226, + 945, + 771, + 957 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Volume 2: Short Papers, pages 1359-1372", + "bbox": [ + 368, + 958, + 630, + 971 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "July 9-14, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 295, + 972, + 700, + 985 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/0782fcf933bf83d430bc48da17f22fb97371b382cc22a33bfee1d81242f29482.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Query RepresentationEdit Representation
SQLselect tweets.text from tweets order by tweets.textToken-Level<ReplaceOld> tweets.text <ReplaceNew> tweets.createDate <ReplaceEnd>
Clause-Level<ReplaceOld> order by tweets.text <ReplaceNew> order by tweets.createDate <ReplaceEnd>
PyDictsql = { "select": "select tweets.text", "from": "from tweets", "orderBy": "order by tweets.text" }Clause-Level<ReplaceOld> "orderBy": "order by tweets.text" <ReplaceNew> "orderBy": "order by tweets.createDate" <ReplaceEnd> sql["orderBy"] = "order by tweets.createDate"
Program
", + "bbox": [ + 117, + 80, + 884, + 235 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1: Example representations for a wrong SQL query and the Replace edit action. The corresponding natural language utterance is \"List the text of all tweets in the order of date.\" For token-level and clause-level representations, we format them as \" Span of wrong tokens/clauses Span of correct tokens/clauses \", where , , and are special tokens.", + "bbox": [ + 110, + 244, + 882, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "with CodeT5; (2) simply changing token-level edits to clause-level edits can effectively improve the performance of our models; and (3) our method improves the exact set match accuracy of different parsers by 2.4-6.5 and obtains up to 4.3 point absolute improvement over two strong baselines.", + "bbox": [ + 112, + 315, + 487, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Text-to-SQL Error Correction", + "text_level": 1, + "bbox": [ + 112, + 422, + 410, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given a natural language utterance $\\mathbf{u}$ , a database schema $\\mathbf{s}$ , and a wrong SQL query $\\mathbf{q}_{-}$ produced by an existing parser, our goal is to develop an error correction model that predicts a sequence of edit actions $\\mathbf{e}$ and the correct query $\\mathbf{q}_{+}$ . Following previous work (Zhang et al., 2023), we formulate our task as sequence-to-sequence generation:", + "bbox": [ + 112, + 448, + 487, + 560 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nP (\\mathbf {y} | \\mathbf {x}) = \\Pi_ {t = 1} ^ {T} P (\\mathbf {y} _ {t} | \\mathbf {x}, \\mathbf {y} _ {1: t - 1}) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 569, + 487, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\mathbf{x} = [\\mathbf{u};\\mathbf{s};\\mathbf{q}_{-}]$ is the concatenation of the given inputs and $\\mathbf{y} = [\\mathbf{e};\\mathbf{q}_{+}]$ is the concatenation of all edit actions and the resulting correct query. In this section, we study different representations of SQL queries (Section 2.1) and edits (Section 2.2) to better leverage language models of code.", + "bbox": [ + 112, + 598, + 487, + 695 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 Query Representation", + "text_level": 1, + "bbox": [ + 112, + 705, + 334, + 721 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We consider two representations for a predicted query: (1) the original SQL format and (2) our proposed PyDict (Python Dictionary) representation. To prepare for editing, we disambiguate each SQL query following Rubin and Berant (2021), including lower-casing non-value tokens, resolving table references, and formatting punctuation. This preprocessing normalizes SQL queries predicted by different base parsers and the gold annotations into the same format. To build our PyDict representation, we parse a SQL query into its abstract syntax tree (AST) with Spider's context-free grammar. We", + "bbox": [ + 112, + 726, + 489, + 917 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "use depth-first search to traverse through the AST, find any nested substructures, and construct the dictionary representation bottom-up. Table 1 shows the \"SQL\" and \"PyDict\" representations of a SQL query (more details in Appendix A).", + "bbox": [ + 507, + 315, + 882, + 395 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2 Edit Representation", + "text_level": 1, + "bbox": [ + 507, + 410, + 714, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We first follow Zhang et al. (2023) to use token-level edit representation with special tokens (Table 1), which have unique entries in the tokenizer and the model's embedding layer to describe Replace, Insert, and Delete edit actions (more examples in Appendix F). However, we realize this representation can sometimes be ambiguous. As shown in Table 1, the span \"tweets.text\" appears twice in the SQL query. This repetition would confuse the error correction model with which span to replace when generating the corrected query. Also, the ambiguity makes it difficult to implement rules and directly carry out the edit actions on the wrong query. Hence, we change the token-level edit representation to clause-level, which includes more context of the query to make different edits more distinguishable. In our experiments (Section 4.1), we demonstrate that this simple modification is already effective. Our program representation further improves the performance because it is more similar to the code pre-training corpora and eliminates the need to learn special tokens' representations.", + "bbox": [ + 505, + 432, + 884, + 788 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3 Experimental Setup", + "text_level": 1, + "bbox": [ + 507, + 803, + 717, + 820 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1 Data Synthesis for SQL Error Correction", + "text_level": 1, + "bbox": [ + 507, + 831, + 880, + 847 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To train a text-to-SQL error correction model, we need to collect a set of wrong SQL parses that reflects a realistic distribution of errors (Section 4.2) as our training data. We synthesize this dataset by", + "bbox": [ + 507, + 854, + 882, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1360", + "bbox": [ + 482, + 928, + 521, + 940 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/315c2a10656cad34aaee909b8e17728dbacd044f2f0177a2034af0a9be635d74.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CodeT5BRIDGEv2SmBoP
# of Train47,02024,77620,083
# of Dev448448448
# of Test430392310
Avg. Train Edits2.343.112.72
Avg. Dev Edits2.703.293.31
Avg. Test Edits1.841.511.47
", + "bbox": [ + 121, + 80, + 480, + 187 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 2: Summary of data statistics.", + "bbox": [ + 176, + 198, + 421, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "performing 5-fold cross-validation on each parser, which approximates the actual evaluation setting.", + "bbox": [ + 112, + 228, + 487, + 259 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following the evaluation setup in Yu et al. (2018), we split Spider's training set into five roughly equal subsets by different databases. For each cross-validation fold, we train a text-to-SQL parser (Section 3.2) on four subsets and evaluate it on the remaining one. At inference time, we perform beam search with size 20 for each example and collect grammatical and executable parses in the beam. If a SQL parse is not an exact set match or execution match to the gold annotation, we label it wrong and include it in our training set for error correction. Having synthesized our training dataset, we randomly sample 8 databases and their associated questions to construct a held-out development set. For development set examples, we only keep incorrect SQL parses with the highest beam confidence. For our error correction test set, we train each parser on the full Spider training set and evaluate it on the original Spider's development set without modifications. We similarly keep SQL parses with exact match or execution match errors. Table 2 summarizes the statistics of our data.", + "bbox": [ + 115, + 261, + 489, + 613 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2 Models", + "text_level": 1, + "bbox": [ + 112, + 630, + 218, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-to-SQL base parsers. We choose three text-to-SQL parsers with different decoding strategies and levels of performance (Table 3). We elaborate on our selection criteria in Appendix B.", + "bbox": [ + 112, + 653, + 487, + 715 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- CodeT5 (Wang et al., 2021): We fine-tune CodeT5-base following Xie et al. (2022). This parser represents those using beam search decoding and having a lower accuracy.", + "BRIDGEv2 (Lin et al., 2020): A representative parser with constrained decoding and achieving a medium-level accuracy.", + "- SmBoP (Rubin and Berant, 2021): A representative parser with bottom-up decoding and achieving higher accuracy." + ], + "bbox": [ + 134, + 719, + 487, + 879 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Error correction models. We use two language models of code in all our experiments:", + "bbox": [ + 507, + 84, + 880, + 115 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "CoditT5 (Zhang et al., 2023): A language model pre-trained for code editing tasks by injecting noises to code snippets in CodeSearchNet (Husain et al., 2020) and then denoising with token-level edit representations.", + "- CodeT5 (Wang et al., 2021): A language model pre-trained for general code understanding and generation with four different pre-training objectives." + ], + "bbox": [ + 531, + 117, + 882, + 260 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We compare the existing SQL+Token-Level representation with our proposed ones: SQL+Clause-Level, PyDict+Clause-Level, and PyDict+Program on CodeT5 and the first three on CoditT5.3 Implementation details are in Appendix C.", + "bbox": [ + 507, + 261, + 884, + 341 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3 Evaluation", + "text_level": 1, + "bbox": [ + 507, + 351, + 640, + 366 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We use the increase in Exact Set Match (EM) and Execution Match (EX) accuracy on our error correction test set to measure each model's performance. Because CoditT5's experiments assume the input program has at least one error, we keep this assumption for fair comparisons. To construct a test set satisfying this assumption, we have to compare parser-generated SQL queries with gold annotations (Section 3.1). Thus, we use the Spider development set as our test set and split the Spider training set to build a held-out development set (Table 2) to select model checkpoints during training. We also include results on our held-out development set in the appendix (Table E.1).", + "bbox": [ + 505, + 372, + 882, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4 Results and Analysis", + "text_level": 1, + "bbox": [ + 507, + 609, + 722, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1 Main Results", + "text_level": 1, + "bbox": [ + 507, + 634, + 660, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We summarize our main results in this section. To ensure robustness, we repeat all experiments with 3 different random seeds and report the average performances with standard deviations. Our model can also be used in an interactive framework that allows users to select edit actions from the top- $k$ beam candidates. We include more experiments with simulated user interactions in Appendix E.", + "bbox": [ + 505, + 655, + 880, + 784 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our representation's perplexity is the smallest. We validate that our PyDict+Program representation adheres more closely to the code pre-training corpora by measuring its zero-shot perplexity on CodeT5 using our development set (Section 3.1).", + "bbox": [ + 507, + 793, + 884, + 873 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "Due to SmBoP's bottom-up decoding, we keep its original beam size and collect the top-20 unique beam predictions.", + "bbox": [ + 112, + 891, + 487, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "3We did not use CoditT5 for PyDict+Program because it was pre-trained on token-level edit representations. Its decoder may be specialized in generating edits instead of programs.", + "bbox": [ + 507, + 879, + 882, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1361", + "bbox": [ + 480, + 928, + 517, + 940 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/232333084ea86b3a93da677a41cefb6a766b303c7f67d3da580c35bf61526f39.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelsQueryEditCodeT5BRIDGEv2SmBoP
EMEXEMEXEMEX
No EditN/AN/A62.7 (-)63.6 (-)70.1 (-)68.2 (-)74.6 (-)75.3 (-)
CodiT5SQLToken-Level64.3 (0.1)64.4 (0.2)65.4 (0.5)66.6 (0.3)74.2 (0.4)75.3 (0.1)
SQLClause-Level67.0 (0.4)65.4 (0.5)71.3 (0.5)70.9 (0.2)76.3 (0.0)77.2 (0.3)
PyDictClause-Level67.1 (0.2)66.5 (0.4)70.6 (0.8)70.8 (0.6)76.3 (0.3)77.0 (0.3)
CodeT5SQLToken-Level66.7 (0.9)65.9 (0.5)68.2 (0.4)69.4 (0.8)75.6 (0.4)76.5 (0.6)
SQLClause-Level68.3 (0.3)\\( \\underline{68.2}^{+}(0.6) \\)71.8+(0.4)72.5+(0.2)76.7 (0.6)77.4 (0.3)
PyDictClause-Level66.6 (0.8)67.1 (0.8)72.0+(0.3)72.4+(0.2)77.3 (0.6)77.8 (0.2)
\\( CodeT5^* \\)CodeT5PyDictProgram\\( 69.2^{+}(0.4) \\)\\( 68.4^{+}(0.2) \\)\\( 72.5^{+}(0.4) \\)\\( 73.1^{+}(0.2) \\)77.3 (0.4)77.6 (0.6)
\\( 69.0^{+}(0.2) \\)\\( 68.2^{+}(0.1) \\)\\( 72.5^{+}(0.3) \\)\\( 73.0^{+}(0.6) \\)\\( 78.0^{+}(0.3) \\)\\( 78.5^{+}(0.3) \\)
", + "bbox": [ + 119, + 80, + 877, + 259 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 3: Exact Set Match (EM) and Execution Match (EX) accuracy on Spider development set. The best performances are in bold and the second bests are underlined. Results with $^+$ are statistically significant (McNemar's; $p < 0.05$ ) compared to CodeT5-SQL+Token-Level (Appendix D). Otherwise, the results are not statistically significant. \\*We fine-tune the model to generate edit programs only (without resulting queries) and use Python interpreter to execute the edit actions.", + "bbox": [ + 112, + 268, + 884, + 341 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a38959892ed71b38c996b10dff3a75e11f3afbcab1432d33e61060f6c2917496.jpg", + "image_caption": [ + "Figure 1: CodeT5's zero-shot perplexity (in log scale) of all four representations on our synthesized SQL error development set." + ], + "image_footnote": [], + "bbox": [ + 139, + 353, + 463, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Figure 1, by representing data in Py-Dict, we can reduce the perplexity of CodeT5 by 2 orders of magnitude. After augmenting it with our program representation, we further reduce the zero-shot perplexity of CodeT5 to only $5.96 \\times 10^{2}$ , 3 orders of magnitude less than the SQL+Token-Level representation $(1.26 \\times 10^{5})$ .", + "bbox": [ + 112, + 556, + 489, + 668 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Clause-level editing is more effective, especially when represented in PyDict+Program. Since CodeT5 consistently outperforms CoditT5 with the same representations, we focus on comparisons among CodeT5 variations. As shown in Table 3, compared to CodeT5-SQL+Token-Level, only CodeT5-PyDict+Program achieves statistically significant improvement on all three parsers, while clause-level models fail McNemar's significance test for some parsers. More concretely, it achieves up to 4.3 point more absolute improvement on EM accuracy (68.2 → 72.5; BRIDGEv2) and 3.7 point more absolute improvement on EX accuracy (69.4 → 73.1; BRIDGEv2). Overall, CodeT5-PyDict+Program can boost the parsers' EM accu", + "bbox": [ + 112, + 677, + 489, + 919 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "racy by 2.4-6.5. Thus, both clause-level editing and PyDict+Program representation can better take advantage of language models of code.", + "bbox": [ + 507, + 351, + 882, + 401 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2 Error Analysis", + "text_level": 1, + "bbox": [ + 507, + 414, + 672, + 430 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Additionally, we conduct an error analysis (Table 4) by sampling 100 wrong parses from all three parsers and classifying them into five categories:", + "bbox": [ + 507, + 438, + 882, + 486 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Database Grounding: A generated SQL query has the correct structure, but some table/column names or entity values are wrong.", + "- Incorrect Structure: A generated SQL query has missing, wrong, or redundant structures.", + "- Syntax & Grammar: A generated SQL query violates the programming language's syntax.", + "- False Negative: A generated SQL query is semantically correct but not captured by evaluation metrics, or the gold annotation is wrong.", + "- Other: All other errors, such as wrong aggregation functions, besides the above categories." + ], + "bbox": [ + 531, + 488, + 882, + 678 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the error distributions for each parser are similar, as an example, we discuss our findings based on the strongest parser, SmBoP:", + "bbox": [ + 507, + 681, + 882, + 728 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Database grounding is the major type of error. Among the 100 samples from SmBoP, we find that 54 of them have database grounding errors. Particularly, SmBoP predicts wrong table/column names in 34 parses, inaccurate entity values in 9 parses, and incorrect JOIN relations in 11 parses. Our CodeT5-PyDict+Program model can successfully fix 16 of the 54 erroneous parses, including 10 parses with wrong table/column names, 4 parses with inaccurate entity values, and 2 parses with incorrect JOIN relations. We hypothesize that", + "bbox": [ + 507, + 741, + 884, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1362", + "bbox": [ + 482, + 928, + 521, + 940 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/6668fd7430620200f8ebe85aeac0e4415b4c2d92ccf59e9f74146938cbc233af.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Error CategoryCodeT5BRIDGEv2SmBoP
ResolvedUnresolvedAllResolvedUnresolvedAllResolvedUnresolvedAll
Database Grounding155166144862163854
Incorrect Structure215172121432326
Syntax & Grammar000000145
False Negative099066088
Other17821618167
", + "bbox": [ + 117, + 80, + 880, + 187 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 4: Analysis of 100 sample errors made by each text-to-SQL parser. We group the errors into 5 categories and examine if our CodeT5-PyDict+Program model resolves them.", + "bbox": [ + 112, + 198, + 882, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "database grounding is also a major category of errors in our synthesized training set, so our model has learned to resolve similar errors. Nevertheless, it still cannot correct the remaining 38 SQL parses. We notice that our current representation for database schema is missing critical information, such as column data types and foreign key relations, for our error correction model to fix database grounding errors. Following our PyDict representation for SQL, we suggest designing a code representation for database schema that includes such information to tackle this issue in future work.", + "bbox": [ + 110, + 239, + 490, + 432 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Structural errors are hard to edit automatically. Besides database grounding, 26 of SmBoP's errors belong to another category, incorrect structure. These 26 samples contain 7 parses with incorrect SQL clauses and 19 parses with incorrect subqueries, but our CodeT5-PyDict+Program model only resolves 1 and 2 of them, respectively. We find that correcting such errors usually involves multiple edit steps, which motivates us to incorporate our model into an interactive framework in future work. As our experiments with simulated user interaction (Appendix E.2) show, when our model interacts with the simulated user to correct one clause at a time, it is able to fully correct more SQL parses. Thus, we deem interactive correction would maximize our model's utility in practice.", + "bbox": [ + 110, + 444, + 489, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5 Related Work", + "text_level": 1, + "bbox": [ + 112, + 715, + 270, + 730 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the release of CodeBERT (Feng et al., 2020), many language models of code have emerged for program understanding and generation (Ahmad et al., 2021; Chen et al., 2021; Guo et al., 2021; Wang et al., 2021; Guo et al., 2022; Fried et al., 2023; Nijkamp et al., 2023). In addition to program-related tasks, recent work shows they also excel at processing natural language structures. Using code as meaning representations (MRs), we can leverage language models of code in various tasks, such as commonsense reasoning (Madaan et al., 2022),", + "bbox": [ + 112, + 741, + 489, + 919 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "action planning (Singh et al., 2022), and event extraction (Wang et al., 2022). In fact, how to design MRs to reduce model learning difficulty is a salient research question in semantic parsing (Guo et al., 2019; Gan et al., 2021b; Nie et al., 2022).", + "bbox": [ + 507, + 239, + 884, + 319 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our work demonstrates that program-related tasks themselves can also benefit from code-based MRs. Specifically, we apply such MRs to SQL error correction, a variant of automatic program repair tasks (Tufano et al., 2019; Panthaplackel et al., 2022; Zhang et al., 2023). Although SQL is a code-based MR, it is much harder for models to learn compared to other MRs, such as FunQL and lambda calculus (Li et al., 2022). Consequently, without many SQL queries in their pre-training corpora, language models of code can underperform state-of-the-art text-to-SQL parsers. By converting SQL queries into Python dictionaries, we can explicitly represent their compositional substructures and define edit actions as programs, which reduces the learning difficulty for language models of code and yields better performance.", + "bbox": [ + 507, + 319, + 884, + 594 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6 Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 507, + 604, + 796, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This paper presents a study on developing a text-to-SQL error correction model with clause-level edits and different representations. Our comprehensive experiments demonstrate that clauses are better semantic units than tokens for editing SQL queries and mimicking patterns in code pre-training corpora helps better leverage language models of code. As a future direction, we plan to incorporate our model into interactive semantic parsing frameworks (Li et al., 2020; Yao et al., 2019, 2020; Zeng et al., 2020) by suggesting possible edits to users once a wrong parse is identified. In this way, users would more efficiently correct parse errors and get better assistance. We also plan to experiment with other language models of code (Fried et al., 2023; Nijkamp et al., 2023) and text-to-SQL datasets (Zelle and Mooney, 1996; Gan et al., 2021a) to verify the generalizability of our method.", + "bbox": [ + 507, + 629, + 884, + 919 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1363", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Limitations", + "text_level": 1, + "bbox": [ + 114, + 84, + 220, + 99 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Actual applications of our model. Our work assumes that input SQL queries to our model are always wrong. This assumption is more feasible in an interactive semantic parsing framework, where the users are expected to decide whether a SQL parse, accompanied by its natural language explanations (Elgohary et al., 2020, 2021; Narechinaia et al., 2021; Mo et al., 2022), has errors or not. Alternatively, to remove this assumption, it would be interesting for future work to study the performance of our error correction model in combination with an automatic error detection model (Chen et al., 2023).", + "bbox": [ + 112, + 108, + 492, + 317 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experiments with more language models of code.", + "text_level": 1, + "bbox": [ + 112, + 326, + 490, + 340 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We have only experimented with two language models of code, CodiT5 and CodeT5, both using T5-base (Raffel et al., 2020) as their underlying model architecture. It would be interesting to test how our conclusions generalize to other language models of code in the future. Based on the strong capabilities of large language models of code, such as Codex (Chen et al., 2021), InCoder (Fried et al., 2023), and CodeGen (Nijkamp et al., 2023), we believe that these models can better exploit their knowledge about data structures and their operations in Python. These models may perform even better on Text-to-SQL error correction with our proposed representations.", + "bbox": [ + 112, + 341, + 489, + 567 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 114, + 577, + 285, + 593 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We would like to thank the anonymous reviewers and colleagues from the OSU NLP group for their thoughtful comments. This research was supported in part by a sponsored award from Cisco Research, NSF IIS-1815674, NSF CAREER #1942980, NSF OAC-2112606, and Ohio Supercomputer Center (Center, 1987). The views and conclusions contained herein are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the U.S. government. The U.S. Government is authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notice herein. Ziru is also supported by The Ohio State University Graduate School through University Fellowship.", + "bbox": [ + 112, + 602, + 489, + 843 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 114, + 869, + 213, + 883 + ], + "page_idx": 5 + }, + { + "type": "ref_text", + "text": "Wasi Ahmad, Saikat Chakraborty, Baishakhi Ray, and Kai-Wei Chang. 2021. Unified pre-training for pro", + "bbox": [ + 114, + 891, + 489, + 919 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "gram understanding and generation. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2655-2668, Online. Association for Computational Linguistics.", + "Ohio Supercomputer Center. 1987. Ohio supercomputer center.", + "Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidi Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott Gray, Nick Ryder, Mikhail Pavlov, Alethea Power, Lukasz Kaiser, Mohammad Bavarian, Clemens Winter, Philippe Tillet, Felipe Petroski Such, Dave Cummings, Matthias Plappert, Fotios Chantzis, Elizabeth Barnes, Ariel Herbert-Voss, William Hebgen Guss, Alex Nichol, Alex Paino, Nikolas Tezak, Jie Tang, Igor Babuschkin, Suchir Balaji, Shantanu Jain, William Saunders, Christopher Hesse, Andrew N. Carr, Jan Leike, Josh Achiam, Vedant Misra, Evan Morikawa, Alec Radford, Matthew Knight, Miles Brundage, Mira Murati, Katie Mayer, Peter Welinder, Bob McGrew, Dario Amodei, Sam McCandlish, Ilya Sutskever, and Wojciech Zaremba. 2021. Evaluating large language models trained on code.", + "Shijie Chen, Ziru Chen, Huan Sun, and Yu Su. 2023. Error detection for text-to-sql semantic parsing.", + "Xiang Deng, Ahmed Hassan Awadallah, Christopher Meek, Oleksandr Polozov, Huan Sun, and Matthew Richardson. 2021. Structure-grounded pretraining for text-to-SQL. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1337-1350, Online. Association for Computational Linguistics.", + "Ahmed Elgohary, Saghar Hosseini, and Ahmed Hassan Awadallah. 2020. Speak to your parser: Interactive text-to-SQL with natural language feedback. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2065-2077, Online. Association for Computational Linguistics.", + "Ahmed Elgohary, Christopher Meek, Matthew Richardson, Adam Fourney, Gonzalo Ramos, and Ahmed Hassan Awadallah. 2021. NL-EDIT: Correcting semantic parse errors through natural language interaction. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5599–5610, Online. Association for Computational Linguistics.", + "Zhangyin Feng, Daya Guo, Duyu Tang, Nan Duan, Xiaocheng Feng, Ming Gong, Linjun Shou, Bing Qin, Ting Liu, Daxin Jiang, and Ming Zhou. 2020. CodeBERT: A pre-trained model for programming and natural languages. In Findings of the Association" + ], + "bbox": [ + 509, + 85, + 884, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1364", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for Computational Linguistics: EMNLP 2020, pages 1536-1547, Online. Association for Computational Linguistics.", + "Daniel Fried, Armen Aghajanyan, Jessy Lin, Sida Wang, Eric Wallace, Freda Shi, Ruiqi Zhong, Scott Yih, Luke Zettlemoyer, and Mike Lewis. 2023. Incoder: A generative model for code infilling and synthesis. In The Eleventh International Conference on Learning Representations.", + "Yujiang Gan, Xinyun Chen, Qiuping Huang, Matthew Purver, John R. Woodward, Jinxia Xie, and Pengsheng Huang. 2021a. Towards robustness of text-to-SQL models against synonym substitution. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 2505-2515, Online. Association for Computational Linguistics.", + "Yujuan Gan, Xinyun Chen, Jinxia Xie, Matthew Purver, John R. Woodward, John Drake, and Qiaofu Zhang. 2021b. Natural SQL: Making SQL easier to infer from natural language specifications. In *Findings of the Association for Computational Linguistics: EMNLP* 2021, pages 2030–2042, Punta Cana, Dominican Republic. Association for Computational Linguistics.", + "Daya Guo, Shuai Lu, Nan Duan, Yanlin Wang, Ming Zhou, and Jian Yin. 2022. UniXcoder: Unified cross-modal pre-training for code representation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7212-7225, Dublin, Ireland. Association for Computational Linguistics.", + "Daya Guo, Shuo Ren, Shuai Lu, Zhangyin Feng, Duyu Tang, Shujie Liu, Long Zhou, Nan Duan, Alexey Svyatkovskiy, Shengyu Fu, Michele Tufano, Shao Kun Deng, Colin Clement, Dawn Drain, Neel Sundaresan, Jian Yin, Daxin Jiang, and Ming Zhou. 2021. GraphCodeBERT: Pre-training code representations with data flow. In International Conference on Learning Representations.", + "Jiaqi Guo, Zecheng Zhan, Yan Gao, Yan Xiao, JianGuang Lou, Ting Liu, and Dongmei Zhang. 2019. Towards complex text-to-SQL in cross-domain database with intermediate representation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4524-4535, Florence, Italy. Association for Computational Linguistics.", + "Hamel Husain, Ho-Hsiang Wu, Tiferet Gazit, Miltiadis Allamanis, and Marc Brockschmidt. 2020. Code-searchnet challenge: Evaluating the state of semantic code search.", + "Magne Jorgensen and Martin Shepperd. 2007. A systematic review of software development cost estimation studies. IEEE Transactions on Software Engineering, 33(1):33-53." + ], + "bbox": [ + 115, + 85, + 489, + 917 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Yuntao Li, Bei Chen, Qian Liu, Yan Gao, Jian-Guang Lou, Yan Zhang, and Dongmei Zhang. 2020. \"what do you mean by that?\" a parser-independent interactive approach for enhancing text-to-SQL. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6913-6922, Online. Association for Computational Linguistics.", + "Zhenwen Li, Jiaqi Guo, Qian Liu, Jian-Guang Lou, and Tao Xie. 2022. Exploring the secrets behind the learning difficulty of meaning representations for semantic parsing. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 3616-3625, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.", + "Xi Victoria Lin, Richard Socher, and Caiming Xiong. 2020. Bridging textual and tabular data for cross-domain text-to-SQL semantic parsing. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 4870-4888, Online. Association for Computational Linguistics.", + "Aman Madaan, Shuyan Zhou, Uri Alon, Yiming Yang, and Graham Neubig. 2022. Language models of code are few-shot commonsense learners. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 1384-1403, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.", + "Quinn McNemar. 1947. Note on the sampling error of the difference between correlated proportions or percentages. In Psychometrika, volume 12, page 153-157.", + "Lingbo Mo, Ashley Lewis, Huan Sun, and Michael White. 2022. Towards transparent interactive semantic parsing via step-by-step correction. In *Findings of the Association for Computational Linguistics: ACL* 2022, pages 322-342, Dublin, Ireland. Association for Computational Linguistics.", + "Marius Mosbach, Maksym Andriushchenko, and Dietrich Klakow. 2021. On the stability of fine-tuning BERT: Misconceptions, explanations, and strong baselines. In International Conference on Learning Representations.", + "Arpit Narechania, Adam Fourney, Bongshin Lee, and Gonzalo Ramos. 2021. Diy: Assessing the correctness of natural language to sql systems. In 26th International Conference on Intelligent User Interfaces, IUI '21, page 597-607, New York, NY, USA. Association for Computing Machinery.", + "Lunyiu Nie, Shulin Cao, Jiaxin Shi, Jiuding Sun, Qi Tian, Lei Hou, Juanzi Li, and Jidong Zhai. 2022. GraphQ IR: Unifying the semantic parsing of graph query languages with one intermediate representation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 5848-5865, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics." + ], + "bbox": [ + 510, + 85, + 884, + 917 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1365", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Erik Nijkamp, Bo Pang, Hiroaki Hayashi, Lifu Tu, Huan Wang, Yingbo Zhou, Silvio Savarese, and Caiming Xiong. 2023. Codegen: An open large language model for code with multi-turn program synthesis. In The Eleventh International Conference on Learning Representations.", + "Sheena Panthaplackel, Milos Gligoric, Junyi Jessy Li, and Raymond Mooney. 2022. Using developer discussions to guide fixing bugs in software. In *Findings of the Association for Computational Linguistics: EMNLP* 2022, pages 2292-2301, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.", + "Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. 2019. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc.", + "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(140):1-67.", + "Ohad Rubin and Jonathan Berant. 2021. SmBoP: Semi-autoregressive bottom-up semantic parsing. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 311-324, Online. Association for Computational Linguistics.", + "Torsten Scholak, Nathan Schucher, and Dzmitry Bahdanau. 2021. PICARD: Parsing incrementally for constrained auto-regressive decoding from language models. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 9895-9901, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.", + "Noam Shazeer and Mitchell Stern. 2018. Adafactor: Adaptive learning rates with sublinear memory cost. In Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pages 4596-4604. PMLR.", + "Ishika Singh, Valts Blukis, Arsalan Mousavian, Ankit Goyal, Danfei Xu, Jonathan Tremblay, Dieter Fox, Jesse Thomason, and Animesh Garg. 2022. Progress: Generating situated robot task plans using large language models. In Workshop on Language and Robotics at CoRL 2022." + ], + "bbox": [ + 115, + 85, + 485, + 917 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Lappoon R. Tang and Raymond J. Mooney. 2000. Automated construction of database interfaces: Integrating statistical and relational learning for semantic parsing. In Proceedings of the 2000 Joint SIGDAT Conference on Empirical Methods in Natural Language Processing and Very Large Corpora: Held in Conjunction with the 38th Annual Meeting of the Association for Computational Linguistics - Volume 13, EMNLP '00, page 133-141, USA. Association for Computational Linguistics.", + "Michele Tufano, Jevgenija Pantiuchina, Cody Watson, Gabriele Bavota, and Denys Poshyvanyk. 2019. On learning meaningful code changes via neural machine translation. In Proceedings of the 41st International Conference on Software Engineering, ICSE '19, page 25-36. IEEE Press.", + "Bailin Wang, Richard Shin, Xiaodong Liu, Oleksandr Polozov, and Matthew Richardson. 2020. RAT-SQL: Relation-aware schema encoding and linking for text-to-SQL parsers. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7567-7578, Online. Association for Computational Linguistics.", + "Xingyao Wang, Sha Li, and Heng Ji. 2022. Code4struct: Code generation for few-shot structured prediction from natural language.", + "Yue Wang, Weishi Wang, Shafiq Joty, and Steven C.H. Hoi. 2021. CodeT5: Identifier-aware unified pretrained encoder-decoder models for code understanding and generation. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 8696-8708, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.", + "Cathrin Weiss, Rahul Premraj, Thomas Zimmermann, and Andreas Zeller. 2007. How long will it take to fix this bug? In Fourth International Workshop on Mining Software Repositories (MSR'07:ICSE Workshops 2007), pages 1-1.", + "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online. Association for Computational Linguistics.", + "Tianbao Xie, Chen Henry Wu, Peng Shi, Ruiqi Zhong, Torsten Scholak, Michihiro Yasunaga, Chien-Sheng Wu, Ming Zhong, Pengcheng Yin, Sida I. Wang, Victor Zhong, Bailin Wang, Chengzu Li, Connor Boyle, Ansong Ni, Ziyu Yao, Dragomir Radev, Caiming Xiong, Lingpeng Kong, Rui Zhang, Noah A. Smith, Luke Zettlemoyer, and Tao Yu. 2022. UnifiedSKG:" + ], + "bbox": [ + 510, + 85, + 880, + 917 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1366", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Unifying and multi-tasking structured knowledge grounding with text-to-text language models. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 602-631, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.", + "bbox": [ + 131, + 85, + 489, + 165 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Ziyu Yao, Yu Su, Huan Sun, and Wen-tau Yih. 2019. Model-based interactive semantic parsing: A unified framework and a text-to-SQL case study. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5447-5458, Hong Kong, China. Association for Computational Linguistics.", + "bbox": [ + 114, + 177, + 489, + 297 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Ziyu Yao, Yiqi Tang, Wen-tau Yih, Huan Sun, and Yu Su. 2020. An imitation game for learning semantic parsers from user interaction. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6883-6902, Online. Association for Computational Linguistics.", + "bbox": [ + 114, + 310, + 489, + 404 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Pengcheng Yin and Graham Neubig. 2017. A syntactic neural model for general-purpose code generation. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 440-450, Vancouver, Canada. Association for Computational Linguistics.", + "bbox": [ + 114, + 414, + 489, + 495 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Tao Yu, Rui Zhang, Kai Yang, Michihiro Yasunaga, Dongxu Wang, Zifan Li, James Ma, Irene Li, Qingning Yao, Shanelle Roman, Zilin Zhang, and Dragomir Radev. 2018. Spider: A large-scale human-labeled dataset for complex and cross-domain semantic parsing and text-to-SQL task. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 3911-3921, Brussels, Belgium. Association for Computational Linguistics.", + "bbox": [ + 114, + 508, + 489, + 629 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "John M. Zelle and Raymond J. Mooney. 1996. Learning to parse database queries using inductive logic programming. In Proceedings of the Thirteenth National Conference on Artificial Intelligence - Volume 2, AAAI'96, page 1050-1055. AAAI Press.", + "bbox": [ + 114, + 640, + 489, + 708 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Jichuan Zeng, Xi Victoria Lin, Steven C.H. Hoi, Richard Socher, Caiming Xiong, Michael Lyu, and Irwin King. 2020. Photon: A robust cross-domain text-to-SQL system. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 204-214, Online. Association for Computational Linguistics.", + "bbox": [ + 114, + 720, + 489, + 813 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Jiyang Zhang, Sheena Panthapackel, Pengyu Nie, Junyi Jessy Li, and Milos Gligoric. 2023. Coditt5: Pretraining for source code and natural language editing. In Proceedings of the 37th IEEE/ACM International Conference on Automated Software Engineering, ASE '22, New York, NY, USA. Association for Computing Machinery.", + "bbox": [ + 114, + 825, + 489, + 919 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Appendices", + "text_level": 1, + "bbox": [ + 509, + 83, + 615, + 101 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We provide more details omitted in the main text as follows:", + "bbox": [ + 507, + 111, + 882, + 141 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Appendix A: SQL PyDict Representation", + "- Appendix B: Text-to-SQL Parser Selection", + "- Appendix C: Implementation Details", + "- Appendix D: Statistical Significance Test", + "- Appendix E: Additional Results", + "- Appendix F: More Representation Examples" + ], + "bbox": [ + 531, + 143, + 877, + 240 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "A SQL PyDict Representation", + "text_level": 1, + "bbox": [ + 509, + 254, + 788, + 272 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We implement the transformation from any SQL query to our PyDict representation in three steps (Section 2.1). First, we use context-free grammar to parse a SQL query and obtain its abstract syntax tree (AST). The AST naturally contains a SQL decomposition where each clause has its unique subtree. In addition, if a clause contains a nested query, it would be represented as another independent subtree, which is a child of the root node in the clause's AST subtree. With these substructures explicitly represented, we use depth-first search to traverse through the AST to build our PyDict representation bottom-up. In other words, if a clause contains a subquery, we process the subquery tree as an independent SQL AST and build a dictionary for it. Then, we combine it with other substructures of the clause with different dictionary keys. For example, in Table F.1, we first build the dictionary for \"subquery0\" and assign this identifier as the key. In the main \"clause,\" we replace the subquery's corresponding span with this identifier. Finally, we use another dictionary to wrap the main \"clause\" and \"subquery0\" together as the final representation of the \"where\" clause. We repeat this procedure for each clause to incrementally add (key, value) pairs to the dictionary and \"store\" it to the variable sql, which we refer to in program edit representations.", + "bbox": [ + 505, + 282, + 885, + 718 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "B Text-to-SQL Parser Selection", + "text_level": 1, + "bbox": [ + 509, + 730, + 801, + 747 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We choose existing text-to-SQL parsers in our experiments according to two principles: the parsers predict database entity values, and they cover different decoding strategies, including grammar-based (BRIDGEv2), bottom-up (SmBop), and token-based (CodeT5). We did not include parsers using top-down decoders because they usually cannot predict entity values in conditional statements, such as RAT-SQL (Wang et al., 2020). Instead, we include BRIDGEv2 because its decoding method mimics", + "bbox": [ + 507, + 758, + 884, + 919 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1367", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "the left-to-right CFG derivation of a program, and it uses SQL syntax-based constraints to prevent grammatical errors. In recent work, such decoders, also used in PICARD (Scholak et al., 2021), are more popular than top-down decoders.", + "bbox": [ + 112, + 84, + 489, + 165 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "C Implementation Details", + "text_level": 1, + "bbox": [ + 112, + 179, + 356, + 195 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Our models (Section 3.2) are implemented in PyTorch (Paszke et al., 2019) using Huggingface (Wolf et al., 2020) and trained on a single NVIDIA RTX A6000 GPU (48GB). We use Adafactor (Shazeer and Stern, 2018) to train all our models with the same hyperparameters adapted from Mosbach et al. (2021):", + "bbox": [ + 112, + 206, + 489, + 317 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Learning rate: $3e - 5$", + "- Batch size: 16", + "- Epochs: 10", + "- Scheduler: Linear decay with $10\\%$ warmup" + ], + "bbox": [ + 136, + 319, + 473, + 384 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "D Statistical Significance Test", + "text_level": 1, + "bbox": [ + 112, + 398, + 386, + 414 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of our three clause-level edit representations (Section 4.1), we perform McNemar's Test (McNemar, 1947) to measure the statistical significance of their results in comparison to CodeT5-SQL+Token-Level. For each significance test between two models, we use the median results among our three runs to calculate the comparison matrix. Then, we compute the $p$ -values using statsmodels. When $p < 0.05$ we reject the null hypothesis. In other words, we consider the accuracy improvement statistically significant when $p < 0.05$ .", + "bbox": [ + 112, + 426, + 489, + 619 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "E Additional Results", + "text_level": 1, + "bbox": [ + 112, + 633, + 312, + 648 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Results on our development set. We report model performances on our held-out development set (Section 3.1) in Table E.1. During training, we select the best model by evaluating its EX and EM accuracy on the development set (Section 3.3) every 500 steps. Surprisingly, we find that CodeT5-SQL+Clause-Level sometimes achieves the best performance. For BRIDGEv2, it obtains 35.9 EM accuracy and 39.3 EX accuracy, while CodeT5-PyDict+Program only obtains 34.5 EM accuracy and 37.1 EX accuracy. A possible explanation is that in comparison to the test set, our development set has SQL structures and databases that are more", + "bbox": [ + 112, + 661, + 489, + 869 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "similar to the training set, while the test set has unseen SQL structures and less similar databases. It may also indicate that CodeT5-SQL+Clause-Level overfits the synthetic training data and fails to generalize to realistic test data.", + "bbox": [ + 507, + 84, + 884, + 164 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Results for simulated interaction experiments. To show the potential of using our model in an interactive framework, we extend our main experiments (Section 4.1) by adding simulated user interactions. Since our model uses beam search to decode the edit actions $\\mathbf{e} = \\{e_1,e_2,\\dots ,e_n\\}$ and the resulting correct SQL query $\\mathbf{q}_{+}$ (Equation 1), we simulate user interactions to select one edit action $e_i$ at a time from the beam results.", + "bbox": [ + 507, + 173, + 884, + 316 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "At each time step $t$ , we prompt the decoder with previously selected edit actions $e_1, \\ldots, e_{t-1}$ to complete the sequence $e_t, \\ldots, e_n$ , $\\mathbf{q}_+$ using beam search with size 3. Then, we use gold SQL annotations to simulate the user interaction, which selects an edit action $e_t$ from the three candidates at step $t$ or chooses to skip the current step when all three candidates are wrong. If skipping, the user continues to check the consequent edit actions $e_{t+j}$ ( $j = 1, 2, \\ldots, n-t$ ) until it selects the next edit action. When the interaction finishes, we append the selected edit action to the prompt and let the model regenerate a completion with the new prompt for the next step's interaction. Having simulated interactions for all edit actions, we do not use the generated $\\mathbf{q}_+$ directly because some edit actions are skipped. Instead, we execute the selected ones on the initial SQL query to derive the final query.", + "bbox": [ + 507, + 318, + 884, + 607 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "As shown in Table E.2, when collaborating with a simulated user, our error correction model can further improve the base parsers' accuracy. Compared to its performance without using any interactions, our model achieves up to 4.1 point more absolute improvement on EM accuracy (72.5 → 76.6; BRIDGEv2) and 5.0 point more absolute improvement on EX accuracy (73.1 → 78.1; BRIDGEv2). With these results for simulated interaction experiments, we deem that incorporating our error correction model into an interactive framework is a promising future direction.", + "bbox": [ + 507, + 608, + 885, + 801 + ], + "page_idx": 9 + }, + { + "type": "page_footnote", + "text": "4https://www.statsmodels.org/dev/generated/ statsmodels.stats.contingency_tables.mcnemar. html", + "bbox": [ + 112, + 879, + 462, + 916 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1368", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/9549c264b7e1e50603a0111fbdca78a8cf938b913e81b8b871f2aca463abdbc5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelsQueryEditCodeT5BRIDGEv2SmBoP
EMEXEMEXEMEX
CodiT5SQLToken-Level26.1 (0.4)28.6 (1.0)25.8 (0.3)27.2 (0.6)28.1 (0.9)30.7 (0.7)
SQLClause-Level28.6 (0.4)31.3 (0.5)28.4 (0.5)30.0 (0.2)30.2 (0.8)33.4 (0.8)
PyDictClause-Level28.9 (0.6)32.3 (0.8)28.0 (0.1)30.1 (0.2)27.6 (0.1)30.9 (0.4)
CodeT5SQLToken-Level32.1 (1.1)34.1 (1.2)31.8 (0.4)34.5 (0.8)34.2 (0.1)37.6 (0.1)
SQLClause-Level36.5 (0.6)38.6 (0.5)35.9 (0.4)39.3 (1.3)36.1 (0.6)38.8 (0.5)
PyDictClause-Level35.6 (0.9)37.9 (0.3)32.9 (1.0)34.8 (0.8)33.0 (0.2)36.3 (0.3)
CodeT5* CodeT5PyDictProgram35.7 (0.8)37.9 (0.3)34.8 (0.8)38.3 (0.7)36.0 (0.3)40.2 (0.5)
36.7 (0.2)38.5 (0.6)34.5 (0.1)37.1 (0.2)35.6 (0.8)39.0 (0.1)
", + "bbox": [ + 139, + 80, + 858, + 237 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/e2a4e56d20ed6972a1fce9e25790764a38d4c71c68877d0ceedef236f3e5fa83.jpg", + "table_caption": [ + "Table E.1: Exact Set Match (EM) and Execution Match (EX) accuracy on our held-out development set (Section 3.1). The best performances are in bold and the second bests are underlined. *We fine-tune the model to generate edit programs only (without resulting queries) and use Python interpreter to execute the edit actions." + ], + "table_footnote": [], + "table_body": "
ModelsQueryEditCodeT5BRIDGEv2SmBoP
EMEXEMEXEMEX
No EditN/AN/A62.7 (-)63.6 (-)70.1 (-)68.2 (-)74.6 (-)75.3 (-)
CodeT5*69.2 (0.4)68.4 (0.2)72.5 (0.4)73.1 (0.2)77.3 (0.4)77.6 (0.6)
CodeT5PyDictProgram69.0 (0.2)68.2 (0.1)72.5 (0.3)73.0 (0.6)78.0 (0.3)78.5 (0.3)
\\(CodeT5^{\\dagger}\\)PyDictProgram73.0 (0.7)72.9 (0.8)76.6 (0.4)78.1 (0.2)80.0 (0.3)81.2 (0.6)
", + "bbox": [ + 152, + 303, + 845, + 414 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table E.2: Exact Set Match (EM) and Execution Match (EX) accuracy on Spider development set. The best performances are in bold. *We fine-tune the model to generate edit programs only (without resulting queries) and use Python interpreter to execute the edit actions. †We simulate user interactions using gold SQL queries to choose edit actions during beam search (size 3) and then execute the chosen actions to get the resulting SQL parse.", + "bbox": [ + 112, + 423, + 882, + 481 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "F More Representation Examples", + "text_level": 1, + "bbox": [ + 112, + 505, + 423, + 521 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We provide two more examples in Table F.1 and F.2 to demonstrate how we represent SQL with subqueries and their edits (Section 2.2). We also show different representations for Insert and Delete edit actions.", + "bbox": [ + 112, + 530, + 487, + 609 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1369", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/c3604543de20f82a0e977c1f24315ce2790444838a2680660d4d4995647d42a4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Query RepresentationEdit Representation
SQLselect count(*) from cars_data where cars_data.accelerate > ( select max(cars_data.horsepower) from cars_data )Token-level<ReplaceOld> max(cars_data.horsepower) <ReplaceNew> cars_data.accelerate <ReplaceEnd> <Insert> order by cars_data.horsepower desc limit 1 <InsertEnd> Clause-level <ReplaceOld> select max(cars_data.horsepower) <ReplaceNew> select cars_data.accelerate <ReplaceEnd> <Insert> order by cars_data.horsepower desc limit 1 <InsertEnd>
PyDictsql = { "select": "select count(*)", "from": "from cars_data", "where": { "clause": "where cars_data.accelerate > (subquery0)", "subquery0": { "select": "select max(cars_data.horsepower)", "from": "from cars_data" } } }Clause-level<ReplaceOld> "select": "select max( cars_data.horsepower)" <ReplaceNew> "select": "select cars_data.accelerate" <ReplaceEnd> <Insert> "orderBy": "order by cars_data.horsepower desc", "limit": "limit 1" <InsertEnd> Programsql["where"},{"subquery0"},{"select"} = "select cars_data.accelerate" sql["where"},{"subquery0"},{"orderBy"} = "order by cars_data.horsepower desc" sql["where"},{"subquery0"},{"limit"} = "limit 1"
", + "bbox": [ + 115, + 128, + 884, + 425 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/d7e577ba3a90351d3d9c74bd227b07af3f0ce94a67ed638680ac7e44342462e1.jpg", + "table_caption": [ + "Table F.1: Example representations for a wrong SQL query that contains a nested subquery and its edit actions (including Insert edits). The corresponding natural language utterance is \"What is the number of cars with a greater acceleration than the one with the most horsepower?\"" + ], + "table_footnote": [], + "table_body": "
Query RepresentationEdit Representation
SQLselect employee.name from employee join evaluation on employee.employee_id = evaluation.employee_id group by evaluation.employee_id" order by sum(evaluationbonus) desc limit 1Token-level<Delete> group by evaluation.employee_id <DeleteEnd> <DeleteSum( <DeleteEnd><Delete>) <DeleteEnd>
Clause-level<Delete> group by evaluation.employee_id <DeleteEnd> <ReplaceOld> order by sum(evaluation;bONUS) desc <ReplaceNew> order by evaluation;bONUS desc <ReplaceEnd>
PyDictsql = { "select": "select employee.name", "from": "from employee join evaluation on employee.employee_id = evaluation.employee_id", "groupBy": "group by evaluation.employee_id", "orderBy": "order by sum(evaluation;bONUS) desc", "limit": "limit 1" }Clause-level<Delete> "groupId": "group by evaluation.employee_id" <DeleteEnd><ReplaceOld> "orderBy": "order by sum(evaluation;bONUS) desc" <ReplaceNew> "orderBy": "order by evaluation;bONUS desc" <ReplaceEnd>
Programsql.pop("groupId") sql["orderBy"] = "order by evaluation;bONUS desc"
", + "bbox": [ + 117, + 579, + 882, + 829 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table F.2: Example representations for a wrong SQL query and its edit actions (including Delete edits). The corresponding natural language utterance is \"Find the name of the employee who got the highest one time bonus.\"", + "bbox": [ + 112, + 838, + 882, + 868 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "1370", + "bbox": [ + 482, + 927, + 521, + 940 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A For every submission:", + "bbox": [ + 115, + 107, + 322, + 122 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "A1. Did you describe the limitations of your work? 6", + "A2. Did you discuss any potential risks of your work? Not applicable. Left blank.", + "A3. Do the abstract and introduction summarize the paper's main claims?", + "A4. Have you used AI writing assistants when working on this paper? Left blank." + ], + "bbox": [ + 127, + 126, + 695, + 287 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "B Did you use or create scientific artifacts?", + "bbox": [ + 114, + 298, + 489, + 316 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "3", + "bbox": [ + 134, + 321, + 146, + 332 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "B1. Did you cite the creators of artifacts you used? 3", + "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Not applicable. Left blank.", + "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Not applicable. Left blank.", + "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Not applicable. Left blank.", + "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? 3", + "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. 3" + ], + "bbox": [ + 127, + 346, + 880, + 752 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "C Did you run computational experiments?", + "bbox": [ + 114, + 764, + 492, + 781 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4", + "bbox": [ + 134, + 788, + 146, + 799 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Appendix B", + "bbox": [ + 129, + 810, + 880, + 860 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "ACL 2023 Responsible NLP Checklist", + "bbox": [ + 132, + 84, + 433, + 99 + ], + "page_idx": 12 + }, + { + "type": "page_footnote", + "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.", + "bbox": [ + 112, + 868, + 877, + 892 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "1371", + "bbox": [ + 482, + 928, + 517, + 940 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Appendix B", + "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?", + "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Appendix B" + ], + "bbox": [ + 129, + 83, + 878, + 282 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank.", + "bbox": [ + 112, + 293, + 877, + 330 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response.", + "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response.", + "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response.", + "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response.", + "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response." + ], + "bbox": [ + 127, + 340, + 880, + 640 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "1372", + "bbox": [ + 482, + 927, + 521, + 940 + ], + "page_idx": 13 + } +] \ No newline at end of file diff --git a/2023/Text-to-SQL Error Correction with Language Models of Code/63d877fb-bf81-4e0e-9af4-85d702a6c368_model.json b/2023/Text-to-SQL Error Correction with Language Models of Code/63d877fb-bf81-4e0e-9af4-85d702a6c368_model.json new file mode 100644 index 0000000000000000000000000000000000000000..756afff10198d272325b462bc7826fb9a110c5c4 --- /dev/null +++ b/2023/Text-to-SQL Error Correction with Language Models of Code/63d877fb-bf81-4e0e-9af4-85d702a6c368_model.json @@ -0,0 +1,2571 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.178, + 0.09, + 0.823, + 0.112 + ], + "angle": 0, + "content": "Text-to-SQL Error Correction with Language Models of Code" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.125, + 0.773, + 0.159 + ], + "angle": 0, + "content": "Ziru Chen\\(^{1}\\), Shijie Chen\\(^{1}\\), Michael White\\(^{1}\\), Raymond Mooney\\(^{2}\\), Ali Payani\\(^{3}\\), Jayanth Srinivasa\\(^{3}\\), Yu Su\\(^{1}\\), Huan Sun\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.392, + 0.16, + 0.612, + 0.176 + ], + "angle": 0, + "content": "1The Ohio State University" + }, + { + "type": "text", + "bbox": [ + 0.28, + 0.177, + 0.721, + 0.192 + ], + "angle": 0, + "content": "\\(^{2}\\)The University of Texas at Austin \\(^{3}\\)Cisco Research" + }, + { + "type": "text", + "bbox": [ + 0.198, + 0.193, + 0.806, + 0.21 + ], + "angle": 0, + "content": "{chen.8336, chen.10216, white.1240, su.809, sun.397}@osu.edu" + }, + { + "type": "text", + "bbox": [ + 0.26, + 0.211, + 0.742, + 0.226 + ], + "angle": 0, + "content": "mooney@cs.utexas.edu {apayani, jasriniv}@cisco" + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.253, + 0.341, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.277, + 0.461, + 0.547 + ], + "angle": 0, + "content": "Despite recent progress in text-to-SQL parsing, current semantic parsers are still not accurate enough for practical use. In this paper, we investigate how to build automatic text-to-SQL error correction models. Noticing that token-level edits are out of context and sometimes ambiguous, we propose building clause-level edit models instead. Besides, while most language models of code are not specifically pre-trained for SQL, they know common data structures and their operations in programming languages such as Python. Thus, we propose a novel representation for SQL queries and their edits that adheres more closely to the pre-training corpora of language models of code. Our error correction model improves the exact set match accuracy of different parsers by 2.4-6.5 and obtains up to 4.3 point absolute improvement over two strong baselines.\\(^1\\)" + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.556, + 0.26, + 0.57 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.581, + 0.49, + 0.852 + ], + "angle": 0, + "content": "Text-to-SQL parsing is a classic semantic parsing task that finds wide applications (Zelle and Mooney, 1996; Tang and Mooney, 2000). Since the release of Spider (Yu et al., 2018), a cross-database text-to-SQL benchmark, many semantic parsers with decent performance have been developed (Lin et al., 2020; Wang et al., 2020; Deng et al., 2021; Rubin and Berant, 2021; Scholak et al., 2021). Nonetheless, state-of-the-art semantic parsers are still not accurate enough. As a result, their users need to constantly correct wrongly predicted SQL queries, which can be as time-consuming and error-prone as writing a SQL query from scratch (Jorgensen and Shepperd, 2007; Weiss et al., 2007). Therefore, in this paper, we study the problem of automatic text-to-SQL error correction to better assist users in querying complex databases." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.855, + 0.489, + 0.886 + ], + "angle": 0, + "content": "We first highlight that it is essential to factor in the compositional substructures within SQL" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.254, + 0.885, + 0.543 + ], + "angle": 0, + "content": "queries, such as abstract syntax trees (Yin and Neubig, 2017; Guo et al., 2022) and data-flow graphs (Guo et al., 2021), instead of treating code snippets as string sequences. Compared to individual tokens, substructures (e.g. SQL clauses) include more context of the entire program and are more semantically meaningful. Consequently, edit patterns of such substructures are more intuitive for humans to understand and easier for language models to learn. Moreover, while the pre-training corpora for language models of code, such as CodeT5 (Wang et al., 2021), do not include many SQL queries based on their documentation, they naturally contain abundant examples of common data structures like dictionaries. Therefore, we hypothesize that transforming unfamiliar SQL queries into familiar data structures can help language models of code better perform structural editing of SQL queries." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.547, + 0.885, + 0.869 + ], + "angle": 0, + "content": "Based on these observations, we develop our error correction model and make two contributions. First, we propose considering SQL clauses instead of tokens as basic semantic units for editing. Using a context-free grammar, we can decompose a SQL query and identify its clauses by traversing its abstract syntax tree. Second, we propose a new representation of SQL queries and their edits that adheres more closely to common code pre-training corpora, including CodeSearchNet (Husain et al., 2020), and makes the structures of a SQL query more explicit. With a decomposed SQL query, we pair each clause with its SQL keyword and represent the entire query as a Python dictionary. Then, we format edits on a wrong SQL query as a program that modifies data of the query's corresponding dictionary. Unlike token-level edits in existing work (Zhang et al., 2023), such dictionary operations define all edits unambiguously and can be directly executed with a Python interpreter." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.871, + 0.884, + 0.92 + ], + "angle": 0, + "content": "Through comprehensive experiments with different representations, we show that: (1) our proposed representation has the lowest zero-shot perplexity" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.892, + 0.49, + 0.919 + ], + "angle": 0, + "content": "1Our code and data are available at https://github. com/OSU-NLP-Group/Auto-SQL-Correction." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1359" + }, + { + "type": "footer", + "bbox": [ + 0.227, + 0.946, + 0.772, + 0.958 + ], + "angle": 0, + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + }, + { + "type": "footer", + "bbox": [ + 0.369, + 0.959, + 0.631, + 0.972 + ], + "angle": 0, + "content": "Volume 2: Short Papers, pages 1359-1372" + }, + { + "type": "footer", + "bbox": [ + 0.297, + 0.973, + 0.702, + 0.986 + ], + "angle": 0, + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.118, + 0.082, + 0.885, + 0.236 + ], + "angle": 0, + "content": "
Query RepresentationEdit Representation
SQLselect tweets.text from tweets order by tweets.textToken-Level<ReplaceOld> tweets.text <ReplaceNew> tweets.createDate <ReplaceEnd>
Clause-Level<ReplaceOld> order by tweets.text <ReplaceNew> order by tweets.createDate <ReplaceEnd>
PyDictsql = { "select": "select tweets.text", "from": "from tweets", "orderBy": "order by tweets.text" }Clause-Level<ReplaceOld> "orderBy": "order by tweets.text" <ReplaceNew> "orderBy": "order by tweets.createDate" <ReplaceEnd> sql["orderBy"] = "order by tweets.createDate"
Program
" + }, + { + "type": "table_caption", + "bbox": [ + 0.112, + 0.246, + 0.884, + 0.304 + ], + "angle": 0, + "content": "Table 1: Example representations for a wrong SQL query and the Replace edit action. The corresponding natural language utterance is \"List the text of all tweets in the order of date.\" For token-level and clause-level representations, we format them as \" Span of wrong tokens/clauses Span of correct tokens/clauses \", where , , and are special tokens." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.316, + 0.489, + 0.412 + ], + "angle": 0, + "content": "with CodeT5; (2) simply changing token-level edits to clause-level edits can effectively improve the performance of our models; and (3) our method improves the exact set match accuracy of different parsers by 2.4-6.5 and obtains up to 4.3 point absolute improvement over two strong baselines." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.423, + 0.411, + 0.44 + ], + "angle": 0, + "content": "2 Text-to-SQL Error Correction" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.449, + 0.488, + 0.561 + ], + "angle": 0, + "content": "Given a natural language utterance \\(\\mathbf{u}\\), a database schema \\(\\mathbf{s}\\), and a wrong SQL query \\(\\mathbf{q}_{-}\\) produced by an existing parser, our goal is to develop an error correction model that predicts a sequence of edit actions \\(\\mathbf{e}\\) and the correct query \\(\\mathbf{q}_{+}\\). Following previous work (Zhang et al., 2023), we formulate our task as sequence-to-sequence generation:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.57, + 0.488, + 0.59 + ], + "angle": 0, + "content": "\\[\nP (\\mathbf {y} | \\mathbf {x}) = \\Pi_ {t = 1} ^ {T} P (\\mathbf {y} _ {t} | \\mathbf {x}, \\mathbf {y} _ {1: t - 1}) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.599, + 0.489, + 0.696 + ], + "angle": 0, + "content": "where \\(\\mathbf{x} = [\\mathbf{u};\\mathbf{s};\\mathbf{q}_{-}]\\) is the concatenation of the given inputs and \\(\\mathbf{y} = [\\mathbf{e};\\mathbf{q}_{+}]\\) is the concatenation of all edit actions and the resulting correct query. In this section, we study different representations of SQL queries (Section 2.1) and edits (Section 2.2) to better leverage language models of code." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.706, + 0.336, + 0.722 + ], + "angle": 0, + "content": "2.1 Query Representation" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.727, + 0.49, + 0.919 + ], + "angle": 0, + "content": "We consider two representations for a predicted query: (1) the original SQL format and (2) our proposed PyDict (Python Dictionary) representation. To prepare for editing, we disambiguate each SQL query following Rubin and Berant (2021), including lower-casing non-value tokens, resolving table references, and formatting punctuation. This preprocessing normalizes SQL queries predicted by different base parsers and the gold annotations into the same format. To build our PyDict representation, we parse a SQL query into its abstract syntax tree (AST) with Spider's context-free grammar. We" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.316, + 0.884, + 0.396 + ], + "angle": 0, + "content": "use depth-first search to traverse through the AST, find any nested substructures, and construct the dictionary representation bottom-up. Table 1 shows the \"SQL\" and \"PyDict\" representations of a SQL query (more details in Appendix A)." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.411, + 0.715, + 0.426 + ], + "angle": 0, + "content": "2.2 Edit Representation" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.434, + 0.885, + 0.789 + ], + "angle": 0, + "content": "We first follow Zhang et al. (2023) to use token-level edit representation with special tokens (Table 1), which have unique entries in the tokenizer and the model's embedding layer to describe Replace, Insert, and Delete edit actions (more examples in Appendix F). However, we realize this representation can sometimes be ambiguous. As shown in Table 1, the span \"tweets.text\" appears twice in the SQL query. This repetition would confuse the error correction model with which span to replace when generating the corrected query. Also, the ambiguity makes it difficult to implement rules and directly carry out the edit actions on the wrong query. Hence, we change the token-level edit representation to clause-level, which includes more context of the query to make different edits more distinguishable. In our experiments (Section 4.1), we demonstrate that this simple modification is already effective. Our program representation further improves the performance because it is more similar to the code pre-training corpora and eliminates the need to learn special tokens' representations." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.804, + 0.719, + 0.821 + ], + "angle": 0, + "content": "3 Experimental Setup" + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.832, + 0.881, + 0.848 + ], + "angle": 0, + "content": "3.1 Data Synthesis for SQL Error Correction" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.855, + 0.884, + 0.919 + ], + "angle": 0, + "content": "To train a text-to-SQL error correction model, we need to collect a set of wrong SQL parses that reflects a realistic distribution of errors (Section 4.2) as our training data. We synthesize this dataset by" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1360" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.122, + 0.082, + 0.482, + 0.189 + ], + "angle": 0, + "content": "
CodeT5BRIDGEv2SmBoP
# of Train47,02024,77620,083
# of Dev448448448
# of Test430392310
Avg. Train Edits2.343.112.72
Avg. Dev Edits2.703.293.31
Avg. Test Edits1.841.511.47
" + }, + { + "type": "table_caption", + "bbox": [ + 0.178, + 0.199, + 0.423, + 0.213 + ], + "angle": 0, + "content": "Table 2: Summary of data statistics." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.229, + 0.489, + 0.26 + ], + "angle": 0, + "content": "performing 5-fold cross-validation on each parser, which approximates the actual evaluation setting." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.262, + 0.49, + 0.614 + ], + "angle": 0, + "content": "Following the evaluation setup in Yu et al. (2018), we split Spider's training set into five roughly equal subsets by different databases. For each cross-validation fold, we train a text-to-SQL parser (Section 3.2) on four subsets and evaluate it on the remaining one. At inference time, we perform beam search with size 20 for each example and collect grammatical and executable parses in the beam. If a SQL parse is not an exact set match or execution match to the gold annotation, we label it wrong and include it in our training set for error correction. Having synthesized our training dataset, we randomly sample 8 databases and their associated questions to construct a held-out development set. For development set examples, we only keep incorrect SQL parses with the highest beam confidence. For our error correction test set, we train each parser on the full Spider training set and evaluate it on the original Spider's development set without modifications. We similarly keep SQL parses with exact match or execution match errors. Table 2 summarizes the statistics of our data." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.631, + 0.219, + 0.644 + ], + "angle": 0, + "content": "3.2 Models" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.654, + 0.489, + 0.717 + ], + "angle": 0, + "content": "Text-to-SQL base parsers. We choose three text-to-SQL parsers with different decoding strategies and levels of performance (Table 3). We elaborate on our selection criteria in Appendix B." + }, + { + "type": "text", + "bbox": [ + 0.136, + 0.72, + 0.488, + 0.783 + ], + "angle": 0, + "content": "- CodeT5 (Wang et al., 2021): We fine-tune CodeT5-base following Xie et al. (2022). This parser represents those using beam search decoding and having a lower accuracy." + }, + { + "type": "text", + "bbox": [ + 0.136, + 0.784, + 0.488, + 0.831 + ], + "angle": 0, + "content": "BRIDGEv2 (Lin et al., 2020): A representative parser with constrained decoding and achieving a medium-level accuracy." + }, + { + "type": "text", + "bbox": [ + 0.136, + 0.833, + 0.488, + 0.88 + ], + "angle": 0, + "content": "- SmBoP (Rubin and Berant, 2021): A representative parser with bottom-up decoding and achieving higher accuracy." + }, + { + "type": "list", + "bbox": [ + 0.136, + 0.72, + 0.488, + 0.88 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.892, + 0.488, + 0.919 + ], + "angle": 0, + "content": "Due to SmBoP's bottom-up decoding, we keep its original beam size and collect the top-20 unique beam predictions." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.085, + 0.882, + 0.116 + ], + "angle": 0, + "content": "Error correction models. We use two language models of code in all our experiments:" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.118, + 0.884, + 0.196 + ], + "angle": 0, + "content": "CoditT5 (Zhang et al., 2023): A language model pre-trained for code editing tasks by injecting noises to code snippets in CodeSearchNet (Husain et al., 2020) and then denoising with token-level edit representations." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.198, + 0.884, + 0.261 + ], + "angle": 0, + "content": "- CodeT5 (Wang et al., 2021): A language model pre-trained for general code understanding and generation with four different pre-training objectives." + }, + { + "type": "list", + "bbox": [ + 0.532, + 0.118, + 0.884, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.262, + 0.885, + 0.342 + ], + "angle": 0, + "content": "We compare the existing SQL+Token-Level representation with our proposed ones: SQL+Clause-Level, PyDict+Clause-Level, and PyDict+Program on CodeT5 and the first three on CoditT5.3 Implementation details are in Appendix C." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.353, + 0.642, + 0.367 + ], + "angle": 0, + "content": "3.3 Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.373, + 0.884, + 0.599 + ], + "angle": 0, + "content": "We use the increase in Exact Set Match (EM) and Execution Match (EX) accuracy on our error correction test set to measure each model's performance. Because CoditT5's experiments assume the input program has at least one error, we keep this assumption for fair comparisons. To construct a test set satisfying this assumption, we have to compare parser-generated SQL queries with gold annotations (Section 3.1). Thus, we use the Spider development set as our test set and split the Spider training set to build a held-out development set (Table 2) to select model checkpoints during training. We also include results on our held-out development set in the appendix (Table E.1)." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.61, + 0.724, + 0.626 + ], + "angle": 0, + "content": "4 Results and Analysis" + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.635, + 0.661, + 0.649 + ], + "angle": 0, + "content": "4.1 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.656, + 0.882, + 0.785 + ], + "angle": 0, + "content": "We summarize our main results in this section. To ensure robustness, we repeat all experiments with 3 different random seeds and report the average performances with standard deviations. Our model can also be used in an interactive framework that allows users to select edit actions from the top-\\(k\\) beam candidates. We include more experiments with simulated user interactions in Appendix E." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.794, + 0.885, + 0.874 + ], + "angle": 0, + "content": "Our representation's perplexity is the smallest. We validate that our PyDict+Program representation adheres more closely to the code pre-training corpora by measuring its zero-shot perplexity on CodeT5 using our development set (Section 3.1)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.508, + 0.881, + 0.884, + 0.919 + ], + "angle": 0, + "content": "3We did not use CoditT5 for PyDict+Program because it was pre-trained on token-level edit representations. Its decoder may be specialized in generating edits instead of programs." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.929, + 0.519, + 0.941 + ], + "angle": 0, + "content": "1361" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.121, + 0.082, + 0.878, + 0.26 + ], + "angle": 0, + "content": "
ModelsQueryEditCodeT5BRIDGEv2SmBoP
EMEXEMEXEMEX
No EditN/AN/A62.7 (-)63.6 (-)70.1 (-)68.2 (-)74.6 (-)75.3 (-)
CodiT5SQLToken-Level64.3 (0.1)64.4 (0.2)65.4 (0.5)66.6 (0.3)74.2 (0.4)75.3 (0.1)
SQLClause-Level67.0 (0.4)65.4 (0.5)71.3 (0.5)70.9 (0.2)76.3 (0.0)77.2 (0.3)
PyDictClause-Level67.1 (0.2)66.5 (0.4)70.6 (0.8)70.8 (0.6)76.3 (0.3)77.0 (0.3)
CodeT5SQLToken-Level66.7 (0.9)65.9 (0.5)68.2 (0.4)69.4 (0.8)75.6 (0.4)76.5 (0.6)
SQLClause-Level68.3 (0.3)\\( \\underline{68.2}^{+}(0.6) \\)71.8+(0.4)72.5+(0.2)76.7 (0.6)77.4 (0.3)
PyDictClause-Level66.6 (0.8)67.1 (0.8)72.0+(0.3)72.4+(0.2)77.3 (0.6)77.8 (0.2)
\\( CodeT5^* \\)CodeT5PyDictProgram\\( 69.2^{+}(0.4) \\)\\( 68.4^{+}(0.2) \\)\\( 72.5^{+}(0.4) \\)\\( 73.1^{+}(0.2) \\)77.3 (0.4)77.6 (0.6)
\\( 69.0^{+}(0.2) \\)\\( 68.2^{+}(0.1) \\)\\( 72.5^{+}(0.3) \\)\\( 73.0^{+}(0.6) \\)\\( 78.0^{+}(0.3) \\)\\( 78.5^{+}(0.3) \\)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.269, + 0.885, + 0.342 + ], + "angle": 0, + "content": "Table 3: Exact Set Match (EM) and Execution Match (EX) accuracy on Spider development set. The best performances are in bold and the second bests are underlined. Results with \\(^+\\) are statistically significant (McNemar's; \\(p < 0.05\\) ) compared to CodeT5-SQL+Token-Level (Appendix D). Otherwise, the results are not statistically significant. \\*We fine-tune the model to generate edit programs only (without resulting queries) and use Python interpreter to execute the edit actions." + }, + { + "type": "image", + "bbox": [ + 0.14, + 0.354, + 0.465, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.502, + 0.49, + 0.547 + ], + "angle": 0, + "content": "Figure 1: CodeT5's zero-shot perplexity (in log scale) of all four representations on our synthesized SQL error development set." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.557, + 0.49, + 0.669 + ], + "angle": 0, + "content": "As shown in Figure 1, by representing data in Py-Dict, we can reduce the perplexity of CodeT5 by 2 orders of magnitude. After augmenting it with our program representation, we further reduce the zero-shot perplexity of CodeT5 to only \\(5.96 \\times 10^{2}\\), 3 orders of magnitude less than the SQL+Token-Level representation \\((1.26 \\times 10^{5})\\)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.678, + 0.49, + 0.92 + ], + "angle": 0, + "content": "Clause-level editing is more effective, especially when represented in PyDict+Program. Since CodeT5 consistently outperforms CoditT5 with the same representations, we focus on comparisons among CodeT5 variations. As shown in Table 3, compared to CodeT5-SQL+Token-Level, only CodeT5-PyDict+Program achieves statistically significant improvement on all three parsers, while clause-level models fail McNemar's significance test for some parsers. More concretely, it achieves up to 4.3 point more absolute improvement on EM accuracy (68.2 → 72.5; BRIDGEv2) and 3.7 point more absolute improvement on EX accuracy (69.4 → 73.1; BRIDGEv2). Overall, CodeT5-PyDict+Program can boost the parsers' EM accu" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.353, + 0.883, + 0.403 + ], + "angle": 0, + "content": "racy by 2.4-6.5. Thus, both clause-level editing and PyDict+Program representation can better take advantage of language models of code." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.416, + 0.673, + 0.431 + ], + "angle": 0, + "content": "4.2 Error Analysis" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.439, + 0.883, + 0.487 + ], + "angle": 0, + "content": "Additionally, we conduct an error analysis (Table 4) by sampling 100 wrong parses from all three parsers and classifying them into five categories:" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.489, + 0.884, + 0.535 + ], + "angle": 0, + "content": "- Database Grounding: A generated SQL query has the correct structure, but some table/column names or entity values are wrong." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.537, + 0.882, + 0.567 + ], + "angle": 0, + "content": "- Incorrect Structure: A generated SQL query has missing, wrong, or redundant structures." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.569, + 0.882, + 0.6 + ], + "angle": 0, + "content": "- Syntax & Grammar: A generated SQL query violates the programming language's syntax." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.601, + 0.884, + 0.648 + ], + "angle": 0, + "content": "- False Negative: A generated SQL query is semantically correct but not captured by evaluation metrics, or the gold annotation is wrong." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.649, + 0.884, + 0.68 + ], + "angle": 0, + "content": "- Other: All other errors, such as wrong aggregation functions, besides the above categories." + }, + { + "type": "list", + "bbox": [ + 0.532, + 0.489, + 0.884, + 0.68 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.682, + 0.883, + 0.73 + ], + "angle": 0, + "content": "Since the error distributions for each parser are similar, as an example, we discuss our findings based on the strongest parser, SmBoP:" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.743, + 0.885, + 0.919 + ], + "angle": 0, + "content": "Database grounding is the major type of error. Among the 100 samples from SmBoP, we find that 54 of them have database grounding errors. Particularly, SmBoP predicts wrong table/column names in 34 parses, inaccurate entity values in 9 parses, and incorrect JOIN relations in 11 parses. Our CodeT5-PyDict+Program model can successfully fix 16 of the 54 erroneous parses, including 10 parses with wrong table/column names, 4 parses with inaccurate entity values, and 2 parses with incorrect JOIN relations. We hypothesize that" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1362" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.118, + 0.082, + 0.881, + 0.189 + ], + "angle": 0, + "content": "
Error CategoryCodeT5BRIDGEv2SmBoP
ResolvedUnresolvedAllResolvedUnresolvedAllResolvedUnresolvedAll
Database Grounding155166144862163854
Incorrect Structure215172121432326
Syntax & Grammar000000145
False Negative099066088
Other17821618167
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.199, + 0.883, + 0.228 + ], + "angle": 0, + "content": "Table 4: Analysis of 100 sample errors made by each text-to-SQL parser. We group the errors into 5 categories and examine if our CodeT5-PyDict+Program model resolves them." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.24, + 0.491, + 0.433 + ], + "angle": 0, + "content": "database grounding is also a major category of errors in our synthesized training set, so our model has learned to resolve similar errors. Nevertheless, it still cannot correct the remaining 38 SQL parses. We notice that our current representation for database schema is missing critical information, such as column data types and foreign key relations, for our error correction model to fix database grounding errors. Following our PyDict representation for SQL, we suggest designing a code representation for database schema that includes such information to tackle this issue in future work." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.445, + 0.49, + 0.703 + ], + "angle": 0, + "content": "Structural errors are hard to edit automatically. Besides database grounding, 26 of SmBoP's errors belong to another category, incorrect structure. These 26 samples contain 7 parses with incorrect SQL clauses and 19 parses with incorrect subqueries, but our CodeT5-PyDict+Program model only resolves 1 and 2 of them, respectively. We find that correcting such errors usually involves multiple edit steps, which motivates us to incorporate our model into an interactive framework in future work. As our experiments with simulated user interaction (Appendix E.2) show, when our model interacts with the simulated user to correct one clause at a time, it is able to fully correct more SQL parses. Thus, we deem interactive correction would maximize our model's utility in practice." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.716, + 0.271, + 0.731 + ], + "angle": 0, + "content": "5 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.743, + 0.49, + 0.92 + ], + "angle": 0, + "content": "Since the release of CodeBERT (Feng et al., 2020), many language models of code have emerged for program understanding and generation (Ahmad et al., 2021; Chen et al., 2021; Guo et al., 2021; Wang et al., 2021; Guo et al., 2022; Fried et al., 2023; Nijkamp et al., 2023). In addition to program-related tasks, recent work shows they also excel at processing natural language structures. Using code as meaning representations (MRs), we can leverage language models of code in various tasks, such as commonsense reasoning (Madaan et al., 2022)," + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.24, + 0.885, + 0.32 + ], + "angle": 0, + "content": "action planning (Singh et al., 2022), and event extraction (Wang et al., 2022). In fact, how to design MRs to reduce model learning difficulty is a salient research question in semantic parsing (Guo et al., 2019; Gan et al., 2021b; Nie et al., 2022)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.321, + 0.885, + 0.595 + ], + "angle": 0, + "content": "Our work demonstrates that program-related tasks themselves can also benefit from code-based MRs. Specifically, we apply such MRs to SQL error correction, a variant of automatic program repair tasks (Tufano et al., 2019; Panthaplackel et al., 2022; Zhang et al., 2023). Although SQL is a code-based MR, it is much harder for models to learn compared to other MRs, such as FunQL and lambda calculus (Li et al., 2022). Consequently, without many SQL queries in their pre-training corpora, language models of code can underperform state-of-the-art text-to-SQL parsers. By converting SQL queries into Python dictionaries, we can explicitly represent their compositional substructures and define edit actions as programs, which reduces the learning difficulty for language models of code and yields better performance." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.605, + 0.797, + 0.62 + ], + "angle": 0, + "content": "6 Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.63, + 0.885, + 0.92 + ], + "angle": 0, + "content": "This paper presents a study on developing a text-to-SQL error correction model with clause-level edits and different representations. Our comprehensive experiments demonstrate that clauses are better semantic units than tokens for editing SQL queries and mimicking patterns in code pre-training corpora helps better leverage language models of code. As a future direction, we plan to incorporate our model into interactive semantic parsing frameworks (Li et al., 2020; Yao et al., 2019, 2020; Zeng et al., 2020) by suggesting possible edits to users once a wrong parse is identified. In this way, users would more efficiently correct parse errors and get better assistance. We also plan to experiment with other language models of code (Fried et al., 2023; Nijkamp et al., 2023) and text-to-SQL datasets (Zelle and Mooney, 1996; Gan et al., 2021a) to verify the generalizability of our method." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1363" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.115, + 0.085, + 0.221, + 0.1 + ], + "angle": 0, + "content": "Limitations" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.109, + 0.493, + 0.318 + ], + "angle": 0, + "content": "Actual applications of our model. Our work assumes that input SQL queries to our model are always wrong. This assumption is more feasible in an interactive semantic parsing framework, where the users are expected to decide whether a SQL parse, accompanied by its natural language explanations (Elgohary et al., 2020, 2021; Narechinaia et al., 2021; Mo et al., 2022), has errors or not. Alternatively, to remove this assumption, it would be interesting for future work to study the performance of our error correction model in combination with an automatic error detection model (Chen et al., 2023)." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.327, + 0.492, + 0.341 + ], + "angle": 0, + "content": "Experiments with more language models of code." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.342, + 0.49, + 0.568 + ], + "angle": 0, + "content": "We have only experimented with two language models of code, CodiT5 and CodeT5, both using T5-base (Raffel et al., 2020) as their underlying model architecture. It would be interesting to test how our conclusions generalize to other language models of code in the future. Based on the strong capabilities of large language models of code, such as Codex (Chen et al., 2021), InCoder (Fried et al., 2023), and CodeGen (Nijkamp et al., 2023), we believe that these models can better exploit their knowledge about data structures and their operations in Python. These models may perform even better on Text-to-SQL error correction with our proposed representations." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.579, + 0.287, + 0.594 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.603, + 0.49, + 0.844 + ], + "angle": 0, + "content": "We would like to thank the anonymous reviewers and colleagues from the OSU NLP group for their thoughtful comments. This research was supported in part by a sponsored award from Cisco Research, NSF IIS-1815674, NSF CAREER #1942980, NSF OAC-2112606, and Ohio Supercomputer Center (Center, 1987). The views and conclusions contained herein are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the U.S. government. The U.S. Government is authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notice herein. Ziru is also supported by The Ohio State University Graduate School through University Fellowship." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.87, + 0.214, + 0.884 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.892, + 0.49, + 0.92 + ], + "angle": 0, + "content": "Wasi Ahmad, Saikat Chakraborty, Baishakhi Ray, and Kai-Wei Chang. 2021. Unified pre-training for pro" + }, + { + "type": "ref_text", + "bbox": [ + 0.527, + 0.086, + 0.885, + 0.153 + ], + "angle": 0, + "content": "gram understanding and generation. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2655-2668, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.162, + 0.883, + 0.189 + ], + "angle": 0, + "content": "Ohio Supercomputer Center. 1987. Ohio supercomputer center." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.199, + 0.885, + 0.461 + ], + "angle": 0, + "content": "Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidi Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott Gray, Nick Ryder, Mikhail Pavlov, Alethea Power, Lukasz Kaiser, Mohammad Bavarian, Clemens Winter, Philippe Tillet, Felipe Petroski Such, Dave Cummings, Matthias Plappert, Fotios Chantzis, Elizabeth Barnes, Ariel Herbert-Voss, William Hebgen Guss, Alex Nichol, Alex Paino, Nikolas Tezak, Jie Tang, Igor Babuschkin, Suchir Balaji, Shantanu Jain, William Saunders, Christopher Hesse, Andrew N. Carr, Jan Leike, Josh Achiam, Vedant Misra, Evan Morikawa, Alec Radford, Matthew Knight, Miles Brundage, Mira Murati, Katie Mayer, Peter Welinder, Bob McGrew, Dario Amodei, Sam McCandlish, Ilya Sutskever, and Wojciech Zaremba. 2021. Evaluating large language models trained on code." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.47, + 0.884, + 0.498 + ], + "angle": 0, + "content": "Shijie Chen, Ziru Chen, Huan Sun, and Yu Su. 2023. Error detection for text-to-sql semantic parsing." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.508, + 0.884, + 0.613 + ], + "angle": 0, + "content": "Xiang Deng, Ahmed Hassan Awadallah, Christopher Meek, Oleksandr Polozov, Huan Sun, and Matthew Richardson. 2021. Structure-grounded pretraining for text-to-SQL. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1337-1350, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.623, + 0.885, + 0.715 + ], + "angle": 0, + "content": "Ahmed Elgohary, Saghar Hosseini, and Ahmed Hassan Awadallah. 2020. Speak to your parser: Interactive text-to-SQL with natural language feedback. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2065-2077, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.725, + 0.885, + 0.844 + ], + "angle": 0, + "content": "Ahmed Elgohary, Christopher Meek, Matthew Richardson, Adam Fourney, Gonzalo Ramos, and Ahmed Hassan Awadallah. 2021. NL-EDIT: Correcting semantic parse errors through natural language interaction. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5599–5610, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.853, + 0.884, + 0.919 + ], + "angle": 0, + "content": "Zhangyin Feng, Daya Guo, Duyu Tang, Nan Duan, Xiaocheng Feng, Ming Gong, Linjun Shou, Bing Qin, Ting Liu, Daxin Jiang, and Ming Zhou. 2020. CodeBERT: A pre-trained model for programming and natural languages. In Findings of the Association" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.086, + 0.885, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1364" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.086, + 0.489, + 0.126 + ], + "angle": 0, + "content": "for Computational Linguistics: EMNLP 2020, pages 1536-1547, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.136, + 0.49, + 0.215 + ], + "angle": 0, + "content": "Daniel Fried, Armen Aghajanyan, Jessy Lin, Sida Wang, Eric Wallace, Freda Shi, Ruiqi Zhong, Scott Yih, Luke Zettlemoyer, and Mike Lewis. 2023. Incoder: A generative model for code infilling and synthesis. In The Eleventh International Conference on Learning Representations." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.226, + 0.49, + 0.357 + ], + "angle": 0, + "content": "Yujiang Gan, Xinyun Chen, Qiuping Huang, Matthew Purver, John R. Woodward, Jinxia Xie, and Pengsheng Huang. 2021a. Towards robustness of text-to-SQL models against synonym substitution. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 2505-2515, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.367, + 0.49, + 0.472 + ], + "angle": 0, + "content": "Yujuan Gan, Xinyun Chen, Jinxia Xie, Matthew Purver, John R. Woodward, John Drake, and Qiaofu Zhang. 2021b. Natural SQL: Making SQL easier to infer from natural language specifications. In *Findings of the Association for Computational Linguistics: EMNLP* 2021, pages 2030–2042, Punta Cana, Dominican Republic. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.482, + 0.49, + 0.575 + ], + "angle": 0, + "content": "Daya Guo, Shuai Lu, Nan Duan, Yanlin Wang, Ming Zhou, and Jian Yin. 2022. UniXcoder: Unified cross-modal pre-training for code representation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7212-7225, Dublin, Ireland. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.585, + 0.49, + 0.689 + ], + "angle": 0, + "content": "Daya Guo, Shuo Ren, Shuai Lu, Zhangyin Feng, Duyu Tang, Shujie Liu, Long Zhou, Nan Duan, Alexey Svyatkovskiy, Shengyu Fu, Michele Tufano, Shao Kun Deng, Colin Clement, Dawn Drain, Neel Sundaresan, Jian Yin, Daxin Jiang, and Ming Zhou. 2021. GraphCodeBERT: Pre-training code representations with data flow. In International Conference on Learning Representations." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.7, + 0.49, + 0.792 + ], + "angle": 0, + "content": "Jiaqi Guo, Zecheng Zhan, Yan Gao, Yan Xiao, JianGuang Lou, Ting Liu, and Dongmei Zhang. 2019. Towards complex text-to-SQL in cross-domain database with intermediate representation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4524-4535, Florence, Italy. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.803, + 0.489, + 0.854 + ], + "angle": 0, + "content": "Hamel Husain, Ho-Hsiang Wu, Tiferet Gazit, Miltiadis Allamanis, and Marc Brockschmidt. 2020. Code-searchnet challenge: Evaluating the state of semantic code search." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.866, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Magne Jorgensen and Martin Shepperd. 2007. A systematic review of software development cost estimation studies. IEEE Transactions on Software Engineering, 33(1):33-53." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.513, + 0.086, + 0.885, + 0.191 + ], + "angle": 0, + "content": "Yuntao Li, Bei Chen, Qian Liu, Yan Gao, Jian-Guang Lou, Yan Zhang, and Dongmei Zhang. 2020. \"what do you mean by that?\" a parser-independent interactive approach for enhancing text-to-SQL. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6913-6922, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.201, + 0.885, + 0.293 + ], + "angle": 0, + "content": "Zhenwen Li, Jiaqi Guo, Qian Liu, Jian-Guang Lou, and Tao Xie. 2022. Exploring the secrets behind the learning difficulty of meaning representations for semantic parsing. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 3616-3625, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.304, + 0.885, + 0.383 + ], + "angle": 0, + "content": "Xi Victoria Lin, Richard Socher, and Caiming Xiong. 2020. Bridging textual and tabular data for cross-domain text-to-SQL semantic parsing. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 4870-4888, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.393, + 0.885, + 0.485 + ], + "angle": 0, + "content": "Aman Madaan, Shuyan Zhou, Uri Alon, Yiming Yang, and Graham Neubig. 2022. Language models of code are few-shot commonsense learners. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 1384-1403, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.496, + 0.885, + 0.547 + ], + "angle": 0, + "content": "Quinn McNemar. 1947. Note on the sampling error of the difference between correlated proportions or percentages. In Psychometrika, volume 12, page 153-157." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.559, + 0.885, + 0.638 + ], + "angle": 0, + "content": "Lingbo Mo, Ashley Lewis, Huan Sun, and Michael White. 2022. Towards transparent interactive semantic parsing via step-by-step correction. In *Findings of the Association for Computational Linguistics: ACL* 2022, pages 322-342, Dublin, Ireland. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.648, + 0.885, + 0.714 + ], + "angle": 0, + "content": "Marius Mosbach, Maksym Andriushchenko, and Dietrich Klakow. 2021. On the stability of fine-tuning BERT: Misconceptions, explanations, and strong baselines. In International Conference on Learning Representations." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.724, + 0.885, + 0.803 + ], + "angle": 0, + "content": "Arpit Narechania, Adam Fourney, Bongshin Lee, and Gonzalo Ramos. 2021. Diy: Assessing the correctness of natural language to sql systems. In 26th International Conference on Intelligent User Interfaces, IUI '21, page 597-607, New York, NY, USA. Association for Computing Machinery." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.814, + 0.885, + 0.919 + ], + "angle": 0, + "content": "Lunyiu Nie, Shulin Cao, Jiaxin Shi, Jiuding Sun, Qi Tian, Lei Hou, Juanzi Li, and Jidong Zhai. 2022. GraphQ IR: Unifying the semantic parsing of graph query languages with one intermediate representation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 5848-5865, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.885, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1365" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.086, + 0.487, + 0.165 + ], + "angle": 0, + "content": "Erik Nijkamp, Bo Pang, Hiroaki Hayashi, Lifu Tu, Huan Wang, Yingbo Zhou, Silvio Savarese, and Caiming Xiong. 2023. Codegen: An open large language model for code with multi-turn program synthesis. In The Eleventh International Conference on Learning Representations." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.179, + 0.487, + 0.271 + ], + "angle": 0, + "content": "Sheena Panthaplackel, Milos Gligoric, Junyi Jessy Li, and Raymond Mooney. 2022. Using developer discussions to guide fixing bugs in software. In *Findings of the Association for Computational Linguistics: EMNLP* 2022, pages 2292-2301, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.285, + 0.487, + 0.415 + ], + "angle": 0, + "content": "Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. 2019. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.43, + 0.487, + 0.508 + ], + "angle": 0, + "content": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(140):1-67." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.523, + 0.487, + 0.614 + ], + "angle": 0, + "content": "Ohad Rubin and Jonathan Berant. 2021. SmBoP: Semi-autoregressive bottom-up semantic parsing. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 311-324, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.628, + 0.487, + 0.733 + ], + "angle": 0, + "content": "Torsten Scholak, Nathan Schucher, and Dzmitry Bahdanau. 2021. PICARD: Parsing incrementally for constrained auto-regressive decoding from language models. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 9895-9901, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.747, + 0.487, + 0.824 + ], + "angle": 0, + "content": "Noam Shazeer and Mitchell Stern. 2018. Adafactor: Adaptive learning rates with sublinear memory cost. In Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pages 4596-4604. PMLR." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.84, + 0.487, + 0.918 + ], + "angle": 0, + "content": "Ishika Singh, Valts Blukis, Arsalan Mousavian, Ankit Goyal, Danfei Xu, Jonathan Tremblay, Dieter Fox, Jesse Thomason, and Animesh Garg. 2022. Progress: Generating situated robot task plans using large language models. In Workshop on Language and Robotics at CoRL 2022." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.487, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.086, + 0.882, + 0.216 + ], + "angle": 0, + "content": "Lappoon R. Tang and Raymond J. Mooney. 2000. Automated construction of database interfaces: Integrating statistical and relational learning for semantic parsing. In Proceedings of the 2000 Joint SIGDAT Conference on Empirical Methods in Natural Language Processing and Very Large Corpora: Held in Conjunction with the 38th Annual Meeting of the Association for Computational Linguistics - Volume 13, EMNLP '00, page 133-141, USA. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.227, + 0.882, + 0.305 + ], + "angle": 0, + "content": "Michele Tufano, Jevgenija Pantiuchina, Cody Watson, Gabriele Bavota, and Denys Poshyvanyk. 2019. On learning meaningful code changes via neural machine translation. In Proceedings of the 41st International Conference on Software Engineering, ICSE '19, page 25-36. IEEE Press." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.317, + 0.882, + 0.408 + ], + "angle": 0, + "content": "Bailin Wang, Richard Shin, Xiaodong Liu, Oleksandr Polozov, and Matthew Richardson. 2020. RAT-SQL: Relation-aware schema encoding and linking for text-to-SQL parsers. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7567-7578, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.419, + 0.882, + 0.459 + ], + "angle": 0, + "content": "Xingyao Wang, Sha Li, and Heng Ji. 2022. Code4struct: Code generation for few-shot structured prediction from natural language." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.468, + 0.882, + 0.573 + ], + "angle": 0, + "content": "Yue Wang, Weishi Wang, Shafiq Joty, and Steven C.H. Hoi. 2021. CodeT5: Identifier-aware unified pretrained encoder-decoder models for code understanding and generation. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 8696-8708, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.584, + 0.882, + 0.649 + ], + "angle": 0, + "content": "Cathrin Weiss, Rahul Premraj, Thomas Zimmermann, and Andreas Zeller. 2007. How long will it take to fix this bug? In Fourth International Workshop on Mining Software Repositories (MSR'07:ICSE Workshops 2007), pages 1-1." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.66, + 0.882, + 0.816 + ], + "angle": 0, + "content": "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.827, + 0.882, + 0.918 + ], + "angle": 0, + "content": "Tianbao Xie, Chen Henry Wu, Peng Shi, Ruiqi Zhong, Torsten Scholak, Michihiro Yasunaga, Chien-Sheng Wu, Ming Zhong, Pengcheng Yin, Sida I. Wang, Victor Zhong, Bailin Wang, Chengzu Li, Connor Boyle, Ansong Ni, Ziyu Yao, Dragomir Radev, Caiming Xiong, Lingpeng Kong, Rui Zhang, Noah A. Smith, Luke Zettlemoyer, and Tao Yu. 2022. UnifiedSKG:" + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.882, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1366" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.132, + 0.086, + 0.49, + 0.166 + ], + "angle": 0, + "content": "Unifying and multi-tasking structured knowledge grounding with text-to-text language models. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 602-631, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics." + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.178, + 0.49, + 0.298 + ], + "angle": 0, + "content": "Ziyu Yao, Yu Su, Huan Sun, and Wen-tau Yih. 2019. Model-based interactive semantic parsing: A unified framework and a text-to-SQL case study. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5447-5458, Hong Kong, China. Association for Computational Linguistics." + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.311, + 0.49, + 0.405 + ], + "angle": 0, + "content": "Ziyu Yao, Yiqi Tang, Wen-tau Yih, Huan Sun, and Yu Su. 2020. An imitation game for learning semantic parsers from user interaction. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6883-6902, Online. Association for Computational Linguistics." + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.416, + 0.49, + 0.497 + ], + "angle": 0, + "content": "Pengcheng Yin and Graham Neubig. 2017. A syntactic neural model for general-purpose code generation. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 440-450, Vancouver, Canada. Association for Computational Linguistics." + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.509, + 0.49, + 0.63 + ], + "angle": 0, + "content": "Tao Yu, Rui Zhang, Kai Yang, Michihiro Yasunaga, Dongxu Wang, Zifan Li, James Ma, Irene Li, Qingning Yao, Shanelle Roman, Zilin Zhang, and Dragomir Radev. 2018. Spider: A large-scale human-labeled dataset for complex and cross-domain semantic parsing and text-to-SQL task. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 3911-3921, Brussels, Belgium. Association for Computational Linguistics." + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.641, + 0.49, + 0.709 + ], + "angle": 0, + "content": "John M. Zelle and Raymond J. Mooney. 1996. Learning to parse database queries using inductive logic programming. In Proceedings of the Thirteenth National Conference on Artificial Intelligence - Volume 2, AAAI'96, page 1050-1055. AAAI Press." + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.721, + 0.49, + 0.814 + ], + "angle": 0, + "content": "Jichuan Zeng, Xi Victoria Lin, Steven C.H. Hoi, Richard Socher, Caiming Xiong, Michael Lyu, and Irwin King. 2020. Photon: A robust cross-domain text-to-SQL system. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 204-214, Online. Association for Computational Linguistics." + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.826, + 0.49, + 0.92 + ], + "angle": 0, + "content": "Jiyang Zhang, Sheena Panthapackel, Pengyu Nie, Junyi Jessy Li, and Milos Gligoric. 2023. Coditt5: Pretraining for source code and natural language editing. In Proceedings of the 37th IEEE/ACM International Conference on Automated Software Engineering, ASE '22, New York, NY, USA. Association for Computing Machinery." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.084, + 0.616, + 0.102 + ], + "angle": 0, + "content": "Appendices" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.112, + 0.883, + 0.142 + ], + "angle": 0, + "content": "We provide more details omitted in the main text as follows:" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.145, + 0.856, + 0.16 + ], + "angle": 0, + "content": "- Appendix A: SQL PyDict Representation" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.162, + 0.865, + 0.176 + ], + "angle": 0, + "content": "- Appendix B: Text-to-SQL Parser Selection" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.178, + 0.823, + 0.192 + ], + "angle": 0, + "content": "- Appendix C: Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.194, + 0.854, + 0.208 + ], + "angle": 0, + "content": "- Appendix D: Statistical Significance Test" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.21, + 0.786, + 0.224 + ], + "angle": 0, + "content": "- Appendix E: Additional Results" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.226, + 0.878, + 0.241 + ], + "angle": 0, + "content": "- Appendix F: More Representation Examples" + }, + { + "type": "list", + "bbox": [ + 0.532, + 0.145, + 0.878, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.255, + 0.789, + 0.273 + ], + "angle": 0, + "content": "A SQL PyDict Representation" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.283, + 0.886, + 0.719 + ], + "angle": 0, + "content": "We implement the transformation from any SQL query to our PyDict representation in three steps (Section 2.1). First, we use context-free grammar to parse a SQL query and obtain its abstract syntax tree (AST). The AST naturally contains a SQL decomposition where each clause has its unique subtree. In addition, if a clause contains a nested query, it would be represented as another independent subtree, which is a child of the root node in the clause's AST subtree. With these substructures explicitly represented, we use depth-first search to traverse through the AST to build our PyDict representation bottom-up. In other words, if a clause contains a subquery, we process the subquery tree as an independent SQL AST and build a dictionary for it. Then, we combine it with other substructures of the clause with different dictionary keys. For example, in Table F.1, we first build the dictionary for \"subquery0\" and assign this identifier as the key. In the main \"clause,\" we replace the subquery's corresponding span with this identifier. Finally, we use another dictionary to wrap the main \"clause\" and \"subquery0\" together as the final representation of the \"where\" clause. We repeat this procedure for each clause to incrementally add (key, value) pairs to the dictionary and \"store\" it to the variable sql, which we refer to in program edit representations." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.731, + 0.802, + 0.748 + ], + "angle": 0, + "content": "B Text-to-SQL Parser Selection" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.759, + 0.885, + 0.92 + ], + "angle": 0, + "content": "We choose existing text-to-SQL parsers in our experiments according to two principles: the parsers predict database entity values, and they cover different decoding strategies, including grammar-based (BRIDGEv2), bottom-up (SmBop), and token-based (CodeT5). We did not include parsers using top-down decoders because they usually cannot predict entity values in conditional statements, such as RAT-SQL (Wang et al., 2020). Instead, we include BRIDGEv2 because its decoding method mimics" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1367" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.49, + 0.166 + ], + "angle": 0, + "content": "the left-to-right CFG derivation of a program, and it uses SQL syntax-based constraints to prevent grammatical errors. In recent work, such decoders, also used in PICARD (Scholak et al., 2021), are more popular than top-down decoders." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.18, + 0.357, + 0.196 + ], + "angle": 0, + "content": "C Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.207, + 0.49, + 0.318 + ], + "angle": 0, + "content": "Our models (Section 3.2) are implemented in PyTorch (Paszke et al., 2019) using Huggingface (Wolf et al., 2020) and trained on a single NVIDIA RTX A6000 GPU (48GB). We use Adafactor (Shazeer and Stern, 2018) to train all our models with the same hyperparameters adapted from Mosbach et al. (2021):" + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.321, + 0.315, + 0.335 + ], + "angle": 0, + "content": "- Learning rate: \\(3e - 5\\)" + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.337, + 0.263, + 0.351 + ], + "angle": 0, + "content": "- Batch size: 16" + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.354, + 0.242, + 0.367 + ], + "angle": 0, + "content": "- Epochs: 10" + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.369, + 0.475, + 0.385 + ], + "angle": 0, + "content": "- Scheduler: Linear decay with \\(10\\%\\) warmup" + }, + { + "type": "list", + "bbox": [ + 0.137, + 0.321, + 0.475, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.399, + 0.388, + 0.416 + ], + "angle": 0, + "content": "D Statistical Significance Test" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.427, + 0.49, + 0.62 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of our three clause-level edit representations (Section 4.1), we perform McNemar's Test (McNemar, 1947) to measure the statistical significance of their results in comparison to CodeT5-SQL+Token-Level. For each significance test between two models, we use the median results among our three runs to calculate the comparison matrix. Then, we compute the \\( p \\)-values using statsmodels. When \\( p < 0.05 \\) we reject the null hypothesis. In other words, we consider the accuracy improvement statistically significant when \\( p < 0.05 \\)." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.634, + 0.314, + 0.649 + ], + "angle": 0, + "content": "E Additional Results" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.662, + 0.49, + 0.87 + ], + "angle": 0, + "content": "Results on our development set. We report model performances on our held-out development set (Section 3.1) in Table E.1. During training, we select the best model by evaluating its EX and EM accuracy on the development set (Section 3.3) every 500 steps. Surprisingly, we find that CodeT5-SQL+Clause-Level sometimes achieves the best performance. For BRIDGEv2, it obtains 35.9 EM accuracy and 39.3 EX accuracy, while CodeT5-PyDict+Program only obtains 34.5 EM accuracy and 37.1 EX accuracy. A possible explanation is that in comparison to the test set, our development set has SQL structures and databases that are more" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.165 + ], + "angle": 0, + "content": "similar to the training set, while the test set has unseen SQL structures and less similar databases. It may also indicate that CodeT5-SQL+Clause-Level overfits the synthetic training data and fails to generalize to realistic test data." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.174, + 0.885, + 0.317 + ], + "angle": 0, + "content": "Results for simulated interaction experiments. To show the potential of using our model in an interactive framework, we extend our main experiments (Section 4.1) by adding simulated user interactions. Since our model uses beam search to decode the edit actions \\(\\mathbf{e} = \\{e_1,e_2,\\dots ,e_n\\}\\) and the resulting correct SQL query \\(\\mathbf{q}_{+}\\) (Equation 1), we simulate user interactions to select one edit action \\(e_i\\) at a time from the beam results." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.319, + 0.885, + 0.608 + ], + "angle": 0, + "content": "At each time step \\( t \\), we prompt the decoder with previously selected edit actions \\( e_1, \\ldots, e_{t-1} \\) to complete the sequence \\( e_t, \\ldots, e_n \\), \\( \\mathbf{q}_+ \\) using beam search with size 3. Then, we use gold SQL annotations to simulate the user interaction, which selects an edit action \\( e_t \\) from the three candidates at step \\( t \\) or chooses to skip the current step when all three candidates are wrong. If skipping, the user continues to check the consequent edit actions \\( e_{t+j} \\) (\\( j = 1, 2, \\ldots, n-t \\)) until it selects the next edit action. When the interaction finishes, we append the selected edit action to the prompt and let the model regenerate a completion with the new prompt for the next step's interaction. Having simulated interactions for all edit actions, we do not use the generated \\( \\mathbf{q}_+ \\) directly because some edit actions are skipped. Instead, we execute the selected ones on the initial SQL query to derive the final query." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.609, + 0.887, + 0.802 + ], + "angle": 0, + "content": "As shown in Table E.2, when collaborating with a simulated user, our error correction model can further improve the base parsers' accuracy. Compared to its performance without using any interactions, our model achieves up to 4.1 point more absolute improvement on EM accuracy (72.5 → 76.6; BRIDGEv2) and 5.0 point more absolute improvement on EX accuracy (73.1 → 78.1; BRIDGEv2). With these results for simulated interaction experiments, we deem that incorporating our error correction model into an interactive framework is a promising future direction." + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.881, + 0.463, + 0.917 + ], + "angle": 0, + "content": "4https://www.statsmodels.org/dev/generated/ statsmodels.stats.contingency_tables.mcnemar. html" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1368" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.14, + 0.082, + 0.859, + 0.239 + ], + "angle": 0, + "content": "
ModelsQueryEditCodeT5BRIDGEv2SmBoP
EMEXEMEXEMEX
CodiT5SQLToken-Level26.1 (0.4)28.6 (1.0)25.8 (0.3)27.2 (0.6)28.1 (0.9)30.7 (0.7)
SQLClause-Level28.6 (0.4)31.3 (0.5)28.4 (0.5)30.0 (0.2)30.2 (0.8)33.4 (0.8)
PyDictClause-Level28.9 (0.6)32.3 (0.8)28.0 (0.1)30.1 (0.2)27.6 (0.1)30.9 (0.4)
CodeT5SQLToken-Level32.1 (1.1)34.1 (1.2)31.8 (0.4)34.5 (0.8)34.2 (0.1)37.6 (0.1)
SQLClause-Level36.5 (0.6)38.6 (0.5)35.9 (0.4)39.3 (1.3)36.1 (0.6)38.8 (0.5)
PyDictClause-Level35.6 (0.9)37.9 (0.3)32.9 (1.0)34.8 (0.8)33.0 (0.2)36.3 (0.3)
CodeT5* CodeT5PyDictProgram35.7 (0.8)37.9 (0.3)34.8 (0.8)38.3 (0.7)36.0 (0.3)40.2 (0.5)
36.7 (0.2)38.5 (0.6)34.5 (0.1)37.1 (0.2)35.6 (0.8)39.0 (0.1)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.248, + 0.883, + 0.293 + ], + "angle": 0, + "content": "Table E.1: Exact Set Match (EM) and Execution Match (EX) accuracy on our held-out development set (Section 3.1). The best performances are in bold and the second bests are underlined. *We fine-tune the model to generate edit programs only (without resulting queries) and use Python interpreter to execute the edit actions." + }, + { + "type": "table", + "bbox": [ + 0.153, + 0.304, + 0.846, + 0.415 + ], + "angle": 0, + "content": "
ModelsQueryEditCodeT5BRIDGEv2SmBoP
EMEXEMEXEMEX
No EditN/AN/A62.7 (-)63.6 (-)70.1 (-)68.2 (-)74.6 (-)75.3 (-)
CodeT5*69.2 (0.4)68.4 (0.2)72.5 (0.4)73.1 (0.2)77.3 (0.4)77.6 (0.6)
CodeT5PyDictProgram69.0 (0.2)68.2 (0.1)72.5 (0.3)73.0 (0.6)78.0 (0.3)78.5 (0.3)
\\(CodeT5^{\\dagger}\\)PyDictProgram73.0 (0.7)72.9 (0.8)76.6 (0.4)78.1 (0.2)80.0 (0.3)81.2 (0.6)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.424, + 0.884, + 0.482 + ], + "angle": 0, + "content": "Table E.2: Exact Set Match (EM) and Execution Match (EX) accuracy on Spider development set. The best performances are in bold. *We fine-tune the model to generate edit programs only (without resulting queries) and use Python interpreter to execute the edit actions. †We simulate user interactions using gold SQL queries to choose edit actions during beam search (size 3) and then execute the chosen actions to get the resulting SQL parse." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.506, + 0.424, + 0.523 + ], + "angle": 0, + "content": "F More Representation Examples" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.531, + 0.489, + 0.61 + ], + "angle": 0, + "content": "We provide two more examples in Table F.1 and F.2 to demonstrate how we represent SQL with subqueries and their edits (Section 2.2). We also show different representations for Insert and Delete edit actions." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1369" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.117, + 0.129, + 0.885, + 0.426 + ], + "angle": 0, + "content": "
Query RepresentationEdit Representation
SQLselect count(*) from cars_data where cars_data.accelerate > ( select max(cars_data.horsepower) from cars_data )Token-level<ReplaceOld> max(cars_data.horsepower) <ReplaceNew> cars_data.accelerate <ReplaceEnd> <Insert> order by cars_data.horsepower desc limit 1 <InsertEnd> Clause-level <ReplaceOld> select max(cars_data.horsepower) <ReplaceNew> select cars_data.accelerate <ReplaceEnd> <Insert> order by cars_data.horsepower desc limit 1 <InsertEnd>
PyDictsql = { "select": "select count(*)", "from": "from cars_data", "where": { "clause": "where cars_data.accelerate > (subquery0)", "subquery0": { "select": "select max(cars_data.horsepower)", "from": "from cars_data" } } }Clause-level<ReplaceOld> "select": "select max( cars_data.horsepower)" <ReplaceNew> "select": "select cars_data.accelerate" <ReplaceEnd> <Insert> "orderBy": "order by cars_data.horsepower desc", "limit": "limit 1" <InsertEnd> Programsql["where"},{"subquery0"},{"select"} = "select cars_data.accelerate" sql["where"},{"subquery0"},{"orderBy"} = "order by cars_data.horsepower desc" sql["where"},{"subquery0"},{"limit"} = "limit 1"
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.435, + 0.884, + 0.479 + ], + "angle": 0, + "content": "Table F.1: Example representations for a wrong SQL query that contains a nested subquery and its edit actions (including Insert edits). The corresponding natural language utterance is \"What is the number of cars with a greater acceleration than the one with the most horsepower?\"" + }, + { + "type": "table", + "bbox": [ + 0.118, + 0.58, + 0.884, + 0.83 + ], + "angle": 0, + "content": "
Query RepresentationEdit Representation
SQLselect employee.name from employee join evaluation on employee.employee_id = evaluation.employee_id group by evaluation.employee_id" order by sum(evaluationbonus) desc limit 1Token-level<Delete> group by evaluation.employee_id <DeleteEnd> <DeleteSum( <DeleteEnd><Delete>) <DeleteEnd>
Clause-level<Delete> group by evaluation.employee_id <DeleteEnd> <ReplaceOld> order by sum(evaluation;bONUS) desc <ReplaceNew> order by evaluation;bONUS desc <ReplaceEnd>
PyDictsql = { "select": "select employee.name", "from": "from employee join evaluation on employee.employee_id = evaluation.employee_id", "groupBy": "group by evaluation.employee_id", "orderBy": "order by sum(evaluation;bONUS) desc", "limit": "limit 1" }Clause-level<Delete> "groupId": "group by evaluation.employee_id" <DeleteEnd><ReplaceOld> "orderBy": "order by sum(evaluation;bONUS) desc" <ReplaceNew> "orderBy": "order by evaluation;bONUS desc" <ReplaceEnd>
Programsql.pop("groupId") sql["orderBy"] = "order by evaluation;bONUS desc"
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.839, + 0.883, + 0.869 + ], + "angle": 0, + "content": "Table F.2: Example representations for a wrong SQL query and its edit actions (including Delete edits). The corresponding natural language utterance is \"Find the name of the employee who got the highest one time bonus.\"" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.928, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1370" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.133, + 0.085, + 0.435, + 0.1 + ], + "angle": 0, + "content": "ACL 2023 Responsible NLP Checklist" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.108, + 0.323, + 0.123 + ], + "angle": 0, + "content": "A For every submission:" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.127, + 0.533, + 0.158 + ], + "angle": 0, + "content": "A1. Did you describe the limitations of your work? 6" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.171, + 0.553, + 0.202 + ], + "angle": 0, + "content": "A2. Did you discuss any potential risks of your work? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.212, + 0.697, + 0.244 + ], + "angle": 0, + "content": "A3. Do the abstract and introduction summarize the paper's main claims?" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.256, + 0.669, + 0.288 + ], + "angle": 0, + "content": "A4. Have you used AI writing assistants when working on this paper? Left blank." + }, + { + "type": "list", + "bbox": [ + 0.129, + 0.127, + 0.697, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.299, + 0.49, + 0.317 + ], + "angle": 0, + "content": "B Did you use or create scientific artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.322, + 0.147, + 0.334 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.347, + 0.53, + 0.378 + ], + "angle": 0, + "content": "B1. Did you cite the creators of artifacts you used? 3" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.391, + 0.779, + 0.423 + ], + "angle": 0, + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.434, + 0.881, + 0.513 + ], + "angle": 0, + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.525, + 0.881, + 0.588 + ], + "angle": 0, + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.599, + 0.881, + 0.645 + ], + "angle": 0, + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? 3" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.658, + 0.881, + 0.753 + ], + "angle": 0, + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. 3" + }, + { + "type": "list", + "bbox": [ + 0.129, + 0.347, + 0.881, + 0.753 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.765, + 0.494, + 0.782 + ], + "angle": 0, + "content": "C Did you run computational experiments?" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.789, + 0.147, + 0.8 + ], + "angle": 0, + "content": "4" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.812, + 0.881, + 0.861 + ], + "angle": 0, + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Appendix B" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.869, + 0.878, + 0.893 + ], + "angle": 0, + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.519, + 0.941 + ], + "angle": 0, + "content": "1371" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.13, + 0.084, + 0.88, + 0.133 + ], + "angle": 0, + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Appendix B" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.143, + 0.88, + 0.205 + ], + "angle": 0, + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.218, + 0.88, + 0.283 + ], + "angle": 0, + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Appendix B" + }, + { + "type": "list", + "bbox": [ + 0.13, + 0.084, + 0.88, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.294, + 0.878, + 0.331 + ], + "angle": 0, + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.341, + 0.881, + 0.388 + ], + "angle": 0, + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.4, + 0.881, + 0.464 + ], + "angle": 0, + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.476, + 0.881, + 0.54 + ], + "angle": 0, + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.55, + 0.873, + 0.582 + ], + "angle": 0, + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.593, + 0.88, + 0.641 + ], + "angle": 0, + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response." + }, + { + "type": "list", + "bbox": [ + 0.128, + 0.341, + 0.881, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.928, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1372" + } + ] +] \ No newline at end of file diff --git a/2023/Text-to-SQL Error Correction with Language Models of Code/63d877fb-bf81-4e0e-9af4-85d702a6c368_origin.pdf b/2023/Text-to-SQL Error Correction with Language Models of Code/63d877fb-bf81-4e0e-9af4-85d702a6c368_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3684fa68f1608d87ee0be28072054325d4eb9932 --- /dev/null +++ b/2023/Text-to-SQL Error Correction with Language Models of Code/63d877fb-bf81-4e0e-9af4-85d702a6c368_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:91a1170817eaf4ffef48b8085451d2347e9aa386ca336756c8e106392bebb9f9 +size 637318 diff --git a/2023/Text-to-SQL Error Correction with Language Models of Code/full.md b/2023/Text-to-SQL Error Correction with Language Models of Code/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8a47743b656dbb2b3e97b18c9ace83fdbdb1698d --- /dev/null +++ b/2023/Text-to-SQL Error Correction with Language Models of Code/full.md @@ -0,0 +1,322 @@ +# Text-to-SQL Error Correction with Language Models of Code + +Ziru Chen $^{1}$ , Shijie Chen $^{1}$ , Michael White $^{1}$ , Raymond Mooney $^{2}$ , Ali Payani $^{3}$ , Jayanth Srinivasa $^{3}$ , Yu Su $^{1}$ , Huan Sun $^{1}$ + +1The Ohio State University + +$^{2}$ The University of Texas at Austin $^{3}$ Cisco Research + +{chen.8336, chen.10216, white.1240, su.809, sun.397}@osu.edu + +mooney@cs.utexas.edu {apayani, jasriniv}@cisco + +# Abstract + +Despite recent progress in text-to-SQL parsing, current semantic parsers are still not accurate enough for practical use. In this paper, we investigate how to build automatic text-to-SQL error correction models. Noticing that token-level edits are out of context and sometimes ambiguous, we propose building clause-level edit models instead. Besides, while most language models of code are not specifically pre-trained for SQL, they know common data structures and their operations in programming languages such as Python. Thus, we propose a novel representation for SQL queries and their edits that adheres more closely to the pre-training corpora of language models of code. Our error correction model improves the exact set match accuracy of different parsers by 2.4-6.5 and obtains up to 4.3 point absolute improvement over two strong baselines. $^1$ + +# 1 Introduction + +Text-to-SQL parsing is a classic semantic parsing task that finds wide applications (Zelle and Mooney, 1996; Tang and Mooney, 2000). Since the release of Spider (Yu et al., 2018), a cross-database text-to-SQL benchmark, many semantic parsers with decent performance have been developed (Lin et al., 2020; Wang et al., 2020; Deng et al., 2021; Rubin and Berant, 2021; Scholak et al., 2021). Nonetheless, state-of-the-art semantic parsers are still not accurate enough. As a result, their users need to constantly correct wrongly predicted SQL queries, which can be as time-consuming and error-prone as writing a SQL query from scratch (Jorgensen and Shepperd, 2007; Weiss et al., 2007). Therefore, in this paper, we study the problem of automatic text-to-SQL error correction to better assist users in querying complex databases. + +We first highlight that it is essential to factor in the compositional substructures within SQL + +queries, such as abstract syntax trees (Yin and Neubig, 2017; Guo et al., 2022) and data-flow graphs (Guo et al., 2021), instead of treating code snippets as string sequences. Compared to individual tokens, substructures (e.g. SQL clauses) include more context of the entire program and are more semantically meaningful. Consequently, edit patterns of such substructures are more intuitive for humans to understand and easier for language models to learn. Moreover, while the pre-training corpora for language models of code, such as CodeT5 (Wang et al., 2021), do not include many SQL queries based on their documentation, they naturally contain abundant examples of common data structures like dictionaries. Therefore, we hypothesize that transforming unfamiliar SQL queries into familiar data structures can help language models of code better perform structural editing of SQL queries. + +Based on these observations, we develop our error correction model and make two contributions. First, we propose considering SQL clauses instead of tokens as basic semantic units for editing. Using a context-free grammar, we can decompose a SQL query and identify its clauses by traversing its abstract syntax tree. Second, we propose a new representation of SQL queries and their edits that adheres more closely to common code pre-training corpora, including CodeSearchNet (Husain et al., 2020), and makes the structures of a SQL query more explicit. With a decomposed SQL query, we pair each clause with its SQL keyword and represent the entire query as a Python dictionary. Then, we format edits on a wrong SQL query as a program that modifies data of the query's corresponding dictionary. Unlike token-level edits in existing work (Zhang et al., 2023), such dictionary operations define all edits unambiguously and can be directly executed with a Python interpreter. + +Through comprehensive experiments with different representations, we show that: (1) our proposed representation has the lowest zero-shot perplexity + +
Query RepresentationEdit Representation
SQLselect tweets.text from tweets order by tweets.textToken-Level<ReplaceOld> tweets.text <ReplaceNew> tweets.createDate <ReplaceEnd>
Clause-Level<ReplaceOld> order by tweets.text <ReplaceNew> order by tweets.createDate <ReplaceEnd>
PyDictsql = { "select": "select tweets.text", "from": "from tweets", "orderBy": "order by tweets.text" }Clause-Level<ReplaceOld> "orderBy": "order by tweets.text" <ReplaceNew> "orderBy": "order by tweets.createDate" <ReplaceEnd> sql["orderBy"] = "order by tweets.createDate"
Program
+ +Table 1: Example representations for a wrong SQL query and the Replace edit action. The corresponding natural language utterance is "List the text of all tweets in the order of date." For token-level and clause-level representations, we format them as " Span of wrong tokens/clauses Span of correct tokens/clauses ", where , , and are special tokens. + +with CodeT5; (2) simply changing token-level edits to clause-level edits can effectively improve the performance of our models; and (3) our method improves the exact set match accuracy of different parsers by 2.4-6.5 and obtains up to 4.3 point absolute improvement over two strong baselines. + +# 2 Text-to-SQL Error Correction + +Given a natural language utterance $\mathbf{u}$ , a database schema $\mathbf{s}$ , and a wrong SQL query $\mathbf{q}_{-}$ produced by an existing parser, our goal is to develop an error correction model that predicts a sequence of edit actions $\mathbf{e}$ and the correct query $\mathbf{q}_{+}$ . Following previous work (Zhang et al., 2023), we formulate our task as sequence-to-sequence generation: + +$$ +P (\mathbf {y} | \mathbf {x}) = \Pi_ {t = 1} ^ {T} P (\mathbf {y} _ {t} | \mathbf {x}, \mathbf {y} _ {1: t - 1}) \tag {1} +$$ + +where $\mathbf{x} = [\mathbf{u};\mathbf{s};\mathbf{q}_{-}]$ is the concatenation of the given inputs and $\mathbf{y} = [\mathbf{e};\mathbf{q}_{+}]$ is the concatenation of all edit actions and the resulting correct query. In this section, we study different representations of SQL queries (Section 2.1) and edits (Section 2.2) to better leverage language models of code. + +# 2.1 Query Representation + +We consider two representations for a predicted query: (1) the original SQL format and (2) our proposed PyDict (Python Dictionary) representation. To prepare for editing, we disambiguate each SQL query following Rubin and Berant (2021), including lower-casing non-value tokens, resolving table references, and formatting punctuation. This preprocessing normalizes SQL queries predicted by different base parsers and the gold annotations into the same format. To build our PyDict representation, we parse a SQL query into its abstract syntax tree (AST) with Spider's context-free grammar. We + +use depth-first search to traverse through the AST, find any nested substructures, and construct the dictionary representation bottom-up. Table 1 shows the "SQL" and "PyDict" representations of a SQL query (more details in Appendix A). + +# 2.2 Edit Representation + +We first follow Zhang et al. (2023) to use token-level edit representation with special tokens (Table 1), which have unique entries in the tokenizer and the model's embedding layer to describe Replace, Insert, and Delete edit actions (more examples in Appendix F). However, we realize this representation can sometimes be ambiguous. As shown in Table 1, the span "tweets.text" appears twice in the SQL query. This repetition would confuse the error correction model with which span to replace when generating the corrected query. Also, the ambiguity makes it difficult to implement rules and directly carry out the edit actions on the wrong query. Hence, we change the token-level edit representation to clause-level, which includes more context of the query to make different edits more distinguishable. In our experiments (Section 4.1), we demonstrate that this simple modification is already effective. Our program representation further improves the performance because it is more similar to the code pre-training corpora and eliminates the need to learn special tokens' representations. + +# 3 Experimental Setup + +# 3.1 Data Synthesis for SQL Error Correction + +To train a text-to-SQL error correction model, we need to collect a set of wrong SQL parses that reflects a realistic distribution of errors (Section 4.2) as our training data. We synthesize this dataset by + +
CodeT5BRIDGEv2SmBoP
# of Train47,02024,77620,083
# of Dev448448448
# of Test430392310
Avg. Train Edits2.343.112.72
Avg. Dev Edits2.703.293.31
Avg. Test Edits1.841.511.47
+ +Table 2: Summary of data statistics. + +performing 5-fold cross-validation on each parser, which approximates the actual evaluation setting. + +Following the evaluation setup in Yu et al. (2018), we split Spider's training set into five roughly equal subsets by different databases. For each cross-validation fold, we train a text-to-SQL parser (Section 3.2) on four subsets and evaluate it on the remaining one. At inference time, we perform beam search with size 20 for each example and collect grammatical and executable parses in the beam. If a SQL parse is not an exact set match or execution match to the gold annotation, we label it wrong and include it in our training set for error correction. Having synthesized our training dataset, we randomly sample 8 databases and their associated questions to construct a held-out development set. For development set examples, we only keep incorrect SQL parses with the highest beam confidence. For our error correction test set, we train each parser on the full Spider training set and evaluate it on the original Spider's development set without modifications. We similarly keep SQL parses with exact match or execution match errors. Table 2 summarizes the statistics of our data. + +# 3.2 Models + +Text-to-SQL base parsers. We choose three text-to-SQL parsers with different decoding strategies and levels of performance (Table 3). We elaborate on our selection criteria in Appendix B. + +- CodeT5 (Wang et al., 2021): We fine-tune CodeT5-base following Xie et al. (2022). This parser represents those using beam search decoding and having a lower accuracy. +BRIDGEv2 (Lin et al., 2020): A representative parser with constrained decoding and achieving a medium-level accuracy. +- SmBoP (Rubin and Berant, 2021): A representative parser with bottom-up decoding and achieving higher accuracy. + +Error correction models. We use two language models of code in all our experiments: + +CoditT5 (Zhang et al., 2023): A language model pre-trained for code editing tasks by injecting noises to code snippets in CodeSearchNet (Husain et al., 2020) and then denoising with token-level edit representations. +- CodeT5 (Wang et al., 2021): A language model pre-trained for general code understanding and generation with four different pre-training objectives. + +We compare the existing SQL+Token-Level representation with our proposed ones: SQL+Clause-Level, PyDict+Clause-Level, and PyDict+Program on CodeT5 and the first three on CoditT5.3 Implementation details are in Appendix C. + +# 3.3 Evaluation + +We use the increase in Exact Set Match (EM) and Execution Match (EX) accuracy on our error correction test set to measure each model's performance. Because CoditT5's experiments assume the input program has at least one error, we keep this assumption for fair comparisons. To construct a test set satisfying this assumption, we have to compare parser-generated SQL queries with gold annotations (Section 3.1). Thus, we use the Spider development set as our test set and split the Spider training set to build a held-out development set (Table 2) to select model checkpoints during training. We also include results on our held-out development set in the appendix (Table E.1). + +# 4 Results and Analysis + +# 4.1 Main Results + +We summarize our main results in this section. To ensure robustness, we repeat all experiments with 3 different random seeds and report the average performances with standard deviations. Our model can also be used in an interactive framework that allows users to select edit actions from the top- $k$ beam candidates. We include more experiments with simulated user interactions in Appendix E. + +Our representation's perplexity is the smallest. We validate that our PyDict+Program representation adheres more closely to the code pre-training corpora by measuring its zero-shot perplexity on CodeT5 using our development set (Section 3.1). + +
ModelsQueryEditCodeT5BRIDGEv2SmBoP
EMEXEMEXEMEX
No EditN/AN/A62.7 (-)63.6 (-)70.1 (-)68.2 (-)74.6 (-)75.3 (-)
CodiT5SQLToken-Level64.3 (0.1)64.4 (0.2)65.4 (0.5)66.6 (0.3)74.2 (0.4)75.3 (0.1)
SQLClause-Level67.0 (0.4)65.4 (0.5)71.3 (0.5)70.9 (0.2)76.3 (0.0)77.2 (0.3)
PyDictClause-Level67.1 (0.2)66.5 (0.4)70.6 (0.8)70.8 (0.6)76.3 (0.3)77.0 (0.3)
CodeT5SQLToken-Level66.7 (0.9)65.9 (0.5)68.2 (0.4)69.4 (0.8)75.6 (0.4)76.5 (0.6)
SQLClause-Level68.3 (0.3)\( \underline{68.2}^{+}(0.6) \)71.8+(0.4)72.5+(0.2)76.7 (0.6)77.4 (0.3)
PyDictClause-Level66.6 (0.8)67.1 (0.8)72.0+(0.3)72.4+(0.2)77.3 (0.6)77.8 (0.2)
\( CodeT5^* \)CodeT5PyDictProgram\( 69.2^{+}(0.4) \)\( 68.4^{+}(0.2) \)\( 72.5^{+}(0.4) \)\( 73.1^{+}(0.2) \)77.3 (0.4)77.6 (0.6)
\( 69.0^{+}(0.2) \)\( 68.2^{+}(0.1) \)\( 72.5^{+}(0.3) \)\( 73.0^{+}(0.6) \)\( 78.0^{+}(0.3) \)\( 78.5^{+}(0.3) \)
+ +Table 3: Exact Set Match (EM) and Execution Match (EX) accuracy on Spider development set. The best performances are in bold and the second bests are underlined. Results with $^+$ are statistically significant (McNemar's; $p < 0.05$ ) compared to CodeT5-SQL+Token-Level (Appendix D). Otherwise, the results are not statistically significant. \*We fine-tune the model to generate edit programs only (without resulting queries) and use Python interpreter to execute the edit actions. + +![](images/a38959892ed71b38c996b10dff3a75e11f3afbcab1432d33e61060f6c2917496.jpg) +Figure 1: CodeT5's zero-shot perplexity (in log scale) of all four representations on our synthesized SQL error development set. + +As shown in Figure 1, by representing data in Py-Dict, we can reduce the perplexity of CodeT5 by 2 orders of magnitude. After augmenting it with our program representation, we further reduce the zero-shot perplexity of CodeT5 to only $5.96 \times 10^{2}$ , 3 orders of magnitude less than the SQL+Token-Level representation $(1.26 \times 10^{5})$ . + +Clause-level editing is more effective, especially when represented in PyDict+Program. Since CodeT5 consistently outperforms CoditT5 with the same representations, we focus on comparisons among CodeT5 variations. As shown in Table 3, compared to CodeT5-SQL+Token-Level, only CodeT5-PyDict+Program achieves statistically significant improvement on all three parsers, while clause-level models fail McNemar's significance test for some parsers. More concretely, it achieves up to 4.3 point more absolute improvement on EM accuracy (68.2 → 72.5; BRIDGEv2) and 3.7 point more absolute improvement on EX accuracy (69.4 → 73.1; BRIDGEv2). Overall, CodeT5-PyDict+Program can boost the parsers' EM accu + +racy by 2.4-6.5. Thus, both clause-level editing and PyDict+Program representation can better take advantage of language models of code. + +# 4.2 Error Analysis + +Additionally, we conduct an error analysis (Table 4) by sampling 100 wrong parses from all three parsers and classifying them into five categories: + +- Database Grounding: A generated SQL query has the correct structure, but some table/column names or entity values are wrong. +- Incorrect Structure: A generated SQL query has missing, wrong, or redundant structures. +- Syntax & Grammar: A generated SQL query violates the programming language's syntax. +- False Negative: A generated SQL query is semantically correct but not captured by evaluation metrics, or the gold annotation is wrong. +- Other: All other errors, such as wrong aggregation functions, besides the above categories. + +Since the error distributions for each parser are similar, as an example, we discuss our findings based on the strongest parser, SmBoP: + +Database grounding is the major type of error. Among the 100 samples from SmBoP, we find that 54 of them have database grounding errors. Particularly, SmBoP predicts wrong table/column names in 34 parses, inaccurate entity values in 9 parses, and incorrect JOIN relations in 11 parses. Our CodeT5-PyDict+Program model can successfully fix 16 of the 54 erroneous parses, including 10 parses with wrong table/column names, 4 parses with inaccurate entity values, and 2 parses with incorrect JOIN relations. We hypothesize that + +
Error CategoryCodeT5BRIDGEv2SmBoP
ResolvedUnresolvedAllResolvedUnresolvedAllResolvedUnresolvedAll
Database Grounding155166144862163854
Incorrect Structure215172121432326
Syntax & Grammar000000145
False Negative099066088
Other17821618167
+ +Table 4: Analysis of 100 sample errors made by each text-to-SQL parser. We group the errors into 5 categories and examine if our CodeT5-PyDict+Program model resolves them. + +database grounding is also a major category of errors in our synthesized training set, so our model has learned to resolve similar errors. Nevertheless, it still cannot correct the remaining 38 SQL parses. We notice that our current representation for database schema is missing critical information, such as column data types and foreign key relations, for our error correction model to fix database grounding errors. Following our PyDict representation for SQL, we suggest designing a code representation for database schema that includes such information to tackle this issue in future work. + +Structural errors are hard to edit automatically. Besides database grounding, 26 of SmBoP's errors belong to another category, incorrect structure. These 26 samples contain 7 parses with incorrect SQL clauses and 19 parses with incorrect subqueries, but our CodeT5-PyDict+Program model only resolves 1 and 2 of them, respectively. We find that correcting such errors usually involves multiple edit steps, which motivates us to incorporate our model into an interactive framework in future work. As our experiments with simulated user interaction (Appendix E.2) show, when our model interacts with the simulated user to correct one clause at a time, it is able to fully correct more SQL parses. Thus, we deem interactive correction would maximize our model's utility in practice. + +# 5 Related Work + +Since the release of CodeBERT (Feng et al., 2020), many language models of code have emerged for program understanding and generation (Ahmad et al., 2021; Chen et al., 2021; Guo et al., 2021; Wang et al., 2021; Guo et al., 2022; Fried et al., 2023; Nijkamp et al., 2023). In addition to program-related tasks, recent work shows they also excel at processing natural language structures. Using code as meaning representations (MRs), we can leverage language models of code in various tasks, such as commonsense reasoning (Madaan et al., 2022), + +action planning (Singh et al., 2022), and event extraction (Wang et al., 2022). In fact, how to design MRs to reduce model learning difficulty is a salient research question in semantic parsing (Guo et al., 2019; Gan et al., 2021b; Nie et al., 2022). + +Our work demonstrates that program-related tasks themselves can also benefit from code-based MRs. Specifically, we apply such MRs to SQL error correction, a variant of automatic program repair tasks (Tufano et al., 2019; Panthaplackel et al., 2022; Zhang et al., 2023). Although SQL is a code-based MR, it is much harder for models to learn compared to other MRs, such as FunQL and lambda calculus (Li et al., 2022). Consequently, without many SQL queries in their pre-training corpora, language models of code can underperform state-of-the-art text-to-SQL parsers. By converting SQL queries into Python dictionaries, we can explicitly represent their compositional substructures and define edit actions as programs, which reduces the learning difficulty for language models of code and yields better performance. + +# 6 Conclusion and Future Work + +This paper presents a study on developing a text-to-SQL error correction model with clause-level edits and different representations. Our comprehensive experiments demonstrate that clauses are better semantic units than tokens for editing SQL queries and mimicking patterns in code pre-training corpora helps better leverage language models of code. As a future direction, we plan to incorporate our model into interactive semantic parsing frameworks (Li et al., 2020; Yao et al., 2019, 2020; Zeng et al., 2020) by suggesting possible edits to users once a wrong parse is identified. In this way, users would more efficiently correct parse errors and get better assistance. We also plan to experiment with other language models of code (Fried et al., 2023; Nijkamp et al., 2023) and text-to-SQL datasets (Zelle and Mooney, 1996; Gan et al., 2021a) to verify the generalizability of our method. + +# Limitations + +Actual applications of our model. Our work assumes that input SQL queries to our model are always wrong. This assumption is more feasible in an interactive semantic parsing framework, where the users are expected to decide whether a SQL parse, accompanied by its natural language explanations (Elgohary et al., 2020, 2021; Narechinaia et al., 2021; Mo et al., 2022), has errors or not. Alternatively, to remove this assumption, it would be interesting for future work to study the performance of our error correction model in combination with an automatic error detection model (Chen et al., 2023). + +# Experiments with more language models of code. + +We have only experimented with two language models of code, CodiT5 and CodeT5, both using T5-base (Raffel et al., 2020) as their underlying model architecture. It would be interesting to test how our conclusions generalize to other language models of code in the future. Based on the strong capabilities of large language models of code, such as Codex (Chen et al., 2021), InCoder (Fried et al., 2023), and CodeGen (Nijkamp et al., 2023), we believe that these models can better exploit their knowledge about data structures and their operations in Python. These models may perform even better on Text-to-SQL error correction with our proposed representations. + +# Acknowledgements + +We would like to thank the anonymous reviewers and colleagues from the OSU NLP group for their thoughtful comments. This research was supported in part by a sponsored award from Cisco Research, NSF IIS-1815674, NSF CAREER #1942980, NSF OAC-2112606, and Ohio Supercomputer Center (Center, 1987). The views and conclusions contained herein are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the U.S. government. The U.S. Government is authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notice herein. Ziru is also supported by The Ohio State University Graduate School through University Fellowship. + +# References + +Wasi Ahmad, Saikat Chakraborty, Baishakhi Ray, and Kai-Wei Chang. 2021. Unified pre-training for pro + +gram understanding and generation. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2655-2668, Online. Association for Computational Linguistics. +Ohio Supercomputer Center. 1987. Ohio supercomputer center. +Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidi Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott Gray, Nick Ryder, Mikhail Pavlov, Alethea Power, Lukasz Kaiser, Mohammad Bavarian, Clemens Winter, Philippe Tillet, Felipe Petroski Such, Dave Cummings, Matthias Plappert, Fotios Chantzis, Elizabeth Barnes, Ariel Herbert-Voss, William Hebgen Guss, Alex Nichol, Alex Paino, Nikolas Tezak, Jie Tang, Igor Babuschkin, Suchir Balaji, Shantanu Jain, William Saunders, Christopher Hesse, Andrew N. Carr, Jan Leike, Josh Achiam, Vedant Misra, Evan Morikawa, Alec Radford, Matthew Knight, Miles Brundage, Mira Murati, Katie Mayer, Peter Welinder, Bob McGrew, Dario Amodei, Sam McCandlish, Ilya Sutskever, and Wojciech Zaremba. 2021. Evaluating large language models trained on code. +Shijie Chen, Ziru Chen, Huan Sun, and Yu Su. 2023. Error detection for text-to-sql semantic parsing. +Xiang Deng, Ahmed Hassan Awadallah, Christopher Meek, Oleksandr Polozov, Huan Sun, and Matthew Richardson. 2021. Structure-grounded pretraining for text-to-SQL. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1337-1350, Online. Association for Computational Linguistics. +Ahmed Elgohary, Saghar Hosseini, and Ahmed Hassan Awadallah. 2020. Speak to your parser: Interactive text-to-SQL with natural language feedback. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2065-2077, Online. Association for Computational Linguistics. +Ahmed Elgohary, Christopher Meek, Matthew Richardson, Adam Fourney, Gonzalo Ramos, and Ahmed Hassan Awadallah. 2021. NL-EDIT: Correcting semantic parse errors through natural language interaction. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5599–5610, Online. Association for Computational Linguistics. +Zhangyin Feng, Daya Guo, Duyu Tang, Nan Duan, Xiaocheng Feng, Ming Gong, Linjun Shou, Bing Qin, Ting Liu, Daxin Jiang, and Ming Zhou. 2020. CodeBERT: A pre-trained model for programming and natural languages. In Findings of the Association + +for Computational Linguistics: EMNLP 2020, pages 1536-1547, Online. Association for Computational Linguistics. +Daniel Fried, Armen Aghajanyan, Jessy Lin, Sida Wang, Eric Wallace, Freda Shi, Ruiqi Zhong, Scott Yih, Luke Zettlemoyer, and Mike Lewis. 2023. Incoder: A generative model for code infilling and synthesis. In The Eleventh International Conference on Learning Representations. +Yujiang Gan, Xinyun Chen, Qiuping Huang, Matthew Purver, John R. Woodward, Jinxia Xie, and Pengsheng Huang. 2021a. Towards robustness of text-to-SQL models against synonym substitution. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 2505-2515, Online. Association for Computational Linguistics. +Yujuan Gan, Xinyun Chen, Jinxia Xie, Matthew Purver, John R. Woodward, John Drake, and Qiaofu Zhang. 2021b. Natural SQL: Making SQL easier to infer from natural language specifications. In *Findings of the Association for Computational Linguistics: EMNLP* 2021, pages 2030–2042, Punta Cana, Dominican Republic. Association for Computational Linguistics. +Daya Guo, Shuai Lu, Nan Duan, Yanlin Wang, Ming Zhou, and Jian Yin. 2022. UniXcoder: Unified cross-modal pre-training for code representation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7212-7225, Dublin, Ireland. Association for Computational Linguistics. +Daya Guo, Shuo Ren, Shuai Lu, Zhangyin Feng, Duyu Tang, Shujie Liu, Long Zhou, Nan Duan, Alexey Svyatkovskiy, Shengyu Fu, Michele Tufano, Shao Kun Deng, Colin Clement, Dawn Drain, Neel Sundaresan, Jian Yin, Daxin Jiang, and Ming Zhou. 2021. GraphCodeBERT: Pre-training code representations with data flow. In International Conference on Learning Representations. +Jiaqi Guo, Zecheng Zhan, Yan Gao, Yan Xiao, JianGuang Lou, Ting Liu, and Dongmei Zhang. 2019. Towards complex text-to-SQL in cross-domain database with intermediate representation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4524-4535, Florence, Italy. Association for Computational Linguistics. +Hamel Husain, Ho-Hsiang Wu, Tiferet Gazit, Miltiadis Allamanis, and Marc Brockschmidt. 2020. Code-searchnet challenge: Evaluating the state of semantic code search. +Magne Jorgensen and Martin Shepperd. 2007. A systematic review of software development cost estimation studies. IEEE Transactions on Software Engineering, 33(1):33-53. + +Yuntao Li, Bei Chen, Qian Liu, Yan Gao, Jian-Guang Lou, Yan Zhang, and Dongmei Zhang. 2020. "what do you mean by that?" a parser-independent interactive approach for enhancing text-to-SQL. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6913-6922, Online. Association for Computational Linguistics. +Zhenwen Li, Jiaqi Guo, Qian Liu, Jian-Guang Lou, and Tao Xie. 2022. Exploring the secrets behind the learning difficulty of meaning representations for semantic parsing. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 3616-3625, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics. +Xi Victoria Lin, Richard Socher, and Caiming Xiong. 2020. Bridging textual and tabular data for cross-domain text-to-SQL semantic parsing. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 4870-4888, Online. Association for Computational Linguistics. +Aman Madaan, Shuyan Zhou, Uri Alon, Yiming Yang, and Graham Neubig. 2022. Language models of code are few-shot commonsense learners. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 1384-1403, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics. +Quinn McNemar. 1947. Note on the sampling error of the difference between correlated proportions or percentages. In Psychometrika, volume 12, page 153-157. +Lingbo Mo, Ashley Lewis, Huan Sun, and Michael White. 2022. Towards transparent interactive semantic parsing via step-by-step correction. In *Findings of the Association for Computational Linguistics: ACL* 2022, pages 322-342, Dublin, Ireland. Association for Computational Linguistics. +Marius Mosbach, Maksym Andriushchenko, and Dietrich Klakow. 2021. On the stability of fine-tuning BERT: Misconceptions, explanations, and strong baselines. In International Conference on Learning Representations. +Arpit Narechania, Adam Fourney, Bongshin Lee, and Gonzalo Ramos. 2021. Diy: Assessing the correctness of natural language to sql systems. In 26th International Conference on Intelligent User Interfaces, IUI '21, page 597-607, New York, NY, USA. Association for Computing Machinery. +Lunyiu Nie, Shulin Cao, Jiaxin Shi, Jiuding Sun, Qi Tian, Lei Hou, Juanzi Li, and Jidong Zhai. 2022. GraphQ IR: Unifying the semantic parsing of graph query languages with one intermediate representation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 5848-5865, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics. + +Erik Nijkamp, Bo Pang, Hiroaki Hayashi, Lifu Tu, Huan Wang, Yingbo Zhou, Silvio Savarese, and Caiming Xiong. 2023. Codegen: An open large language model for code with multi-turn program synthesis. In The Eleventh International Conference on Learning Representations. +Sheena Panthaplackel, Milos Gligoric, Junyi Jessy Li, and Raymond Mooney. 2022. Using developer discussions to guide fixing bugs in software. In *Findings of the Association for Computational Linguistics: EMNLP* 2022, pages 2292-2301, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics. +Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. 2019. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc. +Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(140):1-67. +Ohad Rubin and Jonathan Berant. 2021. SmBoP: Semi-autoregressive bottom-up semantic parsing. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 311-324, Online. Association for Computational Linguistics. +Torsten Scholak, Nathan Schucher, and Dzmitry Bahdanau. 2021. PICARD: Parsing incrementally for constrained auto-regressive decoding from language models. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 9895-9901, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics. +Noam Shazeer and Mitchell Stern. 2018. Adafactor: Adaptive learning rates with sublinear memory cost. In Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pages 4596-4604. PMLR. +Ishika Singh, Valts Blukis, Arsalan Mousavian, Ankit Goyal, Danfei Xu, Jonathan Tremblay, Dieter Fox, Jesse Thomason, and Animesh Garg. 2022. Progress: Generating situated robot task plans using large language models. In Workshop on Language and Robotics at CoRL 2022. + +Lappoon R. Tang and Raymond J. Mooney. 2000. Automated construction of database interfaces: Integrating statistical and relational learning for semantic parsing. In Proceedings of the 2000 Joint SIGDAT Conference on Empirical Methods in Natural Language Processing and Very Large Corpora: Held in Conjunction with the 38th Annual Meeting of the Association for Computational Linguistics - Volume 13, EMNLP '00, page 133-141, USA. Association for Computational Linguistics. +Michele Tufano, Jevgenija Pantiuchina, Cody Watson, Gabriele Bavota, and Denys Poshyvanyk. 2019. On learning meaningful code changes via neural machine translation. In Proceedings of the 41st International Conference on Software Engineering, ICSE '19, page 25-36. IEEE Press. +Bailin Wang, Richard Shin, Xiaodong Liu, Oleksandr Polozov, and Matthew Richardson. 2020. RAT-SQL: Relation-aware schema encoding and linking for text-to-SQL parsers. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7567-7578, Online. Association for Computational Linguistics. +Xingyao Wang, Sha Li, and Heng Ji. 2022. Code4struct: Code generation for few-shot structured prediction from natural language. +Yue Wang, Weishi Wang, Shafiq Joty, and Steven C.H. Hoi. 2021. CodeT5: Identifier-aware unified pretrained encoder-decoder models for code understanding and generation. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 8696-8708, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics. +Cathrin Weiss, Rahul Premraj, Thomas Zimmermann, and Andreas Zeller. 2007. How long will it take to fix this bug? In Fourth International Workshop on Mining Software Repositories (MSR'07:ICSE Workshops 2007), pages 1-1. +Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online. Association for Computational Linguistics. +Tianbao Xie, Chen Henry Wu, Peng Shi, Ruiqi Zhong, Torsten Scholak, Michihiro Yasunaga, Chien-Sheng Wu, Ming Zhong, Pengcheng Yin, Sida I. Wang, Victor Zhong, Bailin Wang, Chengzu Li, Connor Boyle, Ansong Ni, Ziyu Yao, Dragomir Radev, Caiming Xiong, Lingpeng Kong, Rui Zhang, Noah A. Smith, Luke Zettlemoyer, and Tao Yu. 2022. UnifiedSKG: + +Unifying and multi-tasking structured knowledge grounding with text-to-text language models. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 602-631, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics. + +Ziyu Yao, Yu Su, Huan Sun, and Wen-tau Yih. 2019. Model-based interactive semantic parsing: A unified framework and a text-to-SQL case study. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5447-5458, Hong Kong, China. Association for Computational Linguistics. + +Ziyu Yao, Yiqi Tang, Wen-tau Yih, Huan Sun, and Yu Su. 2020. An imitation game for learning semantic parsers from user interaction. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6883-6902, Online. Association for Computational Linguistics. + +Pengcheng Yin and Graham Neubig. 2017. A syntactic neural model for general-purpose code generation. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 440-450, Vancouver, Canada. Association for Computational Linguistics. + +Tao Yu, Rui Zhang, Kai Yang, Michihiro Yasunaga, Dongxu Wang, Zifan Li, James Ma, Irene Li, Qingning Yao, Shanelle Roman, Zilin Zhang, and Dragomir Radev. 2018. Spider: A large-scale human-labeled dataset for complex and cross-domain semantic parsing and text-to-SQL task. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 3911-3921, Brussels, Belgium. Association for Computational Linguistics. + +John M. Zelle and Raymond J. Mooney. 1996. Learning to parse database queries using inductive logic programming. In Proceedings of the Thirteenth National Conference on Artificial Intelligence - Volume 2, AAAI'96, page 1050-1055. AAAI Press. + +Jichuan Zeng, Xi Victoria Lin, Steven C.H. Hoi, Richard Socher, Caiming Xiong, Michael Lyu, and Irwin King. 2020. Photon: A robust cross-domain text-to-SQL system. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 204-214, Online. Association for Computational Linguistics. + +Jiyang Zhang, Sheena Panthapackel, Pengyu Nie, Junyi Jessy Li, and Milos Gligoric. 2023. Coditt5: Pretraining for source code and natural language editing. In Proceedings of the 37th IEEE/ACM International Conference on Automated Software Engineering, ASE '22, New York, NY, USA. Association for Computing Machinery. + +# Appendices + +We provide more details omitted in the main text as follows: + +- Appendix A: SQL PyDict Representation +- Appendix B: Text-to-SQL Parser Selection +- Appendix C: Implementation Details +- Appendix D: Statistical Significance Test +- Appendix E: Additional Results +- Appendix F: More Representation Examples + +# A SQL PyDict Representation + +We implement the transformation from any SQL query to our PyDict representation in three steps (Section 2.1). First, we use context-free grammar to parse a SQL query and obtain its abstract syntax tree (AST). The AST naturally contains a SQL decomposition where each clause has its unique subtree. In addition, if a clause contains a nested query, it would be represented as another independent subtree, which is a child of the root node in the clause's AST subtree. With these substructures explicitly represented, we use depth-first search to traverse through the AST to build our PyDict representation bottom-up. In other words, if a clause contains a subquery, we process the subquery tree as an independent SQL AST and build a dictionary for it. Then, we combine it with other substructures of the clause with different dictionary keys. For example, in Table F.1, we first build the dictionary for "subquery0" and assign this identifier as the key. In the main "clause," we replace the subquery's corresponding span with this identifier. Finally, we use another dictionary to wrap the main "clause" and "subquery0" together as the final representation of the "where" clause. We repeat this procedure for each clause to incrementally add (key, value) pairs to the dictionary and "store" it to the variable sql, which we refer to in program edit representations. + +# B Text-to-SQL Parser Selection + +We choose existing text-to-SQL parsers in our experiments according to two principles: the parsers predict database entity values, and they cover different decoding strategies, including grammar-based (BRIDGEv2), bottom-up (SmBop), and token-based (CodeT5). We did not include parsers using top-down decoders because they usually cannot predict entity values in conditional statements, such as RAT-SQL (Wang et al., 2020). Instead, we include BRIDGEv2 because its decoding method mimics + +the left-to-right CFG derivation of a program, and it uses SQL syntax-based constraints to prevent grammatical errors. In recent work, such decoders, also used in PICARD (Scholak et al., 2021), are more popular than top-down decoders. + +# C Implementation Details + +Our models (Section 3.2) are implemented in PyTorch (Paszke et al., 2019) using Huggingface (Wolf et al., 2020) and trained on a single NVIDIA RTX A6000 GPU (48GB). We use Adafactor (Shazeer and Stern, 2018) to train all our models with the same hyperparameters adapted from Mosbach et al. (2021): + +- Learning rate: $3e - 5$ +- Batch size: 16 +- Epochs: 10 +- Scheduler: Linear decay with $10\%$ warmup + +# D Statistical Significance Test + +To demonstrate the effectiveness of our three clause-level edit representations (Section 4.1), we perform McNemar's Test (McNemar, 1947) to measure the statistical significance of their results in comparison to CodeT5-SQL+Token-Level. For each significance test between two models, we use the median results among our three runs to calculate the comparison matrix. Then, we compute the $p$ -values using statsmodels. When $p < 0.05$ we reject the null hypothesis. In other words, we consider the accuracy improvement statistically significant when $p < 0.05$ . + +# E Additional Results + +Results on our development set. We report model performances on our held-out development set (Section 3.1) in Table E.1. During training, we select the best model by evaluating its EX and EM accuracy on the development set (Section 3.3) every 500 steps. Surprisingly, we find that CodeT5-SQL+Clause-Level sometimes achieves the best performance. For BRIDGEv2, it obtains 35.9 EM accuracy and 39.3 EX accuracy, while CodeT5-PyDict+Program only obtains 34.5 EM accuracy and 37.1 EX accuracy. A possible explanation is that in comparison to the test set, our development set has SQL structures and databases that are more + +similar to the training set, while the test set has unseen SQL structures and less similar databases. It may also indicate that CodeT5-SQL+Clause-Level overfits the synthetic training data and fails to generalize to realistic test data. + +Results for simulated interaction experiments. To show the potential of using our model in an interactive framework, we extend our main experiments (Section 4.1) by adding simulated user interactions. Since our model uses beam search to decode the edit actions $\mathbf{e} = \{e_1,e_2,\dots ,e_n\}$ and the resulting correct SQL query $\mathbf{q}_{+}$ (Equation 1), we simulate user interactions to select one edit action $e_i$ at a time from the beam results. + +At each time step $t$ , we prompt the decoder with previously selected edit actions $e_1, \ldots, e_{t-1}$ to complete the sequence $e_t, \ldots, e_n$ , $\mathbf{q}_+$ using beam search with size 3. Then, we use gold SQL annotations to simulate the user interaction, which selects an edit action $e_t$ from the three candidates at step $t$ or chooses to skip the current step when all three candidates are wrong. If skipping, the user continues to check the consequent edit actions $e_{t+j}$ ( $j = 1, 2, \ldots, n-t$ ) until it selects the next edit action. When the interaction finishes, we append the selected edit action to the prompt and let the model regenerate a completion with the new prompt for the next step's interaction. Having simulated interactions for all edit actions, we do not use the generated $\mathbf{q}_+$ directly because some edit actions are skipped. Instead, we execute the selected ones on the initial SQL query to derive the final query. + +As shown in Table E.2, when collaborating with a simulated user, our error correction model can further improve the base parsers' accuracy. Compared to its performance without using any interactions, our model achieves up to 4.1 point more absolute improvement on EM accuracy (72.5 → 76.6; BRIDGEv2) and 5.0 point more absolute improvement on EX accuracy (73.1 → 78.1; BRIDGEv2). With these results for simulated interaction experiments, we deem that incorporating our error correction model into an interactive framework is a promising future direction. + +
ModelsQueryEditCodeT5BRIDGEv2SmBoP
EMEXEMEXEMEX
CodiT5SQLToken-Level26.1 (0.4)28.6 (1.0)25.8 (0.3)27.2 (0.6)28.1 (0.9)30.7 (0.7)
SQLClause-Level28.6 (0.4)31.3 (0.5)28.4 (0.5)30.0 (0.2)30.2 (0.8)33.4 (0.8)
PyDictClause-Level28.9 (0.6)32.3 (0.8)28.0 (0.1)30.1 (0.2)27.6 (0.1)30.9 (0.4)
CodeT5SQLToken-Level32.1 (1.1)34.1 (1.2)31.8 (0.4)34.5 (0.8)34.2 (0.1)37.6 (0.1)
SQLClause-Level36.5 (0.6)38.6 (0.5)35.9 (0.4)39.3 (1.3)36.1 (0.6)38.8 (0.5)
PyDictClause-Level35.6 (0.9)37.9 (0.3)32.9 (1.0)34.8 (0.8)33.0 (0.2)36.3 (0.3)
CodeT5* CodeT5PyDictProgram35.7 (0.8)37.9 (0.3)34.8 (0.8)38.3 (0.7)36.0 (0.3)40.2 (0.5)
36.7 (0.2)38.5 (0.6)34.5 (0.1)37.1 (0.2)35.6 (0.8)39.0 (0.1)
+ +Table E.1: Exact Set Match (EM) and Execution Match (EX) accuracy on our held-out development set (Section 3.1). The best performances are in bold and the second bests are underlined. *We fine-tune the model to generate edit programs only (without resulting queries) and use Python interpreter to execute the edit actions. + +
ModelsQueryEditCodeT5BRIDGEv2SmBoP
EMEXEMEXEMEX
No EditN/AN/A62.7 (-)63.6 (-)70.1 (-)68.2 (-)74.6 (-)75.3 (-)
CodeT5*69.2 (0.4)68.4 (0.2)72.5 (0.4)73.1 (0.2)77.3 (0.4)77.6 (0.6)
CodeT5PyDictProgram69.0 (0.2)68.2 (0.1)72.5 (0.3)73.0 (0.6)78.0 (0.3)78.5 (0.3)
\(CodeT5^{\dagger}\)PyDictProgram73.0 (0.7)72.9 (0.8)76.6 (0.4)78.1 (0.2)80.0 (0.3)81.2 (0.6)
+ +Table E.2: Exact Set Match (EM) and Execution Match (EX) accuracy on Spider development set. The best performances are in bold. *We fine-tune the model to generate edit programs only (without resulting queries) and use Python interpreter to execute the edit actions. †We simulate user interactions using gold SQL queries to choose edit actions during beam search (size 3) and then execute the chosen actions to get the resulting SQL parse. + +# F More Representation Examples + +We provide two more examples in Table F.1 and F.2 to demonstrate how we represent SQL with subqueries and their edits (Section 2.2). We also show different representations for Insert and Delete edit actions. + +
Query RepresentationEdit Representation
SQLselect count(*) from cars_data where cars_data.accelerate > ( select max(cars_data.horsepower) from cars_data )Token-level<ReplaceOld> max(cars_data.horsepower) <ReplaceNew> cars_data.accelerate <ReplaceEnd> <Insert> order by cars_data.horsepower desc limit 1 <InsertEnd> Clause-level <ReplaceOld> select max(cars_data.horsepower) <ReplaceNew> select cars_data.accelerate <ReplaceEnd> <Insert> order by cars_data.horsepower desc limit 1 <InsertEnd>
PyDictsql = { "select": "select count(*)", "from": "from cars_data", "where": { "clause": "where cars_data.accelerate > (subquery0)", "subquery0": { "select": "select max(cars_data.horsepower)", "from": "from cars_data" } } }Clause-level<ReplaceOld> "select": "select max( cars_data.horsepower)" <ReplaceNew> "select": "select cars_data.accelerate" <ReplaceEnd> <Insert> "orderBy": "order by cars_data.horsepower desc", "limit": "limit 1" <InsertEnd> Programsql["where"},{"subquery0"},{"select"} = "select cars_data.accelerate" sql["where"},{"subquery0"},{"orderBy"} = "order by cars_data.horsepower desc" sql["where"},{"subquery0"},{"limit"} = "limit 1"
+ +Table F.1: Example representations for a wrong SQL query that contains a nested subquery and its edit actions (including Insert edits). The corresponding natural language utterance is "What is the number of cars with a greater acceleration than the one with the most horsepower?" + +
Query RepresentationEdit Representation
SQLselect employee.name from employee join evaluation on employee.employee_id = evaluation.employee_id group by evaluation.employee_id" order by sum(evaluationbonus) desc limit 1Token-level<Delete> group by evaluation.employee_id <DeleteEnd> <DeleteSum( <DeleteEnd><Delete>) <DeleteEnd>
Clause-level<Delete> group by evaluation.employee_id <DeleteEnd> <ReplaceOld> order by sum(evaluation;bONUS) desc <ReplaceNew> order by evaluation;bONUS desc <ReplaceEnd>
PyDictsql = { "select": "select employee.name", "from": "from employee join evaluation on employee.employee_id = evaluation.employee_id", "groupBy": "group by evaluation.employee_id", "orderBy": "order by sum(evaluation;bONUS) desc", "limit": "limit 1" }Clause-level<Delete> "groupId": "group by evaluation.employee_id" <DeleteEnd><ReplaceOld> "orderBy": "order by sum(evaluation;bONUS) desc" <ReplaceNew> "orderBy": "order by evaluation;bONUS desc" <ReplaceEnd>
Programsql.pop("groupId") sql["orderBy"] = "order by evaluation;bONUS desc"
+ +Table F.2: Example representations for a wrong SQL query and its edit actions (including Delete edits). The corresponding natural language utterance is "Find the name of the employee who got the highest one time bonus." + +A For every submission: + +A1. Did you describe the limitations of your work? 6 +A2. Did you discuss any potential risks of your work? Not applicable. Left blank. +A3. Do the abstract and introduction summarize the paper's main claims? +A4. Have you used AI writing assistants when working on this paper? Left blank. + +B Did you use or create scientific artifacts? + +3 + +B1. Did you cite the creators of artifacts you used? 3 +B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Not applicable. Left blank. +B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Not applicable. Left blank. +B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Not applicable. Left blank. +B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? 3 +B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. 3 + +C Did you run computational experiments? + +4 + +C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Appendix B + +C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Appendix B +C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? +C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Appendix B + +D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank. + +D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response. +D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response. +D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response. +D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response. +D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response. \ No newline at end of file diff --git a/2023/Text-to-SQL Error Correction with Language Models of Code/images.zip b/2023/Text-to-SQL Error Correction with Language Models of Code/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c872e03fb99f0925921ed643165f8a5ec991aa5e --- /dev/null +++ b/2023/Text-to-SQL Error Correction with Language Models of Code/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1d7fe9c754c29397f7a6d411ba8554ea4fb5392cdbb5f657e638c7aab3ddbcdc +size 522692 diff --git a/2023/Text-to-SQL Error Correction with Language Models of Code/layout.json b/2023/Text-to-SQL Error Correction with Language Models of Code/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b9b8eb3a27df3fb79c2af442ec4e4e40f6ff2966 --- /dev/null +++ b/2023/Text-to-SQL Error Correction with Language Models of Code/layout.json @@ -0,0 +1,8394 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 105, + 75, + 489, + 94 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 75, + 489, + 94 + ], + "spans": [ + { + "bbox": [ + 105, + 75, + 489, + 94 + ], + "type": "text", + "content": "Text-to-SQL Error Correction with Language Models of Code" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 136, + 105, + 459, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 105, + 459, + 133 + ], + "spans": [ + { + "bbox": [ + 136, + 105, + 459, + 133 + ], + "type": "text", + "content": "Ziru Chen" + }, + { + "bbox": [ + 136, + 105, + 459, + 133 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 136, + 105, + 459, + 133 + ], + "type": "text", + "content": ", Shijie Chen" + }, + { + "bbox": [ + 136, + 105, + 459, + 133 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 136, + 105, + 459, + 133 + ], + "type": "text", + "content": ", Michael White" + }, + { + "bbox": [ + 136, + 105, + 459, + 133 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 136, + 105, + 459, + 133 + ], + "type": "text", + "content": ", Raymond Mooney" + }, + { + "bbox": [ + 136, + 105, + 459, + 133 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 136, + 105, + 459, + 133 + ], + "type": "text", + "content": ", Ali Payani" + }, + { + "bbox": [ + 136, + 105, + 459, + 133 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 136, + 105, + 459, + 133 + ], + "type": "text", + "content": ", Jayanth Srinivasa" + }, + { + "bbox": [ + 136, + 105, + 459, + 133 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 136, + 105, + 459, + 133 + ], + "type": "text", + "content": ", Yu Su" + }, + { + "bbox": [ + 136, + 105, + 459, + 133 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 136, + 105, + 459, + 133 + ], + "type": "text", + "content": ", Huan Sun" + }, + { + "bbox": [ + 136, + 105, + 459, + 133 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 233, + 134, + 364, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 134, + 364, + 148 + ], + "spans": [ + { + "bbox": [ + 233, + 134, + 364, + 148 + ], + "type": "text", + "content": "1The Ohio State University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 166, + 148, + 428, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 148, + 428, + 161 + ], + "spans": [ + { + "bbox": [ + 166, + 148, + 428, + 161 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 166, + 148, + 428, + 161 + ], + "type": "text", + "content": "The University of Texas at Austin " + }, + { + "bbox": [ + 166, + 148, + 428, + 161 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 166, + 148, + 428, + 161 + ], + "type": "text", + "content": "Cisco Research" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 117, + 162, + 479, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 162, + 479, + 176 + ], + "spans": [ + { + "bbox": [ + 117, + 162, + 479, + 176 + ], + "type": "text", + "content": "{chen.8336, chen.10216, white.1240, su.809, sun.397}@osu.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 154, + 177, + 441, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 177, + 441, + 190 + ], + "spans": [ + { + "bbox": [ + 154, + 177, + 441, + 190 + ], + "type": "text", + "content": "mooney@cs.utexas.edu {apayani, jasriniv}@cisco" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 84, + 232, + 274, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 232, + 274, + 460 + ], + "spans": [ + { + "bbox": [ + 84, + 232, + 274, + 460 + ], + "type": "text", + "content": "Despite recent progress in text-to-SQL parsing, current semantic parsers are still not accurate enough for practical use. In this paper, we investigate how to build automatic text-to-SQL error correction models. Noticing that token-level edits are out of context and sometimes ambiguous, we propose building clause-level edit models instead. Besides, while most language models of code are not specifically pre-trained for SQL, they know common data structures and their operations in programming languages such as Python. Thus, we propose a novel representation for SQL queries and their edits that adheres more closely to the pre-training corpora of language models of code. Our error correction model improves the exact set match accuracy of different parsers by 2.4-6.5 and obtains up to 4.3 point absolute improvement over two strong baselines." + }, + { + "bbox": [ + 84, + 232, + 274, + 460 + ], + "type": "inline_equation", + "content": "^1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 68, + 467, + 154, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 467, + 154, + 479 + ], + "spans": [ + { + "bbox": [ + 68, + 467, + 154, + 479 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 488, + 291, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 488, + 291, + 716 + ], + "spans": [ + { + "bbox": [ + 67, + 488, + 291, + 716 + ], + "type": "text", + "content": "Text-to-SQL parsing is a classic semantic parsing task that finds wide applications (Zelle and Mooney, 1996; Tang and Mooney, 2000). Since the release of Spider (Yu et al., 2018), a cross-database text-to-SQL benchmark, many semantic parsers with decent performance have been developed (Lin et al., 2020; Wang et al., 2020; Deng et al., 2021; Rubin and Berant, 2021; Scholak et al., 2021). Nonetheless, state-of-the-art semantic parsers are still not accurate enough. As a result, their users need to constantly correct wrongly predicted SQL queries, which can be as time-consuming and error-prone as writing a SQL query from scratch (Jorgensen and Shepperd, 2007; Weiss et al., 2007). Therefore, in this paper, we study the problem of automatic text-to-SQL error correction to better assist users in querying complex databases." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 719, + 290, + 745 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 719, + 290, + 745 + ], + "spans": [ + { + "bbox": [ + 67, + 719, + 290, + 745 + ], + "type": "text", + "content": "We first highlight that it is essential to factor in the compositional substructures within SQL" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 213, + 526, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 213, + 526, + 456 + ], + "spans": [ + { + "bbox": [ + 302, + 213, + 526, + 456 + ], + "type": "text", + "content": "queries, such as abstract syntax trees (Yin and Neubig, 2017; Guo et al., 2022) and data-flow graphs (Guo et al., 2021), instead of treating code snippets as string sequences. Compared to individual tokens, substructures (e.g. SQL clauses) include more context of the entire program and are more semantically meaningful. Consequently, edit patterns of such substructures are more intuitive for humans to understand and easier for language models to learn. Moreover, while the pre-training corpora for language models of code, such as CodeT5 (Wang et al., 2021), do not include many SQL queries based on their documentation, they naturally contain abundant examples of common data structures like dictionaries. Therefore, we hypothesize that transforming unfamiliar SQL queries into familiar data structures can help language models of code better perform structural editing of SQL queries." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 460, + 526, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 460, + 526, + 730 + ], + "spans": [ + { + "bbox": [ + 302, + 460, + 526, + 730 + ], + "type": "text", + "content": "Based on these observations, we develop our error correction model and make two contributions. First, we propose considering SQL clauses instead of tokens as basic semantic units for editing. Using a context-free grammar, we can decompose a SQL query and identify its clauses by traversing its abstract syntax tree. Second, we propose a new representation of SQL queries and their edits that adheres more closely to common code pre-training corpora, including CodeSearchNet (Husain et al., 2020), and makes the structures of a SQL query more explicit. With a decomposed SQL query, we pair each clause with its SQL keyword and represent the entire query as a Python dictionary. Then, we format edits on a wrong SQL query as a program that modifies data of the query's corresponding dictionary. Unlike token-level edits in existing work (Zhang et al., 2023), such dictionary operations define all edits unambiguously and can be directly executed with a Python interpreter." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 732, + 525, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 732, + 525, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 732, + 525, + 773 + ], + "type": "text", + "content": "Through comprehensive experiments with different representations, we show that: (1) our proposed representation has the lowest zero-shot perplexity" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 750, + 291, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 750, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 750, + 291, + 772 + ], + "type": "text", + "content": "1Our code and data are available at https://github. com/OSU-NLP-Group/Auto-SQL-Correction." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "1359" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 135, + 795, + 459, + 805 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 795, + 459, + 805 + ], + "spans": [ + { + "bbox": [ + 135, + 795, + 459, + 805 + ], + "type": "text", + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 219, + 806, + 375, + 817 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 806, + 375, + 817 + ], + "spans": [ + { + "bbox": [ + 219, + 806, + 375, + 817 + ], + "type": "text", + "content": "Volume 2: Short Papers, pages 1359-1372" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "spans": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "text", + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 68, + 526, + 198 + ], + "blocks": [ + { + "bbox": [ + 70, + 68, + 526, + 198 + ], + "lines": [ + { + "bbox": [ + 70, + 68, + 526, + 198 + ], + "spans": [ + { + "bbox": [ + 70, + 68, + 526, + 198 + ], + "type": "table", + "html": "
Query RepresentationEdit Representation
SQLselect tweets.text from tweets order by tweets.textToken-Level<ReplaceOld> tweets.text <ReplaceNew> tweets.createDate <ReplaceEnd>
Clause-Level<ReplaceOld> order by tweets.text <ReplaceNew> order by tweets.createDate <ReplaceEnd>
PyDictsql = { "select": "select tweets.text", "from": "from tweets", "orderBy": "order by tweets.text" }Clause-Level<ReplaceOld> "orderBy": "order by tweets.text" <ReplaceNew> "orderBy": "order by tweets.createDate" <ReplaceEnd> sql["orderBy"] = "order by tweets.createDate"
Program
", + "image_path": "0782fcf933bf83d430bc48da17f22fb97371b382cc22a33bfee1d81242f29482.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 206, + 525, + 255 + ], + "lines": [ + { + "bbox": [ + 66, + 206, + 525, + 255 + ], + "spans": [ + { + "bbox": [ + 66, + 206, + 525, + 255 + ], + "type": "text", + "content": "Table 1: Example representations for a wrong SQL query and the Replace edit action. The corresponding natural language utterance is \"List the text of all tweets in the order of date.\" For token-level and clause-level representations, we format them as \" Span of wrong tokens/clauses Span of correct tokens/clauses \", where , , and are special tokens." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 265, + 290, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 265, + 290, + 346 + ], + "spans": [ + { + "bbox": [ + 67, + 265, + 290, + 346 + ], + "type": "text", + "content": "with CodeT5; (2) simply changing token-level edits to clause-level edits can effectively improve the performance of our models; and (3) our method improves the exact set match accuracy of different parsers by 2.4-6.5 and obtains up to 4.3 point absolute improvement over two strong baselines." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 355, + 244, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 355, + 244, + 370 + ], + "spans": [ + { + "bbox": [ + 67, + 355, + 244, + 370 + ], + "type": "text", + "content": "2 Text-to-SQL Error Correction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 377, + 290, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 377, + 290, + 471 + ], + "spans": [ + { + "bbox": [ + 67, + 377, + 290, + 471 + ], + "type": "text", + "content": "Given a natural language utterance " + }, + { + "bbox": [ + 67, + 377, + 290, + 471 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 67, + 377, + 290, + 471 + ], + "type": "text", + "content": ", a database schema " + }, + { + "bbox": [ + 67, + 377, + 290, + 471 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 67, + 377, + 290, + 471 + ], + "type": "text", + "content": ", and a wrong SQL query " + }, + { + "bbox": [ + 67, + 377, + 290, + 471 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_{-}" + }, + { + "bbox": [ + 67, + 377, + 290, + 471 + ], + "type": "text", + "content": " produced by an existing parser, our goal is to develop an error correction model that predicts a sequence of edit actions " + }, + { + "bbox": [ + 67, + 377, + 290, + 471 + ], + "type": "inline_equation", + "content": "\\mathbf{e}" + }, + { + "bbox": [ + 67, + 377, + 290, + 471 + ], + "type": "text", + "content": " and the correct query " + }, + { + "bbox": [ + 67, + 377, + 290, + 471 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_{+}" + }, + { + "bbox": [ + 67, + 377, + 290, + 471 + ], + "type": "text", + "content": ". Following previous work (Zhang et al., 2023), we formulate our task as sequence-to-sequence generation:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 479, + 290, + 496 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 479, + 290, + 496 + ], + "spans": [ + { + "bbox": [ + 107, + 479, + 290, + 496 + ], + "type": "interline_equation", + "content": "P (\\mathbf {y} | \\mathbf {x}) = \\Pi_ {t = 1} ^ {T} P (\\mathbf {y} _ {t} | \\mathbf {x}, \\mathbf {y} _ {1: t - 1}) \\tag {1}", + "image_path": "50d881196a540e1c7aceb9547f1709249d5b78199a6d53f3794fcc0718769783.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 503, + 290, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 503, + 290, + 585 + ], + "spans": [ + { + "bbox": [ + 67, + 503, + 290, + 585 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 67, + 503, + 290, + 585 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = [\\mathbf{u};\\mathbf{s};\\mathbf{q}_{-}]" + }, + { + "bbox": [ + 67, + 503, + 290, + 585 + ], + "type": "text", + "content": " is the concatenation of the given inputs and " + }, + { + "bbox": [ + 67, + 503, + 290, + 585 + ], + "type": "inline_equation", + "content": "\\mathbf{y} = [\\mathbf{e};\\mathbf{q}_{+}]" + }, + { + "bbox": [ + 67, + 503, + 290, + 585 + ], + "type": "text", + "content": " is the concatenation of all edit actions and the resulting correct query. In this section, we study different representations of SQL queries (Section 2.1) and edits (Section 2.2) to better leverage language models of code." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 593, + 199, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 593, + 199, + 607 + ], + "spans": [ + { + "bbox": [ + 67, + 593, + 199, + 607 + ], + "type": "text", + "content": "2.1 Query Representation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 611, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 611, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 611, + 291, + 772 + ], + "type": "text", + "content": "We consider two representations for a predicted query: (1) the original SQL format and (2) our proposed PyDict (Python Dictionary) representation. To prepare for editing, we disambiguate each SQL query following Rubin and Berant (2021), including lower-casing non-value tokens, resolving table references, and formatting punctuation. This preprocessing normalizes SQL queries predicted by different base parsers and the gold annotations into the same format. To build our PyDict representation, we parse a SQL query into its abstract syntax tree (AST) with Spider's context-free grammar. We" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 265, + 525, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 265, + 525, + 333 + ], + "spans": [ + { + "bbox": [ + 302, + 265, + 525, + 333 + ], + "type": "text", + "content": "use depth-first search to traverse through the AST, find any nested substructures, and construct the dictionary representation bottom-up. Table 1 shows the \"SQL\" and \"PyDict\" representations of a SQL query (more details in Appendix A)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 345, + 425, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 345, + 425, + 358 + ], + "spans": [ + { + "bbox": [ + 302, + 345, + 425, + 358 + ], + "type": "text", + "content": "2.2 Edit Representation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 301, + 364, + 526, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 364, + 526, + 663 + ], + "spans": [ + { + "bbox": [ + 301, + 364, + 526, + 663 + ], + "type": "text", + "content": "We first follow Zhang et al. (2023) to use token-level edit representation with special tokens (Table 1), which have unique entries in the tokenizer and the model's embedding layer to describe Replace, Insert, and Delete edit actions (more examples in Appendix F). However, we realize this representation can sometimes be ambiguous. As shown in Table 1, the span \"tweets.text\" appears twice in the SQL query. This repetition would confuse the error correction model with which span to replace when generating the corrected query. Also, the ambiguity makes it difficult to implement rules and directly carry out the edit actions on the wrong query. Hence, we change the token-level edit representation to clause-level, which includes more context of the query to make different edits more distinguishable. In our experiments (Section 4.1), we demonstrate that this simple modification is already effective. Our program representation further improves the performance because it is more similar to the code pre-training corpora and eliminates the need to learn special tokens' representations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 676, + 427, + 690 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 676, + 427, + 690 + ], + "spans": [ + { + "bbox": [ + 302, + 676, + 427, + 690 + ], + "type": "text", + "content": "3 Experimental Setup" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 699, + 524, + 713 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 699, + 524, + 713 + ], + "spans": [ + { + "bbox": [ + 302, + 699, + 524, + 713 + ], + "type": "text", + "content": "3.1 Data Synthesis for SQL Error Correction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 719, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 719, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 719, + 525, + 772 + ], + "type": "text", + "content": "To train a text-to-SQL error correction model, we need to collect a set of wrong SQL parses that reflects a realistic distribution of errors (Section 4.2) as our training data. We synthesize this dataset by" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "text", + "content": "1360" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 72, + 68, + 286, + 158 + ], + "blocks": [ + { + "bbox": [ + 72, + 68, + 286, + 158 + ], + "lines": [ + { + "bbox": [ + 72, + 68, + 286, + 158 + ], + "spans": [ + { + "bbox": [ + 72, + 68, + 286, + 158 + ], + "type": "table", + "html": "
CodeT5BRIDGEv2SmBoP
# of Train47,02024,77620,083
# of Dev448448448
# of Test430392310
Avg. Train Edits2.343.112.72
Avg. Dev Edits2.703.293.31
Avg. Test Edits1.841.511.47
", + "image_path": "315c2a10656cad34aaee909b8e17728dbacd044f2f0177a2034af0a9be635d74.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 167, + 251, + 179 + ], + "lines": [ + { + "bbox": [ + 105, + 167, + 251, + 179 + ], + "spans": [ + { + "bbox": [ + 105, + 167, + 251, + 179 + ], + "type": "text", + "content": "Table 2: Summary of data statistics." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 192, + 290, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 192, + 290, + 218 + ], + "spans": [ + { + "bbox": [ + 67, + 192, + 290, + 218 + ], + "type": "text", + "content": "performing 5-fold cross-validation on each parser, which approximates the actual evaluation setting." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 220, + 291, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 220, + 291, + 516 + ], + "spans": [ + { + "bbox": [ + 69, + 220, + 291, + 516 + ], + "type": "text", + "content": "Following the evaluation setup in Yu et al. (2018), we split Spider's training set into five roughly equal subsets by different databases. For each cross-validation fold, we train a text-to-SQL parser (Section 3.2) on four subsets and evaluate it on the remaining one. At inference time, we perform beam search with size 20 for each example and collect grammatical and executable parses in the beam. If a SQL parse is not an exact set match or execution match to the gold annotation, we label it wrong and include it in our training set for error correction. Having synthesized our training dataset, we randomly sample 8 databases and their associated questions to construct a held-out development set. For development set examples, we only keep incorrect SQL parses with the highest beam confidence. For our error correction test set, we train each parser on the full Spider training set and evaluate it on the original Spider's development set without modifications. We similarly keep SQL parses with exact match or execution match errors. Table 2 summarizes the statistics of our data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 530, + 130, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 530, + 130, + 541 + ], + "spans": [ + { + "bbox": [ + 67, + 530, + 130, + 541 + ], + "type": "text", + "content": "3.2 Models" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 550, + 290, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 550, + 290, + 602 + ], + "spans": [ + { + "bbox": [ + 67, + 550, + 290, + 602 + ], + "type": "text", + "content": "Text-to-SQL base parsers. We choose three text-to-SQL parsers with different decoding strategies and levels of performance (Table 3). We elaborate on our selection criteria in Appendix B." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 80, + 605, + 290, + 740 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 80, + 605, + 290, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 605, + 290, + 658 + ], + "spans": [ + { + "bbox": [ + 80, + 605, + 290, + 658 + ], + "type": "text", + "content": "- CodeT5 (Wang et al., 2021): We fine-tune CodeT5-base following Xie et al. (2022). This parser represents those using beam search decoding and having a lower accuracy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 80, + 659, + 290, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 659, + 290, + 698 + ], + "spans": [ + { + "bbox": [ + 80, + 659, + 290, + 698 + ], + "type": "text", + "content": "BRIDGEv2 (Lin et al., 2020): A representative parser with constrained decoding and achieving a medium-level accuracy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 80, + 700, + 290, + 740 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 700, + 290, + 740 + ], + "spans": [ + { + "bbox": [ + 80, + 700, + 290, + 740 + ], + "type": "text", + "content": "- SmBoP (Rubin and Berant, 2021): A representative parser with bottom-up decoding and achieving higher accuracy." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 302, + 71, + 524, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 524, + 97 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 524, + 97 + ], + "type": "text", + "content": "Error correction models. We use two language models of code in all our experiments:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 99, + 525, + 219 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 316, + 99, + 525, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 99, + 525, + 164 + ], + "spans": [ + { + "bbox": [ + 316, + 99, + 525, + 164 + ], + "type": "text", + "content": "CoditT5 (Zhang et al., 2023): A language model pre-trained for code editing tasks by injecting noises to code snippets in CodeSearchNet (Husain et al., 2020) and then denoising with token-level edit representations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 166, + 525, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 166, + 525, + 219 + ], + "spans": [ + { + "bbox": [ + 316, + 166, + 525, + 219 + ], + "type": "text", + "content": "- CodeT5 (Wang et al., 2021): A language model pre-trained for general code understanding and generation with four different pre-training objectives." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 302, + 220, + 526, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 220, + 526, + 287 + ], + "spans": [ + { + "bbox": [ + 302, + 220, + 526, + 287 + ], + "type": "text", + "content": "We compare the existing SQL+Token-Level representation with our proposed ones: SQL+Clause-Level, PyDict+Clause-Level, and PyDict+Program on CodeT5 and the first three on CoditT5.3 Implementation details are in Appendix C." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 296, + 381, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 296, + 381, + 308 + ], + "spans": [ + { + "bbox": [ + 302, + 296, + 381, + 308 + ], + "type": "text", + "content": "3.3 Evaluation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 301, + 313, + 525, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 313, + 525, + 503 + ], + "spans": [ + { + "bbox": [ + 301, + 313, + 525, + 503 + ], + "type": "text", + "content": "We use the increase in Exact Set Match (EM) and Execution Match (EX) accuracy on our error correction test set to measure each model's performance. Because CoditT5's experiments assume the input program has at least one error, we keep this assumption for fair comparisons. To construct a test set satisfying this assumption, we have to compare parser-generated SQL queries with gold annotations (Section 3.1). Thus, we use the Spider development set as our test set and split the Spider training set to build a held-out development set (Table 2) to select model checkpoints during training. We also include results on our held-out development set in the appendix (Table E.1)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 302, + 513, + 430, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 513, + 430, + 526 + ], + "spans": [ + { + "bbox": [ + 302, + 513, + 430, + 526 + ], + "type": "text", + "content": "4 Results and Analysis" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 302, + 534, + 393, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 534, + 393, + 545 + ], + "spans": [ + { + "bbox": [ + 302, + 534, + 393, + 545 + ], + "type": "text", + "content": "4.1 Main Results" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 301, + 551, + 524, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 551, + 524, + 660 + ], + "spans": [ + { + "bbox": [ + 301, + 551, + 524, + 660 + ], + "type": "text", + "content": "We summarize our main results in this section. To ensure robustness, we repeat all experiments with 3 different random seeds and report the average performances with standard deviations. Our model can also be used in an interactive framework that allows users to select edit actions from the top-" + }, + { + "bbox": [ + 301, + 551, + 524, + 660 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 301, + 551, + 524, + 660 + ], + "type": "text", + "content": " beam candidates. We include more experiments with simulated user interactions in Appendix E." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 302, + 667, + 526, + 735 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 667, + 526, + 735 + ], + "spans": [ + { + "bbox": [ + 302, + 667, + 526, + 735 + ], + "type": "text", + "content": "Our representation's perplexity is the smallest. We validate that our PyDict+Program representation adheres more closely to the code pre-training corpora by measuring its zero-shot perplexity on CodeT5 using our development set (Section 3.1)." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 750, + 290, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 750, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 750, + 290, + 772 + ], + "type": "text", + "content": "Due to SmBoP's bottom-up decoding, we keep its original beam size and collect the top-20 unique beam predictions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 740, + 525, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 740, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 740, + 525, + 772 + ], + "type": "text", + "content": "3We did not use CoditT5 for PyDict+Program because it was pre-trained on token-level edit representations. Its decoder may be specialized in generating edits instead of programs." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 286, + 781, + 308, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 308, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 308, + 791 + ], + "type": "text", + "content": "1361" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 71, + 68, + 522, + 218 + ], + "blocks": [ + { + "bbox": [ + 71, + 68, + 522, + 218 + ], + "lines": [ + { + "bbox": [ + 71, + 68, + 522, + 218 + ], + "spans": [ + { + "bbox": [ + 71, + 68, + 522, + 218 + ], + "type": "table", + "html": "
ModelsQueryEditCodeT5BRIDGEv2SmBoP
EMEXEMEXEMEX
No EditN/AN/A62.7 (-)63.6 (-)70.1 (-)68.2 (-)74.6 (-)75.3 (-)
CodiT5SQLToken-Level64.3 (0.1)64.4 (0.2)65.4 (0.5)66.6 (0.3)74.2 (0.4)75.3 (0.1)
SQLClause-Level67.0 (0.4)65.4 (0.5)71.3 (0.5)70.9 (0.2)76.3 (0.0)77.2 (0.3)
PyDictClause-Level67.1 (0.2)66.5 (0.4)70.6 (0.8)70.8 (0.6)76.3 (0.3)77.0 (0.3)
CodeT5SQLToken-Level66.7 (0.9)65.9 (0.5)68.2 (0.4)69.4 (0.8)75.6 (0.4)76.5 (0.6)
SQLClause-Level68.3 (0.3)\\( \\underline{68.2}^{+}(0.6) \\)71.8+(0.4)72.5+(0.2)76.7 (0.6)77.4 (0.3)
PyDictClause-Level66.6 (0.8)67.1 (0.8)72.0+(0.3)72.4+(0.2)77.3 (0.6)77.8 (0.2)
\\( CodeT5^* \\)CodeT5PyDictProgram\\( 69.2^{+}(0.4) \\)\\( 68.4^{+}(0.2) \\)\\( 72.5^{+}(0.4) \\)\\( 73.1^{+}(0.2) \\)77.3 (0.4)77.6 (0.6)
\\( 69.0^{+}(0.2) \\)\\( 68.2^{+}(0.1) \\)\\( 72.5^{+}(0.3) \\)\\( 73.0^{+}(0.6) \\)\\( 78.0^{+}(0.3) \\)\\( 78.5^{+}(0.3) \\)
", + "image_path": "232333084ea86b3a93da677a41cefb6a766b303c7f67d3da580c35bf61526f39.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 226, + 526, + 287 + ], + "lines": [ + { + "bbox": [ + 67, + 226, + 526, + 287 + ], + "spans": [ + { + "bbox": [ + 67, + 226, + 526, + 287 + ], + "type": "text", + "content": "Table 3: Exact Set Match (EM) and Execution Match (EX) accuracy on Spider development set. The best performances are in bold and the second bests are underlined. Results with " + }, + { + "bbox": [ + 67, + 226, + 526, + 287 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 67, + 226, + 526, + 287 + ], + "type": "text", + "content": " are statistically significant (McNemar's; " + }, + { + "bbox": [ + 67, + 226, + 526, + 287 + ], + "type": "inline_equation", + "content": "p < 0.05" + }, + { + "bbox": [ + 67, + 226, + 526, + 287 + ], + "type": "text", + "content": " ) compared to CodeT5-SQL+Token-Level (Appendix D). Otherwise, the results are not statistically significant. \\*We fine-tune the model to generate edit programs only (without resulting queries) and use Python interpreter to execute the edit actions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 83, + 297, + 276, + 413 + ], + "blocks": [ + { + "bbox": [ + 83, + 297, + 276, + 413 + ], + "lines": [ + { + "bbox": [ + 83, + 297, + 276, + 413 + ], + "spans": [ + { + "bbox": [ + 83, + 297, + 276, + 413 + ], + "type": "image", + "image_path": "a38959892ed71b38c996b10dff3a75e11f3afbcab1432d33e61060f6c2917496.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 422, + 291, + 460 + ], + "lines": [ + { + "bbox": [ + 67, + 422, + 291, + 460 + ], + "spans": [ + { + "bbox": [ + 67, + 422, + 291, + 460 + ], + "type": "text", + "content": "Figure 1: CodeT5's zero-shot perplexity (in log scale) of all four representations on our synthesized SQL error development set." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 468, + 291, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 468, + 291, + 562 + ], + "spans": [ + { + "bbox": [ + 67, + 468, + 291, + 562 + ], + "type": "text", + "content": "As shown in Figure 1, by representing data in Py-Dict, we can reduce the perplexity of CodeT5 by 2 orders of magnitude. After augmenting it with our program representation, we further reduce the zero-shot perplexity of CodeT5 to only " + }, + { + "bbox": [ + 67, + 468, + 291, + 562 + ], + "type": "inline_equation", + "content": "5.96 \\times 10^{2}" + }, + { + "bbox": [ + 67, + 468, + 291, + 562 + ], + "type": "text", + "content": ", 3 orders of magnitude less than the SQL+Token-Level representation " + }, + { + "bbox": [ + 67, + 468, + 291, + 562 + ], + "type": "inline_equation", + "content": "(1.26 \\times 10^{5})" + }, + { + "bbox": [ + 67, + 468, + 291, + 562 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 570, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 570, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 570, + 291, + 773 + ], + "type": "text", + "content": "Clause-level editing is more effective, especially when represented in PyDict+Program. Since CodeT5 consistently outperforms CoditT5 with the same representations, we focus on comparisons among CodeT5 variations. As shown in Table 3, compared to CodeT5-SQL+Token-Level, only CodeT5-PyDict+Program achieves statistically significant improvement on all three parsers, while clause-level models fail McNemar's significance test for some parsers. More concretely, it achieves up to 4.3 point more absolute improvement on EM accuracy (68.2 → 72.5; BRIDGEv2) and 3.7 point more absolute improvement on EX accuracy (69.4 → 73.1; BRIDGEv2). Overall, CodeT5-PyDict+Program can boost the parsers' EM accu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 296, + 525, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 296, + 525, + 338 + ], + "spans": [ + { + "bbox": [ + 302, + 296, + 525, + 338 + ], + "type": "text", + "content": "racy by 2.4-6.5. Thus, both clause-level editing and PyDict+Program representation can better take advantage of language models of code." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 349, + 400, + 362 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 349, + 400, + 362 + ], + "spans": [ + { + "bbox": [ + 302, + 349, + 400, + 362 + ], + "type": "text", + "content": "4.2 Error Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 369, + 525, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 369, + 525, + 409 + ], + "spans": [ + { + "bbox": [ + 302, + 369, + 525, + 409 + ], + "type": "text", + "content": "Additionally, we conduct an error analysis (Table 4) by sampling 100 wrong parses from all three parsers and classifying them into five categories:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 316, + 411, + 525, + 571 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 316, + 411, + 525, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 525, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 525, + 449 + ], + "type": "text", + "content": "- Database Grounding: A generated SQL query has the correct structure, but some table/column names or entity values are wrong." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 316, + 451, + 524, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 451, + 524, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 451, + 524, + 476 + ], + "type": "text", + "content": "- Incorrect Structure: A generated SQL query has missing, wrong, or redundant structures." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 316, + 478, + 524, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 524, + 504 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 524, + 504 + ], + "type": "text", + "content": "- Syntax & Grammar: A generated SQL query violates the programming language's syntax." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 505, + 525, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 505, + 525, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 505, + 525, + 544 + ], + "type": "text", + "content": "- False Negative: A generated SQL query is semantically correct but not captured by evaluation metrics, or the gold annotation is wrong." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 545, + 525, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 525, + 571 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 525, + 571 + ], + "type": "text", + "content": "- Other: All other errors, such as wrong aggregation functions, besides the above categories." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 302, + 573, + 525, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 573, + 525, + 613 + ], + "spans": [ + { + "bbox": [ + 302, + 573, + 525, + 613 + ], + "type": "text", + "content": "Since the error distributions for each parser are similar, as an example, we discuss our findings based on the strongest parser, SmBoP:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 624, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 624, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 624, + 526, + 772 + ], + "type": "text", + "content": "Database grounding is the major type of error. Among the 100 samples from SmBoP, we find that 54 of them have database grounding errors. Particularly, SmBoP predicts wrong table/column names in 34 parses, inaccurate entity values in 9 parses, and incorrect JOIN relations in 11 parses. Our CodeT5-PyDict+Program model can successfully fix 16 of the 54 erroneous parses, including 10 parses with wrong table/column names, 4 parses with inaccurate entity values, and 2 parses with incorrect JOIN relations. We hypothesize that" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "text", + "content": "1362" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 68, + 524, + 158 + ], + "blocks": [ + { + "bbox": [ + 70, + 68, + 524, + 158 + ], + "lines": [ + { + "bbox": [ + 70, + 68, + 524, + 158 + ], + "spans": [ + { + "bbox": [ + 70, + 68, + 524, + 158 + ], + "type": "table", + "html": "
Error CategoryCodeT5BRIDGEv2SmBoP
ResolvedUnresolvedAllResolvedUnresolvedAllResolvedUnresolvedAll
Database Grounding155166144862163854
Incorrect Structure215172121432326
Syntax & Grammar000000145
False Negative099066088
Other17821618167
", + "image_path": "6668fd7430620200f8ebe85aeac0e4415b4c2d92ccf59e9f74146938cbc233af.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 167, + 525, + 191 + ], + "lines": [ + { + "bbox": [ + 67, + 167, + 525, + 191 + ], + "spans": [ + { + "bbox": [ + 67, + 167, + 525, + 191 + ], + "type": "text", + "content": "Table 4: Analysis of 100 sample errors made by each text-to-SQL parser. We group the errors into 5 categories and examine if our CodeT5-PyDict+Program model resolves them." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 66, + 201, + 292, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 201, + 292, + 364 + ], + "spans": [ + { + "bbox": [ + 66, + 201, + 292, + 364 + ], + "type": "text", + "content": "database grounding is also a major category of errors in our synthesized training set, so our model has learned to resolve similar errors. Nevertheless, it still cannot correct the remaining 38 SQL parses. We notice that our current representation for database schema is missing critical information, such as column data types and foreign key relations, for our error correction model to fix database grounding errors. Following our PyDict representation for SQL, we suggest designing a code representation for database schema that includes such information to tackle this issue in future work." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 374, + 291, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 374, + 291, + 591 + ], + "spans": [ + { + "bbox": [ + 66, + 374, + 291, + 591 + ], + "type": "text", + "content": "Structural errors are hard to edit automatically. Besides database grounding, 26 of SmBoP's errors belong to another category, incorrect structure. These 26 samples contain 7 parses with incorrect SQL clauses and 19 parses with incorrect subqueries, but our CodeT5-PyDict+Program model only resolves 1 and 2 of them, respectively. We find that correcting such errors usually involves multiple edit steps, which motivates us to incorporate our model into an interactive framework in future work. As our experiments with simulated user interaction (Appendix E.2) show, when our model interacts with the simulated user to correct one clause at a time, it is able to fully correct more SQL parses. Thus, we deem interactive correction would maximize our model's utility in practice." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 602, + 161, + 614 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 602, + 161, + 614 + ], + "spans": [ + { + "bbox": [ + 67, + 602, + 161, + 614 + ], + "type": "text", + "content": "5 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 624, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 624, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 624, + 291, + 773 + ], + "type": "text", + "content": "Since the release of CodeBERT (Feng et al., 2020), many language models of code have emerged for program understanding and generation (Ahmad et al., 2021; Chen et al., 2021; Guo et al., 2021; Wang et al., 2021; Guo et al., 2022; Fried et al., 2023; Nijkamp et al., 2023). In addition to program-related tasks, recent work shows they also excel at processing natural language structures. Using code as meaning representations (MRs), we can leverage language models of code in various tasks, such as commonsense reasoning (Madaan et al., 2022)," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 201, + 526, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 201, + 526, + 269 + ], + "spans": [ + { + "bbox": [ + 302, + 201, + 526, + 269 + ], + "type": "text", + "content": "action planning (Singh et al., 2022), and event extraction (Wang et al., 2022). In fact, how to design MRs to reduce model learning difficulty is a salient research question in semantic parsing (Guo et al., 2019; Gan et al., 2021b; Nie et al., 2022)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 269, + 526, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 269, + 526, + 500 + ], + "spans": [ + { + "bbox": [ + 302, + 269, + 526, + 500 + ], + "type": "text", + "content": "Our work demonstrates that program-related tasks themselves can also benefit from code-based MRs. Specifically, we apply such MRs to SQL error correction, a variant of automatic program repair tasks (Tufano et al., 2019; Panthaplackel et al., 2022; Zhang et al., 2023). Although SQL is a code-based MR, it is much harder for models to learn compared to other MRs, such as FunQL and lambda calculus (Li et al., 2022). Consequently, without many SQL queries in their pre-training corpora, language models of code can underperform state-of-the-art text-to-SQL parsers. By converting SQL queries into Python dictionaries, we can explicitly represent their compositional substructures and define edit actions as programs, which reduces the learning difficulty for language models of code and yields better performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 508, + 474, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 508, + 474, + 521 + ], + "spans": [ + { + "bbox": [ + 302, + 508, + 474, + 521 + ], + "type": "text", + "content": "6 Conclusion and Future Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 529, + 526, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 529, + 526, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 529, + 526, + 773 + ], + "type": "text", + "content": "This paper presents a study on developing a text-to-SQL error correction model with clause-level edits and different representations. Our comprehensive experiments demonstrate that clauses are better semantic units than tokens for editing SQL queries and mimicking patterns in code pre-training corpora helps better leverage language models of code. As a future direction, we plan to incorporate our model into interactive semantic parsing frameworks (Li et al., 2020; Yao et al., 2019, 2020; Zeng et al., 2020) by suggesting possible edits to users once a wrong parse is identified. In this way, users would more efficiently correct parse errors and get better assistance. We also plan to experiment with other language models of code (Fried et al., 2023; Nijkamp et al., 2023) and text-to-SQL datasets (Zelle and Mooney, 1996; Gan et al., 2021a) to verify the generalizability of our method." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1363" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 71, + 131, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 71, + 131, + 84 + ], + "spans": [ + { + "bbox": [ + 68, + 71, + 131, + 84 + ], + "type": "text", + "content": "Limitations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 91, + 293, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 91, + 293, + 267 + ], + "spans": [ + { + "bbox": [ + 67, + 91, + 293, + 267 + ], + "type": "text", + "content": "Actual applications of our model. Our work assumes that input SQL queries to our model are always wrong. This assumption is more feasible in an interactive semantic parsing framework, where the users are expected to decide whether a SQL parse, accompanied by its natural language explanations (Elgohary et al., 2020, 2021; Narechinaia et al., 2021; Mo et al., 2022), has errors or not. Alternatively, to remove this assumption, it would be interesting for future work to study the performance of our error correction model in combination with an automatic error detection model (Chen et al., 2023)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 275, + 292, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 275, + 292, + 286 + ], + "spans": [ + { + "bbox": [ + 67, + 275, + 292, + 286 + ], + "type": "text", + "content": "Experiments with more language models of code." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 287, + 291, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 287, + 291, + 477 + ], + "spans": [ + { + "bbox": [ + 67, + 287, + 291, + 477 + ], + "type": "text", + "content": "We have only experimented with two language models of code, CodiT5 and CodeT5, both using T5-base (Raffel et al., 2020) as their underlying model architecture. It would be interesting to test how our conclusions generalize to other language models of code in the future. Based on the strong capabilities of large language models of code, such as Codex (Chen et al., 2021), InCoder (Fried et al., 2023), and CodeGen (Nijkamp et al., 2023), we believe that these models can better exploit their knowledge about data structures and their operations in Python. These models may perform even better on Text-to-SQL error correction with our proposed representations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 486, + 170, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 486, + 170, + 499 + ], + "spans": [ + { + "bbox": [ + 68, + 486, + 170, + 499 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 507, + 291, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 507, + 291, + 709 + ], + "spans": [ + { + "bbox": [ + 67, + 507, + 291, + 709 + ], + "type": "text", + "content": "We would like to thank the anonymous reviewers and colleagues from the OSU NLP group for their thoughtful comments. This research was supported in part by a sponsored award from Cisco Research, NSF IIS-1815674, NSF CAREER #1942980, NSF OAC-2112606, and Ohio Supercomputer Center (Center, 1987). The views and conclusions contained herein are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the U.S. government. The U.S. Government is authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notice herein. Ziru is also supported by The Ohio State University Graduate School through University Fellowship." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 731, + 127, + 743 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 731, + 127, + 743 + ], + "spans": [ + { + "bbox": [ + 68, + 731, + 127, + 743 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 68, + 750, + 291, + 773 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 750, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 68, + 750, + 291, + 773 + ], + "type": "text", + "content": "Wasi Ahmad, Saikat Chakraborty, Baishakhi Ray, and Kai-Wei Chang. 2021. Unified pre-training for pro" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 303, + 72, + 526, + 772 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 313, + 72, + 526, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 526, + 128 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 526, + 128 + ], + "type": "text", + "content": "gram understanding and generation. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2655-2668, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 303, + 136, + 525, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 136, + 525, + 158 + ], + "spans": [ + { + "bbox": [ + 303, + 136, + 525, + 158 + ], + "type": "text", + "content": "Ohio Supercomputer Center. 1987. Ohio supercomputer center." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 303, + 167, + 526, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 167, + 526, + 387 + ], + "spans": [ + { + "bbox": [ + 303, + 167, + 526, + 387 + ], + "type": "text", + "content": "Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidi Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott Gray, Nick Ryder, Mikhail Pavlov, Alethea Power, Lukasz Kaiser, Mohammad Bavarian, Clemens Winter, Philippe Tillet, Felipe Petroski Such, Dave Cummings, Matthias Plappert, Fotios Chantzis, Elizabeth Barnes, Ariel Herbert-Voss, William Hebgen Guss, Alex Nichol, Alex Paino, Nikolas Tezak, Jie Tang, Igor Babuschkin, Suchir Balaji, Shantanu Jain, William Saunders, Christopher Hesse, Andrew N. Carr, Jan Leike, Josh Achiam, Vedant Misra, Evan Morikawa, Alec Radford, Matthew Knight, Miles Brundage, Mira Murati, Katie Mayer, Peter Welinder, Bob McGrew, Dario Amodei, Sam McCandlish, Ilya Sutskever, and Wojciech Zaremba. 2021. Evaluating large language models trained on code." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 303, + 395, + 525, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 395, + 525, + 418 + ], + "spans": [ + { + "bbox": [ + 303, + 395, + 525, + 418 + ], + "type": "text", + "content": "Shijie Chen, Ziru Chen, Huan Sun, and Yu Su. 2023. Error detection for text-to-sql semantic parsing." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 303, + 427, + 525, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 427, + 525, + 515 + ], + "spans": [ + { + "bbox": [ + 303, + 427, + 525, + 515 + ], + "type": "text", + "content": "Xiang Deng, Ahmed Hassan Awadallah, Christopher Meek, Oleksandr Polozov, Huan Sun, and Matthew Richardson. 2021. Structure-grounded pretraining for text-to-SQL. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1337-1350, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 303, + 523, + 526, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 523, + 526, + 601 + ], + "spans": [ + { + "bbox": [ + 303, + 523, + 526, + 601 + ], + "type": "text", + "content": "Ahmed Elgohary, Saghar Hosseini, and Ahmed Hassan Awadallah. 2020. Speak to your parser: Interactive text-to-SQL with natural language feedback. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2065-2077, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 303, + 609, + 526, + 709 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 609, + 526, + 709 + ], + "spans": [ + { + "bbox": [ + 303, + 609, + 526, + 709 + ], + "type": "text", + "content": "Ahmed Elgohary, Christopher Meek, Matthew Richardson, Adam Fourney, Gonzalo Ramos, and Ahmed Hassan Awadallah. 2021. NL-EDIT: Correcting semantic parse errors through natural language interaction. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5599–5610, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 303, + 717, + 525, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 717, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 303, + 717, + 525, + 772 + ], + "type": "text", + "content": "Zhangyin Feng, Daya Guo, Duyu Tang, Nan Duan, Xiaocheng Feng, Ming Gong, Linjun Shou, Bing Qin, Ting Liu, Daxin Jiang, and Ming Zhou. 2020. CodeBERT: A pre-trained model for programming and natural languages. In Findings of the Association" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1364" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 79, + 72, + 290, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 72, + 290, + 105 + ], + "spans": [ + { + "bbox": [ + 79, + 72, + 290, + 105 + ], + "type": "text", + "content": "for Computational Linguistics: EMNLP 2020, pages 1536-1547, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 114, + 291, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 114, + 291, + 180 + ], + "spans": [ + { + "bbox": [ + 69, + 114, + 291, + 180 + ], + "type": "text", + "content": "Daniel Fried, Armen Aghajanyan, Jessy Lin, Sida Wang, Eric Wallace, Freda Shi, Ruiqi Zhong, Scott Yih, Luke Zettlemoyer, and Mike Lewis. 2023. Incoder: A generative model for code infilling and synthesis. In The Eleventh International Conference on Learning Representations." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 190, + 291, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 190, + 291, + 300 + ], + "spans": [ + { + "bbox": [ + 69, + 190, + 291, + 300 + ], + "type": "text", + "content": "Yujiang Gan, Xinyun Chen, Qiuping Huang, Matthew Purver, John R. Woodward, Jinxia Xie, and Pengsheng Huang. 2021a. Towards robustness of text-to-SQL models against synonym substitution. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 2505-2515, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 308, + 291, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 308, + 291, + 396 + ], + "spans": [ + { + "bbox": [ + 69, + 308, + 291, + 396 + ], + "type": "text", + "content": "Yujuan Gan, Xinyun Chen, Jinxia Xie, Matthew Purver, John R. Woodward, John Drake, and Qiaofu Zhang. 2021b. Natural SQL: Making SQL easier to infer from natural language specifications. In *Findings of the Association for Computational Linguistics: EMNLP* 2021, pages 2030–2042, Punta Cana, Dominican Republic. Association for Computational Linguistics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 405, + 291, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 405, + 291, + 483 + ], + "spans": [ + { + "bbox": [ + 69, + 405, + 291, + 483 + ], + "type": "text", + "content": "Daya Guo, Shuai Lu, Nan Duan, Yanlin Wang, Ming Zhou, and Jian Yin. 2022. UniXcoder: Unified cross-modal pre-training for code representation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7212-7225, Dublin, Ireland. Association for Computational Linguistics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 491, + 291, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 491, + 291, + 579 + ], + "spans": [ + { + "bbox": [ + 69, + 491, + 291, + 579 + ], + "type": "text", + "content": "Daya Guo, Shuo Ren, Shuai Lu, Zhangyin Feng, Duyu Tang, Shujie Liu, Long Zhou, Nan Duan, Alexey Svyatkovskiy, Shengyu Fu, Michele Tufano, Shao Kun Deng, Colin Clement, Dawn Drain, Neel Sundaresan, Jian Yin, Daxin Jiang, and Ming Zhou. 2021. GraphCodeBERT: Pre-training code representations with data flow. In International Conference on Learning Representations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 588, + 291, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 588, + 291, + 666 + ], + "spans": [ + { + "bbox": [ + 69, + 588, + 291, + 666 + ], + "type": "text", + "content": "Jiaqi Guo, Zecheng Zhan, Yan Gao, Yan Xiao, JianGuang Lou, Ting Liu, and Dongmei Zhang. 2019. Towards complex text-to-SQL in cross-domain database with intermediate representation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4524-4535, Florence, Italy. Association for Computational Linguistics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 675, + 290, + 718 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 675, + 290, + 718 + ], + "spans": [ + { + "bbox": [ + 69, + 675, + 290, + 718 + ], + "type": "text", + "content": "Hamel Husain, Ho-Hsiang Wu, Tiferet Gazit, Miltiadis Allamanis, and Marc Brockschmidt. 2020. Code-searchnet challenge: Evaluating the state of semantic code search." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 728, + 291, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 728, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 728, + 291, + 772 + ], + "type": "text", + "content": "Magne Jorgensen and Martin Shepperd. 2007. A systematic review of software development cost estimation studies. IEEE Transactions on Software Engineering, 33(1):33-53." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 526, + 772 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 305, + 72, + 526, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 526, + 160 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 526, + 160 + ], + "type": "text", + "content": "Yuntao Li, Bei Chen, Qian Liu, Yan Gao, Jian-Guang Lou, Yan Zhang, and Dongmei Zhang. 2020. \"what do you mean by that?\" a parser-independent interactive approach for enhancing text-to-SQL. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6913-6922, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 169, + 526, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 169, + 526, + 246 + ], + "spans": [ + { + "bbox": [ + 304, + 169, + 526, + 246 + ], + "type": "text", + "content": "Zhenwen Li, Jiaqi Guo, Qian Liu, Jian-Guang Lou, and Tao Xie. 2022. Exploring the secrets behind the learning difficulty of meaning representations for semantic parsing. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 3616-3625, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 255, + 526, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 526, + 322 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 526, + 322 + ], + "type": "text", + "content": "Xi Victoria Lin, Richard Socher, and Caiming Xiong. 2020. Bridging textual and tabular data for cross-domain text-to-SQL semantic parsing. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 4870-4888, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 330, + 526, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 330, + 526, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 330, + 526, + 407 + ], + "type": "text", + "content": "Aman Madaan, Shuyan Zhou, Uri Alon, Yiming Yang, and Graham Neubig. 2022. Language models of code are few-shot commonsense learners. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 1384-1403, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 417, + 526, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 417, + 526, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 417, + 526, + 460 + ], + "type": "text", + "content": "Quinn McNemar. 1947. Note on the sampling error of the difference between correlated proportions or percentages. In Psychometrika, volume 12, page 153-157." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 470, + 526, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 470, + 526, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 470, + 526, + 536 + ], + "type": "text", + "content": "Lingbo Mo, Ashley Lewis, Huan Sun, and Michael White. 2022. Towards transparent interactive semantic parsing via step-by-step correction. In *Findings of the Association for Computational Linguistics: ACL* 2022, pages 322-342, Dublin, Ireland. Association for Computational Linguistics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 544, + 526, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 526, + 600 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 526, + 600 + ], + "type": "text", + "content": "Marius Mosbach, Maksym Andriushchenko, and Dietrich Klakow. 2021. On the stability of fine-tuning BERT: Misconceptions, explanations, and strong baselines. In International Conference on Learning Representations." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 608, + 526, + 675 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 608, + 526, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 608, + 526, + 675 + ], + "type": "text", + "content": "Arpit Narechania, Adam Fourney, Bongshin Lee, and Gonzalo Ramos. 2021. Diy: Assessing the correctness of natural language to sql systems. In 26th International Conference on Intelligent User Interfaces, IUI '21, page 597-607, New York, NY, USA. Association for Computing Machinery." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 684, + 526, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 684, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 684, + 526, + 772 + ], + "type": "text", + "content": "Lunyiu Nie, Shulin Cao, Jiaxin Shi, Jiuding Sun, Qi Tian, Lei Hou, Juanzi Li, and Jidong Zhai. 2022. GraphQ IR: Unifying the semantic parsing of graph query languages with one intermediate representation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 5848-5865, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics." + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1365" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 289, + 772 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 69, + 72, + 289, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 289, + 138 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 289, + 138 + ], + "type": "text", + "content": "Erik Nijkamp, Bo Pang, Hiroaki Hayashi, Lifu Tu, Huan Wang, Yingbo Zhou, Silvio Savarese, and Caiming Xiong. 2023. Codegen: An open large language model for code with multi-turn program synthesis. In The Eleventh International Conference on Learning Representations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 150, + 289, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 150, + 289, + 227 + ], + "spans": [ + { + "bbox": [ + 69, + 150, + 289, + 227 + ], + "type": "text", + "content": "Sheena Panthaplackel, Milos Gligoric, Junyi Jessy Li, and Raymond Mooney. 2022. Using developer discussions to guide fixing bugs in software. In *Findings of the Association for Computational Linguistics: EMNLP* 2022, pages 2292-2301, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 239, + 289, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 239, + 289, + 349 + ], + "spans": [ + { + "bbox": [ + 69, + 239, + 289, + 349 + ], + "type": "text", + "content": "Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. 2019. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 361, + 289, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 361, + 289, + 427 + ], + "spans": [ + { + "bbox": [ + 69, + 361, + 289, + 427 + ], + "type": "text", + "content": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(140):1-67." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 439, + 289, + 516 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 439, + 289, + 516 + ], + "spans": [ + { + "bbox": [ + 69, + 439, + 289, + 516 + ], + "type": "text", + "content": "Ohad Rubin and Jonathan Berant. 2021. SmBoP: Semi-autoregressive bottom-up semantic parsing. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 311-324, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 528, + 289, + 616 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 528, + 289, + 616 + ], + "spans": [ + { + "bbox": [ + 69, + 528, + 289, + 616 + ], + "type": "text", + "content": "Torsten Scholak, Nathan Schucher, and Dzmitry Bahdanau. 2021. PICARD: Parsing incrementally for constrained auto-regressive decoding from language models. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 9895-9901, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 628, + 289, + 692 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 628, + 289, + 692 + ], + "spans": [ + { + "bbox": [ + 69, + 628, + 289, + 692 + ], + "type": "text", + "content": "Noam Shazeer and Mitchell Stern. 2018. Adafactor: Adaptive learning rates with sublinear memory cost. In Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pages 4596-4604. PMLR." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 706, + 289, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 706, + 289, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 706, + 289, + 772 + ], + "type": "text", + "content": "Ishika Singh, Valts Blukis, Arsalan Mousavian, Ankit Goyal, Danfei Xu, Jonathan Tremblay, Dieter Fox, Jesse Thomason, and Animesh Garg. 2022. Progress: Generating situated robot task plans using large language models. In Workshop on Language and Robotics at CoRL 2022." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 524, + 772 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 304, + 72, + 524, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 524, + 181 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 524, + 181 + ], + "type": "text", + "content": "Lappoon R. Tang and Raymond J. Mooney. 2000. Automated construction of database interfaces: Integrating statistical and relational learning for semantic parsing. In Proceedings of the 2000 Joint SIGDAT Conference on Empirical Methods in Natural Language Processing and Very Large Corpora: Held in Conjunction with the 38th Annual Meeting of the Association for Computational Linguistics - Volume 13, EMNLP '00, page 133-141, USA. Association for Computational Linguistics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 190, + 524, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 190, + 524, + 256 + ], + "spans": [ + { + "bbox": [ + 304, + 190, + 524, + 256 + ], + "type": "text", + "content": "Michele Tufano, Jevgenija Pantiuchina, Cody Watson, Gabriele Bavota, and Denys Poshyvanyk. 2019. On learning meaningful code changes via neural machine translation. In Proceedings of the 41st International Conference on Software Engineering, ICSE '19, page 25-36. IEEE Press." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 266, + 524, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 524, + 343 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 524, + 343 + ], + "type": "text", + "content": "Bailin Wang, Richard Shin, Xiaodong Liu, Oleksandr Polozov, and Matthew Richardson. 2020. RAT-SQL: Relation-aware schema encoding and linking for text-to-SQL parsers. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7567-7578, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 352, + 524, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 352, + 524, + 386 + ], + "spans": [ + { + "bbox": [ + 304, + 352, + 524, + 386 + ], + "type": "text", + "content": "Xingyao Wang, Sha Li, and Heng Ji. 2022. Code4struct: Code generation for few-shot structured prediction from natural language." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 393, + 524, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 393, + 524, + 481 + ], + "spans": [ + { + "bbox": [ + 304, + 393, + 524, + 481 + ], + "type": "text", + "content": "Yue Wang, Weishi Wang, Shafiq Joty, and Steven C.H. Hoi. 2021. CodeT5: Identifier-aware unified pretrained encoder-decoder models for code understanding and generation. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 8696-8708, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 491, + 524, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 524, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 524, + 545 + ], + "type": "text", + "content": "Cathrin Weiss, Rahul Premraj, Thomas Zimmermann, and Andreas Zeller. 2007. How long will it take to fix this bug? In Fourth International Workshop on Mining Software Repositories (MSR'07:ICSE Workshops 2007), pages 1-1." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 555, + 524, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 524, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 524, + 686 + ], + "type": "text", + "content": "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 695, + 524, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 695, + 524, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 695, + 524, + 772 + ], + "type": "text", + "content": "Tianbao Xie, Chen Henry Wu, Peng Shi, Ruiqi Zhong, Torsten Scholak, Michihiro Yasunaga, Chien-Sheng Wu, Ming Zhong, Pengcheng Yin, Sida I. Wang, Victor Zhong, Bailin Wang, Chengzu Li, Connor Boyle, Ansong Ni, Ziyu Yao, Dragomir Radev, Caiming Xiong, Lingpeng Kong, Rui Zhang, Noah A. Smith, Luke Zettlemoyer, and Tao Yu. 2022. UnifiedSKG:" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1366" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 78, + 72, + 291, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 72, + 291, + 139 + ], + "spans": [ + { + "bbox": [ + 78, + 72, + 291, + 139 + ], + "type": "text", + "content": "Unifying and multi-tasking structured knowledge grounding with text-to-text language models. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 602-631, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 149, + 291, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 149, + 291, + 250 + ], + "spans": [ + { + "bbox": [ + 68, + 149, + 291, + 250 + ], + "type": "text", + "content": "Ziyu Yao, Yu Su, Huan Sun, and Wen-tau Yih. 2019. Model-based interactive semantic parsing: A unified framework and a text-to-SQL case study. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5447-5458, Hong Kong, China. Association for Computational Linguistics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 68, + 261, + 291, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 261, + 291, + 340 + ], + "spans": [ + { + "bbox": [ + 68, + 261, + 291, + 340 + ], + "type": "text", + "content": "Ziyu Yao, Yiqi Tang, Wen-tau Yih, Huan Sun, and Yu Su. 2020. An imitation game for learning semantic parsers from user interaction. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6883-6902, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 349, + 291, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 349, + 291, + 417 + ], + "spans": [ + { + "bbox": [ + 68, + 349, + 291, + 417 + ], + "type": "text", + "content": "Pengcheng Yin and Graham Neubig. 2017. A syntactic neural model for general-purpose code generation. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 440-450, Vancouver, Canada. Association for Computational Linguistics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 428, + 291, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 428, + 291, + 529 + ], + "spans": [ + { + "bbox": [ + 68, + 428, + 291, + 529 + ], + "type": "text", + "content": "Tao Yu, Rui Zhang, Kai Yang, Michihiro Yasunaga, Dongxu Wang, Zifan Li, James Ma, Irene Li, Qingning Yao, Shanelle Roman, Zilin Zhang, and Dragomir Radev. 2018. Spider: A large-scale human-labeled dataset for complex and cross-domain semantic parsing and text-to-SQL task. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 3911-3921, Brussels, Belgium. Association for Computational Linguistics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 539, + 291, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 539, + 291, + 596 + ], + "spans": [ + { + "bbox": [ + 68, + 539, + 291, + 596 + ], + "type": "text", + "content": "John M. Zelle and Raymond J. Mooney. 1996. Learning to parse database queries using inductive logic programming. In Proceedings of the Thirteenth National Conference on Artificial Intelligence - Volume 2, AAAI'96, page 1050-1055. AAAI Press." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 606, + 291, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 606, + 291, + 684 + ], + "spans": [ + { + "bbox": [ + 68, + 606, + 291, + 684 + ], + "type": "text", + "content": "Jichuan Zeng, Xi Victoria Lin, Steven C.H. Hoi, Richard Socher, Caiming Xiong, Michael Lyu, and Irwin King. 2020. Photon: A robust cross-domain text-to-SQL system. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 204-214, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 68, + 694, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 694, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 68, + 694, + 291, + 773 + ], + "type": "text", + "content": "Jiyang Zhang, Sheena Panthapackel, Pengyu Nie, Junyi Jessy Li, and Milos Gligoric. 2023. Coditt5: Pretraining for source code and natural language editing. In Proceedings of the 37th IEEE/ACM International Conference on Automated Software Engineering, ASE '22, New York, NY, USA. Association for Computing Machinery." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 303, + 70, + 366, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 70, + 366, + 85 + ], + "spans": [ + { + "bbox": [ + 303, + 70, + 366, + 85 + ], + "type": "text", + "content": "Appendices" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 94, + 525, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 94, + 525, + 119 + ], + "spans": [ + { + "bbox": [ + 302, + 94, + 525, + 119 + ], + "type": "text", + "content": "We provide more details omitted in the main text as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 316, + 121, + 522, + 202 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 316, + 121, + 509, + 134 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 121, + 509, + 134 + ], + "spans": [ + { + "bbox": [ + 316, + 121, + 509, + 134 + ], + "type": "text", + "content": "- Appendix A: SQL PyDict Representation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 316, + 136, + 514, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 136, + 514, + 148 + ], + "spans": [ + { + "bbox": [ + 316, + 136, + 514, + 148 + ], + "type": "text", + "content": "- Appendix B: Text-to-SQL Parser Selection" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 149, + 489, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 149, + 489, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 149, + 489, + 161 + ], + "type": "text", + "content": "- Appendix C: Implementation Details" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 163, + 508, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 508, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 508, + 174 + ], + "type": "text", + "content": "- Appendix D: Statistical Significance Test" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 176, + 467, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 176, + 467, + 188 + ], + "spans": [ + { + "bbox": [ + 316, + 176, + 467, + 188 + ], + "type": "text", + "content": "- Appendix E: Additional Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 190, + 522, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 190, + 522, + 202 + ], + "spans": [ + { + "bbox": [ + 316, + 190, + 522, + 202 + ], + "type": "text", + "content": "- Appendix F: More Representation Examples" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 303, + 214, + 469, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 214, + 469, + 229 + ], + "spans": [ + { + "bbox": [ + 303, + 214, + 469, + 229 + ], + "type": "text", + "content": "A SQL PyDict Representation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 301, + 238, + 527, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 238, + 527, + 604 + ], + "spans": [ + { + "bbox": [ + 301, + 238, + 527, + 604 + ], + "type": "text", + "content": "We implement the transformation from any SQL query to our PyDict representation in three steps (Section 2.1). First, we use context-free grammar to parse a SQL query and obtain its abstract syntax tree (AST). The AST naturally contains a SQL decomposition where each clause has its unique subtree. In addition, if a clause contains a nested query, it would be represented as another independent subtree, which is a child of the root node in the clause's AST subtree. With these substructures explicitly represented, we use depth-first search to traverse through the AST to build our PyDict representation bottom-up. In other words, if a clause contains a subquery, we process the subquery tree as an independent SQL AST and build a dictionary for it. Then, we combine it with other substructures of the clause with different dictionary keys. For example, in Table F.1, we first build the dictionary for \"subquery0\" and assign this identifier as the key. In the main \"clause,\" we replace the subquery's corresponding span with this identifier. Finally, we use another dictionary to wrap the main \"clause\" and \"subquery0\" together as the final representation of the \"where\" clause. We repeat this procedure for each clause to incrementally add (key, value) pairs to the dictionary and \"store\" it to the variable sql, which we refer to in program edit representations." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 303, + 614, + 477, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 614, + 477, + 629 + ], + "spans": [ + { + "bbox": [ + 303, + 614, + 477, + 629 + ], + "type": "text", + "content": "B Text-to-SQL Parser Selection" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 302, + 638, + 526, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 638, + 526, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 638, + 526, + 773 + ], + "type": "text", + "content": "We choose existing text-to-SQL parsers in our experiments according to two principles: the parsers predict database entity values, and they cover different decoding strategies, including grammar-based (BRIDGEv2), bottom-up (SmBop), and token-based (CodeT5). We did not include parsers using top-down decoders because they usually cannot predict entity values in conditional statements, such as RAT-SQL (Wang et al., 2020). Instead, we include BRIDGEv2 because its decoding method mimics" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "1367" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 291, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 291, + 139 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 291, + 139 + ], + "type": "text", + "content": "the left-to-right CFG derivation of a program, and it uses SQL syntax-based constraints to prevent grammatical errors. In recent work, such decoders, also used in PICARD (Scholak et al., 2021), are more popular than top-down decoders." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 151, + 212, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 151, + 212, + 164 + ], + "spans": [ + { + "bbox": [ + 67, + 151, + 212, + 164 + ], + "type": "text", + "content": "C Implementation Details" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 174, + 291, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 174, + 291, + 267 + ], + "spans": [ + { + "bbox": [ + 67, + 174, + 291, + 267 + ], + "type": "text", + "content": "Our models (Section 3.2) are implemented in PyTorch (Paszke et al., 2019) using Huggingface (Wolf et al., 2020) and trained on a single NVIDIA RTX A6000 GPU (48GB). We use Adafactor (Shazeer and Stern, 2018) to train all our models with the same hyperparameters adapted from Mosbach et al. (2021):" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 81, + 269, + 282, + 323 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 81, + 269, + 187, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 269, + 187, + 281 + ], + "spans": [ + { + "bbox": [ + 81, + 269, + 187, + 281 + ], + "type": "text", + "content": "- Learning rate: " + }, + { + "bbox": [ + 81, + 269, + 187, + 281 + ], + "type": "inline_equation", + "content": "3e - 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 81, + 283, + 156, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 283, + 156, + 295 + ], + "spans": [ + { + "bbox": [ + 81, + 283, + 156, + 295 + ], + "type": "text", + "content": "- Batch size: 16" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 81, + 297, + 143, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 297, + 143, + 308 + ], + "spans": [ + { + "bbox": [ + 81, + 297, + 143, + 308 + ], + "type": "text", + "content": "- Epochs: 10" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 81, + 310, + 282, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 310, + 282, + 323 + ], + "spans": [ + { + "bbox": [ + 81, + 310, + 282, + 323 + ], + "type": "text", + "content": "- Scheduler: Linear decay with " + }, + { + "bbox": [ + 81, + 310, + 282, + 323 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 81, + 310, + 282, + 323 + ], + "type": "text", + "content": " warmup" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 67, + 335, + 230, + 349 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 335, + 230, + 349 + ], + "spans": [ + { + "bbox": [ + 67, + 335, + 230, + 349 + ], + "type": "text", + "content": "D Statistical Significance Test" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 359, + 291, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 359, + 291, + 521 + ], + "spans": [ + { + "bbox": [ + 67, + 359, + 291, + 521 + ], + "type": "text", + "content": "To demonstrate the effectiveness of our three clause-level edit representations (Section 4.1), we perform McNemar's Test (McNemar, 1947) to measure the statistical significance of their results in comparison to CodeT5-SQL+Token-Level. For each significance test between two models, we use the median results among our three runs to calculate the comparison matrix. Then, we compute the " + }, + { + "bbox": [ + 67, + 359, + 291, + 521 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 67, + 359, + 291, + 521 + ], + "type": "text", + "content": "-values using statsmodels. When " + }, + { + "bbox": [ + 67, + 359, + 291, + 521 + ], + "type": "inline_equation", + "content": "p < 0.05" + }, + { + "bbox": [ + 67, + 359, + 291, + 521 + ], + "type": "text", + "content": " we reject the null hypothesis. In other words, we consider the accuracy improvement statistically significant when " + }, + { + "bbox": [ + 67, + 359, + 291, + 521 + ], + "type": "inline_equation", + "content": "p < 0.05" + }, + { + "bbox": [ + 67, + 359, + 291, + 521 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 533, + 186, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 533, + 186, + 545 + ], + "spans": [ + { + "bbox": [ + 67, + 533, + 186, + 545 + ], + "type": "text", + "content": "E Additional Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 556, + 291, + 731 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 556, + 291, + 731 + ], + "spans": [ + { + "bbox": [ + 67, + 556, + 291, + 731 + ], + "type": "text", + "content": "Results on our development set. We report model performances on our held-out development set (Section 3.1) in Table E.1. During training, we select the best model by evaluating its EX and EM accuracy on the development set (Section 3.3) every 500 steps. Surprisingly, we find that CodeT5-SQL+Clause-Level sometimes achieves the best performance. For BRIDGEv2, it obtains 35.9 EM accuracy and 39.3 EX accuracy, while CodeT5-PyDict+Program only obtains 34.5 EM accuracy and 37.1 EX accuracy. A possible explanation is that in comparison to the test set, our development set has SQL structures and databases that are more" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 71, + 526, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 138 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 138 + ], + "type": "text", + "content": "similar to the training set, while the test set has unseen SQL structures and less similar databases. It may also indicate that CodeT5-SQL+Clause-Level overfits the synthetic training data and fails to generalize to realistic test data." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 146, + 526, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 146, + 526, + 266 + ], + "spans": [ + { + "bbox": [ + 302, + 146, + 526, + 266 + ], + "type": "text", + "content": "Results for simulated interaction experiments. To show the potential of using our model in an interactive framework, we extend our main experiments (Section 4.1) by adding simulated user interactions. Since our model uses beam search to decode the edit actions " + }, + { + "bbox": [ + 302, + 146, + 526, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{e} = \\{e_1,e_2,\\dots ,e_n\\}" + }, + { + "bbox": [ + 302, + 146, + 526, + 266 + ], + "type": "text", + "content": " and the resulting correct SQL query " + }, + { + "bbox": [ + 302, + 146, + 526, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_{+}" + }, + { + "bbox": [ + 302, + 146, + 526, + 266 + ], + "type": "text", + "content": " (Equation 1), we simulate user interactions to select one edit action " + }, + { + "bbox": [ + 302, + 146, + 526, + 266 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 302, + 146, + 526, + 266 + ], + "type": "text", + "content": " at a time from the beam results." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 268, + 526, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 268, + 526, + 511 + ], + "spans": [ + { + "bbox": [ + 302, + 268, + 526, + 511 + ], + "type": "text", + "content": "At each time step " + }, + { + "bbox": [ + 302, + 268, + 526, + 511 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 302, + 268, + 526, + 511 + ], + "type": "text", + "content": ", we prompt the decoder with previously selected edit actions " + }, + { + "bbox": [ + 302, + 268, + 526, + 511 + ], + "type": "inline_equation", + "content": "e_1, \\ldots, e_{t-1}" + }, + { + "bbox": [ + 302, + 268, + 526, + 511 + ], + "type": "text", + "content": " to complete the sequence " + }, + { + "bbox": [ + 302, + 268, + 526, + 511 + ], + "type": "inline_equation", + "content": "e_t, \\ldots, e_n" + }, + { + "bbox": [ + 302, + 268, + 526, + 511 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 302, + 268, + 526, + 511 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_+" + }, + { + "bbox": [ + 302, + 268, + 526, + 511 + ], + "type": "text", + "content": " using beam search with size 3. Then, we use gold SQL annotations to simulate the user interaction, which selects an edit action " + }, + { + "bbox": [ + 302, + 268, + 526, + 511 + ], + "type": "inline_equation", + "content": "e_t" + }, + { + "bbox": [ + 302, + 268, + 526, + 511 + ], + "type": "text", + "content": " from the three candidates at step " + }, + { + "bbox": [ + 302, + 268, + 526, + 511 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 302, + 268, + 526, + 511 + ], + "type": "text", + "content": " or chooses to skip the current step when all three candidates are wrong. If skipping, the user continues to check the consequent edit actions " + }, + { + "bbox": [ + 302, + 268, + 526, + 511 + ], + "type": "inline_equation", + "content": "e_{t+j}" + }, + { + "bbox": [ + 302, + 268, + 526, + 511 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 302, + 268, + 526, + 511 + ], + "type": "inline_equation", + "content": "j = 1, 2, \\ldots, n-t" + }, + { + "bbox": [ + 302, + 268, + 526, + 511 + ], + "type": "text", + "content": ") until it selects the next edit action. When the interaction finishes, we append the selected edit action to the prompt and let the model regenerate a completion with the new prompt for the next step's interaction. Having simulated interactions for all edit actions, we do not use the generated " + }, + { + "bbox": [ + 302, + 268, + 526, + 511 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_+" + }, + { + "bbox": [ + 302, + 268, + 526, + 511 + ], + "type": "text", + "content": " directly because some edit actions are skipped. Instead, we execute the selected ones on the initial SQL query to derive the final query." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 512, + 527, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 512, + 527, + 674 + ], + "spans": [ + { + "bbox": [ + 302, + 512, + 527, + 674 + ], + "type": "text", + "content": "As shown in Table E.2, when collaborating with a simulated user, our error correction model can further improve the base parsers' accuracy. Compared to its performance without using any interactions, our model achieves up to 4.1 point more absolute improvement on EM accuracy (72.5 → 76.6; BRIDGEv2) and 5.0 point more absolute improvement on EX accuracy (73.1 → 78.1; BRIDGEv2). With these results for simulated interaction experiments, we deem that incorporating our error correction model into an interactive framework is a promising future direction." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 740, + 275, + 771 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 740, + 275, + 771 + ], + "spans": [ + { + "bbox": [ + 67, + 740, + 275, + 771 + ], + "type": "text", + "content": "4https://www.statsmodels.org/dev/generated/ statsmodels.stats.contingency_tables.mcnemar. html" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1368" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 68, + 511, + 200 + ], + "blocks": [ + { + "bbox": [ + 83, + 68, + 511, + 200 + ], + "lines": [ + { + "bbox": [ + 83, + 68, + 511, + 200 + ], + "spans": [ + { + "bbox": [ + 83, + 68, + 511, + 200 + ], + "type": "table", + "html": "
ModelsQueryEditCodeT5BRIDGEv2SmBoP
EMEXEMEXEMEX
CodiT5SQLToken-Level26.1 (0.4)28.6 (1.0)25.8 (0.3)27.2 (0.6)28.1 (0.9)30.7 (0.7)
SQLClause-Level28.6 (0.4)31.3 (0.5)28.4 (0.5)30.0 (0.2)30.2 (0.8)33.4 (0.8)
PyDictClause-Level28.9 (0.6)32.3 (0.8)28.0 (0.1)30.1 (0.2)27.6 (0.1)30.9 (0.4)
CodeT5SQLToken-Level32.1 (1.1)34.1 (1.2)31.8 (0.4)34.5 (0.8)34.2 (0.1)37.6 (0.1)
SQLClause-Level36.5 (0.6)38.6 (0.5)35.9 (0.4)39.3 (1.3)36.1 (0.6)38.8 (0.5)
PyDictClause-Level35.6 (0.9)37.9 (0.3)32.9 (1.0)34.8 (0.8)33.0 (0.2)36.3 (0.3)
CodeT5* CodeT5PyDictProgram35.7 (0.8)37.9 (0.3)34.8 (0.8)38.3 (0.7)36.0 (0.3)40.2 (0.5)
36.7 (0.2)38.5 (0.6)34.5 (0.1)37.1 (0.2)35.6 (0.8)39.0 (0.1)
", + "image_path": "9549c264b7e1e50603a0111fbdca78a8cf938b913e81b8b871f2aca463abdbc5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 91, + 255, + 503, + 349 + ], + "blocks": [ + { + "bbox": [ + 67, + 208, + 525, + 246 + ], + "lines": [ + { + "bbox": [ + 67, + 208, + 525, + 246 + ], + "spans": [ + { + "bbox": [ + 67, + 208, + 525, + 246 + ], + "type": "text", + "content": "Table E.1: Exact Set Match (EM) and Execution Match (EX) accuracy on our held-out development set (Section 3.1). The best performances are in bold and the second bests are underlined. *We fine-tune the model to generate edit programs only (without resulting queries) and use Python interpreter to execute the edit actions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 91, + 255, + 503, + 349 + ], + "lines": [ + { + "bbox": [ + 91, + 255, + 503, + 349 + ], + "spans": [ + { + "bbox": [ + 91, + 255, + 503, + 349 + ], + "type": "table", + "html": "
ModelsQueryEditCodeT5BRIDGEv2SmBoP
EMEXEMEXEMEX
No EditN/AN/A62.7 (-)63.6 (-)70.1 (-)68.2 (-)74.6 (-)75.3 (-)
CodeT5*69.2 (0.4)68.4 (0.2)72.5 (0.4)73.1 (0.2)77.3 (0.4)77.6 (0.6)
CodeT5PyDictProgram69.0 (0.2)68.2 (0.1)72.5 (0.3)73.0 (0.6)78.0 (0.3)78.5 (0.3)
\\(CodeT5^{\\dagger}\\)PyDictProgram73.0 (0.7)72.9 (0.8)76.6 (0.4)78.1 (0.2)80.0 (0.3)81.2 (0.6)
", + "image_path": "e2a4e56d20ed6972a1fce9e25790764a38d4c71c68877d0ceedef236f3e5fa83.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 356, + 525, + 405 + ], + "lines": [ + { + "bbox": [ + 67, + 356, + 525, + 405 + ], + "spans": [ + { + "bbox": [ + 67, + 356, + 525, + 405 + ], + "type": "text", + "content": "Table E.2: Exact Set Match (EM) and Execution Match (EX) accuracy on Spider development set. The best performances are in bold. *We fine-tune the model to generate edit programs only (without resulting queries) and use Python interpreter to execute the edit actions. †We simulate user interactions using gold SQL queries to choose edit actions during beam search (size 3) and then execute the chosen actions to get the resulting SQL parse." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 425, + 252, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 425, + 252, + 439 + ], + "spans": [ + { + "bbox": [ + 67, + 425, + 252, + 439 + ], + "type": "text", + "content": "F More Representation Examples" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 446, + 290, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 446, + 290, + 513 + ], + "spans": [ + { + "bbox": [ + 67, + 446, + 290, + 513 + ], + "type": "text", + "content": "We provide two more examples in Table F.1 and F.2 to demonstrate how we represent SQL with subqueries and their edits (Section 2.2). We also show different representations for Insert and Delete edit actions." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1369" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 108, + 526, + 358 + ], + "blocks": [ + { + "bbox": [ + 69, + 108, + 526, + 358 + ], + "lines": [ + { + "bbox": [ + 69, + 108, + 526, + 358 + ], + "spans": [ + { + "bbox": [ + 69, + 108, + 526, + 358 + ], + "type": "table", + "html": "
Query RepresentationEdit Representation
SQLselect count(*) from cars_data where cars_data.accelerate > ( select max(cars_data.horsepower) from cars_data )Token-level<ReplaceOld> max(cars_data.horsepower) <ReplaceNew> cars_data.accelerate <ReplaceEnd> <Insert> order by cars_data.horsepower desc limit 1 <InsertEnd> Clause-level <ReplaceOld> select max(cars_data.horsepower) <ReplaceNew> select cars_data.accelerate <ReplaceEnd> <Insert> order by cars_data.horsepower desc limit 1 <InsertEnd>
PyDictsql = { "select": "select count(*)", "from": "from cars_data", "where": { "clause": "where cars_data.accelerate > (subquery0)", "subquery0": { "select": "select max(cars_data.horsepower)", "from": "from cars_data" } } }Clause-level<ReplaceOld> "select": "select max( cars_data.horsepower)" <ReplaceNew> "select": "select cars_data.accelerate" <ReplaceEnd> <Insert> "orderBy": "order by cars_data.horsepower desc", "limit": "limit 1" <InsertEnd> Programsql["where"},{"subquery0"},{"select"} = "select cars_data.accelerate" sql["where"},{"subquery0"},{"orderBy"} = "order by cars_data.horsepower desc" sql["where"},{"subquery0"},{"limit"} = "limit 1"
", + "image_path": "c3604543de20f82a0e977c1f24315ce2790444838a2680660d4d4995647d42a4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 70, + 487, + 525, + 698 + ], + "blocks": [ + { + "bbox": [ + 67, + 365, + 525, + 402 + ], + "lines": [ + { + "bbox": [ + 67, + 365, + 525, + 402 + ], + "spans": [ + { + "bbox": [ + 67, + 365, + 525, + 402 + ], + "type": "text", + "content": "Table F.1: Example representations for a wrong SQL query that contains a nested subquery and its edit actions (including Insert edits). The corresponding natural language utterance is \"What is the number of cars with a greater acceleration than the one with the most horsepower?\"" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 487, + 525, + 698 + ], + "lines": [ + { + "bbox": [ + 70, + 487, + 525, + 698 + ], + "spans": [ + { + "bbox": [ + 70, + 487, + 525, + 698 + ], + "type": "table", + "html": "
Query RepresentationEdit Representation
SQLselect employee.name from employee join evaluation on employee.employee_id = evaluation.employee_id group by evaluation.employee_id" order by sum(evaluationbonus) desc limit 1Token-level<Delete> group by evaluation.employee_id <DeleteEnd> <DeleteSum( <DeleteEnd><Delete>) <DeleteEnd>
Clause-level<Delete> group by evaluation.employee_id <DeleteEnd> <ReplaceOld> order by sum(evaluation;bONUS) desc <ReplaceNew> order by evaluation;bONUS desc <ReplaceEnd>
PyDictsql = { "select": "select employee.name", "from": "from employee join evaluation on employee.employee_id = evaluation.employee_id", "groupBy": "group by evaluation.employee_id", "orderBy": "order by sum(evaluation;bONUS) desc", "limit": "limit 1" }Clause-level<Delete> "groupId": "group by evaluation.employee_id" <DeleteEnd><ReplaceOld> "orderBy": "order by sum(evaluation;bONUS) desc" <ReplaceNew> "orderBy": "order by evaluation;bONUS desc" <ReplaceEnd>
Programsql.pop("groupId") sql["orderBy"] = "order by evaluation;bONUS desc"
", + "image_path": "d7e577ba3a90351d3d9c74bd227b07af3f0ce94a67ed638680ac7e44342462e1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 705, + 525, + 730 + ], + "lines": [ + { + "bbox": [ + 67, + 705, + 525, + 730 + ], + "spans": [ + { + "bbox": [ + 67, + 705, + 525, + 730 + ], + "type": "text", + "content": "Table F.2: Example representations for a wrong SQL query and its edit actions (including Delete edits). The corresponding natural language utterance is \"Find the name of the employee who got the highest one time bonus.\"" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 780, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 780, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 780, + 310, + 791 + ], + "type": "text", + "content": "1370" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "spans": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "type": "text", + "content": "A For every submission:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 76, + 106, + 414, + 242 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "spans": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "type": "text", + "content": "A1. Did you describe the limitations of your work? 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 76, + 143, + 329, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 143, + 329, + 169 + ], + "spans": [ + { + "bbox": [ + 76, + 143, + 329, + 169 + ], + "type": "text", + "content": "A2. Did you discuss any potential risks of your work? Not applicable. Left blank." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 178, + 414, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 178, + 414, + 205 + ], + "spans": [ + { + "bbox": [ + 77, + 178, + 414, + 205 + ], + "type": "text", + "content": "A3. Do the abstract and introduction summarize the paper's main claims?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 77, + 215, + 398, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 215, + 398, + 242 + ], + "spans": [ + { + "bbox": [ + 77, + 215, + 398, + 242 + ], + "type": "text", + "content": "A4. Have you used AI writing assistants when working on this paper? Left blank." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 68, + 251, + 291, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 251, + 291, + 266 + ], + "spans": [ + { + "bbox": [ + 68, + 251, + 291, + 266 + ], + "type": "text", + "content": "B Did you use or create scientific artifacts?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 80, + 270, + 87, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 270, + 87, + 280 + ], + "spans": [ + { + "bbox": [ + 80, + 270, + 87, + 280 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 76, + 291, + 524, + 633 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 77, + 291, + 315, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 291, + 315, + 317 + ], + "spans": [ + { + "bbox": [ + 77, + 291, + 315, + 317 + ], + "type": "text", + "content": "B1. Did you cite the creators of artifacts you used? 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 76, + 328, + 463, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 328, + 463, + 355 + ], + "spans": [ + { + "bbox": [ + 76, + 328, + 463, + 355 + ], + "type": "text", + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Not applicable. Left blank." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 76, + 364, + 524, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 364, + 524, + 431 + ], + "spans": [ + { + "bbox": [ + 76, + 364, + 524, + 431 + ], + "type": "text", + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Not applicable. Left blank." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 76, + 441, + 524, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 441, + 524, + 494 + ], + "spans": [ + { + "bbox": [ + 76, + 441, + 524, + 494 + ], + "type": "text", + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Not applicable. Left blank." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 77, + 503, + 524, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 503, + 524, + 542 + ], + "spans": [ + { + "bbox": [ + 77, + 503, + 524, + 542 + ], + "type": "text", + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 77, + 553, + 524, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 553, + 524, + 633 + ], + "spans": [ + { + "bbox": [ + 77, + 553, + 524, + 633 + ], + "type": "text", + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "spans": [ + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "type": "text", + "content": "C Did you run computational experiments?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 80, + 663, + 87, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 663, + 87, + 672 + ], + "spans": [ + { + "bbox": [ + 80, + 663, + 87, + 672 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 77, + 682, + 524, + 724 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 682, + 524, + 724 + ], + "spans": [ + { + "bbox": [ + 77, + 682, + 524, + 724 + ], + "type": "text", + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Appendix B" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "spans": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "text", + "content": "ACL 2023 Responsible NLP Checklist" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "spans": [ + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "type": "text", + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 287, + 781, + 308, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 308, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 308, + 791 + ], + "type": "text", + "content": "1371" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 77, + 70, + 523, + 238 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 77, + 70, + 523, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 70, + 523, + 111 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 523, + 111 + ], + "type": "text", + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Appendix B" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 77, + 120, + 523, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 120, + 523, + 172 + ], + "spans": [ + { + "bbox": [ + 77, + 120, + 523, + 172 + ], + "type": "text", + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 183, + 523, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 183, + 523, + 238 + ], + "spans": [ + { + "bbox": [ + 77, + 183, + 523, + 238 + ], + "type": "text", + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Appendix B" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 67, + 247, + 522, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 247, + 522, + 278 + ], + "spans": [ + { + "bbox": [ + 67, + 247, + 522, + 278 + ], + "type": "text", + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 76, + 286, + 524, + 539 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 76, + 286, + 524, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 286, + 524, + 326 + ], + "spans": [ + { + "bbox": [ + 76, + 286, + 524, + 326 + ], + "type": "text", + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 76, + 336, + 524, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 336, + 524, + 390 + ], + "spans": [ + { + "bbox": [ + 76, + 336, + 524, + 390 + ], + "type": "text", + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 76, + 400, + 524, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 400, + 524, + 454 + ], + "spans": [ + { + "bbox": [ + 76, + 400, + 524, + 454 + ], + "type": "text", + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 76, + 462, + 519, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 462, + 519, + 489 + ], + "spans": [ + { + "bbox": [ + 76, + 462, + 519, + 489 + ], + "type": "text", + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 76, + 498, + 523, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 498, + 523, + 539 + ], + "spans": [ + { + "bbox": [ + 76, + 498, + 523, + 539 + ], + "type": "text", + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 780, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 780, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 780, + 310, + 791 + ], + "type": "text", + "content": "1372" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 13 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/The Art of Prompting_ Event Detection based on Type Specific Prompts/5e1d8751-6f7a-44c9-99a0-fd4f3669ec72_content_list.json b/2023/The Art of Prompting_ Event Detection based on Type Specific Prompts/5e1d8751-6f7a-44c9-99a0-fd4f3669ec72_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4959a5f9235b544758c414348cc9d853b3b8ee67 --- /dev/null +++ b/2023/The Art of Prompting_ Event Detection based on Type Specific Prompts/5e1d8751-6f7a-44c9-99a0-fd4f3669ec72_content_list.json @@ -0,0 +1,1609 @@ +[ + { + "type": "text", + "text": "The Art of Prompting: Event Detection based on Type Specific Prompts", + "text_level": 1, + "bbox": [ + 127, + 89, + 870, + 111 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sijia Wang*, Mo Yu*, Lifu Huang*", + "bbox": [ + 337, + 140, + 660, + 158 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$\\spadesuit$ Virginia Tech, $\\spadesuit$ WeChat AI", + "bbox": [ + 376, + 158, + 623, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$\\clubsuit$ {sijiawang, lifuh}@vt.edu, $\\clubsuit$ moyumyu@tencent.com", + "bbox": [ + 216, + 174, + 784, + 192 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 252, + 339, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We compare various forms of prompts to represent event types and develop a unified framework to incorporate the event type specific prompts for supervised, few-shot, and zero-shot event detection. The experimental results demonstrate that a well-defined and comprehensive event type prompt can significantly improve event detection performance, especially when the annotated data is scarce (few-shot event detection) or not available (zero-shot event detection). By leveraging the semantics of event types, our unified framework shows up to $22.2\\%$ F-score gain over the previous state-of-the-art baselines1.", + "bbox": [ + 141, + 280, + 460, + 478 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 114, + 492, + 258, + 507 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Event detection (ED) (Grishman, 1997; Chinchor and Marsh, 1998; Ahn, 2006) is the task of identifying and typing event mentions from natural language text. Supervised approaches, especially deep neural networks (Chen et al., 2020; Du and Cardie, 2020; Lin et al., 2020; Liu et al., 2020; Li et al., 2020; Lyu et al., 2021), have shown remarkable performance under a critical prerequisite of a large amount of manual annotations. However, they cannot be effectively generalized to new languages, domains or types, especially when the annotations are not enough (Huang et al., 2016; Huang and Ji, 2020; Lai et al., 2020b; Shen et al., 2021) or there is no annotation available (Lyu et al., 2021; Zhang et al., 2021b; Pasupat and Liang, 2014).", + "bbox": [ + 112, + 518, + 489, + 759 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent studies have shown that both the accuracy and generalizability of ED can be improved via leveraging the semantics of event types based on various forms of prompts, such as event type specific queries (Lyu et al., 2021; Du and Cardie, 2020; Liu et al., 2020), definitions (Chen et al., 2020), structures (Lin et al., 2020; Wang et al.,", + "bbox": [ + 112, + 760, + 489, + 872 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2019), or a few prototype event triggers (Wang and Cohen, 2009; Dalvi et al., 2012; Pasupat and Liang, 2014; Bronstein et al., 2015; Lai and Nguyen, 2019; Zhang et al., 2021b; Cong et al., 2021). These studies further encourage us to take another step forward and think about the following three questions: (1) does the choice of prompt matter when the training data is abundant or scarce? (2) what's the best form of ED prompt? (3) how to best leverage the prompt to detect event mentions?", + "bbox": [ + 507, + 253, + 884, + 413 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To answer the above research questions, we conduct extensive experiments with various forms of prompts for each event type, including (a) event type name, (b) prototype seed triggers, (c) definition, (d) event type structure based on both event type name and its predefined argument roles, (e) free parameter based continuous soft prompt, and (f) a more comprehensive event type description (named APEX prompt) that covers all the information of prompts (a)-(d). We observe that (1) by considering the semantics of event types with most forms of prompts, especially seed triggers and the comprehensive event type descriptions, the performance of ED under all settings can be significantly improved; (2) Among all forms of event representations, the comprehensive description based prompts show to be the most effective, especially for few-shot and zero-shot ED; (3) Different forms of event type representations provide complementary improvements, indicating that they capture distinct aspects and knowledge of the event types.", + "bbox": [ + 507, + 414, + 884, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The contributions of this work are as follows:", + "bbox": [ + 526, + 756, + 865, + 770 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "- We investigate various prompts to represent event types for both supervised and weakly supervised ED, and prove that a well-defined and comprehensive event type prompt can dramatically improve the performance of ED and the transferability from old types to new types.", + "bbox": [ + 507, + 774, + 882, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "- A unified framework is developed to leverage the semantics of event types with prompts for supervised, few-shot, and zero-shot ED, and demonstrate", + "bbox": [ + 507, + 871, + 882, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1The source code, model checkpoints and data are publicly available at https://github.com/VT-NLP/Event_APEX.", + "bbox": [ + 112, + 879, + 487, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1286", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics", + "bbox": [ + 226, + 945, + 769, + 957 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Volume 2: Short Papers, pages 1286-1299", + "bbox": [ + 368, + 958, + 628, + 971 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "July 9-14, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 295, + 972, + 700, + 985 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "state-of-the-art performance with up to $22.2\\%$ F-score improvement over the strong baseline methods.", + "bbox": [ + 112, + 84, + 489, + 131 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 114, + 147, + 268, + 162 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Supervised ED: Most of the existing Event Detection studies follow a supervised learning paradigm (Ji and Grishman, 2008; Liao and Grishman, 2010; McClosky et al., 2011; Li et al., 2013; Chen et al., 2015; Cao et al., 2015; Feng et al., 2016; Yang and Mitchell, 2016; Nguyen et al., 2016; Zhang et al., 2017; Lin et al., 2020; Wang et al., 2021b). However, they cannot be directly applied to detect new types of events. Recently studies have shown that, by leveraging the semantics of event types based on type-specific questions (Du and Cardie, 2020; Liu et al., 2020; Li et al., 2020; Lyu et al., 2021) or seed event triggers (Bronstein et al., 2015; Lai and Nguyen, 2019; Wang et al., 2021a), the event detection performance can be improved. However, it is still unknown whether they are the best choices for representing the semantics of event types.", + "bbox": [ + 112, + 175, + 489, + 464 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Few-shot ED: Two primary learning strategies in few-shot classification tasks are Meta-Learning (Kang et al., 2019; Li et al., 2021; Xiao and Marlet, 2020; Yan et al., 2019; Chowdhury et al., 2021) and Metric Learning (Sun et al., 2021; Wang et al., 2020b; Zhang et al., 2021a; Agarwal et al., 2021). Several studies have exploited metric learning to align the semantics of candidate events with a few examples of the novel event types for few-shot event detection (Lai et al., 2020a; Deng et al., 2020; Lai et al., 2020b; Cong et al., 2021; Chen et al., 2021; Shen et al., 2021).", + "bbox": [ + 112, + 476, + 489, + 670 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Zero-shot ED: Huang et al. (2018) first exploited zero-shot event extraction by leveraging Abstract Meaning Representation (Banarescu et al., 2013) to represent event mentions and types into a shared semantic space. Recent studies (Zhang et al., 2021b; Lyu et al., 2021) further demonstrate that by leveraging a large external corpus with abundant anchor triggers, zero-shot event detection can also be achieved with decent performance without using any training data.", + "bbox": [ + 112, + 682, + 489, + 843 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prompt Learning Prompt learning aims to learn a task-specific prompt while keeping most of the model's parameters frozen (Li and Liang, 2021; Hambardzumyan et al., 2021; Brown et al., 2020).", + "bbox": [ + 112, + 854, + 489, + 919 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "It has shown competitive performance in many applications of natural language processing (Raffel et al., 2020; Brown et al., 2020; Shin et al., 2020; Jiang et al., 2020; Lester et al., 2021; Schick and Schütze, 2021b). Previous work either used a manual (Petroni et al., 2019; Brown et al., 2020; Schick and Schütze, 2021a) or automated approach (Jiang et al., 2020; Yuan et al., 2021; Li and Liang, 2021) to create prompts.", + "bbox": [ + 507, + 84, + 884, + 229 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3 Problem Formulation", + "text_level": 1, + "bbox": [ + 507, + 242, + 732, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Here, we first define each setting of the event detection task and then describe the various forms of event type prompts.", + "bbox": [ + 507, + 269, + 884, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1 Settings of ED", + "text_level": 1, + "bbox": [ + 507, + 330, + 668, + 346 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For supervised ED (SED), we follow the conventional supervised event detection setting where the training, validation, and evaluation data sets cover the same set of event types. The goal is to learn a model $f$ to identify and classify event mentions for the target event types.", + "bbox": [ + 507, + 351, + 882, + 448 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For few-shot ED (FSED), there are two separate training data sets for few-shot event detection: (1) A large-scale data set $\\mathcal{D}_{base} = \\{(\\mathbf{x}_i,\\mathbf{y}_i)\\}_{i = 1}^M$ that covers the old event types (named base types) where $M$ denotes the number of base event types; (2) a smaller data set $\\mathcal{D}_{novel} = \\{(\\mathbf{x}_j,\\mathbf{y}_j)\\}_{j = 1}^{N\\times K}$ that covers $N$ novel event types, with $K$ examples each. Note that the base and novel event types are disjoint except for the Other class. The model $f$ will be first optimized on $\\mathcal{D}_{base}$ , and then further fine-tuned on $D_{novel}$ . The goal is to evaluate the generalizability and transferability of the model from base event types to new event types with few annotations.", + "bbox": [ + 507, + 449, + 882, + 673 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For zero-shot ED (ZSED), the training data sets are the only difference between zero-shot and few-shot event detection. In zero-shot event detection, there is only a large-scale base training data set $\\mathcal{D}_{base} = \\{(\\mathbf{x}_i,\\mathbf{y}_i)\\}_{i = 1}^M$ for the base event types. The model $f$ will be only optimized on base event types and evaluated on the novel types.", + "bbox": [ + 507, + 675, + 884, + 788 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.2 Event Type Prompts", + "text_level": 1, + "bbox": [ + 507, + 801, + 717, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We compare the following five forms of prompts to represent the event types: (a) Event Type Name is the event class name, usually consisting of one to three tokens. (b) Definition can be a short sentence that formally describes the meaning of the event types. (c) Prototype Seed Triggers a list of", + "bbox": [ + 507, + 822, + 882, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1287", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ad00dc77d21e6e626ac8784c7edf17c7db361a5b331e9c7a53e86d5455263aab.jpg", + "image_caption": [ + "Figure 1: Overview of the unified framework for event detection based on event type specific prompts." + ], + "image_footnote": [], + "bbox": [ + 126, + 85, + 684, + 233 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a6636a2acdf2de32d3f68ae9a163bd57d22cfdc0ea4d615e461b218bdf15b604.jpg", + "image_caption": [ + "Event Type Prompt" + ], + "image_footnote": [], + "bbox": [ + 705, + 97, + 880, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tokens or phrases that are frequently identified as event triggers. (d) Event Type Structure consists of event key argument roles, indicating the core participants of the target event type. (e) Prompts can also be Continuous Soft Prompt, that is, a free vector of parameters to represent each event type. (f) We further define a more comprehensive description APEX Prompt that is manually written and covers all previous prompts except soft prompts. Examples of all event type prompts are shown in Figure 1 and Appendix A. Detailed prompt token selection is in Appendix B.", + "bbox": [ + 112, + 282, + 489, + 476 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4 A Unified Framework for ED", + "text_level": 1, + "bbox": [ + 112, + 487, + 401, + 502 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We adapt (Wang et al., 2021a) and design a unified event detection framework (as shown in Figure 1) which leverages event type specific prompts to detect events under supervised, few-shot, and zero-shot settings. Formally, given an input sentence $W = \\{w_{1}, w_{2}, \\dots, w_{n}\\}$ , we take each event type prompt $T^{t} = \\{\\tau_{1}^{t}, \\tau_{2}^{t}, \\dots, \\tau_{m}^{t}\\}$ as a query of $M$ tokens to extract triggers for event type $t$ . Specifically, we first concatenate them into a sequence [CLS] $\\tau_{1}^{t} \\dots \\tau_{m}^{t}$ [SEP] $w_{1} \\dots w_{n}$ [SEP]. We use a pre-trained BERT encoder (Devlin et al., 2019) to get contextual representations for the input sentence $W = \\{w_{0}, w_{2}, \\dots, w_{n}\\}$ as well as the event type prompt $T = \\{\\tau_{0}^{t}, \\tau_{1}^{t}, \\dots, \\tau_{m}^{t}\\}^{2}$ .", + "bbox": [ + 112, + 511, + 489, + 736 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a prompt of each event type, we aim to extract corresponding event triggers from the input sentence. To achieve this goal, we need to capture the semantic correlation of each input token to the event type. Thus we learn a weight distribution over the sequence of contextual representations of the event type prompt, to obtain event type $t$ aware contextual representation $A_{i}^{t} = \\sum_{j=1}^{|T^{t}|} \\alpha_{ij} \\cdot \\tau_{j}^{t}$ , where $\\alpha_{ij} = \\cos(\\boldsymbol{w}_{i}, \\tau_{j}^{t})$ , where", + "bbox": [ + 112, + 737, + 489, + 888 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2In our experiments, the representation of each $\\pmb{w}_i$ or $\\pmb{\\tau}_i$ is based on the contextual embedding of the first sub-token.", + "bbox": [ + 112, + 892, + 485, + 917 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\tau_{j}$ is the contextual representation of the $j$ -th prompt token. $\\cos (\\cdot)$ is the cosine similarity function between two vectors.", + "bbox": [ + 507, + 282, + 884, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With that, the event type aware contextual representation $\\mathbf{A}_i^t$ will be concatenated with the original contextual representation $\\mathbf{w}_i$ from the encoder, and classified into a binary label, indicating whether it is a candidate trigger of event type $t$ or not: $\\tilde{\\mathbf{y}}_i^t = \\mathbf{U}_o([ \\mathbf{w}_i; \\mathbf{A}_i^t; \\mathbf{P}_i ])$ , where $[;]$ denotes concatenation operation, $\\mathbf{U}_o$ is a learnable parameter matrix for event trigger detection, and $\\mathbf{P}_i$ is the one-hot part-of-speech (POS) encoding of word $\\mathbf{w}_i$ . For continuous soft prompt based event detection, we follow Li and Liang (2021) where a prefix index $q$ is prepended to the input sequence $W' = [q; W]$ . The prefix embedding is learned by $\\mathbf{q} = \\mathrm{MLP}_{\\theta}(\\mathbf{Q}_{\\theta}[q])$ , where $\\mathbf{Q}_{\\theta} \\in \\mathbb{R}^{|\\mathcal{Q}| \\times k}$ denotes the embedding lookup table for the vocabulary of prefix indices. Both $\\mathrm{MLP}_{\\theta}$ and $\\mathbf{Q}_{\\theta}$ are trainable parameters. Detailed learning strategy is in Appendix C.", + "bbox": [ + 507, + 332, + 884, + 620 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "5 Experiment Setup", + "text_level": 1, + "bbox": [ + 507, + 634, + 700, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We perform experiments on three public benchmark datasets, including ACE05-E $^{+}$ (Automatic Content Extraction), ERE (Entity Relation Event) (Song et al., 2015), and MAVEN (Wang et al., 2020a). On each dataset, we conduct experiments for SED, FSED, and ZSED. For SED, we use the same data split as the previous studies (Li et al., 2013; Wadden et al., 2019; Lin et al., 2020; Du and Cardie, 2020; Lin et al., 2020; Nguyen et al., 2021; Wang et al., 2020a) on all the three benchmark datasets. For FSED and ZSED on MAVEN, we follow the previous study (Chen et al., 2021) and choose 120 event types with the most frequent mentions as the base event types and the rest 45 event types as novel ones. For FSED and ZSED on ACE and ERE, previous studies (Lai et al., 2020b,a;", + "bbox": [ + 505, + 661, + 884, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1288", + "bbox": [ + 482, + 927, + 519, + 940 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/96d92f02992ef9dd1f782a365b502db021551420812024554ffc91819b572827.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSEDFSEDZSED
Previous SOTA73.3\n(Nguyen et al., 2021)35.2*\n(Lai et al., 2020b)49.1*\n(Zhang et al., 2021b)
(a) Event type name72.252.749.8
(b) Definition73.146.745.5
(c) Seed triggers73.753.849.6
(d) Event structure72.850.448.0
(e) Soft prompt68.148.2-
Majority voting of (a-e)73.952.148.7
(f) APEX Prompt74.957.451.2
", + "bbox": [ + 117, + 80, + 485, + 230 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/db8770a7a9101ca525fe6872e9e0455ac14d086c11049df461ed255d5c9929c8.jpg", + "table_caption": [ + "Table 1: Results of event detection (ED) on ACE05 (F1-score, %)* indicates evaluation on our data set split based on the authors' public implementations." + ], + "table_footnote": [], + "table_body": "
MethodSEDFSEDZSED
Previous SOTA59.4(Lu et al., 2021)33.0*(Lai et al., 2020b)41.2*(Zhang et al., 2021b)
(a) Event type Name58.244.840.5
(b) Definition57.944.240.4
(c) Seed triggers60.450.446.2
(d) Event structure59.148.548.7
(e) Soft prompt55.641.7-
Majority voting of (a-e)60.247.945.6
(f) APEX Prompt63.452.648.9
", + "bbox": [ + 117, + 296, + 485, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Chen et al., 2021) follow different data splits and settings, making it hard for a fair comparison. Considering the research goals of FSED and ZSED, we define the following conditions to split the ACE and ERE datasets: (i) The base event types and novel event types should be disjoint except Other. (ii) Each base or novel event type should contain at least 15 instances. (iii) The training set should contain sufficient annotated event mentions.", + "bbox": [ + 112, + 508, + 487, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To meet the above conditions, for ACE, we define the event types of 5 main event categories: Business, Contact, Conflict, Justice and Movement as the base event types, and types of the remaining 3 main categories: Life, Personnel and Transaction as the novel event types. In total, there are 18 qualified base types and 10 qualified novel types (the others do not satisfy the second condition). For ERE, we use the exact same 10 novel event types as ACE, and the rest 25 types as base event types. Detailed data and hyperparameter descriptions are in Appendix D and Appendix E.", + "bbox": [ + 112, + 653, + 489, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "6 Results and Discussion", + "text_level": 1, + "bbox": [ + 112, + 860, + 346, + 876 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overall Results The experimental results for SED, FSED, and ZSED on ACE05, ERE, and", + "bbox": [ + 112, + 887, + 487, + 917 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/dbba2a5306c5f6f7506ad0eda772dd02235a3b9a611fa2cb7317435cdcdbd1d5.jpg", + "table_caption": [ + "Table 2: Results of event detection (ED) on ERE (F1-score, %)." + ], + "table_footnote": [], + "table_body": "
MethodSEDFSEDZSED
Previous SOTA68.5\n(Wang et al., 2021b)57.0\n(Chen et al., 2021)40.2*\n(Zhang et al., 2021b)
(a) Event type name68.863.458.8
(b) Definition67.156.952.9
(c) Seed triggers68.765.159.1
(e) Soft prompt64.538.6-
Majority voting of (a-e)68.463.458.1
(f) APEX Prompt68.868.459.9
", + "bbox": [ + 512, + 80, + 882, + 219 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 3: Results of event detection (ED) on MAVEN (F1-score, %). Event type structure prompts are not applicable to MAVEN as it does not contain any predefined argument roles.", + "bbox": [ + 507, + 228, + 884, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MAVEN are shown in Table 1-3, from which we see that (1) the APEX prompt achieves the best performance among all the forms of prompts under all the settings of the three benchmark datasets. Compared with the previous state of the art, the APEX prompt shows up to $4\\%$ F-score gain for SED (on ERE), $22.2\\%$ F-score gain for FSED (on ACE), and $19.7\\%$ F-score gain for ZSED (on MAVEN); (2) All the forms of prompts provide significant improvement for FSED and ZSED, demonstrating the benefit of leveraging the semantics of event types via various forms of prompts. (3) Except APEX, seed triggers provide more improvements than other forms of event type prompts under most settings, suggesting its potential to represent the semantics of event types accurately. (4) Continuous soft prompt does not provide comparable performance as other forms of event type representations, which proves the necessity of leveraging event type specific prior knowledge to the representations; (5) The majority voting does not show improvement over individual prompts since each prompt captures a particular aspect of the event type semantics.", + "bbox": [ + 507, + 312, + 884, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Supervised Event Detection By carefully investigating the event mentions that are correctly detected by the APEX prompt while missed by other prompts, we find that the APEX prompt is more effective in detecting two types of event mentions: homonyms (multiple-meaning words) and intricate words. General homonyms are usually hard to be detected as event mentions as they usually have dozens of meanings in different contexts. For example, consider the following two examples: (i) Airlines are getting [Transport:Movement] flyers to destinations on time more often. (ii) If the board cannot vote to give [Transaction:Transfer-Money'] themselves present money. Here, \"get\" and \"give\"", + "bbox": [ + 507, + 694, + 885, + 919 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1289", + "bbox": [ + 482, + 928, + 521, + 940 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/940326617a3f0805c4bc018c189d1a6c32c56c66230e62b62b39bb80e6946672.jpg", + "image_caption": [ + "Figure 2: F-score distribution of all novel types based on various event type prompts under the few-shot event detection setting on ACE (Best view in color)" + ], + "image_footnote": [], + "bbox": [ + 233, + 85, + 759, + 265 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "are not detected based on the event type name or seed triggers but are correctly identified by the definition and APEX prompts. The definition and APEX prompts make $10\\%$ and $7\\%$ fewer false predictions than seed triggers on general homonyms. For intricate words, their semantics usually cannot be captured with an individual prompt. In the following two examples: (i) It is reasonable, however, to reimburse board members for legitimate expenses (ii) ... ever having discussed being compensated by the board in the future ... \"reimburse\" and \"compensated\" indicate sophisticated meaning of Transaction:Transfer-Money, which may not be captured by prompts, such as seed triggers. With the event definition and the argument roles in the APEX prompt, the highly correlated contexts, such as \"board members\" and \"legitimate expenses\", can help the model correctly detect reimburse as an event mention of Transaction:Transfer-Money.", + "bbox": [ + 115, + 318, + 489, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Few-shot Event Detection Figure 2 shows the F-score distribution of all novel types based on various forms of event type prompts, from which we observe that: (1) The event type name, seed triggers, and APEX prompt generally perform better than definition and structure, as they carry more straightforward semantics of event types. (2) Event type name based prompts show lower performance on Personnel:End-Position, Personnel:Start-Position and Transaction:Transfer-Money than other event types, as the semantics of these event type names are less indicative than other event types. (3) Seed trigger based prompts perform worse than event type name and APEX prompts on two event types, Life:injure and Life:die, probably because the prototype seed triggers are not properly selected. (4) The structure based prompt outperforms the other", + "bbox": [ + 112, + 645, + 489, + 917 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "prompts on Life:Injure as Life:Injure events require the existence of a person or victim. (5) APEX prompt shows consistently (almost) best performance on all the event types because it combines all the information of other prompts. (6) We also observe that the performance of Life:Be-Born, Life:Die, Life:Marry, and Personnel:Elect based on various forms of prompts are consistently better than the other types as the intrinsic semantics of those types the corresponding event triggers are concentrated.", + "bbox": [ + 507, + 318, + 884, + 495 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Zero-shot Event Detection The proposed prompt-based method is more affordable to be generalized compared with the prior state-of-the-art zero-shot approach (Zhang et al., 2021b). The average length of created APEX prompts is less than 20 tokens. Thus manually creating them will not take much human effort. On the contrary, Zhang et al. (2021b) requires an extensive collection of anchor sentences to perform zero-shot event detection, e.g., 4,556,237 anchor sentences for ACE and ERE. This process is time-consuming and expensive.", + "bbox": [ + 507, + 507, + 882, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "7 Conclusion", + "text_level": 1, + "bbox": [ + 509, + 714, + 640, + 730 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We investigate a variety of prompts to represent the semantics of event types, and leverage them with a unified framework for supervised, few-shot and zero-shot event detection. Experimental results demonstrate that, a well-defined and comprehensive description of event types can significantly improve the performance of event detection, especially when the annotations are limited (few-shot event detection) or even not available (zero-shot event detection), with up to $22.2\\%$ F-score gain over the prior state of the art.", + "bbox": [ + 507, + 741, + 884, + 917 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1290", + "bbox": [ + 482, + 927, + 521, + 940 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Limitations", + "text_level": 1, + "bbox": [ + 114, + 84, + 220, + 98 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We have demonstrated that an accurate description can perform better for both supervised and weakly supervised event detection. However, the event types from most existing ontologies are not properly defined. For example, in ACE annotation guideline (Linguistic Data Consortium, 2005), transfer-money is defined as \"giving, receiving, borrowing, or lending money when it is not in the context of purchasing something\". However, it is hard for the model to interpret it accurately, especially the constraints \"not in the context of purchasing something\". In addition, many event types from MAVEN, e.g., Achieve, Award, and Incident, are not associated with any definitions. A potential future research direction is to leverage mining-based approaches or state-of-the-art generators to automatically generate a comprehensive event type description based on various sources, such as annotation guidelines, example annotations, and external knowledge bases.", + "bbox": [ + 115, + 110, + 489, + 430 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 114, + 444, + 278, + 458 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We thank the anonymous reviewers and area chair for their valuable time and constructive comments. This research is based upon work supported by the Amazon Research Award.", + "bbox": [ + 112, + 470, + 487, + 533 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 114, + 561, + 213, + 576 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Ashutosh Agarwal, Anay Majee, Anbumani Subramanian, and Chetan Arora. 2021. Attention guided cosine margin for overcoming class-imbalance in few-shot road object detection.", + "David Ahn. 2006. The stages of event extraction. In Proceedings of the Workshop on Annotating and Reasoning about Time and Events, pages 1-8.", + "Collin F Baker, Charles J Fillmore, and John B Lowe. 1998. The berkeley framenet project. In 36th Annual Meeting of the Association for Computational Linguistics and 17th International Conference on Computational Linguistics, Volume 1, pages 86-90.", + "Laura Banarescu, Claire Bonial, Shu Cai, Madalina Georgescu, Kira Griffith, Ulf Hermjakob, Kevin Knight, Philipp Koehn, Martha Palmer, and Nathan Schneider. 2013. Abstract Meaning Representation for sembanking. In Proceedings of the 7th Linguistic Annotation Workshop and Interoperability with Discourse, pages 178-186, Sofia, Bulgaria. Association for Computational Linguistics.", + "Ofer Bronstein, Ido Dagan, Qi Li, Heng Ji, and Anette Frank. 2015. Seed-based event trigger labeling: How" + ], + "bbox": [ + 115, + 583, + 487, + 917 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "far can event descriptions get us? In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 372-376.", + "Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems, volume 33, pages 1877-1901. Curran Associates, Inc.", + "Kai Cao, Xiang Li, Miao Fan, and Ralph Grishman. 2015. Improving event detection with active learning. In Proceedings of the International Conference on Recent Advances in Natural Language Processing, pages 72-77, Hissar, Bulgaria. INCOMA Ltd. Shoumen, BULGARIA.", + "Jiawei Chen, Hongyu Lin, Xianpei Han, and Le Sun. 2021. Honey or poison? solving the trigger curse in few-shot event detection via causal intervention.", + "Yubo Chen, Liheng Xu, Kang Liu, Daojian Zeng, and Jun Zhao. 2015. Event extraction via dynamic multipooling convolutional neural networks. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 167-176.", + "Yunmo Chen, Tongfei Chen, Seth Ebner, Aaron Steven White, and Benjamin Van Durme. 2020. Reading the manual: Event extraction as definition comprehension. In Proceedings of the Fourth Workshop on Structured Prediction for NLP, pages 74-83, Online. Association for Computational Linguistics.", + "Nancy Chinchor and Elaine Marsh. 1998. Muc-7 information extraction task definition. In Proceeding of the seventh message understanding conference (MUC-7), Appendices, pages 359-367.", + "Arkabandhu Chowdhury, Mingchao Jiang, and Chris Jermaine. 2021. Few-shot image classification: Just use a library of pre-trained feature extractors and a simple classifier. abs/2101.00562.", + "Xin Cong, Shiyao Cui, Bowen Yu, Tingwen Liu, Yubin Wang, and Bin Wang. 2021. Few-shot event detection with prototypical amortized conditional random field. In Findings of the Association for Computational Linguistics: ACL-IJCNLP.", + "Bhavana Dalvi, William W. Cohen, and Jamie Callan. 2012. Websets: extracting sets of entities from" + ], + "bbox": [ + 510, + 85, + 884, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1291", + "bbox": [ + 482, + 928, + 517, + 940 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the web using unsupervised information extraction. ArXiv, abs/1307.0261.", + "Shumin Deng, Ningyu Zhang, Jiaojian Kang, Yichi Zhang, Wei Zhang, and Huajun Chen. 2020. Meta-learning with dynamic-memory-based prototypical network for few-shot event detection. Proceedings of the 13th International Conference on Web Search and Data Mining.", + "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics.", + "Xinya Du and Claire Cardie. 2020. Event extraction by answering (almost) natural questions. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 671-683, Online. Association for Computational Linguistics.", + "Xiaocheng Feng, Lifu Huang, Duyu Tang, Heng Ji, Bing Qin, and Ting Liu. 2016. A language-independent neural network for event detection. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 66-71, Berlin, Germany. Association for Computational Linguistics.", + "Ralph Grishman. 1997. Information extraction: Techniques and challenges. In International summer school on information extraction, pages 10-27. Springer.", + "Karen Hambardzumyan, Hrant Khachatrian, and Jonathan May. 2021. WARP: Word-level Adversarial ReProgramming. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4921-4933, Online. Association for Computational Linguistics.", + "Lifu Huang, Taylor Cassidy, Xiaocheng Feng, Heng Ji, Clare Voss, Jiawei Han, and Avirup Sil. 2016. Liberal event extraction and event schema induction. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 258-268.", + "Lifu Huang and Heng Ji. 2020. Semi-supervised new event type induction and event detection. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 718-724.", + "Lifu Huang, Heng Ji, Kyunghyun Cho, Ido Dagan, Sebastian Riedel, and Clare Voss. 2018. Zero-shot transfer learning for event extraction. In Proceedings of the 56th Annual Meeting of the Association for" + ], + "bbox": [ + 115, + 85, + 487, + 917 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Computational Linguistics (Volume 1: Long Papers), pages 2160-2170, Melbourne, Australia. Association for Computational Linguistics.", + "Heng Ji and Ralph Grishman. 2008. Refining event extraction through cross-document inference. In Proceedings of ACL-08: Hlt, pages 254-262.", + "Zhengbao Jiang, Frank F. Xu, J. Araki, and Graham Neubig. 2020. How can we know what language models know? Transactions of the Association for Computational Linguistics, 8:423-438.", + "Bingyi Kang, Zhuang Liu, Xin Wang, Fisher Yu, Jiashi Feng, and Trevor Darrell. 2019. Few-shot object detection via feature reweighting. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 8419-8428.", + "Viet Dac Lai, Franck Dernoncourt, and Thien Huu Nguyen. 2020a. Exploiting the matching information in the support set for few shot event classification. Pacific-Asia Conference on Knowledge Discovery and Data Mining, page 233-245.", + "Viet Dac Lai and Thien Huu Nguyen. 2019. Extending event detection to new types with learning from keywords. arXiv preprint arXiv:1910.11368.", + "Viet Dac Lai, Thien Huu Nguyen, and Franck Dernoncourt. 2020b. Extensively matching for few-shot learning event detection. In Proceedings of the First Joint Workshop on Narrative Understanding, Storylines, and Events, pages 38-45, Online. Association for Computational Linguistics.", + "Brian Lester, Rami Al-Rfou, and Noah Constant. 2021. The power of scale for parameter-efficient prompt tuning. In EMNLP.", + "Bohao Li, Boyu Yang, Chang Liu, Feng Liu, Rongrong Ji, and Qixiang Ye. 2021. Beyond max-margin: Class margin equilibrium for few-shot object detection. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7359-7368.", + "Fayuan Li, Weihua Peng, Yuguang Chen, Quan Wang, Lu Pan, Yajuan Lyu, and Yong Zhu. 2020. Event extraction as multi-turn question answering. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 829–838, Online. Association for Computational Linguistics.", + "Qi Li, Heng Ji, and Liang Huang. 2013. Joint event extraction via structured prediction with global features. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 73-82, Sofia, Bulgaria. Association for Computational Linguistics.", + "Xiang Lisa Li and Percy Liang. 2021. Prefix-tuning: Optimizing continuous prompts for generation. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), abs/2101.00190." + ], + "bbox": [ + 510, + 85, + 880, + 917 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1292", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Shasha Liao and Ralph Grishman. 2010. Using document level cross-event inference to improve event extraction. In Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 789-797.", + "Ying Lin, Heng Ji, Fei Huang, and Lingfei Wu. 2020. A joint neural model for information extraction with global features. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7999-8009, Online. Association for Computational Linguistics.", + "Linguistic Data Consortium. 2005. English annotation guidelines for events. https://www.ldc.upenn.edu/sites/www.ldc.upenn.edu/files/english-events-guidelines-v5.4.3.pdf.", + "Jian Liu, Yubo Chen, Kang Liu, Wei Bi, and Xiaojiang Liu. 2020. Event extraction as machine reading comprehension. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1641-1651, Online. Association for Computational Linguistics.", + "Yaojie Lu, Hongyu Lin, Jin Xu, Xianpei Han, Jialong Tang, Annan Li, Le Sun, Meng Liao, and Shaoyi Chen. 2021. Text2Event: Controllable sequence-to-structure generation for end-to-end event extraction. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 2795-2806, Online. Association for Computational Linguistics.", + "Qing Lyu, Hongming Zhang, Elior Sulem, and Dan Roth. 2021. Zero-shot Event Extraction via Transfer Learning: Challenges and Insights. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 322-332, Online. Association for Computational Linguistics.", + "David McClosky, Mihai Surdeanu, and Christopher D Manning. 2011. Event extraction as dependency parsing. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 1626-1635.", + "Minh Van Nguyen, Viet Dac Lai, and Thien Huu Nguyen. 2021. Cross-task instance representation interactions and label dependencies for joint information extraction with graph convolutional networks.", + "Thien Huu Nguyen, Kyunghyun Cho, and Ralph Grishman. 2016. Joint event extraction via recurrent neural networks. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 300-309, San Diego, California. Association for Computational Linguistics." + ], + "bbox": [ + 115, + 85, + 489, + 917 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Panupong Pasupat and Percy Liang. 2014. Zero-shot entity extraction from web pages. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 391-401.", + "Fabio Petroni, Tim Rocttäschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, and Alexander Miller. 2019. Language models as knowledge bases? In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2463-2473, Hong Kong, China. Association for Computational Linguistics.", + "Colin Raffel, Noam M. Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR.", + "Timo Schick and Hinrich Schütze. 2021a. Few-shot text generation with pattern-exploiting training.", + "Timo Schick and Hinrich Schütze. 2021b. It's not just size that matters: Small language models are also few-shot learners. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics, pages 2339-2352.", + "Shirong Shen, Tongtong Wu, Guilin Qi, Yuan-Fang Li, Gholamreza Haffari, and Sheng Bi. 2021. Adaptive knowledge-enhanced bayesian meta-learning for few-shot event detection. In Findings of the Association for Computational Linguistics, page 2417-2429. Association for Computational Linguistics (ACL). Annual Meeting of the Association for Computational Linguistics and the International Joint Conference on Natural Language Processing 2021, ACL-IJCNLP 2021; Conference date: 01-08-2021 Through 06-08-2021.", + "Taylor Shin, Yasaman Razeghi, Robert L. Logan IV, Eric Wallace, and Sameer Singh. 2020. AutoPrompt: Eliciting knowledge from language models with automatically generated prompts. In Empirical Methods in Natural Language Processing (EMNLP).", + "Zhiyi Song, Ann Bies, Stephanie Strassel, Tom Riese, Justin Mott, Joe Ellis, Jonathan Wright, Seth Kulick, Neville Ryant, and Xiaoyi Ma. 2015. From light to rich ere: annotation of entities, relations, and events. In Proceedings of the 3rd Workshop on EVENTS: Definition, Detection, Coreference, and Representation, pages 89-98.", + "Bo Sun, Banghuai Li, Shengcai Cai, Ye Yuan, and Chi Zhang. 2021. Fsce: Few-shot object detection via contrastive proposal encoding. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7348-7358.", + "David Wadden, Ulme Wennberg, Yi Luan, and Hannaneh Hajishirzi. 2019. Entity, relation, and event" + ], + "bbox": [ + 510, + 85, + 882, + 917 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1293", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "extraction with contextualized span representations. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5784-5789, Hong Kong, China. Association for Computational Linguistics.", + "Richard C Wang and William Cohen. 2009. Character-level analysis of semi-structured documents for set expansion. In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 1503–1512.", + "Sijia Wang, Mo Yu, Shiyu Chang, Lichao Sun, and Lifu Huang. 2021a. Query and extract: Refining event extraction as type-oriented binary decoding. arXiv preprint arXiv:2110.07476.", + "Xiao Wang, Houye Ji, Chuan Shi, Bai Wang, Yanfang Ye, Peng Cui, and Philip S Yu. 2019. Heterogeneous graph attention network. In The World Wide Web Conference, WWW '19, page 2022-2032, New York, NY, USA. Association for Computing Machinery.", + "Xiaozhi Wang, Ziqi Wang, Xu Han, Wangyi Jiang, Rong Han, Zhiyuan Liu, Juanzi Li, Peng Li, Yankai Lin, and Jie Zhou. 2020a. MAVEN: A massive general domain event detection dataset. In Proceedings of EMNLP 2020.", + "Xin Wang, Thomas E. Huang, Trevor Darrell, Joseph E Gonzalez, and Fisher Yu. 2020b. Frustratingly simple few-shot object detection.", + "Ziqi Wang, Xiaozhi Wang, Xu Han, Yankai Lin, Lei Hou, Zhiyuan Liu, Peng Li, Juanzi Li, and Jie Zhou. 2021b. CLEVE: Contrastive Pre-training for Event Extraction. In Proceedings of ACL-IJCNLP, pages 6283-6297, Online. Association for Computational Linguistics.", + "Yang Xiao and Renaud Marlet. 2020. Few-shot object detection and viewpoint estimation for objects in the wild. In ECCV.", + "Xiaopeng Yan, Ziliang Chen, Anni Xu, Xiaoxi Wang, Xiaodan Liang, and Liang Lin. 2019. Meta r-cnn: Towards general solver for instance-level low-shot learning. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9576-9585.", + "Bishan Yang and Tom M. Mitchell. 2016. Joint extraction of events and entities within a document context. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 289-299, San Diego, California. Association for Computational Linguistics.", + "Weizhe Yuan, Graham Neubig, and Pengfei Liu. 2021. BARTScore: Evaluating generated text as text generation. In Advances in Neural Information Processing Systems." + ], + "bbox": [ + 115, + 85, + 489, + 917 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Gongjie Zhang, Kaiwen Cui, Rongliang Wu, Shijian Lu, and Yonghong Tian. 2021a. Pnpdet: Efficient few-shot detection without forgetting via plug-and-play sub-networks. 2021 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 3822-3831.", + "Hongming Zhang, Haoyu Wang, and Dan Roth. 2021b. Zero-shot Label-aware Event Trigger and Argument Classification. In *Findings of the Association for Computational Linguistics: ACL-IJCNLP* 2021, pages 1331-1340, Online. Association for Computational Linguistics.", + "Tongtao Zhang, Spencer Whitehead, Hanwang Zhang, Hongzhi Li, Joseph Ellis, Lifu Huang, Wei Liu, Heng Ji, and Shih-Fu Chang. 2017. Improving event extraction via multimodal integration. In Proceedings of the 25th ACM international conference on Multimedia, pages 270-278." + ], + "bbox": [ + 510, + 85, + 880, + 342 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1294", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "A APEX prompt examples for ACE", + "text_level": 1, + "bbox": [ + 112, + 83, + 438, + 99 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 4 and Table 5 show APEX prompt examples for ACE events.", + "bbox": [ + 112, + 109, + 485, + 141 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "B Prompt Token Selection", + "text_level": 1, + "bbox": [ + 112, + 154, + 359, + 171 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In our experiments, the event type names and event type structures are automatically extracted from the target event ontology, such as ACE (Linguistic Data Consortium, 2005), ERE (Song et al., 2015) and MAVEN (Wang et al., 2020a). The prototype seed triggers are automatically selected from the annotated data for supervised and few-shot event extraction. For zero-shot event extraction, we manually select $R$ words from the NLTK synonyms of each event type as its prototype seed triggers. The definitions and APEX prompts are based on the official annotation guides for each target event ontology (Linguistic Data Consortium, 2005; Song et al., 2015; Wang et al., 2020a) and the available definitions in FrameNet (Baker et al., 1998) with manual editing.", + "bbox": [ + 112, + 181, + 489, + 439 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "C Learning Strategy", + "text_level": 1, + "bbox": [ + 112, + 451, + 310, + 468 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The learning strategy varies for supervised, few-shot, and zero-shot learning. For supervised learning, we optimize the following objective for event trigger detection $\\mathcal{L} = -\\frac{1}{|\\mathcal{T}||\\mathcal{N}|}\\sum_{t\\in \\mathcal{T}}\\sum_{i = 1}^{|\\mathcal{N}|}\\boldsymbol{y}_i^t$ $\\log \\tilde{\\boldsymbol{y}}_i^t$ where $\\mathcal{T}$ is the set of target event types and $\\mathcal{N}$ is the set of tokens from the training dataset. $\\boldsymbol{y}_i^t$ denotes the ground truth label vector. For few-shot event detection, we optimize the model on both base training data set and the smaller training data set for novel event types: $\\mathcal{L} = -\\frac{1}{|\\mathcal{T}^B||\\mathcal{N}^B|}\\sum_{t\\in \\mathcal{T}^B}\\sum_{i = 1}^{|\\mathcal{N}^B|}\\boldsymbol{y}_i^t\\cdot \\log \\tilde{\\boldsymbol{y}}_i^t -$ $\\beta \\frac{1}{|\\mathcal{T}^N||\\mathcal{N}^N|}\\sum_{t\\in \\mathcal{T}^N}\\sum_{i = 1}^{|\\mathcal{N}^N|}\\boldsymbol{y}_i^t\\cdot \\log \\tilde{\\boldsymbol{y}}_i^t$ , where $\\mathcal{T}^B$ and $\\mathcal{N}^B$ denote the set of base event types and tokens from the base training data set, respectively. $\\mathcal{T}^N$ is the set of novel event types. $\\mathcal{N}^N$ is the set of tokens from the training data set for novel event types. $\\beta$ is a hyper-parameter to balance the two objectives. For zero-shot event detection, as we only have the base training data set, we minimize the following objective: $\\mathcal{L} = -\\frac{1}{|\\mathcal{T}^B||\\mathcal{N}^B|}\\sum_{t\\in \\mathcal{T}^B}\\sum_{i = 1}^{|\\mathcal{N}^B|}\\boldsymbol{y}_i^t$ $\\log \\tilde{\\boldsymbol{y}}_i^t$", + "bbox": [ + 112, + 476, + 489, + 832 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "D Dataset", + "text_level": 1, + "bbox": [ + 112, + 844, + 220, + 859 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "After defining the base and novel event types, we create the training, validation, and evaluation split for all three datasets. We use the sentences with", + "bbox": [ + 112, + 871, + 487, + 917 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "only base event type mentions as the base training data set for few-shot event detection, and randomly select 10 sentences with novel event type mentions as the additional smaller training data set. We use the sentences with both base and novel event type mentions as the development set and use the remaining sentences with only novel event type mentions as the evaluation dataset. We use the same development and evaluation set as few-shot event detection for zero-shot event detection and remove the instances with novel event mentions from the training set. We randomly split the sentences without any event annotations proportionally to the number of sentences with event mentions in each set for both zero-shot and few-shot event detection. Table 6 shows the detailed data statistics for all the three datasets under the few-shot and zero-shot event extraction settings.", + "bbox": [ + 507, + 84, + 884, + 374 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "E Hyperparameters and Evaluation", + "text_level": 1, + "bbox": [ + 507, + 385, + 836, + 401 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For a fair comparison with the previous baseline approaches, we use the same pre-trained bert-large-uncased model for fine-tuning and optimizing our model with BertAdam. For supervised event detection, we optimize the parameters with grid search: training epoch is 3, learning rate $\\in [3e - 6,1e - 4]$ , training batch size $\\in$ {8, 12, 16, 24, 32}, dropout rate $\\in$ {0.4, 0.5, 0.6}. The running time is up to 3 hours on one Quadro RTX 8000. For evaluation, we use the same criteria as previous studies (Li et al., 2013; Chen et al., 2015; Nguyen et al., 2016; Lin et al., 2020): an event mention is correct if its span and event type match a reference event mention.", + "bbox": [ + 507, + 411, + 884, + 634 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1295", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/f4003b7e10c48c68ad81d17bd1d16fee111d107424f11b5ec7f4d40f31d50f71.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Event Rep TypeComprehensive Prompt
Business:Declare-BankruptcyDeclare Bankruptcy [SEP] bankruptcy bankruptciesbankrupting [SEP] Organizationrequest legal protection from debt collection at a Place
Business:End-OrgEnd Organization [SEP] dissolving disbanded [SEP] an Organization goes out of business at a Place
Business:Merge-OrgMerge Organization [SEP] merging merger [SEP] two or more Organizations come together to form a new organization at a Place
Business:Start-OrgStart Organization [SEP] founded [SEP] an Agent create a new Organization at a Place
Conflict:AttackAttack [SEP] invaded airstrikes overthrew ambushed [SEP] An Attacker physically attacks a Target with Instrument at a Place
Conflict:DemonstrateDemonstrate [SEP] demonstrations protest strikes riots [SEP] Entities come together in a Place to protest or demand official action
Contact:MeetMeet [SEP] reunited retreats [SEP] two or more Entities come together at same Place and interact in person
Contact:Phone-WritePhone Write [SEP] emailed letter [SEP] phone or written communication between two or more Entities
Justice:AcquitAcquit [SEP] acquitted [SEP] a trial of Defendant ends but Adjudicator fails to produce a conviction at a Place
Justice:AppealAppeal [SEP] appeal [SEP] the decision for Defendant of a court is taken to a higher court for Adjudicator review with Prosecutor
Justice:Arrest-JailArrest Jail [SEP] arrested locked [SEP] the Agent takes custody of a Person at a Place
Justice:Charge-IndictCharge Indict [SEP] indictment [SEP] a Defendant is accused of a crime by a Prosecutor for Adjudicator
Justice:ConvictConvict [SEP] pled guilty convicting [SEP] an Defendant found guilty of a crime by Adjudicator at a Place
Justice:ExecuteExecute [SEP] death [SEP] the life of a Person is taken by an Agent at a Place
Justice:ExtraditeExtradite [SEP] extradition [SEP] a Person is sent by an Agent from Origin to Destination
Justice:FineFine [SEP] payouts financial punishment [SEP] a Adjudicator issues a financial punishment Money to an Entity at a Place
Justice:PardonPardon [SEP] pardoned lift sentence [SEP] an Adjudicator lifts a sentence of Defendant at a Place
Justice:Release-ParoleRelease Parole [SEP] parole [SEP] an Entity ends its custody of a Person at a Place
Justice:SentenceSentence [SEP] sentenced punishment [SEP] the punishment for the defendant is issued by a state actor
Justice:SueSue [SEP] lawsuits [SEP] Plaintiff initiate a court proceeding to determine the liability of a Defendant judge by Adjudicator at a Place
Justice:Trial-HearingTrial Hearing [SEP] trial hearings [SEP] a court proceeding initiated to determine the guilty or innocence of a Person with Prosecutor and Adjudicator at a Place
Life:Be-BornBe Born [SEP] childbirth [SEP] a Person is born at a Place
Life:DieDie [SEP] deceased extermination [SEP] life of a Victim ends by an Agent with Instrument at a Place
", + "bbox": [ + 166, + 135, + 831, + 838 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 4: APEX templates for ACE event types", + "bbox": [ + 339, + 845, + 655, + 860 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1296", + "bbox": [ + 482, + 928, + 521, + 940 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/db85a5d2ca78e6e95630c233c483aedd29200088a1a4fb1d0d8e52809303bdfa.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Event Rep TypeComprehensive Prompt
Life:DivorceDivorce [SEP] people divorce [SEP] two Person are officially divorced at a place
Life:InjureInjure [SEP] hospitalised paralyzed dismember [SEP] a Victim experiences physical harm from Agent with Instrument at a Place
Life:MarryMarry [SEP] married marriage marry [SEP] two Person are married at a Place
Movement:TransportTransport [SEP] arrival travels penetrated expelled [SEP] an Agent moves an Artifact from Origin to Destination with Vehicle at Price
Personnel:ElectElect [SEP] reelected elected election [SEP] a candidate Person wins an election by voting Entity at a Place
Personnel:End-PositionEnd Position [SEP] resigning retired resigned [SEP] a Person stops working for an Entity or change office at a Place
Personnel:NominateNominate [SEP] nominate [SEP] a Person is nominated for a new position by another Agent at a Place
Personnel:Start-PositionStart Position [SEP] hiring rehired recruited [SEP] a Person begins working for an Entity or change office at a Place
Transaction:Transfer-MoneyTransfer Money [SEP] donations reimbursing deductions [SEP] transfer Money from the Giver to the Beneficiary or Recipient at a Place
Transaction:Transfer-OwnershipTransfer Ownership [SEP] purchased buy sell loan [SEP] buying selling loaning borrowing giving receiving of Artifacts from Seller to Buyer or Beneficiary at a Place at Price
", + "bbox": [ + 166, + 155, + 831, + 470 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/54b225e7e82ce612a9251ab2cd6f081886751c268200cbb035f9b68557ab1e03.jpg", + "table_caption": [ + "Table 5: APEX templates for ACE event types (continued)" + ], + "table_footnote": [], + "table_body": "
DatasetACE05-E+ERE-ENMAVEN
# TypesBase1825120
Novel101045
# MentionsBase3572544993675
Novel172431833201
TrainFew-shot3216388688085
Zero-shot3116378687635
Validation900(51%/49%)2797(53%/47%)3883(71%/23%)
Evaluation119520121652
", + "bbox": [ + 265, + 653, + 732, + 816 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 6: Data statistics for ACE2005, ERE and MAVEN datasets under few-shot/zero-shot event detection settings.", + "bbox": [ + 112, + 825, + 880, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "1297", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A For every submission:", + "bbox": [ + 115, + 107, + 322, + 122 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "A1. Did you describe the limitations of your work? Section 8", + "A2. Did you discuss any potential risks of your work? Section 8", + "A3. Do the abstract and introduction summarize the paper's main claims? Abstract and Section 1", + "A4. Have you used AI writing assistants when working on this paper? Left blank." + ], + "bbox": [ + 129, + 126, + 695, + 288 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "B Did you use or create scientific artifacts?", + "bbox": [ + 114, + 299, + 487, + 316 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Section 5 and Appendix C", + "bbox": [ + 132, + 321, + 329, + 336 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "B1. Did you cite the creators of artifacts you used? Section 5 and Appendix C", + "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Section 5", + "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Section 5", + "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Left blank.", + "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Section 5", + "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Appendix C" + ], + "bbox": [ + 129, + 346, + 880, + 753 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "C Did you run computational experiments?", + "bbox": [ + 114, + 764, + 492, + 781 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Section 5 and Appendix D", + "bbox": [ + 132, + 785, + 329, + 801 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Appendix D", + "bbox": [ + 129, + 810, + 880, + 860 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "ACL 2023 Responsible NLP Checklist", + "bbox": [ + 132, + 84, + 433, + 99 + ], + "page_idx": 12 + }, + { + "type": "page_footnote", + "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.", + "bbox": [ + 112, + 868, + 877, + 892 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "1298", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?", + "bbox": [ + 129, + 83, + 878, + 115 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Section 5 and Appendix D", + "bbox": [ + 149, + 117, + 347, + 131 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?", + "bbox": [ + 129, + 142, + 882, + 190 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Appendix D", + "bbox": [ + 149, + 192, + 242, + 206 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?", + "bbox": [ + 129, + 217, + 882, + 265 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Appendix B", + "bbox": [ + 149, + 267, + 240, + 282 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?", + "bbox": [ + 112, + 293, + 877, + 309 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 132, + 313, + 213, + 329 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?", + "bbox": [ + 127, + 340, + 882, + 372 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 374, + 248, + 388 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?", + "bbox": [ + 127, + 399, + 882, + 447 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 449, + 248, + 463 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?", + "bbox": [ + 127, + 475, + 882, + 521 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 524, + 248, + 539 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?", + "bbox": [ + 127, + 549, + 873, + 565 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 567, + 248, + 581 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?", + "bbox": [ + 127, + 592, + 880, + 623 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 626, + 248, + 640 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "1299", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 13 + } +] \ No newline at end of file diff --git a/2023/The Art of Prompting_ Event Detection based on Type Specific Prompts/5e1d8751-6f7a-44c9-99a0-fd4f3669ec72_model.json b/2023/The Art of Prompting_ Event Detection based on Type Specific Prompts/5e1d8751-6f7a-44c9-99a0-fd4f3669ec72_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b8198896a6378f212fb544d5c614ab65010c5779 --- /dev/null +++ b/2023/The Art of Prompting_ Event Detection based on Type Specific Prompts/5e1d8751-6f7a-44c9-99a0-fd4f3669ec72_model.json @@ -0,0 +1,2428 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.129, + 0.09, + 0.872, + 0.112 + ], + "angle": 0, + "content": "The Art of Prompting: Event Detection based on Type Specific Prompts" + }, + { + "type": "text", + "bbox": [ + 0.339, + 0.141, + 0.662, + 0.159 + ], + "angle": 0, + "content": "Sijia Wang*, Mo Yu*, Lifu Huang*" + }, + { + "type": "text", + "bbox": [ + 0.377, + 0.159, + 0.625, + 0.175 + ], + "angle": 0, + "content": "\\(\\spadesuit\\) Virginia Tech, \\(\\spadesuit\\) WeChat AI" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.175, + 0.785, + 0.193 + ], + "angle": 0, + "content": "\\(\\clubsuit\\) {sijiawang, lifuh}@vt.edu, \\(\\clubsuit\\) moyumyu@tencent.com" + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.253, + 0.341, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.281, + 0.461, + 0.479 + ], + "angle": 0, + "content": "We compare various forms of prompts to represent event types and develop a unified framework to incorporate the event type specific prompts for supervised, few-shot, and zero-shot event detection. The experimental results demonstrate that a well-defined and comprehensive event type prompt can significantly improve event detection performance, especially when the annotated data is scarce (few-shot event detection) or not available (zero-shot event detection). By leveraging the semantics of event types, our unified framework shows up to \\(22.2\\%\\) F-score gain over the previous state-of-the-art baselines1." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.493, + 0.26, + 0.508 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.519, + 0.49, + 0.76 + ], + "angle": 0, + "content": "Event detection (ED) (Grishman, 1997; Chinchor and Marsh, 1998; Ahn, 2006) is the task of identifying and typing event mentions from natural language text. Supervised approaches, especially deep neural networks (Chen et al., 2020; Du and Cardie, 2020; Lin et al., 2020; Liu et al., 2020; Li et al., 2020; Lyu et al., 2021), have shown remarkable performance under a critical prerequisite of a large amount of manual annotations. However, they cannot be effectively generalized to new languages, domains or types, especially when the annotations are not enough (Huang et al., 2016; Huang and Ji, 2020; Lai et al., 2020b; Shen et al., 2021) or there is no annotation available (Lyu et al., 2021; Zhang et al., 2021b; Pasupat and Liang, 2014)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.761, + 0.49, + 0.873 + ], + "angle": 0, + "content": "Recent studies have shown that both the accuracy and generalizability of ED can be improved via leveraging the semantics of event types based on various forms of prompts, such as event type specific queries (Lyu et al., 2021; Du and Cardie, 2020; Liu et al., 2020), definitions (Chen et al., 2020), structures (Lin et al., 2020; Wang et al.," + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.254, + 0.885, + 0.414 + ], + "angle": 0, + "content": "2019), or a few prototype event triggers (Wang and Cohen, 2009; Dalvi et al., 2012; Pasupat and Liang, 2014; Bronstein et al., 2015; Lai and Nguyen, 2019; Zhang et al., 2021b; Cong et al., 2021). These studies further encourage us to take another step forward and think about the following three questions: (1) does the choice of prompt matter when the training data is abundant or scarce? (2) what's the best form of ED prompt? (3) how to best leverage the prompt to detect event mentions?" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.416, + 0.885, + 0.755 + ], + "angle": 0, + "content": "To answer the above research questions, we conduct extensive experiments with various forms of prompts for each event type, including (a) event type name, (b) prototype seed triggers, (c) definition, (d) event type structure based on both event type name and its predefined argument roles, (e) free parameter based continuous soft prompt, and (f) a more comprehensive event type description (named APEX prompt) that covers all the information of prompts (a)-(d). We observe that (1) by considering the semantics of event types with most forms of prompts, especially seed triggers and the comprehensive event type descriptions, the performance of ED under all settings can be significantly improved; (2) Among all forms of event representations, the comprehensive description based prompts show to be the most effective, especially for few-shot and zero-shot ED; (3) Different forms of event type representations provide complementary improvements, indicating that they capture distinct aspects and knowledge of the event types." + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.757, + 0.866, + 0.771 + ], + "angle": 0, + "content": "The contributions of this work are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.775, + 0.884, + 0.87 + ], + "angle": 0, + "content": "- We investigate various prompts to represent event types for both supervised and weakly supervised ED, and prove that a well-defined and comprehensive event type prompt can dramatically improve the performance of ED and the transferability from old types to new types." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.872, + 0.884, + 0.918 + ], + "angle": 0, + "content": "- A unified framework is developed to leverage the semantics of event types with prompts for supervised, few-shot, and zero-shot ED, and demonstrate" + }, + { + "type": "page_footnote", + "bbox": [ + 0.113, + 0.881, + 0.488, + 0.918 + ], + "angle": 0, + "content": "1The source code, model checkpoints and data are publicly available at https://github.com/VT-NLP/Event_APEX." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1286" + }, + { + "type": "footer", + "bbox": [ + 0.227, + 0.946, + 0.77, + 0.958 + ], + "angle": 0, + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + }, + { + "type": "footer", + "bbox": [ + 0.369, + 0.959, + 0.63, + 0.972 + ], + "angle": 0, + "content": "Volume 2: Short Papers, pages 1286-1299" + }, + { + "type": "footer", + "bbox": [ + 0.297, + 0.973, + 0.702, + 0.986 + ], + "angle": 0, + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.114, + 0.085, + 0.49, + 0.133 + ], + "angle": 0, + "content": "state-of-the-art performance with up to \\(22.2\\%\\) F-score improvement over the strong baseline methods." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.148, + 0.27, + 0.163 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.176, + 0.49, + 0.466 + ], + "angle": 0, + "content": "Supervised ED: Most of the existing Event Detection studies follow a supervised learning paradigm (Ji and Grishman, 2008; Liao and Grishman, 2010; McClosky et al., 2011; Li et al., 2013; Chen et al., 2015; Cao et al., 2015; Feng et al., 2016; Yang and Mitchell, 2016; Nguyen et al., 2016; Zhang et al., 2017; Lin et al., 2020; Wang et al., 2021b). However, they cannot be directly applied to detect new types of events. Recently studies have shown that, by leveraging the semantics of event types based on type-specific questions (Du and Cardie, 2020; Liu et al., 2020; Li et al., 2020; Lyu et al., 2021) or seed event triggers (Bronstein et al., 2015; Lai and Nguyen, 2019; Wang et al., 2021a), the event detection performance can be improved. However, it is still unknown whether they are the best choices for representing the semantics of event types." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.478, + 0.49, + 0.671 + ], + "angle": 0, + "content": "Few-shot ED: Two primary learning strategies in few-shot classification tasks are Meta-Learning (Kang et al., 2019; Li et al., 2021; Xiao and Marlet, 2020; Yan et al., 2019; Chowdhury et al., 2021) and Metric Learning (Sun et al., 2021; Wang et al., 2020b; Zhang et al., 2021a; Agarwal et al., 2021). Several studies have exploited metric learning to align the semantics of candidate events with a few examples of the novel event types for few-shot event detection (Lai et al., 2020a; Deng et al., 2020; Lai et al., 2020b; Cong et al., 2021; Chen et al., 2021; Shen et al., 2021)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.683, + 0.49, + 0.844 + ], + "angle": 0, + "content": "Zero-shot ED: Huang et al. (2018) first exploited zero-shot event extraction by leveraging Abstract Meaning Representation (Banarescu et al., 2013) to represent event mentions and types into a shared semantic space. Recent studies (Zhang et al., 2021b; Lyu et al., 2021) further demonstrate that by leveraging a large external corpus with abundant anchor triggers, zero-shot event detection can also be achieved with decent performance without using any training data." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.856, + 0.49, + 0.92 + ], + "angle": 0, + "content": "Prompt Learning Prompt learning aims to learn a task-specific prompt while keeping most of the model's parameters frozen (Li and Liang, 2021; Hambardzumyan et al., 2021; Brown et al., 2020)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.23 + ], + "angle": 0, + "content": "It has shown competitive performance in many applications of natural language processing (Raffel et al., 2020; Brown et al., 2020; Shin et al., 2020; Jiang et al., 2020; Lester et al., 2021; Schick and Schütze, 2021b). Previous work either used a manual (Petroni et al., 2019; Brown et al., 2020; Schick and Schütze, 2021a) or automated approach (Jiang et al., 2020; Yuan et al., 2021; Li and Liang, 2021) to create prompts." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.243, + 0.733, + 0.259 + ], + "angle": 0, + "content": "3 Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.271, + 0.885, + 0.319 + ], + "angle": 0, + "content": "Here, we first define each setting of the event detection task and then describe the various forms of event type prompts." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.331, + 0.669, + 0.347 + ], + "angle": 0, + "content": "3.1 Settings of ED" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.353, + 0.884, + 0.449 + ], + "angle": 0, + "content": "For supervised ED (SED), we follow the conventional supervised event detection setting where the training, validation, and evaluation data sets cover the same set of event types. The goal is to learn a model \\( f \\) to identify and classify event mentions for the target event types." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.45, + 0.884, + 0.674 + ], + "angle": 0, + "content": "For few-shot ED (FSED), there are two separate training data sets for few-shot event detection: (1) A large-scale data set \\(\\mathcal{D}_{base} = \\{(\\mathbf{x}_i,\\mathbf{y}_i)\\}_{i = 1}^M\\) that covers the old event types (named base types) where \\(M\\) denotes the number of base event types; (2) a smaller data set \\(\\mathcal{D}_{novel} = \\{(\\mathbf{x}_j,\\mathbf{y}_j)\\}_{j = 1}^{N\\times K}\\) that covers \\(N\\) novel event types, with \\(K\\) examples each. Note that the base and novel event types are disjoint except for the Other class. The model \\(f\\) will be first optimized on \\(\\mathcal{D}_{base}\\), and then further fine-tuned on \\(D_{novel}\\). The goal is to evaluate the generalizability and transferability of the model from base event types to new event types with few annotations." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.676, + 0.885, + 0.789 + ], + "angle": 0, + "content": "For zero-shot ED (ZSED), the training data sets are the only difference between zero-shot and few-shot event detection. In zero-shot event detection, there is only a large-scale base training data set \\(\\mathcal{D}_{base} = \\{(\\mathbf{x}_i,\\mathbf{y}_i)\\}_{i = 1}^M\\) for the base event types. The model \\(f\\) will be only optimized on base event types and evaluated on the novel types." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.802, + 0.719, + 0.818 + ], + "angle": 0, + "content": "3.2 Event Type Prompts" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.823, + 0.884, + 0.919 + ], + "angle": 0, + "content": "We compare the following five forms of prompts to represent the event types: (a) Event Type Name is the event class name, usually consisting of one to three tokens. (b) Definition can be a short sentence that formally describes the meaning of the event types. (c) Prototype Seed Triggers a list of" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1287" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.127, + 0.086, + 0.685, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.707, + 0.087, + 0.807, + 0.097 + ], + "angle": 0, + "content": "Event Type Prompt" + }, + { + "type": "image", + "bbox": [ + 0.707, + 0.098, + 0.882, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.151, + 0.244, + 0.843, + 0.26 + ], + "angle": 0, + "content": "Figure 1: Overview of the unified framework for event detection based on event type specific prompts." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.284, + 0.49, + 0.478 + ], + "angle": 0, + "content": "tokens or phrases that are frequently identified as event triggers. (d) Event Type Structure consists of event key argument roles, indicating the core participants of the target event type. (e) Prompts can also be Continuous Soft Prompt, that is, a free vector of parameters to represent each event type. (f) We further define a more comprehensive description APEX Prompt that is manually written and covers all previous prompts except soft prompts. Examples of all event type prompts are shown in Figure 1 and Appendix A. Detailed prompt token selection is in Appendix B." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.488, + 0.402, + 0.503 + ], + "angle": 0, + "content": "4 A Unified Framework for ED" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.512, + 0.49, + 0.737 + ], + "angle": 0, + "content": "We adapt (Wang et al., 2021a) and design a unified event detection framework (as shown in Figure 1) which leverages event type specific prompts to detect events under supervised, few-shot, and zero-shot settings. Formally, given an input sentence \\( W = \\{w_{1}, w_{2}, \\dots, w_{n}\\} \\), we take each event type prompt \\( T^{t} = \\{\\tau_{1}^{t}, \\tau_{2}^{t}, \\dots, \\tau_{m}^{t}\\} \\) as a query of \\( M \\) tokens to extract triggers for event type \\( t \\). Specifically, we first concatenate them into a sequence [CLS] \\( \\tau_{1}^{t} \\dots \\tau_{m}^{t} \\) [SEP] \\( w_{1} \\dots w_{n} \\) [SEP]. We use a pre-trained BERT encoder (Devlin et al., 2019) to get contextual representations for the input sentence \\( W = \\{w_{0}, w_{2}, \\dots, w_{n}\\} \\) as well as the event type prompt \\( T = \\{\\tau_{0}^{t}, \\tau_{1}^{t}, \\dots, \\tau_{m}^{t}\\}^{2} \\)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.738, + 0.49, + 0.889 + ], + "angle": 0, + "content": "Given a prompt of each event type, we aim to extract corresponding event triggers from the input sentence. To achieve this goal, we need to capture the semantic correlation of each input token to the event type. Thus we learn a weight distribution over the sequence of contextual representations of the event type prompt, to obtain event type \\( t \\) aware contextual representation \\( A_{i}^{t} = \\sum_{j=1}^{|T^{t}|} \\alpha_{ij} \\cdot \\tau_{j}^{t} \\), where \\( \\alpha_{ij} = \\cos(\\boldsymbol{w}_{i}, \\tau_{j}^{t}) \\), where" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.893, + 0.486, + 0.919 + ], + "angle": 0, + "content": "2In our experiments, the representation of each \\( \\pmb{w}_i \\) or \\( \\pmb{\\tau}_i \\) is based on the contextual embedding of the first sub-token." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.284, + 0.885, + 0.331 + ], + "angle": 0, + "content": "\\(\\tau_{j}\\) is the contextual representation of the \\(j\\)-th prompt token. \\(\\cos (\\cdot)\\) is the cosine similarity function between two vectors." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.333, + 0.885, + 0.621 + ], + "angle": 0, + "content": "With that, the event type aware contextual representation \\( \\mathbf{A}_i^t \\) will be concatenated with the original contextual representation \\( \\mathbf{w}_i \\) from the encoder, and classified into a binary label, indicating whether it is a candidate trigger of event type \\( t \\) or not: \\( \\tilde{\\mathbf{y}}_i^t = \\mathbf{U}_o([ \\mathbf{w}_i; \\mathbf{A}_i^t; \\mathbf{P}_i ]) \\), where \\( [;] \\) denotes concatenation operation, \\( \\mathbf{U}_o \\) is a learnable parameter matrix for event trigger detection, and \\( \\mathbf{P}_i \\) is the one-hot part-of-speech (POS) encoding of word \\( \\mathbf{w}_i \\). For continuous soft prompt based event detection, we follow Li and Liang (2021) where a prefix index \\( q \\) is prepended to the input sequence \\( W' = [q; W] \\). The prefix embedding is learned by \\( \\mathbf{q} = \\mathrm{MLP}_{\\theta}(\\mathbf{Q}_{\\theta}[q]) \\), where \\( \\mathbf{Q}_{\\theta} \\in \\mathbb{R}^{|\\mathcal{Q}| \\times k} \\) denotes the embedding lookup table for the vocabulary of prefix indices. Both \\( \\mathrm{MLP}_{\\theta} \\) and \\( \\mathbf{Q}_{\\theta} \\) are trainable parameters. Detailed learning strategy is in Appendix C." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.636, + 0.702, + 0.653 + ], + "angle": 0, + "content": "5 Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.662, + 0.885, + 0.919 + ], + "angle": 0, + "content": "We perform experiments on three public benchmark datasets, including ACE05-E\\(^{+}\\) (Automatic Content Extraction), ERE (Entity Relation Event) (Song et al., 2015), and MAVEN (Wang et al., 2020a). On each dataset, we conduct experiments for SED, FSED, and ZSED. For SED, we use the same data split as the previous studies (Li et al., 2013; Wadden et al., 2019; Lin et al., 2020; Du and Cardie, 2020; Lin et al., 2020; Nguyen et al., 2021; Wang et al., 2020a) on all the three benchmark datasets. For FSED and ZSED on MAVEN, we follow the previous study (Chen et al., 2021) and choose 120 event types with the most frequent mentions as the base event types and the rest 45 event types as novel ones. For FSED and ZSED on ACE and ERE, previous studies (Lai et al., 2020b,a;" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1288" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.119, + 0.082, + 0.486, + 0.231 + ], + "angle": 0, + "content": "
MethodSEDFSEDZSED
Previous SOTA73.3\n(Nguyen et al., 2021)35.2*\n(Lai et al., 2020b)49.1*\n(Zhang et al., 2021b)
(a) Event type name72.252.749.8
(b) Definition73.146.745.5
(c) Seed triggers73.753.849.6
(d) Event structure72.850.448.0
(e) Soft prompt68.148.2-
Majority voting of (a-e)73.952.148.7
(f) APEX Prompt74.957.451.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.24, + 0.489, + 0.286 + ], + "angle": 0, + "content": "Table 1: Results of event detection (ED) on ACE05 (F1-score, %)* indicates evaluation on our data set split based on the authors' public implementations." + }, + { + "type": "table", + "bbox": [ + 0.119, + 0.297, + 0.486, + 0.445 + ], + "angle": 0, + "content": "
MethodSEDFSEDZSED
Previous SOTA59.4(Lu et al., 2021)33.0*(Lai et al., 2020b)41.2*(Zhang et al., 2021b)
(a) Event type Name58.244.840.5
(b) Definition57.944.240.4
(c) Seed triggers60.450.446.2
(d) Event structure59.148.548.7
(e) Soft prompt55.641.7-
Majority voting of (a-e)60.247.945.6
(f) APEX Prompt63.452.648.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.455, + 0.49, + 0.484 + ], + "angle": 0, + "content": "Table 2: Results of event detection (ED) on ERE (F1-score, %)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.51, + 0.489, + 0.653 + ], + "angle": 0, + "content": "Chen et al., 2021) follow different data splits and settings, making it hard for a fair comparison. Considering the research goals of FSED and ZSED, we define the following conditions to split the ACE and ERE datasets: (i) The base event types and novel event types should be disjoint except Other. (ii) Each base or novel event type should contain at least 15 instances. (iii) The training set should contain sufficient annotated event mentions." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.655, + 0.49, + 0.849 + ], + "angle": 0, + "content": "To meet the above conditions, for ACE, we define the event types of 5 main event categories: Business, Contact, Conflict, Justice and Movement as the base event types, and types of the remaining 3 main categories: Life, Personnel and Transaction as the novel event types. In total, there are 18 qualified base types and 10 qualified novel types (the others do not satisfy the second condition). For ERE, we use the exact same 10 novel event types as ACE, and the rest 25 types as base event types. Detailed data and hyperparameter descriptions are in Appendix D and Appendix E." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.861, + 0.347, + 0.877 + ], + "angle": 0, + "content": "6 Results and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.888, + 0.489, + 0.919 + ], + "angle": 0, + "content": "Overall Results The experimental results for SED, FSED, and ZSED on ACE05, ERE, and" + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.082, + 0.883, + 0.22 + ], + "angle": 0, + "content": "
MethodSEDFSEDZSED
Previous SOTA68.5\n(Wang et al., 2021b)57.0\n(Chen et al., 2021)40.2*\n(Zhang et al., 2021b)
(a) Event type name68.863.458.8
(b) Definition67.156.952.9
(c) Seed triggers68.765.159.1
(e) Soft prompt64.538.6-
Majority voting of (a-e)68.463.458.1
(f) APEX Prompt68.868.459.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.508, + 0.229, + 0.885, + 0.288 + ], + "angle": 0, + "content": "Table 3: Results of event detection (ED) on MAVEN (F1-score, %). Event type structure prompts are not applicable to MAVEN as it does not contain any predefined argument roles." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.313, + 0.885, + 0.684 + ], + "angle": 0, + "content": "MAVEN are shown in Table 1-3, from which we see that (1) the APEX prompt achieves the best performance among all the forms of prompts under all the settings of the three benchmark datasets. Compared with the previous state of the art, the APEX prompt shows up to \\(4\\%\\) F-score gain for SED (on ERE), \\(22.2\\%\\) F-score gain for FSED (on ACE), and \\(19.7\\%\\) F-score gain for ZSED (on MAVEN); (2) All the forms of prompts provide significant improvement for FSED and ZSED, demonstrating the benefit of leveraging the semantics of event types via various forms of prompts. (3) Except APEX, seed triggers provide more improvements than other forms of event type prompts under most settings, suggesting its potential to represent the semantics of event types accurately. (4) Continuous soft prompt does not provide comparable performance as other forms of event type representations, which proves the necessity of leveraging event type specific prior knowledge to the representations; (5) The majority voting does not show improvement over individual prompts since each prompt captures a particular aspect of the event type semantics." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.695, + 0.887, + 0.92 + ], + "angle": 0, + "content": "Supervised Event Detection By carefully investigating the event mentions that are correctly detected by the APEX prompt while missed by other prompts, we find that the APEX prompt is more effective in detecting two types of event mentions: homonyms (multiple-meaning words) and intricate words. General homonyms are usually hard to be detected as event mentions as they usually have dozens of meanings in different contexts. For example, consider the following two examples: (i) Airlines are getting [Transport:Movement] flyers to destinations on time more often. (ii) If the board cannot vote to give [Transaction:Transfer-Money'] themselves present money. Here, \"get\" and \"give\"" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1289" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.235, + 0.086, + 0.761, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.275, + 0.884, + 0.307 + ], + "angle": 0, + "content": "Figure 2: F-score distribution of all novel types based on various event type prompts under the few-shot event detection setting on ACE (Best view in color)" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.319, + 0.49, + 0.626 + ], + "angle": 0, + "content": "are not detected based on the event type name or seed triggers but are correctly identified by the definition and APEX prompts. The definition and APEX prompts make \\(10\\%\\) and \\(7\\%\\) fewer false predictions than seed triggers on general homonyms. For intricate words, their semantics usually cannot be captured with an individual prompt. In the following two examples: (i) It is reasonable, however, to reimburse board members for legitimate expenses (ii) ... ever having discussed being compensated by the board in the future ... \"reimburse\" and \"compensated\" indicate sophisticated meaning of Transaction:Transfer-Money, which may not be captured by prompts, such as seed triggers. With the event definition and the argument roles in the APEX prompt, the highly correlated contexts, such as \"board members\" and \"legitimate expenses\", can help the model correctly detect reimburse as an event mention of Transaction:Transfer-Money." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.646, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Few-shot Event Detection Figure 2 shows the F-score distribution of all novel types based on various forms of event type prompts, from which we observe that: (1) The event type name, seed triggers, and APEX prompt generally perform better than definition and structure, as they carry more straightforward semantics of event types. (2) Event type name based prompts show lower performance on Personnel:End-Position, Personnel:Start-Position and Transaction:Transfer-Money than other event types, as the semantics of these event type names are less indicative than other event types. (3) Seed trigger based prompts perform worse than event type name and APEX prompts on two event types, Life:injure and Life:die, probably because the prototype seed triggers are not properly selected. (4) The structure based prompt outperforms the other" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.319, + 0.885, + 0.496 + ], + "angle": 0, + "content": "prompts on Life:Injure as Life:Injure events require the existence of a person or victim. (5) APEX prompt shows consistently (almost) best performance on all the event types because it combines all the information of other prompts. (6) We also observe that the performance of Life:Be-Born, Life:Die, Life:Marry, and Personnel:Elect based on various forms of prompts are consistently better than the other types as the intrinsic semantics of those types the corresponding event triggers are concentrated." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.508, + 0.884, + 0.701 + ], + "angle": 0, + "content": "Zero-shot Event Detection The proposed prompt-based method is more affordable to be generalized compared with the prior state-of-the-art zero-shot approach (Zhang et al., 2021b). The average length of created APEX prompts is less than 20 tokens. Thus manually creating them will not take much human effort. On the contrary, Zhang et al. (2021b) requires an extensive collection of anchor sentences to perform zero-shot event detection, e.g., 4,556,237 anchor sentences for ACE and ERE. This process is time-consuming and expensive." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.715, + 0.642, + 0.731 + ], + "angle": 0, + "content": "7 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.743, + 0.885, + 0.919 + ], + "angle": 0, + "content": "We investigate a variety of prompts to represent the semantics of event types, and leverage them with a unified framework for supervised, few-shot and zero-shot event detection. Experimental results demonstrate that, a well-defined and comprehensive description of event types can significantly improve the performance of event detection, especially when the annotations are limited (few-shot event detection) or even not available (zero-shot event detection), with up to \\(22.2\\%\\) F-score gain over the prior state of the art." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.928, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1290" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.115, + 0.085, + 0.221, + 0.099 + ], + "angle": 0, + "content": "Limitations" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.111, + 0.49, + 0.431 + ], + "angle": 0, + "content": "We have demonstrated that an accurate description can perform better for both supervised and weakly supervised event detection. However, the event types from most existing ontologies are not properly defined. For example, in ACE annotation guideline (Linguistic Data Consortium, 2005), transfer-money is defined as \"giving, receiving, borrowing, or lending money when it is not in the context of purchasing something\". However, it is hard for the model to interpret it accurately, especially the constraints \"not in the context of purchasing something\". In addition, many event types from MAVEN, e.g., Achieve, Award, and Incident, are not associated with any definitions. A potential future research direction is to leverage mining-based approaches or state-of-the-art generators to automatically generate a comprehensive event type description based on various sources, such as annotation guidelines, example annotations, and external knowledge bases." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.445, + 0.279, + 0.46 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.471, + 0.489, + 0.534 + ], + "angle": 0, + "content": "We thank the anonymous reviewers and area chair for their valuable time and constructive comments. This research is based upon work supported by the Amazon Research Award." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.562, + 0.214, + 0.577 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.585, + 0.489, + 0.638 + ], + "angle": 0, + "content": "Ashutosh Agarwal, Anay Majee, Anbumani Subramanian, and Chetan Arora. 2021. Attention guided cosine margin for overcoming class-imbalance in few-shot road object detection." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.649, + 0.488, + 0.689 + ], + "angle": 0, + "content": "David Ahn. 2006. The stages of event extraction. In Proceedings of the Workshop on Annotating and Reasoning about Time and Events, pages 1-8." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.7, + 0.489, + 0.766 + ], + "angle": 0, + "content": "Collin F Baker, Charles J Fillmore, and John B Lowe. 1998. The berkeley framenet project. In 36th Annual Meeting of the Association for Computational Linguistics and 17th International Conference on Computational Linguistics, Volume 1, pages 86-90." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.776, + 0.488, + 0.881 + ], + "angle": 0, + "content": "Laura Banarescu, Claire Bonial, Shu Cai, Madalina Georgescu, Kira Griffith, Ulf Hermjakob, Kevin Knight, Philipp Koehn, Martha Palmer, and Nathan Schneider. 2013. Abstract Meaning Representation for sembanking. In Proceedings of the 7th Linguistic Annotation Workshop and Interoperability with Discourse, pages 178-186, Sofia, Bulgaria. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.892, + 0.488, + 0.919 + ], + "angle": 0, + "content": "Ofer Bronstein, Ido Dagan, Qi Li, Heng Ji, and Anette Frank. 2015. Seed-based event trigger labeling: How" + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.585, + 0.489, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.529, + 0.086, + 0.883, + 0.152 + ], + "angle": 0, + "content": "far can event descriptions get us? In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 372-376." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.163, + 0.885, + 0.345 + ], + "angle": 0, + "content": "Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems, volume 33, pages 1877-1901. Curran Associates, Inc." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.356, + 0.884, + 0.435 + ], + "angle": 0, + "content": "Kai Cao, Xiang Li, Miao Fan, and Ralph Grishman. 2015. Improving event detection with active learning. In Proceedings of the International Conference on Recent Advances in Natural Language Processing, pages 72-77, Hissar, Bulgaria. INCOMA Ltd. Shoumen, BULGARIA." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.446, + 0.883, + 0.485 + ], + "angle": 0, + "content": "Jiawei Chen, Hongyu Lin, Xianpei Han, and Le Sun. 2021. Honey or poison? solving the trigger curse in few-shot event detection via causal intervention." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.496, + 0.884, + 0.589 + ], + "angle": 0, + "content": "Yubo Chen, Liheng Xu, Kang Liu, Daojian Zeng, and Jun Zhao. 2015. Event extraction via dynamic multipooling convolutional neural networks. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 167-176." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.599, + 0.883, + 0.678 + ], + "angle": 0, + "content": "Yunmo Chen, Tongfei Chen, Seth Ebner, Aaron Steven White, and Benjamin Van Durme. 2020. Reading the manual: Event extraction as definition comprehension. In Proceedings of the Fourth Workshop on Structured Prediction for NLP, pages 74-83, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.688, + 0.883, + 0.742 + ], + "angle": 0, + "content": "Nancy Chinchor and Elaine Marsh. 1998. Muc-7 information extraction task definition. In Proceeding of the seventh message understanding conference (MUC-7), Appendices, pages 359-367." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.752, + 0.883, + 0.805 + ], + "angle": 0, + "content": "Arkabandhu Chowdhury, Mingchao Jiang, and Chris Jermaine. 2021. Few-shot image classification: Just use a library of pre-trained feature extractors and a simple classifier. abs/2101.00562." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.815, + 0.883, + 0.881 + ], + "angle": 0, + "content": "Xin Cong, Shiyao Cui, Bowen Yu, Tingwen Liu, Yubin Wang, and Bin Wang. 2021. Few-shot event detection with prototypical amortized conditional random field. In Findings of the Association for Computational Linguistics: ACL-IJCNLP." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.892, + 0.883, + 0.919 + ], + "angle": 0, + "content": "Bhavana Dalvi, William W. Cohen, and Jamie Callan. 2012. Websets: extracting sets of entities from" + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.885, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.519, + 0.941 + ], + "angle": 0, + "content": "1291" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.135, + 0.086, + 0.489, + 0.113 + ], + "angle": 0, + "content": "the web using unsupervised information extraction. ArXiv, abs/1307.0261." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.124, + 0.487, + 0.203 + ], + "angle": 0, + "content": "Shumin Deng, Ningyu Zhang, Jiaojian Kang, Yichi Zhang, Wei Zhang, and Huajun Chen. 2020. Meta-learning with dynamic-memory-based prototypical network for few-shot event detection. Proceedings of the 13th International Conference on Web Search and Data Mining." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.213, + 0.487, + 0.331 + ], + "angle": 0, + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.342, + 0.487, + 0.407 + ], + "angle": 0, + "content": "Xinya Du and Claire Cardie. 2020. Event extraction by answering (almost) natural questions. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 671-683, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.418, + 0.487, + 0.511 + ], + "angle": 0, + "content": "Xiaocheng Feng, Lifu Huang, Duyu Tang, Heng Ji, Bing Qin, and Ting Liu. 2016. A language-independent neural network for event detection. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 66-71, Berlin, Germany. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.521, + 0.487, + 0.574 + ], + "angle": 0, + "content": "Ralph Grishman. 1997. Information extraction: Techniques and challenges. In International summer school on information extraction, pages 10-27. Springer." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.584, + 0.487, + 0.689 + ], + "angle": 0, + "content": "Karen Hambardzumyan, Hrant Khachatrian, and Jonathan May. 2021. WARP: Word-level Adversarial ReProgramming. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4921-4933, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.7, + 0.487, + 0.779 + ], + "angle": 0, + "content": "Lifu Huang, Taylor Cassidy, Xiaocheng Feng, Heng Ji, Clare Voss, Jiawei Han, and Avirup Sil. 2016. Liberal event extraction and event schema induction. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 258-268." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.79, + 0.487, + 0.854 + ], + "angle": 0, + "content": "Lifu Huang and Heng Ji. 2020. Semi-supervised new event type induction and event detection. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 718-724." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.866, + 0.487, + 0.919 + ], + "angle": 0, + "content": "Lifu Huang, Heng Ji, Kyunghyun Cho, Ido Dagan, Sebastian Riedel, and Clare Voss. 2018. Zero-shot transfer learning for event extraction. In Proceedings of the 56th Annual Meeting of the Association for" + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.489, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.53, + 0.086, + 0.882, + 0.126 + ], + "angle": 0, + "content": "Computational Linguistics (Volume 1: Long Papers), pages 2160-2170, Melbourne, Australia. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.136, + 0.882, + 0.175 + ], + "angle": 0, + "content": "Heng Ji and Ralph Grishman. 2008. Refining event extraction through cross-document inference. In Proceedings of ACL-08: Hlt, pages 254-262." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.186, + 0.882, + 0.238 + ], + "angle": 0, + "content": "Zhengbao Jiang, Frank F. Xu, J. Araki, and Graham Neubig. 2020. How can we know what language models know? Transactions of the Association for Computational Linguistics, 8:423-438." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.248, + 0.882, + 0.313 + ], + "angle": 0, + "content": "Bingyi Kang, Zhuang Liu, Xin Wang, Fisher Yu, Jiashi Feng, and Trevor Darrell. 2019. Few-shot object detection via feature reweighting. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 8419-8428." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.324, + 0.882, + 0.39 + ], + "angle": 0, + "content": "Viet Dac Lai, Franck Dernoncourt, and Thien Huu Nguyen. 2020a. Exploiting the matching information in the support set for few shot event classification. Pacific-Asia Conference on Knowledge Discovery and Data Mining, page 233-245." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.399, + 0.882, + 0.439 + ], + "angle": 0, + "content": "Viet Dac Lai and Thien Huu Nguyen. 2019. Extending event detection to new types with learning from keywords. arXiv preprint arXiv:1910.11368." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.449, + 0.882, + 0.527 + ], + "angle": 0, + "content": "Viet Dac Lai, Thien Huu Nguyen, and Franck Dernoncourt. 2020b. Extensively matching for few-shot learning event detection. In Proceedings of the First Joint Workshop on Narrative Understanding, Storylines, and Events, pages 38-45, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.538, + 0.882, + 0.577 + ], + "angle": 0, + "content": "Brian Lester, Rami Al-Rfou, and Noah Constant. 2021. The power of scale for parameter-efficient prompt tuning. In EMNLP." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.587, + 0.882, + 0.653 + ], + "angle": 0, + "content": "Bohao Li, Boyu Yang, Chang Liu, Feng Liu, Rongrong Ji, and Qixiang Ye. 2021. Beyond max-margin: Class margin equilibrium for few-shot object detection. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7359-7368." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.663, + 0.882, + 0.741 + ], + "angle": 0, + "content": "Fayuan Li, Weihua Peng, Yuguang Chen, Quan Wang, Lu Pan, Yajuan Lyu, and Yong Zhu. 2020. Event extraction as multi-turn question answering. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 829–838, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.752, + 0.882, + 0.83 + ], + "angle": 0, + "content": "Qi Li, Heng Ji, and Liang Huang. 2013. Joint event extraction via structured prediction with global features. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 73-82, Sofia, Bulgaria. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.84, + 0.882, + 0.919 + ], + "angle": 0, + "content": "Xiang Lisa Li and Percy Liang. 2021. Prefix-tuning: Optimizing continuous prompts for generation. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), abs/2101.00190." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.882, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1292" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.153 + ], + "angle": 0, + "content": "Shasha Liao and Ralph Grishman. 2010. Using document level cross-event inference to improve event extraction. In Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 789-797." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.164, + 0.489, + 0.243 + ], + "angle": 0, + "content": "Ying Lin, Heng Ji, Fei Huang, and Lingfei Wu. 2020. A joint neural model for information extraction with global features. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7999-8009, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.255, + 0.49, + 0.321 + ], + "angle": 0, + "content": "Linguistic Data Consortium. 2005. English annotation guidelines for events. https://www.ldc.upenn.edu/sites/www.ldc.upenn.edu/files/english-events-guidelines-v5.4.3.pdf." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.333, + 0.489, + 0.412 + ], + "angle": 0, + "content": "Jian Liu, Yubo Chen, Kang Liu, Wei Bi, and Xiaojiang Liu. 2020. Event extraction as machine reading comprehension. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1641-1651, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.424, + 0.489, + 0.554 + ], + "angle": 0, + "content": "Yaojie Lu, Hongyu Lin, Jin Xu, Xianpei Han, Jialong Tang, Annan Li, Le Sun, Meng Liao, and Shaoyi Chen. 2021. Text2Event: Controllable sequence-to-structure generation for end-to-end event extraction. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 2795-2806, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.567, + 0.489, + 0.672 + ], + "angle": 0, + "content": "Qing Lyu, Hongming Zhang, Elior Sulem, and Dan Roth. 2021. Zero-shot Event Extraction via Transfer Learning: Challenges and Insights. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 322-332, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.684, + 0.489, + 0.75 + ], + "angle": 0, + "content": "David McClosky, Mihai Surdeanu, and Christopher D Manning. 2011. Event extraction as dependency parsing. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 1626-1635." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.762, + 0.489, + 0.814 + ], + "angle": 0, + "content": "Minh Van Nguyen, Viet Dac Lai, and Thien Huu Nguyen. 2021. Cross-task instance representation interactions and label dependencies for joint information extraction with graph convolutional networks." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.826, + 0.489, + 0.919 + ], + "angle": 0, + "content": "Thien Huu Nguyen, Kyunghyun Cho, and Ralph Grishman. 2016. Joint event extraction via recurrent neural networks. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 300-309, San Diego, California. Association for Computational Linguistics." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.086, + 0.883, + 0.152 + ], + "angle": 0, + "content": "Panupong Pasupat and Percy Liang. 2014. Zero-shot entity extraction from web pages. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 391-401." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.163, + 0.883, + 0.281 + ], + "angle": 0, + "content": "Fabio Petroni, Tim Rocttäschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, and Alexander Miller. 2019. Language models as knowledge bases? In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2463-2473, Hong Kong, China. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.292, + 0.883, + 0.356 + ], + "angle": 0, + "content": "Colin Raffel, Noam M. Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.368, + 0.883, + 0.395 + ], + "angle": 0, + "content": "Timo Schick and Hinrich Schütze. 2021a. Few-shot text generation with pattern-exploiting training." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.405, + 0.883, + 0.471 + ], + "angle": 0, + "content": "Timo Schick and Hinrich Schütze. 2021b. It's not just size that matters: Small language models are also few-shot learners. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics, pages 2339-2352." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.482, + 0.883, + 0.624 + ], + "angle": 0, + "content": "Shirong Shen, Tongtong Wu, Guilin Qi, Yuan-Fang Li, Gholamreza Haffari, and Sheng Bi. 2021. Adaptive knowledge-enhanced bayesian meta-learning for few-shot event detection. In Findings of the Association for Computational Linguistics, page 2417-2429. Association for Computational Linguistics (ACL). Annual Meeting of the Association for Computational Linguistics and the International Joint Conference on Natural Language Processing 2021, ACL-IJCNLP 2021; Conference date: 01-08-2021 Through 06-08-2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.636, + 0.883, + 0.702 + ], + "angle": 0, + "content": "Taylor Shin, Yasaman Razeghi, Robert L. Logan IV, Eric Wallace, and Sameer Singh. 2020. AutoPrompt: Eliciting knowledge from language models with automatically generated prompts. In Empirical Methods in Natural Language Processing (EMNLP)." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.713, + 0.883, + 0.805 + ], + "angle": 0, + "content": "Zhiyi Song, Ann Bies, Stephanie Strassel, Tom Riese, Justin Mott, Joe Ellis, Jonathan Wright, Seth Kulick, Neville Ryant, and Xiaoyi Ma. 2015. From light to rich ere: annotation of entities, relations, and events. In Proceedings of the 3rd Workshop on EVENTS: Definition, Detection, Coreference, and Representation, pages 89-98." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.816, + 0.883, + 0.881 + ], + "angle": 0, + "content": "Bo Sun, Banghuai Li, Shengcai Cai, Ye Yuan, and Chi Zhang. 2021. Fsce: Few-shot object detection via contrastive proposal encoding. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7348-7358." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.892, + 0.883, + 0.919 + ], + "angle": 0, + "content": "David Wadden, Ulme Wennberg, Yi Luan, and Hannaneh Hajishirzi. 2019. Entity, relation, and event" + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.883, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1293" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.135, + 0.086, + 0.49, + 0.179 + ], + "angle": 0, + "content": "extraction with contextualized span representations. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5784-5789, Hong Kong, China. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.19, + 0.49, + 0.256 + ], + "angle": 0, + "content": "Richard C Wang and William Cohen. 2009. Character-level analysis of semi-structured documents for set expansion. In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 1503–1512." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.268, + 0.488, + 0.322 + ], + "angle": 0, + "content": "Sijia Wang, Mo Yu, Shiyu Chang, Lichao Sun, and Lifu Huang. 2021a. Query and extract: Refining event extraction as type-oriented binary decoding. arXiv preprint arXiv:2110.07476." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.333, + 0.488, + 0.4 + ], + "angle": 0, + "content": "Xiao Wang, Houye Ji, Chuan Shi, Bai Wang, Yanfang Ye, Peng Cui, and Philip S Yu. 2019. Heterogeneous graph attention network. In The World Wide Web Conference, WWW '19, page 2022-2032, New York, NY, USA. Association for Computing Machinery." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.411, + 0.488, + 0.476 + ], + "angle": 0, + "content": "Xiaozhi Wang, Ziqi Wang, Xu Han, Wangyi Jiang, Rong Han, Zhiyuan Liu, Juanzi Li, Peng Li, Yankai Lin, and Jie Zhou. 2020a. MAVEN: A massive general domain event detection dataset. In Proceedings of EMNLP 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.489, + 0.488, + 0.529 + ], + "angle": 0, + "content": "Xin Wang, Thomas E. Huang, Trevor Darrell, Joseph E Gonzalez, and Fisher Yu. 2020b. Frustratingly simple few-shot object detection." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.541, + 0.488, + 0.62 + ], + "angle": 0, + "content": "Ziqi Wang, Xiaozhi Wang, Xu Han, Yankai Lin, Lei Hou, Zhiyuan Liu, Peng Li, Juanzi Li, and Jie Zhou. 2021b. CLEVE: Contrastive Pre-training for Event Extraction. In Proceedings of ACL-IJCNLP, pages 6283-6297, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.632, + 0.488, + 0.671 + ], + "angle": 0, + "content": "Yang Xiao and Renaud Marlet. 2020. Few-shot object detection and viewpoint estimation for objects in the wild. In ECCV." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.684, + 0.488, + 0.75 + ], + "angle": 0, + "content": "Xiaopeng Yan, Ziliang Chen, Anni Xu, Xiaoxi Wang, Xiaodan Liang, and Liang Lin. 2019. Meta r-cnn: Towards general solver for instance-level low-shot learning. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9576-9585." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.762, + 0.488, + 0.854 + ], + "angle": 0, + "content": "Bishan Yang and Tom M. Mitchell. 2016. Joint extraction of events and entities within a document context. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 289-299, San Diego, California. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.866, + 0.488, + 0.918 + ], + "angle": 0, + "content": "Weizhe Yuan, Graham Neubig, and Pengfei Liu. 2021. BARTScore: Evaluating generated text as text generation. In Advances in Neural Information Processing Systems." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.086, + 0.882, + 0.164 + ], + "angle": 0, + "content": "Gongjie Zhang, Kaiwen Cui, Rongliang Wu, Shijian Lu, and Yonghong Tian. 2021a. Pnpdet: Efficient few-shot detection without forgetting via plug-and-play sub-networks. 2021 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 3822-3831." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.175, + 0.882, + 0.254 + ], + "angle": 0, + "content": "Hongming Zhang, Haoyu Wang, and Dan Roth. 2021b. Zero-shot Label-aware Event Trigger and Argument Classification. In *Findings of the Association for Computational Linguistics: ACL-IJCNLP* 2021, pages 1331-1340, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.264, + 0.882, + 0.343 + ], + "angle": 0, + "content": "Tongtao Zhang, Spencer Whitehead, Hanwang Zhang, Hongzhi Li, Joseph Ellis, Lifu Huang, Wei Liu, Heng Ji, and Shih-Fu Chang. 2017. Improving event extraction via multimodal integration. In Proceedings of the 25th ACM international conference on Multimedia, pages 270-278." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.882, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1294" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.114, + 0.084, + 0.44, + 0.101 + ], + "angle": 0, + "content": "A APEX prompt examples for ACE" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.11, + 0.486, + 0.142 + ], + "angle": 0, + "content": "Table 4 and Table 5 show APEX prompt examples for ACE events." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.155, + 0.36, + 0.172 + ], + "angle": 0, + "content": "B Prompt Token Selection" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.182, + 0.49, + 0.44 + ], + "angle": 0, + "content": "In our experiments, the event type names and event type structures are automatically extracted from the target event ontology, such as ACE (Linguistic Data Consortium, 2005), ERE (Song et al., 2015) and MAVEN (Wang et al., 2020a). The prototype seed triggers are automatically selected from the annotated data for supervised and few-shot event extraction. For zero-shot event extraction, we manually select \\( R \\) words from the NLTK synonyms of each event type as its prototype seed triggers. The definitions and APEX prompts are based on the official annotation guides for each target event ontology (Linguistic Data Consortium, 2005; Song et al., 2015; Wang et al., 2020a) and the available definitions in FrameNet (Baker et al., 1998) with manual editing." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.452, + 0.312, + 0.469 + ], + "angle": 0, + "content": "C Learning Strategy" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.478, + 0.49, + 0.833 + ], + "angle": 0, + "content": "The learning strategy varies for supervised, few-shot, and zero-shot learning. For supervised learning, we optimize the following objective for event trigger detection \\(\\mathcal{L} = -\\frac{1}{|\\mathcal{T}||\\mathcal{N}|}\\sum_{t\\in \\mathcal{T}}\\sum_{i = 1}^{|\\mathcal{N}|}\\boldsymbol{y}_i^t\\) \\(\\log \\tilde{\\boldsymbol{y}}_i^t\\) where \\(\\mathcal{T}\\) is the set of target event types and \\(\\mathcal{N}\\) is the set of tokens from the training dataset. \\(\\boldsymbol{y}_i^t\\) denotes the ground truth label vector. For few-shot event detection, we optimize the model on both base training data set and the smaller training data set for novel event types: \\(\\mathcal{L} = -\\frac{1}{|\\mathcal{T}^B||\\mathcal{N}^B|}\\sum_{t\\in \\mathcal{T}^B}\\sum_{i = 1}^{|\\mathcal{N}^B|}\\boldsymbol{y}_i^t\\cdot \\log \\tilde{\\boldsymbol{y}}_i^t -\\) \\(\\beta \\frac{1}{|\\mathcal{T}^N||\\mathcal{N}^N|}\\sum_{t\\in \\mathcal{T}^N}\\sum_{i = 1}^{|\\mathcal{N}^N|}\\boldsymbol{y}_i^t\\cdot \\log \\tilde{\\boldsymbol{y}}_i^t\\) , where \\(\\mathcal{T}^B\\) and \\(\\mathcal{N}^B\\) denote the set of base event types and tokens from the base training data set, respectively. \\(\\mathcal{T}^N\\) is the set of novel event types. \\(\\mathcal{N}^N\\) is the set of tokens from the training data set for novel event types. \\(\\beta\\) is a hyper-parameter to balance the two objectives. For zero-shot event detection, as we only have the base training data set, we minimize the following objective: \\(\\mathcal{L} = -\\frac{1}{|\\mathcal{T}^B||\\mathcal{N}^B|}\\sum_{t\\in \\mathcal{T}^B}\\sum_{i = 1}^{|\\mathcal{N}^B|}\\boldsymbol{y}_i^t\\) \\(\\log \\tilde{\\boldsymbol{y}}_i^t\\)" + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.845, + 0.221, + 0.86 + ], + "angle": 0, + "content": "D Dataset" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.872, + 0.489, + 0.919 + ], + "angle": 0, + "content": "After defining the base and novel event types, we create the training, validation, and evaluation split for all three datasets. We use the sentences with" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.375 + ], + "angle": 0, + "content": "only base event type mentions as the base training data set for few-shot event detection, and randomly select 10 sentences with novel event type mentions as the additional smaller training data set. We use the sentences with both base and novel event type mentions as the development set and use the remaining sentences with only novel event type mentions as the evaluation dataset. We use the same development and evaluation set as few-shot event detection for zero-shot event detection and remove the instances with novel event mentions from the training set. We randomly split the sentences without any event annotations proportionally to the number of sentences with event mentions in each set for both zero-shot and few-shot event detection. Table 6 shows the detailed data statistics for all the three datasets under the few-shot and zero-shot event extraction settings." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.386, + 0.838, + 0.403 + ], + "angle": 0, + "content": "E Hyperparameters and Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.412, + 0.885, + 0.636 + ], + "angle": 0, + "content": "For a fair comparison with the previous baseline approaches, we use the same pre-trained bert-large-uncased model for fine-tuning and optimizing our model with BertAdam. For supervised event detection, we optimize the parameters with grid search: training epoch is 3, learning rate \\(\\in [3e - 6,1e - 4]\\), training batch size \\(\\in\\) {8, 12, 16, 24, 32}, dropout rate \\(\\in\\) {0.4, 0.5, 0.6}. The running time is up to 3 hours on one Quadro RTX 8000. For evaluation, we use the same criteria as previous studies (Li et al., 2013; Chen et al., 2015; Nguyen et al., 2016; Lin et al., 2020): an event mention is correct if its span and event type match a reference event mention." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1295" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.167, + 0.136, + 0.833, + 0.839 + ], + "angle": 0, + "content": "
Event Rep TypeComprehensive Prompt
Business:Declare-BankruptcyDeclare Bankruptcy [SEP] bankruptcy bankruptciesbankrupting [SEP] Organizationrequest legal protection from debt collection at a Place
Business:End-OrgEnd Organization [SEP] dissolving disbanded [SEP] an Organization goes out of business at a Place
Business:Merge-OrgMerge Organization [SEP] merging merger [SEP] two or more Organizations come together to form a new organization at a Place
Business:Start-OrgStart Organization [SEP] founded [SEP] an Agent create a new Organization at a Place
Conflict:AttackAttack [SEP] invaded airstrikes overthrew ambushed [SEP] An Attacker physically attacks a Target with Instrument at a Place
Conflict:DemonstrateDemonstrate [SEP] demonstrations protest strikes riots [SEP] Entities come together in a Place to protest or demand official action
Contact:MeetMeet [SEP] reunited retreats [SEP] two or more Entities come together at same Place and interact in person
Contact:Phone-WritePhone Write [SEP] emailed letter [SEP] phone or written communication between two or more Entities
Justice:AcquitAcquit [SEP] acquitted [SEP] a trial of Defendant ends but Adjudicator fails to produce a conviction at a Place
Justice:AppealAppeal [SEP] appeal [SEP] the decision for Defendant of a court is taken to a higher court for Adjudicator review with Prosecutor
Justice:Arrest-JailArrest Jail [SEP] arrested locked [SEP] the Agent takes custody of a Person at a Place
Justice:Charge-IndictCharge Indict [SEP] indictment [SEP] a Defendant is accused of a crime by a Prosecutor for Adjudicator
Justice:ConvictConvict [SEP] pled guilty convicting [SEP] an Defendant found guilty of a crime by Adjudicator at a Place
Justice:ExecuteExecute [SEP] death [SEP] the life of a Person is taken by an Agent at a Place
Justice:ExtraditeExtradite [SEP] extradition [SEP] a Person is sent by an Agent from Origin to Destination
Justice:FineFine [SEP] payouts financial punishment [SEP] a Adjudicator issues a financial punishment Money to an Entity at a Place
Justice:PardonPardon [SEP] pardoned lift sentence [SEP] an Adjudicator lifts a sentence of Defendant at a Place
Justice:Release-ParoleRelease Parole [SEP] parole [SEP] an Entity ends its custody of a Person at a Place
Justice:SentenceSentence [SEP] sentenced punishment [SEP] the punishment for the defendant is issued by a state actor
Justice:SueSue [SEP] lawsuits [SEP] Plaintiff initiate a court proceeding to determine the liability of a Defendant judge by Adjudicator at a Place
Justice:Trial-HearingTrial Hearing [SEP] trial hearings [SEP] a court proceeding initiated to determine the guilty or innocence of a Person with Prosecutor and Adjudicator at a Place
Life:Be-BornBe Born [SEP] childbirth [SEP] a Person is born at a Place
Life:DieDie [SEP] deceased extermination [SEP] life of a Victim ends by an Agent with Instrument at a Place
" + }, + { + "type": "table_caption", + "bbox": [ + 0.34, + 0.846, + 0.657, + 0.861 + ], + "angle": 0, + "content": "Table 4: APEX templates for ACE event types" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1296" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.167, + 0.156, + 0.833, + 0.472 + ], + "angle": 0, + "content": "
Event Rep TypeComprehensive Prompt
Life:DivorceDivorce [SEP] people divorce [SEP] two Person are officially divorced at a place
Life:InjureInjure [SEP] hospitalised paralyzed dismember [SEP] a Victim experiences physical harm from Agent with Instrument at a Place
Life:MarryMarry [SEP] married marriage marry [SEP] two Person are married at a Place
Movement:TransportTransport [SEP] arrival travels penetrated expelled [SEP] an Agent moves an Artifact from Origin to Destination with Vehicle at Price
Personnel:ElectElect [SEP] reelected elected election [SEP] a candidate Person wins an election by voting Entity at a Place
Personnel:End-PositionEnd Position [SEP] resigning retired resigned [SEP] a Person stops working for an Entity or change office at a Place
Personnel:NominateNominate [SEP] nominate [SEP] a Person is nominated for a new position by another Agent at a Place
Personnel:Start-PositionStart Position [SEP] hiring rehired recruited [SEP] a Person begins working for an Entity or change office at a Place
Transaction:Transfer-MoneyTransfer Money [SEP] donations reimbursing deductions [SEP] transfer Money from the Giver to the Beneficiary or Recipient at a Place
Transaction:Transfer-OwnershipTransfer Ownership [SEP] purchased buy sell loan [SEP] buying selling loaning borrowing giving receiving of Artifacts from Seller to Buyer or Beneficiary at a Place at Price
" + }, + { + "type": "table_caption", + "bbox": [ + 0.3, + 0.482, + 0.698, + 0.497 + ], + "angle": 0, + "content": "Table 5: APEX templates for ACE event types (continued)" + }, + { + "type": "table", + "bbox": [ + 0.267, + 0.654, + 0.734, + 0.817 + ], + "angle": 0, + "content": "
DatasetACE05-E+ERE-ENMAVEN
# TypesBase1825120
Novel101045
# MentionsBase3572544993675
Novel172431833201
TrainFew-shot3216388688085
Zero-shot3116378687635
Validation900(51%/49%)2797(53%/47%)3883(71%/23%)
Evaluation119520121652
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.826, + 0.881, + 0.841 + ], + "angle": 0, + "content": "Table 6: Data statistics for ACE2005, ERE and MAVEN datasets under few-shot/zero-shot event detection settings." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1297" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.133, + 0.085, + 0.435, + 0.1 + ], + "angle": 0, + "content": "ACL 2023 Responsible NLP Checklist" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.108, + 0.323, + 0.123 + ], + "angle": 0, + "content": "A For every submission:" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.127, + 0.533, + 0.158 + ], + "angle": 0, + "content": "A1. Did you describe the limitations of your work? Section 8" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.169, + 0.553, + 0.201 + ], + "angle": 0, + "content": "A2. Did you discuss any potential risks of your work? Section 8" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.213, + 0.697, + 0.245 + ], + "angle": 0, + "content": "A3. Do the abstract and introduction summarize the paper's main claims? Abstract and Section 1" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.256, + 0.669, + 0.289 + ], + "angle": 0, + "content": "A4. Have you used AI writing assistants when working on this paper? Left blank." + }, + { + "type": "list", + "bbox": [ + 0.13, + 0.127, + 0.697, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.3, + 0.489, + 0.317 + ], + "angle": 0, + "content": "B Did you use or create scientific artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.322, + 0.33, + 0.337 + ], + "angle": 0, + "content": "Section 5 and Appendix C" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.347, + 0.53, + 0.38 + ], + "angle": 0, + "content": "B1. Did you cite the creators of artifacts you used? Section 5 and Appendix C" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.39, + 0.779, + 0.422 + ], + "angle": 0, + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Section 5" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.433, + 0.882, + 0.512 + ], + "angle": 0, + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Section 5" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.525, + 0.882, + 0.588 + ], + "angle": 0, + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Left blank." + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.599, + 0.882, + 0.647 + ], + "angle": 0, + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Section 5" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.658, + 0.882, + 0.755 + ], + "angle": 0, + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Appendix C" + }, + { + "type": "list", + "bbox": [ + 0.13, + 0.347, + 0.882, + 0.755 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.765, + 0.494, + 0.782 + ], + "angle": 0, + "content": "C Did you run computational experiments?" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.787, + 0.33, + 0.802 + ], + "angle": 0, + "content": "Section 5 and Appendix D" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.812, + 0.882, + 0.861 + ], + "angle": 0, + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Appendix D" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.869, + 0.878, + 0.893 + ], + "angle": 0, + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1298" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.13, + 0.084, + 0.88, + 0.116 + ], + "angle": 0, + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.118, + 0.348, + 0.133 + ], + "angle": 0, + "content": "Section 5 and Appendix D" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.143, + 0.884, + 0.191 + ], + "angle": 0, + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.193, + 0.243, + 0.208 + ], + "angle": 0, + "content": "Appendix D" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.218, + 0.884, + 0.266 + ], + "angle": 0, + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.268, + 0.241, + 0.283 + ], + "angle": 0, + "content": "Appendix B" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.294, + 0.878, + 0.31 + ], + "angle": 0, + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.315, + 0.215, + 0.33 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.341, + 0.883, + 0.373 + ], + "angle": 0, + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.375, + 0.249, + 0.39 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.4, + 0.884, + 0.448 + ], + "angle": 0, + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.45, + 0.25, + 0.464 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.476, + 0.884, + 0.523 + ], + "angle": 0, + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.525, + 0.25, + 0.54 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.55, + 0.875, + 0.566 + ], + "angle": 0, + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.568, + 0.249, + 0.582 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.593, + 0.881, + 0.624 + ], + "angle": 0, + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.627, + 0.25, + 0.642 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1299" + } + ] +] \ No newline at end of file diff --git a/2023/The Art of Prompting_ Event Detection based on Type Specific Prompts/5e1d8751-6f7a-44c9-99a0-fd4f3669ec72_origin.pdf b/2023/The Art of Prompting_ Event Detection based on Type Specific Prompts/5e1d8751-6f7a-44c9-99a0-fd4f3669ec72_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..573f13e526a09ef1d67f2ae383b3a241ee1cbe0d --- /dev/null +++ b/2023/The Art of Prompting_ Event Detection based on Type Specific Prompts/5e1d8751-6f7a-44c9-99a0-fd4f3669ec72_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4cb71e69e73e08e33df17a45f39a695ddc42fd74a826a08db94429ae194eee38 +size 635692 diff --git a/2023/The Art of Prompting_ Event Detection based on Type Specific Prompts/full.md b/2023/The Art of Prompting_ Event Detection based on Type Specific Prompts/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1a72be6324e5ae48305b901d94d5d84309e7c8c0 --- /dev/null +++ b/2023/The Art of Prompting_ Event Detection based on Type Specific Prompts/full.md @@ -0,0 +1,308 @@ +# The Art of Prompting: Event Detection based on Type Specific Prompts + +Sijia Wang*, Mo Yu*, Lifu Huang* + +$\spadesuit$ Virginia Tech, $\spadesuit$ WeChat AI + +$\clubsuit$ {sijiawang, lifuh}@vt.edu, $\clubsuit$ moyumyu@tencent.com + +# Abstract + +We compare various forms of prompts to represent event types and develop a unified framework to incorporate the event type specific prompts for supervised, few-shot, and zero-shot event detection. The experimental results demonstrate that a well-defined and comprehensive event type prompt can significantly improve event detection performance, especially when the annotated data is scarce (few-shot event detection) or not available (zero-shot event detection). By leveraging the semantics of event types, our unified framework shows up to $22.2\%$ F-score gain over the previous state-of-the-art baselines1. + +# 1 Introduction + +Event detection (ED) (Grishman, 1997; Chinchor and Marsh, 1998; Ahn, 2006) is the task of identifying and typing event mentions from natural language text. Supervised approaches, especially deep neural networks (Chen et al., 2020; Du and Cardie, 2020; Lin et al., 2020; Liu et al., 2020; Li et al., 2020; Lyu et al., 2021), have shown remarkable performance under a critical prerequisite of a large amount of manual annotations. However, they cannot be effectively generalized to new languages, domains or types, especially when the annotations are not enough (Huang et al., 2016; Huang and Ji, 2020; Lai et al., 2020b; Shen et al., 2021) or there is no annotation available (Lyu et al., 2021; Zhang et al., 2021b; Pasupat and Liang, 2014). + +Recent studies have shown that both the accuracy and generalizability of ED can be improved via leveraging the semantics of event types based on various forms of prompts, such as event type specific queries (Lyu et al., 2021; Du and Cardie, 2020; Liu et al., 2020), definitions (Chen et al., 2020), structures (Lin et al., 2020; Wang et al., + +2019), or a few prototype event triggers (Wang and Cohen, 2009; Dalvi et al., 2012; Pasupat and Liang, 2014; Bronstein et al., 2015; Lai and Nguyen, 2019; Zhang et al., 2021b; Cong et al., 2021). These studies further encourage us to take another step forward and think about the following three questions: (1) does the choice of prompt matter when the training data is abundant or scarce? (2) what's the best form of ED prompt? (3) how to best leverage the prompt to detect event mentions? + +To answer the above research questions, we conduct extensive experiments with various forms of prompts for each event type, including (a) event type name, (b) prototype seed triggers, (c) definition, (d) event type structure based on both event type name and its predefined argument roles, (e) free parameter based continuous soft prompt, and (f) a more comprehensive event type description (named APEX prompt) that covers all the information of prompts (a)-(d). We observe that (1) by considering the semantics of event types with most forms of prompts, especially seed triggers and the comprehensive event type descriptions, the performance of ED under all settings can be significantly improved; (2) Among all forms of event representations, the comprehensive description based prompts show to be the most effective, especially for few-shot and zero-shot ED; (3) Different forms of event type representations provide complementary improvements, indicating that they capture distinct aspects and knowledge of the event types. + +The contributions of this work are as follows: + +- We investigate various prompts to represent event types for both supervised and weakly supervised ED, and prove that a well-defined and comprehensive event type prompt can dramatically improve the performance of ED and the transferability from old types to new types. + +- A unified framework is developed to leverage the semantics of event types with prompts for supervised, few-shot, and zero-shot ED, and demonstrate + +state-of-the-art performance with up to $22.2\%$ F-score improvement over the strong baseline methods. + +# 2 Related Work + +Supervised ED: Most of the existing Event Detection studies follow a supervised learning paradigm (Ji and Grishman, 2008; Liao and Grishman, 2010; McClosky et al., 2011; Li et al., 2013; Chen et al., 2015; Cao et al., 2015; Feng et al., 2016; Yang and Mitchell, 2016; Nguyen et al., 2016; Zhang et al., 2017; Lin et al., 2020; Wang et al., 2021b). However, they cannot be directly applied to detect new types of events. Recently studies have shown that, by leveraging the semantics of event types based on type-specific questions (Du and Cardie, 2020; Liu et al., 2020; Li et al., 2020; Lyu et al., 2021) or seed event triggers (Bronstein et al., 2015; Lai and Nguyen, 2019; Wang et al., 2021a), the event detection performance can be improved. However, it is still unknown whether they are the best choices for representing the semantics of event types. + +Few-shot ED: Two primary learning strategies in few-shot classification tasks are Meta-Learning (Kang et al., 2019; Li et al., 2021; Xiao and Marlet, 2020; Yan et al., 2019; Chowdhury et al., 2021) and Metric Learning (Sun et al., 2021; Wang et al., 2020b; Zhang et al., 2021a; Agarwal et al., 2021). Several studies have exploited metric learning to align the semantics of candidate events with a few examples of the novel event types for few-shot event detection (Lai et al., 2020a; Deng et al., 2020; Lai et al., 2020b; Cong et al., 2021; Chen et al., 2021; Shen et al., 2021). + +Zero-shot ED: Huang et al. (2018) first exploited zero-shot event extraction by leveraging Abstract Meaning Representation (Banarescu et al., 2013) to represent event mentions and types into a shared semantic space. Recent studies (Zhang et al., 2021b; Lyu et al., 2021) further demonstrate that by leveraging a large external corpus with abundant anchor triggers, zero-shot event detection can also be achieved with decent performance without using any training data. + +Prompt Learning Prompt learning aims to learn a task-specific prompt while keeping most of the model's parameters frozen (Li and Liang, 2021; Hambardzumyan et al., 2021; Brown et al., 2020). + +It has shown competitive performance in many applications of natural language processing (Raffel et al., 2020; Brown et al., 2020; Shin et al., 2020; Jiang et al., 2020; Lester et al., 2021; Schick and Schütze, 2021b). Previous work either used a manual (Petroni et al., 2019; Brown et al., 2020; Schick and Schütze, 2021a) or automated approach (Jiang et al., 2020; Yuan et al., 2021; Li and Liang, 2021) to create prompts. + +# 3 Problem Formulation + +Here, we first define each setting of the event detection task and then describe the various forms of event type prompts. + +# 3.1 Settings of ED + +For supervised ED (SED), we follow the conventional supervised event detection setting where the training, validation, and evaluation data sets cover the same set of event types. The goal is to learn a model $f$ to identify and classify event mentions for the target event types. + +For few-shot ED (FSED), there are two separate training data sets for few-shot event detection: (1) A large-scale data set $\mathcal{D}_{base} = \{(\mathbf{x}_i,\mathbf{y}_i)\}_{i = 1}^M$ that covers the old event types (named base types) where $M$ denotes the number of base event types; (2) a smaller data set $\mathcal{D}_{novel} = \{(\mathbf{x}_j,\mathbf{y}_j)\}_{j = 1}^{N\times K}$ that covers $N$ novel event types, with $K$ examples each. Note that the base and novel event types are disjoint except for the Other class. The model $f$ will be first optimized on $\mathcal{D}_{base}$ , and then further fine-tuned on $D_{novel}$ . The goal is to evaluate the generalizability and transferability of the model from base event types to new event types with few annotations. + +For zero-shot ED (ZSED), the training data sets are the only difference between zero-shot and few-shot event detection. In zero-shot event detection, there is only a large-scale base training data set $\mathcal{D}_{base} = \{(\mathbf{x}_i,\mathbf{y}_i)\}_{i = 1}^M$ for the base event types. The model $f$ will be only optimized on base event types and evaluated on the novel types. + +# 3.2 Event Type Prompts + +We compare the following five forms of prompts to represent the event types: (a) Event Type Name is the event class name, usually consisting of one to three tokens. (b) Definition can be a short sentence that formally describes the meaning of the event types. (c) Prototype Seed Triggers a list of + +![](images/ad00dc77d21e6e626ac8784c7edf17c7db361a5b331e9c7a53e86d5455263aab.jpg) +Figure 1: Overview of the unified framework for event detection based on event type specific prompts. + +![](images/a6636a2acdf2de32d3f68ae9a163bd57d22cfdc0ea4d615e461b218bdf15b604.jpg) +Event Type Prompt + +tokens or phrases that are frequently identified as event triggers. (d) Event Type Structure consists of event key argument roles, indicating the core participants of the target event type. (e) Prompts can also be Continuous Soft Prompt, that is, a free vector of parameters to represent each event type. (f) We further define a more comprehensive description APEX Prompt that is manually written and covers all previous prompts except soft prompts. Examples of all event type prompts are shown in Figure 1 and Appendix A. Detailed prompt token selection is in Appendix B. + +# 4 A Unified Framework for ED + +We adapt (Wang et al., 2021a) and design a unified event detection framework (as shown in Figure 1) which leverages event type specific prompts to detect events under supervised, few-shot, and zero-shot settings. Formally, given an input sentence $W = \{w_{1}, w_{2}, \dots, w_{n}\}$ , we take each event type prompt $T^{t} = \{\tau_{1}^{t}, \tau_{2}^{t}, \dots, \tau_{m}^{t}\}$ as a query of $M$ tokens to extract triggers for event type $t$ . Specifically, we first concatenate them into a sequence [CLS] $\tau_{1}^{t} \dots \tau_{m}^{t}$ [SEP] $w_{1} \dots w_{n}$ [SEP]. We use a pre-trained BERT encoder (Devlin et al., 2019) to get contextual representations for the input sentence $W = \{w_{0}, w_{2}, \dots, w_{n}\}$ as well as the event type prompt $T = \{\tau_{0}^{t}, \tau_{1}^{t}, \dots, \tau_{m}^{t}\}^{2}$ . + +Given a prompt of each event type, we aim to extract corresponding event triggers from the input sentence. To achieve this goal, we need to capture the semantic correlation of each input token to the event type. Thus we learn a weight distribution over the sequence of contextual representations of the event type prompt, to obtain event type $t$ aware contextual representation $A_{i}^{t} = \sum_{j=1}^{|T^{t}|} \alpha_{ij} \cdot \tau_{j}^{t}$ , where $\alpha_{ij} = \cos(\boldsymbol{w}_{i}, \tau_{j}^{t})$ , where + +2In our experiments, the representation of each $\pmb{w}_i$ or $\pmb{\tau}_i$ is based on the contextual embedding of the first sub-token. + +$\tau_{j}$ is the contextual representation of the $j$ -th prompt token. $\cos (\cdot)$ is the cosine similarity function between two vectors. + +With that, the event type aware contextual representation $\mathbf{A}_i^t$ will be concatenated with the original contextual representation $\mathbf{w}_i$ from the encoder, and classified into a binary label, indicating whether it is a candidate trigger of event type $t$ or not: $\tilde{\mathbf{y}}_i^t = \mathbf{U}_o([ \mathbf{w}_i; \mathbf{A}_i^t; \mathbf{P}_i ])$ , where $[;]$ denotes concatenation operation, $\mathbf{U}_o$ is a learnable parameter matrix for event trigger detection, and $\mathbf{P}_i$ is the one-hot part-of-speech (POS) encoding of word $\mathbf{w}_i$ . For continuous soft prompt based event detection, we follow Li and Liang (2021) where a prefix index $q$ is prepended to the input sequence $W' = [q; W]$ . The prefix embedding is learned by $\mathbf{q} = \mathrm{MLP}_{\theta}(\mathbf{Q}_{\theta}[q])$ , where $\mathbf{Q}_{\theta} \in \mathbb{R}^{|\mathcal{Q}| \times k}$ denotes the embedding lookup table for the vocabulary of prefix indices. Both $\mathrm{MLP}_{\theta}$ and $\mathbf{Q}_{\theta}$ are trainable parameters. Detailed learning strategy is in Appendix C. + +# 5 Experiment Setup + +We perform experiments on three public benchmark datasets, including ACE05-E $^{+}$ (Automatic Content Extraction), ERE (Entity Relation Event) (Song et al., 2015), and MAVEN (Wang et al., 2020a). On each dataset, we conduct experiments for SED, FSED, and ZSED. For SED, we use the same data split as the previous studies (Li et al., 2013; Wadden et al., 2019; Lin et al., 2020; Du and Cardie, 2020; Lin et al., 2020; Nguyen et al., 2021; Wang et al., 2020a) on all the three benchmark datasets. For FSED and ZSED on MAVEN, we follow the previous study (Chen et al., 2021) and choose 120 event types with the most frequent mentions as the base event types and the rest 45 event types as novel ones. For FSED and ZSED on ACE and ERE, previous studies (Lai et al., 2020b,a; + +
MethodSEDFSEDZSED
Previous SOTA73.3 +(Nguyen et al., 2021)35.2* +(Lai et al., 2020b)49.1* +(Zhang et al., 2021b)
(a) Event type name72.252.749.8
(b) Definition73.146.745.5
(c) Seed triggers73.753.849.6
(d) Event structure72.850.448.0
(e) Soft prompt68.148.2-
Majority voting of (a-e)73.952.148.7
(f) APEX Prompt74.957.451.2
+ +Table 1: Results of event detection (ED) on ACE05 (F1-score, %)* indicates evaluation on our data set split based on the authors' public implementations. + +
MethodSEDFSEDZSED
Previous SOTA59.4(Lu et al., 2021)33.0*(Lai et al., 2020b)41.2*(Zhang et al., 2021b)
(a) Event type Name58.244.840.5
(b) Definition57.944.240.4
(c) Seed triggers60.450.446.2
(d) Event structure59.148.548.7
(e) Soft prompt55.641.7-
Majority voting of (a-e)60.247.945.6
(f) APEX Prompt63.452.648.9
+ +Chen et al., 2021) follow different data splits and settings, making it hard for a fair comparison. Considering the research goals of FSED and ZSED, we define the following conditions to split the ACE and ERE datasets: (i) The base event types and novel event types should be disjoint except Other. (ii) Each base or novel event type should contain at least 15 instances. (iii) The training set should contain sufficient annotated event mentions. + +To meet the above conditions, for ACE, we define the event types of 5 main event categories: Business, Contact, Conflict, Justice and Movement as the base event types, and types of the remaining 3 main categories: Life, Personnel and Transaction as the novel event types. In total, there are 18 qualified base types and 10 qualified novel types (the others do not satisfy the second condition). For ERE, we use the exact same 10 novel event types as ACE, and the rest 25 types as base event types. Detailed data and hyperparameter descriptions are in Appendix D and Appendix E. + +# 6 Results and Discussion + +Overall Results The experimental results for SED, FSED, and ZSED on ACE05, ERE, and + +Table 2: Results of event detection (ED) on ERE (F1-score, %). + +
MethodSEDFSEDZSED
Previous SOTA68.5 +(Wang et al., 2021b)57.0 +(Chen et al., 2021)40.2* +(Zhang et al., 2021b)
(a) Event type name68.863.458.8
(b) Definition67.156.952.9
(c) Seed triggers68.765.159.1
(e) Soft prompt64.538.6-
Majority voting of (a-e)68.463.458.1
(f) APEX Prompt68.868.459.9
+ +Table 3: Results of event detection (ED) on MAVEN (F1-score, %). Event type structure prompts are not applicable to MAVEN as it does not contain any predefined argument roles. + +MAVEN are shown in Table 1-3, from which we see that (1) the APEX prompt achieves the best performance among all the forms of prompts under all the settings of the three benchmark datasets. Compared with the previous state of the art, the APEX prompt shows up to $4\%$ F-score gain for SED (on ERE), $22.2\%$ F-score gain for FSED (on ACE), and $19.7\%$ F-score gain for ZSED (on MAVEN); (2) All the forms of prompts provide significant improvement for FSED and ZSED, demonstrating the benefit of leveraging the semantics of event types via various forms of prompts. (3) Except APEX, seed triggers provide more improvements than other forms of event type prompts under most settings, suggesting its potential to represent the semantics of event types accurately. (4) Continuous soft prompt does not provide comparable performance as other forms of event type representations, which proves the necessity of leveraging event type specific prior knowledge to the representations; (5) The majority voting does not show improvement over individual prompts since each prompt captures a particular aspect of the event type semantics. + +Supervised Event Detection By carefully investigating the event mentions that are correctly detected by the APEX prompt while missed by other prompts, we find that the APEX prompt is more effective in detecting two types of event mentions: homonyms (multiple-meaning words) and intricate words. General homonyms are usually hard to be detected as event mentions as they usually have dozens of meanings in different contexts. For example, consider the following two examples: (i) Airlines are getting [Transport:Movement] flyers to destinations on time more often. (ii) If the board cannot vote to give [Transaction:Transfer-Money'] themselves present money. Here, "get" and "give" + +![](images/940326617a3f0805c4bc018c189d1a6c32c56c66230e62b62b39bb80e6946672.jpg) +Figure 2: F-score distribution of all novel types based on various event type prompts under the few-shot event detection setting on ACE (Best view in color) + +are not detected based on the event type name or seed triggers but are correctly identified by the definition and APEX prompts. The definition and APEX prompts make $10\%$ and $7\%$ fewer false predictions than seed triggers on general homonyms. For intricate words, their semantics usually cannot be captured with an individual prompt. In the following two examples: (i) It is reasonable, however, to reimburse board members for legitimate expenses (ii) ... ever having discussed being compensated by the board in the future ... "reimburse" and "compensated" indicate sophisticated meaning of Transaction:Transfer-Money, which may not be captured by prompts, such as seed triggers. With the event definition and the argument roles in the APEX prompt, the highly correlated contexts, such as "board members" and "legitimate expenses", can help the model correctly detect reimburse as an event mention of Transaction:Transfer-Money. + +Few-shot Event Detection Figure 2 shows the F-score distribution of all novel types based on various forms of event type prompts, from which we observe that: (1) The event type name, seed triggers, and APEX prompt generally perform better than definition and structure, as they carry more straightforward semantics of event types. (2) Event type name based prompts show lower performance on Personnel:End-Position, Personnel:Start-Position and Transaction:Transfer-Money than other event types, as the semantics of these event type names are less indicative than other event types. (3) Seed trigger based prompts perform worse than event type name and APEX prompts on two event types, Life:injure and Life:die, probably because the prototype seed triggers are not properly selected. (4) The structure based prompt outperforms the other + +prompts on Life:Injure as Life:Injure events require the existence of a person or victim. (5) APEX prompt shows consistently (almost) best performance on all the event types because it combines all the information of other prompts. (6) We also observe that the performance of Life:Be-Born, Life:Die, Life:Marry, and Personnel:Elect based on various forms of prompts are consistently better than the other types as the intrinsic semantics of those types the corresponding event triggers are concentrated. + +Zero-shot Event Detection The proposed prompt-based method is more affordable to be generalized compared with the prior state-of-the-art zero-shot approach (Zhang et al., 2021b). The average length of created APEX prompts is less than 20 tokens. Thus manually creating them will not take much human effort. On the contrary, Zhang et al. (2021b) requires an extensive collection of anchor sentences to perform zero-shot event detection, e.g., 4,556,237 anchor sentences for ACE and ERE. This process is time-consuming and expensive. + +# 7 Conclusion + +We investigate a variety of prompts to represent the semantics of event types, and leverage them with a unified framework for supervised, few-shot and zero-shot event detection. Experimental results demonstrate that, a well-defined and comprehensive description of event types can significantly improve the performance of event detection, especially when the annotations are limited (few-shot event detection) or even not available (zero-shot event detection), with up to $22.2\%$ F-score gain over the prior state of the art. + +# Limitations + +We have demonstrated that an accurate description can perform better for both supervised and weakly supervised event detection. However, the event types from most existing ontologies are not properly defined. For example, in ACE annotation guideline (Linguistic Data Consortium, 2005), transfer-money is defined as "giving, receiving, borrowing, or lending money when it is not in the context of purchasing something". However, it is hard for the model to interpret it accurately, especially the constraints "not in the context of purchasing something". In addition, many event types from MAVEN, e.g., Achieve, Award, and Incident, are not associated with any definitions. A potential future research direction is to leverage mining-based approaches or state-of-the-art generators to automatically generate a comprehensive event type description based on various sources, such as annotation guidelines, example annotations, and external knowledge bases. + +# Acknowledgments + +We thank the anonymous reviewers and area chair for their valuable time and constructive comments. This research is based upon work supported by the Amazon Research Award. + +# References + +Ashutosh Agarwal, Anay Majee, Anbumani Subramanian, and Chetan Arora. 2021. Attention guided cosine margin for overcoming class-imbalance in few-shot road object detection. +David Ahn. 2006. The stages of event extraction. In Proceedings of the Workshop on Annotating and Reasoning about Time and Events, pages 1-8. +Collin F Baker, Charles J Fillmore, and John B Lowe. 1998. The berkeley framenet project. In 36th Annual Meeting of the Association for Computational Linguistics and 17th International Conference on Computational Linguistics, Volume 1, pages 86-90. +Laura Banarescu, Claire Bonial, Shu Cai, Madalina Georgescu, Kira Griffith, Ulf Hermjakob, Kevin Knight, Philipp Koehn, Martha Palmer, and Nathan Schneider. 2013. Abstract Meaning Representation for sembanking. In Proceedings of the 7th Linguistic Annotation Workshop and Interoperability with Discourse, pages 178-186, Sofia, Bulgaria. Association for Computational Linguistics. +Ofer Bronstein, Ido Dagan, Qi Li, Heng Ji, and Anette Frank. 2015. Seed-based event trigger labeling: How + +far can event descriptions get us? In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 372-376. +Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems, volume 33, pages 1877-1901. Curran Associates, Inc. +Kai Cao, Xiang Li, Miao Fan, and Ralph Grishman. 2015. Improving event detection with active learning. In Proceedings of the International Conference on Recent Advances in Natural Language Processing, pages 72-77, Hissar, Bulgaria. INCOMA Ltd. Shoumen, BULGARIA. +Jiawei Chen, Hongyu Lin, Xianpei Han, and Le Sun. 2021. Honey or poison? solving the trigger curse in few-shot event detection via causal intervention. +Yubo Chen, Liheng Xu, Kang Liu, Daojian Zeng, and Jun Zhao. 2015. Event extraction via dynamic multipooling convolutional neural networks. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 167-176. +Yunmo Chen, Tongfei Chen, Seth Ebner, Aaron Steven White, and Benjamin Van Durme. 2020. Reading the manual: Event extraction as definition comprehension. In Proceedings of the Fourth Workshop on Structured Prediction for NLP, pages 74-83, Online. Association for Computational Linguistics. +Nancy Chinchor and Elaine Marsh. 1998. Muc-7 information extraction task definition. In Proceeding of the seventh message understanding conference (MUC-7), Appendices, pages 359-367. +Arkabandhu Chowdhury, Mingchao Jiang, and Chris Jermaine. 2021. Few-shot image classification: Just use a library of pre-trained feature extractors and a simple classifier. abs/2101.00562. +Xin Cong, Shiyao Cui, Bowen Yu, Tingwen Liu, Yubin Wang, and Bin Wang. 2021. Few-shot event detection with prototypical amortized conditional random field. In Findings of the Association for Computational Linguistics: ACL-IJCNLP. +Bhavana Dalvi, William W. Cohen, and Jamie Callan. 2012. Websets: extracting sets of entities from + +the web using unsupervised information extraction. ArXiv, abs/1307.0261. +Shumin Deng, Ningyu Zhang, Jiaojian Kang, Yichi Zhang, Wei Zhang, and Huajun Chen. 2020. Meta-learning with dynamic-memory-based prototypical network for few-shot event detection. Proceedings of the 13th International Conference on Web Search and Data Mining. +Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics. +Xinya Du and Claire Cardie. 2020. Event extraction by answering (almost) natural questions. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 671-683, Online. Association for Computational Linguistics. +Xiaocheng Feng, Lifu Huang, Duyu Tang, Heng Ji, Bing Qin, and Ting Liu. 2016. A language-independent neural network for event detection. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 66-71, Berlin, Germany. Association for Computational Linguistics. +Ralph Grishman. 1997. Information extraction: Techniques and challenges. In International summer school on information extraction, pages 10-27. Springer. +Karen Hambardzumyan, Hrant Khachatrian, and Jonathan May. 2021. WARP: Word-level Adversarial ReProgramming. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4921-4933, Online. Association for Computational Linguistics. +Lifu Huang, Taylor Cassidy, Xiaocheng Feng, Heng Ji, Clare Voss, Jiawei Han, and Avirup Sil. 2016. Liberal event extraction and event schema induction. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 258-268. +Lifu Huang and Heng Ji. 2020. Semi-supervised new event type induction and event detection. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 718-724. +Lifu Huang, Heng Ji, Kyunghyun Cho, Ido Dagan, Sebastian Riedel, and Clare Voss. 2018. Zero-shot transfer learning for event extraction. In Proceedings of the 56th Annual Meeting of the Association for + +Computational Linguistics (Volume 1: Long Papers), pages 2160-2170, Melbourne, Australia. Association for Computational Linguistics. +Heng Ji and Ralph Grishman. 2008. Refining event extraction through cross-document inference. In Proceedings of ACL-08: Hlt, pages 254-262. +Zhengbao Jiang, Frank F. Xu, J. Araki, and Graham Neubig. 2020. How can we know what language models know? Transactions of the Association for Computational Linguistics, 8:423-438. +Bingyi Kang, Zhuang Liu, Xin Wang, Fisher Yu, Jiashi Feng, and Trevor Darrell. 2019. Few-shot object detection via feature reweighting. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 8419-8428. +Viet Dac Lai, Franck Dernoncourt, and Thien Huu Nguyen. 2020a. Exploiting the matching information in the support set for few shot event classification. Pacific-Asia Conference on Knowledge Discovery and Data Mining, page 233-245. +Viet Dac Lai and Thien Huu Nguyen. 2019. Extending event detection to new types with learning from keywords. arXiv preprint arXiv:1910.11368. +Viet Dac Lai, Thien Huu Nguyen, and Franck Dernoncourt. 2020b. Extensively matching for few-shot learning event detection. In Proceedings of the First Joint Workshop on Narrative Understanding, Storylines, and Events, pages 38-45, Online. Association for Computational Linguistics. +Brian Lester, Rami Al-Rfou, and Noah Constant. 2021. The power of scale for parameter-efficient prompt tuning. In EMNLP. +Bohao Li, Boyu Yang, Chang Liu, Feng Liu, Rongrong Ji, and Qixiang Ye. 2021. Beyond max-margin: Class margin equilibrium for few-shot object detection. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7359-7368. +Fayuan Li, Weihua Peng, Yuguang Chen, Quan Wang, Lu Pan, Yajuan Lyu, and Yong Zhu. 2020. Event extraction as multi-turn question answering. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 829–838, Online. Association for Computational Linguistics. +Qi Li, Heng Ji, and Liang Huang. 2013. Joint event extraction via structured prediction with global features. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 73-82, Sofia, Bulgaria. Association for Computational Linguistics. +Xiang Lisa Li and Percy Liang. 2021. Prefix-tuning: Optimizing continuous prompts for generation. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), abs/2101.00190. + +Shasha Liao and Ralph Grishman. 2010. Using document level cross-event inference to improve event extraction. In Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 789-797. +Ying Lin, Heng Ji, Fei Huang, and Lingfei Wu. 2020. A joint neural model for information extraction with global features. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7999-8009, Online. Association for Computational Linguistics. +Linguistic Data Consortium. 2005. English annotation guidelines for events. https://www.ldc.upenn.edu/sites/www.ldc.upenn.edu/files/english-events-guidelines-v5.4.3.pdf. +Jian Liu, Yubo Chen, Kang Liu, Wei Bi, and Xiaojiang Liu. 2020. Event extraction as machine reading comprehension. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1641-1651, Online. Association for Computational Linguistics. +Yaojie Lu, Hongyu Lin, Jin Xu, Xianpei Han, Jialong Tang, Annan Li, Le Sun, Meng Liao, and Shaoyi Chen. 2021. Text2Event: Controllable sequence-to-structure generation for end-to-end event extraction. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 2795-2806, Online. Association for Computational Linguistics. +Qing Lyu, Hongming Zhang, Elior Sulem, and Dan Roth. 2021. Zero-shot Event Extraction via Transfer Learning: Challenges and Insights. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 322-332, Online. Association for Computational Linguistics. +David McClosky, Mihai Surdeanu, and Christopher D Manning. 2011. Event extraction as dependency parsing. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 1626-1635. +Minh Van Nguyen, Viet Dac Lai, and Thien Huu Nguyen. 2021. Cross-task instance representation interactions and label dependencies for joint information extraction with graph convolutional networks. +Thien Huu Nguyen, Kyunghyun Cho, and Ralph Grishman. 2016. Joint event extraction via recurrent neural networks. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 300-309, San Diego, California. Association for Computational Linguistics. + +Panupong Pasupat and Percy Liang. 2014. Zero-shot entity extraction from web pages. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 391-401. +Fabio Petroni, Tim Rocttäschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, and Alexander Miller. 2019. Language models as knowledge bases? In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2463-2473, Hong Kong, China. Association for Computational Linguistics. +Colin Raffel, Noam M. Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR. +Timo Schick and Hinrich Schütze. 2021a. Few-shot text generation with pattern-exploiting training. +Timo Schick and Hinrich Schütze. 2021b. It's not just size that matters: Small language models are also few-shot learners. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics, pages 2339-2352. +Shirong Shen, Tongtong Wu, Guilin Qi, Yuan-Fang Li, Gholamreza Haffari, and Sheng Bi. 2021. Adaptive knowledge-enhanced bayesian meta-learning for few-shot event detection. In Findings of the Association for Computational Linguistics, page 2417-2429. Association for Computational Linguistics (ACL). Annual Meeting of the Association for Computational Linguistics and the International Joint Conference on Natural Language Processing 2021, ACL-IJCNLP 2021; Conference date: 01-08-2021 Through 06-08-2021. +Taylor Shin, Yasaman Razeghi, Robert L. Logan IV, Eric Wallace, and Sameer Singh. 2020. AutoPrompt: Eliciting knowledge from language models with automatically generated prompts. In Empirical Methods in Natural Language Processing (EMNLP). +Zhiyi Song, Ann Bies, Stephanie Strassel, Tom Riese, Justin Mott, Joe Ellis, Jonathan Wright, Seth Kulick, Neville Ryant, and Xiaoyi Ma. 2015. From light to rich ere: annotation of entities, relations, and events. In Proceedings of the 3rd Workshop on EVENTS: Definition, Detection, Coreference, and Representation, pages 89-98. +Bo Sun, Banghuai Li, Shengcai Cai, Ye Yuan, and Chi Zhang. 2021. Fsce: Few-shot object detection via contrastive proposal encoding. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7348-7358. +David Wadden, Ulme Wennberg, Yi Luan, and Hannaneh Hajishirzi. 2019. Entity, relation, and event + +extraction with contextualized span representations. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5784-5789, Hong Kong, China. Association for Computational Linguistics. +Richard C Wang and William Cohen. 2009. Character-level analysis of semi-structured documents for set expansion. In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 1503–1512. +Sijia Wang, Mo Yu, Shiyu Chang, Lichao Sun, and Lifu Huang. 2021a. Query and extract: Refining event extraction as type-oriented binary decoding. arXiv preprint arXiv:2110.07476. +Xiao Wang, Houye Ji, Chuan Shi, Bai Wang, Yanfang Ye, Peng Cui, and Philip S Yu. 2019. Heterogeneous graph attention network. In The World Wide Web Conference, WWW '19, page 2022-2032, New York, NY, USA. Association for Computing Machinery. +Xiaozhi Wang, Ziqi Wang, Xu Han, Wangyi Jiang, Rong Han, Zhiyuan Liu, Juanzi Li, Peng Li, Yankai Lin, and Jie Zhou. 2020a. MAVEN: A massive general domain event detection dataset. In Proceedings of EMNLP 2020. +Xin Wang, Thomas E. Huang, Trevor Darrell, Joseph E Gonzalez, and Fisher Yu. 2020b. Frustratingly simple few-shot object detection. +Ziqi Wang, Xiaozhi Wang, Xu Han, Yankai Lin, Lei Hou, Zhiyuan Liu, Peng Li, Juanzi Li, and Jie Zhou. 2021b. CLEVE: Contrastive Pre-training for Event Extraction. In Proceedings of ACL-IJCNLP, pages 6283-6297, Online. Association for Computational Linguistics. +Yang Xiao and Renaud Marlet. 2020. Few-shot object detection and viewpoint estimation for objects in the wild. In ECCV. +Xiaopeng Yan, Ziliang Chen, Anni Xu, Xiaoxi Wang, Xiaodan Liang, and Liang Lin. 2019. Meta r-cnn: Towards general solver for instance-level low-shot learning. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9576-9585. +Bishan Yang and Tom M. Mitchell. 2016. Joint extraction of events and entities within a document context. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 289-299, San Diego, California. Association for Computational Linguistics. +Weizhe Yuan, Graham Neubig, and Pengfei Liu. 2021. BARTScore: Evaluating generated text as text generation. In Advances in Neural Information Processing Systems. + +Gongjie Zhang, Kaiwen Cui, Rongliang Wu, Shijian Lu, and Yonghong Tian. 2021a. Pnpdet: Efficient few-shot detection without forgetting via plug-and-play sub-networks. 2021 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 3822-3831. +Hongming Zhang, Haoyu Wang, and Dan Roth. 2021b. Zero-shot Label-aware Event Trigger and Argument Classification. In *Findings of the Association for Computational Linguistics: ACL-IJCNLP* 2021, pages 1331-1340, Online. Association for Computational Linguistics. +Tongtao Zhang, Spencer Whitehead, Hanwang Zhang, Hongzhi Li, Joseph Ellis, Lifu Huang, Wei Liu, Heng Ji, and Shih-Fu Chang. 2017. Improving event extraction via multimodal integration. In Proceedings of the 25th ACM international conference on Multimedia, pages 270-278. + +# A APEX prompt examples for ACE + +Table 4 and Table 5 show APEX prompt examples for ACE events. + +# B Prompt Token Selection + +In our experiments, the event type names and event type structures are automatically extracted from the target event ontology, such as ACE (Linguistic Data Consortium, 2005), ERE (Song et al., 2015) and MAVEN (Wang et al., 2020a). The prototype seed triggers are automatically selected from the annotated data for supervised and few-shot event extraction. For zero-shot event extraction, we manually select $R$ words from the NLTK synonyms of each event type as its prototype seed triggers. The definitions and APEX prompts are based on the official annotation guides for each target event ontology (Linguistic Data Consortium, 2005; Song et al., 2015; Wang et al., 2020a) and the available definitions in FrameNet (Baker et al., 1998) with manual editing. + +# C Learning Strategy + +The learning strategy varies for supervised, few-shot, and zero-shot learning. For supervised learning, we optimize the following objective for event trigger detection $\mathcal{L} = -\frac{1}{|\mathcal{T}||\mathcal{N}|}\sum_{t\in \mathcal{T}}\sum_{i = 1}^{|\mathcal{N}|}\boldsymbol{y}_i^t$ $\log \tilde{\boldsymbol{y}}_i^t$ where $\mathcal{T}$ is the set of target event types and $\mathcal{N}$ is the set of tokens from the training dataset. $\boldsymbol{y}_i^t$ denotes the ground truth label vector. For few-shot event detection, we optimize the model on both base training data set and the smaller training data set for novel event types: $\mathcal{L} = -\frac{1}{|\mathcal{T}^B||\mathcal{N}^B|}\sum_{t\in \mathcal{T}^B}\sum_{i = 1}^{|\mathcal{N}^B|}\boldsymbol{y}_i^t\cdot \log \tilde{\boldsymbol{y}}_i^t -$ $\beta \frac{1}{|\mathcal{T}^N||\mathcal{N}^N|}\sum_{t\in \mathcal{T}^N}\sum_{i = 1}^{|\mathcal{N}^N|}\boldsymbol{y}_i^t\cdot \log \tilde{\boldsymbol{y}}_i^t$ , where $\mathcal{T}^B$ and $\mathcal{N}^B$ denote the set of base event types and tokens from the base training data set, respectively. $\mathcal{T}^N$ is the set of novel event types. $\mathcal{N}^N$ is the set of tokens from the training data set for novel event types. $\beta$ is a hyper-parameter to balance the two objectives. For zero-shot event detection, as we only have the base training data set, we minimize the following objective: $\mathcal{L} = -\frac{1}{|\mathcal{T}^B||\mathcal{N}^B|}\sum_{t\in \mathcal{T}^B}\sum_{i = 1}^{|\mathcal{N}^B|}\boldsymbol{y}_i^t$ $\log \tilde{\boldsymbol{y}}_i^t$ + +# D Dataset + +After defining the base and novel event types, we create the training, validation, and evaluation split for all three datasets. We use the sentences with + +only base event type mentions as the base training data set for few-shot event detection, and randomly select 10 sentences with novel event type mentions as the additional smaller training data set. We use the sentences with both base and novel event type mentions as the development set and use the remaining sentences with only novel event type mentions as the evaluation dataset. We use the same development and evaluation set as few-shot event detection for zero-shot event detection and remove the instances with novel event mentions from the training set. We randomly split the sentences without any event annotations proportionally to the number of sentences with event mentions in each set for both zero-shot and few-shot event detection. Table 6 shows the detailed data statistics for all the three datasets under the few-shot and zero-shot event extraction settings. + +# E Hyperparameters and Evaluation + +For a fair comparison with the previous baseline approaches, we use the same pre-trained bert-large-uncased model for fine-tuning and optimizing our model with BertAdam. For supervised event detection, we optimize the parameters with grid search: training epoch is 3, learning rate $\in [3e - 6,1e - 4]$ , training batch size $\in$ {8, 12, 16, 24, 32}, dropout rate $\in$ {0.4, 0.5, 0.6}. The running time is up to 3 hours on one Quadro RTX 8000. For evaluation, we use the same criteria as previous studies (Li et al., 2013; Chen et al., 2015; Nguyen et al., 2016; Lin et al., 2020): an event mention is correct if its span and event type match a reference event mention. + +
Event Rep TypeComprehensive Prompt
Business:Declare-BankruptcyDeclare Bankruptcy [SEP] bankruptcy bankruptciesbankrupting [SEP] Organizationrequest legal protection from debt collection at a Place
Business:End-OrgEnd Organization [SEP] dissolving disbanded [SEP] an Organization goes out of business at a Place
Business:Merge-OrgMerge Organization [SEP] merging merger [SEP] two or more Organizations come together to form a new organization at a Place
Business:Start-OrgStart Organization [SEP] founded [SEP] an Agent create a new Organization at a Place
Conflict:AttackAttack [SEP] invaded airstrikes overthrew ambushed [SEP] An Attacker physically attacks a Target with Instrument at a Place
Conflict:DemonstrateDemonstrate [SEP] demonstrations protest strikes riots [SEP] Entities come together in a Place to protest or demand official action
Contact:MeetMeet [SEP] reunited retreats [SEP] two or more Entities come together at same Place and interact in person
Contact:Phone-WritePhone Write [SEP] emailed letter [SEP] phone or written communication between two or more Entities
Justice:AcquitAcquit [SEP] acquitted [SEP] a trial of Defendant ends but Adjudicator fails to produce a conviction at a Place
Justice:AppealAppeal [SEP] appeal [SEP] the decision for Defendant of a court is taken to a higher court for Adjudicator review with Prosecutor
Justice:Arrest-JailArrest Jail [SEP] arrested locked [SEP] the Agent takes custody of a Person at a Place
Justice:Charge-IndictCharge Indict [SEP] indictment [SEP] a Defendant is accused of a crime by a Prosecutor for Adjudicator
Justice:ConvictConvict [SEP] pled guilty convicting [SEP] an Defendant found guilty of a crime by Adjudicator at a Place
Justice:ExecuteExecute [SEP] death [SEP] the life of a Person is taken by an Agent at a Place
Justice:ExtraditeExtradite [SEP] extradition [SEP] a Person is sent by an Agent from Origin to Destination
Justice:FineFine [SEP] payouts financial punishment [SEP] a Adjudicator issues a financial punishment Money to an Entity at a Place
Justice:PardonPardon [SEP] pardoned lift sentence [SEP] an Adjudicator lifts a sentence of Defendant at a Place
Justice:Release-ParoleRelease Parole [SEP] parole [SEP] an Entity ends its custody of a Person at a Place
Justice:SentenceSentence [SEP] sentenced punishment [SEP] the punishment for the defendant is issued by a state actor
Justice:SueSue [SEP] lawsuits [SEP] Plaintiff initiate a court proceeding to determine the liability of a Defendant judge by Adjudicator at a Place
Justice:Trial-HearingTrial Hearing [SEP] trial hearings [SEP] a court proceeding initiated to determine the guilty or innocence of a Person with Prosecutor and Adjudicator at a Place
Life:Be-BornBe Born [SEP] childbirth [SEP] a Person is born at a Place
Life:DieDie [SEP] deceased extermination [SEP] life of a Victim ends by an Agent with Instrument at a Place
+ +Table 4: APEX templates for ACE event types + +
Event Rep TypeComprehensive Prompt
Life:DivorceDivorce [SEP] people divorce [SEP] two Person are officially divorced at a place
Life:InjureInjure [SEP] hospitalised paralyzed dismember [SEP] a Victim experiences physical harm from Agent with Instrument at a Place
Life:MarryMarry [SEP] married marriage marry [SEP] two Person are married at a Place
Movement:TransportTransport [SEP] arrival travels penetrated expelled [SEP] an Agent moves an Artifact from Origin to Destination with Vehicle at Price
Personnel:ElectElect [SEP] reelected elected election [SEP] a candidate Person wins an election by voting Entity at a Place
Personnel:End-PositionEnd Position [SEP] resigning retired resigned [SEP] a Person stops working for an Entity or change office at a Place
Personnel:NominateNominate [SEP] nominate [SEP] a Person is nominated for a new position by another Agent at a Place
Personnel:Start-PositionStart Position [SEP] hiring rehired recruited [SEP] a Person begins working for an Entity or change office at a Place
Transaction:Transfer-MoneyTransfer Money [SEP] donations reimbursing deductions [SEP] transfer Money from the Giver to the Beneficiary or Recipient at a Place
Transaction:Transfer-OwnershipTransfer Ownership [SEP] purchased buy sell loan [SEP] buying selling loaning borrowing giving receiving of Artifacts from Seller to Buyer or Beneficiary at a Place at Price
+ +Table 5: APEX templates for ACE event types (continued) + +
DatasetACE05-E+ERE-ENMAVEN
# TypesBase1825120
Novel101045
# MentionsBase3572544993675
Novel172431833201
TrainFew-shot3216388688085
Zero-shot3116378687635
Validation900(51%/49%)2797(53%/47%)3883(71%/23%)
Evaluation119520121652
+ +Table 6: Data statistics for ACE2005, ERE and MAVEN datasets under few-shot/zero-shot event detection settings. + +A For every submission: + +A1. Did you describe the limitations of your work? Section 8 +A2. Did you discuss any potential risks of your work? Section 8 +A3. Do the abstract and introduction summarize the paper's main claims? Abstract and Section 1 +A4. Have you used AI writing assistants when working on this paper? Left blank. + +B Did you use or create scientific artifacts? + +Section 5 and Appendix C + +B1. Did you cite the creators of artifacts you used? Section 5 and Appendix C +B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Section 5 +B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Section 5 +B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Left blank. +B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Section 5 +B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Appendix C + +C Did you run computational experiments? + +Section 5 and Appendix D + +C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Appendix D + +C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? + +Section 5 and Appendix D + +C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? + +Appendix D + +C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? + +Appendix B + +D Did you use human annotators (e.g., crowdworkers) or research with human participants? + +Left blank. + +D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? + +No response. + +D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? + +No response. + +D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? + +No response. + +D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? + +No response. + +D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? + +No response. \ No newline at end of file diff --git a/2023/The Art of Prompting_ Event Detection based on Type Specific Prompts/images.zip b/2023/The Art of Prompting_ Event Detection based on Type Specific Prompts/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d8ecde788ba20c64306de95e05d2ff09714f4a6e --- /dev/null +++ b/2023/The Art of Prompting_ Event Detection based on Type Specific Prompts/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:24682c2b1956845d8825e6ed10ae62786f0f0ffd20a28f3b409ec9101bd8debf +size 663597 diff --git a/2023/The Art of Prompting_ Event Detection based on Type Specific Prompts/layout.json b/2023/The Art of Prompting_ Event Detection based on Type Specific Prompts/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0d31c696c7aeb6f6120ed61ac884b9ead3af6997 --- /dev/null +++ b/2023/The Art of Prompting_ Event Detection based on Type Specific Prompts/layout.json @@ -0,0 +1,8648 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 76, + 75, + 518, + 94 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 75, + 518, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 75, + 518, + 94 + ], + "type": "text", + "content": "The Art of Prompting: Event Detection based on Type Specific Prompts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 201, + 118, + 393, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 118, + 393, + 133 + ], + "spans": [ + { + "bbox": [ + 201, + 118, + 393, + 133 + ], + "type": "text", + "content": "Sijia Wang*, Mo Yu*, Lifu Huang*" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 224, + 133, + 371, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 133, + 371, + 147 + ], + "spans": [ + { + "bbox": [ + 224, + 133, + 371, + 147 + ], + "type": "inline_equation", + "content": "\\spadesuit" + }, + { + "bbox": [ + 224, + 133, + 371, + 147 + ], + "type": "text", + "content": " Virginia Tech, " + }, + { + "bbox": [ + 224, + 133, + 371, + 147 + ], + "type": "inline_equation", + "content": "\\spadesuit" + }, + { + "bbox": [ + 224, + 133, + 371, + 147 + ], + "type": "text", + "content": " WeChat AI" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 129, + 147, + 467, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 147, + 467, + 162 + ], + "spans": [ + { + "bbox": [ + 129, + 147, + 467, + 162 + ], + "type": "inline_equation", + "content": "\\clubsuit" + }, + { + "bbox": [ + 129, + 147, + 467, + 162 + ], + "type": "text", + "content": " {sijiawang, lifuh}@vt.edu, " + }, + { + "bbox": [ + 129, + 147, + 467, + 162 + ], + "type": "inline_equation", + "content": "\\clubsuit" + }, + { + "bbox": [ + 129, + 147, + 467, + 162 + ], + "type": "text", + "content": " moyumyu@tencent.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 84, + 236, + 274, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 236, + 274, + 402 + ], + "spans": [ + { + "bbox": [ + 84, + 236, + 274, + 402 + ], + "type": "text", + "content": "We compare various forms of prompts to represent event types and develop a unified framework to incorporate the event type specific prompts for supervised, few-shot, and zero-shot event detection. The experimental results demonstrate that a well-defined and comprehensive event type prompt can significantly improve event detection performance, especially when the annotated data is scarce (few-shot event detection) or not available (zero-shot event detection). By leveraging the semantics of event types, our unified framework shows up to " + }, + { + "bbox": [ + 84, + 236, + 274, + 402 + ], + "type": "inline_equation", + "content": "22.2\\%" + }, + { + "bbox": [ + 84, + 236, + 274, + 402 + ], + "type": "text", + "content": " F-score gain over the previous state-of-the-art baselines1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 414, + 154, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 414, + 154, + 427 + ], + "spans": [ + { + "bbox": [ + 68, + 414, + 154, + 427 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 436, + 291, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 436, + 291, + 639 + ], + "spans": [ + { + "bbox": [ + 67, + 436, + 291, + 639 + ], + "type": "text", + "content": "Event detection (ED) (Grishman, 1997; Chinchor and Marsh, 1998; Ahn, 2006) is the task of identifying and typing event mentions from natural language text. Supervised approaches, especially deep neural networks (Chen et al., 2020; Du and Cardie, 2020; Lin et al., 2020; Liu et al., 2020; Li et al., 2020; Lyu et al., 2021), have shown remarkable performance under a critical prerequisite of a large amount of manual annotations. However, they cannot be effectively generalized to new languages, domains or types, especially when the annotations are not enough (Huang et al., 2016; Huang and Ji, 2020; Lai et al., 2020b; Shen et al., 2021) or there is no annotation available (Lyu et al., 2021; Zhang et al., 2021b; Pasupat and Liang, 2014)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 640, + 291, + 734 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 640, + 291, + 734 + ], + "spans": [ + { + "bbox": [ + 67, + 640, + 291, + 734 + ], + "type": "text", + "content": "Recent studies have shown that both the accuracy and generalizability of ED can be improved via leveraging the semantics of event types based on various forms of prompts, such as event type specific queries (Lyu et al., 2021; Du and Cardie, 2020; Liu et al., 2020), definitions (Chen et al., 2020), structures (Lin et al., 2020; Wang et al.," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 213, + 526, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 213, + 526, + 348 + ], + "spans": [ + { + "bbox": [ + 302, + 213, + 526, + 348 + ], + "type": "text", + "content": "2019), or a few prototype event triggers (Wang and Cohen, 2009; Dalvi et al., 2012; Pasupat and Liang, 2014; Bronstein et al., 2015; Lai and Nguyen, 2019; Zhang et al., 2021b; Cong et al., 2021). These studies further encourage us to take another step forward and think about the following three questions: (1) does the choice of prompt matter when the training data is abundant or scarce? (2) what's the best form of ED prompt? (3) how to best leverage the prompt to detect event mentions?" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 349, + 526, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 349, + 526, + 634 + ], + "spans": [ + { + "bbox": [ + 302, + 349, + 526, + 634 + ], + "type": "text", + "content": "To answer the above research questions, we conduct extensive experiments with various forms of prompts for each event type, including (a) event type name, (b) prototype seed triggers, (c) definition, (d) event type structure based on both event type name and its predefined argument roles, (e) free parameter based continuous soft prompt, and (f) a more comprehensive event type description (named APEX prompt) that covers all the information of prompts (a)-(d). We observe that (1) by considering the semantics of event types with most forms of prompts, especially seed triggers and the comprehensive event type descriptions, the performance of ED under all settings can be significantly improved; (2) Among all forms of event representations, the comprehensive description based prompts show to be the most effective, especially for few-shot and zero-shot ED; (3) Different forms of event type representations provide complementary improvements, indicating that they capture distinct aspects and knowledge of the event types." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 636, + 515, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 636, + 515, + 648 + ], + "spans": [ + { + "bbox": [ + 313, + 636, + 515, + 648 + ], + "type": "text", + "content": "The contributions of this work are as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 651, + 525, + 731 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 651, + 525, + 731 + ], + "spans": [ + { + "bbox": [ + 302, + 651, + 525, + 731 + ], + "type": "text", + "content": "- We investigate various prompts to represent event types for both supervised and weakly supervised ED, and prove that a well-defined and comprehensive event type prompt can dramatically improve the performance of ED and the transferability from old types to new types." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 733, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 733, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 733, + 525, + 772 + ], + "type": "text", + "content": "- A unified framework is developed to leverage the semantics of event types with prompts for supervised, few-shot, and zero-shot ED, and demonstrate" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 740, + 290, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 740, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 740, + 290, + 772 + ], + "type": "text", + "content": "1The source code, model checkpoints and data are publicly available at https://github.com/VT-NLP/Event_APEX." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "1286" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "spans": [ + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "type": "text", + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 219, + 806, + 374, + 817 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 806, + 374, + 817 + ], + "spans": [ + { + "bbox": [ + 219, + 806, + 374, + 817 + ], + "type": "text", + "content": "Volume 2: Short Papers, pages 1286-1299" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "spans": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "text", + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 291, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 291, + 111 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 291, + 111 + ], + "type": "text", + "content": "state-of-the-art performance with up to " + }, + { + "bbox": [ + 67, + 71, + 291, + 111 + ], + "type": "inline_equation", + "content": "22.2\\%" + }, + { + "bbox": [ + 67, + 71, + 291, + 111 + ], + "type": "text", + "content": " F-score improvement over the strong baseline methods." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 124, + 160, + 137 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 124, + 160, + 137 + ], + "spans": [ + { + "bbox": [ + 68, + 124, + 160, + 137 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 148, + 291, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 148, + 291, + 391 + ], + "spans": [ + { + "bbox": [ + 67, + 148, + 291, + 391 + ], + "type": "text", + "content": "Supervised ED: Most of the existing Event Detection studies follow a supervised learning paradigm (Ji and Grishman, 2008; Liao and Grishman, 2010; McClosky et al., 2011; Li et al., 2013; Chen et al., 2015; Cao et al., 2015; Feng et al., 2016; Yang and Mitchell, 2016; Nguyen et al., 2016; Zhang et al., 2017; Lin et al., 2020; Wang et al., 2021b). However, they cannot be directly applied to detect new types of events. Recently studies have shown that, by leveraging the semantics of event types based on type-specific questions (Du and Cardie, 2020; Liu et al., 2020; Li et al., 2020; Lyu et al., 2021) or seed event triggers (Bronstein et al., 2015; Lai and Nguyen, 2019; Wang et al., 2021a), the event detection performance can be improved. However, it is still unknown whether they are the best choices for representing the semantics of event types." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 401, + 291, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 401, + 291, + 564 + ], + "spans": [ + { + "bbox": [ + 67, + 401, + 291, + 564 + ], + "type": "text", + "content": "Few-shot ED: Two primary learning strategies in few-shot classification tasks are Meta-Learning (Kang et al., 2019; Li et al., 2021; Xiao and Marlet, 2020; Yan et al., 2019; Chowdhury et al., 2021) and Metric Learning (Sun et al., 2021; Wang et al., 2020b; Zhang et al., 2021a; Agarwal et al., 2021). Several studies have exploited metric learning to align the semantics of candidate events with a few examples of the novel event types for few-shot event detection (Lai et al., 2020a; Deng et al., 2020; Lai et al., 2020b; Cong et al., 2021; Chen et al., 2021; Shen et al., 2021)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 574, + 291, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 574, + 291, + 709 + ], + "spans": [ + { + "bbox": [ + 67, + 574, + 291, + 709 + ], + "type": "text", + "content": "Zero-shot ED: Huang et al. (2018) first exploited zero-shot event extraction by leveraging Abstract Meaning Representation (Banarescu et al., 2013) to represent event mentions and types into a shared semantic space. Recent studies (Zhang et al., 2021b; Lyu et al., 2021) further demonstrate that by leveraging a large external corpus with abundant anchor triggers, zero-shot event detection can also be achieved with decent performance without using any training data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 719, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 719, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 719, + 291, + 773 + ], + "type": "text", + "content": "Prompt Learning Prompt learning aims to learn a task-specific prompt while keeping most of the model's parameters frozen (Li and Liang, 2021; Hambardzumyan et al., 2021; Brown et al., 2020)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 71, + 526, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 193 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 193 + ], + "type": "text", + "content": "It has shown competitive performance in many applications of natural language processing (Raffel et al., 2020; Brown et al., 2020; Shin et al., 2020; Jiang et al., 2020; Lester et al., 2021; Schick and Schütze, 2021b). Previous work either used a manual (Petroni et al., 2019; Brown et al., 2020; Schick and Schütze, 2021a) or automated approach (Jiang et al., 2020; Yuan et al., 2021; Li and Liang, 2021) to create prompts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 204, + 436, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 204, + 436, + 217 + ], + "spans": [ + { + "bbox": [ + 302, + 204, + 436, + 217 + ], + "type": "text", + "content": "3 Problem Formulation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 227, + 526, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 227, + 526, + 268 + ], + "spans": [ + { + "bbox": [ + 302, + 227, + 526, + 268 + ], + "type": "text", + "content": "Here, we first define each setting of the event detection task and then describe the various forms of event type prompts." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 278, + 398, + 291 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 278, + 398, + 291 + ], + "spans": [ + { + "bbox": [ + 302, + 278, + 398, + 291 + ], + "type": "text", + "content": "3.1 Settings of ED" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 296, + 525, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 296, + 525, + 377 + ], + "spans": [ + { + "bbox": [ + 302, + 296, + 525, + 377 + ], + "type": "text", + "content": "For supervised ED (SED), we follow the conventional supervised event detection setting where the training, validation, and evaluation data sets cover the same set of event types. The goal is to learn a model " + }, + { + "bbox": [ + 302, + 296, + 525, + 377 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 302, + 296, + 525, + 377 + ], + "type": "text", + "content": " to identify and classify event mentions for the target event types." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 378, + 525, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 378, + 525, + 566 + ], + "spans": [ + { + "bbox": [ + 302, + 378, + 525, + 566 + ], + "type": "text", + "content": "For few-shot ED (FSED), there are two separate training data sets for few-shot event detection: (1) A large-scale data set " + }, + { + "bbox": [ + 302, + 378, + 525, + 566 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{base} = \\{(\\mathbf{x}_i,\\mathbf{y}_i)\\}_{i = 1}^M" + }, + { + "bbox": [ + 302, + 378, + 525, + 566 + ], + "type": "text", + "content": " that covers the old event types (named base types) where " + }, + { + "bbox": [ + 302, + 378, + 525, + 566 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 302, + 378, + 525, + 566 + ], + "type": "text", + "content": " denotes the number of base event types; (2) a smaller data set " + }, + { + "bbox": [ + 302, + 378, + 525, + 566 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{novel} = \\{(\\mathbf{x}_j,\\mathbf{y}_j)\\}_{j = 1}^{N\\times K}" + }, + { + "bbox": [ + 302, + 378, + 525, + 566 + ], + "type": "text", + "content": " that covers " + }, + { + "bbox": [ + 302, + 378, + 525, + 566 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 302, + 378, + 525, + 566 + ], + "type": "text", + "content": " novel event types, with " + }, + { + "bbox": [ + 302, + 378, + 525, + 566 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 302, + 378, + 525, + 566 + ], + "type": "text", + "content": " examples each. Note that the base and novel event types are disjoint except for the Other class. The model " + }, + { + "bbox": [ + 302, + 378, + 525, + 566 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 302, + 378, + 525, + 566 + ], + "type": "text", + "content": " will be first optimized on " + }, + { + "bbox": [ + 302, + 378, + 525, + 566 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{base}" + }, + { + "bbox": [ + 302, + 378, + 525, + 566 + ], + "type": "text", + "content": ", and then further fine-tuned on " + }, + { + "bbox": [ + 302, + 378, + 525, + 566 + ], + "type": "inline_equation", + "content": "D_{novel}" + }, + { + "bbox": [ + 302, + 378, + 525, + 566 + ], + "type": "text", + "content": ". The goal is to evaluate the generalizability and transferability of the model from base event types to new event types with few annotations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 568, + 526, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 568, + 526, + 663 + ], + "spans": [ + { + "bbox": [ + 302, + 568, + 526, + 663 + ], + "type": "text", + "content": "For zero-shot ED (ZSED), the training data sets are the only difference between zero-shot and few-shot event detection. In zero-shot event detection, there is only a large-scale base training data set " + }, + { + "bbox": [ + 302, + 568, + 526, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{base} = \\{(\\mathbf{x}_i,\\mathbf{y}_i)\\}_{i = 1}^M" + }, + { + "bbox": [ + 302, + 568, + 526, + 663 + ], + "type": "text", + "content": " for the base event types. The model " + }, + { + "bbox": [ + 302, + 568, + 526, + 663 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 302, + 568, + 526, + 663 + ], + "type": "text", + "content": " will be only optimized on base event types and evaluated on the novel types." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 674, + 427, + 687 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 674, + 427, + 687 + ], + "spans": [ + { + "bbox": [ + 302, + 674, + 427, + 687 + ], + "type": "text", + "content": "3.2 Event Type Prompts" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 692, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 692, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 692, + 525, + 772 + ], + "type": "text", + "content": "We compare the following five forms of prompts to represent the event types: (a) Event Type Name is the event class name, usually consisting of one to three tokens. (b) Definition can be a short sentence that formally describes the meaning of the event types. (c) Prototype Seed Triggers a list of" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "1287" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 72, + 407, + 196 + ], + "blocks": [ + { + "bbox": [ + 75, + 72, + 407, + 196 + ], + "lines": [ + { + "bbox": [ + 75, + 72, + 407, + 196 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 407, + 196 + ], + "type": "image", + "image_path": "ad00dc77d21e6e626ac8784c7edf17c7db361a5b331e9c7a53e86d5455263aab.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 205, + 501, + 218 + ], + "lines": [ + { + "bbox": [ + 89, + 205, + 501, + 218 + ], + "spans": [ + { + "bbox": [ + 89, + 205, + 501, + 218 + ], + "type": "text", + "content": "Figure 1: Overview of the unified framework for event detection based on event type specific prompts." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 420, + 82, + 524, + 191 + ], + "blocks": [ + { + "bbox": [ + 420, + 73, + 480, + 81 + ], + "lines": [ + { + "bbox": [ + 420, + 73, + 480, + 81 + ], + "spans": [ + { + "bbox": [ + 420, + 73, + 480, + 81 + ], + "type": "text", + "content": "Event Type Prompt" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 420, + 82, + 524, + 191 + ], + "lines": [ + { + "bbox": [ + 420, + 82, + 524, + 191 + ], + "spans": [ + { + "bbox": [ + 420, + 82, + 524, + 191 + ], + "type": "image", + "image_path": "a6636a2acdf2de32d3f68ae9a163bd57d22cfdc0ea4d615e461b218bdf15b604.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 238, + 291, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 238, + 291, + 401 + ], + "spans": [ + { + "bbox": [ + 67, + 238, + 291, + 401 + ], + "type": "text", + "content": "tokens or phrases that are frequently identified as event triggers. (d) Event Type Structure consists of event key argument roles, indicating the core participants of the target event type. (e) Prompts can also be Continuous Soft Prompt, that is, a free vector of parameters to represent each event type. (f) We further define a more comprehensive description APEX Prompt that is manually written and covers all previous prompts except soft prompts. Examples of all event type prompts are shown in Figure 1 and Appendix A. Detailed prompt token selection is in Appendix B." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 410, + 239, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 410, + 239, + 423 + ], + "spans": [ + { + "bbox": [ + 67, + 410, + 239, + 423 + ], + "type": "text", + "content": "4 A Unified Framework for ED" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 430, + 291, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 430, + 291, + 619 + ], + "spans": [ + { + "bbox": [ + 67, + 430, + 291, + 619 + ], + "type": "text", + "content": "We adapt (Wang et al., 2021a) and design a unified event detection framework (as shown in Figure 1) which leverages event type specific prompts to detect events under supervised, few-shot, and zero-shot settings. Formally, given an input sentence " + }, + { + "bbox": [ + 67, + 430, + 291, + 619 + ], + "type": "inline_equation", + "content": "W = \\{w_{1}, w_{2}, \\dots, w_{n}\\}" + }, + { + "bbox": [ + 67, + 430, + 291, + 619 + ], + "type": "text", + "content": ", we take each event type prompt " + }, + { + "bbox": [ + 67, + 430, + 291, + 619 + ], + "type": "inline_equation", + "content": "T^{t} = \\{\\tau_{1}^{t}, \\tau_{2}^{t}, \\dots, \\tau_{m}^{t}\\}" + }, + { + "bbox": [ + 67, + 430, + 291, + 619 + ], + "type": "text", + "content": " as a query of " + }, + { + "bbox": [ + 67, + 430, + 291, + 619 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 67, + 430, + 291, + 619 + ], + "type": "text", + "content": " tokens to extract triggers for event type " + }, + { + "bbox": [ + 67, + 430, + 291, + 619 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 67, + 430, + 291, + 619 + ], + "type": "text", + "content": ". Specifically, we first concatenate them into a sequence [CLS] " + }, + { + "bbox": [ + 67, + 430, + 291, + 619 + ], + "type": "inline_equation", + "content": "\\tau_{1}^{t} \\dots \\tau_{m}^{t}" + }, + { + "bbox": [ + 67, + 430, + 291, + 619 + ], + "type": "text", + "content": " [SEP] " + }, + { + "bbox": [ + 67, + 430, + 291, + 619 + ], + "type": "inline_equation", + "content": "w_{1} \\dots w_{n}" + }, + { + "bbox": [ + 67, + 430, + 291, + 619 + ], + "type": "text", + "content": " [SEP]. We use a pre-trained BERT encoder (Devlin et al., 2019) to get contextual representations for the input sentence " + }, + { + "bbox": [ + 67, + 430, + 291, + 619 + ], + "type": "inline_equation", + "content": "W = \\{w_{0}, w_{2}, \\dots, w_{n}\\}" + }, + { + "bbox": [ + 67, + 430, + 291, + 619 + ], + "type": "text", + "content": " as well as the event type prompt " + }, + { + "bbox": [ + 67, + 430, + 291, + 619 + ], + "type": "inline_equation", + "content": "T = \\{\\tau_{0}^{t}, \\tau_{1}^{t}, \\dots, \\tau_{m}^{t}\\}^{2}" + }, + { + "bbox": [ + 67, + 430, + 291, + 619 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 620, + 291, + 747 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 620, + 291, + 747 + ], + "spans": [ + { + "bbox": [ + 67, + 620, + 291, + 747 + ], + "type": "text", + "content": "Given a prompt of each event type, we aim to extract corresponding event triggers from the input sentence. To achieve this goal, we need to capture the semantic correlation of each input token to the event type. Thus we learn a weight distribution over the sequence of contextual representations of the event type prompt, to obtain event type " + }, + { + "bbox": [ + 67, + 620, + 291, + 747 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 67, + 620, + 291, + 747 + ], + "type": "text", + "content": " aware contextual representation " + }, + { + "bbox": [ + 67, + 620, + 291, + 747 + ], + "type": "inline_equation", + "content": "A_{i}^{t} = \\sum_{j=1}^{|T^{t}|} \\alpha_{ij} \\cdot \\tau_{j}^{t}" + }, + { + "bbox": [ + 67, + 620, + 291, + 747 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 67, + 620, + 291, + 747 + ], + "type": "inline_equation", + "content": "\\alpha_{ij} = \\cos(\\boldsymbol{w}_{i}, \\tau_{j}^{t})" + }, + { + "bbox": [ + 67, + 620, + 291, + 747 + ], + "type": "text", + "content": ", where" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 751, + 289, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 751, + 289, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 751, + 289, + 772 + ], + "type": "text", + "content": "2In our experiments, the representation of each " + }, + { + "bbox": [ + 67, + 751, + 289, + 772 + ], + "type": "inline_equation", + "content": "\\pmb{w}_i" + }, + { + "bbox": [ + 67, + 751, + 289, + 772 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 67, + 751, + 289, + 772 + ], + "type": "inline_equation", + "content": "\\pmb{\\tau}_i" + }, + { + "bbox": [ + 67, + 751, + 289, + 772 + ], + "type": "text", + "content": " is based on the contextual embedding of the first sub-token." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 238, + 526, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 238, + 526, + 278 + ], + "spans": [ + { + "bbox": [ + 302, + 238, + 526, + 278 + ], + "type": "inline_equation", + "content": "\\tau_{j}" + }, + { + "bbox": [ + 302, + 238, + 526, + 278 + ], + "type": "text", + "content": " is the contextual representation of the " + }, + { + "bbox": [ + 302, + 238, + 526, + 278 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 302, + 238, + 526, + 278 + ], + "type": "text", + "content": "-th prompt token. " + }, + { + "bbox": [ + 302, + 238, + 526, + 278 + ], + "type": "inline_equation", + "content": "\\cos (\\cdot)" + }, + { + "bbox": [ + 302, + 238, + 526, + 278 + ], + "type": "text", + "content": " is the cosine similarity function between two vectors." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 280, + 526, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 280, + 526, + 522 + ], + "spans": [ + { + "bbox": [ + 302, + 280, + 526, + 522 + ], + "type": "text", + "content": "With that, the event type aware contextual representation " + }, + { + "bbox": [ + 302, + 280, + 526, + 522 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i^t" + }, + { + "bbox": [ + 302, + 280, + 526, + 522 + ], + "type": "text", + "content": " will be concatenated with the original contextual representation " + }, + { + "bbox": [ + 302, + 280, + 526, + 522 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_i" + }, + { + "bbox": [ + 302, + 280, + 526, + 522 + ], + "type": "text", + "content": " from the encoder, and classified into a binary label, indicating whether it is a candidate trigger of event type " + }, + { + "bbox": [ + 302, + 280, + 526, + 522 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 302, + 280, + 526, + 522 + ], + "type": "text", + "content": " or not: " + }, + { + "bbox": [ + 302, + 280, + 526, + 522 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{y}}_i^t = \\mathbf{U}_o([ \\mathbf{w}_i; \\mathbf{A}_i^t; \\mathbf{P}_i ])" + }, + { + "bbox": [ + 302, + 280, + 526, + 522 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 302, + 280, + 526, + 522 + ], + "type": "inline_equation", + "content": "[;]" + }, + { + "bbox": [ + 302, + 280, + 526, + 522 + ], + "type": "text", + "content": " denotes concatenation operation, " + }, + { + "bbox": [ + 302, + 280, + 526, + 522 + ], + "type": "inline_equation", + "content": "\\mathbf{U}_o" + }, + { + "bbox": [ + 302, + 280, + 526, + 522 + ], + "type": "text", + "content": " is a learnable parameter matrix for event trigger detection, and " + }, + { + "bbox": [ + 302, + 280, + 526, + 522 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_i" + }, + { + "bbox": [ + 302, + 280, + 526, + 522 + ], + "type": "text", + "content": " is the one-hot part-of-speech (POS) encoding of word " + }, + { + "bbox": [ + 302, + 280, + 526, + 522 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_i" + }, + { + "bbox": [ + 302, + 280, + 526, + 522 + ], + "type": "text", + "content": ". For continuous soft prompt based event detection, we follow Li and Liang (2021) where a prefix index " + }, + { + "bbox": [ + 302, + 280, + 526, + 522 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 302, + 280, + 526, + 522 + ], + "type": "text", + "content": " is prepended to the input sequence " + }, + { + "bbox": [ + 302, + 280, + 526, + 522 + ], + "type": "inline_equation", + "content": "W' = [q; W]" + }, + { + "bbox": [ + 302, + 280, + 526, + 522 + ], + "type": "text", + "content": ". The prefix embedding is learned by " + }, + { + "bbox": [ + 302, + 280, + 526, + 522 + ], + "type": "inline_equation", + "content": "\\mathbf{q} = \\mathrm{MLP}_{\\theta}(\\mathbf{Q}_{\\theta}[q])" + }, + { + "bbox": [ + 302, + 280, + 526, + 522 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 302, + 280, + 526, + 522 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\theta} \\in \\mathbb{R}^{|\\mathcal{Q}| \\times k}" + }, + { + "bbox": [ + 302, + 280, + 526, + 522 + ], + "type": "text", + "content": " denotes the embedding lookup table for the vocabulary of prefix indices. Both " + }, + { + "bbox": [ + 302, + 280, + 526, + 522 + ], + "type": "inline_equation", + "content": "\\mathrm{MLP}_{\\theta}" + }, + { + "bbox": [ + 302, + 280, + 526, + 522 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 302, + 280, + 526, + 522 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\theta}" + }, + { + "bbox": [ + 302, + 280, + 526, + 522 + ], + "type": "text", + "content": " are trainable parameters. Detailed learning strategy is in Appendix C." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 534, + 417, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 534, + 417, + 549 + ], + "spans": [ + { + "bbox": [ + 302, + 534, + 417, + 549 + ], + "type": "text", + "content": "5 Experiment Setup" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 301, + 556, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 556, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 301, + 556, + 526, + 772 + ], + "type": "text", + "content": "We perform experiments on three public benchmark datasets, including ACE05-E" + }, + { + "bbox": [ + 301, + 556, + 526, + 772 + ], + "type": "inline_equation", + "content": "^{+}" + }, + { + "bbox": [ + 301, + 556, + 526, + 772 + ], + "type": "text", + "content": " (Automatic Content Extraction), ERE (Entity Relation Event) (Song et al., 2015), and MAVEN (Wang et al., 2020a). On each dataset, we conduct experiments for SED, FSED, and ZSED. For SED, we use the same data split as the previous studies (Li et al., 2013; Wadden et al., 2019; Lin et al., 2020; Du and Cardie, 2020; Lin et al., 2020; Nguyen et al., 2021; Wang et al., 2020a) on all the three benchmark datasets. For FSED and ZSED on MAVEN, we follow the previous study (Chen et al., 2021) and choose 120 event types with the most frequent mentions as the base event types and the rest 45 event types as novel ones. For FSED and ZSED on ACE and ERE, previous studies (Lai et al., 2020b,a;" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 780, + 309, + 791 + ], + "type": "text", + "content": "1288" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 68, + 289, + 194 + ], + "blocks": [ + { + "bbox": [ + 70, + 68, + 289, + 194 + ], + "lines": [ + { + "bbox": [ + 70, + 68, + 289, + 194 + ], + "spans": [ + { + "bbox": [ + 70, + 68, + 289, + 194 + ], + "type": "table", + "html": "
MethodSEDFSEDZSED
Previous SOTA73.3\n(Nguyen et al., 2021)35.2*\n(Lai et al., 2020b)49.1*\n(Zhang et al., 2021b)
(a) Event type name72.252.749.8
(b) Definition73.146.745.5
(c) Seed triggers73.753.849.6
(d) Event structure72.850.448.0
(e) Soft prompt68.148.2-
Majority voting of (a-e)73.952.148.7
(f) APEX Prompt74.957.451.2
", + "image_path": "96d92f02992ef9dd1f782a365b502db021551420812024554ffc91819b572827.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 70, + 249, + 289, + 374 + ], + "blocks": [ + { + "bbox": [ + 67, + 201, + 290, + 240 + ], + "lines": [ + { + "bbox": [ + 67, + 201, + 290, + 240 + ], + "spans": [ + { + "bbox": [ + 67, + 201, + 290, + 240 + ], + "type": "text", + "content": "Table 1: Results of event detection (ED) on ACE05 (F1-score, %)* indicates evaluation on our data set split based on the authors' public implementations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 249, + 289, + 374 + ], + "lines": [ + { + "bbox": [ + 70, + 249, + 289, + 374 + ], + "spans": [ + { + "bbox": [ + 70, + 249, + 289, + 374 + ], + "type": "table", + "html": "
MethodSEDFSEDZSED
Previous SOTA59.4(Lu et al., 2021)33.0*(Lai et al., 2020b)41.2*(Zhang et al., 2021b)
(a) Event type Name58.244.840.5
(b) Definition57.944.240.4
(c) Seed triggers60.450.446.2
(d) Event structure59.148.548.7
(e) Soft prompt55.641.7-
Majority voting of (a-e)60.247.945.6
(f) APEX Prompt63.452.648.9
", + "image_path": "db8770a7a9101ca525fe6872e9e0455ac14d086c11049df461ed255d5c9929c8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 428, + 290, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 428, + 290, + 549 + ], + "spans": [ + { + "bbox": [ + 67, + 428, + 290, + 549 + ], + "type": "text", + "content": "Chen et al., 2021) follow different data splits and settings, making it hard for a fair comparison. Considering the research goals of FSED and ZSED, we define the following conditions to split the ACE and ERE datasets: (i) The base event types and novel event types should be disjoint except Other. (ii) Each base or novel event type should contain at least 15 instances. (iii) The training set should contain sufficient annotated event mentions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 550, + 291, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 550, + 291, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 550, + 291, + 714 + ], + "type": "text", + "content": "To meet the above conditions, for ACE, we define the event types of 5 main event categories: Business, Contact, Conflict, Justice and Movement as the base event types, and types of the remaining 3 main categories: Life, Personnel and Transaction as the novel event types. In total, there are 18 qualified base types and 10 qualified novel types (the others do not satisfy the second condition). For ERE, we use the exact same 10 novel event types as ACE, and the rest 25 types as base event types. Detailed data and hyperparameter descriptions are in Appendix D and Appendix E." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 724, + 206, + 737 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 724, + 206, + 737 + ], + "spans": [ + { + "bbox": [ + 67, + 724, + 206, + 737 + ], + "type": "text", + "content": "6 Results and Discussion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "type": "text", + "content": "Overall Results The experimental results for SED, FSED, and ZSED on ACE05, ERE, and" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 305, + 68, + 525, + 185 + ], + "blocks": [ + { + "bbox": [ + 67, + 382, + 291, + 407 + ], + "lines": [ + { + "bbox": [ + 67, + 382, + 291, + 407 + ], + "spans": [ + { + "bbox": [ + 67, + 382, + 291, + 407 + ], + "type": "text", + "content": "Table 2: Results of event detection (ED) on ERE (F1-score, %)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 305, + 68, + 525, + 185 + ], + "lines": [ + { + "bbox": [ + 305, + 68, + 525, + 185 + ], + "spans": [ + { + "bbox": [ + 305, + 68, + 525, + 185 + ], + "type": "table", + "html": "
MethodSEDFSEDZSED
Previous SOTA68.5\n(Wang et al., 2021b)57.0\n(Chen et al., 2021)40.2*\n(Zhang et al., 2021b)
(a) Event type name68.863.458.8
(b) Definition67.156.952.9
(c) Seed triggers68.765.159.1
(e) Soft prompt64.538.6-
Majority voting of (a-e)68.463.458.1
(f) APEX Prompt68.868.459.9
", + "image_path": "dbba2a5306c5f6f7506ad0eda772dd02235a3b9a611fa2cb7317435cdcdbd1d5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 192, + 526, + 242 + ], + "lines": [ + { + "bbox": [ + 302, + 192, + 526, + 242 + ], + "spans": [ + { + "bbox": [ + 302, + 192, + 526, + 242 + ], + "type": "text", + "content": "Table 3: Results of event detection (ED) on MAVEN (F1-score, %). Event type structure prompts are not applicable to MAVEN as it does not contain any predefined argument roles." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 302, + 263, + 526, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 263, + 526, + 575 + ], + "spans": [ + { + "bbox": [ + 302, + 263, + 526, + 575 + ], + "type": "text", + "content": "MAVEN are shown in Table 1-3, from which we see that (1) the APEX prompt achieves the best performance among all the forms of prompts under all the settings of the three benchmark datasets. Compared with the previous state of the art, the APEX prompt shows up to " + }, + { + "bbox": [ + 302, + 263, + 526, + 575 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 302, + 263, + 526, + 575 + ], + "type": "text", + "content": " F-score gain for SED (on ERE), " + }, + { + "bbox": [ + 302, + 263, + 526, + 575 + ], + "type": "inline_equation", + "content": "22.2\\%" + }, + { + "bbox": [ + 302, + 263, + 526, + 575 + ], + "type": "text", + "content": " F-score gain for FSED (on ACE), and " + }, + { + "bbox": [ + 302, + 263, + 526, + 575 + ], + "type": "inline_equation", + "content": "19.7\\%" + }, + { + "bbox": [ + 302, + 263, + 526, + 575 + ], + "type": "text", + "content": " F-score gain for ZSED (on MAVEN); (2) All the forms of prompts provide significant improvement for FSED and ZSED, demonstrating the benefit of leveraging the semantics of event types via various forms of prompts. (3) Except APEX, seed triggers provide more improvements than other forms of event type prompts under most settings, suggesting its potential to represent the semantics of event types accurately. (4) Continuous soft prompt does not provide comparable performance as other forms of event type representations, which proves the necessity of leveraging event type specific prior knowledge to the representations; (5) The majority voting does not show improvement over individual prompts since each prompt captures a particular aspect of the event type semantics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 584, + 527, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 584, + 527, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 584, + 527, + 773 + ], + "type": "text", + "content": "Supervised Event Detection By carefully investigating the event mentions that are correctly detected by the APEX prompt while missed by other prompts, we find that the APEX prompt is more effective in detecting two types of event mentions: homonyms (multiple-meaning words) and intricate words. General homonyms are usually hard to be detected as event mentions as they usually have dozens of meanings in different contexts. For example, consider the following two examples: (i) Airlines are getting [Transport:Movement] flyers to destinations on time more often. (ii) If the board cannot vote to give [Transaction:Transfer-Money'] themselves present money. Here, \"get\" and \"give\"" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "text", + "content": "1289" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 139, + 72, + 452, + 223 + ], + "blocks": [ + { + "bbox": [ + 139, + 72, + 452, + 223 + ], + "lines": [ + { + "bbox": [ + 139, + 72, + 452, + 223 + ], + "spans": [ + { + "bbox": [ + 139, + 72, + 452, + 223 + ], + "type": "image", + "image_path": "940326617a3f0805c4bc018c189d1a6c32c56c66230e62b62b39bb80e6946672.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 231, + 525, + 258 + ], + "lines": [ + { + "bbox": [ + 67, + 231, + 525, + 258 + ], + "spans": [ + { + "bbox": [ + 67, + 231, + 525, + 258 + ], + "type": "text", + "content": "Figure 2: F-score distribution of all novel types based on various event type prompts under the few-shot event detection setting on ACE (Best view in color)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 268, + 291, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 268, + 291, + 526 + ], + "spans": [ + { + "bbox": [ + 69, + 268, + 291, + 526 + ], + "type": "text", + "content": "are not detected based on the event type name or seed triggers but are correctly identified by the definition and APEX prompts. The definition and APEX prompts make " + }, + { + "bbox": [ + 69, + 268, + 291, + 526 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 69, + 268, + 291, + 526 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 69, + 268, + 291, + 526 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 69, + 268, + 291, + 526 + ], + "type": "text", + "content": " fewer false predictions than seed triggers on general homonyms. For intricate words, their semantics usually cannot be captured with an individual prompt. In the following two examples: (i) It is reasonable, however, to reimburse board members for legitimate expenses (ii) ... ever having discussed being compensated by the board in the future ... \"reimburse\" and \"compensated\" indicate sophisticated meaning of Transaction:Transfer-Money, which may not be captured by prompts, such as seed triggers. With the event definition and the argument roles in the APEX prompt, the highly correlated contexts, such as \"board members\" and \"legitimate expenses\", can help the model correctly detect reimburse as an event mention of Transaction:Transfer-Money." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 543, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 543, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 543, + 291, + 772 + ], + "type": "text", + "content": "Few-shot Event Detection Figure 2 shows the F-score distribution of all novel types based on various forms of event type prompts, from which we observe that: (1) The event type name, seed triggers, and APEX prompt generally perform better than definition and structure, as they carry more straightforward semantics of event types. (2) Event type name based prompts show lower performance on Personnel:End-Position, Personnel:Start-Position and Transaction:Transfer-Money than other event types, as the semantics of these event type names are less indicative than other event types. (3) Seed trigger based prompts perform worse than event type name and APEX prompts on two event types, Life:injure and Life:die, probably because the prototype seed triggers are not properly selected. (4) The structure based prompt outperforms the other" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 302, + 268, + 526, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 268, + 526, + 417 + ], + "spans": [ + { + "bbox": [ + 302, + 268, + 526, + 417 + ], + "type": "text", + "content": "prompts on Life:Injure as Life:Injure events require the existence of a person or victim. (5) APEX prompt shows consistently (almost) best performance on all the event types because it combines all the information of other prompts. (6) We also observe that the performance of Life:Be-Born, Life:Die, Life:Marry, and Personnel:Elect based on various forms of prompts are consistently better than the other types as the intrinsic semantics of those types the corresponding event triggers are concentrated." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 302, + 427, + 525, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 427, + 525, + 589 + ], + "spans": [ + { + "bbox": [ + 302, + 427, + 525, + 589 + ], + "type": "text", + "content": "Zero-shot Event Detection The proposed prompt-based method is more affordable to be generalized compared with the prior state-of-the-art zero-shot approach (Zhang et al., 2021b). The average length of created APEX prompts is less than 20 tokens. Thus manually creating them will not take much human effort. On the contrary, Zhang et al. (2021b) requires an extensive collection of anchor sentences to perform zero-shot event detection, e.g., 4,556,237 anchor sentences for ACE and ERE. This process is time-consuming and expensive." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 303, + 601, + 381, + 614 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 601, + 381, + 614 + ], + "spans": [ + { + "bbox": [ + 303, + 601, + 381, + 614 + ], + "type": "text", + "content": "7 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 624, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 624, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 624, + 526, + 772 + ], + "type": "text", + "content": "We investigate a variety of prompts to represent the semantics of event types, and leverage them with a unified framework for supervised, few-shot and zero-shot event detection. Experimental results demonstrate that, a well-defined and comprehensive description of event types can significantly improve the performance of event detection, especially when the annotations are limited (few-shot event detection) or even not available (zero-shot event detection), with up to " + }, + { + "bbox": [ + 302, + 624, + 526, + 772 + ], + "type": "inline_equation", + "content": "22.2\\%" + }, + { + "bbox": [ + 302, + 624, + 526, + 772 + ], + "type": "text", + "content": " F-score gain over the prior state of the art." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 780, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 780, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 780, + 310, + 791 + ], + "type": "text", + "content": "1290" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 71, + 131, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 71, + 131, + 83 + ], + "spans": [ + { + "bbox": [ + 68, + 71, + 131, + 83 + ], + "type": "text", + "content": "Limitations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 93, + 291, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 93, + 291, + 362 + ], + "spans": [ + { + "bbox": [ + 69, + 93, + 291, + 362 + ], + "type": "text", + "content": "We have demonstrated that an accurate description can perform better for both supervised and weakly supervised event detection. However, the event types from most existing ontologies are not properly defined. For example, in ACE annotation guideline (Linguistic Data Consortium, 2005), transfer-money is defined as \"giving, receiving, borrowing, or lending money when it is not in the context of purchasing something\". However, it is hard for the model to interpret it accurately, especially the constraints \"not in the context of purchasing something\". In addition, many event types from MAVEN, e.g., Achieve, Award, and Incident, are not associated with any definitions. A potential future research direction is to leverage mining-based approaches or state-of-the-art generators to automatically generate a comprehensive event type description based on various sources, such as annotation guidelines, example annotations, and external knowledge bases." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 68, + 374, + 166, + 386 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 374, + 166, + 386 + ], + "spans": [ + { + "bbox": [ + 68, + 374, + 166, + 386 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 396, + 290, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 396, + 290, + 449 + ], + "spans": [ + { + "bbox": [ + 67, + 396, + 290, + 449 + ], + "type": "text", + "content": "We thank the anonymous reviewers and area chair for their valuable time and constructive comments. This research is based upon work supported by the Amazon Research Award." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 472, + 127, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 472, + 127, + 485 + ], + "spans": [ + { + "bbox": [ + 68, + 472, + 127, + 485 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 491, + 290, + 772 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 69, + 491, + 290, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 491, + 290, + 536 + ], + "spans": [ + { + "bbox": [ + 69, + 491, + 290, + 536 + ], + "type": "text", + "content": "Ashutosh Agarwal, Anay Majee, Anbumani Subramanian, and Chetan Arora. 2021. Attention guided cosine margin for overcoming class-imbalance in few-shot road object detection." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 545, + 290, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 545, + 290, + 579 + ], + "spans": [ + { + "bbox": [ + 69, + 545, + 290, + 579 + ], + "type": "text", + "content": "David Ahn. 2006. The stages of event extraction. In Proceedings of the Workshop on Annotating and Reasoning about Time and Events, pages 1-8." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 588, + 290, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 588, + 290, + 644 + ], + "spans": [ + { + "bbox": [ + 69, + 588, + 290, + 644 + ], + "type": "text", + "content": "Collin F Baker, Charles J Fillmore, and John B Lowe. 1998. The berkeley framenet project. In 36th Annual Meeting of the Association for Computational Linguistics and 17th International Conference on Computational Linguistics, Volume 1, pages 86-90." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 652, + 290, + 740 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 652, + 290, + 740 + ], + "spans": [ + { + "bbox": [ + 69, + 652, + 290, + 740 + ], + "type": "text", + "content": "Laura Banarescu, Claire Bonial, Shu Cai, Madalina Georgescu, Kira Griffith, Ulf Hermjakob, Kevin Knight, Philipp Koehn, Martha Palmer, and Nathan Schneider. 2013. Abstract Meaning Representation for sembanking. In Proceedings of the 7th Linguistic Annotation Workshop and Interoperability with Discourse, pages 178-186, Sofia, Bulgaria. Association for Computational Linguistics." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 750, + 290, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 750, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 750, + 290, + 772 + ], + "type": "text", + "content": "Ofer Bronstein, Ido Dagan, Qi Li, Heng Ji, and Anette Frank. 2015. Seed-based event trigger labeling: How" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 526, + 772 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 314, + 72, + 525, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 525, + 127 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 525, + 127 + ], + "type": "text", + "content": "far can event descriptions get us? In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 372-376." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 137, + 526, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 137, + 526, + 290 + ], + "spans": [ + { + "bbox": [ + 304, + 137, + 526, + 290 + ], + "type": "text", + "content": "Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems, volume 33, pages 1877-1901. Curran Associates, Inc." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 299, + 525, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 299, + 525, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 299, + 525, + 365 + ], + "type": "text", + "content": "Kai Cao, Xiang Li, Miao Fan, and Ralph Grishman. 2015. Improving event detection with active learning. In Proceedings of the International Conference on Recent Advances in Natural Language Processing, pages 72-77, Hissar, Bulgaria. INCOMA Ltd. Shoumen, BULGARIA." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 375, + 525, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 375, + 525, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 375, + 525, + 407 + ], + "type": "text", + "content": "Jiawei Chen, Hongyu Lin, Xianpei Han, and Le Sun. 2021. Honey or poison? solving the trigger curse in few-shot event detection via causal intervention." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 417, + 525, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 417, + 525, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 417, + 525, + 495 + ], + "type": "text", + "content": "Yubo Chen, Liheng Xu, Kang Liu, Daojian Zeng, and Jun Zhao. 2015. Event extraction via dynamic multipooling convolutional neural networks. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 167-176." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 503, + 525, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 503, + 525, + 570 + ], + "spans": [ + { + "bbox": [ + 304, + 503, + 525, + 570 + ], + "type": "text", + "content": "Yunmo Chen, Tongfei Chen, Seth Ebner, Aaron Steven White, and Benjamin Van Durme. 2020. Reading the manual: Event extraction as definition comprehension. In Proceedings of the Fourth Workshop on Structured Prediction for NLP, pages 74-83, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 578, + 525, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 525, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 525, + 624 + ], + "type": "text", + "content": "Nancy Chinchor and Elaine Marsh. 1998. Muc-7 information extraction task definition. In Proceeding of the seventh message understanding conference (MUC-7), Appendices, pages 359-367." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 632, + 525, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 632, + 525, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 632, + 525, + 677 + ], + "type": "text", + "content": "Arkabandhu Chowdhury, Mingchao Jiang, and Chris Jermaine. 2021. Few-shot image classification: Just use a library of pre-trained feature extractors and a simple classifier. abs/2101.00562." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 685, + 525, + 740 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 685, + 525, + 740 + ], + "spans": [ + { + "bbox": [ + 304, + 685, + 525, + 740 + ], + "type": "text", + "content": "Xin Cong, Shiyao Cui, Bowen Yu, Tingwen Liu, Yubin Wang, and Bin Wang. 2021. Few-shot event detection with prototypical amortized conditional random field. In Findings of the Association for Computational Linguistics: ACL-IJCNLP." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 750, + 525, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 750, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 750, + 525, + 772 + ], + "type": "text", + "content": "Bhavana Dalvi, William W. Cohen, and Jamie Callan. 2012. Websets: extracting sets of entities from" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 308, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 308, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 308, + 791 + ], + "type": "text", + "content": "1291" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 290, + 772 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 80, + 72, + 290, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 72, + 290, + 95 + ], + "spans": [ + { + "bbox": [ + 80, + 72, + 290, + 95 + ], + "type": "text", + "content": "the web using unsupervised information extraction. ArXiv, abs/1307.0261." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 104, + 289, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 104, + 289, + 170 + ], + "spans": [ + { + "bbox": [ + 69, + 104, + 289, + 170 + ], + "type": "text", + "content": "Shumin Deng, Ningyu Zhang, Jiaojian Kang, Yichi Zhang, Wei Zhang, and Huajun Chen. 2020. Meta-learning with dynamic-memory-based prototypical network for few-shot event detection. Proceedings of the 13th International Conference on Web Search and Data Mining." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 179, + 289, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 179, + 289, + 278 + ], + "spans": [ + { + "bbox": [ + 69, + 179, + 289, + 278 + ], + "type": "text", + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 287, + 289, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 287, + 289, + 342 + ], + "spans": [ + { + "bbox": [ + 69, + 287, + 289, + 342 + ], + "type": "text", + "content": "Xinya Du and Claire Cardie. 2020. Event extraction by answering (almost) natural questions. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 671-683, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 351, + 289, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 351, + 289, + 429 + ], + "spans": [ + { + "bbox": [ + 69, + 351, + 289, + 429 + ], + "type": "text", + "content": "Xiaocheng Feng, Lifu Huang, Duyu Tang, Heng Ji, Bing Qin, and Ting Liu. 2016. A language-independent neural network for event detection. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 66-71, Berlin, Germany. Association for Computational Linguistics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 438, + 289, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 438, + 289, + 482 + ], + "spans": [ + { + "bbox": [ + 69, + 438, + 289, + 482 + ], + "type": "text", + "content": "Ralph Grishman. 1997. Information extraction: Techniques and challenges. In International summer school on information extraction, pages 10-27. Springer." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 491, + 289, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 491, + 289, + 579 + ], + "spans": [ + { + "bbox": [ + 69, + 491, + 289, + 579 + ], + "type": "text", + "content": "Karen Hambardzumyan, Hrant Khachatrian, and Jonathan May. 2021. WARP: Word-level Adversarial ReProgramming. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4921-4933, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 588, + 289, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 588, + 289, + 655 + ], + "spans": [ + { + "bbox": [ + 69, + 588, + 289, + 655 + ], + "type": "text", + "content": "Lifu Huang, Taylor Cassidy, Xiaocheng Feng, Heng Ji, Clare Voss, Jiawei Han, and Avirup Sil. 2016. Liberal event extraction and event schema induction. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 258-268." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 664, + 289, + 718 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 664, + 289, + 718 + ], + "spans": [ + { + "bbox": [ + 69, + 664, + 289, + 718 + ], + "type": "text", + "content": "Lifu Huang and Heng Ji. 2020. Semi-supervised new event type induction and event detection. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 718-724." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 728, + 289, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 728, + 289, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 728, + 289, + 772 + ], + "type": "text", + "content": "Lifu Huang, Heng Ji, Kyunghyun Cho, Ido Dagan, Sebastian Riedel, and Clare Voss. 2018. Zero-shot transfer learning for event extraction. In Proceedings of the 56th Annual Meeting of the Association for" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 524, + 772 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 315, + 72, + 524, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 524, + 105 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 524, + 105 + ], + "type": "text", + "content": "Computational Linguistics (Volume 1: Long Papers), pages 2160-2170, Melbourne, Australia. Association for Computational Linguistics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 114, + 524, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 114, + 524, + 147 + ], + "spans": [ + { + "bbox": [ + 304, + 114, + 524, + 147 + ], + "type": "text", + "content": "Heng Ji and Ralph Grishman. 2008. Refining event extraction through cross-document inference. In Proceedings of ACL-08: Hlt, pages 254-262." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 156, + 524, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 156, + 524, + 200 + ], + "spans": [ + { + "bbox": [ + 304, + 156, + 524, + 200 + ], + "type": "text", + "content": "Zhengbao Jiang, Frank F. Xu, J. Araki, and Graham Neubig. 2020. How can we know what language models know? Transactions of the Association for Computational Linguistics, 8:423-438." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 208, + 524, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 208, + 524, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 208, + 524, + 263 + ], + "type": "text", + "content": "Bingyi Kang, Zhuang Liu, Xin Wang, Fisher Yu, Jiashi Feng, and Trevor Darrell. 2019. Few-shot object detection via feature reweighting. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 8419-8428." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 272, + 524, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 272, + 524, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 272, + 524, + 327 + ], + "type": "text", + "content": "Viet Dac Lai, Franck Dernoncourt, and Thien Huu Nguyen. 2020a. Exploiting the matching information in the support set for few shot event classification. Pacific-Asia Conference on Knowledge Discovery and Data Mining, page 233-245." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 335, + 524, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 524, + 369 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 524, + 369 + ], + "type": "text", + "content": "Viet Dac Lai and Thien Huu Nguyen. 2019. Extending event detection to new types with learning from keywords. arXiv preprint arXiv:1910.11368." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 377, + 524, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 524, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 524, + 443 + ], + "type": "text", + "content": "Viet Dac Lai, Thien Huu Nguyen, and Franck Dernoncourt. 2020b. Extensively matching for few-shot learning event detection. In Proceedings of the First Joint Workshop on Narrative Understanding, Storylines, and Events, pages 38-45, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 452, + 524, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 452, + 524, + 485 + ], + "spans": [ + { + "bbox": [ + 304, + 452, + 524, + 485 + ], + "type": "text", + "content": "Brian Lester, Rami Al-Rfou, and Noah Constant. 2021. The power of scale for parameter-efficient prompt tuning. In EMNLP." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 493, + 524, + 549 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 493, + 524, + 549 + ], + "spans": [ + { + "bbox": [ + 304, + 493, + 524, + 549 + ], + "type": "text", + "content": "Bohao Li, Boyu Yang, Chang Liu, Feng Liu, Rongrong Ji, and Qixiang Ye. 2021. Beyond max-margin: Class margin equilibrium for few-shot object detection. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7359-7368." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 557, + 524, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 524, + 623 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 524, + 623 + ], + "type": "text", + "content": "Fayuan Li, Weihua Peng, Yuguang Chen, Quan Wang, Lu Pan, Yajuan Lyu, and Yong Zhu. 2020. Event extraction as multi-turn question answering. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 829–838, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 632, + 524, + 698 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 632, + 524, + 698 + ], + "spans": [ + { + "bbox": [ + 304, + 632, + 524, + 698 + ], + "type": "text", + "content": "Qi Li, Heng Ji, and Liang Huang. 2013. Joint event extraction via structured prediction with global features. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 73-82, Sofia, Bulgaria. Association for Computational Linguistics." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 706, + 524, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 706, + 524, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 706, + 524, + 772 + ], + "type": "text", + "content": "Xiang Lisa Li and Percy Liang. 2021. Prefix-tuning: Optimizing continuous prompts for generation. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), abs/2101.00190." + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1292" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 291, + 128 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 291, + 128 + ], + "type": "text", + "content": "Shasha Liao and Ralph Grishman. 2010. Using document level cross-event inference to improve event extraction. In Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 789-797." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 137, + 290, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 137, + 290, + 204 + ], + "spans": [ + { + "bbox": [ + 69, + 137, + 290, + 204 + ], + "type": "text", + "content": "Ying Lin, Heng Ji, Fei Huang, and Lingfei Wu. 2020. A joint neural model for information extraction with global features. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7999-8009, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 214, + 291, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 214, + 291, + 269 + ], + "spans": [ + { + "bbox": [ + 69, + 214, + 291, + 269 + ], + "type": "text", + "content": "Linguistic Data Consortium. 2005. English annotation guidelines for events. https://www.ldc.upenn.edu/sites/www.ldc.upenn.edu/files/english-events-guidelines-v5.4.3.pdf." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 280, + 290, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 280, + 290, + 346 + ], + "spans": [ + { + "bbox": [ + 69, + 280, + 290, + 346 + ], + "type": "text", + "content": "Jian Liu, Yubo Chen, Kang Liu, Wei Bi, and Xiaojiang Liu. 2020. Event extraction as machine reading comprehension. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1641-1651, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 356, + 290, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 356, + 290, + 465 + ], + "spans": [ + { + "bbox": [ + 69, + 356, + 290, + 465 + ], + "type": "text", + "content": "Yaojie Lu, Hongyu Lin, Jin Xu, Xianpei Han, Jialong Tang, Annan Li, Le Sun, Meng Liao, and Shaoyi Chen. 2021. Text2Event: Controllable sequence-to-structure generation for end-to-end event extraction. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 2795-2806, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 476, + 290, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 476, + 290, + 565 + ], + "spans": [ + { + "bbox": [ + 69, + 476, + 290, + 565 + ], + "type": "text", + "content": "Qing Lyu, Hongming Zhang, Elior Sulem, and Dan Roth. 2021. Zero-shot Event Extraction via Transfer Learning: Challenges and Insights. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 322-332, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 575, + 290, + 630 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 575, + 290, + 630 + ], + "spans": [ + { + "bbox": [ + 69, + 575, + 290, + 630 + ], + "type": "text", + "content": "David McClosky, Mihai Surdeanu, and Christopher D Manning. 2011. Event extraction as dependency parsing. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 1626-1635." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 640, + 290, + 684 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 640, + 290, + 684 + ], + "spans": [ + { + "bbox": [ + 69, + 640, + 290, + 684 + ], + "type": "text", + "content": "Minh Van Nguyen, Viet Dac Lai, and Thien Huu Nguyen. 2021. Cross-task instance representation interactions and label dependencies for joint information extraction with graph convolutional networks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 694, + 290, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 694, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 694, + 290, + 772 + ], + "type": "text", + "content": "Thien Huu Nguyen, Kyunghyun Cho, and Ralph Grishman. 2016. Joint event extraction via recurrent neural networks. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 300-309, San Diego, California. Association for Computational Linguistics." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 525, + 772 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 304, + 72, + 525, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 525, + 127 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 525, + 127 + ], + "type": "text", + "content": "Panupong Pasupat and Percy Liang. 2014. Zero-shot entity extraction from web pages. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 391-401." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 137, + 525, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 137, + 525, + 236 + ], + "spans": [ + { + "bbox": [ + 304, + 137, + 525, + 236 + ], + "type": "text", + "content": "Fabio Petroni, Tim Rocttäschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, and Alexander Miller. 2019. Language models as knowledge bases? In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2463-2473, Hong Kong, China. Association for Computational Linguistics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 245, + 525, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 245, + 525, + 299 + ], + "spans": [ + { + "bbox": [ + 304, + 245, + 525, + 299 + ], + "type": "text", + "content": "Colin Raffel, Noam M. Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 309, + 525, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 309, + 525, + 332 + ], + "spans": [ + { + "bbox": [ + 304, + 309, + 525, + 332 + ], + "type": "text", + "content": "Timo Schick and Hinrich Schütze. 2021a. Few-shot text generation with pattern-exploiting training." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 340, + 525, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 340, + 525, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 525, + 396 + ], + "type": "text", + "content": "Timo Schick and Hinrich Schütze. 2021b. It's not just size that matters: Small language models are also few-shot learners. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics, pages 2339-2352." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 405, + 525, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 405, + 525, + 524 + ], + "spans": [ + { + "bbox": [ + 304, + 405, + 525, + 524 + ], + "type": "text", + "content": "Shirong Shen, Tongtong Wu, Guilin Qi, Yuan-Fang Li, Gholamreza Haffari, and Sheng Bi. 2021. Adaptive knowledge-enhanced bayesian meta-learning for few-shot event detection. In Findings of the Association for Computational Linguistics, page 2417-2429. Association for Computational Linguistics (ACL). Annual Meeting of the Association for Computational Linguistics and the International Joint Conference on Natural Language Processing 2021, ACL-IJCNLP 2021; Conference date: 01-08-2021 Through 06-08-2021." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 534, + 525, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 525, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 525, + 590 + ], + "type": "text", + "content": "Taylor Shin, Yasaman Razeghi, Robert L. Logan IV, Eric Wallace, and Sameer Singh. 2020. AutoPrompt: Eliciting knowledge from language models with automatically generated prompts. In Empirical Methods in Natural Language Processing (EMNLP)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 599, + 525, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 599, + 525, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 599, + 525, + 677 + ], + "type": "text", + "content": "Zhiyi Song, Ann Bies, Stephanie Strassel, Tom Riese, Justin Mott, Joe Ellis, Jonathan Wright, Seth Kulick, Neville Ryant, and Xiaoyi Ma. 2015. From light to rich ere: annotation of entities, relations, and events. In Proceedings of the 3rd Workshop on EVENTS: Definition, Detection, Coreference, and Representation, pages 89-98." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 686, + 525, + 740 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 686, + 525, + 740 + ], + "spans": [ + { + "bbox": [ + 304, + 686, + 525, + 740 + ], + "type": "text", + "content": "Bo Sun, Banghuai Li, Shengcai Cai, Ye Yuan, and Chi Zhang. 2021. Fsce: Few-shot object detection via contrastive proposal encoding. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7348-7358." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 750, + 525, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 750, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 750, + 525, + 772 + ], + "type": "text", + "content": "David Wadden, Ulme Wennberg, Yi Luan, and Hannaneh Hajishirzi. 2019. Entity, relation, and event" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1293" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 80, + 72, + 291, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 72, + 291, + 150 + ], + "spans": [ + { + "bbox": [ + 80, + 72, + 291, + 150 + ], + "type": "text", + "content": "extraction with contextualized span representations. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5784-5789, Hong Kong, China. Association for Computational Linguistics." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 159, + 291, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 159, + 291, + 215 + ], + "spans": [ + { + "bbox": [ + 69, + 159, + 291, + 215 + ], + "type": "text", + "content": "Richard C Wang and William Cohen. 2009. Character-level analysis of semi-structured documents for set expansion. In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 1503–1512." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 225, + 290, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 225, + 290, + 270 + ], + "spans": [ + { + "bbox": [ + 69, + 225, + 290, + 270 + ], + "type": "text", + "content": "Sijia Wang, Mo Yu, Shiyu Chang, Lichao Sun, and Lifu Huang. 2021a. Query and extract: Refining event extraction as type-oriented binary decoding. arXiv preprint arXiv:2110.07476." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 280, + 290, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 280, + 290, + 336 + ], + "spans": [ + { + "bbox": [ + 69, + 280, + 290, + 336 + ], + "type": "text", + "content": "Xiao Wang, Houye Ji, Chuan Shi, Bai Wang, Yanfang Ye, Peng Cui, and Philip S Yu. 2019. Heterogeneous graph attention network. In The World Wide Web Conference, WWW '19, page 2022-2032, New York, NY, USA. Association for Computing Machinery." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 345, + 290, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 345, + 290, + 400 + ], + "spans": [ + { + "bbox": [ + 69, + 345, + 290, + 400 + ], + "type": "text", + "content": "Xiaozhi Wang, Ziqi Wang, Xu Han, Wangyi Jiang, Rong Han, Zhiyuan Liu, Juanzi Li, Peng Li, Yankai Lin, and Jie Zhou. 2020a. MAVEN: A massive general domain event detection dataset. In Proceedings of EMNLP 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 411, + 290, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 411, + 290, + 444 + ], + "spans": [ + { + "bbox": [ + 69, + 411, + 290, + 444 + ], + "type": "text", + "content": "Xin Wang, Thomas E. Huang, Trevor Darrell, Joseph E Gonzalez, and Fisher Yu. 2020b. Frustratingly simple few-shot object detection." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 454, + 290, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 454, + 290, + 521 + ], + "spans": [ + { + "bbox": [ + 69, + 454, + 290, + 521 + ], + "type": "text", + "content": "Ziqi Wang, Xiaozhi Wang, Xu Han, Yankai Lin, Lei Hou, Zhiyuan Liu, Peng Li, Juanzi Li, and Jie Zhou. 2021b. CLEVE: Contrastive Pre-training for Event Extraction. In Proceedings of ACL-IJCNLP, pages 6283-6297, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 531, + 290, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 531, + 290, + 564 + ], + "spans": [ + { + "bbox": [ + 69, + 531, + 290, + 564 + ], + "type": "text", + "content": "Yang Xiao and Renaud Marlet. 2020. Few-shot object detection and viewpoint estimation for objects in the wild. In ECCV." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 575, + 290, + 630 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 575, + 290, + 630 + ], + "spans": [ + { + "bbox": [ + 69, + 575, + 290, + 630 + ], + "type": "text", + "content": "Xiaopeng Yan, Ziliang Chen, Anni Xu, Xiaoxi Wang, Xiaodan Liang, and Liang Lin. 2019. Meta r-cnn: Towards general solver for instance-level low-shot learning. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9576-9585." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 640, + 290, + 718 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 640, + 290, + 718 + ], + "spans": [ + { + "bbox": [ + 69, + 640, + 290, + 718 + ], + "type": "text", + "content": "Bishan Yang and Tom M. Mitchell. 2016. Joint extraction of events and entities within a document context. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 289-299, San Diego, California. Association for Computational Linguistics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 728, + 290, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 728, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 728, + 290, + 772 + ], + "type": "text", + "content": "Weizhe Yuan, Graham Neubig, and Pengfei Liu. 2021. BARTScore: Evaluating generated text as text generation. In Advances in Neural Information Processing Systems." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 524, + 288 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 304, + 72, + 524, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 524, + 137 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 524, + 137 + ], + "type": "text", + "content": "Gongjie Zhang, Kaiwen Cui, Rongliang Wu, Shijian Lu, and Yonghong Tian. 2021a. Pnpdet: Efficient few-shot detection without forgetting via plug-and-play sub-networks. 2021 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 3822-3831." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 147, + 524, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 147, + 524, + 213 + ], + "spans": [ + { + "bbox": [ + 304, + 147, + 524, + 213 + ], + "type": "text", + "content": "Hongming Zhang, Haoyu Wang, and Dan Roth. 2021b. Zero-shot Label-aware Event Trigger and Argument Classification. In *Findings of the Association for Computational Linguistics: ACL-IJCNLP* 2021, pages 1331-1340, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 222, + 524, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 222, + 524, + 288 + ], + "spans": [ + { + "bbox": [ + 304, + 222, + 524, + 288 + ], + "type": "text", + "content": "Tongtao Zhang, Spencer Whitehead, Hanwang Zhang, Hongzhi Li, Joseph Ellis, Lifu Huang, Wei Liu, Heng Ji, and Shih-Fu Chang. 2017. Improving event extraction via multimodal integration. In Proceedings of the 25th ACM international conference on Multimedia, pages 270-278." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1294" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 70, + 261, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 70, + 261, + 84 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 261, + 84 + ], + "type": "text", + "content": "A APEX prompt examples for ACE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 92, + 289, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 92, + 289, + 119 + ], + "spans": [ + { + "bbox": [ + 67, + 92, + 289, + 119 + ], + "type": "text", + "content": "Table 4 and Table 5 show APEX prompt examples for ACE events." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 130, + 214, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 130, + 214, + 144 + ], + "spans": [ + { + "bbox": [ + 67, + 130, + 214, + 144 + ], + "type": "text", + "content": "B Prompt Token Selection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 153, + 291, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 153, + 291, + 370 + ], + "spans": [ + { + "bbox": [ + 67, + 153, + 291, + 370 + ], + "type": "text", + "content": "In our experiments, the event type names and event type structures are automatically extracted from the target event ontology, such as ACE (Linguistic Data Consortium, 2005), ERE (Song et al., 2015) and MAVEN (Wang et al., 2020a). The prototype seed triggers are automatically selected from the annotated data for supervised and few-shot event extraction. For zero-shot event extraction, we manually select " + }, + { + "bbox": [ + 67, + 153, + 291, + 370 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 67, + 153, + 291, + 370 + ], + "type": "text", + "content": " words from the NLTK synonyms of each event type as its prototype seed triggers. The definitions and APEX prompts are based on the official annotation guides for each target event ontology (Linguistic Data Consortium, 2005; Song et al., 2015; Wang et al., 2020a) and the available definitions in FrameNet (Baker et al., 1998) with manual editing." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 380, + 185, + 394 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 380, + 185, + 394 + ], + "spans": [ + { + "bbox": [ + 67, + 380, + 185, + 394 + ], + "type": "text", + "content": "C Learning Strategy" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 401, + 291, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 401, + 291, + 700 + ], + "spans": [ + { + "bbox": [ + 67, + 401, + 291, + 700 + ], + "type": "text", + "content": "The learning strategy varies for supervised, few-shot, and zero-shot learning. For supervised learning, we optimize the following objective for event trigger detection " + }, + { + "bbox": [ + 67, + 401, + 291, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = -\\frac{1}{|\\mathcal{T}||\\mathcal{N}|}\\sum_{t\\in \\mathcal{T}}\\sum_{i = 1}^{|\\mathcal{N}|}\\boldsymbol{y}_i^t" + }, + { + "bbox": [ + 67, + 401, + 291, + 700 + ], + "type": "inline_equation", + "content": "\\log \\tilde{\\boldsymbol{y}}_i^t" + }, + { + "bbox": [ + 67, + 401, + 291, + 700 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 67, + 401, + 291, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 67, + 401, + 291, + 700 + ], + "type": "text", + "content": " is the set of target event types and " + }, + { + "bbox": [ + 67, + 401, + 291, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 67, + 401, + 291, + 700 + ], + "type": "text", + "content": " is the set of tokens from the training dataset. " + }, + { + "bbox": [ + 67, + 401, + 291, + 700 + ], + "type": "inline_equation", + "content": "\\boldsymbol{y}_i^t" + }, + { + "bbox": [ + 67, + 401, + 291, + 700 + ], + "type": "text", + "content": " denotes the ground truth label vector. For few-shot event detection, we optimize the model on both base training data set and the smaller training data set for novel event types: " + }, + { + "bbox": [ + 67, + 401, + 291, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = -\\frac{1}{|\\mathcal{T}^B||\\mathcal{N}^B|}\\sum_{t\\in \\mathcal{T}^B}\\sum_{i = 1}^{|\\mathcal{N}^B|}\\boldsymbol{y}_i^t\\cdot \\log \\tilde{\\boldsymbol{y}}_i^t -" + }, + { + "bbox": [ + 67, + 401, + 291, + 700 + ], + "type": "inline_equation", + "content": "\\beta \\frac{1}{|\\mathcal{T}^N||\\mathcal{N}^N|}\\sum_{t\\in \\mathcal{T}^N}\\sum_{i = 1}^{|\\mathcal{N}^N|}\\boldsymbol{y}_i^t\\cdot \\log \\tilde{\\boldsymbol{y}}_i^t" + }, + { + "bbox": [ + 67, + 401, + 291, + 700 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 67, + 401, + 291, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^B" + }, + { + "bbox": [ + 67, + 401, + 291, + 700 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 401, + 291, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{N}^B" + }, + { + "bbox": [ + 67, + 401, + 291, + 700 + ], + "type": "text", + "content": " denote the set of base event types and tokens from the base training data set, respectively. " + }, + { + "bbox": [ + 67, + 401, + 291, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^N" + }, + { + "bbox": [ + 67, + 401, + 291, + 700 + ], + "type": "text", + "content": " is the set of novel event types. " + }, + { + "bbox": [ + 67, + 401, + 291, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{N}^N" + }, + { + "bbox": [ + 67, + 401, + 291, + 700 + ], + "type": "text", + "content": " is the set of tokens from the training data set for novel event types. " + }, + { + "bbox": [ + 67, + 401, + 291, + 700 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 67, + 401, + 291, + 700 + ], + "type": "text", + "content": " is a hyper-parameter to balance the two objectives. For zero-shot event detection, as we only have the base training data set, we minimize the following objective: " + }, + { + "bbox": [ + 67, + 401, + 291, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = -\\frac{1}{|\\mathcal{T}^B||\\mathcal{N}^B|}\\sum_{t\\in \\mathcal{T}^B}\\sum_{i = 1}^{|\\mathcal{N}^B|}\\boldsymbol{y}_i^t" + }, + { + "bbox": [ + 67, + 401, + 291, + 700 + ], + "type": "inline_equation", + "content": "\\log \\tilde{\\boldsymbol{y}}_i^t" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 710, + 131, + 723 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 710, + 131, + 723 + ], + "spans": [ + { + "bbox": [ + 67, + 710, + 131, + 723 + ], + "type": "text", + "content": "D Dataset" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 733, + 290, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 733, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 733, + 290, + 772 + ], + "type": "text", + "content": "After defining the base and novel event types, we create the training, validation, and evaluation split for all three datasets. We use the sentences with" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 71, + 526, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 315 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 315 + ], + "type": "text", + "content": "only base event type mentions as the base training data set for few-shot event detection, and randomly select 10 sentences with novel event type mentions as the additional smaller training data set. We use the sentences with both base and novel event type mentions as the development set and use the remaining sentences with only novel event type mentions as the evaluation dataset. We use the same development and evaluation set as few-shot event detection for zero-shot event detection and remove the instances with novel event mentions from the training set. We randomly split the sentences without any event annotations proportionally to the number of sentences with event mentions in each set for both zero-shot and few-shot event detection. Table 6 shows the detailed data statistics for all the three datasets under the few-shot and zero-shot event extraction settings." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 324, + 498, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 324, + 498, + 338 + ], + "spans": [ + { + "bbox": [ + 302, + 324, + 498, + 338 + ], + "type": "text", + "content": "E Hyperparameters and Evaluation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 346, + 526, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 346, + 526, + 534 + ], + "spans": [ + { + "bbox": [ + 302, + 346, + 526, + 534 + ], + "type": "text", + "content": "For a fair comparison with the previous baseline approaches, we use the same pre-trained bert-large-uncased model for fine-tuning and optimizing our model with BertAdam. For supervised event detection, we optimize the parameters with grid search: training epoch is 3, learning rate " + }, + { + "bbox": [ + 302, + 346, + 526, + 534 + ], + "type": "inline_equation", + "content": "\\in [3e - 6,1e - 4]" + }, + { + "bbox": [ + 302, + 346, + 526, + 534 + ], + "type": "text", + "content": ", training batch size " + }, + { + "bbox": [ + 302, + 346, + 526, + 534 + ], + "type": "inline_equation", + "content": "\\in" + }, + { + "bbox": [ + 302, + 346, + 526, + 534 + ], + "type": "text", + "content": " {8, 12, 16, 24, 32}, dropout rate " + }, + { + "bbox": [ + 302, + 346, + 526, + 534 + ], + "type": "inline_equation", + "content": "\\in" + }, + { + "bbox": [ + 302, + 346, + 526, + 534 + ], + "type": "text", + "content": " {0.4, 0.5, 0.6}. The running time is up to 3 hours on one Quadro RTX 8000. For evaluation, we use the same criteria as previous studies (Li et al., 2013; Chen et al., 2015; Nguyen et al., 2016; Lin et al., 2020): an event mention is correct if its span and event type match a reference event mention." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1295" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 99, + 114, + 495, + 705 + ], + "blocks": [ + { + "bbox": [ + 99, + 114, + 495, + 705 + ], + "lines": [ + { + "bbox": [ + 99, + 114, + 495, + 705 + ], + "spans": [ + { + "bbox": [ + 99, + 114, + 495, + 705 + ], + "type": "table", + "html": "
Event Rep TypeComprehensive Prompt
Business:Declare-BankruptcyDeclare Bankruptcy [SEP] bankruptcy bankruptciesbankrupting [SEP] Organizationrequest legal protection from debt collection at a Place
Business:End-OrgEnd Organization [SEP] dissolving disbanded [SEP] an Organization goes out of business at a Place
Business:Merge-OrgMerge Organization [SEP] merging merger [SEP] two or more Organizations come together to form a new organization at a Place
Business:Start-OrgStart Organization [SEP] founded [SEP] an Agent create a new Organization at a Place
Conflict:AttackAttack [SEP] invaded airstrikes overthrew ambushed [SEP] An Attacker physically attacks a Target with Instrument at a Place
Conflict:DemonstrateDemonstrate [SEP] demonstrations protest strikes riots [SEP] Entities come together in a Place to protest or demand official action
Contact:MeetMeet [SEP] reunited retreats [SEP] two or more Entities come together at same Place and interact in person
Contact:Phone-WritePhone Write [SEP] emailed letter [SEP] phone or written communication between two or more Entities
Justice:AcquitAcquit [SEP] acquitted [SEP] a trial of Defendant ends but Adjudicator fails to produce a conviction at a Place
Justice:AppealAppeal [SEP] appeal [SEP] the decision for Defendant of a court is taken to a higher court for Adjudicator review with Prosecutor
Justice:Arrest-JailArrest Jail [SEP] arrested locked [SEP] the Agent takes custody of a Person at a Place
Justice:Charge-IndictCharge Indict [SEP] indictment [SEP] a Defendant is accused of a crime by a Prosecutor for Adjudicator
Justice:ConvictConvict [SEP] pled guilty convicting [SEP] an Defendant found guilty of a crime by Adjudicator at a Place
Justice:ExecuteExecute [SEP] death [SEP] the life of a Person is taken by an Agent at a Place
Justice:ExtraditeExtradite [SEP] extradition [SEP] a Person is sent by an Agent from Origin to Destination
Justice:FineFine [SEP] payouts financial punishment [SEP] a Adjudicator issues a financial punishment Money to an Entity at a Place
Justice:PardonPardon [SEP] pardoned lift sentence [SEP] an Adjudicator lifts a sentence of Defendant at a Place
Justice:Release-ParoleRelease Parole [SEP] parole [SEP] an Entity ends its custody of a Person at a Place
Justice:SentenceSentence [SEP] sentenced punishment [SEP] the punishment for the defendant is issued by a state actor
Justice:SueSue [SEP] lawsuits [SEP] Plaintiff initiate a court proceeding to determine the liability of a Defendant judge by Adjudicator at a Place
Justice:Trial-HearingTrial Hearing [SEP] trial hearings [SEP] a court proceeding initiated to determine the guilty or innocence of a Person with Prosecutor and Adjudicator at a Place
Life:Be-BornBe Born [SEP] childbirth [SEP] a Person is born at a Place
Life:DieDie [SEP] deceased extermination [SEP] life of a Victim ends by an Agent with Instrument at a Place
", + "image_path": "f4003b7e10c48c68ad81d17bd1d16fee111d107424f11b5ec7f4d40f31d50f71.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 202, + 711, + 390, + 724 + ], + "lines": [ + { + "bbox": [ + 202, + 711, + 390, + 724 + ], + "spans": [ + { + "bbox": [ + 202, + 711, + 390, + 724 + ], + "type": "text", + "content": "Table 4: APEX templates for ACE event types" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "text", + "content": "1296" + } + ] + } + ], + "index": 2 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 99, + 131, + 495, + 396 + ], + "blocks": [ + { + "bbox": [ + 99, + 131, + 495, + 396 + ], + "lines": [ + { + "bbox": [ + 99, + 131, + 495, + 396 + ], + "spans": [ + { + "bbox": [ + 99, + 131, + 495, + 396 + ], + "type": "table", + "html": "
Event Rep TypeComprehensive Prompt
Life:DivorceDivorce [SEP] people divorce [SEP] two Person are officially divorced at a place
Life:InjureInjure [SEP] hospitalised paralyzed dismember [SEP] a Victim experiences physical harm from Agent with Instrument at a Place
Life:MarryMarry [SEP] married marriage marry [SEP] two Person are married at a Place
Movement:TransportTransport [SEP] arrival travels penetrated expelled [SEP] an Agent moves an Artifact from Origin to Destination with Vehicle at Price
Personnel:ElectElect [SEP] reelected elected election [SEP] a candidate Person wins an election by voting Entity at a Place
Personnel:End-PositionEnd Position [SEP] resigning retired resigned [SEP] a Person stops working for an Entity or change office at a Place
Personnel:NominateNominate [SEP] nominate [SEP] a Person is nominated for a new position by another Agent at a Place
Personnel:Start-PositionStart Position [SEP] hiring rehired recruited [SEP] a Person begins working for an Entity or change office at a Place
Transaction:Transfer-MoneyTransfer Money [SEP] donations reimbursing deductions [SEP] transfer Money from the Giver to the Beneficiary or Recipient at a Place
Transaction:Transfer-OwnershipTransfer Ownership [SEP] purchased buy sell loan [SEP] buying selling loaning borrowing giving receiving of Artifacts from Seller to Buyer or Beneficiary at a Place at Price
", + "image_path": "db85a5d2ca78e6e95630c233c483aedd29200088a1a4fb1d0d8e52809303bdfa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 158, + 550, + 436, + 687 + ], + "blocks": [ + { + "bbox": [ + 178, + 405, + 415, + 417 + ], + "lines": [ + { + "bbox": [ + 178, + 405, + 415, + 417 + ], + "spans": [ + { + "bbox": [ + 178, + 405, + 415, + 417 + ], + "type": "text", + "content": "Table 5: APEX templates for ACE event types (continued)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 158, + 550, + 436, + 687 + ], + "lines": [ + { + "bbox": [ + 158, + 550, + 436, + 687 + ], + "spans": [ + { + "bbox": [ + 158, + 550, + 436, + 687 + ], + "type": "table", + "html": "
DatasetACE05-E+ERE-ENMAVEN
# TypesBase1825120
Novel101045
# MentionsBase3572544993675
Novel172431833201
TrainFew-shot3216388688085
Zero-shot3116378687635
Validation900(51%/49%)2797(53%/47%)3883(71%/23%)
Evaluation119520121652
", + "image_path": "54b225e7e82ce612a9251ab2cd6f081886751c268200cbb035f9b68557ab1e03.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 694, + 524, + 707 + ], + "lines": [ + { + "bbox": [ + 67, + 694, + 524, + 707 + ], + "spans": [ + { + "bbox": [ + 67, + 694, + 524, + 707 + ], + "type": "text", + "content": "Table 6: Data statistics for ACE2005, ERE and MAVEN datasets under few-shot/zero-shot event detection settings." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1297" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "spans": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "type": "text", + "content": "A For every submission:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 106, + 414, + 243 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "spans": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "type": "text", + "content": "A1. Did you describe the limitations of your work? Section 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 142, + 329, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 142, + 329, + 169 + ], + "spans": [ + { + "bbox": [ + 77, + 142, + 329, + 169 + ], + "type": "text", + "content": "A2. Did you discuss any potential risks of your work? Section 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 179, + 414, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 179, + 414, + 206 + ], + "spans": [ + { + "bbox": [ + 77, + 179, + 414, + 206 + ], + "type": "text", + "content": "A3. Do the abstract and introduction summarize the paper's main claims? Abstract and Section 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 77, + 215, + 398, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 215, + 398, + 243 + ], + "spans": [ + { + "bbox": [ + 77, + 215, + 398, + 243 + ], + "type": "text", + "content": "A4. Have you used AI writing assistants when working on this paper? Left blank." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 68, + 252, + 290, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 252, + 290, + 266 + ], + "spans": [ + { + "bbox": [ + 68, + 252, + 290, + 266 + ], + "type": "text", + "content": "B Did you use or create scientific artifacts?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 79, + 270, + 196, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 270, + 196, + 283 + ], + "spans": [ + { + "bbox": [ + 79, + 270, + 196, + 283 + ], + "type": "text", + "content": "Section 5 and Appendix C" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 77, + 291, + 524, + 634 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 77, + 291, + 315, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 291, + 315, + 319 + ], + "spans": [ + { + "bbox": [ + 77, + 291, + 315, + 319 + ], + "type": "text", + "content": "B1. Did you cite the creators of artifacts you used? Section 5 and Appendix C" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 77, + 327, + 463, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 327, + 463, + 354 + ], + "spans": [ + { + "bbox": [ + 77, + 327, + 463, + 354 + ], + "type": "text", + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Section 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 77, + 364, + 524, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 364, + 524, + 430 + ], + "spans": [ + { + "bbox": [ + 77, + 364, + 524, + 430 + ], + "type": "text", + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Section 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 77, + 441, + 524, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 441, + 524, + 494 + ], + "spans": [ + { + "bbox": [ + 77, + 441, + 524, + 494 + ], + "type": "text", + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Left blank." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 77, + 503, + 524, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 503, + 524, + 544 + ], + "spans": [ + { + "bbox": [ + 77, + 503, + 524, + 544 + ], + "type": "text", + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Section 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 77, + 553, + 524, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 553, + 524, + 634 + ], + "spans": [ + { + "bbox": [ + 77, + 553, + 524, + 634 + ], + "type": "text", + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Appendix C" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "spans": [ + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "type": "text", + "content": "C Did you run computational experiments?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 79, + 661, + 196, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 661, + 196, + 674 + ], + "spans": [ + { + "bbox": [ + 79, + 661, + 196, + 674 + ], + "type": "text", + "content": "Section 5 and Appendix D" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 77, + 682, + 524, + 724 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 682, + 524, + 724 + ], + "spans": [ + { + "bbox": [ + 77, + 682, + 524, + 724 + ], + "type": "text", + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Appendix D" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "spans": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "text", + "content": "ACL 2023 Responsible NLP Checklist" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "spans": [ + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "type": "text", + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1298" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 77, + 70, + 523, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 70, + 523, + 97 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 523, + 97 + ], + "type": "text", + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 89, + 99, + 207, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 99, + 207, + 111 + ], + "spans": [ + { + "bbox": [ + 89, + 99, + 207, + 111 + ], + "type": "text", + "content": "Section 5 and Appendix D" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 120, + 525, + 160 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 120, + 525, + 160 + ], + "spans": [ + { + "bbox": [ + 77, + 120, + 525, + 160 + ], + "type": "text", + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 162, + 144, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 162, + 144, + 174 + ], + "spans": [ + { + "bbox": [ + 89, + 162, + 144, + 174 + ], + "type": "text", + "content": "Appendix D" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 183, + 525, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 183, + 525, + 223 + ], + "spans": [ + { + "bbox": [ + 77, + 183, + 525, + 223 + ], + "type": "text", + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 225, + 143, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 225, + 143, + 238 + ], + "spans": [ + { + "bbox": [ + 89, + 225, + 143, + 238 + ], + "type": "text", + "content": "Appendix B" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 247, + 522, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 247, + 522, + 260 + ], + "spans": [ + { + "bbox": [ + 67, + 247, + 522, + 260 + ], + "type": "text", + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 79, + 264, + 127, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 264, + 127, + 277 + ], + "spans": [ + { + "bbox": [ + 79, + 264, + 127, + 277 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 76, + 286, + 525, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 286, + 525, + 313 + ], + "spans": [ + { + "bbox": [ + 76, + 286, + 525, + 313 + ], + "type": "text", + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 89, + 315, + 148, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 315, + 148, + 327 + ], + "spans": [ + { + "bbox": [ + 89, + 315, + 148, + 327 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 76, + 336, + 525, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 336, + 525, + 376 + ], + "spans": [ + { + "bbox": [ + 76, + 336, + 525, + 376 + ], + "type": "text", + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 89, + 378, + 148, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 378, + 148, + 390 + ], + "spans": [ + { + "bbox": [ + 89, + 378, + 148, + 390 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 76, + 400, + 525, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 400, + 525, + 439 + ], + "spans": [ + { + "bbox": [ + 76, + 400, + 525, + 439 + ], + "type": "text", + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 89, + 441, + 148, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 441, + 148, + 454 + ], + "spans": [ + { + "bbox": [ + 89, + 441, + 148, + 454 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 76, + 462, + 520, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 462, + 520, + 476 + ], + "spans": [ + { + "bbox": [ + 76, + 462, + 520, + 476 + ], + "type": "text", + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 89, + 477, + 148, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 477, + 148, + 489 + ], + "spans": [ + { + "bbox": [ + 89, + 477, + 148, + 489 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 76, + 498, + 524, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 498, + 524, + 524 + ], + "spans": [ + { + "bbox": [ + 76, + 498, + 524, + 524 + ], + "type": "text", + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 89, + 527, + 148, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 527, + 148, + 539 + ], + "spans": [ + { + "bbox": [ + 89, + 527, + 148, + 539 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1299" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 13 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/The Ecological Fallacy in Annotation_ Modeling Human Label Variation goes beyond Sociodemographics/8b624391-c7ce-443d-9caf-875f1d838500_content_list.json b/2023/The Ecological Fallacy in Annotation_ Modeling Human Label Variation goes beyond Sociodemographics/8b624391-c7ce-443d-9caf-875f1d838500_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a39fb48e920ca393e81c67fc1e78a198791a7440 --- /dev/null +++ b/2023/The Ecological Fallacy in Annotation_ Modeling Human Label Variation goes beyond Sociodemographics/8b624391-c7ce-443d-9caf-875f1d838500_content_list.json @@ -0,0 +1,1851 @@ +[ + { + "type": "text", + "text": "The Ecological Fallacy in Annotation: Modelling Human Label Variation goes beyond Sociodemographics", + "text_level": 1, + "bbox": [ + 149, + 90, + 848, + 130 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Matthias Orlikowski1, Paul Röttger2, Philipp Cimiano1, and Dirk Hovy3", + "bbox": [ + 189, + 143, + 811, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Bielefeld University", + "bbox": [ + 416, + 174, + 586, + 191 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ University of Oxford", + "bbox": [ + 411, + 191, + 589, + 206 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Computing Sciences Department, Bocconi University, Milan, Italy", + "bbox": [ + 228, + 206, + 773, + 224 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 252, + 339, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Many NLP tasks exhibit human label variation, where different annotators give different labels to the same texts. This variation is known to depend, at least in part, on the sociodemographics of annotators. Recent research aims to model individual annotator behaviour rather than predicting aggregated labels, and we would expect that sociodemographic information is useful for these models. On the other hand, the ecological fallacy states that aggregate group behaviour, such as the behaviour of the average female annotator, does not necessarily explain individual behaviour. To account for sociodemographics in models of individual annotator behaviour, we introduce group-specific layers to multi-annotator models. In a series of experiments for toxic content detection, we find that explicitly accounting for sociodemographic attributes in this way does not significantly improve model performance. This result shows that individual annotation behaviour depends on much more than just sociodemographics.", + "bbox": [ + 141, + 280, + 460, + 593 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 114, + 606, + 258, + 621 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Different annotators will not necessarily assign the same labels to the same texts, resulting in human label variation (Plank, 2022). Previous work finds that this variation depends at least in part on the sociodemographics of annotators, such as their age and gender (Binns et al., 2017; Al Kuwatly et al., 2020; Excell and Al Moubayed, 2021; Shen and Rose, 2021). These results are particularly pronounced for subjective tasks like toxic content detection (Sap et al., 2019; Kumar et al., 2021; Sap et al., 2022; Goyal et al., 2022). Since human label variation is relevant to a wide range of NLP tasks, recent research has begun to model individual annotator behaviour, rather than predicting aggregated labels (Davani et al., 2022; Gordon et al., 2022). In this setting, we would expect sociodemographic attributes to help explain annotator decisions. Therefore, we investigate whether explicitly", + "bbox": [ + 112, + 629, + 489, + 917 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0261e22efde8dee493e93beebcd49541625a911d2b7967e05ba5f1df0447ab78.jpg", + "image_caption": [ + "Figure 1: Group-specific layers representing annotator sociodemographics in multi-annotator models." + ], + "image_footnote": [], + "bbox": [ + 515, + 252, + 877, + 365 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "accounting for the sociodemographic attributes of annotators leads to better predictions of their annotation behaviour1.", + "bbox": [ + 507, + 432, + 882, + 478 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "There is a risk of misreading these efforts as an example of the ecological fallacy: aggregate group behaviour does not necessarily explain individual behaviour (Robinson, 1950; Freedman, 2015). For example, while on average, white annotators may be more likely to label African-American Vernacular English as toxic (Sap et al., 2019), that does not mean it is true for every white annotator individually. However, we aim at exactly this distinction to discuss the relevance of sociodemographic groups in models of individual annotator behaviour. Likewise, we do not assume prior work to commit ecological fallacies, even if a less-nuanced read might suggest it.", + "bbox": [ + 507, + 481, + 884, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Davani et al. (2022) introduce a simple multi-annotator model, where each annotator is modelled with a separate classification head. We expand their model with group-specific layers, which are activated for each annotator based on their sociodemographic attributes. We compare the two model setups to a control setup where we randomise group assignments. All comparisons use annotator-level toxicity data from Kumar et al. (2021). We find that find that explicitly accounting for sociodem", + "bbox": [ + 507, + 707, + 882, + 868 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Code to run our experiments and analyses is available at https://github.com/morlikowski/ecological-fallacy", + "bbox": [ + 507, + 879, + 882, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1017", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics", + "bbox": [ + 226, + 945, + 769, + 957 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Volume 2: Short Papers, pages 1017-1029", + "bbox": [ + 368, + 958, + 628, + 971 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "July 9-14, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 295, + 972, + 700, + 985 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "graphic attributes does not significantly improve model performance. This result suggests that human label variation happens at a more individual level than sociodemographics, and that annotator decisions are even more complex.", + "bbox": [ + 112, + 84, + 487, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions 1) We introduce group-specific layers to model groups of annotators with shared attributes in multi-annotator models. 2) We evaluate the effect of group-specific layers for toxic content detection, and show that explicitly accounting for sociodemographic attributes does not significantly improve performance, thus highlighting the risk of the ecological fallacy in annotator modelling.", + "bbox": [ + 112, + 171, + 487, + 300 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As a corollary, we show that multi-annotator models can be applied to many times more annotators than in prior work.", + "bbox": [ + 112, + 300, + 489, + 350 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 112, + 357, + 270, + 372 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Sociodemographics in Annotation Behaviour A growing body of research studies how annotator sociodemographics relate to their annotation decisions, for tasks ranging from natural language inference (Biester et al., 2022) to the detection of racist (Larimore et al., 2021) or generally toxic (Sap et al., 2022) language. Goyal et al. (2022), for example, find that annotators from certain sociodemographic groups (e.g., LGBTQ people) tend to find content attacking their own groups (e.g., homophobic content) to be more toxic. This motivates our research into explicitly accounting for sociodemographics to model annotation behaviour. However, the link between sociodemographics and behaviour is not uncontested. Biester et al. (2022), for example, do not find significant differences in annotation behaviour between annotators of different genders for four different tasks.", + "bbox": [ + 112, + 380, + 487, + 670 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Predicting Annotators' Decisions on Text Different from analyses of annotation behaviour, a recent line of research attempts to learn models based on individual annotations (Plank et al., 2014; Jamison and Gurevych, 2015; Akhtar et al., 2020; Fornaciari et al., 2021; Cercas Curry et al., 2021). These models are motivated by the concern that aggregating labels into a single \"truth\" is too simplistic for many tasks (Uma et al., 2021; Basile et al., 2021) and might introduce uneven representation of perspectives (Prabhakaran et al., 2021; Abercrombie et al., 2022).", + "bbox": [ + 112, + 677, + 489, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A particular way of learning from disaggregated labels are models that predict individual annotator decisions for an example. Our work builds directly", + "bbox": [ + 112, + 871, + 487, + 919 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "on such a model, multi-annotator models (Davani et al., 2022), which we describe in more detail separately ( $\\S 4$ ). Gordon et al. (2022) present a model which also predicts individual annotations and allows a user to interactively aggregate them based on \"a jury\" inspired by the US judicial system. Their work is similar to ours in central aspects as they explicitly model annotators' sociodemographics and use the same dataset as we do (Kumar et al., 2021). Different from our work, they frame the task as a regression problem and develop a model based on recommender systems. While they also explore ecological fallacies, they focus on usage risks of their system and countermeasures. In contrast, we consider the issue of the ecological fallacy in modelling annotation behaviour more generally. We compare our findings to their results ( $\\S 6$ ).", + "bbox": [ + 507, + 84, + 884, + 357 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3 Data", + "text_level": 1, + "bbox": [ + 509, + 363, + 588, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We use a sample of the Kumar et al. (2021) dataset for our experiments. The full dataset contains 107,620 English comments from Twitter, Reddit, and 4Chan, annotated for toxicity by 17,280 annotators. The annotation process encouraged annotator subjectivity (Röttger et al., 2022) which is a desired feature for modelling annotator behaviour. For each annotator, there is extensive sociodemographic information, collected with a survey. Annotations are given as ratings on a five-point scale which we convert to binary annotations by mapping ratings of 2 to 4 to toxic, and ratings 0 and 1 to non-toxic.", + "bbox": [ + 507, + 387, + 884, + 595 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We randomly sample comments from the dataset until we reach annotations from more than 5,000 annotators. We then add all other annotations by these annotators. This approach maximizes the number of examples while controlling the number of annotators in our sample.", + "bbox": [ + 507, + 596, + 882, + 692 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our final sample contains 111,780 annotations from 5,002 annotators on 22,360 comments with 20 to 120 annotations per annotator (mean 22.35). Most comments have five annotations. 20 comments have four because we removed any underage annotators before sampling. In total 78,357 annotations $(70.10\\%)$ are toxic, and 33,423 annotations $(29.90\\%)$ are non-toxic.", + "bbox": [ + 507, + 693, + 884, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We focus on four sociodemographic attributes: gender, age, education, and sexual orientation. Group sizes vary by attribute. For gender, 2,450 annotators $(48.98\\%)$ identify as female, 2,116 $(42.30\\%)$ as male, 23 $(0.46\\%)$ as non-binary (rest in residual categories, full statistics in A.1).", + "bbox": [ + 507, + 822, + 882, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1018", + "bbox": [ + 482, + 927, + 519, + 940 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 112, + 84, + 258, + 99 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We compare three models. The baseline model is the multi-annotator model by Davani et al. (2022). We use their multi-task variant: For each annotator, there is a separate classification layer trained on annotations from that annotator. All annotator layers share a pre-trained language model used to encode the input. We use RoBERTa (Liu et al., 2019) for this, motivated by computational constraints. The other models in our experiments build on this baseline model.", + "bbox": [ + 112, + 107, + 489, + 267 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the sociodemographic models, we add group-specific layers based on sociodemographic attributes of the annotators. A single attribute, e.g., age, implies several groups, e.g., ages 25-34, ages 35-44. We add the group-specific layers between the pre-trained model and the annotator layers. Each group of annotators shares a separate group-specific layer. We implement group-specific layers as fully-connected, linear layers, each learning a feature transformation applied for one group of annotators.", + "bbox": [ + 112, + 269, + 489, + 445 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, for the random models, we shuffle the assignment of annotators to groups from the sociodemographic model, retaining the relative group sizes. In other words, the probability of each annotator staying in the same group or being reassigned to another group corresponds to the relative size of each group. This approach keeps the model architecture constant while removing the connection between actual sociodemographic attributes and group assignment. It allows us to distinguish the effects of additional parameters, which group-specific layers add in comparison to the baseline, from the effects of sociodemographic information.", + "bbox": [ + 112, + 447, + 489, + 655 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1 Evaluation Setup", + "text_level": 1, + "bbox": [ + 112, + 664, + 294, + 678 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We evaluate all models on individual annotations from gender, age, education, and sexual orientation groups. This setup is comparable to the \"individual label\" evaluations in Davani et al. (2022) and Gordon et al. (2022), but with scores calculated per group of annotators. We measure performance in macro-average $F_{1}$ , to weigh each class equally.", + "bbox": [ + 112, + 686, + 489, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Cross-Validation As there is no standard split available for our dataset, we perform three iterations of a four-fold cross-validation with different seeds (training details in Appendix A.3). We choose four folds, so that even very small groups have more than a hundred annotations in each test set. Across folds, the numbers of annotations per", + "bbox": [ + 112, + 806, + 489, + 919 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sociodemographic group are similar (see Appendix A.4). We construct test sets that only contain comments unseen by the annotators in the training set. We also ensure that all test sets have similar proportions of toxic or non-toxic comments (assigned by the majority of annotators) to address the class imbalance in the dataset (70.62% toxic, see §3).", + "bbox": [ + 507, + 84, + 884, + 197 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Statistical Significance We test for statistical significance of our results from multiple runs of k-fold cross-validation via replicability analysis (Dror et al., 2017). We report the number of significant folds and the Bonferroni-corrected count (Dror et al., 2018) in Appendix A.2. We compute the p-values for each fold via a paired bootstrap-sampling test with BooStSa (Fornaciari et al., 2022). We set the significance level $\\alpha = 0.05$ , draw 1000 bootstrap samples per fold, and use a sample size of $50\\%$ of the respective test set.", + "bbox": [ + 507, + 205, + 884, + 381 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Remarks on Groups Annotators from different groups of the same attribute will in most cases not have annotated the same examples. Therefore, comparisons between models are only meaningful within each group.", + "bbox": [ + 507, + 388, + 882, + 469 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The groups modeled via group-specific layers and those in the result tables are always the same. For example, if we report scores for gender groups, then the sociodemographic and randomized models are also based on gender groups. In the following, we focus on a subset of groups, omitting, e.g., \"Prefer not to say\" (see Appendix A.5).", + "bbox": [ + 507, + 470, + 882, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "5 Results", + "text_level": 1, + "bbox": [ + 509, + 590, + 606, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1 shows the results for gender, age, education, and sexual orientation. A naive majority class baseline that predicts all input to be toxic performs worse than all other models with a large margin (exact results in Appendix A.5).", + "bbox": [ + 507, + 614, + 882, + 695 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sociodemographics vs. Baseline Across attributes, the average scores of the sociodemographic model and the baseline are similar. The sociodemographic model often has a slightly higher average macro F1 than the baseline, but no statistically significant gains. Where average performance is better by several points, as for homosexual annotators, this gain is offset by a large variance in performance (a consequence of small group sizes).", + "bbox": [ + 507, + 702, + 884, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sociodemographics vs. Random We also do not find significant performance differences between sociodemographic group-layer models and the corresponding random group assignment models. For", + "bbox": [ + 507, + 854, + 882, + 919 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1019", + "bbox": [ + 482, + 927, + 519, + 940 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "most groups, the randomized models achieve the highest average scores, but differences to the sociodemographic model are never statistically significant.", + "bbox": [ + 112, + 84, + 489, + 148 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/c24ea2cb73b40d6d979e210364b98816e24e54af3f43d4058505b0f5ce60fa31.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
GenderBaselineSoc-Dem.Random
Male68.00±0.4967.66±0.4667.63±0.53
Female62.23±0.5362.25±1.1962.41±0.92
Nonbinary56.33±6.0056.80±7.2458.00±7.49
", + "bbox": [ + 147, + 155, + 450, + 219 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/346b1a1e433ecaf0ce38f83b147608910c5f7a10c767d1d9bcdd7c808b1dd0e1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AgeBaselineSoc-Dem.Random
18 - 2459.39±1.5860.44±1.0560.52±1.37
25 - 3466.72±0.5666.63±0.8366.92±0.51
35 - 4464.50±0.5964.94±1.3365.24±0.89
45 - 5465.68±0.6665.88±1.3965.98±0.83
55 - 6464.37±1.2264.94±1.6664.84±1.30
65 or older63.34±2.0764.70±2.2162.77±2.39
", + "bbox": [ + 147, + 228, + 450, + 329 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/2e360ddab0beb967cacf0c9f50c8ea0eb13bda8f9b2f7cef1bfc636d3a7869f2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
EducationBaselineSoc-Dem.Random
Associate degree60.69±1.4460.54±2.3560.78±1.62
Bachelor's degree66.16±0.5166.23±0.8266.80±0.54
Doctoral degree61.93±3.8263.79±5.0363.27±3.67
High school60.53±1.3960.47±2.2260.55±1.87
Below high school58.28±4.6862.12±4.9060.17±4.25
Master's degree69.71±0.8669.58±0.9369.45±0.96
Professional degree66.75±2.3767.84±3.3268.62±2.84
College, no degree58.65±1.1959.40±1.7959.99±2.19
", + "bbox": [ + 122, + 338, + 475, + 461 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/18f8ad51bbc02072cbb4e2a4a00d78630d9522133575aea6cc0f461e550fb257.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SexualityBaselineSoc-Dem.Random
Bisexual71.83±1.1471.42±1.5169.46±1.95
Heterosexual63.25±0.3963.32±1.2163.82±0.55
Homosexual64.43±1.7566.11±2.2065.12±1.94
", + "bbox": [ + 139, + 470, + 457, + 536 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1: Average and standard deviation of macro $F_{1}$ from three runs of four-fold stratified cross-validation. Separate table for each attribute. Bold results are the highest averages per group. However, no difference is statistically significant (see Appendix A.2)", + "bbox": [ + 112, + 544, + 489, + 617 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "6 Discussion", + "text_level": 1, + "bbox": [ + 112, + 639, + 240, + 653 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We do not find strong evidence that explicitly modelling sociodemographics helps to predict annotation behaviour with multi-annotator models. These results might seem counter-intuitive, given the evidence of systematic annotation differences between sociodemographic groups (see §2). This discrepancy, however, echoes the issue highlighted by ecological fallacies (Robinson, 1950): Not every annotator will be a perfect representative of their group, so we will not necessarily learn additional information based on their group identity. This seems especially true if we already have access to individual behaviour (i.e., individual annotations).", + "bbox": [ + 112, + 661, + 489, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In contrast to Davani et al. (2022), we made sociodemographic information explicit in our experiments, as one of the factors influencing annotation", + "bbox": [ + 112, + 871, + 489, + 917 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "behaviour. Group-specific layers can be seen as an inductive bias putting emphasis on the sociodemographic relations between annotators. However, there are potentially many other factors influencing annotation behaviour (e.g., attitudes, moral values, cognitive biases, psychological traits). In light of our results, it seems plausible that multi-annotator models learn about these factors implicitly as part of predicting individual behaviour, so that making one factor explicit does not change prediction quality, at least in the case of sociodemographics.", + "bbox": [ + 505, + 84, + 884, + 261 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Still, we also know that generally group attributes can help predict individual decisions, i.e., as base rates or priors. To avoid ecological fallacies in modelling annotation, we therefore need to better understand when and how modelling sociodemographic information is useful in predicting an individual annotator's decisions. For example, we have only evaluated group-specific layers for single attributes. In contrast, social scientists have long adopted the idea of intersectionality (Crenshaw, 1989), which also informs research on fairness in machine learning (Wang et al., 2022). Intersectionality means that the effect of interactions between sociodemographic attributes enables specific experiences that are not captured by the attributes in isolation. For example, identifying as a man means something different depending on the person's education. Groups derived from single attributes might simply be too coarse to improve classifiers learnt from individual labels, as in multi-annotator models.", + "bbox": [ + 507, + 265, + 884, + 602 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The dataset we use (Kumar et al., 2021) has many characteristics which are ideal for our study (see §3). However, it uses a broad notion of toxicity, in contrast to other studies of toxic language (Larimore et al., 2021; Sap et al., 2022), which match content and analysed groups. When modeling the groups frequently referenced in the datasets themselves, we would expect greater benefits from group-specific layers. Similar to us, Biester et al. (2022) who do not find significant differences between annotators of different genders, do so in a more general setting.", + "bbox": [ + 507, + 608, + 882, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We can only partially compare to Gordon et al. (2022), despite using the same dataset. In addition to differences in approach (see §2), our and their work also differ in their research questions and thus experimental conditions. Gordon et al. (2022) compare their full model (group and individual) against using group information alone.", + "bbox": [ + 507, + 806, + 884, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1020", + "bbox": [ + 482, + 927, + 521, + 940 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We compare our full model (group and individual) against using individual information alone. So it is unclear if their model would benefit from group information in comparison to individual-level information alone. While they find an improvement from group information it is only in comparison to a baseline predicting not individual but aggregated labels. Additionally, the composition of test sets sampled from the full dataset differs between the studies: Gordon et al. (2022) use a test set of 5,000 comments, while we use 22,360 comments in a four-fold cross-validation. We leave an explicit comparison to future work.", + "bbox": [ + 112, + 84, + 492, + 294 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Group-specific layers (§4) are a natural extension of annotator-specific classification layers in multi-annotator models. However, other architectures to predict annotator-level labels use different ways to represent sociodemographic information, e.g., via embeddings in a recommender system (Gordon et al., 2022). Future work could explore additional representations of annotator attributes (e.g., as part of the input, either textual or as separate features) and other approaches to modelling the relation of individual labeling decisions and attributes (e.g., probabilistic graphical models).", + "bbox": [ + 115, + 294, + 489, + 487 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "7 Conclusion", + "text_level": 1, + "bbox": [ + 112, + 494, + 247, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We ask how relevant modelling explicit sociodemographic information is in learning from individual annotators. Our experiments with group-specific layers for four sociodemographic attributes on social media data with toxicity annotations (Kumar et al., 2021) show no significant benefit of modelling sociodemographic groups in multi-annotator models. However, as the issue of ecological fallacies highlights, it is not implausible that these models do not learn additional information from group information beyond the inherent variation. However, our results do not refute the usefulness of sociodemographic attributes in modelling annotation, but underscore the importance of their judicious use. Different tasks and model architectures will likely benefit to different extents. Ultimately, annotation behaviour is driven by complex factors and we will need to consider more than annotators' sociodemographics.", + "bbox": [ + 112, + 518, + 489, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 112, + 832, + 285, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We thank Deepak Kumar for providing access to the disaggregated dataset and his continued support. We also thank Aida Mostafazadeh Davani for providing information on implementation de", + "bbox": [ + 112, + 854, + 489, + 919 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tails of multi-annotator models. Members of MilaNLP (Bocconi) and the Semantic Computing Group (Bielefeld) provided feedback on earlier versions of this paper, for which we thank them again.", + "bbox": [ + 507, + 84, + 884, + 149 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This work has in part been funded by the European Research Council (ERC) under the European Union's Horizon 2020 research and innovation program (grant agreement No. 949944, INTEGRACTOR). Likewise, this work has in part been funded by the VolkswagenStiftung as part of the \"3B Bots Building Bridges\" project.", + "bbox": [ + 507, + 149, + 885, + 261 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Limitations", + "text_level": 1, + "bbox": [ + 509, + 268, + 613, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While the dataset by Kumar et al. (2021) enabled us to test models for a range of often overlooked groups (e.g., non-binary or bisexual annotators), we ultimately modelled only four specific attributes (gender, age, education, sexual orientation). There are likely to be more factors that could play a role. Additionally, annotators in the Kumar et al. (2021) dataset are exclusively from the United States of America, so that results do not necessarily hold for other countries or cultures (Hovy and Yang, 2021). Specifically perceptions of harmful content online are known to vary across countries (Jiang et al., 2021).", + "bbox": [ + 507, + 290, + 884, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We used only the (Kumar et al., 2021) dataset. This is mainly due to our strict criteria regarding dataset size and availability of annotator-level labels and sociodemographic information. These characteristics were a prerequisite for our experiments across different attributes with sufficient numbers of annotators. Most datasets which include annotator-level labels and sociodemographic information contain much smaller numbers of annotators and attributes. Nevertheless, with the Measuring Hate Speech Corpus there is at least one additional dataset (Sachdeva et al., 2022) with comparable characteristics that could be used in future experiments. Also, additional small-scale, more focused experiments could use datasets like Sap et al. (2022) or HS-Brexit (Akhtar et al., 2021) which was annotated by 6 annotators, each from one of two sociodemographic groups.", + "bbox": [ + 507, + 500, + 884, + 790 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We do not study the aggregation of individual predictions or evaluate against majority labels, as these are not directly relevant to our investigation of sociodemographic attributes in models of annotation behaviour. Consequently, we cannot derive a conclusion about performance in those settings from our results. This is a noteworthy limitation, because part of the experiments introducing", + "bbox": [ + 507, + 790, + 884, + 919 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1021", + "bbox": [ + 482, + 928, + 517, + 940 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "multi-annotator models in Davani et al. (2022) compare labels aggregated from multi-annotator models against predictions from a standard classifier (directly trained on aggregated labels).", + "bbox": [ + 112, + 84, + 487, + 148 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For computational reasons, our experiments use a comparatively small pre-trained language model (RoBERTa, Liu et al. 2019). Thus, results might differ with larger models.", + "bbox": [ + 112, + 149, + 487, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ethics Statement", + "text_level": 1, + "bbox": [ + 114, + 221, + 265, + 236 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As sociodemographic attributes are sensitive information, we do not infer attributes, but build on a self-reported, IRB-reviewed dataset (Kumar et al., 2021). We also see potential for a discussion of \"privacy by design\" in modelling human label variation based on our results: There can be circumstances in which knowing more about annotators is not relevant, and indeed might lead to violations of privacy.", + "bbox": [ + 110, + 246, + 487, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As multi-annotator models attempt to capture the preferences of individual annotators, there are valid concerns around privacy and anonymity. As discussed in Davani et al. (2022), increasing the annotator count can be one option to reduce privacy risks. We show it is feasible to learn a model for a large number of individual annotators (5002 vs. 18 and 82 in their work). But a prerequisite for improved privacy is to apply effective aggregation on top of individual predictions, which we do not study in the present work.", + "bbox": [ + 112, + 392, + 489, + 568 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 114, + 592, + 213, + 607 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Gavin Abercrombie, Valerio Basile, Sara Tonelli, Verena Rieser, and Alexandra Uma, editors. 2022. Proceedings of the 1st Workshop on Perspectivist Approaches to NLP @LREC2022. European Language Resources Association, Marseille, France.", + "Sohail Akhtar, Valerio Basile, and Viviana Patti. 2020. Modeling annotator perspective and polarized opinions to improve hate speech detection. In Proceedings of the AAAI Conference on Human Computation and Crowdsourcing, volume 8, pages 151-154.", + "Sohail Akhtar, Valerio Basile, and Viviana Patti. 2021. Whose opinions matter? perspective-aware models to identify opinions of hate speech victims in abusive language detection. Preprint arXiv:2106.15896.", + "Hala Al Kuwatly, Maximilian Wich, and Georg Groh. 2020. Identifying and measuring annotator bias based on annotators' demographic characteristics. In Proceedings of the Fourth Workshop on Online Abuse and Harms, pages 184-190, Online. Association for Computational Linguistics." + ], + "bbox": [ + 115, + 614, + 489, + 917 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Valerio Basile, Michael Fell, Tommaso Fornaciari, Dirk Hovy, Silviu Paun, Barbara Plank, Massimo Poesio, and Alexandra Uma. 2021. We need to consider disagreement in evaluation. In Proceedings of the 1st Workshop on Benchmarking: Past, Present and Future, pages 15-21, Online. Association for Computational Linguistics.", + "Laura Biester, Vanita Sharma, Ashkan Kazemi, Naihao Deng, Steven Wilson, and Rada Mihalcea. 2022. Analyzing the effects of annotator gender across NLP tasks. In Proceedings of the 1st Workshop on Perspectivist Approaches to NLP @LREC2022, pages 10-19, Marseille, France. European Language Resources Association.", + "Reuben Binns, Michael Veale, Max Van Kleek, and Nigel Shadbolt. 2017. Like trainer, like bot? inheritance of bias in algorithmic content moderation. In Social Informatics, Lecture Notes in Computer Science, pages 405-415. Springer International Publishing.", + "Amanda Cercas Curry, Gavin Abercrombie, and Verena Rieser. 2021. ConvAbuse: Data, analysis, and benchmarks for nuanced abuse detection in conversational AI. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7388-7403, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.", + "Kimberle Crenshaw. 1989. Demarginalizing the intersection of race and sex: A black feminist critique of antidiscrimination doctrine, feminist theory and antiracist politics. University of Chicago Legal Forum, 1989(1):Article 8.", + "Aida Mostafazadeh Davani, Mark Diaz, and Vinodkumar Prabhakaran. 2022. Dealing with disagreements: Looking beyond the majority vote in subjective annotations. Transactions of the Association for Computational Linguistics, 10:92-110.", + "Rotem Dror, Gili Baumer, Marina Bogomolov, and Roi Reichart. 2017. Replicability analysis for natural language processing: Testing significance with multiple datasets. Transactions of the Association for Computational Linguistics, 5:471-486.", + "Rotem Dror, Gili Baumer, Segev Shlomov, and Roi Reichart. 2018. The hitchhiker's guide to testing statistical significance in natural language processing. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1383-1392, Melbourne, Australia. Association for Computational Linguistics.", + "Elizabeth Excell and Noura Al Moubayed. 2021. Towards equal gender representation in the annotations of toxic language detection. In Proceedings of the 3rd Workshop on Gender Bias in Natural Language Processing, pages 55-65, Online. Association for Computational Linguistics." + ], + "bbox": [ + 510, + 84, + 884, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1022", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Tommaso Fornaciari, Alexandra Uma, Silviu Paun, Barbara Plank, Dirk Hovy, and Massimo Poesio. 2021. Beyond black & white: Leveraging annotator disagreement via soft-label multi-task learning. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2591-2597, Online. Association for Computational Linguistics.", + "Tommaso Fornaciari, Alexandra Uma, Massimo Poesio, and Dirk Hovy. 2022. Hard and soft evaluation of NLP models with BOOtSTrap SAmpling - BooStSa. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 127-134, Dublin, Ireland. Association for Computational Linguistics.", + "David A. Freedman. 2015. Ecological inference. In James D. Wright, editor, International Encyclopedia of the Social & Behavioral Sciences (Second Edition), pages 868-870. Elsevier.", + "Mitchell L. Gordon, Michelle S. Lam, Joon Sung Park, Kayur Patel, Jeff Hancock, Tatsunori Hashimoto, and Michael S. Bernstein. 2022. Jury learning: Integrating dissenting voices into machine learning models. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems, CHI '22, pages 1-19. Association for Computing Machinery.", + "Nitesh Goyal, Ian D. Kivlichan, Rachel Rosen, and Lucy Vasserman. 2022. Is your toxicity my toxicity? exploring the impact of rater identity on toxicity annotation. Proceedings of the ACM on Human-Computer Interaction, 6:1-28.", + "Dirk Hovy and Diyi Yang. 2021. The importance of modeling social factors of language: Theory and practice. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 588-602, Online. Association for Computational Linguistics.", + "Emily Jamison and Iryna Gurevych. 2015. Noise or additional information? leveraging crowdsourced annotation item agreement for natural language tasks. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 291-297, Lisbon, Portugal. Association for Computational Linguistics.", + "Jialun Aaron Jiang, Morgan Klaus Scheuerman, Casey Fiesler, and Jed R. Brubaker. 2021. Understanding international perceptions of the severity of harmful content online. PLOS ONE, 16(8).", + "Deepak Kumar, Patrick Gage Kelley, Sunny Consolvo, Joshua Mason, Elie Bursztein, Zakir Durmeric, Kurt Thomas, and Michael Bailey. 2021. Designing toxic content classification for a diversity of perspectives. In Seventeenth Symposium on Usable Privacy and Security (SOUPS 2021), pages 299-318. USENIX Association." + ], + "bbox": [ + 115, + 85, + 489, + 917 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Savannah Larimore, Ian Kennedy, Breon Haskett, and Alina Arseniev-Koehler. 2021. Reconsidering annotator disagreement about racist language: Noise or signal? In Proceedings of the Ninth International Workshop on Natural Language Processing for Social Media, pages 81-90, Online. Association for Computational Linguistics.", + "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. Preprint arXiv:1907.11692.", + "F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. 2011. Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12:2825-2830.", + "Barbara Plank. 2022. The \"problem\" of human label variation: On ground truth in data, modeling and evaluation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 10671-10682, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.", + "Barbara Plank, Dirk Hovy, and Anders Søgaard. 2014. Learning part-of-speech taggers with inter-annotator agreement loss. In Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 742–751, Gothenburg, Sweden. Association for Computational Linguistics.", + "Vinodkumar Prabhakaran, Aida Mostafazadeh Davani, and Mark Diaz. 2021. On releasing annotator-level labels and information in datasets. In Proceedings of the Joint 15th Linguistic Annotation Workshop (LAW) and 3rd Designing Meaning Representations (DMR) Workshop, pages 133-138, Punta Cana, Dominican Republic. Association for Computational Linguistics.", + "W. S. Robinson. 1950. Ecological correlations and the behavior of individuals. American Sociological Review, 15(3):351-357.", + "Paul Röttger, Bertie Vidgen, Dirk Hovy, and Janet Pierrehumbert. 2022. Two contrasting data annotation paradigms for subjective NLP tasks. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 175-190, Seattle, United States. Association for Computational Linguistics.", + "Pratik Sachdeva, Renata Barreto, Geoff Bacon, Alexander Sahn, Claudia von Vacano, and Chris Kennedy. 2022. The measuring hate speech corpus: Leveraging rasch measurement theory for data perspectivism. In Proceedings of the 1st Workshop on Perspectivist Approaches to NLP @LREC2022, pages 83-94, Marseille, France. European Language Resources Association." + ], + "bbox": [ + 510, + 85, + 882, + 917 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1023", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 6 + }, + { + "type": "ref_text", + "text": "Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. 2019. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. In 5th Workshop on Energy Efficient Machine Learning and Cognitive Computing @ NeurIPS 2019.", + "bbox": [ + 115, + 85, + 487, + 151 + ], + "page_idx": 7 + }, + { + "type": "ref_text", + "text": "Maarten Sap, Dallas Card, Saadia Gabriel, Yejin Choi, and Noah A. Smith. 2019. The risk of racial bias in hate speech detection. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1668-1678, Florence, Italy. Association for Computational Linguistics.", + "bbox": [ + 115, + 161, + 487, + 240 + ], + "page_idx": 7 + }, + { + "type": "ref_text", + "text": "Maarten Sap, Swabha Swayamdipta, Laura Vianna, Xuhui Zhou, Yejin Choi, and Noah A. Smith. 2022. Annotators with attitudes: How annotator beliefs and identities bias toxic language detection. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5884-5906, Seattle, United States. Association for Computational Linguistics.", + "bbox": [ + 115, + 249, + 487, + 367 + ], + "page_idx": 7 + }, + { + "type": "ref_text", + "text": "Qinlan Shen and Carolyn Rose. 2021. What sounds \"right\" to me? experiential factors in the perception of political ideology. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 1762-1771, Online. Association for Computational Linguistics.", + "bbox": [ + 115, + 376, + 487, + 468 + ], + "page_idx": 7 + }, + { + "type": "ref_text", + "text": "Alexandra N. Uma, Tommaso Fornaciari, Dirk Hovy, Silviu Paun, Barbara Plank, and Massimo Poesio. 2021. Learning from disagreement: A survey. Journal of Artificial Intelligence Research, 72:1385-1470.", + "bbox": [ + 115, + 479, + 487, + 543 + ], + "page_idx": 7 + }, + { + "type": "ref_text", + "text": "Angelina Wang, Vikram V Ramaswamy, and Olga Russakovsky. 2022. Towards intersectionality in machine learning: Including more identities, handling underrepresentation, and performing evaluation. In 2022 ACM Conference on Fairness, Accountability, and Transparency, FAccT '22, pages 336-349. Association for Computing Machinery.", + "bbox": [ + 115, + 554, + 487, + 646 + ], + "page_idx": 7 + }, + { + "type": "ref_text", + "text": "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online. Association for Computational Linguistics.", + "bbox": [ + 115, + 655, + 487, + 812 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A Appendix", + "text_level": 1, + "bbox": [ + 115, + 825, + 236, + 841 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A.1 Annotator Sociodemographics in Sample", + "text_level": 1, + "bbox": [ + 114, + 848, + 485, + 864 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2 shows how many annotators the sample contains. Counts are given per group of the four attributes gender, age, education and sexuality.", + "bbox": [ + 112, + 871, + 485, + 917 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the Kumar et al. (2021) dataset, sociodemographic attributes are given for each individual annotation - not once per annotator. For some annotators, conflicting attribute values exist (e.g., two different age groups). As the data collection spanned several months (Kumar et al., 2021), these value changes can in principle be reasonable (e.g., because an annotator got older, finished a degree, changed sexual preference or gender identity). However, as reasonable changes can not easily be discerned from erroneous input, we disambiguate values based on a heuristic: If an annotator reports several values for an attribute, we assume the most frequent value to be valid. In cases of no clear most frequent value, we set the attribute to \"Prefer not to say\". Thus, the main results do not contain annotators with ambiguous attributes.", + "bbox": [ + 507, + 84, + 882, + 357 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A.2 Significance Tests", + "text_level": 1, + "bbox": [ + 509, + 366, + 697, + 380 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results of a replicability analysis (Dror et al., 2017) testing for significant differences in macro $F_{1}$ on scores from three runs of four-fold cross-validation. Table 3 shows results for a comparison of the sociodemographic models against the baseline models. Table 4 shows results for a comparison of the sociodemographic models against the randomized assignment models. The Bonferroni correction for the corrected count of significant folds $\\hat{k}_{Bonferroni}$ is used to account for the fact that we have overlapping test sets from multiple runs of four-fold cross-validation.", + "bbox": [ + 507, + 388, + 882, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A.3 Training Details, Hyperparameters and Computational Resources", + "text_level": 1, + "bbox": [ + 509, + 589, + 868, + 621 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We implement models and the training loop using the Hugging Face Transformers library (version 4.19.2, Wolf et al. 2020). Maximum sequence length is 512 tokens, with truncation and padding to the maximum length. We train for 3 epochs with a batch size of 8 and an initial learning rate of 0.00001. Otherwise, we used default parameters. We found results to particularly depend on the learning rate, with higher or lower values leading to worse results.", + "bbox": [ + 507, + 627, + 882, + 788 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We use a weighted loss function. Label weights are calculated per annotator on the training set of each fold. Label weights, evaluation scores and the four-fold dataset splits (StratifiedKFold) are calculated using the scikit-learn library (version 1.0.2, Pedregosa et al. 2011). The folds are based on a fixed random seed per iteration: 2803636207, 165043843, 2923262358", + "bbox": [ + 507, + 790, + 882, + 917 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1024", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/caf24a1aec721b7e03cec00eef451e8e68c6963e40fc0accd6dd331759ac6b4f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Number of Annotators
Gender
Female2450
Male2116
Prefer not to say412
Nonbinary23
Other1
", + "bbox": [ + 171, + 80, + 428, + 181 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/725b26b53e9f011b4cb410c17f8c9a5999e171badfdb916080aa142b5ac9cb1e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Number of Annotators
Age
18 - 24489
25 - 341861
35 - 441115
45 - 54529
55 - 64321
65 or older119
Prefer not to say568
", + "bbox": [ + 171, + 187, + 426, + 313 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/4d2f443630932d6e0215455b61f115dbfe0bff95fba3b7a6b7d56694573f5c94.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Number of Annotators
Sexuality
Heterosexual4018
Bisexual469
Prefer not to say346
Homosexual134
Other35
", + "bbox": [ + 171, + 322, + 428, + 423 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/6519765e784f01feb0edab7fe68fff85ff643b928196d32c7956cb702ebb37d8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
EducationNumber of Annotators
Bachelor's degree1879
College, no degree861
Prefer not to say647
Master's degree642
Associate degree460
High school363
Professional degree68
Doctoral degree51
Below high school25
Other6
", + "bbox": [ + 164, + 432, + 435, + 592 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The majority of parameters in our model belong to the pre-trained language model shared between all group-specific and annotator-specific layers. Specifically, RoBERTa (Liu et al., 2019) in the roberta-base variant has 125 Million parameters. We keep the pre-trained model's default output dimensionality of 768, so that each group-specific layer adds $768 * 768 + 768 = 590$ , 592 parameters and each annotator layer adds $768 * 2 + 2 = 1$ , 538 parameters.", + "bbox": [ + 112, + 674, + 487, + 835 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "All experiments ran on a single GPU (GeForce GTX 1080 Ti, 12GB GPU RAM). Per fold, training and evaluation together take about three and a half hours in our setting. Three runs of four-fold cross-validation (12 folds), thus take around 42 hours", + "bbox": [ + 112, + 838, + 489, + 917 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/32ad7e940605d95071ac8ced8895f2a1491b15b7f03ce436050c304998f2f1b0.jpg", + "table_caption": [ + "Table 2: Number of annotators per group for attributes gender, age, sexuality and education. Counts refer to the entire sample" + ], + "table_footnote": [], + "table_body": "
hatkcounthatkBonf.
Female20
Male00
Nonbinary10
", + "bbox": [ + 606, + 80, + 786, + 147 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/e3c0792bfae62e9417a852c39d1d4cc6e0619d8c96dbaa40b5a98dd6d2835c4e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
hatkcounthatkBonf.
18 - 2420
25 - 3420
35 - 4410
45 - 5400
55 - 6410
65 or older10
", + "bbox": [ + 606, + 156, + 786, + 259 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/8f80c5e137d1b0101a4ea36672c28c390623e4b188053e8071a5d8918a8b2473.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ˆkcountˆkBonf.
Bisexual20
Heterosexual42
Homosexual10
", + "bbox": [ + 606, + 268, + 786, + 332 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/a8a67b4a978d772342cef4159df4b52f82005757c861476a929c0158454e9f49.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
kcountkBonf.
Associate degree00
Bachelor's degree10
Doctoral degree20
High school00
Belowhigh school00
Master's degree00
Professional degree00
College, no degree22
", + "bbox": [ + 581, + 344, + 806, + 468 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 3: Results of a replicability analysis of baseline vs sociodemographic models. Raw and Bonferroni-corrected counts of significant folds out of 12 folds from three runs of four-fold cross-validation. P-values for each fold are computed via a paired bootstrap test with significance level $\\alpha = 0.05$ , 1000 bootstrap samples per fold and a sample size of $50\\%$ of the respective test set.", + "bbox": [ + 507, + 478, + 882, + 592 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "(1.75 days). With four attributes and three trainable models the combined run time of the reported experiments is estimated to be 21 days. Including preliminary experiments, which, however, mostly were not full runs of k-fold cross-validation and also utilized DistilBERT (Sanh et al., 2019) with slightly faster run times, it will be many times more. There is no discernible difference in experiment run times between multi-annotator models with or without groups or different numbers of groups.", + "bbox": [ + 507, + 623, + 882, + 785 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "A.4 Number of Annotations per Group across all Test Sets", + "text_level": 1, + "bbox": [ + 507, + 797, + 880, + 826 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 5 contains the number of annotations we have per group across the total of 12 folds (from three runs of four-fold cross-validation). This number of annotations is the effective test set size per group. As the numbers do not vary substantially, perfor", + "bbox": [ + 507, + 838, + 882, + 917 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1025", + "bbox": [ + 482, + 927, + 519, + 940 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/fe0ce9572f53a63635954facf9e6ea438bda1ce5b49524888e22863ae688e876.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
kcountkBonf.
Female22
Male10
Nonbinary10
", + "bbox": [ + 211, + 82, + 386, + 146 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/dc82f1d9db26486c59056b62e1ded50a8075f393987e5dbed661e13e1f823f5f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
kcountkBonf.
18 - 2410
25 - 3400
35 - 4410
45 - 5410
55 - 6430
65 or older10
", + "bbox": [ + 211, + 158, + 384, + 258 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/ab65613339085586161cb49be1db61fb9bcdbd71f381a4fbda16c5de3ee162bb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
hatcounthatBonf.
Bisexual62
Heterosexual11
Homosexual00
", + "bbox": [ + 206, + 269, + 391, + 331 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/702e850c9b2a192fecd3a5c53aec251af17118f60f15c9db6834da30e3abe6c5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
kcountkBonf.
Associate degree20
Bachelor's degree10
Doctoral degree00
High school20
Belowhigh school20
Master's degree00
Professional degree00
College, no degree11
", + "bbox": [ + 186, + 344, + 411, + 468 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "mance on each fold is equally representative for all groups.", + "bbox": [ + 112, + 618, + 487, + 651 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "A.5 Full Results", + "text_level": 1, + "bbox": [ + 114, + 656, + 258, + 671 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 6 shows full results of experiments (see 4), including results for all residual categories and a naive baseline which always predicts toxic.", + "bbox": [ + 112, + 678, + 489, + 726 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/7a27e45775e194b4ba54cfcacbc2a55e9f36cf3371e54d597e6cbd612b75c2a9.jpg", + "table_caption": [ + "Table 4: Results from replicability analysis of randomized vs sociodemographic models. Raw and Bonferroni-corrected counts of significant folds out of 12 folds from three runs of four-fold cross-validation. P-values for each fold are computed via a paired bootstrap test with significance level $\\alpha = 0.05$ , 1000 bootstrap samples per fold and a sample size of $50\\%$ of the respective test set." + ], + "table_footnote": [], + "table_body": "
GenderNumber Of AnnotationsMinMax
Female13555±86.4413383.013664.0
Male11925±61.6511843.012062.0
Nonbinary115±6.03104.0122.0
Other5±1.952.08.0
Prefer not to say2345±51.192281.02453.0
", + "bbox": [ + 510, + 200, + 885, + 300 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/a722e27081a3aea739b5da7b93cd401028b9d7cff73204fe2a029abc8d5d9b92.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AgeNumber Of AnnotationsMinMax
18 - 242615±50.8825212697
25 - 3410315±61.451024410457
35 - 446250±51.0661796324
45 - 543025±47.2329293083
55 - 641865±25.4818311903
65 or older675±19.31643704
Prefer not to say3200±55.2831313289
", + "bbox": [ + 515, + 310, + 870, + 434 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/9e3544005ebc4e866aea52dd31404f4b2c48337a98939bfc114a3b6f90b0417d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SexualityNumber Of AnnotationsMinMax
Bisexual2445±39.2623832501
Heterosexual22630±63.002250722726
Homosexual725±26.57670759
Other190±7.91173201
Prefer not to say1955±35.3918782009
", + "bbox": [ + 517, + 443, + 870, + 543 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/b1601f8aef7178a374f3956a192719185a06e4c50c72992aada3967166b2da4a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
EducationNumber Of AnnotationsMinMax
Associate professor2605±47.5925162697
Bachelor's degree10510±84.791034810700
Doctoral degree305±18.83270332
High school2080±37.0120152139
Below high school165±11.17144184
Master's degree3515±48.0834253580
Other30±3.442536
Prefer not to say3690±52.9236033808
Professional degree380±17.87352411
College, no degree4665±71.3645394776
", + "bbox": [ + 510, + 552, + 880, + 712 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 5: Average, standard deviation, minimum and maximum of number of annotations per fold. All information given per group of gender, age, education and sexuality. Statistics are calculated across 12 folds from three runs of four-fold cross-validation.", + "bbox": [ + 507, + 722, + 882, + 794 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1026", + "bbox": [ + 482, + 928, + 521, + 940 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/21a17ff1f2a07c1b08ed9915e4a34c9b984dad2935a8eb8a879ca50f559747f3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
GenderMajority BaselineBaselineSoc-Dem.Random
Female41.79±0.1262.23±0.5362.25±1.1962.41±0.92
Male40.53±0.1168.00±0.4967.66±0.4667.63±0.53
Nonbinary44.69±1.3956.33±6.0056.80±7.2458.00±7.49
Other45.50±4.6948.56±10.7850.53±14.6343.66±7.25
Prefer not to say41.05±0.3664.54±1.1365.05±1.5265.08±1.86
", + "bbox": [ + 263, + 247, + 729, + 335 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/e0a9733f35c388dcd059b70d3d757e06b472b3a615fa43b0d42bf81291aafafc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AgeMajority BaselineBaselineSoc-Dem.Random
18 - 2442.49±0.2859.39±1.5860.44±1.0560.52±1.37
25 - 3440.49±0.0966.72±0.5666.63±0.8366.92±0.51
35 - 4441.87±0.1564.50±0.5964.94±1.3365.24±0.89
45 - 5440.63±0.2665.68±0.6665.88±1.3965.98±0.83
55 - 6441.65±0.3964.37±1.2264.94±1.6664.84±1.30
65 or older41.46±0.5463.34±2.0764.70±2.2162.77±2.39
Prefer not to say41.37±0.3263.99±1.3265.24±1.1864.73±1.33
", + "bbox": [ + 263, + 344, + 722, + 457 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/b02b56a006df2ea473f73cf387d47a11d02b3d9517886c04b4f75943232454e1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
EducationMajority BaselineBaselineSoc-Dem.Random
Associate degree43.16±0.1960.69±1.4460.54±2.3560.78±1.62
Bachelor's degree40.38±0.1066.16±0.5166.23±0.8266.80±0.54
Doctoral degree43.34±0.9461.93±3.8263.79±5.0363.27±3.67
High school43.02±0.2660.53±1.3960.47±2.2260.55±1.87
Below high school43.10±1.4458.28±4.6862.12±4.9060.17±4.25
Master's degree37.55±0.3269.71±0.8669.58±0.9369.45±0.96
Other42.95±2.3156.56±10.8857.59±9.8657.71±12.28
Prefer not to say40.97±0.2765.07±1.1665.69±1.0565.74±1.09
Professional degree40.43±0.8066.75±2.3767.84±3.3268.62±2.84
College, no degree43.61±0.1858.65±1.1959.40±1.7959.99±2.19
", + "bbox": [ + 253, + 466, + 739, + 612 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/78c4e3df5636f674d9c687d72c3665a6ec9be5fa90bb8faeda716302a6b3b16c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SexualityMajority BaselineBaselineSoc-Dem.Random
Bisexual34.69±0.5071.83±1.1471.42±1.5169.46±1.95
Heterosexual41.99±0.0663.25±0.3963.32±1.2163.82±0.55
Homosexual41.15±0.4164.43±1.7566.11±2.2065.12±1.94
Other43.53±0.7857.55±3.7960.57±4.5158.69±4.72
Prefer not to say39.12±0.2467.80±1.5667.27±1.5267.46±1.11
", + "bbox": [ + 270, + 621, + 722, + 709 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 6: Average and standard deviation of macro $F_{1}$ from three runs of four-fold stratified cross-validation. Separate table for each attribute. Bold results are the highest average per group. Full results including naive majority baseline", + "bbox": [ + 112, + 719, + 882, + 750 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1027", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "A For every submission:", + "bbox": [ + 115, + 107, + 322, + 122 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "A1. Did you describe the limitations of your work? Limitations, 8", + "A2. Did you discuss any potential risks of your work? Ethics Statement, 9", + "A3. Do the abstract and introduction summarize the paper's main claims?", + "A4. Have you used AI writing assistants when working on this paper? Left blank." + ], + "bbox": [ + 129, + 126, + 695, + 287 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "B Did you use or create scientific artifacts?", + "bbox": [ + 114, + 299, + 487, + 316 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "3, Appendix A.3", + "bbox": [ + 132, + 321, + 253, + 336 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "B1. Did you cite the creators of artifacts you used? 3, Appendix A.3", + "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Clear from context, citations", + "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Clear from context, citations", + "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? 3, Ethics Statement 9", + "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? 3, Appendix A.1", + "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. 3, 4, Appendix A.4" + ], + "bbox": [ + 129, + 346, + 880, + 753 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "C Did you run computational experiments?", + "bbox": [ + 114, + 764, + 492, + 781 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4", + "bbox": [ + 132, + 788, + 146, + 799 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Appendix A.3", + "bbox": [ + 129, + 810, + 880, + 860 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "ACL 2023 Responsible NLP Checklist", + "bbox": [ + 132, + 84, + 433, + 99 + ], + "page_idx": 11 + }, + { + "type": "page_footnote", + "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.", + "bbox": [ + 112, + 868, + 877, + 892 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "1028", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? No response.", + "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? No response.", + "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Appendix A.3" + ], + "bbox": [ + 127, + 84, + 880, + 282 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank.", + "text_level": 1, + "bbox": [ + 112, + 293, + 877, + 330 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response.", + "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response.", + "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response.", + "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response.", + "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response." + ], + "bbox": [ + 127, + 341, + 880, + 640 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "1029", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2023/The Ecological Fallacy in Annotation_ Modeling Human Label Variation goes beyond Sociodemographics/8b624391-c7ce-443d-9caf-875f1d838500_model.json b/2023/The Ecological Fallacy in Annotation_ Modeling Human Label Variation goes beyond Sociodemographics/8b624391-c7ce-443d-9caf-875f1d838500_model.json new file mode 100644 index 0000000000000000000000000000000000000000..edcae478ab8949b03b6355197b8857a5ecf56b60 --- /dev/null +++ b/2023/The Ecological Fallacy in Annotation_ Modeling Human Label Variation goes beyond Sociodemographics/8b624391-c7ce-443d-9caf-875f1d838500_model.json @@ -0,0 +1,2283 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.15, + 0.091, + 0.849, + 0.131 + ], + "angle": 0, + "content": "The Ecological Fallacy in Annotation: Modelling Human Label Variation goes beyond Sociodemographics" + }, + { + "type": "text", + "bbox": [ + 0.191, + 0.145, + 0.812, + 0.163 + ], + "angle": 0, + "content": "Matthias Orlikowski1, Paul Röttger2, Philipp Cimiano1, and Dirk Hovy3" + }, + { + "type": "text", + "bbox": [ + 0.417, + 0.175, + 0.588, + 0.192 + ], + "angle": 0, + "content": "\\(^{1}\\)Bielefeld University" + }, + { + "type": "text", + "bbox": [ + 0.413, + 0.192, + 0.591, + 0.208 + ], + "angle": 0, + "content": "\\(^{2}\\)University of Oxford" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.208, + 0.774, + 0.225 + ], + "angle": 0, + "content": "3Computing Sciences Department, Bocconi University, Milan, Italy" + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.253, + 0.341, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.281, + 0.461, + 0.594 + ], + "angle": 0, + "content": "Many NLP tasks exhibit human label variation, where different annotators give different labels to the same texts. This variation is known to depend, at least in part, on the sociodemographics of annotators. Recent research aims to model individual annotator behaviour rather than predicting aggregated labels, and we would expect that sociodemographic information is useful for these models. On the other hand, the ecological fallacy states that aggregate group behaviour, such as the behaviour of the average female annotator, does not necessarily explain individual behaviour. To account for sociodemographics in models of individual annotator behaviour, we introduce group-specific layers to multi-annotator models. In a series of experiments for toxic content detection, we find that explicitly accounting for sociodemographic attributes in this way does not significantly improve model performance. This result shows that individual annotation behaviour depends on much more than just sociodemographics." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.607, + 0.26, + 0.622 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.63, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Different annotators will not necessarily assign the same labels to the same texts, resulting in human label variation (Plank, 2022). Previous work finds that this variation depends at least in part on the sociodemographics of annotators, such as their age and gender (Binns et al., 2017; Al Kuwatly et al., 2020; Excell and Al Moubayed, 2021; Shen and Rose, 2021). These results are particularly pronounced for subjective tasks like toxic content detection (Sap et al., 2019; Kumar et al., 2021; Sap et al., 2022; Goyal et al., 2022). Since human label variation is relevant to a wide range of NLP tasks, recent research has begun to model individual annotator behaviour, rather than predicting aggregated labels (Davani et al., 2022; Gordon et al., 2022). In this setting, we would expect sociodemographic attributes to help explain annotator decisions. Therefore, we investigate whether explicitly" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.253, + 0.878, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.377, + 0.883, + 0.405 + ], + "angle": 0, + "content": "Figure 1: Group-specific layers representing annotator sociodemographics in multi-annotator models." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.433, + 0.884, + 0.479 + ], + "angle": 0, + "content": "accounting for the sociodemographic attributes of annotators leads to better predictions of their annotation behaviour1." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.482, + 0.885, + 0.706 + ], + "angle": 0, + "content": "There is a risk of misreading these efforts as an example of the ecological fallacy: aggregate group behaviour does not necessarily explain individual behaviour (Robinson, 1950; Freedman, 2015). For example, while on average, white annotators may be more likely to label African-American Vernacular English as toxic (Sap et al., 2019), that does not mean it is true for every white annotator individually. However, we aim at exactly this distinction to discuss the relevance of sociodemographic groups in models of individual annotator behaviour. Likewise, we do not assume prior work to commit ecological fallacies, even if a less-nuanced read might suggest it." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.708, + 0.884, + 0.869 + ], + "angle": 0, + "content": "Davani et al. (2022) introduce a simple multi-annotator model, where each annotator is modelled with a separate classification head. We expand their model with group-specific layers, which are activated for each annotator based on their sociodemographic attributes. We compare the two model setups to a control setup where we randomise group assignments. All comparisons use annotator-level toxicity data from Kumar et al. (2021). We find that find that explicitly accounting for sociodem" + }, + { + "type": "page_footnote", + "bbox": [ + 0.508, + 0.881, + 0.883, + 0.919 + ], + "angle": 0, + "content": "\\(^{1}\\)Code to run our experiments and analyses is available at https://github.com/morlikowski/ecological-fallacy" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1017" + }, + { + "type": "footer", + "bbox": [ + 0.227, + 0.946, + 0.77, + 0.958 + ], + "angle": 0, + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + }, + { + "type": "footer", + "bbox": [ + 0.369, + 0.959, + 0.63, + 0.972 + ], + "angle": 0, + "content": "Volume 2: Short Papers, pages 1017-1029" + }, + { + "type": "footer", + "bbox": [ + 0.297, + 0.973, + 0.702, + 0.986 + ], + "angle": 0, + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.489, + 0.166 + ], + "angle": 0, + "content": "graphic attributes does not significantly improve model performance. This result suggests that human label variation happens at a more individual level than sociodemographics, and that annotator decisions are even more complex." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.172, + 0.489, + 0.302 + ], + "angle": 0, + "content": "Contributions 1) We introduce group-specific layers to model groups of annotators with shared attributes in multi-annotator models. 2) We evaluate the effect of group-specific layers for toxic content detection, and show that explicitly accounting for sociodemographic attributes does not significantly improve performance, thus highlighting the risk of the ecological fallacy in annotator modelling." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.302, + 0.49, + 0.351 + ], + "angle": 0, + "content": "As a corollary, we show that multi-annotator models can be applied to many times more annotators than in prior work." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.358, + 0.271, + 0.373 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.381, + 0.489, + 0.671 + ], + "angle": 0, + "content": "Sociodemographics in Annotation Behaviour A growing body of research studies how annotator sociodemographics relate to their annotation decisions, for tasks ranging from natural language inference (Biester et al., 2022) to the detection of racist (Larimore et al., 2021) or generally toxic (Sap et al., 2022) language. Goyal et al. (2022), for example, find that annotators from certain sociodemographic groups (e.g., LGBTQ people) tend to find content attacking their own groups (e.g., homophobic content) to be more toxic. This motivates our research into explicitly accounting for sociodemographics to model annotation behaviour. However, the link between sociodemographics and behaviour is not uncontested. Biester et al. (2022), for example, do not find significant differences in annotation behaviour between annotators of different genders for four different tasks." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.678, + 0.49, + 0.87 + ], + "angle": 0, + "content": "Predicting Annotators' Decisions on Text Different from analyses of annotation behaviour, a recent line of research attempts to learn models based on individual annotations (Plank et al., 2014; Jamison and Gurevych, 2015; Akhtar et al., 2020; Fornaciari et al., 2021; Cercas Curry et al., 2021). These models are motivated by the concern that aggregating labels into a single \"truth\" is too simplistic for many tasks (Uma et al., 2021; Basile et al., 2021) and might introduce uneven representation of perspectives (Prabhakaran et al., 2021; Abercrombie et al., 2022)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.872, + 0.489, + 0.92 + ], + "angle": 0, + "content": "A particular way of learning from disaggregated labels are models that predict individual annotator decisions for an example. Our work builds directly" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.358 + ], + "angle": 0, + "content": "on such a model, multi-annotator models (Davani et al., 2022), which we describe in more detail separately (\\(\\S 4\\)). Gordon et al. (2022) present a model which also predicts individual annotations and allows a user to interactively aggregate them based on \"a jury\" inspired by the US judicial system. Their work is similar to ours in central aspects as they explicitly model annotators' sociodemographics and use the same dataset as we do (Kumar et al., 2021). Different from our work, they frame the task as a regression problem and develop a model based on recommender systems. While they also explore ecological fallacies, they focus on usage risks of their system and countermeasures. In contrast, we consider the issue of the ecological fallacy in modelling annotation behaviour more generally. We compare our findings to their results (\\(\\S 6\\))." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.365, + 0.589, + 0.379 + ], + "angle": 0, + "content": "3 Data" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.388, + 0.885, + 0.596 + ], + "angle": 0, + "content": "We use a sample of the Kumar et al. (2021) dataset for our experiments. The full dataset contains 107,620 English comments from Twitter, Reddit, and 4Chan, annotated for toxicity by 17,280 annotators. The annotation process encouraged annotator subjectivity (Röttger et al., 2022) which is a desired feature for modelling annotator behaviour. For each annotator, there is extensive sociodemographic information, collected with a survey. Annotations are given as ratings on a five-point scale which we convert to binary annotations by mapping ratings of 2 to 4 to toxic, and ratings 0 and 1 to non-toxic." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.598, + 0.884, + 0.693 + ], + "angle": 0, + "content": "We randomly sample comments from the dataset until we reach annotations from more than 5,000 annotators. We then add all other annotations by these annotators. This approach maximizes the number of examples while controlling the number of annotators in our sample." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.694, + 0.885, + 0.822 + ], + "angle": 0, + "content": "Our final sample contains 111,780 annotations from 5,002 annotators on 22,360 comments with 20 to 120 annotations per annotator (mean 22.35). Most comments have five annotations. 20 comments have four because we removed any underage annotators before sampling. In total 78,357 annotations \\((70.10\\%)\\) are toxic, and 33,423 annotations \\((29.90\\%)\\) are non-toxic." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.823, + 0.884, + 0.919 + ], + "angle": 0, + "content": "We focus on four sociodemographic attributes: gender, age, education, and sexual orientation. Group sizes vary by attribute. For gender, 2,450 annotators \\((48.98\\%)\\) identify as female, 2,116 \\((42.30\\%)\\) as male, 23 \\((0.46\\%)\\) as non-binary (rest in residual categories, full statistics in A.1)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1018" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.114, + 0.085, + 0.26, + 0.101 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.108, + 0.49, + 0.268 + ], + "angle": 0, + "content": "We compare three models. The baseline model is the multi-annotator model by Davani et al. (2022). We use their multi-task variant: For each annotator, there is a separate classification layer trained on annotations from that annotator. All annotator layers share a pre-trained language model used to encode the input. We use RoBERTa (Liu et al., 2019) for this, motivated by computational constraints. The other models in our experiments build on this baseline model." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.27, + 0.49, + 0.446 + ], + "angle": 0, + "content": "For the sociodemographic models, we add group-specific layers based on sociodemographic attributes of the annotators. A single attribute, e.g., age, implies several groups, e.g., ages 25-34, ages 35-44. We add the group-specific layers between the pre-trained model and the annotator layers. Each group of annotators shares a separate group-specific layer. We implement group-specific layers as fully-connected, linear layers, each learning a feature transformation applied for one group of annotators." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.448, + 0.49, + 0.656 + ], + "angle": 0, + "content": "Finally, for the random models, we shuffle the assignment of annotators to groups from the sociodemographic model, retaining the relative group sizes. In other words, the probability of each annotator staying in the same group or being reassigned to another group corresponds to the relative size of each group. This approach keeps the model architecture constant while removing the connection between actual sociodemographic attributes and group assignment. It allows us to distinguish the effects of additional parameters, which group-specific layers add in comparison to the baseline, from the effects of sociodemographic information." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.665, + 0.295, + 0.68 + ], + "angle": 0, + "content": "4.1 Evaluation Setup" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.687, + 0.49, + 0.801 + ], + "angle": 0, + "content": "We evaluate all models on individual annotations from gender, age, education, and sexual orientation groups. This setup is comparable to the \"individual label\" evaluations in Davani et al. (2022) and Gordon et al. (2022), but with scores calculated per group of annotators. We measure performance in macro-average \\( F_{1} \\), to weigh each class equally." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.807, + 0.49, + 0.92 + ], + "angle": 0, + "content": "Cross-Validation As there is no standard split available for our dataset, we perform three iterations of a four-fold cross-validation with different seeds (training details in Appendix A.3). We choose four folds, so that even very small groups have more than a hundred annotations in each test set. Across folds, the numbers of annotations per" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.198 + ], + "angle": 0, + "content": "sociodemographic group are similar (see Appendix A.4). We construct test sets that only contain comments unseen by the annotators in the training set. We also ensure that all test sets have similar proportions of toxic or non-toxic comments (assigned by the majority of annotators) to address the class imbalance in the dataset (70.62% toxic, see §3)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.206, + 0.885, + 0.382 + ], + "angle": 0, + "content": "Statistical Significance We test for statistical significance of our results from multiple runs of k-fold cross-validation via replicability analysis (Dror et al., 2017). We report the number of significant folds and the Bonferroni-corrected count (Dror et al., 2018) in Appendix A.2. We compute the p-values for each fold via a paired bootstrap-sampling test with BooStSa (Fornaciari et al., 2022). We set the significance level \\(\\alpha = 0.05\\), draw 1000 bootstrap samples per fold, and use a sample size of \\(50\\%\\) of the respective test set." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.39, + 0.884, + 0.47 + ], + "angle": 0, + "content": "Remarks on Groups Annotators from different groups of the same attribute will in most cases not have annotated the same examples. Therefore, comparisons between models are only meaningful within each group." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.471, + 0.884, + 0.584 + ], + "angle": 0, + "content": "The groups modeled via group-specific layers and those in the result tables are always the same. For example, if we report scores for gender groups, then the sociodemographic and randomized models are also based on gender groups. In the following, we focus on a subset of groups, omitting, e.g., \"Prefer not to say\" (see Appendix A.5)." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.591, + 0.608, + 0.607 + ], + "angle": 0, + "content": "5 Results" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.615, + 0.884, + 0.696 + ], + "angle": 0, + "content": "Table 1 shows the results for gender, age, education, and sexual orientation. A naive majority class baseline that predicts all input to be toxic performs worse than all other models with a large margin (exact results in Appendix A.5)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.703, + 0.885, + 0.849 + ], + "angle": 0, + "content": "Sociodemographics vs. Baseline Across attributes, the average scores of the sociodemographic model and the baseline are similar. The sociodemographic model often has a slightly higher average macro F1 than the baseline, but no statistically significant gains. Where average performance is better by several points, as for homosexual annotators, this gain is offset by a large variance in performance (a consequence of small group sizes)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.856, + 0.884, + 0.92 + ], + "angle": 0, + "content": "Sociodemographics vs. Random We also do not find significant performance differences between sociodemographic group-layer models and the corresponding random group assignment models. For" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1019" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.49, + 0.149 + ], + "angle": 0, + "content": "most groups, the randomized models achieve the highest average scores, but differences to the sociodemographic model are never statistically significant." + }, + { + "type": "table", + "bbox": [ + 0.148, + 0.156, + 0.451, + 0.22 + ], + "angle": 0, + "content": "
GenderBaselineSoc-Dem.Random
Male68.00±0.4967.66±0.4667.63±0.53
Female62.23±0.5362.25±1.1962.41±0.92
Nonbinary56.33±6.0056.80±7.2458.00±7.49
" + }, + { + "type": "table", + "bbox": [ + 0.148, + 0.229, + 0.451, + 0.33 + ], + "angle": 0, + "content": "
AgeBaselineSoc-Dem.Random
18 - 2459.39±1.5860.44±1.0560.52±1.37
25 - 3466.72±0.5666.63±0.8366.92±0.51
35 - 4464.50±0.5964.94±1.3365.24±0.89
45 - 5465.68±0.6665.88±1.3965.98±0.83
55 - 6464.37±1.2264.94±1.6664.84±1.30
65 or older63.34±2.0764.70±2.2162.77±2.39
" + }, + { + "type": "table", + "bbox": [ + 0.123, + 0.339, + 0.477, + 0.462 + ], + "angle": 0, + "content": "
EducationBaselineSoc-Dem.Random
Associate degree60.69±1.4460.54±2.3560.78±1.62
Bachelor's degree66.16±0.5166.23±0.8266.80±0.54
Doctoral degree61.93±3.8263.79±5.0363.27±3.67
High school60.53±1.3960.47±2.2260.55±1.87
Below high school58.28±4.6862.12±4.9060.17±4.25
Master's degree69.71±0.8669.58±0.9369.45±0.96
Professional degree66.75±2.3767.84±3.3268.62±2.84
College, no degree58.65±1.1959.40±1.7959.99±2.19
" + }, + { + "type": "table", + "bbox": [ + 0.141, + 0.472, + 0.458, + 0.537 + ], + "angle": 0, + "content": "
SexualityBaselineSoc-Dem.Random
Bisexual71.83±1.1471.42±1.5169.46±1.95
Heterosexual63.25±0.3963.32±1.2163.82±0.55
Homosexual64.43±1.7566.11±2.2065.12±1.94
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.545, + 0.49, + 0.618 + ], + "angle": 0, + "content": "Table 1: Average and standard deviation of macro \\( F_{1} \\) from three runs of four-fold stratified cross-validation. Separate table for each attribute. Bold results are the highest averages per group. However, no difference is statistically significant (see Appendix A.2)" + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.64, + 0.241, + 0.655 + ], + "angle": 0, + "content": "6 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.662, + 0.49, + 0.87 + ], + "angle": 0, + "content": "We do not find strong evidence that explicitly modelling sociodemographics helps to predict annotation behaviour with multi-annotator models. These results might seem counter-intuitive, given the evidence of systematic annotation differences between sociodemographic groups (see §2). This discrepancy, however, echoes the issue highlighted by ecological fallacies (Robinson, 1950): Not every annotator will be a perfect representative of their group, so we will not necessarily learn additional information based on their group identity. This seems especially true if we already have access to individual behaviour (i.e., individual annotations)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.872, + 0.49, + 0.919 + ], + "angle": 0, + "content": "In contrast to Davani et al. (2022), we made sociodemographic information explicit in our experiments, as one of the factors influencing annotation" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.085, + 0.885, + 0.262 + ], + "angle": 0, + "content": "behaviour. Group-specific layers can be seen as an inductive bias putting emphasis on the sociodemographic relations between annotators. However, there are potentially many other factors influencing annotation behaviour (e.g., attitudes, moral values, cognitive biases, psychological traits). In light of our results, it seems plausible that multi-annotator models learn about these factors implicitly as part of predicting individual behaviour, so that making one factor explicit does not change prediction quality, at least in the case of sociodemographics." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.266, + 0.885, + 0.603 + ], + "angle": 0, + "content": "Still, we also know that generally group attributes can help predict individual decisions, i.e., as base rates or priors. To avoid ecological fallacies in modelling annotation, we therefore need to better understand when and how modelling sociodemographic information is useful in predicting an individual annotator's decisions. For example, we have only evaluated group-specific layers for single attributes. In contrast, social scientists have long adopted the idea of intersectionality (Crenshaw, 1989), which also informs research on fairness in machine learning (Wang et al., 2022). Intersectionality means that the effect of interactions between sociodemographic attributes enables specific experiences that are not captured by the attributes in isolation. For example, identifying as a man means something different depending on the person's education. Groups derived from single attributes might simply be too coarse to improve classifiers learnt from individual labels, as in multi-annotator models." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.609, + 0.884, + 0.802 + ], + "angle": 0, + "content": "The dataset we use (Kumar et al., 2021) has many characteristics which are ideal for our study (see §3). However, it uses a broad notion of toxicity, in contrast to other studies of toxic language (Larimore et al., 2021; Sap et al., 2022), which match content and analysed groups. When modeling the groups frequently referenced in the datasets themselves, we would expect greater benefits from group-specific layers. Similar to us, Biester et al. (2022) who do not find significant differences between annotators of different genders, do so in a more general setting." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.807, + 0.885, + 0.919 + ], + "angle": 0, + "content": "We can only partially compare to Gordon et al. (2022), despite using the same dataset. In addition to differences in approach (see §2), our and their work also differ in their research questions and thus experimental conditions. Gordon et al. (2022) compare their full model (group and individual) against using group information alone." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.928, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1020" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.493, + 0.295 + ], + "angle": 0, + "content": "We compare our full model (group and individual) against using individual information alone. So it is unclear if their model would benefit from group information in comparison to individual-level information alone. While they find an improvement from group information it is only in comparison to a baseline predicting not individual but aggregated labels. Additionally, the composition of test sets sampled from the full dataset differs between the studies: Gordon et al. (2022) use a test set of 5,000 comments, while we use 22,360 comments in a four-fold cross-validation. We leave an explicit comparison to future work." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.296, + 0.49, + 0.488 + ], + "angle": 0, + "content": "Group-specific layers (§4) are a natural extension of annotator-specific classification layers in multi-annotator models. However, other architectures to predict annotator-level labels use different ways to represent sociodemographic information, e.g., via embeddings in a recommender system (Gordon et al., 2022). Future work could explore additional representations of annotator attributes (e.g., as part of the input, either textual or as separate features) and other approaches to modelling the relation of individual labeling decisions and attributes (e.g., probabilistic graphical models)." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.495, + 0.248, + 0.51 + ], + "angle": 0, + "content": "7 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.519, + 0.49, + 0.826 + ], + "angle": 0, + "content": "We ask how relevant modelling explicit sociodemographic information is in learning from individual annotators. Our experiments with group-specific layers for four sociodemographic attributes on social media data with toxicity annotations (Kumar et al., 2021) show no significant benefit of modelling sociodemographic groups in multi-annotator models. However, as the issue of ecological fallacies highlights, it is not implausible that these models do not learn additional information from group information beyond the inherent variation. However, our results do not refute the usefulness of sociodemographic attributes in modelling annotation, but underscore the importance of their judicious use. Different tasks and model architectures will likely benefit to different extents. Ultimately, annotation behaviour is driven by complex factors and we will need to consider more than annotators' sociodemographics." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.833, + 0.287, + 0.849 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.855, + 0.49, + 0.92 + ], + "angle": 0, + "content": "We thank Deepak Kumar for providing access to the disaggregated dataset and his continued support. We also thank Aida Mostafazadeh Davani for providing information on implementation de" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.15 + ], + "angle": 0, + "content": "tails of multi-annotator models. Members of MilaNLP (Bocconi) and the Semantic Computing Group (Bielefeld) provided feedback on earlier versions of this paper, for which we thank them again." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.15, + 0.886, + 0.262 + ], + "angle": 0, + "content": "This work has in part been funded by the European Research Council (ERC) under the European Union's Horizon 2020 research and innovation program (grant agreement No. 949944, INTEGRACTOR). Likewise, this work has in part been funded by the VolkswagenStiftung as part of the \"3B Bots Building Bridges\" project." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.269, + 0.615, + 0.284 + ], + "angle": 0, + "content": "Limitations" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.291, + 0.885, + 0.501 + ], + "angle": 0, + "content": "While the dataset by Kumar et al. (2021) enabled us to test models for a range of often overlooked groups (e.g., non-binary or bisexual annotators), we ultimately modelled only four specific attributes (gender, age, education, sexual orientation). There are likely to be more factors that could play a role. Additionally, annotators in the Kumar et al. (2021) dataset are exclusively from the United States of America, so that results do not necessarily hold for other countries or cultures (Hovy and Yang, 2021). Specifically perceptions of harmful content online are known to vary across countries (Jiang et al., 2021)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.501, + 0.885, + 0.791 + ], + "angle": 0, + "content": "We used only the (Kumar et al., 2021) dataset. This is mainly due to our strict criteria regarding dataset size and availability of annotator-level labels and sociodemographic information. These characteristics were a prerequisite for our experiments across different attributes with sufficient numbers of annotators. Most datasets which include annotator-level labels and sociodemographic information contain much smaller numbers of annotators and attributes. Nevertheless, with the Measuring Hate Speech Corpus there is at least one additional dataset (Sachdeva et al., 2022) with comparable characteristics that could be used in future experiments. Also, additional small-scale, more focused experiments could use datasets like Sap et al. (2022) or HS-Brexit (Akhtar et al., 2021) which was annotated by 6 annotators, each from one of two sociodemographic groups." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.791, + 0.885, + 0.92 + ], + "angle": 0, + "content": "We do not study the aggregation of individual predictions or evaluate against majority labels, as these are not directly relevant to our investigation of sociodemographic attributes in models of annotation behaviour. Consequently, we cannot derive a conclusion about performance in those settings from our results. This is a noteworthy limitation, because part of the experiments introducing" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.519, + 0.941 + ], + "angle": 0, + "content": "1021" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.489, + 0.149 + ], + "angle": 0, + "content": "multi-annotator models in Davani et al. (2022) compare labels aggregated from multi-annotator models against predictions from a standard classifier (directly trained on aggregated labels)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.151, + 0.488, + 0.214 + ], + "angle": 0, + "content": "For computational reasons, our experiments use a comparatively small pre-trained language model (RoBERTa, Liu et al. 2019). Thus, results might differ with larger models." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.222, + 0.266, + 0.237 + ], + "angle": 0, + "content": "Ethics Statement" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.247, + 0.489, + 0.39 + ], + "angle": 0, + "content": "As sociodemographic attributes are sensitive information, we do not infer attributes, but build on a self-reported, IRB-reviewed dataset (Kumar et al., 2021). We also see potential for a discussion of \"privacy by design\" in modelling human label variation based on our results: There can be circumstances in which knowing more about annotators is not relevant, and indeed might lead to violations of privacy." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.393, + 0.49, + 0.569 + ], + "angle": 0, + "content": "As multi-annotator models attempt to capture the preferences of individual annotators, there are valid concerns around privacy and anonymity. As discussed in Davani et al. (2022), increasing the annotator count can be one option to reduce privacy risks. We show it is feasible to learn a model for a large number of individual annotators (5002 vs. 18 and 82 in their work). But a prerequisite for improved privacy is to apply effective aggregation on top of individual predictions, which we do not study in the present work." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.593, + 0.214, + 0.608 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.615, + 0.49, + 0.682 + ], + "angle": 0, + "content": "Gavin Abercrombie, Valerio Basile, Sara Tonelli, Verena Rieser, and Alexandra Uma, editors. 2022. Proceedings of the 1st Workshop on Perspectivist Approaches to NLP @LREC2022. European Language Resources Association, Marseille, France." + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.694, + 0.49, + 0.761 + ], + "angle": 0, + "content": "Sohail Akhtar, Valerio Basile, and Viviana Patti. 2020. Modeling annotator perspective and polarized opinions to improve hate speech detection. In Proceedings of the AAAI Conference on Human Computation and Crowdsourcing, volume 8, pages 151-154." + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.773, + 0.49, + 0.827 + ], + "angle": 0, + "content": "Sohail Akhtar, Valerio Basile, and Viviana Patti. 2021. Whose opinions matter? perspective-aware models to identify opinions of hate speech victims in abusive language detection. Preprint arXiv:2106.15896." + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.84, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Hala Al Kuwatly, Maximilian Wich, and Georg Groh. 2020. Identifying and measuring annotator bias based on annotators' demographic characteristics. In Proceedings of the Fourth Workshop on Online Abuse and Harms, pages 184-190, Online. Association for Computational Linguistics." + }, + { + "type": "list", + "bbox": [ + 0.116, + 0.615, + 0.49, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.085, + 0.885, + 0.179 + ], + "angle": 0, + "content": "Valerio Basile, Michael Fell, Tommaso Fornaciari, Dirk Hovy, Silviu Paun, Barbara Plank, Massimo Poesio, and Alexandra Uma. 2021. We need to consider disagreement in evaluation. In Proceedings of the 1st Workshop on Benchmarking: Past, Present and Future, pages 15-21, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.191, + 0.885, + 0.283 + ], + "angle": 0, + "content": "Laura Biester, Vanita Sharma, Ashkan Kazemi, Naihao Deng, Steven Wilson, and Rada Mihalcea. 2022. Analyzing the effects of annotator gender across NLP tasks. In Proceedings of the 1st Workshop on Perspectivist Approaches to NLP @LREC2022, pages 10-19, Marseille, France. European Language Resources Association." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.297, + 0.885, + 0.376 + ], + "angle": 0, + "content": "Reuben Binns, Michael Veale, Max Van Kleek, and Nigel Shadbolt. 2017. Like trainer, like bot? inheritance of bias in algorithmic content moderation. In Social Informatics, Lecture Notes in Computer Science, pages 405-415. Springer International Publishing." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.39, + 0.885, + 0.482 + ], + "angle": 0, + "content": "Amanda Cercas Curry, Gavin Abercrombie, and Verena Rieser. 2021. ConvAbuse: Data, analysis, and benchmarks for nuanced abuse detection in conversational AI. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7388-7403, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.495, + 0.885, + 0.561 + ], + "angle": 0, + "content": "Kimberle Crenshaw. 1989. Demarginalizing the intersection of race and sex: A black feminist critique of antidiscrimination doctrine, feminist theory and antiracist politics. University of Chicago Legal Forum, 1989(1):Article 8." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.575, + 0.885, + 0.641 + ], + "angle": 0, + "content": "Aida Mostafazadeh Davani, Mark Diaz, and Vinodkumar Prabhakaran. 2022. Dealing with disagreements: Looking beyond the majority vote in subjective annotations. Transactions of the Association for Computational Linguistics, 10:92-110." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.655, + 0.885, + 0.721 + ], + "angle": 0, + "content": "Rotem Dror, Gili Baumer, Marina Bogomolov, and Roi Reichart. 2017. Replicability analysis for natural language processing: Testing significance with multiple datasets. Transactions of the Association for Computational Linguistics, 5:471-486." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.734, + 0.885, + 0.827 + ], + "angle": 0, + "content": "Rotem Dror, Gili Baumer, Segev Shlomov, and Roi Reichart. 2018. The hitchhiker's guide to testing statistical significance in natural language processing. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1383-1392, Melbourne, Australia. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.84, + 0.885, + 0.919 + ], + "angle": 0, + "content": "Elizabeth Excell and Noura Al Moubayed. 2021. Towards equal gender representation in the annotations of toxic language detection. In Proceedings of the 3rd Workshop on Gender Bias in Natural Language Processing, pages 55-65, Online. Association for Computational Linguistics." + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.085, + 0.885, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1022" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.204 + ], + "angle": 0, + "content": "Tommaso Fornaciari, Alexandra Uma, Silviu Paun, Barbara Plank, Dirk Hovy, and Massimo Poesio. 2021. Beyond black & white: Leveraging annotator disagreement via soft-label multi-task learning. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2591-2597, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.215, + 0.488, + 0.307 + ], + "angle": 0, + "content": "Tommaso Fornaciari, Alexandra Uma, Massimo Poesio, and Dirk Hovy. 2022. Hard and soft evaluation of NLP models with BOOtSTrap SAmpling - BooStSa. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 127-134, Dublin, Ireland. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.317, + 0.488, + 0.37 + ], + "angle": 0, + "content": "David A. Freedman. 2015. Ecological inference. In James D. Wright, editor, International Encyclopedia of the Social & Behavioral Sciences (Second Edition), pages 868-870. Elsevier." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.38, + 0.488, + 0.472 + ], + "angle": 0, + "content": "Mitchell L. Gordon, Michelle S. Lam, Joon Sung Park, Kayur Patel, Jeff Hancock, Tatsunori Hashimoto, and Michael S. Bernstein. 2022. Jury learning: Integrating dissenting voices into machine learning models. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems, CHI '22, pages 1-19. Association for Computing Machinery." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.482, + 0.488, + 0.548 + ], + "angle": 0, + "content": "Nitesh Goyal, Ian D. Kivlichan, Rachel Rosen, and Lucy Vasserman. 2022. Is your toxicity my toxicity? exploring the impact of rater identity on toxicity annotation. Proceedings of the ACM on Human-Computer Interaction, 6:1-28." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.559, + 0.488, + 0.65 + ], + "angle": 0, + "content": "Dirk Hovy and Diyi Yang. 2021. The importance of modeling social factors of language: Theory and practice. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 588-602, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.661, + 0.488, + 0.753 + ], + "angle": 0, + "content": "Emily Jamison and Iryna Gurevych. 2015. Noise or additional information? leveraging crowdsourced annotation item agreement for natural language tasks. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 291-297, Lisbon, Portugal. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.763, + 0.488, + 0.816 + ], + "angle": 0, + "content": "Jialun Aaron Jiang, Morgan Klaus Scheuerman, Casey Fiesler, and Jed R. Brubaker. 2021. Understanding international perceptions of the severity of harmful content online. PLOS ONE, 16(8)." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.826, + 0.488, + 0.918 + ], + "angle": 0, + "content": "Deepak Kumar, Patrick Gage Kelley, Sunny Consolvo, Joshua Mason, Elie Bursztein, Zakir Durmeric, Kurt Thomas, and Michael Bailey. 2021. Designing toxic content classification for a diversity of perspectives. In Seventeenth Symposium on Usable Privacy and Security (SOUPS 2021), pages 299-318. USENIX Association." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.086, + 0.883, + 0.178 + ], + "angle": 0, + "content": "Savannah Larimore, Ian Kennedy, Breon Haskett, and Alina Arseniev-Koehler. 2021. Reconsidering annotator disagreement about racist language: Noise or signal? In Proceedings of the Ninth International Workshop on Natural Language Processing for Social Media, pages 81-90, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.187, + 0.883, + 0.253 + ], + "angle": 0, + "content": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. Preprint arXiv:1907.11692." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.261, + 0.883, + 0.352 + ], + "angle": 0, + "content": "F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. 2011. Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12:2825-2830." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.362, + 0.883, + 0.441 + ], + "angle": 0, + "content": "Barbara Plank. 2022. The \"problem\" of human label variation: On ground truth in data, modeling and evaluation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 10671-10682, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.449, + 0.883, + 0.542 + ], + "angle": 0, + "content": "Barbara Plank, Dirk Hovy, and Anders Søgaard. 2014. Learning part-of-speech taggers with inter-annotator agreement loss. In Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 742–751, Gothenburg, Sweden. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.551, + 0.883, + 0.643 + ], + "angle": 0, + "content": "Vinodkumar Prabhakaran, Aida Mostafazadeh Davani, and Mark Diaz. 2021. On releasing annotator-level labels and information in datasets. In Proceedings of the Joint 15th Linguistic Annotation Workshop (LAW) and 3rd Designing Meaning Representations (DMR) Workshop, pages 133-138, Punta Cana, Dominican Republic. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.651, + 0.883, + 0.691 + ], + "angle": 0, + "content": "W. S. Robinson. 1950. Ecological correlations and the behavior of individuals. American Sociological Review, 15(3):351-357." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.7, + 0.883, + 0.805 + ], + "angle": 0, + "content": "Paul Röttger, Bertie Vidgen, Dirk Hovy, and Janet Pierrehumbert. 2022. Two contrasting data annotation paradigms for subjective NLP tasks. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 175-190, Seattle, United States. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.814, + 0.883, + 0.918 + ], + "angle": 0, + "content": "Pratik Sachdeva, Renata Barreto, Geoff Bacon, Alexander Sahn, Claudia von Vacano, and Chris Kennedy. 2022. The measuring hate speech corpus: Leveraging rasch measurement theory for data perspectivism. In Proceedings of the 1st Workshop on Perspectivist Approaches to NLP @LREC2022, pages 83-94, Marseille, France. European Language Resources Association." + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.086, + 0.883, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1023" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.086, + 0.488, + 0.152 + ], + "angle": 0, + "content": "Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. 2019. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. In 5th Workshop on Energy Efficient Machine Learning and Cognitive Computing @ NeurIPS 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.162, + 0.488, + 0.241 + ], + "angle": 0, + "content": "Maarten Sap, Dallas Card, Saadia Gabriel, Yejin Choi, and Noah A. Smith. 2019. The risk of racial bias in hate speech detection. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1668-1678, Florence, Italy. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.25, + 0.488, + 0.368 + ], + "angle": 0, + "content": "Maarten Sap, Swabha Swayamdipta, Laura Vianna, Xuhui Zhou, Yejin Choi, and Noah A. Smith. 2022. Annotators with attitudes: How annotator beliefs and identities bias toxic language detection. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5884-5906, Seattle, United States. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.378, + 0.488, + 0.469 + ], + "angle": 0, + "content": "Qinlan Shen and Carolyn Rose. 2021. What sounds \"right\" to me? experiential factors in the perception of political ideology. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 1762-1771, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.48, + 0.488, + 0.544 + ], + "angle": 0, + "content": "Alexandra N. Uma, Tommaso Fornaciari, Dirk Hovy, Silviu Paun, Barbara Plank, and Massimo Poesio. 2021. Learning from disagreement: A survey. Journal of Artificial Intelligence Research, 72:1385-1470." + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.555, + 0.488, + 0.647 + ], + "angle": 0, + "content": "Angelina Wang, Vikram V Ramaswamy, and Olga Russakovsky. 2022. Towards intersectionality in machine learning: Including more identities, handling underrepresentation, and performing evaluation. In 2022 ACM Conference on Fairness, Accountability, and Transparency, FAccT '22, pages 336-349. Association for Computing Machinery." + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.656, + 0.488, + 0.813 + ], + "angle": 0, + "content": "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online. Association for Computational Linguistics." + }, + { + "type": "title", + "bbox": [ + 0.116, + 0.826, + 0.237, + 0.842 + ], + "angle": 0, + "content": "A Appendix" + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.85, + 0.486, + 0.865 + ], + "angle": 0, + "content": "A.1 Annotator Sociodemographics in Sample" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.872, + 0.486, + 0.919 + ], + "angle": 0, + "content": "Table 2 shows how many annotators the sample contains. Counts are given per group of the four attributes gender, age, education and sexuality." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.085, + 0.884, + 0.358 + ], + "angle": 0, + "content": "In the Kumar et al. (2021) dataset, sociodemographic attributes are given for each individual annotation - not once per annotator. For some annotators, conflicting attribute values exist (e.g., two different age groups). As the data collection spanned several months (Kumar et al., 2021), these value changes can in principle be reasonable (e.g., because an annotator got older, finished a degree, changed sexual preference or gender identity). However, as reasonable changes can not easily be discerned from erroneous input, we disambiguate values based on a heuristic: If an annotator reports several values for an attribute, we assume the most frequent value to be valid. In cases of no clear most frequent value, we set the attribute to \"Prefer not to say\". Thus, the main results do not contain annotators with ambiguous attributes." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.367, + 0.698, + 0.381 + ], + "angle": 0, + "content": "A.2 Significance Tests" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.389, + 0.884, + 0.581 + ], + "angle": 0, + "content": "Results of a replicability analysis (Dror et al., 2017) testing for significant differences in macro \\(F_{1}\\) on scores from three runs of four-fold cross-validation. Table 3 shows results for a comparison of the sociodemographic models against the baseline models. Table 4 shows results for a comparison of the sociodemographic models against the randomized assignment models. The Bonferroni correction for the corrected count of significant folds \\(\\hat{k}_{Bonferroni}\\) is used to account for the fact that we have overlapping test sets from multiple runs of four-fold cross-validation." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.59, + 0.87, + 0.622 + ], + "angle": 0, + "content": "A.3 Training Details, Hyperparameters and Computational Resources" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.629, + 0.884, + 0.789 + ], + "angle": 0, + "content": "We implement models and the training loop using the Hugging Face Transformers library (version 4.19.2, Wolf et al. 2020). Maximum sequence length is 512 tokens, with truncation and padding to the maximum length. We train for 3 epochs with a batch size of 8 and an initial learning rate of 0.00001. Otherwise, we used default parameters. We found results to particularly depend on the learning rate, with higher or lower values leading to worse results." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.791, + 0.884, + 0.918 + ], + "angle": 0, + "content": "We use a weighted loss function. Label weights are calculated per annotator on the training set of each fold. Label weights, evaluation scores and the four-fold dataset splits (StratifiedKFold) are calculated using the scikit-learn library (version 1.0.2, Pedregosa et al. 2011). The folds are based on a fixed random seed per iteration: 2803636207, 165043843, 2923262358" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1024" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.172, + 0.081, + 0.429, + 0.182 + ], + "angle": 0, + "content": "
Number of Annotators
Gender
Female2450
Male2116
Prefer not to say412
Nonbinary23
Other1
" + }, + { + "type": "table", + "bbox": [ + 0.172, + 0.189, + 0.427, + 0.314 + ], + "angle": 0, + "content": "
Number of Annotators
Age
18 - 24489
25 - 341861
35 - 441115
45 - 54529
55 - 64321
65 or older119
Prefer not to say568
" + }, + { + "type": "table", + "bbox": [ + 0.172, + 0.323, + 0.429, + 0.424 + ], + "angle": 0, + "content": "
Number of Annotators
Sexuality
Heterosexual4018
Bisexual469
Prefer not to say346
Homosexual134
Other35
" + }, + { + "type": "table", + "bbox": [ + 0.165, + 0.434, + 0.436, + 0.593 + ], + "angle": 0, + "content": "
EducationNumber of Annotators
Bachelor's degree1879
College, no degree861
Prefer not to say647
Master's degree642
Associate degree460
High school363
Professional degree68
Doctoral degree51
Below high school25
Other6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.602, + 0.489, + 0.645 + ], + "angle": 0, + "content": "Table 2: Number of annotators per group for attributes gender, age, sexuality and education. Counts refer to the entire sample" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.675, + 0.489, + 0.836 + ], + "angle": 0, + "content": "The majority of parameters in our model belong to the pre-trained language model shared between all group-specific and annotator-specific layers. Specifically, RoBERTa (Liu et al., 2019) in the roberta-base variant has 125 Million parameters. We keep the pre-trained model's default output dimensionality of 768, so that each group-specific layer adds \\(768 * 768 + 768 = 590\\), 592 parameters and each annotator layer adds \\(768 * 2 + 2 = 1\\), 538 parameters." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.839, + 0.49, + 0.919 + ], + "angle": 0, + "content": "All experiments ran on a single GPU (GeForce GTX 1080 Ti, 12GB GPU RAM). Per fold, training and evaluation together take about three and a half hours in our setting. Three runs of four-fold cross-validation (12 folds), thus take around 42 hours" + }, + { + "type": "table", + "bbox": [ + 0.607, + 0.082, + 0.787, + 0.148 + ], + "angle": 0, + "content": "
hatkcounthatkBonf.
Female20
Male00
Nonbinary10
" + }, + { + "type": "table", + "bbox": [ + 0.607, + 0.157, + 0.787, + 0.26 + ], + "angle": 0, + "content": "
hatkcounthatkBonf.
18 - 2420
25 - 3420
35 - 4410
45 - 5400
55 - 6410
65 or older10
" + }, + { + "type": "table", + "bbox": [ + 0.607, + 0.269, + 0.787, + 0.334 + ], + "angle": 0, + "content": "
ˆkcountˆkBonf.
Bisexual20
Heterosexual42
Homosexual10
" + }, + { + "type": "table", + "bbox": [ + 0.582, + 0.345, + 0.808, + 0.469 + ], + "angle": 0, + "content": "
kcountkBonf.
Associate degree00
Bachelor's degree10
Doctoral degree20
High school00
Belowhigh school00
Master's degree00
Professional degree00
College, no degree22
" + }, + { + "type": "table_caption", + "bbox": [ + 0.508, + 0.479, + 0.884, + 0.593 + ], + "angle": 0, + "content": "Table 3: Results of a replicability analysis of baseline vs sociodemographic models. Raw and Bonferroni-corrected counts of significant folds out of 12 folds from three runs of four-fold cross-validation. P-values for each fold are computed via a paired bootstrap test with significance level \\(\\alpha = 0.05\\), 1000 bootstrap samples per fold and a sample size of \\(50\\%\\) of the respective test set." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.624, + 0.884, + 0.786 + ], + "angle": 0, + "content": "(1.75 days). With four attributes and three trainable models the combined run time of the reported experiments is estimated to be 21 days. Including preliminary experiments, which, however, mostly were not full runs of k-fold cross-validation and also utilized DistilBERT (Sanh et al., 2019) with slightly faster run times, it will be many times more. There is no discernible difference in experiment run times between multi-annotator models with or without groups or different numbers of groups." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.798, + 0.882, + 0.827 + ], + "angle": 0, + "content": "A.4 Number of Annotations per Group across all Test Sets" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.839, + 0.884, + 0.919 + ], + "angle": 0, + "content": "Table 5 contains the number of annotations we have per group across the total of 12 folds (from three runs of four-fold cross-validation). This number of annotations is the effective test set size per group. As the numbers do not vary substantially, perfor" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1025" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.212, + 0.083, + 0.388, + 0.147 + ], + "angle": 0, + "content": "
kcountkBonf.
Female22
Male10
Nonbinary10
" + }, + { + "type": "table", + "bbox": [ + 0.212, + 0.159, + 0.386, + 0.259 + ], + "angle": 0, + "content": "
kcountkBonf.
18 - 2410
25 - 3400
35 - 4410
45 - 5410
55 - 6430
65 or older10
" + }, + { + "type": "table", + "bbox": [ + 0.207, + 0.271, + 0.393, + 0.332 + ], + "angle": 0, + "content": "
hatcounthatBonf.
Bisexual62
Heterosexual11
Homosexual00
" + }, + { + "type": "table", + "bbox": [ + 0.188, + 0.345, + 0.412, + 0.469 + ], + "angle": 0, + "content": "
kcountkBonf.
Associate degree20
Bachelor's degree10
Doctoral degree00
High school20
Belowhigh school20
Master's degree00
Professional degree00
College, no degree11
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.479, + 0.49, + 0.593 + ], + "angle": 0, + "content": "Table 4: Results from replicability analysis of randomized vs sociodemographic models. Raw and Bonferroni-corrected counts of significant folds out of 12 folds from three runs of four-fold cross-validation. P-values for each fold are computed via a paired bootstrap test with significance level \\(\\alpha = 0.05\\), 1000 bootstrap samples per fold and a sample size of \\(50\\%\\) of the respective test set." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.619, + 0.489, + 0.652 + ], + "angle": 0, + "content": "mance on each fold is equally representative for all groups." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.657, + 0.26, + 0.672 + ], + "angle": 0, + "content": "A.5 Full Results" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.679, + 0.49, + 0.727 + ], + "angle": 0, + "content": "Table 6 shows full results of experiments (see 4), including results for all residual categories and a naive baseline which always predicts toxic." + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.202, + 0.886, + 0.302 + ], + "angle": 0, + "content": "
GenderNumber Of AnnotationsMinMax
Female13555±86.4413383.013664.0
Male11925±61.6511843.012062.0
Nonbinary115±6.03104.0122.0
Other5±1.952.08.0
Prefer not to say2345±51.192281.02453.0
" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.311, + 0.872, + 0.435 + ], + "angle": 0, + "content": "
AgeNumber Of AnnotationsMinMax
18 - 242615±50.8825212697
25 - 3410315±61.451024410457
35 - 446250±51.0661796324
45 - 543025±47.2329293083
55 - 641865±25.4818311903
65 or older675±19.31643704
Prefer not to say3200±55.2831313289
" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.444, + 0.872, + 0.544 + ], + "angle": 0, + "content": "
SexualityNumber Of AnnotationsMinMax
Bisexual2445±39.2623832501
Heterosexual22630±63.002250722726
Homosexual725±26.57670759
Other190±7.91173201
Prefer not to say1955±35.3918782009
" + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.554, + 0.882, + 0.713 + ], + "angle": 0, + "content": "
EducationNumber Of AnnotationsMinMax
Associate professor2605±47.5925162697
Bachelor's degree10510±84.791034810700
Doctoral degree305±18.83270332
High school2080±37.0120152139
Below high school165±11.17144184
Master's degree3515±48.0834253580
Other30±3.442536
Prefer not to say3690±52.9236033808
Professional degree380±17.87352411
College, no degree4665±71.3645394776
" + }, + { + "type": "table_caption", + "bbox": [ + 0.509, + 0.724, + 0.884, + 0.795 + ], + "angle": 0, + "content": "Table 5: Average, standard deviation, minimum and maximum of number of annotations per fold. All information given per group of gender, age, education and sexuality. Statistics are calculated across 12 folds from three runs of four-fold cross-validation." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1026" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.264, + 0.248, + 0.731, + 0.336 + ], + "angle": 0, + "content": "
GenderMajority BaselineBaselineSoc-Dem.Random
Female41.79±0.1262.23±0.5362.25±1.1962.41±0.92
Male40.53±0.1168.00±0.4967.66±0.4667.63±0.53
Nonbinary44.69±1.3956.33±6.0056.80±7.2458.00±7.49
Other45.50±4.6948.56±10.7850.53±14.6343.66±7.25
Prefer not to say41.05±0.3664.54±1.1365.05±1.5265.08±1.86
" + }, + { + "type": "table", + "bbox": [ + 0.264, + 0.345, + 0.724, + 0.458 + ], + "angle": 0, + "content": "
AgeMajority BaselineBaselineSoc-Dem.Random
18 - 2442.49±0.2859.39±1.5860.44±1.0560.52±1.37
25 - 3440.49±0.0966.72±0.5666.63±0.8366.92±0.51
35 - 4441.87±0.1564.50±0.5964.94±1.3365.24±0.89
45 - 5440.63±0.2665.68±0.6665.88±1.3965.98±0.83
55 - 6441.65±0.3964.37±1.2264.94±1.6664.84±1.30
65 or older41.46±0.5463.34±2.0764.70±2.2162.77±2.39
Prefer not to say41.37±0.3263.99±1.3265.24±1.1864.73±1.33
" + }, + { + "type": "table", + "bbox": [ + 0.255, + 0.467, + 0.74, + 0.613 + ], + "angle": 0, + "content": "
EducationMajority BaselineBaselineSoc-Dem.Random
Associate degree43.16±0.1960.69±1.4460.54±2.3560.78±1.62
Bachelor's degree40.38±0.1066.16±0.5166.23±0.8266.80±0.54
Doctoral degree43.34±0.9461.93±3.8263.79±5.0363.27±3.67
High school43.02±0.2660.53±1.3960.47±2.2260.55±1.87
Below high school43.10±1.4458.28±4.6862.12±4.9060.17±4.25
Master's degree37.55±0.3269.71±0.8669.58±0.9369.45±0.96
Other42.95±2.3156.56±10.8857.59±9.8657.71±12.28
Prefer not to say40.97±0.2765.07±1.1665.69±1.0565.74±1.09
Professional degree40.43±0.8066.75±2.3767.84±3.3268.62±2.84
College, no degree43.61±0.1858.65±1.1959.40±1.7959.99±2.19
" + }, + { + "type": "table", + "bbox": [ + 0.271, + 0.622, + 0.724, + 0.711 + ], + "angle": 0, + "content": "
SexualityMajority BaselineBaselineSoc-Dem.Random
Bisexual34.69±0.5071.83±1.1471.42±1.5169.46±1.95
Heterosexual41.99±0.0663.25±0.3963.32±1.2163.82±0.55
Homosexual41.15±0.4164.43±1.7566.11±2.2065.12±1.94
Other43.53±0.7857.55±3.7960.57±4.5158.69±4.72
Prefer not to say39.12±0.2467.80±1.5667.27±1.5267.46±1.11
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.72, + 0.883, + 0.751 + ], + "angle": 0, + "content": "Table 6: Average and standard deviation of macro \\( F_{1} \\) from three runs of four-fold stratified cross-validation. Separate table for each attribute. Bold results are the highest average per group. Full results including naive majority baseline" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1027" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.133, + 0.085, + 0.435, + 0.1 + ], + "angle": 0, + "content": "ACL 2023 Responsible NLP Checklist" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.108, + 0.323, + 0.123 + ], + "angle": 0, + "content": "A For every submission:" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.127, + 0.533, + 0.158 + ], + "angle": 0, + "content": "A1. Did you describe the limitations of your work? Limitations, 8" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.169, + 0.553, + 0.201 + ], + "angle": 0, + "content": "A2. Did you discuss any potential risks of your work? Ethics Statement, 9" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.212, + 0.696, + 0.244 + ], + "angle": 0, + "content": "A3. Do the abstract and introduction summarize the paper's main claims?" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.256, + 0.669, + 0.288 + ], + "angle": 0, + "content": "A4. Have you used AI writing assistants when working on this paper? Left blank." + }, + { + "type": "list", + "bbox": [ + 0.13, + 0.127, + 0.696, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.3, + 0.489, + 0.317 + ], + "angle": 0, + "content": "B Did you use or create scientific artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.322, + 0.255, + 0.337 + ], + "angle": 0, + "content": "3, Appendix A.3" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.347, + 0.53, + 0.38 + ], + "angle": 0, + "content": "B1. Did you cite the creators of artifacts you used? 3, Appendix A.3" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.391, + 0.779, + 0.423 + ], + "angle": 0, + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Clear from context, citations" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.434, + 0.881, + 0.513 + ], + "angle": 0, + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Clear from context, citations" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.524, + 0.881, + 0.588 + ], + "angle": 0, + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? 3, Ethics Statement 9" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.599, + 0.881, + 0.648 + ], + "angle": 0, + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? 3, Appendix A.1" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.658, + 0.881, + 0.755 + ], + "angle": 0, + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. 3, 4, Appendix A.4" + }, + { + "type": "list", + "bbox": [ + 0.13, + 0.347, + 0.881, + 0.755 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.765, + 0.494, + 0.782 + ], + "angle": 0, + "content": "C Did you run computational experiments?" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.789, + 0.147, + 0.8 + ], + "angle": 0, + "content": "4" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.812, + 0.881, + 0.861 + ], + "angle": 0, + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Appendix A.3" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.869, + 0.878, + 0.893 + ], + "angle": 0, + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1028" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.129, + 0.085, + 0.88, + 0.133 + ], + "angle": 0, + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.144, + 0.881, + 0.207 + ], + "angle": 0, + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? No response." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.218, + 0.881, + 0.283 + ], + "angle": 0, + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Appendix A.3" + }, + { + "type": "list", + "bbox": [ + 0.129, + 0.085, + 0.881, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.294, + 0.878, + 0.331 + ], + "angle": 0, + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.342, + 0.881, + 0.388 + ], + "angle": 0, + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.401, + 0.881, + 0.464 + ], + "angle": 0, + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.476, + 0.881, + 0.54 + ], + "angle": 0, + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.551, + 0.873, + 0.582 + ], + "angle": 0, + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.593, + 0.881, + 0.641 + ], + "angle": 0, + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response." + }, + { + "type": "list", + "bbox": [ + 0.129, + 0.342, + 0.881, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1029" + } + ] +] \ No newline at end of file diff --git a/2023/The Ecological Fallacy in Annotation_ Modeling Human Label Variation goes beyond Sociodemographics/8b624391-c7ce-443d-9caf-875f1d838500_origin.pdf b/2023/The Ecological Fallacy in Annotation_ Modeling Human Label Variation goes beyond Sociodemographics/8b624391-c7ce-443d-9caf-875f1d838500_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a93a3021aac6cbec7c1f12cfe42c3c920a713f18 --- /dev/null +++ b/2023/The Ecological Fallacy in Annotation_ Modeling Human Label Variation goes beyond Sociodemographics/8b624391-c7ce-443d-9caf-875f1d838500_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3f83769564b1459a6bb48a22826a7e4a65a7d42bfa0849bfcaf36e884126fbd5 +size 348180 diff --git a/2023/The Ecological Fallacy in Annotation_ Modeling Human Label Variation goes beyond Sociodemographics/full.md b/2023/The Ecological Fallacy in Annotation_ Modeling Human Label Variation goes beyond Sociodemographics/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0184c8e917d6da6dac2a73aad80b693be1280400 --- /dev/null +++ b/2023/The Ecological Fallacy in Annotation_ Modeling Human Label Variation goes beyond Sociodemographics/full.md @@ -0,0 +1,313 @@ +# The Ecological Fallacy in Annotation: Modelling Human Label Variation goes beyond Sociodemographics + +Matthias Orlikowski1, Paul Röttger2, Philipp Cimiano1, and Dirk Hovy3 + +$^{1}$ Bielefeld University + +$^{2}$ University of Oxford + +3Computing Sciences Department, Bocconi University, Milan, Italy + +# Abstract + +Many NLP tasks exhibit human label variation, where different annotators give different labels to the same texts. This variation is known to depend, at least in part, on the sociodemographics of annotators. Recent research aims to model individual annotator behaviour rather than predicting aggregated labels, and we would expect that sociodemographic information is useful for these models. On the other hand, the ecological fallacy states that aggregate group behaviour, such as the behaviour of the average female annotator, does not necessarily explain individual behaviour. To account for sociodemographics in models of individual annotator behaviour, we introduce group-specific layers to multi-annotator models. In a series of experiments for toxic content detection, we find that explicitly accounting for sociodemographic attributes in this way does not significantly improve model performance. This result shows that individual annotation behaviour depends on much more than just sociodemographics. + +# 1 Introduction + +Different annotators will not necessarily assign the same labels to the same texts, resulting in human label variation (Plank, 2022). Previous work finds that this variation depends at least in part on the sociodemographics of annotators, such as their age and gender (Binns et al., 2017; Al Kuwatly et al., 2020; Excell and Al Moubayed, 2021; Shen and Rose, 2021). These results are particularly pronounced for subjective tasks like toxic content detection (Sap et al., 2019; Kumar et al., 2021; Sap et al., 2022; Goyal et al., 2022). Since human label variation is relevant to a wide range of NLP tasks, recent research has begun to model individual annotator behaviour, rather than predicting aggregated labels (Davani et al., 2022; Gordon et al., 2022). In this setting, we would expect sociodemographic attributes to help explain annotator decisions. Therefore, we investigate whether explicitly + +![](images/0261e22efde8dee493e93beebcd49541625a911d2b7967e05ba5f1df0447ab78.jpg) +Figure 1: Group-specific layers representing annotator sociodemographics in multi-annotator models. + +accounting for the sociodemographic attributes of annotators leads to better predictions of their annotation behaviour1. + +There is a risk of misreading these efforts as an example of the ecological fallacy: aggregate group behaviour does not necessarily explain individual behaviour (Robinson, 1950; Freedman, 2015). For example, while on average, white annotators may be more likely to label African-American Vernacular English as toxic (Sap et al., 2019), that does not mean it is true for every white annotator individually. However, we aim at exactly this distinction to discuss the relevance of sociodemographic groups in models of individual annotator behaviour. Likewise, we do not assume prior work to commit ecological fallacies, even if a less-nuanced read might suggest it. + +Davani et al. (2022) introduce a simple multi-annotator model, where each annotator is modelled with a separate classification head. We expand their model with group-specific layers, which are activated for each annotator based on their sociodemographic attributes. We compare the two model setups to a control setup where we randomise group assignments. All comparisons use annotator-level toxicity data from Kumar et al. (2021). We find that find that explicitly accounting for sociodem + +graphic attributes does not significantly improve model performance. This result suggests that human label variation happens at a more individual level than sociodemographics, and that annotator decisions are even more complex. + +Contributions 1) We introduce group-specific layers to model groups of annotators with shared attributes in multi-annotator models. 2) We evaluate the effect of group-specific layers for toxic content detection, and show that explicitly accounting for sociodemographic attributes does not significantly improve performance, thus highlighting the risk of the ecological fallacy in annotator modelling. + +As a corollary, we show that multi-annotator models can be applied to many times more annotators than in prior work. + +# 2 Related Work + +Sociodemographics in Annotation Behaviour A growing body of research studies how annotator sociodemographics relate to their annotation decisions, for tasks ranging from natural language inference (Biester et al., 2022) to the detection of racist (Larimore et al., 2021) or generally toxic (Sap et al., 2022) language. Goyal et al. (2022), for example, find that annotators from certain sociodemographic groups (e.g., LGBTQ people) tend to find content attacking their own groups (e.g., homophobic content) to be more toxic. This motivates our research into explicitly accounting for sociodemographics to model annotation behaviour. However, the link between sociodemographics and behaviour is not uncontested. Biester et al. (2022), for example, do not find significant differences in annotation behaviour between annotators of different genders for four different tasks. + +Predicting Annotators' Decisions on Text Different from analyses of annotation behaviour, a recent line of research attempts to learn models based on individual annotations (Plank et al., 2014; Jamison and Gurevych, 2015; Akhtar et al., 2020; Fornaciari et al., 2021; Cercas Curry et al., 2021). These models are motivated by the concern that aggregating labels into a single "truth" is too simplistic for many tasks (Uma et al., 2021; Basile et al., 2021) and might introduce uneven representation of perspectives (Prabhakaran et al., 2021; Abercrombie et al., 2022). + +A particular way of learning from disaggregated labels are models that predict individual annotator decisions for an example. Our work builds directly + +on such a model, multi-annotator models (Davani et al., 2022), which we describe in more detail separately ( $\S 4$ ). Gordon et al. (2022) present a model which also predicts individual annotations and allows a user to interactively aggregate them based on "a jury" inspired by the US judicial system. Their work is similar to ours in central aspects as they explicitly model annotators' sociodemographics and use the same dataset as we do (Kumar et al., 2021). Different from our work, they frame the task as a regression problem and develop a model based on recommender systems. While they also explore ecological fallacies, they focus on usage risks of their system and countermeasures. In contrast, we consider the issue of the ecological fallacy in modelling annotation behaviour more generally. We compare our findings to their results ( $\S 6$ ). + +# 3 Data + +We use a sample of the Kumar et al. (2021) dataset for our experiments. The full dataset contains 107,620 English comments from Twitter, Reddit, and 4Chan, annotated for toxicity by 17,280 annotators. The annotation process encouraged annotator subjectivity (Röttger et al., 2022) which is a desired feature for modelling annotator behaviour. For each annotator, there is extensive sociodemographic information, collected with a survey. Annotations are given as ratings on a five-point scale which we convert to binary annotations by mapping ratings of 2 to 4 to toxic, and ratings 0 and 1 to non-toxic. + +We randomly sample comments from the dataset until we reach annotations from more than 5,000 annotators. We then add all other annotations by these annotators. This approach maximizes the number of examples while controlling the number of annotators in our sample. + +Our final sample contains 111,780 annotations from 5,002 annotators on 22,360 comments with 20 to 120 annotations per annotator (mean 22.35). Most comments have five annotations. 20 comments have four because we removed any underage annotators before sampling. In total 78,357 annotations $(70.10\%)$ are toxic, and 33,423 annotations $(29.90\%)$ are non-toxic. + +We focus on four sociodemographic attributes: gender, age, education, and sexual orientation. Group sizes vary by attribute. For gender, 2,450 annotators $(48.98\%)$ identify as female, 2,116 $(42.30\%)$ as male, 23 $(0.46\%)$ as non-binary (rest in residual categories, full statistics in A.1). + +# 4 Experiments + +We compare three models. The baseline model is the multi-annotator model by Davani et al. (2022). We use their multi-task variant: For each annotator, there is a separate classification layer trained on annotations from that annotator. All annotator layers share a pre-trained language model used to encode the input. We use RoBERTa (Liu et al., 2019) for this, motivated by computational constraints. The other models in our experiments build on this baseline model. + +For the sociodemographic models, we add group-specific layers based on sociodemographic attributes of the annotators. A single attribute, e.g., age, implies several groups, e.g., ages 25-34, ages 35-44. We add the group-specific layers between the pre-trained model and the annotator layers. Each group of annotators shares a separate group-specific layer. We implement group-specific layers as fully-connected, linear layers, each learning a feature transformation applied for one group of annotators. + +Finally, for the random models, we shuffle the assignment of annotators to groups from the sociodemographic model, retaining the relative group sizes. In other words, the probability of each annotator staying in the same group or being reassigned to another group corresponds to the relative size of each group. This approach keeps the model architecture constant while removing the connection between actual sociodemographic attributes and group assignment. It allows us to distinguish the effects of additional parameters, which group-specific layers add in comparison to the baseline, from the effects of sociodemographic information. + +# 4.1 Evaluation Setup + +We evaluate all models on individual annotations from gender, age, education, and sexual orientation groups. This setup is comparable to the "individual label" evaluations in Davani et al. (2022) and Gordon et al. (2022), but with scores calculated per group of annotators. We measure performance in macro-average $F_{1}$ , to weigh each class equally. + +Cross-Validation As there is no standard split available for our dataset, we perform three iterations of a four-fold cross-validation with different seeds (training details in Appendix A.3). We choose four folds, so that even very small groups have more than a hundred annotations in each test set. Across folds, the numbers of annotations per + +sociodemographic group are similar (see Appendix A.4). We construct test sets that only contain comments unseen by the annotators in the training set. We also ensure that all test sets have similar proportions of toxic or non-toxic comments (assigned by the majority of annotators) to address the class imbalance in the dataset (70.62% toxic, see §3). + +Statistical Significance We test for statistical significance of our results from multiple runs of k-fold cross-validation via replicability analysis (Dror et al., 2017). We report the number of significant folds and the Bonferroni-corrected count (Dror et al., 2018) in Appendix A.2. We compute the p-values for each fold via a paired bootstrap-sampling test with BooStSa (Fornaciari et al., 2022). We set the significance level $\alpha = 0.05$ , draw 1000 bootstrap samples per fold, and use a sample size of $50\%$ of the respective test set. + +Remarks on Groups Annotators from different groups of the same attribute will in most cases not have annotated the same examples. Therefore, comparisons between models are only meaningful within each group. + +The groups modeled via group-specific layers and those in the result tables are always the same. For example, if we report scores for gender groups, then the sociodemographic and randomized models are also based on gender groups. In the following, we focus on a subset of groups, omitting, e.g., "Prefer not to say" (see Appendix A.5). + +# 5 Results + +Table 1 shows the results for gender, age, education, and sexual orientation. A naive majority class baseline that predicts all input to be toxic performs worse than all other models with a large margin (exact results in Appendix A.5). + +Sociodemographics vs. Baseline Across attributes, the average scores of the sociodemographic model and the baseline are similar. The sociodemographic model often has a slightly higher average macro F1 than the baseline, but no statistically significant gains. Where average performance is better by several points, as for homosexual annotators, this gain is offset by a large variance in performance (a consequence of small group sizes). + +Sociodemographics vs. Random We also do not find significant performance differences between sociodemographic group-layer models and the corresponding random group assignment models. For + +most groups, the randomized models achieve the highest average scores, but differences to the sociodemographic model are never statistically significant. + +
GenderBaselineSoc-Dem.Random
Male68.00±0.4967.66±0.4667.63±0.53
Female62.23±0.5362.25±1.1962.41±0.92
Nonbinary56.33±6.0056.80±7.2458.00±7.49
+ +
AgeBaselineSoc-Dem.Random
18 - 2459.39±1.5860.44±1.0560.52±1.37
25 - 3466.72±0.5666.63±0.8366.92±0.51
35 - 4464.50±0.5964.94±1.3365.24±0.89
45 - 5465.68±0.6665.88±1.3965.98±0.83
55 - 6464.37±1.2264.94±1.6664.84±1.30
65 or older63.34±2.0764.70±2.2162.77±2.39
+ +
EducationBaselineSoc-Dem.Random
Associate degree60.69±1.4460.54±2.3560.78±1.62
Bachelor's degree66.16±0.5166.23±0.8266.80±0.54
Doctoral degree61.93±3.8263.79±5.0363.27±3.67
High school60.53±1.3960.47±2.2260.55±1.87
Below high school58.28±4.6862.12±4.9060.17±4.25
Master's degree69.71±0.8669.58±0.9369.45±0.96
Professional degree66.75±2.3767.84±3.3268.62±2.84
College, no degree58.65±1.1959.40±1.7959.99±2.19
+ +
SexualityBaselineSoc-Dem.Random
Bisexual71.83±1.1471.42±1.5169.46±1.95
Heterosexual63.25±0.3963.32±1.2163.82±0.55
Homosexual64.43±1.7566.11±2.2065.12±1.94
+ +Table 1: Average and standard deviation of macro $F_{1}$ from three runs of four-fold stratified cross-validation. Separate table for each attribute. Bold results are the highest averages per group. However, no difference is statistically significant (see Appendix A.2) + +# 6 Discussion + +We do not find strong evidence that explicitly modelling sociodemographics helps to predict annotation behaviour with multi-annotator models. These results might seem counter-intuitive, given the evidence of systematic annotation differences between sociodemographic groups (see §2). This discrepancy, however, echoes the issue highlighted by ecological fallacies (Robinson, 1950): Not every annotator will be a perfect representative of their group, so we will not necessarily learn additional information based on their group identity. This seems especially true if we already have access to individual behaviour (i.e., individual annotations). + +In contrast to Davani et al. (2022), we made sociodemographic information explicit in our experiments, as one of the factors influencing annotation + +behaviour. Group-specific layers can be seen as an inductive bias putting emphasis on the sociodemographic relations between annotators. However, there are potentially many other factors influencing annotation behaviour (e.g., attitudes, moral values, cognitive biases, psychological traits). In light of our results, it seems plausible that multi-annotator models learn about these factors implicitly as part of predicting individual behaviour, so that making one factor explicit does not change prediction quality, at least in the case of sociodemographics. + +Still, we also know that generally group attributes can help predict individual decisions, i.e., as base rates or priors. To avoid ecological fallacies in modelling annotation, we therefore need to better understand when and how modelling sociodemographic information is useful in predicting an individual annotator's decisions. For example, we have only evaluated group-specific layers for single attributes. In contrast, social scientists have long adopted the idea of intersectionality (Crenshaw, 1989), which also informs research on fairness in machine learning (Wang et al., 2022). Intersectionality means that the effect of interactions between sociodemographic attributes enables specific experiences that are not captured by the attributes in isolation. For example, identifying as a man means something different depending on the person's education. Groups derived from single attributes might simply be too coarse to improve classifiers learnt from individual labels, as in multi-annotator models. + +The dataset we use (Kumar et al., 2021) has many characteristics which are ideal for our study (see §3). However, it uses a broad notion of toxicity, in contrast to other studies of toxic language (Larimore et al., 2021; Sap et al., 2022), which match content and analysed groups. When modeling the groups frequently referenced in the datasets themselves, we would expect greater benefits from group-specific layers. Similar to us, Biester et al. (2022) who do not find significant differences between annotators of different genders, do so in a more general setting. + +We can only partially compare to Gordon et al. (2022), despite using the same dataset. In addition to differences in approach (see §2), our and their work also differ in their research questions and thus experimental conditions. Gordon et al. (2022) compare their full model (group and individual) against using group information alone. + +We compare our full model (group and individual) against using individual information alone. So it is unclear if their model would benefit from group information in comparison to individual-level information alone. While they find an improvement from group information it is only in comparison to a baseline predicting not individual but aggregated labels. Additionally, the composition of test sets sampled from the full dataset differs between the studies: Gordon et al. (2022) use a test set of 5,000 comments, while we use 22,360 comments in a four-fold cross-validation. We leave an explicit comparison to future work. + +Group-specific layers (§4) are a natural extension of annotator-specific classification layers in multi-annotator models. However, other architectures to predict annotator-level labels use different ways to represent sociodemographic information, e.g., via embeddings in a recommender system (Gordon et al., 2022). Future work could explore additional representations of annotator attributes (e.g., as part of the input, either textual or as separate features) and other approaches to modelling the relation of individual labeling decisions and attributes (e.g., probabilistic graphical models). + +# 7 Conclusion + +We ask how relevant modelling explicit sociodemographic information is in learning from individual annotators. Our experiments with group-specific layers for four sociodemographic attributes on social media data with toxicity annotations (Kumar et al., 2021) show no significant benefit of modelling sociodemographic groups in multi-annotator models. However, as the issue of ecological fallacies highlights, it is not implausible that these models do not learn additional information from group information beyond the inherent variation. However, our results do not refute the usefulness of sociodemographic attributes in modelling annotation, but underscore the importance of their judicious use. Different tasks and model architectures will likely benefit to different extents. Ultimately, annotation behaviour is driven by complex factors and we will need to consider more than annotators' sociodemographics. + +# Acknowledgements + +We thank Deepak Kumar for providing access to the disaggregated dataset and his continued support. We also thank Aida Mostafazadeh Davani for providing information on implementation de + +tails of multi-annotator models. Members of MilaNLP (Bocconi) and the Semantic Computing Group (Bielefeld) provided feedback on earlier versions of this paper, for which we thank them again. + +This work has in part been funded by the European Research Council (ERC) under the European Union's Horizon 2020 research and innovation program (grant agreement No. 949944, INTEGRACTOR). Likewise, this work has in part been funded by the VolkswagenStiftung as part of the "3B Bots Building Bridges" project. + +# Limitations + +While the dataset by Kumar et al. (2021) enabled us to test models for a range of often overlooked groups (e.g., non-binary or bisexual annotators), we ultimately modelled only four specific attributes (gender, age, education, sexual orientation). There are likely to be more factors that could play a role. Additionally, annotators in the Kumar et al. (2021) dataset are exclusively from the United States of America, so that results do not necessarily hold for other countries or cultures (Hovy and Yang, 2021). Specifically perceptions of harmful content online are known to vary across countries (Jiang et al., 2021). + +We used only the (Kumar et al., 2021) dataset. This is mainly due to our strict criteria regarding dataset size and availability of annotator-level labels and sociodemographic information. These characteristics were a prerequisite for our experiments across different attributes with sufficient numbers of annotators. Most datasets which include annotator-level labels and sociodemographic information contain much smaller numbers of annotators and attributes. Nevertheless, with the Measuring Hate Speech Corpus there is at least one additional dataset (Sachdeva et al., 2022) with comparable characteristics that could be used in future experiments. Also, additional small-scale, more focused experiments could use datasets like Sap et al. (2022) or HS-Brexit (Akhtar et al., 2021) which was annotated by 6 annotators, each from one of two sociodemographic groups. + +We do not study the aggregation of individual predictions or evaluate against majority labels, as these are not directly relevant to our investigation of sociodemographic attributes in models of annotation behaviour. Consequently, we cannot derive a conclusion about performance in those settings from our results. This is a noteworthy limitation, because part of the experiments introducing + +multi-annotator models in Davani et al. (2022) compare labels aggregated from multi-annotator models against predictions from a standard classifier (directly trained on aggregated labels). + +For computational reasons, our experiments use a comparatively small pre-trained language model (RoBERTa, Liu et al. 2019). Thus, results might differ with larger models. + +# Ethics Statement + +As sociodemographic attributes are sensitive information, we do not infer attributes, but build on a self-reported, IRB-reviewed dataset (Kumar et al., 2021). We also see potential for a discussion of "privacy by design" in modelling human label variation based on our results: There can be circumstances in which knowing more about annotators is not relevant, and indeed might lead to violations of privacy. + +As multi-annotator models attempt to capture the preferences of individual annotators, there are valid concerns around privacy and anonymity. As discussed in Davani et al. (2022), increasing the annotator count can be one option to reduce privacy risks. We show it is feasible to learn a model for a large number of individual annotators (5002 vs. 18 and 82 in their work). But a prerequisite for improved privacy is to apply effective aggregation on top of individual predictions, which we do not study in the present work. + +# References + +Gavin Abercrombie, Valerio Basile, Sara Tonelli, Verena Rieser, and Alexandra Uma, editors. 2022. Proceedings of the 1st Workshop on Perspectivist Approaches to NLP @LREC2022. European Language Resources Association, Marseille, France. +Sohail Akhtar, Valerio Basile, and Viviana Patti. 2020. Modeling annotator perspective and polarized opinions to improve hate speech detection. In Proceedings of the AAAI Conference on Human Computation and Crowdsourcing, volume 8, pages 151-154. +Sohail Akhtar, Valerio Basile, and Viviana Patti. 2021. Whose opinions matter? perspective-aware models to identify opinions of hate speech victims in abusive language detection. Preprint arXiv:2106.15896. +Hala Al Kuwatly, Maximilian Wich, and Georg Groh. 2020. Identifying and measuring annotator bias based on annotators' demographic characteristics. In Proceedings of the Fourth Workshop on Online Abuse and Harms, pages 184-190, Online. Association for Computational Linguistics. + +Valerio Basile, Michael Fell, Tommaso Fornaciari, Dirk Hovy, Silviu Paun, Barbara Plank, Massimo Poesio, and Alexandra Uma. 2021. We need to consider disagreement in evaluation. In Proceedings of the 1st Workshop on Benchmarking: Past, Present and Future, pages 15-21, Online. Association for Computational Linguistics. +Laura Biester, Vanita Sharma, Ashkan Kazemi, Naihao Deng, Steven Wilson, and Rada Mihalcea. 2022. Analyzing the effects of annotator gender across NLP tasks. In Proceedings of the 1st Workshop on Perspectivist Approaches to NLP @LREC2022, pages 10-19, Marseille, France. European Language Resources Association. +Reuben Binns, Michael Veale, Max Van Kleek, and Nigel Shadbolt. 2017. Like trainer, like bot? inheritance of bias in algorithmic content moderation. In Social Informatics, Lecture Notes in Computer Science, pages 405-415. Springer International Publishing. +Amanda Cercas Curry, Gavin Abercrombie, and Verena Rieser. 2021. ConvAbuse: Data, analysis, and benchmarks for nuanced abuse detection in conversational AI. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7388-7403, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics. +Kimberle Crenshaw. 1989. Demarginalizing the intersection of race and sex: A black feminist critique of antidiscrimination doctrine, feminist theory and antiracist politics. University of Chicago Legal Forum, 1989(1):Article 8. +Aida Mostafazadeh Davani, Mark Diaz, and Vinodkumar Prabhakaran. 2022. Dealing with disagreements: Looking beyond the majority vote in subjective annotations. Transactions of the Association for Computational Linguistics, 10:92-110. +Rotem Dror, Gili Baumer, Marina Bogomolov, and Roi Reichart. 2017. Replicability analysis for natural language processing: Testing significance with multiple datasets. Transactions of the Association for Computational Linguistics, 5:471-486. +Rotem Dror, Gili Baumer, Segev Shlomov, and Roi Reichart. 2018. The hitchhiker's guide to testing statistical significance in natural language processing. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1383-1392, Melbourne, Australia. Association for Computational Linguistics. +Elizabeth Excell and Noura Al Moubayed. 2021. Towards equal gender representation in the annotations of toxic language detection. In Proceedings of the 3rd Workshop on Gender Bias in Natural Language Processing, pages 55-65, Online. Association for Computational Linguistics. + +Tommaso Fornaciari, Alexandra Uma, Silviu Paun, Barbara Plank, Dirk Hovy, and Massimo Poesio. 2021. Beyond black & white: Leveraging annotator disagreement via soft-label multi-task learning. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2591-2597, Online. Association for Computational Linguistics. +Tommaso Fornaciari, Alexandra Uma, Massimo Poesio, and Dirk Hovy. 2022. Hard and soft evaluation of NLP models with BOOtSTrap SAmpling - BooStSa. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 127-134, Dublin, Ireland. Association for Computational Linguistics. +David A. Freedman. 2015. Ecological inference. In James D. Wright, editor, International Encyclopedia of the Social & Behavioral Sciences (Second Edition), pages 868-870. Elsevier. +Mitchell L. Gordon, Michelle S. Lam, Joon Sung Park, Kayur Patel, Jeff Hancock, Tatsunori Hashimoto, and Michael S. Bernstein. 2022. Jury learning: Integrating dissenting voices into machine learning models. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems, CHI '22, pages 1-19. Association for Computing Machinery. +Nitesh Goyal, Ian D. Kivlichan, Rachel Rosen, and Lucy Vasserman. 2022. Is your toxicity my toxicity? exploring the impact of rater identity on toxicity annotation. Proceedings of the ACM on Human-Computer Interaction, 6:1-28. +Dirk Hovy and Diyi Yang. 2021. The importance of modeling social factors of language: Theory and practice. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 588-602, Online. Association for Computational Linguistics. +Emily Jamison and Iryna Gurevych. 2015. Noise or additional information? leveraging crowdsourced annotation item agreement for natural language tasks. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 291-297, Lisbon, Portugal. Association for Computational Linguistics. +Jialun Aaron Jiang, Morgan Klaus Scheuerman, Casey Fiesler, and Jed R. Brubaker. 2021. Understanding international perceptions of the severity of harmful content online. PLOS ONE, 16(8). +Deepak Kumar, Patrick Gage Kelley, Sunny Consolvo, Joshua Mason, Elie Bursztein, Zakir Durmeric, Kurt Thomas, and Michael Bailey. 2021. Designing toxic content classification for a diversity of perspectives. In Seventeenth Symposium on Usable Privacy and Security (SOUPS 2021), pages 299-318. USENIX Association. + +Savannah Larimore, Ian Kennedy, Breon Haskett, and Alina Arseniev-Koehler. 2021. Reconsidering annotator disagreement about racist language: Noise or signal? In Proceedings of the Ninth International Workshop on Natural Language Processing for Social Media, pages 81-90, Online. Association for Computational Linguistics. +Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. Preprint arXiv:1907.11692. +F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. 2011. Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12:2825-2830. +Barbara Plank. 2022. The "problem" of human label variation: On ground truth in data, modeling and evaluation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 10671-10682, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics. +Barbara Plank, Dirk Hovy, and Anders Søgaard. 2014. Learning part-of-speech taggers with inter-annotator agreement loss. In Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 742–751, Gothenburg, Sweden. Association for Computational Linguistics. +Vinodkumar Prabhakaran, Aida Mostafazadeh Davani, and Mark Diaz. 2021. On releasing annotator-level labels and information in datasets. In Proceedings of the Joint 15th Linguistic Annotation Workshop (LAW) and 3rd Designing Meaning Representations (DMR) Workshop, pages 133-138, Punta Cana, Dominican Republic. Association for Computational Linguistics. +W. S. Robinson. 1950. Ecological correlations and the behavior of individuals. American Sociological Review, 15(3):351-357. +Paul Röttger, Bertie Vidgen, Dirk Hovy, and Janet Pierrehumbert. 2022. Two contrasting data annotation paradigms for subjective NLP tasks. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 175-190, Seattle, United States. Association for Computational Linguistics. +Pratik Sachdeva, Renata Barreto, Geoff Bacon, Alexander Sahn, Claudia von Vacano, and Chris Kennedy. 2022. The measuring hate speech corpus: Leveraging rasch measurement theory for data perspectivism. In Proceedings of the 1st Workshop on Perspectivist Approaches to NLP @LREC2022, pages 83-94, Marseille, France. European Language Resources Association. + +Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. 2019. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. In 5th Workshop on Energy Efficient Machine Learning and Cognitive Computing @ NeurIPS 2019. + +Maarten Sap, Dallas Card, Saadia Gabriel, Yejin Choi, and Noah A. Smith. 2019. The risk of racial bias in hate speech detection. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1668-1678, Florence, Italy. Association for Computational Linguistics. + +Maarten Sap, Swabha Swayamdipta, Laura Vianna, Xuhui Zhou, Yejin Choi, and Noah A. Smith. 2022. Annotators with attitudes: How annotator beliefs and identities bias toxic language detection. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5884-5906, Seattle, United States. Association for Computational Linguistics. + +Qinlan Shen and Carolyn Rose. 2021. What sounds "right" to me? experiential factors in the perception of political ideology. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 1762-1771, Online. Association for Computational Linguistics. + +Alexandra N. Uma, Tommaso Fornaciari, Dirk Hovy, Silviu Paun, Barbara Plank, and Massimo Poesio. 2021. Learning from disagreement: A survey. Journal of Artificial Intelligence Research, 72:1385-1470. + +Angelina Wang, Vikram V Ramaswamy, and Olga Russakovsky. 2022. Towards intersectionality in machine learning: Including more identities, handling underrepresentation, and performing evaluation. In 2022 ACM Conference on Fairness, Accountability, and Transparency, FAccT '22, pages 336-349. Association for Computing Machinery. + +Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online. Association for Computational Linguistics. + +# A Appendix + +# A.1 Annotator Sociodemographics in Sample + +Table 2 shows how many annotators the sample contains. Counts are given per group of the four attributes gender, age, education and sexuality. + +In the Kumar et al. (2021) dataset, sociodemographic attributes are given for each individual annotation - not once per annotator. For some annotators, conflicting attribute values exist (e.g., two different age groups). As the data collection spanned several months (Kumar et al., 2021), these value changes can in principle be reasonable (e.g., because an annotator got older, finished a degree, changed sexual preference or gender identity). However, as reasonable changes can not easily be discerned from erroneous input, we disambiguate values based on a heuristic: If an annotator reports several values for an attribute, we assume the most frequent value to be valid. In cases of no clear most frequent value, we set the attribute to "Prefer not to say". Thus, the main results do not contain annotators with ambiguous attributes. + +# A.2 Significance Tests + +Results of a replicability analysis (Dror et al., 2017) testing for significant differences in macro $F_{1}$ on scores from three runs of four-fold cross-validation. Table 3 shows results for a comparison of the sociodemographic models against the baseline models. Table 4 shows results for a comparison of the sociodemographic models against the randomized assignment models. The Bonferroni correction for the corrected count of significant folds $\hat{k}_{Bonferroni}$ is used to account for the fact that we have overlapping test sets from multiple runs of four-fold cross-validation. + +# A.3 Training Details, Hyperparameters and Computational Resources + +We implement models and the training loop using the Hugging Face Transformers library (version 4.19.2, Wolf et al. 2020). Maximum sequence length is 512 tokens, with truncation and padding to the maximum length. We train for 3 epochs with a batch size of 8 and an initial learning rate of 0.00001. Otherwise, we used default parameters. We found results to particularly depend on the learning rate, with higher or lower values leading to worse results. + +We use a weighted loss function. Label weights are calculated per annotator on the training set of each fold. Label weights, evaluation scores and the four-fold dataset splits (StratifiedKFold) are calculated using the scikit-learn library (version 1.0.2, Pedregosa et al. 2011). The folds are based on a fixed random seed per iteration: 2803636207, 165043843, 2923262358 + +
Number of Annotators
Gender
Female2450
Male2116
Prefer not to say412
Nonbinary23
Other1
+ +
Number of Annotators
Age
18 - 24489
25 - 341861
35 - 441115
45 - 54529
55 - 64321
65 or older119
Prefer not to say568
+ +
Number of Annotators
Sexuality
Heterosexual4018
Bisexual469
Prefer not to say346
Homosexual134
Other35
+ +
EducationNumber of Annotators
Bachelor's degree1879
College, no degree861
Prefer not to say647
Master's degree642
Associate degree460
High school363
Professional degree68
Doctoral degree51
Below high school25
Other6
+ +The majority of parameters in our model belong to the pre-trained language model shared between all group-specific and annotator-specific layers. Specifically, RoBERTa (Liu et al., 2019) in the roberta-base variant has 125 Million parameters. We keep the pre-trained model's default output dimensionality of 768, so that each group-specific layer adds $768 * 768 + 768 = 590$ , 592 parameters and each annotator layer adds $768 * 2 + 2 = 1$ , 538 parameters. + +All experiments ran on a single GPU (GeForce GTX 1080 Ti, 12GB GPU RAM). Per fold, training and evaluation together take about three and a half hours in our setting. Three runs of four-fold cross-validation (12 folds), thus take around 42 hours + +Table 2: Number of annotators per group for attributes gender, age, sexuality and education. Counts refer to the entire sample + +
hatkcounthatkBonf.
Female20
Male00
Nonbinary10
+ +
hatkcounthatkBonf.
18 - 2420
25 - 3420
35 - 4410
45 - 5400
55 - 6410
65 or older10
+ +
ˆkcountˆkBonf.
Bisexual20
Heterosexual42
Homosexual10
+ +
kcountkBonf.
Associate degree00
Bachelor's degree10
Doctoral degree20
High school00
Belowhigh school00
Master's degree00
Professional degree00
College, no degree22
+ +Table 3: Results of a replicability analysis of baseline vs sociodemographic models. Raw and Bonferroni-corrected counts of significant folds out of 12 folds from three runs of four-fold cross-validation. P-values for each fold are computed via a paired bootstrap test with significance level $\alpha = 0.05$ , 1000 bootstrap samples per fold and a sample size of $50\%$ of the respective test set. + +(1.75 days). With four attributes and three trainable models the combined run time of the reported experiments is estimated to be 21 days. Including preliminary experiments, which, however, mostly were not full runs of k-fold cross-validation and also utilized DistilBERT (Sanh et al., 2019) with slightly faster run times, it will be many times more. There is no discernible difference in experiment run times between multi-annotator models with or without groups or different numbers of groups. + +# A.4 Number of Annotations per Group across all Test Sets + +Table 5 contains the number of annotations we have per group across the total of 12 folds (from three runs of four-fold cross-validation). This number of annotations is the effective test set size per group. As the numbers do not vary substantially, perfor + +
kcountkBonf.
Female22
Male10
Nonbinary10
+ +
kcountkBonf.
18 - 2410
25 - 3400
35 - 4410
45 - 5410
55 - 6430
65 or older10
+ +
hatcounthatBonf.
Bisexual62
Heterosexual11
Homosexual00
+ +
kcountkBonf.
Associate degree20
Bachelor's degree10
Doctoral degree00
High school20
Belowhigh school20
Master's degree00
Professional degree00
College, no degree11
+ +mance on each fold is equally representative for all groups. + +# A.5 Full Results + +Table 6 shows full results of experiments (see 4), including results for all residual categories and a naive baseline which always predicts toxic. + +Table 4: Results from replicability analysis of randomized vs sociodemographic models. Raw and Bonferroni-corrected counts of significant folds out of 12 folds from three runs of four-fold cross-validation. P-values for each fold are computed via a paired bootstrap test with significance level $\alpha = 0.05$ , 1000 bootstrap samples per fold and a sample size of $50\%$ of the respective test set. + +
GenderNumber Of AnnotationsMinMax
Female13555±86.4413383.013664.0
Male11925±61.6511843.012062.0
Nonbinary115±6.03104.0122.0
Other5±1.952.08.0
Prefer not to say2345±51.192281.02453.0
+ +
AgeNumber Of AnnotationsMinMax
18 - 242615±50.8825212697
25 - 3410315±61.451024410457
35 - 446250±51.0661796324
45 - 543025±47.2329293083
55 - 641865±25.4818311903
65 or older675±19.31643704
Prefer not to say3200±55.2831313289
+ +
SexualityNumber Of AnnotationsMinMax
Bisexual2445±39.2623832501
Heterosexual22630±63.002250722726
Homosexual725±26.57670759
Other190±7.91173201
Prefer not to say1955±35.3918782009
+ +
EducationNumber Of AnnotationsMinMax
Associate professor2605±47.5925162697
Bachelor's degree10510±84.791034810700
Doctoral degree305±18.83270332
High school2080±37.0120152139
Below high school165±11.17144184
Master's degree3515±48.0834253580
Other30±3.442536
Prefer not to say3690±52.9236033808
Professional degree380±17.87352411
College, no degree4665±71.3645394776
+ +Table 5: Average, standard deviation, minimum and maximum of number of annotations per fold. All information given per group of gender, age, education and sexuality. Statistics are calculated across 12 folds from three runs of four-fold cross-validation. + +
GenderMajority BaselineBaselineSoc-Dem.Random
Female41.79±0.1262.23±0.5362.25±1.1962.41±0.92
Male40.53±0.1168.00±0.4967.66±0.4667.63±0.53
Nonbinary44.69±1.3956.33±6.0056.80±7.2458.00±7.49
Other45.50±4.6948.56±10.7850.53±14.6343.66±7.25
Prefer not to say41.05±0.3664.54±1.1365.05±1.5265.08±1.86
+ +
AgeMajority BaselineBaselineSoc-Dem.Random
18 - 2442.49±0.2859.39±1.5860.44±1.0560.52±1.37
25 - 3440.49±0.0966.72±0.5666.63±0.8366.92±0.51
35 - 4441.87±0.1564.50±0.5964.94±1.3365.24±0.89
45 - 5440.63±0.2665.68±0.6665.88±1.3965.98±0.83
55 - 6441.65±0.3964.37±1.2264.94±1.6664.84±1.30
65 or older41.46±0.5463.34±2.0764.70±2.2162.77±2.39
Prefer not to say41.37±0.3263.99±1.3265.24±1.1864.73±1.33
+ +
EducationMajority BaselineBaselineSoc-Dem.Random
Associate degree43.16±0.1960.69±1.4460.54±2.3560.78±1.62
Bachelor's degree40.38±0.1066.16±0.5166.23±0.8266.80±0.54
Doctoral degree43.34±0.9461.93±3.8263.79±5.0363.27±3.67
High school43.02±0.2660.53±1.3960.47±2.2260.55±1.87
Below high school43.10±1.4458.28±4.6862.12±4.9060.17±4.25
Master's degree37.55±0.3269.71±0.8669.58±0.9369.45±0.96
Other42.95±2.3156.56±10.8857.59±9.8657.71±12.28
Prefer not to say40.97±0.2765.07±1.1665.69±1.0565.74±1.09
Professional degree40.43±0.8066.75±2.3767.84±3.3268.62±2.84
College, no degree43.61±0.1858.65±1.1959.40±1.7959.99±2.19
+ +
SexualityMajority BaselineBaselineSoc-Dem.Random
Bisexual34.69±0.5071.83±1.1471.42±1.5169.46±1.95
Heterosexual41.99±0.0663.25±0.3963.32±1.2163.82±0.55
Homosexual41.15±0.4164.43±1.7566.11±2.2065.12±1.94
Other43.53±0.7857.55±3.7960.57±4.5158.69±4.72
Prefer not to say39.12±0.2467.80±1.5667.27±1.5267.46±1.11
+ +Table 6: Average and standard deviation of macro $F_{1}$ from three runs of four-fold stratified cross-validation. Separate table for each attribute. Bold results are the highest average per group. Full results including naive majority baseline + +A For every submission: + +A1. Did you describe the limitations of your work? Limitations, 8 +A2. Did you discuss any potential risks of your work? Ethics Statement, 9 +A3. Do the abstract and introduction summarize the paper's main claims? +A4. Have you used AI writing assistants when working on this paper? Left blank. + +B Did you use or create scientific artifacts? + +3, Appendix A.3 + +B1. Did you cite the creators of artifacts you used? 3, Appendix A.3 +B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Clear from context, citations +B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Clear from context, citations +B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? 3, Ethics Statement 9 +B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? 3, Appendix A.1 +B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. 3, 4, Appendix A.4 + +C Did you run computational experiments? + +4 + +C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Appendix A.3 + +C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? No response. +C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? No response. +C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Appendix A.3 + +# D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank. + +D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response. +D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response. +D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response. +D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response. +D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response. \ No newline at end of file diff --git a/2023/The Ecological Fallacy in Annotation_ Modeling Human Label Variation goes beyond Sociodemographics/images.zip b/2023/The Ecological Fallacy in Annotation_ Modeling Human Label Variation goes beyond Sociodemographics/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f9f0e9736dd29aa2ddd7a963c0199a18affa9bc1 --- /dev/null +++ b/2023/The Ecological Fallacy in Annotation_ Modeling Human Label Variation goes beyond Sociodemographics/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:36361d1863fc30edf97fa77ee0992b4c43ba88231866e9cd131b0c485a59b347 +size 593034 diff --git a/2023/The Ecological Fallacy in Annotation_ Modeling Human Label Variation goes beyond Sociodemographics/layout.json b/2023/The Ecological Fallacy in Annotation_ Modeling Human Label Variation goes beyond Sociodemographics/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c60cc53e5f058a796ccbe3e69898f1b3671ab112 --- /dev/null +++ b/2023/The Ecological Fallacy in Annotation_ Modeling Human Label Variation goes beyond Sociodemographics/layout.json @@ -0,0 +1,7516 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 89, + 76, + 505, + 110 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 76, + 505, + 110 + ], + "spans": [ + { + "bbox": [ + 89, + 76, + 505, + 110 + ], + "type": "text", + "content": "The Ecological Fallacy in Annotation: Modelling Human Label Variation goes beyond Sociodemographics" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 113, + 121, + 483, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 121, + 483, + 137 + ], + "spans": [ + { + "bbox": [ + 113, + 121, + 483, + 137 + ], + "type": "text", + "content": "Matthias Orlikowski1, Paul Röttger2, Philipp Cimiano1, and Dirk Hovy3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 248, + 147, + 349, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 147, + 349, + 161 + ], + "spans": [ + { + "bbox": [ + 248, + 147, + 349, + 161 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 248, + 147, + 349, + 161 + ], + "type": "text", + "content": "Bielefeld University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 245, + 161, + 351, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 245, + 161, + 351, + 174 + ], + "spans": [ + { + "bbox": [ + 245, + 161, + 351, + 174 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 245, + 161, + 351, + 174 + ], + "type": "text", + "content": "University of Oxford" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 136, + 174, + 460, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 174, + 460, + 189 + ], + "spans": [ + { + "bbox": [ + 136, + 174, + 460, + 189 + ], + "type": "text", + "content": "3Computing Sciences Department, Bocconi University, Milan, Italy" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 84, + 236, + 274, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 236, + 274, + 499 + ], + "spans": [ + { + "bbox": [ + 84, + 236, + 274, + 499 + ], + "type": "text", + "content": "Many NLP tasks exhibit human label variation, where different annotators give different labels to the same texts. This variation is known to depend, at least in part, on the sociodemographics of annotators. Recent research aims to model individual annotator behaviour rather than predicting aggregated labels, and we would expect that sociodemographic information is useful for these models. On the other hand, the ecological fallacy states that aggregate group behaviour, such as the behaviour of the average female annotator, does not necessarily explain individual behaviour. To account for sociodemographics in models of individual annotator behaviour, we introduce group-specific layers to multi-annotator models. In a series of experiments for toxic content detection, we find that explicitly accounting for sociodemographic attributes in this way does not significantly improve model performance. This result shows that individual annotation behaviour depends on much more than just sociodemographics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 68, + 510, + 154, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 510, + 154, + 523 + ], + "spans": [ + { + "bbox": [ + 68, + 510, + 154, + 523 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 529, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 529, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 529, + 291, + 772 + ], + "type": "text", + "content": "Different annotators will not necessarily assign the same labels to the same texts, resulting in human label variation (Plank, 2022). Previous work finds that this variation depends at least in part on the sociodemographics of annotators, such as their age and gender (Binns et al., 2017; Al Kuwatly et al., 2020; Excell and Al Moubayed, 2021; Shen and Rose, 2021). These results are particularly pronounced for subjective tasks like toxic content detection (Sap et al., 2019; Kumar et al., 2021; Sap et al., 2022; Goyal et al., 2022). Since human label variation is relevant to a wide range of NLP tasks, recent research has begun to model individual annotator behaviour, rather than predicting aggregated labels (Davani et al., 2022; Gordon et al., 2022). In this setting, we would expect sociodemographic attributes to help explain annotator decisions. Therefore, we investigate whether explicitly" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 212, + 522, + 307 + ], + "blocks": [ + { + "bbox": [ + 307, + 212, + 522, + 307 + ], + "lines": [ + { + "bbox": [ + 307, + 212, + 522, + 307 + ], + "spans": [ + { + "bbox": [ + 307, + 212, + 522, + 307 + ], + "type": "image", + "image_path": "0261e22efde8dee493e93beebcd49541625a911d2b7967e05ba5f1df0447ab78.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 317, + 525, + 340 + ], + "lines": [ + { + "bbox": [ + 302, + 317, + 525, + 340 + ], + "spans": [ + { + "bbox": [ + 302, + 317, + 525, + 340 + ], + "type": "text", + "content": "Figure 1: Group-specific layers representing annotator sociodemographics in multi-annotator models." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 364, + 525, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 364, + 525, + 402 + ], + "spans": [ + { + "bbox": [ + 302, + 364, + 525, + 402 + ], + "type": "text", + "content": "accounting for the sociodemographic attributes of annotators leads to better predictions of their annotation behaviour1." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 405, + 526, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 405, + 526, + 593 + ], + "spans": [ + { + "bbox": [ + 302, + 405, + 526, + 593 + ], + "type": "text", + "content": "There is a risk of misreading these efforts as an example of the ecological fallacy: aggregate group behaviour does not necessarily explain individual behaviour (Robinson, 1950; Freedman, 2015). For example, while on average, white annotators may be more likely to label African-American Vernacular English as toxic (Sap et al., 2019), that does not mean it is true for every white annotator individually. However, we aim at exactly this distinction to discuss the relevance of sociodemographic groups in models of individual annotator behaviour. Likewise, we do not assume prior work to commit ecological fallacies, even if a less-nuanced read might suggest it." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 595, + 525, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 595, + 525, + 730 + ], + "spans": [ + { + "bbox": [ + 302, + 595, + 525, + 730 + ], + "type": "text", + "content": "Davani et al. (2022) introduce a simple multi-annotator model, where each annotator is modelled with a separate classification head. We expand their model with group-specific layers, which are activated for each annotator based on their sociodemographic attributes. We compare the two model setups to a control setup where we randomise group assignments. All comparisons use annotator-level toxicity data from Kumar et al. (2021). We find that find that explicitly accounting for sociodem" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 740, + 525, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 740, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 740, + 525, + 772 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 302, + 740, + 525, + 772 + ], + "type": "text", + "content": "Code to run our experiments and analyses is available at https://github.com/morlikowski/ecological-fallacy" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "1017" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "spans": [ + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "type": "text", + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 219, + 806, + 374, + 817 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 806, + 374, + 817 + ], + "spans": [ + { + "bbox": [ + 219, + 806, + 374, + 817 + ], + "type": "text", + "content": "Volume 2: Short Papers, pages 1017-1029" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "spans": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "text", + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 290, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 290, + 139 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 290, + 139 + ], + "type": "text", + "content": "graphic attributes does not significantly improve model performance. This result suggests that human label variation happens at a more individual level than sociodemographics, and that annotator decisions are even more complex." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 144, + 290, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 144, + 290, + 253 + ], + "spans": [ + { + "bbox": [ + 67, + 144, + 290, + 253 + ], + "type": "text", + "content": "Contributions 1) We introduce group-specific layers to model groups of annotators with shared attributes in multi-annotator models. 2) We evaluate the effect of group-specific layers for toxic content detection, and show that explicitly accounting for sociodemographic attributes does not significantly improve performance, thus highlighting the risk of the ecological fallacy in annotator modelling." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 253, + 291, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 253, + 291, + 295 + ], + "spans": [ + { + "bbox": [ + 67, + 253, + 291, + 295 + ], + "type": "text", + "content": "As a corollary, we show that multi-annotator models can be applied to many times more annotators than in prior work." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 301, + 161, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 301, + 161, + 313 + ], + "spans": [ + { + "bbox": [ + 67, + 301, + 161, + 313 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 320, + 290, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 320, + 290, + 564 + ], + "spans": [ + { + "bbox": [ + 67, + 320, + 290, + 564 + ], + "type": "text", + "content": "Sociodemographics in Annotation Behaviour A growing body of research studies how annotator sociodemographics relate to their annotation decisions, for tasks ranging from natural language inference (Biester et al., 2022) to the detection of racist (Larimore et al., 2021) or generally toxic (Sap et al., 2022) language. Goyal et al. (2022), for example, find that annotators from certain sociodemographic groups (e.g., LGBTQ people) tend to find content attacking their own groups (e.g., homophobic content) to be more toxic. This motivates our research into explicitly accounting for sociodemographics to model annotation behaviour. However, the link between sociodemographics and behaviour is not uncontested. Biester et al. (2022), for example, do not find significant differences in annotation behaviour between annotators of different genders for four different tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 570, + 291, + 731 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 570, + 291, + 731 + ], + "spans": [ + { + "bbox": [ + 67, + 570, + 291, + 731 + ], + "type": "text", + "content": "Predicting Annotators' Decisions on Text Different from analyses of annotation behaviour, a recent line of research attempts to learn models based on individual annotations (Plank et al., 2014; Jamison and Gurevych, 2015; Akhtar et al., 2020; Fornaciari et al., 2021; Cercas Curry et al., 2021). These models are motivated by the concern that aggregating labels into a single \"truth\" is too simplistic for many tasks (Uma et al., 2021; Basile et al., 2021) and might introduce uneven representation of perspectives (Prabhakaran et al., 2021; Abercrombie et al., 2022)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 733, + 290, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 733, + 290, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 733, + 290, + 773 + ], + "type": "text", + "content": "A particular way of learning from disaggregated labels are models that predict individual annotator decisions for an example. Our work builds directly" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 71, + 526, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 301 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 301 + ], + "type": "text", + "content": "on such a model, multi-annotator models (Davani et al., 2022), which we describe in more detail separately (" + }, + { + "bbox": [ + 302, + 71, + 526, + 301 + ], + "type": "inline_equation", + "content": "\\S 4" + }, + { + "bbox": [ + 302, + 71, + 526, + 301 + ], + "type": "text", + "content": "). Gordon et al. (2022) present a model which also predicts individual annotations and allows a user to interactively aggregate them based on \"a jury\" inspired by the US judicial system. Their work is similar to ours in central aspects as they explicitly model annotators' sociodemographics and use the same dataset as we do (Kumar et al., 2021). Different from our work, they frame the task as a regression problem and develop a model based on recommender systems. While they also explore ecological fallacies, they focus on usage risks of their system and countermeasures. In contrast, we consider the issue of the ecological fallacy in modelling annotation behaviour more generally. We compare our findings to their results (" + }, + { + "bbox": [ + 302, + 71, + 526, + 301 + ], + "type": "inline_equation", + "content": "\\S 6" + }, + { + "bbox": [ + 302, + 71, + 526, + 301 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 303, + 306, + 350, + 318 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 306, + 350, + 318 + ], + "spans": [ + { + "bbox": [ + 303, + 306, + 350, + 318 + ], + "type": "text", + "content": "3 Data" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 326, + 526, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 326, + 526, + 501 + ], + "spans": [ + { + "bbox": [ + 302, + 326, + 526, + 501 + ], + "type": "text", + "content": "We use a sample of the Kumar et al. (2021) dataset for our experiments. The full dataset contains 107,620 English comments from Twitter, Reddit, and 4Chan, annotated for toxicity by 17,280 annotators. The annotation process encouraged annotator subjectivity (Röttger et al., 2022) which is a desired feature for modelling annotator behaviour. For each annotator, there is extensive sociodemographic information, collected with a survey. Annotations are given as ratings on a five-point scale which we convert to binary annotations by mapping ratings of 2 to 4 to toxic, and ratings 0 and 1 to non-toxic." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 502, + 525, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 502, + 525, + 582 + ], + "spans": [ + { + "bbox": [ + 302, + 502, + 525, + 582 + ], + "type": "text", + "content": "We randomly sample comments from the dataset until we reach annotations from more than 5,000 annotators. We then add all other annotations by these annotators. This approach maximizes the number of examples while controlling the number of annotators in our sample." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 583, + 526, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 583, + 526, + 691 + ], + "spans": [ + { + "bbox": [ + 302, + 583, + 526, + 691 + ], + "type": "text", + "content": "Our final sample contains 111,780 annotations from 5,002 annotators on 22,360 comments with 20 to 120 annotations per annotator (mean 22.35). Most comments have five annotations. 20 comments have four because we removed any underage annotators before sampling. In total 78,357 annotations " + }, + { + "bbox": [ + 302, + 583, + 526, + 691 + ], + "type": "inline_equation", + "content": "(70.10\\%)" + }, + { + "bbox": [ + 302, + 583, + 526, + 691 + ], + "type": "text", + "content": " are toxic, and 33,423 annotations " + }, + { + "bbox": [ + 302, + 583, + 526, + 691 + ], + "type": "inline_equation", + "content": "(29.90\\%)" + }, + { + "bbox": [ + 302, + 583, + 526, + 691 + ], + "type": "text", + "content": " are non-toxic." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 692, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 692, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 692, + 525, + 772 + ], + "type": "text", + "content": "We focus on four sociodemographic attributes: gender, age, education, and sexual orientation. Group sizes vary by attribute. For gender, 2,450 annotators " + }, + { + "bbox": [ + 302, + 692, + 525, + 772 + ], + "type": "inline_equation", + "content": "(48.98\\%)" + }, + { + "bbox": [ + 302, + 692, + 525, + 772 + ], + "type": "text", + "content": " identify as female, 2,116 " + }, + { + "bbox": [ + 302, + 692, + 525, + 772 + ], + "type": "inline_equation", + "content": "(42.30\\%)" + }, + { + "bbox": [ + 302, + 692, + 525, + 772 + ], + "type": "text", + "content": " as male, 23 " + }, + { + "bbox": [ + 302, + 692, + 525, + 772 + ], + "type": "inline_equation", + "content": "(0.46\\%)" + }, + { + "bbox": [ + 302, + 692, + 525, + 772 + ], + "type": "text", + "content": " as non-binary (rest in residual categories, full statistics in A.1)." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 780, + 309, + 791 + ], + "type": "text", + "content": "1018" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 154, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 154, + 84 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 154, + 84 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 90, + 291, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 90, + 291, + 225 + ], + "spans": [ + { + "bbox": [ + 67, + 90, + 291, + 225 + ], + "type": "text", + "content": "We compare three models. The baseline model is the multi-annotator model by Davani et al. (2022). We use their multi-task variant: For each annotator, there is a separate classification layer trained on annotations from that annotator. All annotator layers share a pre-trained language model used to encode the input. We use RoBERTa (Liu et al., 2019) for this, motivated by computational constraints. The other models in our experiments build on this baseline model." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 227, + 291, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 227, + 291, + 375 + ], + "spans": [ + { + "bbox": [ + 67, + 227, + 291, + 375 + ], + "type": "text", + "content": "For the sociodemographic models, we add group-specific layers based on sociodemographic attributes of the annotators. A single attribute, e.g., age, implies several groups, e.g., ages 25-34, ages 35-44. We add the group-specific layers between the pre-trained model and the annotator layers. Each group of annotators shares a separate group-specific layer. We implement group-specific layers as fully-connected, linear layers, each learning a feature transformation applied for one group of annotators." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 376, + 291, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 376, + 291, + 551 + ], + "spans": [ + { + "bbox": [ + 67, + 376, + 291, + 551 + ], + "type": "text", + "content": "Finally, for the random models, we shuffle the assignment of annotators to groups from the sociodemographic model, retaining the relative group sizes. In other words, the probability of each annotator staying in the same group or being reassigned to another group corresponds to the relative size of each group. This approach keeps the model architecture constant while removing the connection between actual sociodemographic attributes and group assignment. It allows us to distinguish the effects of additional parameters, which group-specific layers add in comparison to the baseline, from the effects of sociodemographic information." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 559, + 175, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 559, + 175, + 571 + ], + "spans": [ + { + "bbox": [ + 67, + 559, + 175, + 571 + ], + "type": "text", + "content": "4.1 Evaluation Setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 577, + 291, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 577, + 291, + 673 + ], + "spans": [ + { + "bbox": [ + 67, + 577, + 291, + 673 + ], + "type": "text", + "content": "We evaluate all models on individual annotations from gender, age, education, and sexual orientation groups. This setup is comparable to the \"individual label\" evaluations in Davani et al. (2022) and Gordon et al. (2022), but with scores calculated per group of annotators. We measure performance in macro-average " + }, + { + "bbox": [ + 67, + 577, + 291, + 673 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 67, + 577, + 291, + 673 + ], + "type": "text", + "content": ", to weigh each class equally." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 678, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 678, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 678, + 291, + 773 + ], + "type": "text", + "content": "Cross-Validation As there is no standard split available for our dataset, we perform three iterations of a four-fold cross-validation with different seeds (training details in Appendix A.3). We choose four folds, so that even very small groups have more than a hundred annotations in each test set. Across folds, the numbers of annotations per" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 71, + 526, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 166 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 166 + ], + "type": "text", + "content": "sociodemographic group are similar (see Appendix A.4). We construct test sets that only contain comments unseen by the annotators in the training set. We also ensure that all test sets have similar proportions of toxic or non-toxic comments (assigned by the majority of annotators) to address the class imbalance in the dataset (70.62% toxic, see §3)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 173, + 526, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 173, + 526, + 321 + ], + "spans": [ + { + "bbox": [ + 302, + 173, + 526, + 321 + ], + "type": "text", + "content": "Statistical Significance We test for statistical significance of our results from multiple runs of k-fold cross-validation via replicability analysis (Dror et al., 2017). We report the number of significant folds and the Bonferroni-corrected count (Dror et al., 2018) in Appendix A.2. We compute the p-values for each fold via a paired bootstrap-sampling test with BooStSa (Fornaciari et al., 2022). We set the significance level " + }, + { + "bbox": [ + 302, + 173, + 526, + 321 + ], + "type": "inline_equation", + "content": "\\alpha = 0.05" + }, + { + "bbox": [ + 302, + 173, + 526, + 321 + ], + "type": "text", + "content": ", draw 1000 bootstrap samples per fold, and use a sample size of " + }, + { + "bbox": [ + 302, + 173, + 526, + 321 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 302, + 173, + 526, + 321 + ], + "type": "text", + "content": " of the respective test set." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 327, + 525, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 327, + 525, + 395 + ], + "spans": [ + { + "bbox": [ + 302, + 327, + 525, + 395 + ], + "type": "text", + "content": "Remarks on Groups Annotators from different groups of the same attribute will in most cases not have annotated the same examples. Therefore, comparisons between models are only meaningful within each group." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 396, + 525, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 396, + 525, + 491 + ], + "spans": [ + { + "bbox": [ + 302, + 396, + 525, + 491 + ], + "type": "text", + "content": "The groups modeled via group-specific layers and those in the result tables are always the same. For example, if we report scores for gender groups, then the sociodemographic and randomized models are also based on gender groups. In the following, we focus on a subset of groups, omitting, e.g., \"Prefer not to say\" (see Appendix A.5)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 303, + 497, + 361, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 497, + 361, + 510 + ], + "spans": [ + { + "bbox": [ + 303, + 497, + 361, + 510 + ], + "type": "text", + "content": "5 Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 517, + 525, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 517, + 525, + 585 + ], + "spans": [ + { + "bbox": [ + 302, + 517, + 525, + 585 + ], + "type": "text", + "content": "Table 1 shows the results for gender, age, education, and sexual orientation. A naive majority class baseline that predicts all input to be toxic performs worse than all other models with a large margin (exact results in Appendix A.5)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 591, + 526, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 591, + 526, + 714 + ], + "spans": [ + { + "bbox": [ + 302, + 591, + 526, + 714 + ], + "type": "text", + "content": "Sociodemographics vs. Baseline Across attributes, the average scores of the sociodemographic model and the baseline are similar. The sociodemographic model often has a slightly higher average macro F1 than the baseline, but no statistically significant gains. Where average performance is better by several points, as for homosexual annotators, this gain is offset by a large variance in performance (a consequence of small group sizes)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 719, + 525, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 719, + 525, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 719, + 525, + 773 + ], + "type": "text", + "content": "Sociodemographics vs. Random We also do not find significant performance differences between sociodemographic group-layer models and the corresponding random group assignment models. For" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 780, + 309, + 791 + ], + "type": "text", + "content": "1019" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 291, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 291, + 125 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 291, + 125 + ], + "type": "text", + "content": "most groups, the randomized models achieve the highest average scores, but differences to the sociodemographic model are never statistically significant." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 88, + 131, + 268, + 185 + ], + "blocks": [ + { + "bbox": [ + 88, + 131, + 268, + 185 + ], + "lines": [ + { + "bbox": [ + 88, + 131, + 268, + 185 + ], + "spans": [ + { + "bbox": [ + 88, + 131, + 268, + 185 + ], + "type": "table", + "html": "
GenderBaselineSoc-Dem.Random
Male68.00±0.4967.66±0.4667.63±0.53
Female62.23±0.5362.25±1.1962.41±0.92
Nonbinary56.33±6.0056.80±7.2458.00±7.49
", + "image_path": "c24ea2cb73b40d6d979e210364b98816e24e54af3f43d4058505b0f5ce60fa31.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 88, + 192, + 268, + 277 + ], + "blocks": [ + { + "bbox": [ + 88, + 192, + 268, + 277 + ], + "lines": [ + { + "bbox": [ + 88, + 192, + 268, + 277 + ], + "spans": [ + { + "bbox": [ + 88, + 192, + 268, + 277 + ], + "type": "table", + "html": "
AgeBaselineSoc-Dem.Random
18 - 2459.39±1.5860.44±1.0560.52±1.37
25 - 3466.72±0.5666.63±0.8366.92±0.51
35 - 4464.50±0.5964.94±1.3365.24±0.89
45 - 5465.68±0.6665.88±1.3965.98±0.83
55 - 6464.37±1.2264.94±1.6664.84±1.30
65 or older63.34±2.0764.70±2.2162.77±2.39
", + "image_path": "346b1a1e433ecaf0ce38f83b147608910c5f7a10c767d1d9bcdd7c808b1dd0e1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 73, + 285, + 283, + 388 + ], + "blocks": [ + { + "bbox": [ + 73, + 285, + 283, + 388 + ], + "lines": [ + { + "bbox": [ + 73, + 285, + 283, + 388 + ], + "spans": [ + { + "bbox": [ + 73, + 285, + 283, + 388 + ], + "type": "table", + "html": "
EducationBaselineSoc-Dem.Random
Associate degree60.69±1.4460.54±2.3560.78±1.62
Bachelor's degree66.16±0.5166.23±0.8266.80±0.54
Doctoral degree61.93±3.8263.79±5.0363.27±3.67
High school60.53±1.3960.47±2.2260.55±1.87
Below high school58.28±4.6862.12±4.9060.17±4.25
Master's degree69.71±0.8669.58±0.9369.45±0.96
Professional degree66.75±2.3767.84±3.3268.62±2.84
College, no degree58.65±1.1959.40±1.7959.99±2.19
", + "image_path": "2e360ddab0beb967cacf0c9f50c8ea0eb13bda8f9b2f7cef1bfc636d3a7869f2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 83, + 396, + 272, + 451 + ], + "blocks": [ + { + "bbox": [ + 83, + 396, + 272, + 451 + ], + "lines": [ + { + "bbox": [ + 83, + 396, + 272, + 451 + ], + "spans": [ + { + "bbox": [ + 83, + 396, + 272, + 451 + ], + "type": "table", + "html": "
SexualityBaselineSoc-Dem.Random
Bisexual71.83±1.1471.42±1.5169.46±1.95
Heterosexual63.25±0.3963.32±1.2163.82±0.55
Homosexual64.43±1.7566.11±2.2065.12±1.94
", + "image_path": "18f8ad51bbc02072cbb4e2a4a00d78630d9522133575aea6cc0f461e550fb257.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 458, + 291, + 519 + ], + "lines": [ + { + "bbox": [ + 67, + 458, + 291, + 519 + ], + "spans": [ + { + "bbox": [ + 67, + 458, + 291, + 519 + ], + "type": "text", + "content": "Table 1: Average and standard deviation of macro " + }, + { + "bbox": [ + 67, + 458, + 291, + 519 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 67, + 458, + 291, + 519 + ], + "type": "text", + "content": " from three runs of four-fold stratified cross-validation. Separate table for each attribute. Bold results are the highest averages per group. However, no difference is statistically significant (see Appendix A.2)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 538, + 143, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 538, + 143, + 550 + ], + "spans": [ + { + "bbox": [ + 67, + 538, + 143, + 550 + ], + "type": "text", + "content": "6 Discussion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 556, + 291, + 731 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 556, + 291, + 731 + ], + "spans": [ + { + "bbox": [ + 67, + 556, + 291, + 731 + ], + "type": "text", + "content": "We do not find strong evidence that explicitly modelling sociodemographics helps to predict annotation behaviour with multi-annotator models. These results might seem counter-intuitive, given the evidence of systematic annotation differences between sociodemographic groups (see §2). This discrepancy, however, echoes the issue highlighted by ecological fallacies (Robinson, 1950): Not every annotator will be a perfect representative of their group, so we will not necessarily learn additional information based on their group identity. This seems especially true if we already have access to individual behaviour (i.e., individual annotations)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 733, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 733, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 733, + 291, + 772 + ], + "type": "text", + "content": "In contrast to Davani et al. (2022), we made sociodemographic information explicit in our experiments, as one of the factors influencing annotation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 301, + 71, + 526, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 71, + 526, + 220 + ], + "spans": [ + { + "bbox": [ + 301, + 71, + 526, + 220 + ], + "type": "text", + "content": "behaviour. Group-specific layers can be seen as an inductive bias putting emphasis on the sociodemographic relations between annotators. However, there are potentially many other factors influencing annotation behaviour (e.g., attitudes, moral values, cognitive biases, psychological traits). In light of our results, it seems plausible that multi-annotator models learn about these factors implicitly as part of predicting individual behaviour, so that making one factor explicit does not change prediction quality, at least in the case of sociodemographics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 223, + 526, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 223, + 526, + 507 + ], + "spans": [ + { + "bbox": [ + 302, + 223, + 526, + 507 + ], + "type": "text", + "content": "Still, we also know that generally group attributes can help predict individual decisions, i.e., as base rates or priors. To avoid ecological fallacies in modelling annotation, we therefore need to better understand when and how modelling sociodemographic information is useful in predicting an individual annotator's decisions. For example, we have only evaluated group-specific layers for single attributes. In contrast, social scientists have long adopted the idea of intersectionality (Crenshaw, 1989), which also informs research on fairness in machine learning (Wang et al., 2022). Intersectionality means that the effect of interactions between sociodemographic attributes enables specific experiences that are not captured by the attributes in isolation. For example, identifying as a man means something different depending on the person's education. Groups derived from single attributes might simply be too coarse to improve classifiers learnt from individual labels, as in multi-annotator models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 512, + 525, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 512, + 525, + 674 + ], + "spans": [ + { + "bbox": [ + 302, + 512, + 525, + 674 + ], + "type": "text", + "content": "The dataset we use (Kumar et al., 2021) has many characteristics which are ideal for our study (see §3). However, it uses a broad notion of toxicity, in contrast to other studies of toxic language (Larimore et al., 2021; Sap et al., 2022), which match content and analysed groups. When modeling the groups frequently referenced in the datasets themselves, we would expect greater benefits from group-specific layers. Similar to us, Biester et al. (2022) who do not find significant differences between annotators of different genders, do so in a more general setting." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 678, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 678, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 678, + 526, + 772 + ], + "type": "text", + "content": "We can only partially compare to Gordon et al. (2022), despite using the same dataset. In addition to differences in approach (see §2), our and their work also differ in their research questions and thus experimental conditions. Gordon et al. (2022) compare their full model (group and individual) against using group information alone." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 780, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 780, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 780, + 310, + 791 + ], + "type": "text", + "content": "1020" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 293, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 293, + 248 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 293, + 248 + ], + "type": "text", + "content": "We compare our full model (group and individual) against using individual information alone. So it is unclear if their model would benefit from group information in comparison to individual-level information alone. While they find an improvement from group information it is only in comparison to a baseline predicting not individual but aggregated labels. Additionally, the composition of test sets sampled from the full dataset differs between the studies: Gordon et al. (2022) use a test set of 5,000 comments, while we use 22,360 comments in a four-fold cross-validation. We leave an explicit comparison to future work." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 248, + 291, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 248, + 291, + 410 + ], + "spans": [ + { + "bbox": [ + 69, + 248, + 291, + 410 + ], + "type": "text", + "content": "Group-specific layers (§4) are a natural extension of annotator-specific classification layers in multi-annotator models. However, other architectures to predict annotator-level labels use different ways to represent sociodemographic information, e.g., via embeddings in a recommender system (Gordon et al., 2022). Future work could explore additional representations of annotator attributes (e.g., as part of the input, either textual or as separate features) and other approaches to modelling the relation of individual labeling decisions and attributes (e.g., probabilistic graphical models)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 416, + 147, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 416, + 147, + 428 + ], + "spans": [ + { + "bbox": [ + 67, + 416, + 147, + 428 + ], + "type": "text", + "content": "7 Conclusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 436, + 291, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 436, + 291, + 694 + ], + "spans": [ + { + "bbox": [ + 67, + 436, + 291, + 694 + ], + "type": "text", + "content": "We ask how relevant modelling explicit sociodemographic information is in learning from individual annotators. Our experiments with group-specific layers for four sociodemographic attributes on social media data with toxicity annotations (Kumar et al., 2021) show no significant benefit of modelling sociodemographic groups in multi-annotator models. However, as the issue of ecological fallacies highlights, it is not implausible that these models do not learn additional information from group information beyond the inherent variation. However, our results do not refute the usefulness of sociodemographic attributes in modelling annotation, but underscore the importance of their judicious use. Different tasks and model architectures will likely benefit to different extents. Ultimately, annotation behaviour is driven by complex factors and we will need to consider more than annotators' sociodemographics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 700, + 170, + 714 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 700, + 170, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 700, + 170, + 714 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 719, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 719, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 719, + 291, + 773 + ], + "type": "text", + "content": "We thank Deepak Kumar for providing access to the disaggregated dataset and his continued support. We also thank Aida Mostafazadeh Davani for providing information on implementation de" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 71, + 526, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 126 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 126 + ], + "type": "text", + "content": "tails of multi-annotator models. Members of MilaNLP (Bocconi) and the Semantic Computing Group (Bielefeld) provided feedback on earlier versions of this paper, for which we thank them again." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 126, + 527, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 126, + 527, + 220 + ], + "spans": [ + { + "bbox": [ + 302, + 126, + 527, + 220 + ], + "type": "text", + "content": "This work has in part been funded by the European Research Council (ERC) under the European Union's Horizon 2020 research and innovation program (grant agreement No. 949944, INTEGRACTOR). Likewise, this work has in part been funded by the VolkswagenStiftung as part of the \"3B Bots Building Bridges\" project." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 303, + 226, + 365, + 238 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 226, + 365, + 238 + ], + "spans": [ + { + "bbox": [ + 303, + 226, + 365, + 238 + ], + "type": "text", + "content": "Limitations" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 244, + 526, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 244, + 526, + 421 + ], + "spans": [ + { + "bbox": [ + 302, + 244, + 526, + 421 + ], + "type": "text", + "content": "While the dataset by Kumar et al. (2021) enabled us to test models for a range of often overlooked groups (e.g., non-binary or bisexual annotators), we ultimately modelled only four specific attributes (gender, age, education, sexual orientation). There are likely to be more factors that could play a role. Additionally, annotators in the Kumar et al. (2021) dataset are exclusively from the United States of America, so that results do not necessarily hold for other countries or cultures (Hovy and Yang, 2021). Specifically perceptions of harmful content online are known to vary across countries (Jiang et al., 2021)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 421, + 526, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 421, + 526, + 665 + ], + "spans": [ + { + "bbox": [ + 302, + 421, + 526, + 665 + ], + "type": "text", + "content": "We used only the (Kumar et al., 2021) dataset. This is mainly due to our strict criteria regarding dataset size and availability of annotator-level labels and sociodemographic information. These characteristics were a prerequisite for our experiments across different attributes with sufficient numbers of annotators. Most datasets which include annotator-level labels and sociodemographic information contain much smaller numbers of annotators and attributes. Nevertheless, with the Measuring Hate Speech Corpus there is at least one additional dataset (Sachdeva et al., 2022) with comparable characteristics that could be used in future experiments. Also, additional small-scale, more focused experiments could use datasets like Sap et al. (2022) or HS-Brexit (Akhtar et al., 2021) which was annotated by 6 annotators, each from one of two sociodemographic groups." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 665, + 526, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 665, + 526, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 665, + 526, + 773 + ], + "type": "text", + "content": "We do not study the aggregation of individual predictions or evaluate against majority labels, as these are not directly relevant to our investigation of sociodemographic attributes in models of annotation behaviour. Consequently, we cannot derive a conclusion about performance in those settings from our results. This is a noteworthy limitation, because part of the experiments introducing" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 308, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 308, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 308, + 791 + ], + "type": "text", + "content": "1021" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 290, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 290, + 125 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 290, + 125 + ], + "type": "text", + "content": "multi-annotator models in Davani et al. (2022) compare labels aggregated from multi-annotator models against predictions from a standard classifier (directly trained on aggregated labels)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 126, + 290, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 126, + 290, + 179 + ], + "spans": [ + { + "bbox": [ + 67, + 126, + 290, + 179 + ], + "type": "text", + "content": "For computational reasons, our experiments use a comparatively small pre-trained language model (RoBERTa, Liu et al. 2019). Thus, results might differ with larger models." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 68, + 186, + 158, + 199 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 186, + 158, + 199 + ], + "spans": [ + { + "bbox": [ + 68, + 186, + 158, + 199 + ], + "type": "text", + "content": "Ethics Statement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 207, + 290, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 207, + 290, + 327 + ], + "spans": [ + { + "bbox": [ + 66, + 207, + 290, + 327 + ], + "type": "text", + "content": "As sociodemographic attributes are sensitive information, we do not infer attributes, but build on a self-reported, IRB-reviewed dataset (Kumar et al., 2021). We also see potential for a discussion of \"privacy by design\" in modelling human label variation based on our results: There can be circumstances in which knowing more about annotators is not relevant, and indeed might lead to violations of privacy." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 330, + 291, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 330, + 291, + 478 + ], + "spans": [ + { + "bbox": [ + 67, + 330, + 291, + 478 + ], + "type": "text", + "content": "As multi-annotator models attempt to capture the preferences of individual annotators, there are valid concerns around privacy and anonymity. As discussed in Davani et al. (2022), increasing the annotator count can be one option to reduce privacy risks. We show it is feasible to learn a model for a large number of individual annotators (5002 vs. 18 and 82 in their work). But a prerequisite for improved privacy is to apply effective aggregation on top of individual predictions, which we do not study in the present work." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 498, + 127, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 498, + 127, + 511 + ], + "spans": [ + { + "bbox": [ + 68, + 498, + 127, + 511 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 517, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 69, + 517, + 291, + 573 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 517, + 291, + 573 + ], + "spans": [ + { + "bbox": [ + 69, + 517, + 291, + 573 + ], + "type": "text", + "content": "Gavin Abercrombie, Valerio Basile, Sara Tonelli, Verena Rieser, and Alexandra Uma, editors. 2022. Proceedings of the 1st Workshop on Perspectivist Approaches to NLP @LREC2022. European Language Resources Association, Marseille, France." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 583, + 291, + 640 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 583, + 291, + 640 + ], + "spans": [ + { + "bbox": [ + 69, + 583, + 291, + 640 + ], + "type": "text", + "content": "Sohail Akhtar, Valerio Basile, and Viviana Patti. 2020. Modeling annotator perspective and polarized opinions to improve hate speech detection. In Proceedings of the AAAI Conference on Human Computation and Crowdsourcing, volume 8, pages 151-154." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 650, + 291, + 695 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 650, + 291, + 695 + ], + "spans": [ + { + "bbox": [ + 69, + 650, + 291, + 695 + ], + "type": "text", + "content": "Sohail Akhtar, Valerio Basile, and Viviana Patti. 2021. Whose opinions matter? perspective-aware models to identify opinions of hate speech victims in abusive language detection. Preprint arXiv:2106.15896." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 706, + 291, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 706, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 706, + 291, + 772 + ], + "type": "text", + "content": "Hala Al Kuwatly, Maximilian Wich, and Georg Groh. 2020. Identifying and measuring annotator bias based on annotators' demographic characteristics. In Proceedings of the Fourth Workshop on Online Abuse and Harms, pages 184-190, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 71, + 526, + 772 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 304, + 71, + 526, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 71, + 526, + 150 + ], + "spans": [ + { + "bbox": [ + 304, + 71, + 526, + 150 + ], + "type": "text", + "content": "Valerio Basile, Michael Fell, Tommaso Fornaciari, Dirk Hovy, Silviu Paun, Barbara Plank, Massimo Poesio, and Alexandra Uma. 2021. We need to consider disagreement in evaluation. In Proceedings of the 1st Workshop on Benchmarking: Past, Present and Future, pages 15-21, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 160, + 526, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 160, + 526, + 238 + ], + "spans": [ + { + "bbox": [ + 304, + 160, + 526, + 238 + ], + "type": "text", + "content": "Laura Biester, Vanita Sharma, Ashkan Kazemi, Naihao Deng, Steven Wilson, and Rada Mihalcea. 2022. Analyzing the effects of annotator gender across NLP tasks. In Proceedings of the 1st Workshop on Perspectivist Approaches to NLP @LREC2022, pages 10-19, Marseille, France. European Language Resources Association." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 249, + 526, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 249, + 526, + 316 + ], + "spans": [ + { + "bbox": [ + 304, + 249, + 526, + 316 + ], + "type": "text", + "content": "Reuben Binns, Michael Veale, Max Van Kleek, and Nigel Shadbolt. 2017. Like trainer, like bot? inheritance of bias in algorithmic content moderation. In Social Informatics, Lecture Notes in Computer Science, pages 405-415. Springer International Publishing." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 327, + 526, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 526, + 405 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 526, + 405 + ], + "type": "text", + "content": "Amanda Cercas Curry, Gavin Abercrombie, and Verena Rieser. 2021. ConvAbuse: Data, analysis, and benchmarks for nuanced abuse detection in conversational AI. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7388-7403, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 416, + 526, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 416, + 526, + 471 + ], + "spans": [ + { + "bbox": [ + 304, + 416, + 526, + 471 + ], + "type": "text", + "content": "Kimberle Crenshaw. 1989. Demarginalizing the intersection of race and sex: A black feminist critique of antidiscrimination doctrine, feminist theory and antiracist politics. University of Chicago Legal Forum, 1989(1):Article 8." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 483, + 526, + 539 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 526, + 539 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 526, + 539 + ], + "type": "text", + "content": "Aida Mostafazadeh Davani, Mark Diaz, and Vinodkumar Prabhakaran. 2022. Dealing with disagreements: Looking beyond the majority vote in subjective annotations. Transactions of the Association for Computational Linguistics, 10:92-110." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 550, + 526, + 606 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 550, + 526, + 606 + ], + "spans": [ + { + "bbox": [ + 304, + 550, + 526, + 606 + ], + "type": "text", + "content": "Rotem Dror, Gili Baumer, Marina Bogomolov, and Roi Reichart. 2017. Replicability analysis for natural language processing: Testing significance with multiple datasets. Transactions of the Association for Computational Linguistics, 5:471-486." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 617, + 526, + 695 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 526, + 695 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 526, + 695 + ], + "type": "text", + "content": "Rotem Dror, Gili Baumer, Segev Shlomov, and Roi Reichart. 2018. The hitchhiker's guide to testing statistical significance in natural language processing. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1383-1392, Melbourne, Australia. Association for Computational Linguistics." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 706, + 526, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 706, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 706, + 526, + 772 + ], + "type": "text", + "content": "Elizabeth Excell and Noura Al Moubayed. 2021. Towards equal gender representation in the annotations of toxic language detection. In Proceedings of the 3rd Workshop on Gender Bias in Natural Language Processing, pages 55-65, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1022" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 291, + 171 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 291, + 171 + ], + "type": "text", + "content": "Tommaso Fornaciari, Alexandra Uma, Silviu Paun, Barbara Plank, Dirk Hovy, and Massimo Poesio. 2021. Beyond black & white: Leveraging annotator disagreement via soft-label multi-task learning. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2591-2597, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 180, + 290, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 180, + 290, + 258 + ], + "spans": [ + { + "bbox": [ + 69, + 180, + 290, + 258 + ], + "type": "text", + "content": "Tommaso Fornaciari, Alexandra Uma, Massimo Poesio, and Dirk Hovy. 2022. Hard and soft evaluation of NLP models with BOOtSTrap SAmpling - BooStSa. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 127-134, Dublin, Ireland. Association for Computational Linguistics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 266, + 290, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 266, + 290, + 311 + ], + "spans": [ + { + "bbox": [ + 69, + 266, + 290, + 311 + ], + "type": "text", + "content": "David A. Freedman. 2015. Ecological inference. In James D. Wright, editor, International Encyclopedia of the Social & Behavioral Sciences (Second Edition), pages 868-870. Elsevier." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 319, + 290, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 319, + 290, + 396 + ], + "spans": [ + { + "bbox": [ + 69, + 319, + 290, + 396 + ], + "type": "text", + "content": "Mitchell L. Gordon, Michelle S. Lam, Joon Sung Park, Kayur Patel, Jeff Hancock, Tatsunori Hashimoto, and Michael S. Bernstein. 2022. Jury learning: Integrating dissenting voices into machine learning models. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems, CHI '22, pages 1-19. Association for Computing Machinery." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 405, + 290, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 405, + 290, + 460 + ], + "spans": [ + { + "bbox": [ + 69, + 405, + 290, + 460 + ], + "type": "text", + "content": "Nitesh Goyal, Ian D. Kivlichan, Rachel Rosen, and Lucy Vasserman. 2022. Is your toxicity my toxicity? exploring the impact of rater identity on toxicity annotation. Proceedings of the ACM on Human-Computer Interaction, 6:1-28." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 470, + 290, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 470, + 290, + 546 + ], + "spans": [ + { + "bbox": [ + 69, + 470, + 290, + 546 + ], + "type": "text", + "content": "Dirk Hovy and Diyi Yang. 2021. The importance of modeling social factors of language: Theory and practice. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 588-602, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 555, + 290, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 555, + 290, + 633 + ], + "spans": [ + { + "bbox": [ + 69, + 555, + 290, + 633 + ], + "type": "text", + "content": "Emily Jamison and Iryna Gurevych. 2015. Noise or additional information? leveraging crowdsourced annotation item agreement for natural language tasks. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 291-297, Lisbon, Portugal. Association for Computational Linguistics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 641, + 290, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 641, + 290, + 686 + ], + "spans": [ + { + "bbox": [ + 69, + 641, + 290, + 686 + ], + "type": "text", + "content": "Jialun Aaron Jiang, Morgan Klaus Scheuerman, Casey Fiesler, and Jed R. Brubaker. 2021. Understanding international perceptions of the severity of harmful content online. PLOS ONE, 16(8)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 694, + 290, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 694, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 694, + 290, + 772 + ], + "type": "text", + "content": "Deepak Kumar, Patrick Gage Kelley, Sunny Consolvo, Joshua Mason, Elie Bursztein, Zakir Durmeric, Kurt Thomas, and Michael Bailey. 2021. Designing toxic content classification for a diversity of perspectives. In Seventeenth Symposium on Usable Privacy and Security (SOUPS 2021), pages 299-318. USENIX Association." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 525, + 772 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 304, + 72, + 525, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 525, + 149 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 525, + 149 + ], + "type": "text", + "content": "Savannah Larimore, Ian Kennedy, Breon Haskett, and Alina Arseniev-Koehler. 2021. Reconsidering annotator disagreement about racist language: Noise or signal? In Proceedings of the Ninth International Workshop on Natural Language Processing for Social Media, pages 81-90, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 157, + 525, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 157, + 525, + 212 + ], + "spans": [ + { + "bbox": [ + 304, + 157, + 525, + 212 + ], + "type": "text", + "content": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. Preprint arXiv:1907.11692." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 219, + 525, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 219, + 525, + 296 + ], + "spans": [ + { + "bbox": [ + 304, + 219, + 525, + 296 + ], + "type": "text", + "content": "F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. 2011. Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12:2825-2830." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 304, + 525, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 525, + 370 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 525, + 370 + ], + "type": "text", + "content": "Barbara Plank. 2022. The \"problem\" of human label variation: On ground truth in data, modeling and evaluation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 10671-10682, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 377, + 525, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 525, + 455 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 525, + 455 + ], + "type": "text", + "content": "Barbara Plank, Dirk Hovy, and Anders Søgaard. 2014. Learning part-of-speech taggers with inter-annotator agreement loss. In Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 742–751, Gothenburg, Sweden. Association for Computational Linguistics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 463, + 525, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 463, + 525, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 463, + 525, + 540 + ], + "type": "text", + "content": "Vinodkumar Prabhakaran, Aida Mostafazadeh Davani, and Mark Diaz. 2021. On releasing annotator-level labels and information in datasets. In Proceedings of the Joint 15th Linguistic Annotation Workshop (LAW) and 3rd Designing Meaning Representations (DMR) Workshop, pages 133-138, Punta Cana, Dominican Republic. Association for Computational Linguistics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 547, + 525, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 547, + 525, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 547, + 525, + 581 + ], + "type": "text", + "content": "W. S. Robinson. 1950. Ecological correlations and the behavior of individuals. American Sociological Review, 15(3):351-357." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 588, + 525, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 588, + 525, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 588, + 525, + 677 + ], + "type": "text", + "content": "Paul Röttger, Bertie Vidgen, Dirk Hovy, and Janet Pierrehumbert. 2022. Two contrasting data annotation paradigms for subjective NLP tasks. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 175-190, Seattle, United States. Association for Computational Linguistics." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 684, + 525, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 684, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 684, + 525, + 772 + ], + "type": "text", + "content": "Pratik Sachdeva, Renata Barreto, Geoff Bacon, Alexander Sahn, Claudia von Vacano, and Chris Kennedy. 2022. The measuring hate speech corpus: Leveraging rasch measurement theory for data perspectivism. In Proceedings of the 1st Workshop on Perspectivist Approaches to NLP @LREC2022, pages 83-94, Marseille, France. European Language Resources Association." + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1023" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 290, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 290, + 127 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 290, + 127 + ], + "type": "text", + "content": "Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. 2019. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. In 5th Workshop on Energy Efficient Machine Learning and Cognitive Computing @ NeurIPS 2019." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 136, + 290, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 136, + 290, + 202 + ], + "spans": [ + { + "bbox": [ + 69, + 136, + 290, + 202 + ], + "type": "text", + "content": "Maarten Sap, Dallas Card, Saadia Gabriel, Yejin Choi, and Noah A. Smith. 2019. The risk of racial bias in hate speech detection. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1668-1678, Florence, Italy. Association for Computational Linguistics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 210, + 290, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 210, + 290, + 309 + ], + "spans": [ + { + "bbox": [ + 69, + 210, + 290, + 309 + ], + "type": "text", + "content": "Maarten Sap, Swabha Swayamdipta, Laura Vianna, Xuhui Zhou, Yejin Choi, and Noah A. Smith. 2022. Annotators with attitudes: How annotator beliefs and identities bias toxic language detection. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5884-5906, Seattle, United States. Association for Computational Linguistics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 317, + 290, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 317, + 290, + 394 + ], + "spans": [ + { + "bbox": [ + 69, + 317, + 290, + 394 + ], + "type": "text", + "content": "Qinlan Shen and Carolyn Rose. 2021. What sounds \"right\" to me? experiential factors in the perception of political ideology. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 1762-1771, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 403, + 290, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 403, + 290, + 457 + ], + "spans": [ + { + "bbox": [ + 69, + 403, + 290, + 457 + ], + "type": "text", + "content": "Alexandra N. Uma, Tommaso Fornaciari, Dirk Hovy, Silviu Paun, Barbara Plank, and Massimo Poesio. 2021. Learning from disagreement: A survey. Journal of Artificial Intelligence Research, 72:1385-1470." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 466, + 290, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 466, + 290, + 544 + ], + "spans": [ + { + "bbox": [ + 69, + 466, + 290, + 544 + ], + "type": "text", + "content": "Angelina Wang, Vikram V Ramaswamy, and Olga Russakovsky. 2022. Towards intersectionality in machine learning: Including more identities, handling underrepresentation, and performing evaluation. In 2022 ACM Conference on Fairness, Accountability, and Transparency, FAccT '22, pages 336-349. Association for Computing Machinery." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 551, + 290, + 683 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 551, + 290, + 683 + ], + "spans": [ + { + "bbox": [ + 69, + 551, + 290, + 683 + ], + "type": "text", + "content": "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 694, + 141, + 708 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 694, + 141, + 708 + ], + "spans": [ + { + "bbox": [ + 69, + 694, + 141, + 708 + ], + "type": "text", + "content": "A Appendix" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 68, + 714, + 289, + 727 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 714, + 289, + 727 + ], + "spans": [ + { + "bbox": [ + 68, + 714, + 289, + 727 + ], + "type": "text", + "content": "A.1 Annotator Sociodemographics in Sample" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 733, + 289, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 733, + 289, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 733, + 289, + 772 + ], + "type": "text", + "content": "Table 2 shows how many annotators the sample contains. Counts are given per group of the four attributes gender, age, education and sexuality." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 71, + 525, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 525, + 301 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 525, + 301 + ], + "type": "text", + "content": "In the Kumar et al. (2021) dataset, sociodemographic attributes are given for each individual annotation - not once per annotator. For some annotators, conflicting attribute values exist (e.g., two different age groups). As the data collection spanned several months (Kumar et al., 2021), these value changes can in principle be reasonable (e.g., because an annotator got older, finished a degree, changed sexual preference or gender identity). However, as reasonable changes can not easily be discerned from erroneous input, we disambiguate values based on a heuristic: If an annotator reports several values for an attribute, we assume the most frequent value to be valid. In cases of no clear most frequent value, we set the attribute to \"Prefer not to say\". Thus, the main results do not contain annotators with ambiguous attributes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 303, + 308, + 415, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 308, + 415, + 320 + ], + "spans": [ + { + "bbox": [ + 303, + 308, + 415, + 320 + ], + "type": "text", + "content": "A.2 Significance Tests" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 327, + 525, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 327, + 525, + 488 + ], + "spans": [ + { + "bbox": [ + 302, + 327, + 525, + 488 + ], + "type": "text", + "content": "Results of a replicability analysis (Dror et al., 2017) testing for significant differences in macro " + }, + { + "bbox": [ + 302, + 327, + 525, + 488 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 302, + 327, + 525, + 488 + ], + "type": "text", + "content": " on scores from three runs of four-fold cross-validation. Table 3 shows results for a comparison of the sociodemographic models against the baseline models. Table 4 shows results for a comparison of the sociodemographic models against the randomized assignment models. The Bonferroni correction for the corrected count of significant folds " + }, + { + "bbox": [ + 302, + 327, + 525, + 488 + ], + "type": "inline_equation", + "content": "\\hat{k}_{Bonferroni}" + }, + { + "bbox": [ + 302, + 327, + 525, + 488 + ], + "type": "text", + "content": " is used to account for the fact that we have overlapping test sets from multiple runs of four-fold cross-validation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 303, + 496, + 517, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 496, + 517, + 523 + ], + "spans": [ + { + "bbox": [ + 303, + 496, + 517, + 523 + ], + "type": "text", + "content": "A.3 Training Details, Hyperparameters and Computational Resources" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 528, + 525, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 528, + 525, + 663 + ], + "spans": [ + { + "bbox": [ + 302, + 528, + 525, + 663 + ], + "type": "text", + "content": "We implement models and the training loop using the Hugging Face Transformers library (version 4.19.2, Wolf et al. 2020). Maximum sequence length is 512 tokens, with truncation and padding to the maximum length. We train for 3 epochs with a batch size of 8 and an initial learning rate of 0.00001. Otherwise, we used default parameters. We found results to particularly depend on the learning rate, with higher or lower values leading to worse results." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 665, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 665, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 665, + 525, + 772 + ], + "type": "text", + "content": "We use a weighted loss function. Label weights are calculated per annotator on the training set of each fold. Label weights, evaluation scores and the four-fold dataset splits (StratifiedKFold) are calculated using the scikit-learn library (version 1.0.2, Pedregosa et al. 2011). The folds are based on a fixed random seed per iteration: 2803636207, 165043843, 2923262358" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1024" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 102, + 68, + 255, + 153 + ], + "blocks": [ + { + "bbox": [ + 102, + 68, + 255, + 153 + ], + "lines": [ + { + "bbox": [ + 102, + 68, + 255, + 153 + ], + "spans": [ + { + "bbox": [ + 102, + 68, + 255, + 153 + ], + "type": "table", + "html": "
Number of Annotators
Gender
Female2450
Male2116
Prefer not to say412
Nonbinary23
Other1
", + "image_path": "caf24a1aec721b7e03cec00eef451e8e68c6963e40fc0accd6dd331759ac6b4f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 102, + 158, + 254, + 264 + ], + "blocks": [ + { + "bbox": [ + 102, + 158, + 254, + 264 + ], + "lines": [ + { + "bbox": [ + 102, + 158, + 254, + 264 + ], + "spans": [ + { + "bbox": [ + 102, + 158, + 254, + 264 + ], + "type": "table", + "html": "
Number of Annotators
Age
18 - 24489
25 - 341861
35 - 441115
45 - 54529
55 - 64321
65 or older119
Prefer not to say568
", + "image_path": "725b26b53e9f011b4cb410c17f8c9a5999e171badfdb916080aa142b5ac9cb1e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 102, + 271, + 255, + 356 + ], + "blocks": [ + { + "bbox": [ + 102, + 271, + 255, + 356 + ], + "lines": [ + { + "bbox": [ + 102, + 271, + 255, + 356 + ], + "spans": [ + { + "bbox": [ + 102, + 271, + 255, + 356 + ], + "type": "table", + "html": "
Number of Annotators
Sexuality
Heterosexual4018
Bisexual469
Prefer not to say346
Homosexual134
Other35
", + "image_path": "4d2f443630932d6e0215455b61f115dbfe0bff95fba3b7a6b7d56694573f5c94.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 98, + 364, + 259, + 498 + ], + "blocks": [ + { + "bbox": [ + 98, + 364, + 259, + 498 + ], + "lines": [ + { + "bbox": [ + 98, + 364, + 259, + 498 + ], + "spans": [ + { + "bbox": [ + 98, + 364, + 259, + 498 + ], + "type": "table", + "html": "
EducationNumber of Annotators
Bachelor's degree1879
College, no degree861
Prefer not to say647
Master's degree642
Associate degree460
High school363
Professional degree68
Doctoral degree51
Below high school25
Other6
", + "image_path": "6519765e784f01feb0edab7fe68fff85ff643b928196d32c7956cb702ebb37d8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 567, + 290, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 567, + 290, + 703 + ], + "spans": [ + { + "bbox": [ + 67, + 567, + 290, + 703 + ], + "type": "text", + "content": "The majority of parameters in our model belong to the pre-trained language model shared between all group-specific and annotator-specific layers. Specifically, RoBERTa (Liu et al., 2019) in the roberta-base variant has 125 Million parameters. We keep the pre-trained model's default output dimensionality of 768, so that each group-specific layer adds " + }, + { + "bbox": [ + 67, + 567, + 290, + 703 + ], + "type": "inline_equation", + "content": "768 * 768 + 768 = 590" + }, + { + "bbox": [ + 67, + 567, + 290, + 703 + ], + "type": "text", + "content": ", 592 parameters and each annotator layer adds " + }, + { + "bbox": [ + 67, + 567, + 290, + 703 + ], + "type": "inline_equation", + "content": "768 * 2 + 2 = 1" + }, + { + "bbox": [ + 67, + 567, + 290, + 703 + ], + "type": "text", + "content": ", 538 parameters." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 705, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 705, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 705, + 291, + 772 + ], + "type": "text", + "content": "All experiments ran on a single GPU (GeForce GTX 1080 Ti, 12GB GPU RAM). Per fold, training and evaluation together take about three and a half hours in our setting. Three runs of four-fold cross-validation (12 folds), thus take around 42 hours" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 361, + 68, + 468, + 124 + ], + "blocks": [ + { + "bbox": [ + 67, + 506, + 290, + 542 + ], + "lines": [ + { + "bbox": [ + 67, + 506, + 290, + 542 + ], + "spans": [ + { + "bbox": [ + 67, + 506, + 290, + 542 + ], + "type": "text", + "content": "Table 2: Number of annotators per group for attributes gender, age, sexuality and education. Counts refer to the entire sample" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 361, + 68, + 468, + 124 + ], + "lines": [ + { + "bbox": [ + 361, + 68, + 468, + 124 + ], + "spans": [ + { + "bbox": [ + 361, + 68, + 468, + 124 + ], + "type": "table", + "html": "
hatkcounthatkBonf.
Female20
Male00
Nonbinary10
", + "image_path": "32ad7e940605d95071ac8ced8895f2a1491b15b7f03ce436050c304998f2f1b0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 361, + 132, + 468, + 218 + ], + "blocks": [ + { + "bbox": [ + 361, + 132, + 468, + 218 + ], + "lines": [ + { + "bbox": [ + 361, + 132, + 468, + 218 + ], + "spans": [ + { + "bbox": [ + 361, + 132, + 468, + 218 + ], + "type": "table", + "html": "
hatkcounthatkBonf.
18 - 2420
25 - 3420
35 - 4410
45 - 5400
55 - 6410
65 or older10
", + "image_path": "e3c0792bfae62e9417a852c39d1d4cc6e0619d8c96dbaa40b5a98dd6d2835c4e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 361, + 226, + 468, + 280 + ], + "blocks": [ + { + "bbox": [ + 361, + 226, + 468, + 280 + ], + "lines": [ + { + "bbox": [ + 361, + 226, + 468, + 280 + ], + "spans": [ + { + "bbox": [ + 361, + 226, + 468, + 280 + ], + "type": "table", + "html": "
ˆkcountˆkBonf.
Bisexual20
Heterosexual42
Homosexual10
", + "image_path": "8f80c5e137d1b0101a4ea36672c28c390623e4b188053e8071a5d8918a8b2473.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 346, + 290, + 480, + 394 + ], + "blocks": [ + { + "bbox": [ + 346, + 290, + 480, + 394 + ], + "lines": [ + { + "bbox": [ + 346, + 290, + 480, + 394 + ], + "spans": [ + { + "bbox": [ + 346, + 290, + 480, + 394 + ], + "type": "table", + "html": "
kcountkBonf.
Associate degree00
Bachelor's degree10
Doctoral degree20
High school00
Belowhigh school00
Master's degree00
Professional degree00
College, no degree22
", + "image_path": "a8a67b4a978d772342cef4159df4b52f82005757c861476a929c0158454e9f49.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 402, + 525, + 498 + ], + "lines": [ + { + "bbox": [ + 302, + 402, + 525, + 498 + ], + "spans": [ + { + "bbox": [ + 302, + 402, + 525, + 498 + ], + "type": "text", + "content": "Table 3: Results of a replicability analysis of baseline vs sociodemographic models. Raw and Bonferroni-corrected counts of significant folds out of 12 folds from three runs of four-fold cross-validation. P-values for each fold are computed via a paired bootstrap test with significance level " + }, + { + "bbox": [ + 302, + 402, + 525, + 498 + ], + "type": "inline_equation", + "content": "\\alpha = 0.05" + }, + { + "bbox": [ + 302, + 402, + 525, + 498 + ], + "type": "text", + "content": ", 1000 bootstrap samples per fold and a sample size of " + }, + { + "bbox": [ + 302, + 402, + 525, + 498 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 302, + 402, + 525, + 498 + ], + "type": "text", + "content": " of the respective test set." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 302, + 524, + 525, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 524, + 525, + 661 + ], + "spans": [ + { + "bbox": [ + 302, + 524, + 525, + 661 + ], + "type": "text", + "content": "(1.75 days). With four attributes and three trainable models the combined run time of the reported experiments is estimated to be 21 days. Including preliminary experiments, which, however, mostly were not full runs of k-fold cross-validation and also utilized DistilBERT (Sanh et al., 2019) with slightly faster run times, it will be many times more. There is no discernible difference in experiment run times between multi-annotator models with or without groups or different numbers of groups." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 671, + 524, + 695 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 671, + 524, + 695 + ], + "spans": [ + { + "bbox": [ + 302, + 671, + 524, + 695 + ], + "type": "text", + "content": "A.4 Number of Annotations per Group across all Test Sets" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 705, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 705, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 705, + 525, + 772 + ], + "type": "text", + "content": "Table 5 contains the number of annotations we have per group across the total of 12 folds (from three runs of four-fold cross-validation). This number of annotations is the effective test set size per group. As the numbers do not vary substantially, perfor" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 780, + 309, + 791 + ], + "type": "text", + "content": "1025" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 126, + 69, + 230, + 123 + ], + "blocks": [ + { + "bbox": [ + 126, + 69, + 230, + 123 + ], + "lines": [ + { + "bbox": [ + 126, + 69, + 230, + 123 + ], + "spans": [ + { + "bbox": [ + 126, + 69, + 230, + 123 + ], + "type": "table", + "html": "
kcountkBonf.
Female22
Male10
Nonbinary10
", + "image_path": "fe0ce9572f53a63635954facf9e6ea438bda1ce5b49524888e22863ae688e876.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 126, + 133, + 229, + 217 + ], + "blocks": [ + { + "bbox": [ + 126, + 133, + 229, + 217 + ], + "lines": [ + { + "bbox": [ + 126, + 133, + 229, + 217 + ], + "spans": [ + { + "bbox": [ + 126, + 133, + 229, + 217 + ], + "type": "table", + "html": "
kcountkBonf.
18 - 2410
25 - 3400
35 - 4410
45 - 5410
55 - 6430
65 or older10
", + "image_path": "dc82f1d9db26486c59056b62e1ded50a8075f393987e5dbed661e13e1f823f5f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 123, + 227, + 233, + 279 + ], + "blocks": [ + { + "bbox": [ + 123, + 227, + 233, + 279 + ], + "lines": [ + { + "bbox": [ + 123, + 227, + 233, + 279 + ], + "spans": [ + { + "bbox": [ + 123, + 227, + 233, + 279 + ], + "type": "table", + "html": "
hatcounthatBonf.
Bisexual62
Heterosexual11
Homosexual00
", + "image_path": "ab65613339085586161cb49be1db61fb9bcdbd71f381a4fbda16c5de3ee162bb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 111, + 290, + 245, + 394 + ], + "blocks": [ + { + "bbox": [ + 111, + 290, + 245, + 394 + ], + "lines": [ + { + "bbox": [ + 111, + 290, + 245, + 394 + ], + "spans": [ + { + "bbox": [ + 111, + 290, + 245, + 394 + ], + "type": "table", + "html": "
kcountkBonf.
Associate degree20
Bachelor's degree10
Doctoral degree00
High school20
Belowhigh school20
Master's degree00
Professional degree00
College, no degree11
", + "image_path": "702e850c9b2a192fecd3a5c53aec251af17118f60f15c9db6834da30e3abe6c5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 520, + 290, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 520, + 290, + 548 + ], + "spans": [ + { + "bbox": [ + 67, + 520, + 290, + 548 + ], + "type": "text", + "content": "mance on each fold is equally representative for all groups." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 552, + 154, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 552, + 154, + 565 + ], + "spans": [ + { + "bbox": [ + 68, + 552, + 154, + 565 + ], + "type": "text", + "content": "A.5 Full Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 571, + 291, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 571, + 291, + 611 + ], + "spans": [ + { + "bbox": [ + 67, + 571, + 291, + 611 + ], + "type": "text", + "content": "Table 6 shows full results of experiments (see 4), including results for all residual categories and a naive baseline which always predicts toxic." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 304, + 169, + 527, + 253 + ], + "blocks": [ + { + "bbox": [ + 67, + 402, + 291, + 498 + ], + "lines": [ + { + "bbox": [ + 67, + 402, + 291, + 498 + ], + "spans": [ + { + "bbox": [ + 67, + 402, + 291, + 498 + ], + "type": "text", + "content": "Table 4: Results from replicability analysis of randomized vs sociodemographic models. Raw and Bonferroni-corrected counts of significant folds out of 12 folds from three runs of four-fold cross-validation. P-values for each fold are computed via a paired bootstrap test with significance level " + }, + { + "bbox": [ + 67, + 402, + 291, + 498 + ], + "type": "inline_equation", + "content": "\\alpha = 0.05" + }, + { + "bbox": [ + 67, + 402, + 291, + 498 + ], + "type": "text", + "content": ", 1000 bootstrap samples per fold and a sample size of " + }, + { + "bbox": [ + 67, + 402, + 291, + 498 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 67, + 402, + 291, + 498 + ], + "type": "text", + "content": " of the respective test set." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 304, + 169, + 527, + 253 + ], + "lines": [ + { + "bbox": [ + 304, + 169, + 527, + 253 + ], + "spans": [ + { + "bbox": [ + 304, + 169, + 527, + 253 + ], + "type": "table", + "html": "
GenderNumber Of AnnotationsMinMax
Female13555±86.4413383.013664.0
Male11925±61.6511843.012062.0
Nonbinary115±6.03104.0122.0
Other5±1.952.08.0
Prefer not to say2345±51.192281.02453.0
", + "image_path": "7a27e45775e194b4ba54cfcacbc2a55e9f36cf3371e54d597e6cbd612b75c2a9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 307, + 261, + 518, + 365 + ], + "blocks": [ + { + "bbox": [ + 307, + 261, + 518, + 365 + ], + "lines": [ + { + "bbox": [ + 307, + 261, + 518, + 365 + ], + "spans": [ + { + "bbox": [ + 307, + 261, + 518, + 365 + ], + "type": "table", + "html": "
AgeNumber Of AnnotationsMinMax
18 - 242615±50.8825212697
25 - 3410315±61.451024410457
35 - 446250±51.0661796324
45 - 543025±47.2329293083
55 - 641865±25.4818311903
65 or older675±19.31643704
Prefer not to say3200±55.2831313289
", + "image_path": "a722e27081a3aea739b5da7b93cd401028b9d7cff73204fe2a029abc8d5d9b92.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 308, + 373, + 518, + 457 + ], + "blocks": [ + { + "bbox": [ + 308, + 373, + 518, + 457 + ], + "lines": [ + { + "bbox": [ + 308, + 373, + 518, + 457 + ], + "spans": [ + { + "bbox": [ + 308, + 373, + 518, + 457 + ], + "type": "table", + "html": "
SexualityNumber Of AnnotationsMinMax
Bisexual2445±39.2623832501
Heterosexual22630±63.002250722726
Homosexual725±26.57670759
Other190±7.91173201
Prefer not to say1955±35.3918782009
", + "image_path": "9e3544005ebc4e866aea52dd31404f4b2c48337a98939bfc114a3b6f90b0417d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 304, + 465, + 524, + 599 + ], + "blocks": [ + { + "bbox": [ + 304, + 465, + 524, + 599 + ], + "lines": [ + { + "bbox": [ + 304, + 465, + 524, + 599 + ], + "spans": [ + { + "bbox": [ + 304, + 465, + 524, + 599 + ], + "type": "table", + "html": "
EducationNumber Of AnnotationsMinMax
Associate professor2605±47.5925162697
Bachelor's degree10510±84.791034810700
Doctoral degree305±18.83270332
High school2080±37.0120152139
Below high school165±11.17144184
Master's degree3515±48.0834253580
Other30±3.442536
Prefer not to say3690±52.9236033808
Professional degree380±17.87352411
College, no degree4665±71.3645394776
", + "image_path": "b1601f8aef7178a374f3956a192719185a06e4c50c72992aada3967166b2da4a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 608, + 525, + 668 + ], + "lines": [ + { + "bbox": [ + 302, + 608, + 525, + 668 + ], + "spans": [ + { + "bbox": [ + 302, + 608, + 525, + 668 + ], + "type": "text", + "content": "Table 5: Average, standard deviation, minimum and maximum of number of annotations per fold. All information given per group of gender, age, education and sexuality. Statistics are calculated across 12 folds from three runs of four-fold cross-validation." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "text", + "content": "1026" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 157, + 208, + 434, + 282 + ], + "blocks": [ + { + "bbox": [ + 157, + 208, + 434, + 282 + ], + "lines": [ + { + "bbox": [ + 157, + 208, + 434, + 282 + ], + "spans": [ + { + "bbox": [ + 157, + 208, + 434, + 282 + ], + "type": "table", + "html": "
GenderMajority BaselineBaselineSoc-Dem.Random
Female41.79±0.1262.23±0.5362.25±1.1962.41±0.92
Male40.53±0.1168.00±0.4967.66±0.4667.63±0.53
Nonbinary44.69±1.3956.33±6.0056.80±7.2458.00±7.49
Other45.50±4.6948.56±10.7850.53±14.6343.66±7.25
Prefer not to say41.05±0.3664.54±1.1365.05±1.5265.08±1.86
", + "image_path": "21a17ff1f2a07c1b08ed9915e4a34c9b984dad2935a8eb8a879ca50f559747f3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 157, + 290, + 430, + 385 + ], + "blocks": [ + { + "bbox": [ + 157, + 290, + 430, + 385 + ], + "lines": [ + { + "bbox": [ + 157, + 290, + 430, + 385 + ], + "spans": [ + { + "bbox": [ + 157, + 290, + 430, + 385 + ], + "type": "table", + "html": "
AgeMajority BaselineBaselineSoc-Dem.Random
18 - 2442.49±0.2859.39±1.5860.44±1.0560.52±1.37
25 - 3440.49±0.0966.72±0.5666.63±0.8366.92±0.51
35 - 4441.87±0.1564.50±0.5964.94±1.3365.24±0.89
45 - 5440.63±0.2665.68±0.6665.88±1.3965.98±0.83
55 - 6441.65±0.3964.37±1.2264.94±1.6664.84±1.30
65 or older41.46±0.5463.34±2.0764.70±2.2162.77±2.39
Prefer not to say41.37±0.3263.99±1.3265.24±1.1864.73±1.33
", + "image_path": "e0a9733f35c388dcd059b70d3d757e06b472b3a615fa43b0d42bf81291aafafc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 151, + 392, + 440, + 515 + ], + "blocks": [ + { + "bbox": [ + 151, + 392, + 440, + 515 + ], + "lines": [ + { + "bbox": [ + 151, + 392, + 440, + 515 + ], + "spans": [ + { + "bbox": [ + 151, + 392, + 440, + 515 + ], + "type": "table", + "html": "
EducationMajority BaselineBaselineSoc-Dem.Random
Associate degree43.16±0.1960.69±1.4460.54±2.3560.78±1.62
Bachelor's degree40.38±0.1066.16±0.5166.23±0.8266.80±0.54
Doctoral degree43.34±0.9461.93±3.8263.79±5.0363.27±3.67
High school43.02±0.2660.53±1.3960.47±2.2260.55±1.87
Below high school43.10±1.4458.28±4.6862.12±4.9060.17±4.25
Master's degree37.55±0.3269.71±0.8669.58±0.9369.45±0.96
Other42.95±2.3156.56±10.8857.59±9.8657.71±12.28
Prefer not to say40.97±0.2765.07±1.1665.69±1.0565.74±1.09
Professional degree40.43±0.8066.75±2.3767.84±3.3268.62±2.84
College, no degree43.61±0.1858.65±1.1959.40±1.7959.99±2.19
", + "image_path": "b02b56a006df2ea473f73cf387d47a11d02b3d9517886c04b4f75943232454e1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 161, + 523, + 430, + 597 + ], + "blocks": [ + { + "bbox": [ + 161, + 523, + 430, + 597 + ], + "lines": [ + { + "bbox": [ + 161, + 523, + 430, + 597 + ], + "spans": [ + { + "bbox": [ + 161, + 523, + 430, + 597 + ], + "type": "table", + "html": "
SexualityMajority BaselineBaselineSoc-Dem.Random
Bisexual34.69±0.5071.83±1.1471.42±1.5169.46±1.95
Heterosexual41.99±0.0663.25±0.3963.32±1.2163.82±0.55
Homosexual41.15±0.4164.43±1.7566.11±2.2065.12±1.94
Other43.53±0.7857.55±3.7960.57±4.5158.69±4.72
Prefer not to say39.12±0.2467.80±1.5667.27±1.5267.46±1.11
", + "image_path": "78c4e3df5636f674d9c687d72c3665a6ec9be5fa90bb8faeda716302a6b3b16c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 605, + 525, + 631 + ], + "lines": [ + { + "bbox": [ + 67, + 605, + 525, + 631 + ], + "spans": [ + { + "bbox": [ + 67, + 605, + 525, + 631 + ], + "type": "text", + "content": "Table 6: Average and standard deviation of macro " + }, + { + "bbox": [ + 67, + 605, + 525, + 631 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 67, + 605, + 525, + 631 + ], + "type": "text", + "content": " from three runs of four-fold stratified cross-validation. Separate table for each attribute. Bold results are the highest average per group. Full results including naive majority baseline" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1027" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "spans": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "type": "text", + "content": "A For every submission:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 106, + 414, + 242 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "spans": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "type": "text", + "content": "A1. Did you describe the limitations of your work? Limitations, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 142, + 329, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 142, + 329, + 169 + ], + "spans": [ + { + "bbox": [ + 77, + 142, + 329, + 169 + ], + "type": "text", + "content": "A2. Did you discuss any potential risks of your work? Ethics Statement, 9" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 178, + 414, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 178, + 414, + 205 + ], + "spans": [ + { + "bbox": [ + 77, + 178, + 414, + 205 + ], + "type": "text", + "content": "A3. Do the abstract and introduction summarize the paper's main claims?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 77, + 215, + 398, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 215, + 398, + 242 + ], + "spans": [ + { + "bbox": [ + 77, + 215, + 398, + 242 + ], + "type": "text", + "content": "A4. Have you used AI writing assistants when working on this paper? Left blank." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 68, + 252, + 290, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 252, + 290, + 266 + ], + "spans": [ + { + "bbox": [ + 68, + 252, + 290, + 266 + ], + "type": "text", + "content": "B Did you use or create scientific artifacts?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 79, + 270, + 151, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 270, + 151, + 283 + ], + "spans": [ + { + "bbox": [ + 79, + 270, + 151, + 283 + ], + "type": "text", + "content": "3, Appendix A.3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 77, + 291, + 524, + 634 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 77, + 291, + 315, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 291, + 315, + 319 + ], + "spans": [ + { + "bbox": [ + 77, + 291, + 315, + 319 + ], + "type": "text", + "content": "B1. Did you cite the creators of artifacts you used? 3, Appendix A.3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 77, + 328, + 463, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 328, + 463, + 355 + ], + "spans": [ + { + "bbox": [ + 77, + 328, + 463, + 355 + ], + "type": "text", + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Clear from context, citations" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 77, + 364, + 524, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 364, + 524, + 431 + ], + "spans": [ + { + "bbox": [ + 77, + 364, + 524, + 431 + ], + "type": "text", + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Clear from context, citations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 77, + 440, + 524, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 440, + 524, + 494 + ], + "spans": [ + { + "bbox": [ + 77, + 440, + 524, + 494 + ], + "type": "text", + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? 3, Ethics Statement 9" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 77, + 503, + 524, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 503, + 524, + 544 + ], + "spans": [ + { + "bbox": [ + 77, + 503, + 524, + 544 + ], + "type": "text", + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? 3, Appendix A.1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 77, + 553, + 524, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 553, + 524, + 634 + ], + "spans": [ + { + "bbox": [ + 77, + 553, + 524, + 634 + ], + "type": "text", + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. 3, 4, Appendix A.4" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "spans": [ + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "type": "text", + "content": "C Did you run computational experiments?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 79, + 663, + 87, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 663, + 87, + 672 + ], + "spans": [ + { + "bbox": [ + 79, + 663, + 87, + 672 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 77, + 682, + 524, + 724 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 682, + 524, + 724 + ], + "spans": [ + { + "bbox": [ + 77, + 682, + 524, + 724 + ], + "type": "text", + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Appendix A.3" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "spans": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "text", + "content": "ACL 2023 Responsible NLP Checklist" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "spans": [ + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "type": "text", + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1028" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 76, + 71, + 524, + 238 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 76, + 71, + 523, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 71, + 523, + 111 + ], + "spans": [ + { + "bbox": [ + 76, + 71, + 523, + 111 + ], + "type": "text", + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? No response." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 76, + 121, + 524, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 121, + 524, + 174 + ], + "spans": [ + { + "bbox": [ + 76, + 121, + 524, + 174 + ], + "type": "text", + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? No response." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 183, + 524, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 183, + 524, + 238 + ], + "spans": [ + { + "bbox": [ + 77, + 183, + 524, + 238 + ], + "type": "text", + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Appendix A.3" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 67, + 247, + 522, + 278 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 247, + 522, + 278 + ], + "spans": [ + { + "bbox": [ + 67, + 247, + 522, + 278 + ], + "type": "text", + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 76, + 287, + 524, + 539 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 76, + 287, + 524, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 287, + 524, + 326 + ], + "spans": [ + { + "bbox": [ + 76, + 287, + 524, + 326 + ], + "type": "text", + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 76, + 337, + 524, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 337, + 524, + 390 + ], + "spans": [ + { + "bbox": [ + 76, + 337, + 524, + 390 + ], + "type": "text", + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 76, + 400, + 524, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 400, + 524, + 454 + ], + "spans": [ + { + "bbox": [ + 76, + 400, + 524, + 454 + ], + "type": "text", + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 76, + 463, + 519, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 463, + 519, + 489 + ], + "spans": [ + { + "bbox": [ + 76, + 463, + 519, + 489 + ], + "type": "text", + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 76, + 498, + 524, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 498, + 524, + 539 + ], + "spans": [ + { + "bbox": [ + 76, + 498, + 524, + 539 + ], + "type": "text", + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1029" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/The Inside Story_ Towards Better Understanding of Machine Translation Neural Evaluation Metrics/d25ab395-c57a-4c68-8e6d-00772d415f43_content_list.json b/2023/The Inside Story_ Towards Better Understanding of Machine Translation Neural Evaluation Metrics/d25ab395-c57a-4c68-8e6d-00772d415f43_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ce6f7c0426f4fd4e83a83055fa209a80444d0a86 --- /dev/null +++ b/2023/The Inside Story_ Towards Better Understanding of Machine Translation Neural Evaluation Metrics/d25ab395-c57a-4c68-8e6d-00772d415f43_content_list.json @@ -0,0 +1,3220 @@ +[ + { + "type": "text", + "text": "The Inside Story: Towards Better Understanding of Machine Translation Neural Evaluation Metrics", + "text_level": 1, + "bbox": [ + 235, + 87, + 759, + 124 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ricardo Rei\\*1,2,4, Nuno M. Guerreiro\\*3,4, Marcos Treviso\\*3,4, Alon Lavie\\*1, Luisa Coheur\\*2,4, Andre F. T. Martins\\*1,3,4", + "bbox": [ + 245, + 141, + 759, + 174 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1Unbabel, Lisbon, Portugal, 2INESC-ID, Lisbon, Portugal \n3Instituto de Telecomunicações, Lisbon, Portugal", + "$^{4}$ Instituto Superior Técnico, University of Lisbon, Portugal" + ], + "bbox": [ + 260, + 175, + 739, + 225 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 252, + 339, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural metrics for machine translation evaluation, such as COMET, exhibit significant improvements in their correlation with human judgments compared to traditional metrics based on lexical overlap, such as BLEU. Yet neural metrics are, to a great extent, \"black boxes\" that return a single sentence-level score without transparency about the decision-making process. In this work, we develop and compare several neural explainability methods and demonstrate their effectiveness for interpreting state-of-the-art fine-tuned neural metrics. Our study reveals that these metrics leverage token-level information that can be directly attributed to translation errors, as assessed through comparison of token-level neural saliency maps with Multidimensional Quality Metrics (MQM) annotations and with synthetically-generated critical translation errors. To ease future research, we release our code at https://github.com/Unbabel/COMET/tree/explainable-metrics.", + "bbox": [ + 139, + 278, + 460, + 592 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 114, + 602, + 258, + 618 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reference-based neural metrics for machine translation evaluation are achieving evergrowing success, demonstrating superior results over traditional lexical overlap-based metrics, such as BLEU (Papineni et al., 2002) and CHRF (Popovic, 2015), in terms of both their correlation with human ratings and their robustness across diverse domains (Callison-Burch et al., 2006; Smith et al., 2016; Mathur et al., 2020; Kocmi et al., 2021; Freitag et al., 2022). However, lexical overlap-based metrics remain popular for evaluating the performance and progress of translation systems and algorithms. Concerns regarding trust and interpretability may help explain this (Leiter et al., 2022): contrary to traditional metrics, neural metrics are considered \"black boxes\" as they often use", + "bbox": [ + 112, + 627, + 489, + 885 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c81eeb83a2f01d61f4e4656faed8f88c417f549a3f6570b985690eea98bdeaf0.jpg", + "image_caption": [ + "Figure 1: Illustration of our approach. In this example, the metric assigns the translation a low score. We aim to better understand this sentence-level assessment by examining the correspondence between our token-level explanations and human annotated error spans." + ], + "image_footnote": [], + "bbox": [ + 515, + 250, + 867, + 373 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "increasingly large models (e.g., the winning metric of the WMT 22 Metrics shared task was a 10B parameter model (Freitag et al., 2022)).", + "bbox": [ + 507, + 483, + 880, + 532 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While some recent work has focused on explaining the predictions made by reference-free quality estimation (QE) systems (Fomicheva et al., 2021; Zerva et al., 2022), explaining reference-based metrics has remained a largely overlooked problem (Leiter et al., 2022). It is an open question whether the observations from studies of explainable QE carry over to this scenario. Thus, in this work, we fill that gap by turning to state-of-the-art reference-based metrics—we aim to interpret their decision-making process by exploiting the fact that these metrics show consistently good correlations with Multidimensional Quality Metrics (MQM) (Freitag et al., 2021, 2022; Sai et al., 2022), which are fine-grained quality assessments that result from experts identifying error spans in translation outputs (Lommel et al., 2014). We hypothesize that reference-based metrics leverage this token-level information to produce sentence-level scores. To test this hypothesis, we assess whether our explanations – measures of token-level importance obtained via attribution and input attribution methods such as attention weights and gradient scores (Treviso et al., 2021; Rei et al., 2022b) – align with", + "bbox": [ + 507, + 533, + 884, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution. Corresponding author: ricardo.rei@unbabel.com", + "bbox": [ + 112, + 892, + 487, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1089", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics", + "bbox": [ + 226, + 945, + 769, + 957 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Volume 2: Short Papers, pages 1089-1105", + "bbox": [ + 368, + 958, + 628, + 971 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "July 9-14, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 295, + 972, + 700, + 985 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "human-annotated spans (Fomicheva et al., 2021, 2022; Zerva et al., 2022), as illustrated in Figure 1.", + "bbox": [ + 112, + 84, + 489, + 115 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our analysis focuses on two main vectors: (i) understanding the impact of the reference information on the quality of the explanations; and (ii) finding whether the explanations can help to identify potential weaknesses in the metrics. Our main contributions are:", + "bbox": [ + 112, + 117, + 489, + 212 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We provide a comparison between multiple explainability methods for different metrics on all types of evaluation: src-only, ref-only, and src+ref joint evaluation.", + "- We find that explanations are related to the underlying metric architecture, and that leveraging reference information improves the explanations.", + "- We show that explanations for critical translation errors can reveal weaknesses in the metrics." + ], + "bbox": [ + 115, + 216, + 489, + 384 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Explaining Neural Metrics", + "text_level": 1, + "bbox": [ + 112, + 397, + 379, + 414 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We aim to explain sentence-level quality assessments of reference-based metrics by producing token-level explanations that align to translation errors. In what follows, we describe the metrics and how we produce the explanations that we study.", + "bbox": [ + 112, + 423, + 489, + 502 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 Metrics", + "text_level": 1, + "bbox": [ + 112, + 514, + 221, + 529 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We focus our analysis on two state-of-the-art neural metrics: COMET (Rei et al., 2020) and UNITE (Wan et al., 2022). While both metrics use a multilingual encoder model based on XLMR (Conneau et al., 2020), they employ distinct strategies to obtain sentence-level quality scores. On the one hand, COMET separately encodes the source, translation and reference to obtain their respective sentence embeddings; these embeddings are then combined to compute a quality score. On the other, UNITE jointly encodes the sentences to compute a contextualized representation that is subsequently used to compute the quality score. Interestingly, UNITE is trained to obtain quality scores for different input combinations: [mt; src] (SRC), [mt; ref] (REF), and [mt; src; ref] (SRC+REF). In fact, when the input is SRC, UNITE works like TransQuest (Ranasinghe et al., 2020); REF, like BLEURT (Sellam et al., 2020); and SRC+REF, like ROBLEURT (Wan et al., 2021).", + "bbox": [ + 112, + 536, + 489, + 858 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2 Explanations via Attribution Methods", + "text_level": 1, + "bbox": [ + 507, + 84, + 853, + 99 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we produce explanations using attribution methods that assign a scalar value to each translation token (i.e. a token-level attribution) to represent its importance. While many input attribution methods exist and have been extensively studied in the literature (Ribeiro et al., 2016; Shrikumar et al., 2017; Sundararajan et al., 2017; Jain and Wallace, 2019; Atanasova et al., 2020; Zaman and Belinkov, 2022), we focus specifically on those that have been demonstrated to be effective for explaining the predictions of QE models (Treviso et al., 2021; Fomicheva et al., 2022; Fernandes et al., 2022; Zerva et al., 2022) and extend them to our reference-based scenario. Concretely, we use the following techniques to extract explanations:", + "bbox": [ + 507, + 105, + 884, + 347 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- embed-align: the maximum cosine similarity between each translation token embedding and the reference and/or source token embeddings (Tao et al., 2022);", + "- grad $\\ell_2$ : the $\\ell_2$ -norm of gradients with respect to the word embeddings of the translation tokens (Arras et al., 2019);", + "- attention: the attention weights of the translation tokens for each attention head of the encoder (Treviso et al., 2021);", + "- $\\mathbf{attn} \\times \\mathbf{grad}$ : the attention weights of each head scaled by the $\\ell_2$ -norm of the gradients of the value vectors of that head (Rei et al., 2022b)." + ], + "bbox": [ + 510, + 357, + 884, + 600 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3 Experimental Setting", + "text_level": 1, + "bbox": [ + 507, + 625, + 727, + 640 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "MQM annotations. We use MQM annotations from the WMT 2021 Metrics shared task (Freitag et al., 2021),3 covering three language pairs — English-German (en→de), English-Russian (en→ru), and Chinese-English (zh→en) —in two different domains: News and TED Talks. For each incorrect translation, human experts marked the corresponding error spans. In our framework, these error spans should align with the words that the attribution methods assign higher importance to.", + "bbox": [ + 505, + 650, + 882, + 810 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "2For all attention-based methods, we ensemble the explanations from the top 5 heads as this has shown to improve performance consistently over selecting just the best head (Treviso et al., 2021; Rei et al., 2022b). Moreover, we use the full attention matrix, instead of relying only on cross attention information.", + "bbox": [ + 507, + 820, + 884, + 891 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "3https://github.com/google/wmt-mqm-human-evaluation", + "bbox": [ + 509, + 892, + 752, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "Ensembles composed of these two metrics were respectively ranked second and third in WMT 2022 Metrics shared task. The winner of WMT 2022 Metrics task — METRICXXL — is not publicly available (Freitag et al., 2022).", + "bbox": [ + 112, + 869, + 489, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1090", + "bbox": [ + 482, + 928, + 521, + 940 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/278a1cc6bcc5d3756ab7097a5c32d3bd20bb51325d4cb0c0e5041f2fe5e733c3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
METRICEXPLAINABILITY METHODen→dezh→enen→ruAvg.
AUCR@KAUCR@KAUCR@KAUCR@K
src-only* evaluation
UNITE SRCembed-align[mt, src]0.5870.3390.6440.2810.5830.1670.6040.262
grad ℓ20.5720.2930.5350.2000.6200.1690.5760.221
attention0.6360.3220.6120.2530.6120.1890.6200.254
attn × grad0.7070.3760.6390.2940.6330.2110.6600.294
ref-only evaluation
UNITE REFembed-align[mt, ref]0.6580.3960.6670.3280.6350.2180.6530.314
grad ℓ20.5960.3190.5710.2600.6610.2020.6090.261
attention0.6370.3440.6700.3350.6520.2240.6530.301
attn × grad0.7250.4250.6670.3800.6600.2480.6840.351
src, ref joint evaluation
UNITE SRC+REFembed-align[mt, src; ref]0.6500.3830.6700.3300.6180.2130.6460.309
grad ℓ20.5950.3250.5790.2570.6430.1910.6060.257
attention0.6570.4210.6700.3830.6490.2230.6590.342
attn × grad0.7360.4210.6740.3830.6710.2480.6930.351
COMETembed-align[mt, src]0.5900.3710.6740.3140.5770.2200.6140.301
embed-align[mt, ref]0.6940.4250.6960.3550.6470.2750.6790.352
embed-align[mt, src; ref]0.6880.4160.6970.3570.6220.2790.6690.350
grad ℓ20.6030.3120.5400.2520.6040.1850.5820.250
attention0.6040.3510.5920.2590.6330.2090.6080.268
attn × grad0.7100.3650.6330.2780.6620.2440.6690.295
", + "bbox": [ + 181, + 80, + 818, + 385 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1: AUC and Recall@K of explanations obtained via different attribution methods for COMET and UNITE models on the MQM data. *Although UNITE SRC is a src-only evaluation metric, it was trained with reference information (Wan et al., 2022).", + "bbox": [ + 112, + 394, + 882, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Models. For COMET, we use the latest publicly available model: wmt22-comet-da (Rei et al., 2022a). For UNITE, we train our own model using the same data used to train COMET in order to have a comparable setup. We provide full details (training data, correlations with human annotations, and hyperparameters) in Appendix A. Overall, the resulting reference-based UNITE models (REF and SRC+REF) are on par with COMET.", + "bbox": [ + 112, + 457, + 487, + 602 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Evaluation. We want our explanations to be directly attributed to the annotated error spans, in the style of an error detection task. Thus, we report Area Under Curve (AUC) and Recall@K.6 These metrics have been used as the main metrics in previous works on explainable QE (Fomicheva et al., 2021, 2022; Zerva et al., 2022).", + "bbox": [ + 112, + 613, + 489, + 725 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4 Results", + "text_level": 1, + "bbox": [ + 112, + 739, + 213, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1 High-level analysis", + "text_level": 1, + "bbox": [ + 112, + 766, + 307, + 782 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Explanations are tightly related to the underlying metric architecture. The results in Ta-", + "text_level": 1, + "bbox": [ + 112, + 789, + 489, + 820 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ble 1 show that the predictive power of the attribution methods differ between UNITE and COMET: attn $\\times$ grad is the best method for UNITE-based models, while embed-align works best for COMET. This is expected as UNITE constructs a joint representation for the input sentences, thus allowing attention to flow across them; COMET, in contrast, encodes the sentences separately, so it relies heavily on the separate contextualized embeddings that are subsequently combined via elementwise operations such as multiplication and absolute difference. Interestingly, embed-align and attn $\\times$ grad were the winning explainability approaches of the WMT 2022 Shared-Task on Quality Estimation (Zerva et al., 2022). This suggests that explainability methods developed for QE systems can translate well to reference-based metrics. We provide examples of explanations in Appendix C.", + "bbox": [ + 507, + 457, + 884, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reference information boosts explainability power. Table 1 also shows that, across all metrics, using reference information brings substantial improvements over using only the source information. Moreover, while reference-based attributions significantly outperform source-based attributions, combining the source and reference information to", + "bbox": [ + 507, + 758, + 882, + 869 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "In Appendix B, we provide a comparison between the explanations obtained via embed-align with COMET and with its pretrained encoder model, XLM-R.", + "bbox": [ + 507, + 879, + 882, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "4https://huggingface.co/Un babel/wmt22-comet-da", + "bbox": [ + 112, + 831, + 400, + 854 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "5Our implementation differs from the original work by Wan et al. (2022), See Appendix A for full details.", + "bbox": [ + 112, + 856, + 485, + 879 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "In this setup, Recall@K is the proportion of words with the highest attribution that correspond to translation errors against the total number of errors in the annotated error span.", + "bbox": [ + 112, + 881, + 485, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1091", + "bbox": [ + 482, + 928, + 517, + 940 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0683cceb75a1ceb80cc27a61b4f761c4802d986859633f65fa8e3a4a5e01adbc.jpg", + "image_caption": [ + "Figure 2: Performance of the best attribution methods for COMET, UNITE REF and UNITE SRC+REF in terms of Recall@K on translations with critical errors: negations (NEG), hallucinations (HALL), named entity errors (NE), and errors in numbers (NUM)." + ], + "image_footnote": [], + "bbox": [ + 117, + 80, + 487, + 195 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "obtain token-level attributions does not consistently yield superior results over using the reference alone. Notably, the best attribution method for COMET does not require any source information. This is interesting: in some cases, reference-based metrics may largely ignore source information, relying heavily on the reference instead.", + "bbox": [ + 112, + 296, + 489, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2 How do the explanations fare for critical translation errors?", + "text_level": 1, + "bbox": [ + 112, + 420, + 478, + 451 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The MQM data analyzed until now consists primarily of high quality translations, with the majority of annotated errors being non-critical. However, it is important to assess whether our explanations can be accurately attributed to critical errors, as this may reveal potential metric shortcomings. To this end, we employ SMAUG (Alves et al., 2022)8, a tool designed to generate synthetic data for stress-testing metrics, to create corrupted translations that contain critical errors. Concretely, we generate translations with the following pathologies: negation errors, hallucinations via insertions, named entity errors, and errors in numbers.9", + "bbox": [ + 112, + 458, + 489, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Explanations identify critical errors more easily than non-critical errors. Figure 2 shows that explanations are more effective in identifying critical errors compared to other non-critical errors (see Table 1). Specifically, we find significant performance improvements up to nearly $30\\%$ in Recall@K for certain critical errors. Overall, hallucinations are the easiest errors to identify across all neural metrics. This suggests that neural", + "bbox": [ + 112, + 677, + 489, + 822 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "metrics appropriately identify and penalize hallucinated translations, which aligns with the findings of Guerreiro et al. (2022). Moreover, explanations for both UNITE models behave similarly for all errors except numbers, where the source information plays a key role in improving the explanations. Notably, contrary to what we observed for data with non-critical errors, COMET explanations are less effective than those of UNITE REF and UNITE SRC+REF for identifying critical errors.", + "bbox": [ + 507, + 84, + 884, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Explanations can reveal potential metric weaknesses. Figure 2 suggests that COMET explanations struggle to identify localized errors (negation errors, named entity errors and discrepancies in numbers). We hypothesize that this behavior is related to the underlying architecture. Unlike UNITE-based metrics, COMET does not rely on soft alignments via attention between the sentences in the encoding process. This process may be key to identify local misalignments during the encoding process. In fact, the attention-based attributions for UNITE metrics can more easily identify these errors. COMET, however, encodes the sentences separately, which may result in grammatical features (e.g. numbers) being encoded similarly across sentences (Chi et al., 2020; Chang et al., 2022). As such, explanations obtained via embedding alignments will not properly identify these misalignments on similar features. Importantly, these findings align with observations made in (Amrhein and Sennrich, 2022; Raunak et al., 2022). This showcases how explanations can be used to diagnose and reveal shortcomings of neural-based metrics.", + "bbox": [ + 507, + 254, + 884, + 624 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5 Conclusions and Future Work", + "text_level": 1, + "bbox": [ + 507, + 636, + 803, + 651 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this paper, we investigated the use of explainability methods to better understand widely used neural metrics for machine translation evaluation, such as COMET and UNITE. Concretely, we analyzed how explanations are impacted by the reference information, and how they can be used to reveal weaknesses of these metrics. Our analysis shows that the quality of the explanations is tightly related to the underlying metric architecture. Interestingly, we also provide evidence that neural metrics like COMET may heavily rely on reference information over source information. Additionally, we show that explanations can be used to reveal reference-based metrics weaknesses such as failing to severely penalize localized critical errors. This opens up promising opportunities for future", + "bbox": [ + 507, + 661, + 884, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "8https://github.com/Unbabel/smaug", + "bbox": [ + 141, + 832, + 406, + 846 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "9We corrupt fully correct translations that are not an exact copy of the reference translation. Moreover, as the full suit of SMAUG transformations can only be applied to English data, we focus solely on zh→en translations. Overall, the synthetic dataset consists of 2610 translations. Full statistics about the corrupted data and examples are shown in Appendix A.2.", + "bbox": [ + 115, + 846, + 487, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1092", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "research on leveraging explanations to diagnose reference-based metrics errors. To support these studies, we call for future datasets illustrating critical errors (e.g., challenge sets (Karpinska et al., 2022)) to be accompanied by annotated error spans.", + "bbox": [ + 112, + 84, + 490, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Limitations", + "text_level": 1, + "bbox": [ + 114, + 177, + 220, + 192 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We highlight three main limitations of our work.", + "bbox": [ + 112, + 203, + 475, + 218 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "First, although we have explored gradient-based explanations that take the whole network into consideration and have been shown to be faithful in previous work (Bastings et al., 2021), we do not explicitly explore how COMET combines the sentence representations in the feed-forward that precedes the encoder model to produce the sentence-level score.", + "bbox": [ + 112, + 219, + 489, + 344 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Second, we have shown that combining attention with gradient information results in the best explanations for UNITE-based metrics. However, from a practical standpoint, running inference and extracting the explainability scores simultaneously may be more computationally expensive than other techniques: gradient-based metrics benefit from GPU infrastructure and require storing all gradient information.", + "bbox": [ + 112, + 348, + 489, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Third, we have not explored extracting explanations in low-resource settings. That is because high-quality MQM annotations for such language pairs are not yet available. Nevertheless, further research in those settings is needed to access the broader validity of our claims.", + "bbox": [ + 112, + 493, + 489, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 114, + 602, + 285, + 618 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This work was partially supported by the P2020 programs (MAIA, contract 045909), the Portuguese Recovery and Resilience Plan (PRR) through project C645008882-00000055, Center for Responsible AI, by the European Research Council (ERC StG DeepSPIN, 758969), by EU's Horizon Europe Research and Innovation Actions (UTTER, contract 101070631), and by the Fundação para a Ciência e Tecnologia (contracts UIDB/50021/2020 and UIDB/50008/2020).", + "bbox": [ + 112, + 627, + 489, + 788 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 114, + 816, + 213, + 831 + ], + "page_idx": 4 + }, + { + "type": "ref_text", + "text": "Duarte Alves, Ricardo Rei, Ana C Farinha, José G. C. de Souza, and André F. T. Martins. 2022. Robust MT Evaluation with Sentence-level Multilingual Augmentation. In Proceedings of the Seventh Conference on Machine Translation, pages 469-478, Abu Dhabi. Association for Computational Linguistics.", + "bbox": [ + 115, + 838, + 489, + 917 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Chantal Amrhein and Rico Sennrich. 2022. Identifying Weaknesses in Machine Translation Metrics Through Minimum Bayes Risk Decoding: A Case Study for COMET. In Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1125-1141, Online only. Association for Computational Linguistics.", + "Leila Arras, Ahmed Osman, Klaus-Robert Müller, and Wojciech Samek. 2019. Evaluating recurrent neural network explanations. In Proceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 113-126, Florence, Italy. Association for Computational Linguistics.", + "Pepa Atanasova, Jakob Grue Simonsen, Christina Lioma, and Isabelle Augenstein. 2020. A diagnostic study of explainability techniques for text classification. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3256-3274, Online. Association for Computational Linguistics.", + "Jasmijn Bastings, Sebastian Ebert, Polina Zablotskaia, Anders Sandholm, and Katja Filippova. 2021. \"will you find these shortcuts?\" a protocol for evaluating the faithfulness of input salience methods for text classification.", + "Chris Callison-Burch, Miles Osborne, and Philipp Koehn. 2006. Re-evaluating the role of Bleu in machine translation research. In 11th Conference of the European Chapter of the Association for Computational Linguistics, pages 249-256, Trento, Italy. Association for Computational Linguistics.", + "Tyler A. Chang, Zhuowen Tu, and Benjamin K. Bergen. 2022. The geometry of multilingual language model representations.", + "Ethan A. Chi, John Hewitt, and Christopher D. Manning. 2020. Finding universal grammatical relations in multilingual BERT. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5564-5577, Online. Association for Computational Linguistics.", + "Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettle-moyer, and Veselin Stoyanov. 2020. Unsupervised cross-lingual representation learning at scale. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8440-8451, Online. Association for Computational Linguistics.", + "Daniel Deutsch, Rotem Dror, and Dan Roth. 2021. A statistical analysis of summarization evaluation metrics using resampling methods. Transactions of the Association for Computational Linguistics, 9:1132-1146." + ], + "bbox": [ + 509, + 85, + 884, + 917 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1093", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Patrick Fernandes, Marcos Treviso, Danish Pruthi, André F. T. Martins, and Graham Neubig. 2022. Learning to scaffold: Optimizing model explanations for teaching.", + "Marina Fomicheva, Piyawat Lertvittayakumjorn, Wei Zhao, Steffen Eger, and Yang Gao. 2021. The Eval4NLP shared task on explainable quality estimation: Overview and results. In Proceedings of the 2nd Workshop on Evaluation and Comparison of NLP Systems, pages 165-178, Punta Cana, Dominican Republic. Association for Computational Linguistics.", + "Marina Fomicheva, Lucia Specia, and Nikolaos Aletras. 2022. Translation error detection as rationale extraction. In Findings of the Association for Computational Linguistics: ACL 2022, pages 4148-4159, Dublin, Ireland. Association for Computational Linguistics.", + "Markus Freitag, Ricardo Rei, Nitika Mathur, Chi-kiu Lo, Craig Stewart, Eleftherios Avramidis, Tom Kocmi, George Foster, Alon Lavie, and André F. T. Martins. 2022. Results of WMT22 Metrics Shared Task: Stop Using BLEU – Neural Metrics Are Better and More Robust. In Proceedings of the Seventh Conference on Machine Translation, pages 46–68, Abu Dhabi. Association for Computational Linguistics.", + "Markus Freitag, Ricardo Rei, Nitika Mathur, Chi-kiu Lo, Craig Stewart, George Foster, Alon Lavie, and Ondrej Bojar. 2021. Results of the WMT21 metrics shared task: Evaluating metrics with expert-based human evaluations on TED and news domain. In Proceedings of the Sixth Conference on Machine Translation, pages 733-774, Online. Association for Computational Linguistics.", + "Nuno M. Guerreiro, Elena Voita, and André F. T. Martins. 2022. Looking for a needle in a haystack: A comprehensive study of hallucinations in neural machine translation.", + "Sarthak Jain and Byron C. Wallace. 2019. Attention is not Explanation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 3543-3556, Minneapolis, Minnesota. Association for Computational Linguistics.", + "Marzena Karpinska, Nishant Raj, Katherine Thai, Yixiao Song, Ankita Gupta, and Mohit Iyyer. 2022. Demetr: Diagnosing evaluation metrics for translation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, page 9540-9561, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.", + "Tom Kocmi, Christian Federmann, Roman Grundkiewicz, Marcin Junczys-Dowmunt, Hitokazu Matsushita, and Arul Menezes. 2021. To ship or not to ship: An extensive evaluation of automatic metrics for machine translation. In Proceedings of the Sixth" + ], + "bbox": [ + 115, + 85, + 489, + 917 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Machine Translation, pages 478-494, Online. Association for Computational Linguistics.", + "Christoph Leiter, Piyawat Lertvittayakumjorn, Marina Fomicheva, Wei Zhao, Yang Gao, and Steffen Eger. 2022. Towards explainable evaluation metrics for natural language generation.", + "Arle Lommel, Hans Uszkoreit, and Aljoscha Burchardt. 2014. Multidimensional Quality Metrics (MQM): A Framework for Declaring and Describing Translation Quality Metrics. Tradumàtica, pages 0455-463.", + "Nitika Mathur, Timothy Baldwin, and Trevor Cohn. 2020. Tangled up in BLEU: Reevaluating the evaluation of automatic machine translation evaluation metrics. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4984-4997, Online. Association for Computational Linguistics.", + "Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311-318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics.", + "Maja Popovic. 2015. *chrF: character n-gram F-score* for automatic MT evaluation. In *Proceedings of the Tenth Workshop on Statistical Machine Translation*, pages 392–395, Lisbon, Portugal. Association for Computational Linguistics.", + "Tharindu Ranasinghe, Constantin Orasan, and Ruslan Mitkov. 2020. *TransQuest: Translation Quality Estimation with Cross-lingual Transformers*. In *Proceedings of the 28th International Conference on Computational Linguistics*, pages 5070–5081, Barcelona, Spain (Online). International Committee on Computational Linguistics.", + "Vikas Raunak, Matt Post, and Arul Menezes. 2022. Salted: A framework for salient long-tail translation error detection.", + "Ricardo Rei, José G. C. de Souza, Duarte Alves, Chrysoula Zerva, Ana C Farinha, Taisiya Glushkova, Alon Lavie, Luisa Coheur, and André F. T. Martins. 2022a. COMET-22: Unbabel-IST 2022 Submission for the Metrics Shared Task. In Proceedings of the Seventh Conference on Machine Translation, pages 578-585, Abu Dhabi. Association for Computational Linguistics.", + "Ricardo Rei, Craig Stewart, Ana C Farinha, and Alon Lavie. 2020. COMET: A neural framework for MT evaluation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2685-2702, Online. Association for Computational Linguistics.", + "Ricardo Rei, Marcos Treviso, Nuno M. Guerreiro, Chrysoula Zerva, Ana C Farinha, Christine Maroti, José G. C. de Souza, Taisiya Glushkova, Duarte" + ], + "bbox": [ + 510, + 85, + 882, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1094", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Alves, Luisa Coheur, Alon Lavie, and Andre F. T. Martins. 2022b. CometKiwi: IST-Unbabel 2022 Submission for the Quality Estimation Shared Task. In Proceedings of the Seventh Conference on Machine Translation, pages 634-645, Abu Dhabi. Association for Computational Linguistics.", + "Marco Ribeiro, Sameer Singh, and Carlos Guestrin. 2016. \"why should I trust you?\": Explaining the predictions of any classifier. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Demonstrations, pages 97-101, San Diego, California. Association for Computational Linguistics.", + "Ananya B. Sai, Vignesh Nagarajan, Tanay Dixit, Raj Dabre, Anoop Kunchukuttan, Pratyush Kumar, and Mitesh M. Khapra. 2022. IndicMT Eval: A Dataset to Meta-Evaluate Machine Translation metrics for Indian Languages.", + "Thibault Sellam, Dipanjan Das, and Ankur Parikh. 2020. BLEURT: Learning robust metrics for text generation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7881-7892, Online. Association for Computational Linguistics.", + "Avanti Shrikumar, Peyton Greenside, and Anshul Kundaje. 2017. Learning Important Features Through Propagating Activation Differences. In Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 3145-3153. PMLR.", + "Aaron Smith, Christian Hardmeier, and Joerg Tiedemann. 2016. Climbing mont BLEU: The strange world of reachable high-BLEU translations. In Proceedings of the 19th Annual Conference of the European Association for Machine Translation, pages 269-281.", + "Mukund Sundararajan, Ankur Taly, and Qiqi Yan. 2017. Axiomatic Attribution for Deep Networks. In Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 3319-3328. PMLR.", + "Shimin Tao, Su Chang, Ma Miaomiao, Hao Yang, Xiang Geng, Shujian Huang, Min Zhang, Jiaxin Guo, Minghan Wang, and Yinglu Li. 2022. CrossQE: HW-TSC 2022 Submission for the Quality Estimation Shared Task. In Proceedings of the Seventh Conference on Machine Translation, pages 646-652, Abu Dhabi. Association for Computational Linguistics.", + "Marcos Treviso, Nuno M. Guerreiro, Ricardo Rei, and Andre F. T. Martins. 2021. IST-unbabel 2021 submission for the explainable quality estimation shared task. In Proceedings of the 2nd Workshop on Evaluation and Comparison of NLP Systems, pages 133-145, Punta Cana, Dominican Republic. Association for Computational Linguistics." + ], + "bbox": [ + 115, + 85, + 485, + 917 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Yu Wan, Dayiheng Liu, Baosong Yang, Tianchi Bi, Haibo Zhang, Boxing Chen, Weihua Luo, Derek F. Wong, and Lidia S. Chao. 2021. RoBLEURT submission for WMT2021 metrics task. In Proceedings of the Sixth Conference on Machine Translation, pages 1053-1058, Online. Association for Computational Linguistics.", + "Yu Wan, Dayiheng Liu, Baosong Yang, Haibo Zhang, Boxing Chen, Derek Wong, and Lidia Chao. 2022. UniTE: Unified translation evaluation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8117-8127, Dublin, Ireland. Association for Computational Linguistics.", + "Kerem Zaman and Yonatan Belinkov. 2022. A Multilingual Perspective Towards the Evaluation of Attribution Methods in Natural Language Inference.", + "Chrysoula Zerva, Frédéric Blain, Ricardo Rei, Piyawat Lertvittayakumjorn, José G. C. de Souza, Steffen Eger, Diptesh Kanojia, Duarte Alves, Constantin Orasan, Marina Fomicheva, André F. T. Martins, and Lucia Specia. 2022. Findings of the WMT 2022 Shared Task on Quality Estimation. In Proceedings of the Seventh Conference on Machine Translation, pages 69-99, Abu Dhabi. Association for Computational Linguistics." + ], + "bbox": [ + 510, + 85, + 880, + 456 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1095", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A Model Details", + "text_level": 1, + "bbox": [ + 114, + 84, + 273, + 98 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Section 2.1, we employed the latest publicly available model (wmt22-comet-da) for COMET, which emerged as a top-performing metric in the WMT 2022 Metrics task (Freitag et al., 2022). To ensure a comparable setting for UNITE (Wan et al., 2022), we trained our own model. In doing so, we utilized the same data employed in the development of the COMET model by (Rei et al., 2022a), without pretraining any synthetic data, as originally suggested. Additionally, our implementation did not incorporate monotonic regional attention, as our preliminary experiments revealed no discernible benefits from its usage. The hyperparameters used are summarized in Table 3, while Table 4 presents the number of Direct Assessments utilized during training. Furthermore, Table 5 displays the segment-level correlations with WMT 2021 MQM data for the News and TED domains.", + "bbox": [ + 115, + 109, + 489, + 397 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Regarding computational infrastructure, a single NVIDIA A10G GPU with 23GB memory was used. The resulting UNITE model has 565M parameters while COMET has 581M parameters.", + "bbox": [ + 112, + 399, + 489, + 463 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A.1 Output Distribution", + "text_level": 1, + "bbox": [ + 112, + 475, + 321, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To better understand the output of the models and what scores are deemed low, we plotted the output distributions for the two models we used in our study. The average score for English $\\rightarrow$ German data is 0.856 for the COMET model and 0.870 for the UNITE model we trained. From Figure 3 we can observe the distribution of scores. This means that the 0.6692 score from the example in Figure 1 corresponds to a low quality output (5th percentile).", + "bbox": [ + 112, + 495, + 489, + 640 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A.2 SMAUG Corpus", + "text_level": 1, + "bbox": [ + 112, + 651, + 294, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As we have seen in Section 4.2, we have created synthetic translation errors for the following pathologies: negation errors, hallucinations via insertions, named entity errors, and errors in numbers. Table 7 presents a summary of the examples created using SMAUG and in Table 8 we show examples of each error category.", + "bbox": [ + 112, + 671, + 489, + 784 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "B Comparison between COMET and XLM-R Alignments", + "text_level": 1, + "bbox": [ + 112, + 796, + 445, + 829 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "From Table 1, it is evident that the alignments between the reference and/or source and the translation yield effective explanations for COMET. This raises the question of how these alignments compare to the underlying encoder of COMET before", + "bbox": [ + 112, + 839, + 489, + 917 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the fine-tuning process with human annotations. To investigate this, we examine the results for XLM-R without any fine-tuning, as presented in Table 2.", + "bbox": [ + 507, + 84, + 880, + 131 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Overall, the explanations derived from the alignments of COMET prove to be more predictive of error spans than those obtained from XLM-R alignments. This suggests that during the fine-tuning phase, COMET models modify the underlying XLM-R representations to achieve better alignment with translation errors.", + "bbox": [ + 507, + 133, + 884, + 243 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "C Examples", + "text_level": 1, + "bbox": [ + 507, + 256, + 633, + 273 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Tables 9 and 10, we show examples of COMET explanations for Chinese $\\rightarrow$ English and English $\\rightarrow$ German language pairs, respectively. We highlight in gray the corresponding MQM annotation performed by an expert linguist and we sort the examples from highest to lowest COMET scores. From these examples we can observe the following:", + "bbox": [ + 507, + 282, + 882, + 395 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "- Highlights provided by COMET explanations have a high recall with human annotations. In all examples, subword tokens corresponding to translation errors are highlighted in red but we often see that not everything is incorrect.", + "bbox": [ + 507, + 406, + 880, + 485 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "- Explanations are consistent with scores. For example, in the third example from Table 10, the red highlights do not correspond to errors and in fact the translation only has a major error griffen. Nonetheless, the score assigned by COMET is a low score of 0.68 which is faithful to the explanations that was given even if the assessment does not agree with human experts.", + "bbox": [ + 507, + 497, + 882, + 625 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1096", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5deaf02c57aa10bc91a1fe7afb1edd9323f633af8a7892b568eaa348d146d7c0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
METRICEXPLAINABILITY METHODen→dezh→enen→ruAvg.
AUCR@KAUCR@KAUCR@KAUCR@K
XLM-Rembed-align[mt, src]0.5870.3590.6680.3110.5760.1990.6100.289
embed-align[mt, ref]0.6710.4050.6890.3450.6340.2440.6640.331
embed-align[mt, src; ref]0.6660.3950.6900.3470.6160.2420.6570.328
COMETembed-align[mt, src]0.5900.3710.6740.3140.5770.2200.6140.301
embed-align[mt, ref]0.6940.4250.6960.3550.6470.2750.6790.352
embed-align[mt, src; ref]0.6880.4160.6970.3570.6220.2790.6690.350
", + "bbox": [ + 184, + 80, + 811, + 199 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/0cbf1b21ea95c671b3079594768b57863e608636c78f31c584e9f913777664e5.jpg", + "table_caption": [ + "Table 2: AUC and Recall@K of explanations obtained via alignments for COMET and XLM-R without any further fine-tuning on human annotations." + ], + "table_footnote": [], + "table_body": "
HyperparameterUNITECOMET
Encoder ModelXLM-R (large)
OptimizerAdamW
No. frozen epochs0.3
Learning rate (LR)1.5e-05
Encoder LR.1.0e-06
Layerwise Decay0.95
Batch size16
Loss functionMSE
Dropout0.1
Hidden sizes[3072, 1024]
Embedding layerFrozen
FP precision16
No. Epochs12
", + "bbox": [ + 164, + 451, + 436, + 634 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/dec55668b4599bce00e1410c924ed299a4a47f11abb1e74e2ed7db35ceb50d18.jpg", + "table_caption": [ + "Table 3: Hyperparameters used to train UNITE and COMET checkpoints used in this work. The only difference between the two is the number of training epochs due to the fact that, for UNITE, the best validation checkpoint is the first one." + ], + "table_footnote": [], + "table_body": "
Language PairSIZE
zh-en126947
en-de121420
de-en99183
en-zh90805
ru-en79280
en-ru62749
en-CS60937
fi-en46145
en-fi34335
tr-en30186
et-en29496
cs-en27847
en-mr26000
de-CS13804
en-et13376
pl-en11816
en-pl10572
lt-en10315
en-ja9578
gu-en9063
si-en9000
ro-en9000
ne-en9000
en-lt8959
ja-en8939
en-kk8219
en-ta7890
ta-en7577
en-gu6924
kk-en6789
de-fr6691
en-lv5810
en-tr5171
km-en4722
ps-en4611
fr-de3999
Total1027155
", + "bbox": [ + 600, + 325, + 793, + 788 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 4: Number of direct assessments per language pair used to train COMET (Rei et al., 2022a) and the UNITE model used in this work.", + "bbox": [ + 507, + 799, + 880, + 841 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1097", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/81d67d2e1c777b6fda5691781782228f7808597ccd85601e5940c69164d79852.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 117, + 158, + 371, + 294 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/6537587ec22518537216de18a08f3630e956f491abbef7698b8fe26ab93de066.jpg", + "image_caption": [ + "(a)COMET" + ], + "image_footnote": [], + "bbox": [ + 371, + 158, + 623, + 294 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/7f12e5fb24bafeccede2948320d364ece3f8a22a64f142b40f6f3c81028c9be2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 623, + 158, + 875, + 294 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/3e79cca46a0be2cd78a78a86a54cbab459146283d1cd328d1bf3942c7e0b1b2d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 117, + 324, + 371, + 460 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/9894dcd3aa7a90714367ee71223d1e29b6dc2fe1264af58fa21b90c378d8648b.jpg", + "image_caption": [ + "(b) UNITE SRC" + ], + "image_footnote": [], + "bbox": [ + 371, + 324, + 621, + 460 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/ac54ae4e2d1b3fddfb076b8e1840a50715d1ccd9029eac80c23417c604b171c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 623, + 324, + 875, + 460 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/6d21005af26e50d9d8e404c809c12ef33821201eb759ff136e58f4816fcf1134.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 119, + 489, + 371, + 626 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/29fe38098bf71cc10cf0ff1d8da33f403b880c10f40eba4e14304dc71215a2a7.jpg", + "image_caption": [ + "(c) UNITE REF" + ], + "image_footnote": [], + "bbox": [ + 371, + 489, + 621, + 626 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/f70cd84ebf577212a3b3c178be04701601dd5f362429892d4029eade3c8b3ba8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 623, + 489, + 873, + 626 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/e955daaa78864d91e6fdc0a5108bac795d99971ab97f7ec85f7b85c91433dcb8.jpg", + "image_caption": [ + "Figure 3: Distribution of scores for all metrics obtained on the MQM data (for all language pairs)." + ], + "image_footnote": [], + "bbox": [ + 119, + 657, + 371, + 791 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/9c020dee88cd623534f02d3180f8840f6d305d9394786d788b40554663a6624f.jpg", + "image_caption": [ + "(d) UNITE SRC+REF" + ], + "image_footnote": [], + "bbox": [ + 371, + 657, + 621, + 791 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/0921c070257ccf3a1203c6298b58fc5ee2d32d96b279faaac8e3a4c0561808c7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 623, + 657, + 873, + 791 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1098", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/610f05d45f03472047468581cd4ea495d2eb518bbebdcea25940166bd75c6c34.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BLEUCHRFYISI-1BLEURTUNITE\nSRCUNITE\nREFUNITE\nSRC+REFCOMET\nwmt22-comet-da
EN→DENEWSρ0.0770.0920.1630.3070.2740.3210.304
τ0.0690.0920.1440.2400.2220.2480.241
ρ0.1510.1580.2360.3250.3110.3350.338
τ0.1130.1460.2120.2830.2640.3010.298
EN→RUTED Newsρ0.1530.2520.2630.3590.3330.3910.382
τ0.1060.1780.2160.2760.2760.2980.297
ρ0.1540.2680.2350.2860.2390.2890.318
τ0.1120.1890.2040.2550.2320.2620.264
ZH→ENTED Newsρ0.2150.2310.3010.4280.4130.4380.426
τ0.1650.1880.2890.3410.3310.3580.352
ρ0.1550.1810.2870.2950.2440.3010.310
τ0.1130.1440.2160.2460.2240.2650.266
", + "bbox": [ + 114, + 344, + 882, + 583 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 5: Segment-level correlations for WMT 2021 MQM annotations over News and TED domains (Freitag et al., 2021). The metrics are Pearson $(\\rho)$ and Kendall Tau $(\\tau)$ . Results in bold indicate which metrics are top-performing for that specific language pair, domain and metric according to Perm-Both hypothesis test (Deutsch et al., 2021), using 500 re-sampling runs, and setting $p = 0.05$ .", + "bbox": [ + 112, + 593, + 882, + 650 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1099", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/0fc3685d75d2fcdaafc985b00e846cc579438685b7215213252d450128e1cb67.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Error TypeNUM EXAMPLES
NE978
NEG669
HALL530
NUM432
Total2609
", + "bbox": [ + 188, + 221, + 415, + 316 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/414856d8bbbd30168d04dcf936c02d7ab892e2e69e5f9a6dd70f04f559d22a28.jpg", + "table_caption": [ + "Table 6: Number of examples for each category, synthetically-created using SMAUG (Alves et al., 2022)." + ], + "table_footnote": [], + "table_body": "
Language PairTOKENS / SENT.ERRORS / SPANS
en-de528704 / 1531025712 / 3567
en-ru525938 / 1507417620 / 7172
zh-en603258 / 1650643984 / 10042
", + "bbox": [ + 115, + 657, + 487, + 722 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 7: Statistics about MQM data from WMT 2021 Metrics task (Freitag et al., 2021) used in our experiments.", + "bbox": [ + 112, + 731, + 489, + 774 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "1100", + "bbox": [ + 482, + 928, + 521, + 940 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Source:", + "text_level": 1, + "bbox": [ + 126, + 237, + 176, + 249 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "格里沃里表示，分析人士对越南所提出的和平倡议给予认可。", + "bbox": [ + 126, + 250, + 527, + 262 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Translation:", + "text_level": 1, + "bbox": [ + 126, + 267, + 206, + 277 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Grivory said that analysts recognize the peace initiative proposed by Vietnam.", + "bbox": [ + 126, + 278, + 596, + 290 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Reference:", + "text_level": 1, + "bbox": [ + 126, + 292, + 196, + 303 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Grigory said that analysts endorse the peace initiative proposed by Vietnam.", + "bbox": [ + 126, + 304, + 586, + 317 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "NE Error:", + "text_level": 1, + "bbox": [ + 126, + 319, + 196, + 330 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Grivory said that analysts recognize the peace initiative proposed by Russia.", + "bbox": [ + 126, + 331, + 591, + 344 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Source:", + "text_level": 1, + "bbox": [ + 126, + 359, + 176, + 369 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "英国的这一决定预计将会使西班牙的旅游业大受影响。", + "bbox": [ + 126, + 369, + 482, + 382 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Translation:", + "text_level": 1, + "bbox": [ + 126, + 386, + 206, + 395 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "This decision by the United Kingdom is expected to greatly affect Spain's tourism industry.", + "bbox": [ + 126, + 397, + 678, + 410 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Reference:", + "text_level": 1, + "bbox": [ + 126, + 412, + 196, + 423 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "This decision by the UK is expected to have a significant impact on tourism in Spain.", + "bbox": [ + 126, + 424, + 640, + 437 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "NEG Error:", + "text_level": 1, + "bbox": [ + 126, + 439, + 203, + 450 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "This decision by the United Kingdom is expected to greatly benefit Spain's tourism industry.", + "bbox": [ + 126, + 451, + 690, + 464 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Source:", + "text_level": 1, + "bbox": [ + 126, + 478, + 176, + 489 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "由于疫情，人们开始在互联网上花费更多的时间。”", + "bbox": [ + 126, + 489, + 467, + 502 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Translation:", + "text_level": 1, + "bbox": [ + 126, + 506, + 206, + 517 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "\"Because of the epidemic, people are starting to spend more time on the Internet.\"", + "bbox": [ + 126, + 518, + 616, + 530 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Reference:", + "text_level": 1, + "bbox": [ + 126, + 533, + 196, + 543 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "For reason of the pandemic, people are starting to spend more time on the Internet.", + "bbox": [ + 126, + 544, + 638, + 557 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "HALL Error:", + "text_level": 1, + "bbox": [ + 126, + 560, + 213, + 570 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Because we have a lot of friends around during the epidemic, people are starting to spend more time on the mobile devices than on the Internet.\"", + "bbox": [ + 126, + 571, + 855, + 598 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Source:", + "text_level": 1, + "bbox": [ + 126, + 613, + 176, + 624 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "展销区将展至7月29日。", + "bbox": [ + 126, + 624, + 278, + 637 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Translation:", + "text_level": 1, + "bbox": [ + 126, + 640, + 206, + 651 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The exhibition and sales area will be open until July 29.", + "bbox": [ + 126, + 652, + 463, + 665 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Reference:", + "text_level": 1, + "bbox": [ + 126, + 667, + 196, + 678 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The exhibition will last until July 29.", + "bbox": [ + 126, + 678, + 349, + 690 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "NUM Error:", + "text_level": 1, + "bbox": [ + 126, + 694, + 206, + 705 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The exhibition and sales area will be open until July 2018", + "bbox": [ + 126, + 706, + 480, + 719 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 8: Synthetically-generated critical errors (highlighted in gray) created with SMAUG (Alves et al., 2022) to assess whether our explanations can be accurately attributed to critical errors.", + "bbox": [ + 114, + 734, + 880, + 762 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "1101", + "bbox": [ + 482, + 928, + 517, + 940 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Source:", + "text_level": 1, + "bbox": [ + 126, + 181, + 176, + 192 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "And yet, the universe is not a silent movie because the universe isn't silent.", + "bbox": [ + 126, + 193, + 579, + 205 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Translation:", + "text_level": 1, + "bbox": [ + 126, + 208, + 206, + 219 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Und Dennoch ist das Universum kein Stummfilm, weil das Universum nicht still ist.", + "bbox": [ + 126, + 219, + 636, + 231 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "COMET score: 0.8595", + "text_level": 1, + "bbox": [ + 126, + 235, + 270, + 246 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "COMET explanation:", + "text_level": 1, + "bbox": [ + 126, + 247, + 263, + 259 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "_Und _dennoch _ist _das _Univers um _kein _Stu mm film , _weil _das _Univers um _nicht _still _ist .", + "bbox": [ + 129, + 260, + 845, + 272 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Source:", + "text_level": 1, + "bbox": [ + 126, + 284, + 176, + 294 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "And yet black holes may be heard even if they're not seen, and that's because they bang on space-time like a drum.", + "bbox": [ + 126, + 294, + 820, + 307 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Translation:", + "text_level": 1, + "bbox": [ + 126, + 310, + 206, + 321 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Und dennoch werden Schwarze Locher weitereicht gehört, auch wenn sie nicht gesehen werden, und das liegt daran, dass sie wie eine Trommel auf die Raumzeit schlagen.", + "bbox": [ + 126, + 322, + 853, + 346 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "COMET score: 0.7150", + "text_level": 1, + "bbox": [ + 126, + 349, + 268, + 361 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "COMET explanation:", + "text_level": 1, + "bbox": [ + 126, + 362, + 263, + 373 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Und dennoch werden Schwarz e LÖcher vielleicht gehört , auch wenn sie nicht gesehenwerden , und das liegt daran , dass sie wie eine Tro mmel auf die Raum zeit schlagen .", + "bbox": [ + 126, + 374, + 853, + 401 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Source:", + "text_level": 1, + "bbox": [ + 126, + 413, + 176, + 423 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Ash O'Brien and husband Jarett Kelley say they were grabbing a bite to eat at Dusty Rhodes dog park in San Diego on Thursday, with their three-month-old pug in tow.", + "bbox": [ + 126, + 424, + 853, + 448 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Translation:", + "text_level": 1, + "bbox": [ + 126, + 451, + 206, + 461 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Ash O'Brien und Ehemann Jarett Kelley sagen, dass sie am Donnerstag im Hundepark Dusty Rhodes in San Diego einen Happen zu essen griffen, mit ihrem drei Monate alten Mops im Schleptau.", + "bbox": [ + 126, + 463, + 853, + 487 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "COMET score: 0.6835", + "text_level": 1, + "bbox": [ + 126, + 489, + 268, + 501 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "COMET explanation:", + "text_level": 1, + "bbox": [ + 126, + 502, + 263, + 513 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "_Ash_O' Brien_Und_Ehe mann_Ja rett_Kelley_sagen, dass sie_am_Donnerstag_im_Hunde park_Dusty_Rhod es_in_San_Diego_einen_Happ_en_zu_essen_griff_en_, _mit_threm_drei_Monate_alten_M ops_im_Schleppt au.", + "bbox": [ + 126, + 514, + 853, + 556 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Source:", + "text_level": 1, + "bbox": [ + 126, + 567, + 176, + 577 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "It was Einstein's great general theory of relativity.", + "bbox": [ + 126, + 579, + 428, + 592 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Translation:", + "text_level": 1, + "bbox": [ + 126, + 594, + 206, + 605 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Es war Einsteins große allgemeine Forschungen vor Relativitätstheorie.", + "bbox": [ + 126, + 606, + 564, + 619 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "COMET score: 0.6692", + "text_level": 1, + "bbox": [ + 126, + 621, + 268, + 633 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "COMET explanation:", + "text_level": 1, + "bbox": [ + 126, + 634, + 263, + 646 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "_Es _war _Einstein s _große _allgemeine e _Forschung en _vor _Relativ itäts the ori e .", + "bbox": [ + 126, + 646, + 736, + 659 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Source:", + "text_level": 1, + "bbox": [ + 126, + 670, + 176, + 681 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "There's mask-shaming and then there's full on assault.", + "bbox": [ + 126, + 682, + 455, + 694 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Translation:", + "text_level": 1, + "bbox": [ + 126, + 697, + 206, + 708 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Es gibt Maskenschämen und dann ist es voll bei Angriff.", + "bbox": [ + 126, + 709, + 485, + 722 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "COMET score: 0.2318", + "text_level": 1, + "bbox": [ + 126, + 724, + 268, + 734 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "COMET explanation:", + "text_level": 1, + "bbox": [ + 126, + 736, + 263, + 747 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "_Es_gibt_Mask en schä men_und_dann_ist es_voll_bei_Angriff_.", + "bbox": [ + 126, + 747, + 647, + 762 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Table 9: Saliency map for COMET explanation scores for a set of en→de examples. Comparing the token-level explanations with the MQM annotation (highlighted in gray) reveals the source of correspondence between specific token-level translation errors and the resulting scores.", + "bbox": [ + 115, + 777, + 880, + 819 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "1102", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Source:", + "text_level": 1, + "bbox": [ + 124, + 184, + 176, + 193 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "我想告诉大家宇宙有着自己的配乐，而宇宙自身正在不停地播放着。因为太空可以想鼓一样振动。", + "bbox": [ + 124, + 195, + 781, + 208 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Translation:", + "text_level": 1, + "bbox": [ + 126, + 211, + 206, + 222 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "I want to tell you that the universe has its own iconic soundtrack and the universe itself is constantly playing non-stop because space can vibrate like a drum.", + "bbox": [ + 124, + 223, + 855, + 247 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "COMET score: 0.8634", + "text_level": 1, + "bbox": [ + 126, + 250, + 268, + 261 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "COMET explanation:", + "text_level": 1, + "bbox": [ + 126, + 263, + 263, + 274 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "_I_want_to_tell你想_the_univers_e_has_its_own(iconic soundtrack_and_the_univers_e_itself_is_constantly-playing_non-stop_because_space_can_vibrate_to_like_a Drum.", + "bbox": [ + 126, + 274, + 853, + 302 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Source:", + "text_level": 1, + "bbox": [ + 126, + 313, + 176, + 324 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "另外，吉克隽逸和刘烨作为运动助理，也围绕运动少年制造了不少爆笑话题。", + "bbox": [ + 126, + 324, + 610, + 338 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Translation:", + "text_level": 1, + "bbox": [ + 126, + 341, + 206, + 351 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "In addition, as sports assistants, Ji Kejunyi and Liu Ye have also created a lot of hilarious topics around sports teenagers.", + "bbox": [ + 126, + 353, + 855, + 366 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "COMET score: 0.8214", + "text_level": 1, + "bbox": [ + 126, + 368, + 268, + 380 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "COMET explanation:", + "text_level": 1, + "bbox": [ + 126, + 381, + 263, + 393 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "_In _addition , _as _sports _assistant s , _Ji _Ke ju nyi _and _Li u _Ye _have _also _created _a _lot _of_ hila rious _topic s _around _sports _teenager s .", + "bbox": [ + 126, + 393, + 853, + 420 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Source:", + "text_level": 1, + "bbox": [ + 126, + 432, + 176, + 442 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "一番言论让场上的少年和运动领队们都倒吸一口凉气。", + "bbox": [ + 126, + 443, + 482, + 456 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Translation:", + "text_level": 1, + "bbox": [ + 126, + 458, + 206, + 470 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "The remarks made the teenagers and the sports leaders on the field gasp a sigh of relief.", + "bbox": [ + 126, + 470, + 660, + 483 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "COMET score: 0.7793", + "text_level": 1, + "bbox": [ + 126, + 486, + 268, + 498 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "COMET explanation:", + "text_level": 1, + "bbox": [ + 126, + 499, + 263, + 511 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "The re marks made the teenager s and the sports leaders on the field gas p a sig h of _relief", + "bbox": [ + 126, + 511, + 855, + 539 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Source:", + "text_level": 1, + "bbox": [ + 126, + 550, + 176, + 560 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "强烈的阳光是如此地刺眼，", + "bbox": [ + 126, + 561, + 299, + 574 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Translation:", + "text_level": 1, + "bbox": [ + 126, + 577, + 206, + 588 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "The intense sunlight is so harsh;", + "bbox": [ + 126, + 589, + 326, + 602 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "COMET score: 0.7561", + "text_level": 1, + "bbox": [ + 126, + 604, + 268, + 615 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "COMET explanation:", + "text_level": 1, + "bbox": [ + 126, + 615, + 263, + 627 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "The intense sun light is so har sh", + "bbox": [ + 126, + 627, + 423, + 642 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Source:", + "text_level": 1, + "bbox": [ + 126, + 653, + 176, + 664 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "如今，我们希望能够给这部关于宇宙的宏伟的视觉作品配上声音。", + "bbox": [ + 126, + 664, + 569, + 677 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Translation:", + "text_level": 1, + "bbox": [ + 126, + 680, + 206, + 690 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Today, we hope to be able to give this magnificent visual work of the universe a sound.", + "bbox": [ + 126, + 692, + 665, + 706 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "COMET score: 0.7073", + "text_level": 1, + "bbox": [ + 126, + 708, + 268, + 720 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "COMET explanation:", + "text_level": 1, + "bbox": [ + 126, + 720, + 263, + 732 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Today,we hope to be able to give this magnificent ent visual work of the univers e a sound.", + "bbox": [ + 126, + 732, + 855, + 760 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Table 10: Saliency map for COMET explanation scores for a set of $\\mathrm{{zh}} \\rightarrow \\mathrm{{en}}$ examples. Comparing the token-level explanations with the MQM annotation (highlighted in gray) reveals the source of correspondence between specific token-level translation errors and the resulting scores.", + "bbox": [ + 115, + 774, + 880, + 816 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "1103", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "A For every submission:", + "text_level": 1, + "bbox": [ + 114, + 107, + 322, + 122 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "A1. Did you describe the limitations of your work?", + "bbox": [ + 129, + 126, + 532, + 142 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Yes. Section 6", + "bbox": [ + 149, + 143, + 257, + 156 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "A2. Did you discuss any potential risks of your work?", + "bbox": [ + 127, + 168, + 552, + 185 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Not applicable. Left blank.", + "bbox": [ + 149, + 186, + 351, + 200 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "A3. Do the abstract and introduction summarize the paper's main claims?", + "bbox": [ + 129, + 211, + 695, + 227 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Abstract and Section 1", + "bbox": [ + 149, + 229, + 321, + 242 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "A4. Have you used AI writing assistants when working on this paper?", + "bbox": [ + 129, + 253, + 670, + 269 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Assistance purely with the language of the paper along every section. Grammarly and DeepL write", + "bbox": [ + 147, + 271, + 880, + 287 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "B Did you use or create scientific artifacts?", + "text_level": 1, + "bbox": [ + 112, + 297, + 489, + 313 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Section 3 explains the methods we used. We will release the adaptations required to use the explainability methods over COMET framework, the UniTE model we trained, and all data synthetically-generated data.", + "bbox": [ + 112, + 318, + 882, + 350 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "B1. Did you cite the creators of artifacts you used?", + "bbox": [ + 129, + 359, + 529, + 375 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Section 2", + "bbox": [ + 151, + 376, + 223, + 390 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?", + "bbox": [ + 129, + 401, + 778, + 418 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "footnote on the first page. The License will be Apache 2.0", + "bbox": [ + 149, + 419, + 579, + 434 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?", + "bbox": [ + 127, + 445, + 880, + 508 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Not applicable. Left blank.", + "bbox": [ + 149, + 510, + 351, + 525 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?", + "bbox": [ + 127, + 536, + 880, + 583 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Not applicable. Left blank.", + "bbox": [ + 149, + 585, + 351, + 600 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?", + "bbox": [ + 129, + 609, + 880, + 642 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "in the Appendix we have several statistics for training and testing data.", + "bbox": [ + 149, + 643, + 675, + 658 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be.", + "bbox": [ + 129, + 668, + 882, + 749 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 149, + 750, + 231, + 765 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "C Did you run computational experiments?", + "text_level": 1, + "bbox": [ + 112, + 774, + 492, + 791 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Appendix provides detailed information about the trained model.", + "bbox": [ + 129, + 796, + 611, + 812 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?", + "bbox": [ + 129, + 820, + 880, + 854 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Appendix provides detailed information about the trained model including GPU infrastructure and total number of parameters.", + "bbox": [ + 147, + 854, + 880, + 885 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "ACL 2023 Responsible NLP Checklist", + "bbox": [ + 132, + 84, + 433, + 99 + ], + "page_idx": 15 + }, + { + "type": "footer", + "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.", + "bbox": [ + 112, + 892, + 877, + 917 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "1104", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?", + "bbox": [ + 127, + 84, + 878, + 115 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Not applicable. Left blank.", + "bbox": [ + 149, + 117, + 349, + 131 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?", + "bbox": [ + 129, + 142, + 882, + 190 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Appendix has all information needed about test data and performance of the models.", + "bbox": [ + 147, + 192, + 774, + 208 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?", + "bbox": [ + 129, + 217, + 882, + 265 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Section 2 and Appendix", + "bbox": [ + 149, + 267, + 329, + 282 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?", + "bbox": [ + 112, + 293, + 877, + 310 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 132, + 313, + 213, + 329 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?", + "bbox": [ + 127, + 340, + 882, + 372 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Not applicable. Left blank.", + "bbox": [ + 149, + 373, + 349, + 388 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?", + "bbox": [ + 127, + 399, + 882, + 447 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Not applicable. Left blank.", + "bbox": [ + 149, + 448, + 349, + 464 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?", + "bbox": [ + 127, + 475, + 882, + 521 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Not applicable. Left blank.", + "bbox": [ + 149, + 523, + 349, + 539 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?", + "bbox": [ + 127, + 549, + 873, + 565 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Not applicable. Left blank.", + "bbox": [ + 149, + 565, + 349, + 582 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?", + "bbox": [ + 127, + 592, + 880, + 623 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Not applicable. Left blank.", + "bbox": [ + 149, + 625, + 349, + 640 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "1105", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2023/The Inside Story_ Towards Better Understanding of Machine Translation Neural Evaluation Metrics/d25ab395-c57a-4c68-8e6d-00772d415f43_model.json b/2023/The Inside Story_ Towards Better Understanding of Machine Translation Neural Evaluation Metrics/d25ab395-c57a-4c68-8e6d-00772d415f43_model.json new file mode 100644 index 0000000000000000000000000000000000000000..eb1b6816f63f534f0e53d2b0c798b20bdab4b7f8 --- /dev/null +++ b/2023/The Inside Story_ Towards Better Understanding of Machine Translation Neural Evaluation Metrics/d25ab395-c57a-4c68-8e6d-00772d415f43_model.json @@ -0,0 +1,3743 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.236, + 0.089, + 0.761, + 0.126 + ], + "angle": 0, + "content": "The Inside Story: Towards Better Understanding of Machine Translation Neural Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.246, + 0.142, + 0.761, + 0.175 + ], + "angle": 0, + "content": "Ricardo Rei\\*1,2,4, Nuno M. Guerreiro\\*3,4, Marcos Treviso\\*3,4, Alon Lavie\\*1, Luisa Coheur\\*2,4, Andre F. T. Martins\\*1,3,4" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.177, + 0.74, + 0.209 + ], + "angle": 0, + "content": "1Unbabel, Lisbon, Portugal, 2INESC-ID, Lisbon, Portugal \n3Instituto de Telecomunicações, Lisbon, Portugal" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.21, + 0.74, + 0.227 + ], + "angle": 0, + "content": "\\(^{4}\\)Instituto Superior Técnico, University of Lisbon, Portugal" + }, + { + "type": "list", + "bbox": [ + 0.261, + 0.177, + 0.74, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.253, + 0.341, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.141, + 0.279, + 0.461, + 0.593 + ], + "angle": 0, + "content": "Neural metrics for machine translation evaluation, such as COMET, exhibit significant improvements in their correlation with human judgments compared to traditional metrics based on lexical overlap, such as BLEU. Yet neural metrics are, to a great extent, \"black boxes\" that return a single sentence-level score without transparency about the decision-making process. In this work, we develop and compare several neural explainability methods and demonstrate their effectiveness for interpreting state-of-the-art fine-tuned neural metrics. Our study reveals that these metrics leverage token-level information that can be directly attributed to translation errors, as assessed through comparison of token-level neural saliency maps with Multidimensional Quality Metrics (MQM) annotations and with synthetically-generated critical translation errors. To ease future research, we release our code at https://github.com/Unbabel/COMET/tree/explainable-metrics." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.604, + 0.26, + 0.619 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.629, + 0.49, + 0.887 + ], + "angle": 0, + "content": "Reference-based neural metrics for machine translation evaluation are achieving evergrowing success, demonstrating superior results over traditional lexical overlap-based metrics, such as BLEU (Papineni et al., 2002) and CHRF (Popovic, 2015), in terms of both their correlation with human ratings and their robustness across diverse domains (Callison-Burch et al., 2006; Smith et al., 2016; Mathur et al., 2020; Kocmi et al., 2021; Freitag et al., 2022). However, lexical overlap-based metrics remain popular for evaluating the performance and progress of translation systems and algorithms. Concerns regarding trust and interpretability may help explain this (Leiter et al., 2022): contrary to traditional metrics, neural metrics are considered \"black boxes\" as they often use" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.252, + 0.868, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.388, + 0.883, + 0.459 + ], + "angle": 0, + "content": "Figure 1: Illustration of our approach. In this example, the metric assigns the translation a low score. We aim to better understand this sentence-level assessment by examining the correspondence between our token-level explanations and human annotated error spans." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.485, + 0.882, + 0.533 + ], + "angle": 0, + "content": "increasingly large models (e.g., the winning metric of the WMT 22 Metrics shared task was a 10B parameter model (Freitag et al., 2022))." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.534, + 0.885, + 0.919 + ], + "angle": 0, + "content": "While some recent work has focused on explaining the predictions made by reference-free quality estimation (QE) systems (Fomicheva et al., 2021; Zerva et al., 2022), explaining reference-based metrics has remained a largely overlooked problem (Leiter et al., 2022). It is an open question whether the observations from studies of explainable QE carry over to this scenario. Thus, in this work, we fill that gap by turning to state-of-the-art reference-based metrics—we aim to interpret their decision-making process by exploiting the fact that these metrics show consistently good correlations with Multidimensional Quality Metrics (MQM) (Freitag et al., 2021, 2022; Sai et al., 2022), which are fine-grained quality assessments that result from experts identifying error spans in translation outputs (Lommel et al., 2014). We hypothesize that reference-based metrics leverage this token-level information to produce sentence-level scores. To test this hypothesis, we assess whether our explanations – measures of token-level importance obtained via attribution and input attribution methods such as attention weights and gradient scores (Treviso et al., 2021; Rei et al., 2022b) – align with" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.893, + 0.488, + 0.918 + ], + "angle": 0, + "content": "* Equal contribution. Corresponding author: ricardo.rei@unbabel.com" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1089" + }, + { + "type": "footer", + "bbox": [ + 0.227, + 0.946, + 0.77, + 0.958 + ], + "angle": 0, + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + }, + { + "type": "footer", + "bbox": [ + 0.369, + 0.959, + 0.63, + 0.972 + ], + "angle": 0, + "content": "Volume 2: Short Papers, pages 1089-1105" + }, + { + "type": "footer", + "bbox": [ + 0.297, + 0.973, + 0.702, + 0.986 + ], + "angle": 0, + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.49, + 0.116 + ], + "angle": 0, + "content": "human-annotated spans (Fomicheva et al., 2021, 2022; Zerva et al., 2022), as illustrated in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.118, + 0.49, + 0.213 + ], + "angle": 0, + "content": "Our analysis focuses on two main vectors: (i) understanding the impact of the reference information on the quality of the explanations; and (ii) finding whether the explanations can help to identify potential weaknesses in the metrics. Our main contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.217, + 0.49, + 0.281 + ], + "angle": 0, + "content": "- We provide a comparison between multiple explainability methods for different metrics on all types of evaluation: src-only, ref-only, and src+ref joint evaluation." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.293, + 0.49, + 0.341 + ], + "angle": 0, + "content": "- We find that explanations are related to the underlying metric architecture, and that leveraging reference information improves the explanations." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.353, + 0.488, + 0.385 + ], + "angle": 0, + "content": "- We show that explanations for critical translation errors can reveal weaknesses in the metrics." + }, + { + "type": "list", + "bbox": [ + 0.116, + 0.217, + 0.49, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.398, + 0.381, + 0.415 + ], + "angle": 0, + "content": "2 Explaining Neural Metrics" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.424, + 0.49, + 0.504 + ], + "angle": 0, + "content": "We aim to explain sentence-level quality assessments of reference-based metrics by producing token-level explanations that align to translation errors. In what follows, we describe the metrics and how we produce the explanations that we study." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.516, + 0.222, + 0.53 + ], + "angle": 0, + "content": "2.1 Metrics" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.537, + 0.49, + 0.859 + ], + "angle": 0, + "content": "We focus our analysis on two state-of-the-art neural metrics: COMET (Rei et al., 2020) and UNITE (Wan et al., 2022). While both metrics use a multilingual encoder model based on XLMR (Conneau et al., 2020), they employ distinct strategies to obtain sentence-level quality scores. On the one hand, COMET separately encodes the source, translation and reference to obtain their respective sentence embeddings; these embeddings are then combined to compute a quality score. On the other, UNITE jointly encodes the sentences to compute a contextualized representation that is subsequently used to compute the quality score. Interestingly, UNITE is trained to obtain quality scores for different input combinations: [mt; src] (SRC), [mt; ref] (REF), and [mt; src; ref] (SRC+REF). In fact, when the input is SRC, UNITE works like TransQuest (Ranasinghe et al., 2020); REF, like BLEURT (Sellam et al., 2020); and SRC+REF, like ROBLEURT (Wan et al., 2021)." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.085, + 0.855, + 0.1 + ], + "angle": 0, + "content": "2.2 Explanations via Attribution Methods" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.106, + 0.885, + 0.348 + ], + "angle": 0, + "content": "In this work, we produce explanations using attribution methods that assign a scalar value to each translation token (i.e. a token-level attribution) to represent its importance. While many input attribution methods exist and have been extensively studied in the literature (Ribeiro et al., 2016; Shrikumar et al., 2017; Sundararajan et al., 2017; Jain and Wallace, 2019; Atanasova et al., 2020; Zaman and Belinkov, 2022), we focus specifically on those that have been demonstrated to be effective for explaining the predictions of QE models (Treviso et al., 2021; Fomicheva et al., 2022; Fernandes et al., 2022; Zerva et al., 2022) and extend them to our reference-based scenario. Concretely, we use the following techniques to extract explanations:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.359, + 0.885, + 0.424 + ], + "angle": 0, + "content": "- embed-align: the maximum cosine similarity between each translation token embedding and the reference and/or source token embeddings (Tao et al., 2022);" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.435, + 0.884, + 0.482 + ], + "angle": 0, + "content": "- grad \\(\\ell_2\\): the \\(\\ell_2\\)-norm of gradients with respect to the word embeddings of the translation tokens (Arras et al., 2019);" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.495, + 0.884, + 0.543 + ], + "angle": 0, + "content": "- attention: the attention weights of the translation tokens for each attention head of the encoder (Treviso et al., 2021);" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.883, + 0.601 + ], + "angle": 0, + "content": "- \\(\\mathbf{attn} \\times \\mathbf{grad}\\): the attention weights of each head scaled by the \\(\\ell_2\\)-norm of the gradients of the value vectors of that head (Rei et al., 2022b)." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.359, + 0.885, + 0.601 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.626, + 0.729, + 0.642 + ], + "angle": 0, + "content": "3 Experimental Setting" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.651, + 0.884, + 0.812 + ], + "angle": 0, + "content": "MQM annotations. We use MQM annotations from the WMT 2021 Metrics shared task (Freitag et al., 2021),3 covering three language pairs — English-German (en→de), English-Russian (en→ru), and Chinese-English (zh→en) —in two different domains: News and TED Talks. For each incorrect translation, human experts marked the corresponding error spans. In our framework, these error spans should align with the words that the attribution methods assign higher importance to." + }, + { + "type": "page_footnote", + "bbox": [ + 0.508, + 0.821, + 0.885, + 0.892 + ], + "angle": 0, + "content": "2For all attention-based methods, we ensemble the explanations from the top 5 heads as this has shown to improve performance consistently over selecting just the best head (Treviso et al., 2021; Rei et al., 2022b). Moreover, we use the full attention matrix, instead of relying only on cross attention information." + }, + { + "type": "page_footnote", + "bbox": [ + 0.51, + 0.893, + 0.753, + 0.919 + ], + "angle": 0, + "content": "3https://github.com/google/wmt-mqm-human-evaluation" + }, + { + "type": "list", + "bbox": [ + 0.508, + 0.821, + 0.885, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.113, + 0.87, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Ensembles composed of these two metrics were respectively ranked second and third in WMT 2022 Metrics shared task. The winner of WMT 2022 Metrics task — METRICXXL — is not publicly available (Freitag et al., 2022)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1090" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.182, + 0.082, + 0.819, + 0.386 + ], + "angle": 0, + "content": "
METRICEXPLAINABILITY METHODen→dezh→enen→ruAvg.
AUCR@KAUCR@KAUCR@KAUCR@K
src-only* evaluation
UNITE SRCembed-align[mt, src]0.5870.3390.6440.2810.5830.1670.6040.262
grad ℓ20.5720.2930.5350.2000.6200.1690.5760.221
attention0.6360.3220.6120.2530.6120.1890.6200.254
attn × grad0.7070.3760.6390.2940.6330.2110.6600.294
ref-only evaluation
UNITE REFembed-align[mt, ref]0.6580.3960.6670.3280.6350.2180.6530.314
grad ℓ20.5960.3190.5710.2600.6610.2020.6090.261
attention0.6370.3440.6700.3350.6520.2240.6530.301
attn × grad0.7250.4250.6670.3800.6600.2480.6840.351
src, ref joint evaluation
UNITE SRC+REFembed-align[mt, src; ref]0.6500.3830.6700.3300.6180.2130.6460.309
grad ℓ20.5950.3250.5790.2570.6430.1910.6060.257
attention0.6570.4210.6700.3830.6490.2230.6590.342
attn × grad0.7360.4210.6740.3830.6710.2480.6930.351
COMETembed-align[mt, src]0.5900.3710.6740.3140.5770.2200.6140.301
embed-align[mt, ref]0.6940.4250.6960.3550.6470.2750.6790.352
embed-align[mt, src; ref]0.6880.4160.6970.3570.6220.2790.6690.350
grad ℓ20.6030.3120.5400.2520.6040.1850.5820.250
attention0.6040.3510.5920.2590.6330.2090.6080.268
attn × grad0.7100.3650.6330.2780.6620.2440.6690.295
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.395, + 0.884, + 0.44 + ], + "angle": 0, + "content": "Table 1: AUC and Recall@K of explanations obtained via different attribution methods for COMET and UNITE models on the MQM data. *Although UNITE SRC is a src-only evaluation metric, it was trained with reference information (Wan et al., 2022)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.458, + 0.489, + 0.603 + ], + "angle": 0, + "content": "Models. For COMET, we use the latest publicly available model: wmt22-comet-da (Rei et al., 2022a). For UNITE, we train our own model using the same data used to train COMET in order to have a comparable setup. We provide full details (training data, correlations with human annotations, and hyperparameters) in Appendix A. Overall, the resulting reference-based UNITE models (REF and SRC+REF) are on par with COMET." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.614, + 0.49, + 0.726 + ], + "angle": 0, + "content": "Evaluation. We want our explanations to be directly attributed to the annotated error spans, in the style of an error detection task. Thus, we report Area Under Curve (AUC) and Recall@K.6 These metrics have been used as the main metrics in previous works on explainable QE (Fomicheva et al., 2021, 2022; Zerva et al., 2022)." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.74, + 0.214, + 0.755 + ], + "angle": 0, + "content": "4 Results" + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.767, + 0.308, + 0.783 + ], + "angle": 0, + "content": "4.1 High-level analysis" + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.79, + 0.49, + 0.821 + ], + "angle": 0, + "content": "Explanations are tightly related to the underlying metric architecture. The results in Ta-" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.458, + 0.885, + 0.749 + ], + "angle": 0, + "content": "ble 1 show that the predictive power of the attribution methods differ between UNITE and COMET: attn \\(\\times\\) grad is the best method for UNITE-based models, while embed-align works best for COMET. This is expected as UNITE constructs a joint representation for the input sentences, thus allowing attention to flow across them; COMET, in contrast, encodes the sentences separately, so it relies heavily on the separate contextualized embeddings that are subsequently combined via elementwise operations such as multiplication and absolute difference. Interestingly, embed-align and attn \\(\\times\\) grad were the winning explainability approaches of the WMT 2022 Shared-Task on Quality Estimation (Zerva et al., 2022). This suggests that explainability methods developed for QE systems can translate well to reference-based metrics. We provide examples of explanations in Appendix C." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.759, + 0.884, + 0.87 + ], + "angle": 0, + "content": "Reference information boosts explainability power. Table 1 also shows that, across all metrics, using reference information brings substantial improvements over using only the source information. Moreover, while reference-based attributions significantly outperform source-based attributions, combining the source and reference information to" + }, + { + "type": "page_footnote", + "bbox": [ + 0.509, + 0.881, + 0.883, + 0.918 + ], + "angle": 0, + "content": "In Appendix B, we provide a comparison between the explanations obtained via embed-align with COMET and with its pretrained encoder model, XLM-R." + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.832, + 0.4, + 0.856 + ], + "angle": 0, + "content": "4https://huggingface.co/Un babel/wmt22-comet-da" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.857, + 0.486, + 0.881 + ], + "angle": 0, + "content": "5Our implementation differs from the original work by Wan et al. (2022), See Appendix A for full details." + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.882, + 0.486, + 0.919 + ], + "angle": 0, + "content": "In this setup, Recall@K is the proportion of words with the highest attribution that correspond to translation errors against the total number of errors in the annotated error span." + }, + { + "type": "list", + "bbox": [ + 0.114, + 0.832, + 0.486, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.519, + 0.941 + ], + "angle": 0, + "content": "1091" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.119, + 0.082, + 0.488, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.205, + 0.49, + 0.276 + ], + "angle": 0, + "content": "Figure 2: Performance of the best attribution methods for COMET, UNITE REF and UNITE SRC+REF in terms of Recall@K on translations with critical errors: negations (NEG), hallucinations (HALL), named entity errors (NE), and errors in numbers (NUM)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.297, + 0.49, + 0.409 + ], + "angle": 0, + "content": "obtain token-level attributions does not consistently yield superior results over using the reference alone. Notably, the best attribution method for COMET does not require any source information. This is interesting: in some cases, reference-based metrics may largely ignore source information, relying heavily on the reference instead." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.421, + 0.479, + 0.452 + ], + "angle": 0, + "content": "4.2 How do the explanations fare for critical translation errors?" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.459, + 0.49, + 0.668 + ], + "angle": 0, + "content": "The MQM data analyzed until now consists primarily of high quality translations, with the majority of annotated errors being non-critical. However, it is important to assess whether our explanations can be accurately attributed to critical errors, as this may reveal potential metric shortcomings. To this end, we employ SMAUG (Alves et al., 2022)8, a tool designed to generate synthetic data for stress-testing metrics, to create corrupted translations that contain critical errors. Concretely, we generate translations with the following pathologies: negation errors, hallucinations via insertions, named entity errors, and errors in numbers.9" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.678, + 0.49, + 0.823 + ], + "angle": 0, + "content": "Explanations identify critical errors more easily than non-critical errors. Figure 2 shows that explanations are more effective in identifying critical errors compared to other non-critical errors (see Table 1). Specifically, we find significant performance improvements up to nearly \\(30\\%\\) in Recall@K for certain critical errors. Overall, hallucinations are the easiest errors to identify across all neural metrics. This suggests that neural" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.246 + ], + "angle": 0, + "content": "metrics appropriately identify and penalize hallucinated translations, which aligns with the findings of Guerreiro et al. (2022). Moreover, explanations for both UNITE models behave similarly for all errors except numbers, where the source information plays a key role in improving the explanations. Notably, contrary to what we observed for data with non-critical errors, COMET explanations are less effective than those of UNITE REF and UNITE SRC+REF for identifying critical errors." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.255, + 0.885, + 0.625 + ], + "angle": 0, + "content": "Explanations can reveal potential metric weaknesses. Figure 2 suggests that COMET explanations struggle to identify localized errors (negation errors, named entity errors and discrepancies in numbers). We hypothesize that this behavior is related to the underlying architecture. Unlike UNITE-based metrics, COMET does not rely on soft alignments via attention between the sentences in the encoding process. This process may be key to identify local misalignments during the encoding process. In fact, the attention-based attributions for UNITE metrics can more easily identify these errors. COMET, however, encodes the sentences separately, which may result in grammatical features (e.g. numbers) being encoded similarly across sentences (Chi et al., 2020; Chang et al., 2022). As such, explanations obtained via embedding alignments will not properly identify these misalignments on similar features. Importantly, these findings align with observations made in (Amrhein and Sennrich, 2022; Raunak et al., 2022). This showcases how explanations can be used to diagnose and reveal shortcomings of neural-based metrics." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.637, + 0.804, + 0.652 + ], + "angle": 0, + "content": "5 Conclusions and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.662, + 0.885, + 0.919 + ], + "angle": 0, + "content": "In this paper, we investigated the use of explainability methods to better understand widely used neural metrics for machine translation evaluation, such as COMET and UNITE. Concretely, we analyzed how explanations are impacted by the reference information, and how they can be used to reveal weaknesses of these metrics. Our analysis shows that the quality of the explanations is tightly related to the underlying metric architecture. Interestingly, we also provide evidence that neural metrics like COMET may heavily rely on reference information over source information. Additionally, we show that explanations can be used to reveal reference-based metrics weaknesses such as failing to severely penalize localized critical errors. This opens up promising opportunities for future" + }, + { + "type": "page_footnote", + "bbox": [ + 0.142, + 0.833, + 0.408, + 0.847 + ], + "angle": 0, + "content": "8https://github.com/Unbabel/smaug" + }, + { + "type": "page_footnote", + "bbox": [ + 0.116, + 0.847, + 0.488, + 0.919 + ], + "angle": 0, + "content": "9We corrupt fully correct translations that are not an exact copy of the reference translation. Moreover, as the full suit of SMAUG transformations can only be applied to English data, we focus solely on zh→en translations. Overall, the synthetic dataset consists of 2610 translations. Full statistics about the corrupted data and examples are shown in Appendix A.2." + }, + { + "type": "list", + "bbox": [ + 0.116, + 0.833, + 0.488, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1092" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.491, + 0.168 + ], + "angle": 0, + "content": "research on leveraging explanations to diagnose reference-based metrics errors. To support these studies, we call for future datasets illustrating critical errors (e.g., challenge sets (Karpinska et al., 2022)) to be accompanied by annotated error spans." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.178, + 0.221, + 0.193 + ], + "angle": 0, + "content": "Limitations" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.204, + 0.476, + 0.219 + ], + "angle": 0, + "content": "We highlight three main limitations of our work." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.22, + 0.49, + 0.346 + ], + "angle": 0, + "content": "First, although we have explored gradient-based explanations that take the whole network into consideration and have been shown to be faithful in previous work (Bastings et al., 2021), we do not explicitly explore how COMET combines the sentence representations in the feed-forward that precedes the encoder model to produce the sentence-level score." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.349, + 0.49, + 0.492 + ], + "angle": 0, + "content": "Second, we have shown that combining attention with gradient information results in the best explanations for UNITE-based metrics. However, from a practical standpoint, running inference and extracting the explainability scores simultaneously may be more computationally expensive than other techniques: gradient-based metrics benefit from GPU infrastructure and require storing all gradient information." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.494, + 0.49, + 0.59 + ], + "angle": 0, + "content": "Third, we have not explored extracting explanations in low-resource settings. That is because high-quality MQM annotations for such language pairs are not yet available. Nevertheless, further research in those settings is needed to access the broader validity of our claims." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.603, + 0.287, + 0.619 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.628, + 0.49, + 0.789 + ], + "angle": 0, + "content": "This work was partially supported by the P2020 programs (MAIA, contract 045909), the Portuguese Recovery and Resilience Plan (PRR) through project C645008882-00000055, Center for Responsible AI, by the European Research Council (ERC StG DeepSPIN, 758969), by EU's Horizon Europe Research and Innovation Actions (UTTER, contract 101070631), and by the Fundação para a Ciência e Tecnologia (contracts UIDB/50021/2020 and UIDB/50008/2020)." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.817, + 0.214, + 0.832 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.839, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Duarte Alves, Ricardo Rei, Ana C Farinha, José G. C. de Souza, and André F. T. Martins. 2022. Robust MT Evaluation with Sentence-level Multilingual Augmentation. In Proceedings of the Seventh Conference on Machine Translation, pages 469-478, Abu Dhabi. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.086, + 0.883, + 0.204 + ], + "angle": 0, + "content": "Chantal Amrhein and Rico Sennrich. 2022. Identifying Weaknesses in Machine Translation Metrics Through Minimum Bayes Risk Decoding: A Case Study for COMET. In Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1125-1141, Online only. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.214, + 0.885, + 0.307 + ], + "angle": 0, + "content": "Leila Arras, Ahmed Osman, Klaus-Robert Müller, and Wojciech Samek. 2019. Evaluating recurrent neural network explanations. In Proceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 113-126, Florence, Italy. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.317, + 0.884, + 0.41 + ], + "angle": 0, + "content": "Pepa Atanasova, Jakob Grue Simonsen, Christina Lioma, and Isabelle Augenstein. 2020. A diagnostic study of explainability techniques for text classification. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3256-3274, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.419, + 0.884, + 0.484 + ], + "angle": 0, + "content": "Jasmijn Bastings, Sebastian Ebert, Polina Zablotskaia, Anders Sandholm, and Katja Filippova. 2021. \"will you find these shortcuts?\" a protocol for evaluating the faithfulness of input salience methods for text classification." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.495, + 0.885, + 0.575 + ], + "angle": 0, + "content": "Chris Callison-Burch, Miles Osborne, and Philipp Koehn. 2006. Re-evaluating the role of Bleu in machine translation research. In 11th Conference of the European Chapter of the Association for Computational Linguistics, pages 249-256, Trento, Italy. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.585, + 0.884, + 0.625 + ], + "angle": 0, + "content": "Tyler A. Chang, Zhuowen Tu, and Benjamin K. Bergen. 2022. The geometry of multilingual language model representations." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.634, + 0.884, + 0.714 + ], + "angle": 0, + "content": "Ethan A. Chi, John Hewitt, and Christopher D. Manning. 2020. Finding universal grammatical relations in multilingual BERT. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5564-5577, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.724, + 0.885, + 0.843 + ], + "angle": 0, + "content": "Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettle-moyer, and Veselin Stoyanov. 2020. Unsupervised cross-lingual representation learning at scale. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8440-8451, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.852, + 0.884, + 0.918 + ], + "angle": 0, + "content": "Daniel Deutsch, Rotem Dror, and Dan Roth. 2021. A statistical analysis of summarization evaluation metrics using resampling methods. Transactions of the Association for Computational Linguistics, 9:1132-1146." + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.086, + 0.885, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1093" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.14 + ], + "angle": 0, + "content": "Patrick Fernandes, Marcos Treviso, Danish Pruthi, André F. T. Martins, and Graham Neubig. 2022. Learning to scaffold: Optimizing model explanations for teaching." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.149, + 0.49, + 0.255 + ], + "angle": 0, + "content": "Marina Fomicheva, Piyawat Lertvittayakumjorn, Wei Zhao, Steffen Eger, and Yang Gao. 2021. The Eval4NLP shared task on explainable quality estimation: Overview and results. In Proceedings of the 2nd Workshop on Evaluation and Comparison of NLP Systems, pages 165-178, Punta Cana, Dominican Republic. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.265, + 0.49, + 0.343 + ], + "angle": 0, + "content": "Marina Fomicheva, Lucia Specia, and Nikolaos Aletras. 2022. Translation error detection as rationale extraction. In Findings of the Association for Computational Linguistics: ACL 2022, pages 4148-4159, Dublin, Ireland. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.354, + 0.49, + 0.459 + ], + "angle": 0, + "content": "Markus Freitag, Ricardo Rei, Nitika Mathur, Chi-kiu Lo, Craig Stewart, Eleftherios Avramidis, Tom Kocmi, George Foster, Alon Lavie, and André F. T. Martins. 2022. Results of WMT22 Metrics Shared Task: Stop Using BLEU – Neural Metrics Are Better and More Robust. In Proceedings of the Seventh Conference on Machine Translation, pages 46–68, Abu Dhabi. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.469, + 0.49, + 0.575 + ], + "angle": 0, + "content": "Markus Freitag, Ricardo Rei, Nitika Mathur, Chi-kiu Lo, Craig Stewart, George Foster, Alon Lavie, and Ondrej Bojar. 2021. Results of the WMT21 metrics shared task: Evaluating metrics with expert-based human evaluations on TED and news domain. In Proceedings of the Sixth Conference on Machine Translation, pages 733-774, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.585, + 0.49, + 0.637 + ], + "angle": 0, + "content": "Nuno M. Guerreiro, Elena Voita, and André F. T. Martins. 2022. Looking for a needle in a haystack: A comprehensive study of hallucinations in neural machine translation." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.648, + 0.49, + 0.74 + ], + "angle": 0, + "content": "Sarthak Jain and Byron C. Wallace. 2019. Attention is not Explanation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 3543-3556, Minneapolis, Minnesota. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.75, + 0.49, + 0.843 + ], + "angle": 0, + "content": "Marzena Karpinska, Nishant Raj, Katherine Thai, Yixiao Song, Ankita Gupta, and Mohit Iyyer. 2022. Demetr: Diagnosing evaluation metrics for translation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, page 9540-9561, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.853, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Tom Kocmi, Christian Federmann, Roman Grundkiewicz, Marcin Junczys-Dowmunt, Hitokazu Matsushita, and Arul Menezes. 2021. To ship or not to ship: An extensive evaluation of automatic metrics for machine translation. In Proceedings of the Sixth" + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.53, + 0.086, + 0.884, + 0.113 + ], + "angle": 0, + "content": "Conference on Machine Translation, pages 478-494, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.122, + 0.884, + 0.176 + ], + "angle": 0, + "content": "Christoph Leiter, Piyawat Lertvittayakumjorn, Marina Fomicheva, Wei Zhao, Yang Gao, and Steffen Eger. 2022. Towards explainable evaluation metrics for natural language generation." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.185, + 0.884, + 0.238 + ], + "angle": 0, + "content": "Arle Lommel, Hans Uszkoreit, and Aljoscha Burchardt. 2014. Multidimensional Quality Metrics (MQM): A Framework for Declaring and Describing Translation Quality Metrics. Tradumàtica, pages 0455-463." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.247, + 0.884, + 0.34 + ], + "angle": 0, + "content": "Nitika Mathur, Timothy Baldwin, and Trevor Cohn. 2020. Tangled up in BLEU: Reevaluating the evaluation of automatic machine translation evaluation metrics. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4984-4997, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.348, + 0.884, + 0.441 + ], + "angle": 0, + "content": "Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311-318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.45, + 0.884, + 0.516 + ], + "angle": 0, + "content": "Maja Popovic. 2015. *chrF: character n-gram F-score* for automatic MT evaluation. In *Proceedings of the Tenth Workshop on Statistical Machine Translation*, pages 392–395, Lisbon, Portugal. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.525, + 0.884, + 0.617 + ], + "angle": 0, + "content": "Tharindu Ranasinghe, Constantin Orasan, and Ruslan Mitkov. 2020. *TransQuest: Translation Quality Estimation with Cross-lingual Transformers*. In *Proceedings of the 28th International Conference on Computational Linguistics*, pages 5070–5081, Barcelona, Spain (Online). International Committee on Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.627, + 0.884, + 0.666 + ], + "angle": 0, + "content": "Vikas Raunak, Matt Post, and Arul Menezes. 2022. Salted: A framework for salient long-tail translation error detection." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.676, + 0.884, + 0.781 + ], + "angle": 0, + "content": "Ricardo Rei, José G. C. de Souza, Duarte Alves, Chrysoula Zerva, Ana C Farinha, Taisiya Glushkova, Alon Lavie, Luisa Coheur, and André F. T. Martins. 2022a. COMET-22: Unbabel-IST 2022 Submission for the Metrics Shared Task. In Proceedings of the Seventh Conference on Machine Translation, pages 578-585, Abu Dhabi. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.791, + 0.884, + 0.87 + ], + "angle": 0, + "content": "Ricardo Rei, Craig Stewart, Ana C Farinha, and Alon Lavie. 2020. COMET: A neural framework for MT evaluation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2685-2702, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.879, + 0.884, + 0.919 + ], + "angle": 0, + "content": "Ricardo Rei, Marcos Treviso, Nuno M. Guerreiro, Chrysoula Zerva, Ana C Farinha, Christine Maroti, José G. C. de Souza, Taisiya Glushkova, Duarte" + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.884, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1094" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.135, + 0.086, + 0.487, + 0.165 + ], + "angle": 0, + "content": "Alves, Luisa Coheur, Alon Lavie, and Andre F. T. Martins. 2022b. CometKiwi: IST-Unbabel 2022 Submission for the Quality Estimation Shared Task. In Proceedings of the Seventh Conference on Machine Translation, pages 634-645, Abu Dhabi. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.179, + 0.487, + 0.27 + ], + "angle": 0, + "content": "Marco Ribeiro, Sameer Singh, and Carlos Guestrin. 2016. \"why should I trust you?\": Explaining the predictions of any classifier. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Demonstrations, pages 97-101, San Diego, California. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.284, + 0.486, + 0.35 + ], + "angle": 0, + "content": "Ananya B. Sai, Vignesh Nagarajan, Tanay Dixit, Raj Dabre, Anoop Kunchukuttan, Pratyush Kumar, and Mitesh M. Khapra. 2022. IndicMT Eval: A Dataset to Meta-Evaluate Machine Translation metrics for Indian Languages." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.364, + 0.487, + 0.442 + ], + "angle": 0, + "content": "Thibault Sellam, Dipanjan Das, and Ankur Parikh. 2020. BLEURT: Learning robust metrics for text generation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7881-7892, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.456, + 0.487, + 0.535 + ], + "angle": 0, + "content": "Avanti Shrikumar, Peyton Greenside, and Anshul Kundaje. 2017. Learning Important Features Through Propagating Activation Differences. In Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 3145-3153. PMLR." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.549, + 0.487, + 0.626 + ], + "angle": 0, + "content": "Aaron Smith, Christian Hardmeier, and Joerg Tiedemann. 2016. Climbing mont BLEU: The strange world of reachable high-BLEU translations. In Proceedings of the 19th Annual Conference of the European Association for Machine Translation, pages 269-281." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.641, + 0.487, + 0.707 + ], + "angle": 0, + "content": "Mukund Sundararajan, Ankur Taly, and Qiqi Yan. 2017. Axiomatic Attribution for Deep Networks. In Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 3319-3328. PMLR." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.721, + 0.487, + 0.812 + ], + "angle": 0, + "content": "Shimin Tao, Su Chang, Ma Miaomiao, Hao Yang, Xiang Geng, Shujian Huang, Min Zhang, Jiaxin Guo, Minghan Wang, and Yinglu Li. 2022. CrossQE: HW-TSC 2022 Submission for the Quality Estimation Shared Task. In Proceedings of the Seventh Conference on Machine Translation, pages 646-652, Abu Dhabi. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.826, + 0.487, + 0.918 + ], + "angle": 0, + "content": "Marcos Treviso, Nuno M. Guerreiro, Ricardo Rei, and Andre F. T. Martins. 2021. IST-unbabel 2021 submission for the explainable quality estimation shared task. In Proceedings of the 2nd Workshop on Evaluation and Comparison of NLP Systems, pages 133-145, Punta Cana, Dominican Republic. Association for Computational Linguistics." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.487, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.513, + 0.086, + 0.882, + 0.178 + ], + "angle": 0, + "content": "Yu Wan, Dayiheng Liu, Baosong Yang, Tianchi Bi, Haibo Zhang, Boxing Chen, Weihua Luo, Derek F. Wong, and Lidia S. Chao. 2021. RoBLEURT submission for WMT2021 metrics task. In Proceedings of the Sixth Conference on Machine Translation, pages 1053-1058, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.188, + 0.882, + 0.28 + ], + "angle": 0, + "content": "Yu Wan, Dayiheng Liu, Baosong Yang, Haibo Zhang, Boxing Chen, Derek Wong, and Lidia Chao. 2022. UniTE: Unified translation evaluation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8117-8127, Dublin, Ireland. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.29, + 0.882, + 0.33 + ], + "angle": 0, + "content": "Kerem Zaman and Yonatan Belinkov. 2022. A Multilingual Perspective Towards the Evaluation of Attribution Methods in Natural Language Inference." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.339, + 0.882, + 0.457 + ], + "angle": 0, + "content": "Chrysoula Zerva, Frédéric Blain, Ricardo Rei, Piyawat Lertvittayakumjorn, José G. C. de Souza, Steffen Eger, Diptesh Kanojia, Duarte Alves, Constantin Orasan, Marina Fomicheva, André F. T. Martins, and Lucia Specia. 2022. Findings of the WMT 2022 Shared Task on Quality Estimation. In Proceedings of the Seventh Conference on Machine Translation, pages 69-99, Abu Dhabi. Association for Computational Linguistics." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.882, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1095" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.115, + 0.085, + 0.274, + 0.099 + ], + "angle": 0, + "content": "A Model Details" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.11, + 0.49, + 0.398 + ], + "angle": 0, + "content": "In Section 2.1, we employed the latest publicly available model (wmt22-comet-da) for COMET, which emerged as a top-performing metric in the WMT 2022 Metrics task (Freitag et al., 2022). To ensure a comparable setting for UNITE (Wan et al., 2022), we trained our own model. In doing so, we utilized the same data employed in the development of the COMET model by (Rei et al., 2022a), without pretraining any synthetic data, as originally suggested. Additionally, our implementation did not incorporate monotonic regional attention, as our preliminary experiments revealed no discernible benefits from its usage. The hyperparameters used are summarized in Table 3, while Table 4 presents the number of Direct Assessments utilized during training. Furthermore, Table 5 displays the segment-level correlations with WMT 2021 MQM data for the News and TED domains." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.4, + 0.49, + 0.464 + ], + "angle": 0, + "content": "Regarding computational infrastructure, a single NVIDIA A10G GPU with 23GB memory was used. The resulting UNITE model has 565M parameters while COMET has 581M parameters." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.476, + 0.322, + 0.491 + ], + "angle": 0, + "content": "A.1 Output Distribution" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.496, + 0.49, + 0.641 + ], + "angle": 0, + "content": "To better understand the output of the models and what scores are deemed low, we plotted the output distributions for the two models we used in our study. The average score for English \\(\\rightarrow\\) German data is 0.856 for the COMET model and 0.870 for the UNITE model we trained. From Figure 3 we can observe the distribution of scores. This means that the 0.6692 score from the example in Figure 1 corresponds to a low quality output (5th percentile)." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.652, + 0.295, + 0.668 + ], + "angle": 0, + "content": "A.2 SMAUG Corpus" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.673, + 0.49, + 0.785 + ], + "angle": 0, + "content": "As we have seen in Section 4.2, we have created synthetic translation errors for the following pathologies: negation errors, hallucinations via insertions, named entity errors, and errors in numbers. Table 7 presents a summary of the examples created using SMAUG and in Table 8 we show examples of each error category." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.797, + 0.446, + 0.831 + ], + "angle": 0, + "content": "B Comparison between COMET and XLM-R Alignments" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.84, + 0.49, + 0.919 + ], + "angle": 0, + "content": "From Table 1, it is evident that the alignments between the reference and/or source and the translation yield effective explanations for COMET. This raises the question of how these alignments compare to the underlying encoder of COMET before" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.882, + 0.133 + ], + "angle": 0, + "content": "the fine-tuning process with human annotations. To investigate this, we examine the results for XLM-R without any fine-tuning, as presented in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.134, + 0.885, + 0.244 + ], + "angle": 0, + "content": "Overall, the explanations derived from the alignments of COMET prove to be more predictive of error spans than those obtained from XLM-R alignments. This suggests that during the fine-tuning phase, COMET models modify the underlying XLM-R representations to achieve better alignment with translation errors." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.257, + 0.634, + 0.274 + ], + "angle": 0, + "content": "C Examples" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.283, + 0.884, + 0.396 + ], + "angle": 0, + "content": "In Tables 9 and 10, we show examples of COMET explanations for Chinese \\(\\rightarrow\\) English and English \\(\\rightarrow\\) German language pairs, respectively. We highlight in gray the corresponding MQM annotation performed by an expert linguist and we sort the examples from highest to lowest COMET scores. From these examples we can observe the following:" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.407, + 0.882, + 0.486 + ], + "angle": 0, + "content": "- Highlights provided by COMET explanations have a high recall with human annotations. In all examples, subword tokens corresponding to translation errors are highlighted in red but we often see that not everything is incorrect." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.498, + 0.884, + 0.626 + ], + "angle": 0, + "content": "- Explanations are consistent with scores. For example, in the third example from Table 10, the red highlights do not correspond to errors and in fact the translation only has a major error griffen. Nonetheless, the score assigned by COMET is a low score of 0.68 which is faithful to the explanations that was given even if the assessment does not agree with human experts." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1096" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.186, + 0.082, + 0.813, + 0.2 + ], + "angle": 0, + "content": "
METRICEXPLAINABILITY METHODen→dezh→enen→ruAvg.
AUCR@KAUCR@KAUCR@KAUCR@K
XLM-Rembed-align[mt, src]0.5870.3590.6680.3110.5760.1990.6100.289
embed-align[mt, ref]0.6710.4050.6890.3450.6340.2440.6640.331
embed-align[mt, src; ref]0.6660.3950.6900.3470.6160.2420.6570.328
COMETembed-align[mt, src]0.5900.3710.6740.3140.5770.2200.6140.301
embed-align[mt, ref]0.6940.4250.6960.3550.6470.2750.6790.352
embed-align[mt, src; ref]0.6880.4160.6970.3570.6220.2790.6690.350
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.209, + 0.884, + 0.237 + ], + "angle": 0, + "content": "Table 2: AUC and Recall@K of explanations obtained via alignments for COMET and XLM-R without any further fine-tuning on human annotations." + }, + { + "type": "table", + "bbox": [ + 0.166, + 0.453, + 0.437, + 0.635 + ], + "angle": 0, + "content": "
HyperparameterUNITECOMET
Encoder ModelXLM-R (large)
OptimizerAdamW
No. frozen epochs0.3
Learning rate (LR)1.5e-05
Encoder LR.1.0e-06
Layerwise Decay0.95
Batch size16
Loss functionMSE
Dropout0.1
Hidden sizes[3072, 1024]
Embedding layerFrozen
FP precision16
No. Epochs12
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.646, + 0.49, + 0.717 + ], + "angle": 0, + "content": "Table 3: Hyperparameters used to train UNITE and COMET checkpoints used in this work. The only difference between the two is the number of training epochs due to the fact that, for UNITE, the best validation checkpoint is the first one." + }, + { + "type": "table", + "bbox": [ + 0.601, + 0.326, + 0.794, + 0.789 + ], + "angle": 0, + "content": "
Language PairSIZE
zh-en126947
en-de121420
de-en99183
en-zh90805
ru-en79280
en-ru62749
en-CS60937
fi-en46145
en-fi34335
tr-en30186
et-en29496
cs-en27847
en-mr26000
de-CS13804
en-et13376
pl-en11816
en-pl10572
lt-en10315
en-ja9578
gu-en9063
si-en9000
ro-en9000
ne-en9000
en-lt8959
ja-en8939
en-kk8219
en-ta7890
ta-en7577
en-gu6924
kk-en6789
de-fr6691
en-lv5810
en-tr5171
km-en4722
ps-en4611
fr-de3999
Total1027155
" + }, + { + "type": "table_caption", + "bbox": [ + 0.509, + 0.8, + 0.882, + 0.843 + ], + "angle": 0, + "content": "Table 4: Number of direct assessments per language pair used to train COMET (Rei et al., 2022a) and the UNITE model used in this work." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1097" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.159, + 0.372, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.159, + 0.625, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.159, + 0.876, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.463, + 0.304, + 0.536, + 0.316 + ], + "angle": 0, + "content": "(a)COMET" + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.325, + 0.372, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.325, + 0.622, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.325, + 0.876, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.446, + 0.469, + 0.552, + 0.482 + ], + "angle": 0, + "content": "(b) UNITE SRC" + }, + { + "type": "image", + "bbox": [ + 0.12, + 0.491, + 0.372, + 0.627 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.491, + 0.622, + 0.627 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.491, + 0.874, + 0.627 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.449, + 0.636, + 0.55, + 0.648 + ], + "angle": 0, + "content": "(c) UNITE REF" + }, + { + "type": "image", + "bbox": [ + 0.12, + 0.658, + 0.372, + 0.793 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.658, + 0.622, + 0.793 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.658, + 0.874, + 0.793 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.431, + 0.801, + 0.567, + 0.814 + ], + "angle": 0, + "content": "(d) UNITE SRC+REF" + }, + { + "type": "image_caption", + "bbox": [ + 0.167, + 0.825, + 0.828, + 0.84 + ], + "angle": 0, + "content": "Figure 3: Distribution of scores for all metrics obtained on the MQM data (for all language pairs)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1098" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.115, + 0.346, + 0.884, + 0.584 + ], + "angle": 0, + "content": "
BLEUCHRFYISI-1BLEURTUNITE\nSRCUNITE\nREFUNITE\nSRC+REFCOMET\nwmt22-comet-da
EN→DENEWSρ0.0770.0920.1630.3070.2740.3210.304
τ0.0690.0920.1440.2400.2220.2480.241
ρ0.1510.1580.2360.3250.3110.3350.338
τ0.1130.1460.2120.2830.2640.3010.298
EN→RUTED Newsρ0.1530.2520.2630.3590.3330.3910.382
τ0.1060.1780.2160.2760.2760.2980.297
ρ0.1540.2680.2350.2860.2390.2890.318
τ0.1120.1890.2040.2550.2320.2620.264
ZH→ENTED Newsρ0.2150.2310.3010.4280.4130.4380.426
τ0.1650.1880.2890.3410.3310.3580.352
ρ0.1550.1810.2870.2950.2440.3010.310
τ0.1130.1440.2160.2460.2240.2650.266
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.594, + 0.884, + 0.651 + ], + "angle": 0, + "content": "Table 5: Segment-level correlations for WMT 2021 MQM annotations over News and TED domains (Freitag et al., 2021). The metrics are Pearson \\((\\rho)\\) and Kendall Tau \\((\\tau)\\). Results in bold indicate which metrics are top-performing for that specific language pair, domain and metric according to Perm-Both hypothesis test (Deutsch et al., 2021), using 500 re-sampling runs, and setting \\(p = 0.05\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1099" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.189, + 0.222, + 0.416, + 0.317 + ], + "angle": 0, + "content": "
Error TypeNUM EXAMPLES
NE978
NEG669
HALL530
NUM432
Total2609
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.327, + 0.49, + 0.37 + ], + "angle": 0, + "content": "Table 6: Number of examples for each category, synthetically-created using SMAUG (Alves et al., 2022)." + }, + { + "type": "table", + "bbox": [ + 0.116, + 0.658, + 0.489, + 0.724 + ], + "angle": 0, + "content": "
Language PairTOKENS / SENT.ERRORS / SPANS
en-de528704 / 1531025712 / 3567
en-ru525938 / 1507417620 / 7172
zh-en603258 / 1650643984 / 10042
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.732, + 0.49, + 0.775 + ], + "angle": 0, + "content": "Table 7: Statistics about MQM data from WMT 2021 Metrics task (Freitag et al., 2021) used in our experiments." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1100" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.127, + 0.239, + 0.178, + 0.25 + ], + "angle": 0, + "content": "Source:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.251, + 0.529, + 0.263 + ], + "angle": 0, + "content": "格里沃里表示，分析人士对越南所提出的和平倡议给予认可。" + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.268, + 0.208, + 0.278 + ], + "angle": 0, + "content": "Translation:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.279, + 0.598, + 0.291 + ], + "angle": 0, + "content": "Grivory said that analysts recognize the peace initiative proposed by Vietnam." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.293, + 0.198, + 0.304 + ], + "angle": 0, + "content": "Reference:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.305, + 0.587, + 0.318 + ], + "angle": 0, + "content": "Grigory said that analysts endorse the peace initiative proposed by Vietnam." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.32, + 0.197, + 0.331 + ], + "angle": 0, + "content": "NE Error:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.332, + 0.593, + 0.345 + ], + "angle": 0, + "content": "Grivory said that analysts recognize the peace initiative proposed by Russia." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.36, + 0.178, + 0.37 + ], + "angle": 0, + "content": "Source:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.37, + 0.484, + 0.383 + ], + "angle": 0, + "content": "英国的这一决定预计将会使西班牙的旅游业大受影响。" + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.387, + 0.208, + 0.397 + ], + "angle": 0, + "content": "Translation:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.398, + 0.679, + 0.411 + ], + "angle": 0, + "content": "This decision by the United Kingdom is expected to greatly affect Spain's tourism industry." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.413, + 0.198, + 0.424 + ], + "angle": 0, + "content": "Reference:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.425, + 0.642, + 0.438 + ], + "angle": 0, + "content": "This decision by the UK is expected to have a significant impact on tourism in Spain." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.44, + 0.205, + 0.451 + ], + "angle": 0, + "content": "NEG Error:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.452, + 0.692, + 0.466 + ], + "angle": 0, + "content": "This decision by the United Kingdom is expected to greatly benefit Spain's tourism industry." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.479, + 0.178, + 0.49 + ], + "angle": 0, + "content": "Source:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.491, + 0.468, + 0.504 + ], + "angle": 0, + "content": "由于疫情，人们开始在互联网上花费更多的时间。”" + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.507, + 0.208, + 0.518 + ], + "angle": 0, + "content": "Translation:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.519, + 0.617, + 0.531 + ], + "angle": 0, + "content": "\"Because of the epidemic, people are starting to spend more time on the Internet.\"" + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.534, + 0.198, + 0.544 + ], + "angle": 0, + "content": "Reference:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.545, + 0.639, + 0.558 + ], + "angle": 0, + "content": "For reason of the pandemic, people are starting to spend more time on the Internet." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.561, + 0.214, + 0.571 + ], + "angle": 0, + "content": "HALL Error:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.573, + 0.856, + 0.599 + ], + "angle": 0, + "content": "Because we have a lot of friends around during the epidemic, people are starting to spend more time on the mobile devices than on the Internet.\"" + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.614, + 0.178, + 0.625 + ], + "angle": 0, + "content": "Source:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.625, + 0.28, + 0.638 + ], + "angle": 0, + "content": "展销区将展至7月29日。" + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.641, + 0.208, + 0.652 + ], + "angle": 0, + "content": "Translation:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.653, + 0.465, + 0.666 + ], + "angle": 0, + "content": "The exhibition and sales area will be open until July 29." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.668, + 0.198, + 0.679 + ], + "angle": 0, + "content": "Reference:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.68, + 0.351, + 0.692 + ], + "angle": 0, + "content": "The exhibition will last until July 29." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.695, + 0.208, + 0.706 + ], + "angle": 0, + "content": "NUM Error:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.707, + 0.481, + 0.72 + ], + "angle": 0, + "content": "The exhibition and sales area will be open until July 2018" + }, + { + "type": "table_caption", + "bbox": [ + 0.115, + 0.735, + 0.882, + 0.763 + ], + "angle": 0, + "content": "Table 8: Synthetically-generated critical errors (highlighted in gray) created with SMAUG (Alves et al., 2022) to assess whether our explanations can be accurately attributed to critical errors." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.519, + 0.941 + ], + "angle": 0, + "content": "1101" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.127, + 0.182, + 0.178, + 0.193 + ], + "angle": 0, + "content": "Source:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.194, + 0.581, + 0.206 + ], + "angle": 0, + "content": "And yet, the universe is not a silent movie because the universe isn't silent." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.209, + 0.208, + 0.22 + ], + "angle": 0, + "content": "Translation:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.221, + 0.637, + 0.233 + ], + "angle": 0, + "content": "Und Dennoch ist das Universum kein Stummfilm, weil das Universum nicht still ist." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.236, + 0.271, + 0.247 + ], + "angle": 0, + "content": "COMET score: 0.8595" + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.248, + 0.265, + 0.26 + ], + "angle": 0, + "content": "COMET explanation:" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.261, + 0.846, + 0.273 + ], + "angle": 0, + "content": "_Und _dennoch _ist _das _Univers um _kein _Stu mm film , _weil _das _Univers um _nicht _still _ist ." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.285, + 0.178, + 0.295 + ], + "angle": 0, + "content": "Source:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.296, + 0.821, + 0.308 + ], + "angle": 0, + "content": "And yet black holes may be heard even if they're not seen, and that's because they bang on space-time like a drum." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.311, + 0.208, + 0.322 + ], + "angle": 0, + "content": "Translation:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.323, + 0.855, + 0.347 + ], + "angle": 0, + "content": "Und dennoch werden Schwarze Locher weitereicht gehört, auch wenn sie nicht gesehen werden, und das liegt daran, dass sie wie eine Trommel auf die Raumzeit schlagen." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.35, + 0.27, + 0.362 + ], + "angle": 0, + "content": "COMET score: 0.7150" + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.363, + 0.265, + 0.374 + ], + "angle": 0, + "content": "COMET explanation:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.375, + 0.854, + 0.402 + ], + "angle": 0, + "content": "Und dennoch werden Schwarz e LÖcher vielleicht gehört , auch wenn sie nicht gesehenwerden , und das liegt daran , dass sie wie eine Tro mmel auf die Raum zeit schlagen ." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.414, + 0.178, + 0.424 + ], + "angle": 0, + "content": "Source:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.425, + 0.855, + 0.449 + ], + "angle": 0, + "content": "Ash O'Brien and husband Jarett Kelley say they were grabbing a bite to eat at Dusty Rhodes dog park in San Diego on Thursday, with their three-month-old pug in tow." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.452, + 0.208, + 0.462 + ], + "angle": 0, + "content": "Translation:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.464, + 0.855, + 0.488 + ], + "angle": 0, + "content": "Ash O'Brien und Ehemann Jarett Kelley sagen, dass sie am Donnerstag im Hundepark Dusty Rhodes in San Diego einen Happen zu essen griffen, mit ihrem drei Monate alten Mops im Schleptau." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.491, + 0.27, + 0.502 + ], + "angle": 0, + "content": "COMET score: 0.6835" + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.503, + 0.265, + 0.514 + ], + "angle": 0, + "content": "COMET explanation:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.515, + 0.854, + 0.557 + ], + "angle": 0, + "content": "_Ash_O' Brien_Und_Ehe mann_Ja rett_Kelley_sagen, dass sie_am_Donnerstag_im_Hunde park_Dusty_Rhod es_in_San_Diego_einen_Happ_en_zu_essen_griff_en_, _mit_threm_drei_Monate_alten_M ops_im_Schleppt au." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.568, + 0.178, + 0.579 + ], + "angle": 0, + "content": "Source:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.58, + 0.43, + 0.593 + ], + "angle": 0, + "content": "It was Einstein's great general theory of relativity." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.595, + 0.208, + 0.606 + ], + "angle": 0, + "content": "Translation:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.607, + 0.565, + 0.62 + ], + "angle": 0, + "content": "Es war Einsteins große allgemeine Forschungen vor Relativitätstheorie." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.622, + 0.27, + 0.634 + ], + "angle": 0, + "content": "COMET score: 0.6692" + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.635, + 0.265, + 0.647 + ], + "angle": 0, + "content": "COMET explanation:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.647, + 0.737, + 0.66 + ], + "angle": 0, + "content": "_Es _war _Einstein s _große _allgemeine e _Forschung en _vor _Relativ itäts the ori e ." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.671, + 0.178, + 0.682 + ], + "angle": 0, + "content": "Source:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.683, + 0.457, + 0.695 + ], + "angle": 0, + "content": "There's mask-shaming and then there's full on assault." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.698, + 0.208, + 0.709 + ], + "angle": 0, + "content": "Translation:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.71, + 0.486, + 0.723 + ], + "angle": 0, + "content": "Es gibt Maskenschämen und dann ist es voll bei Angriff." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.725, + 0.27, + 0.736 + ], + "angle": 0, + "content": "COMET score: 0.2318" + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.737, + 0.265, + 0.749 + ], + "angle": 0, + "content": "COMET explanation:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.749, + 0.648, + 0.763 + ], + "angle": 0, + "content": "_Es_gibt_Mask en schä men_und_dann_ist es_voll_bei_Angriff_." + }, + { + "type": "table_caption", + "bbox": [ + 0.117, + 0.778, + 0.882, + 0.82 + ], + "angle": 0, + "content": "Table 9: Saliency map for COMET explanation scores for a set of en→de examples. Comparing the token-level explanations with the MQM annotation (highlighted in gray) reveals the source of correspondence between specific token-level translation errors and the resulting scores." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1102" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.126, + 0.185, + 0.178, + 0.195 + ], + "angle": 0, + "content": "Source:" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.196, + 0.782, + 0.209 + ], + "angle": 0, + "content": "我想告诉大家宇宙有着自己的配乐，而宇宙自身正在不停地播放着。因为太空可以想鼓一样振动。" + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.212, + 0.208, + 0.223 + ], + "angle": 0, + "content": "Translation:" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.224, + 0.856, + 0.248 + ], + "angle": 0, + "content": "I want to tell you that the universe has its own iconic soundtrack and the universe itself is constantly playing non-stop because space can vibrate like a drum." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.251, + 0.27, + 0.262 + ], + "angle": 0, + "content": "COMET score: 0.8634" + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.264, + 0.265, + 0.275 + ], + "angle": 0, + "content": "COMET explanation:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.275, + 0.854, + 0.303 + ], + "angle": 0, + "content": "_I_want_to_tell你想_the_univers_e_has_its_own(iconic soundtrack_and_the_univers_e_itself_is_constantly-playing_non-stop_because_space_can_vibrate_to_like_a Drum." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.314, + 0.178, + 0.325 + ], + "angle": 0, + "content": "Source:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.325, + 0.611, + 0.339 + ], + "angle": 0, + "content": "另外，吉克隽逸和刘烨作为运动助理，也围绕运动少年制造了不少爆笑话题。" + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.342, + 0.208, + 0.353 + ], + "angle": 0, + "content": "Translation:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.354, + 0.856, + 0.367 + ], + "angle": 0, + "content": "In addition, as sports assistants, Ji Kejunyi and Liu Ye have also created a lot of hilarious topics around sports teenagers." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.369, + 0.27, + 0.381 + ], + "angle": 0, + "content": "COMET score: 0.8214" + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.382, + 0.265, + 0.394 + ], + "angle": 0, + "content": "COMET explanation:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.394, + 0.854, + 0.422 + ], + "angle": 0, + "content": "_In _addition , _as _sports _assistant s , _Ji _Ke ju nyi _and _Li u _Ye _have _also _created _a _lot _of_ hila rious _topic s _around _sports _teenager s ." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.433, + 0.178, + 0.443 + ], + "angle": 0, + "content": "Source:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.444, + 0.484, + 0.457 + ], + "angle": 0, + "content": "一番言论让场上的少年和运动领队们都倒吸一口凉气。" + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.46, + 0.208, + 0.471 + ], + "angle": 0, + "content": "Translation:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.472, + 0.661, + 0.485 + ], + "angle": 0, + "content": "The remarks made the teenagers and the sports leaders on the field gasp a sigh of relief." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.487, + 0.27, + 0.499 + ], + "angle": 0, + "content": "COMET score: 0.7793" + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.5, + 0.265, + 0.512 + ], + "angle": 0, + "content": "COMET explanation:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.512, + 0.856, + 0.54 + ], + "angle": 0, + "content": "The re marks made the teenager s and the sports leaders on the field gas p a sig h of _relief" + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.551, + 0.178, + 0.561 + ], + "angle": 0, + "content": "Source:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.562, + 0.3, + 0.575 + ], + "angle": 0, + "content": "强烈的阳光是如此地刺眼，" + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.578, + 0.208, + 0.589 + ], + "angle": 0, + "content": "Translation:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.59, + 0.327, + 0.603 + ], + "angle": 0, + "content": "The intense sunlight is so harsh;" + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.605, + 0.269, + 0.617 + ], + "angle": 0, + "content": "COMET score: 0.7561" + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.617, + 0.265, + 0.629 + ], + "angle": 0, + "content": "COMET explanation:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.629, + 0.424, + 0.643 + ], + "angle": 0, + "content": "The intense sun light is so har sh" + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.655, + 0.178, + 0.665 + ], + "angle": 0, + "content": "Source:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.665, + 0.57, + 0.678 + ], + "angle": 0, + "content": "如今，我们希望能够给这部关于宇宙的宏伟的视觉作品配上声音。" + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.681, + 0.208, + 0.692 + ], + "angle": 0, + "content": "Translation:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.693, + 0.667, + 0.707 + ], + "angle": 0, + "content": "Today, we hope to be able to give this magnificent visual work of the universe a sound." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.709, + 0.27, + 0.721 + ], + "angle": 0, + "content": "COMET score: 0.7073" + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.721, + 0.265, + 0.733 + ], + "angle": 0, + "content": "COMET explanation:" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.733, + 0.856, + 0.761 + ], + "angle": 0, + "content": "Today,we hope to be able to give this magnificent ent visual work of the univers e a sound." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.775, + 0.882, + 0.818 + ], + "angle": 0, + "content": "Table 10: Saliency map for COMET explanation scores for a set of \\( \\mathrm{{zh}} \\rightarrow \\mathrm{{en}} \\) examples. Comparing the token-level explanations with the MQM annotation (highlighted in gray) reveals the source of correspondence between specific token-level translation errors and the resulting scores." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1103" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.133, + 0.085, + 0.435, + 0.1 + ], + "angle": 0, + "content": "ACL 2023 Responsible NLP Checklist" + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.108, + 0.323, + 0.123 + ], + "angle": 0, + "content": "A For every submission:" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.127, + 0.533, + 0.143 + ], + "angle": 0, + "content": "A1. Did you describe the limitations of your work?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.145, + 0.258, + 0.158 + ], + "angle": 0, + "content": "Yes. Section 6" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.17, + 0.553, + 0.186 + ], + "angle": 0, + "content": "A2. Did you discuss any potential risks of your work?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.187, + 0.352, + 0.202 + ], + "angle": 0, + "content": "Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.212, + 0.697, + 0.228 + ], + "angle": 0, + "content": "A3. Do the abstract and introduction summarize the paper's main claims?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.23, + 0.322, + 0.243 + ], + "angle": 0, + "content": "Abstract and Section 1" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.254, + 0.671, + 0.271 + ], + "angle": 0, + "content": "A4. Have you used AI writing assistants when working on this paper?" + }, + { + "type": "text", + "bbox": [ + 0.148, + 0.272, + 0.882, + 0.288 + ], + "angle": 0, + "content": "Assistance purely with the language of the paper along every section. Grammarly and DeepL write" + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.298, + 0.49, + 0.314 + ], + "angle": 0, + "content": "B Did you use or create scientific artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.319, + 0.884, + 0.351 + ], + "angle": 0, + "content": "Section 3 explains the methods we used. We will release the adaptations required to use the explainability methods over COMET framework, the UniTE model we trained, and all data synthetically-generated data." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.36, + 0.531, + 0.376 + ], + "angle": 0, + "content": "B1. Did you cite the creators of artifacts you used?" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.378, + 0.224, + 0.391 + ], + "angle": 0, + "content": "Section 2" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.402, + 0.779, + 0.419 + ], + "angle": 0, + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.42, + 0.58, + 0.435 + ], + "angle": 0, + "content": "footnote on the first page. The License will be Apache 2.0" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.446, + 0.882, + 0.509 + ], + "angle": 0, + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.511, + 0.352, + 0.526 + ], + "angle": 0, + "content": "Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.537, + 0.882, + 0.585 + ], + "angle": 0, + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.586, + 0.352, + 0.601 + ], + "angle": 0, + "content": "Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.611, + 0.882, + 0.643 + ], + "angle": 0, + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.644, + 0.677, + 0.659 + ], + "angle": 0, + "content": "in the Appendix we have several statistics for training and testing data." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.669, + 0.884, + 0.75 + ], + "angle": 0, + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be." + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.751, + 0.233, + 0.766 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.775, + 0.494, + 0.792 + ], + "angle": 0, + "content": "C Did you run computational experiments?" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.797, + 0.612, + 0.813 + ], + "angle": 0, + "content": "Appendix provides detailed information about the trained model." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.821, + 0.882, + 0.855 + ], + "angle": 0, + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?" + }, + { + "type": "text", + "bbox": [ + 0.148, + 0.856, + 0.882, + 0.887 + ], + "angle": 0, + "content": "Appendix provides detailed information about the trained model including GPU infrastructure and total number of parameters." + }, + { + "type": "footer", + "bbox": [ + 0.114, + 0.894, + 0.878, + 0.918 + ], + "angle": 0, + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1104" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.129, + 0.085, + 0.88, + 0.116 + ], + "angle": 0, + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.118, + 0.351, + 0.133 + ], + "angle": 0, + "content": "Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.143, + 0.884, + 0.191 + ], + "angle": 0, + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?" + }, + { + "type": "text", + "bbox": [ + 0.148, + 0.193, + 0.775, + 0.209 + ], + "angle": 0, + "content": "Appendix has all information needed about test data and performance of the models." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.218, + 0.884, + 0.266 + ], + "angle": 0, + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.268, + 0.331, + 0.284 + ], + "angle": 0, + "content": "Section 2 and Appendix" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.294, + 0.878, + 0.311 + ], + "angle": 0, + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.315, + 0.215, + 0.33 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.341, + 0.883, + 0.373 + ], + "angle": 0, + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.374, + 0.351, + 0.39 + ], + "angle": 0, + "content": "Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.4, + 0.884, + 0.448 + ], + "angle": 0, + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.449, + 0.351, + 0.465 + ], + "angle": 0, + "content": "Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.476, + 0.884, + 0.523 + ], + "angle": 0, + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.524, + 0.351, + 0.54 + ], + "angle": 0, + "content": "Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.55, + 0.875, + 0.566 + ], + "angle": 0, + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.567, + 0.351, + 0.583 + ], + "angle": 0, + "content": "Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.593, + 0.881, + 0.624 + ], + "angle": 0, + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.626, + 0.351, + 0.642 + ], + "angle": 0, + "content": "Not applicable. Left blank." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1105" + } + ] +] \ No newline at end of file diff --git a/2023/The Inside Story_ Towards Better Understanding of Machine Translation Neural Evaluation Metrics/d25ab395-c57a-4c68-8e6d-00772d415f43_origin.pdf b/2023/The Inside Story_ Towards Better Understanding of Machine Translation Neural Evaluation Metrics/d25ab395-c57a-4c68-8e6d-00772d415f43_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..17094f0639593d4d82d17a99d98b5b0fcf6c2066 --- /dev/null +++ b/2023/The Inside Story_ Towards Better Understanding of Machine Translation Neural Evaluation Metrics/d25ab395-c57a-4c68-8e6d-00772d415f43_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b034a429926c05c9ce95b679d8b5303e401a87603a0168553c65028e81fb9bc3 +size 1017018 diff --git a/2023/The Inside Story_ Towards Better Understanding of Machine Translation Neural Evaluation Metrics/full.md b/2023/The Inside Story_ Towards Better Understanding of Machine Translation Neural Evaluation Metrics/full.md new file mode 100644 index 0000000000000000000000000000000000000000..75db8bfba33b15094e874ac1277b42430aff9a62 --- /dev/null +++ b/2023/The Inside Story_ Towards Better Understanding of Machine Translation Neural Evaluation Metrics/full.md @@ -0,0 +1,542 @@ +# The Inside Story: Towards Better Understanding of Machine Translation Neural Evaluation Metrics + +Ricardo Rei\*1,2,4, Nuno M. Guerreiro\*3,4, Marcos Treviso\*3,4, Alon Lavie\*1, Luisa Coheur\*2,4, Andre F. T. Martins\*1,3,4 + +1Unbabel, Lisbon, Portugal, 2INESC-ID, Lisbon, Portugal +3Instituto de Telecomunicações, Lisbon, Portugal +$^{4}$ Instituto Superior Técnico, University of Lisbon, Portugal + +# Abstract + +Neural metrics for machine translation evaluation, such as COMET, exhibit significant improvements in their correlation with human judgments compared to traditional metrics based on lexical overlap, such as BLEU. Yet neural metrics are, to a great extent, "black boxes" that return a single sentence-level score without transparency about the decision-making process. In this work, we develop and compare several neural explainability methods and demonstrate their effectiveness for interpreting state-of-the-art fine-tuned neural metrics. Our study reveals that these metrics leverage token-level information that can be directly attributed to translation errors, as assessed through comparison of token-level neural saliency maps with Multidimensional Quality Metrics (MQM) annotations and with synthetically-generated critical translation errors. To ease future research, we release our code at https://github.com/Unbabel/COMET/tree/explainable-metrics. + +# 1 Introduction + +Reference-based neural metrics for machine translation evaluation are achieving evergrowing success, demonstrating superior results over traditional lexical overlap-based metrics, such as BLEU (Papineni et al., 2002) and CHRF (Popovic, 2015), in terms of both their correlation with human ratings and their robustness across diverse domains (Callison-Burch et al., 2006; Smith et al., 2016; Mathur et al., 2020; Kocmi et al., 2021; Freitag et al., 2022). However, lexical overlap-based metrics remain popular for evaluating the performance and progress of translation systems and algorithms. Concerns regarding trust and interpretability may help explain this (Leiter et al., 2022): contrary to traditional metrics, neural metrics are considered "black boxes" as they often use + +![](images/c81eeb83a2f01d61f4e4656faed8f88c417f549a3f6570b985690eea98bdeaf0.jpg) +Figure 1: Illustration of our approach. In this example, the metric assigns the translation a low score. We aim to better understand this sentence-level assessment by examining the correspondence between our token-level explanations and human annotated error spans. + +increasingly large models (e.g., the winning metric of the WMT 22 Metrics shared task was a 10B parameter model (Freitag et al., 2022)). + +While some recent work has focused on explaining the predictions made by reference-free quality estimation (QE) systems (Fomicheva et al., 2021; Zerva et al., 2022), explaining reference-based metrics has remained a largely overlooked problem (Leiter et al., 2022). It is an open question whether the observations from studies of explainable QE carry over to this scenario. Thus, in this work, we fill that gap by turning to state-of-the-art reference-based metrics—we aim to interpret their decision-making process by exploiting the fact that these metrics show consistently good correlations with Multidimensional Quality Metrics (MQM) (Freitag et al., 2021, 2022; Sai et al., 2022), which are fine-grained quality assessments that result from experts identifying error spans in translation outputs (Lommel et al., 2014). We hypothesize that reference-based metrics leverage this token-level information to produce sentence-level scores. To test this hypothesis, we assess whether our explanations – measures of token-level importance obtained via attribution and input attribution methods such as attention weights and gradient scores (Treviso et al., 2021; Rei et al., 2022b) – align with + +human-annotated spans (Fomicheva et al., 2021, 2022; Zerva et al., 2022), as illustrated in Figure 1. + +Our analysis focuses on two main vectors: (i) understanding the impact of the reference information on the quality of the explanations; and (ii) finding whether the explanations can help to identify potential weaknesses in the metrics. Our main contributions are: + +- We provide a comparison between multiple explainability methods for different metrics on all types of evaluation: src-only, ref-only, and src+ref joint evaluation. +- We find that explanations are related to the underlying metric architecture, and that leveraging reference information improves the explanations. +- We show that explanations for critical translation errors can reveal weaknesses in the metrics. + +# 2 Explaining Neural Metrics + +We aim to explain sentence-level quality assessments of reference-based metrics by producing token-level explanations that align to translation errors. In what follows, we describe the metrics and how we produce the explanations that we study. + +# 2.1 Metrics + +We focus our analysis on two state-of-the-art neural metrics: COMET (Rei et al., 2020) and UNITE (Wan et al., 2022). While both metrics use a multilingual encoder model based on XLMR (Conneau et al., 2020), they employ distinct strategies to obtain sentence-level quality scores. On the one hand, COMET separately encodes the source, translation and reference to obtain their respective sentence embeddings; these embeddings are then combined to compute a quality score. On the other, UNITE jointly encodes the sentences to compute a contextualized representation that is subsequently used to compute the quality score. Interestingly, UNITE is trained to obtain quality scores for different input combinations: [mt; src] (SRC), [mt; ref] (REF), and [mt; src; ref] (SRC+REF). In fact, when the input is SRC, UNITE works like TransQuest (Ranasinghe et al., 2020); REF, like BLEURT (Sellam et al., 2020); and SRC+REF, like ROBLEURT (Wan et al., 2021). + +# 2.2 Explanations via Attribution Methods + +In this work, we produce explanations using attribution methods that assign a scalar value to each translation token (i.e. a token-level attribution) to represent its importance. While many input attribution methods exist and have been extensively studied in the literature (Ribeiro et al., 2016; Shrikumar et al., 2017; Sundararajan et al., 2017; Jain and Wallace, 2019; Atanasova et al., 2020; Zaman and Belinkov, 2022), we focus specifically on those that have been demonstrated to be effective for explaining the predictions of QE models (Treviso et al., 2021; Fomicheva et al., 2022; Fernandes et al., 2022; Zerva et al., 2022) and extend them to our reference-based scenario. Concretely, we use the following techniques to extract explanations: + +- embed-align: the maximum cosine similarity between each translation token embedding and the reference and/or source token embeddings (Tao et al., 2022); +- grad $\ell_2$ : the $\ell_2$ -norm of gradients with respect to the word embeddings of the translation tokens (Arras et al., 2019); +- attention: the attention weights of the translation tokens for each attention head of the encoder (Treviso et al., 2021); +- $\mathbf{attn} \times \mathbf{grad}$ : the attention weights of each head scaled by the $\ell_2$ -norm of the gradients of the value vectors of that head (Rei et al., 2022b). + +# 3 Experimental Setting + +MQM annotations. We use MQM annotations from the WMT 2021 Metrics shared task (Freitag et al., 2021),3 covering three language pairs — English-German (en→de), English-Russian (en→ru), and Chinese-English (zh→en) —in two different domains: News and TED Talks. For each incorrect translation, human experts marked the corresponding error spans. In our framework, these error spans should align with the words that the attribution methods assign higher importance to. + +
METRICEXPLAINABILITY METHODen→dezh→enen→ruAvg.
AUCR@KAUCR@KAUCR@KAUCR@K
src-only* evaluation
UNITE SRCembed-align[mt, src]0.5870.3390.6440.2810.5830.1670.6040.262
grad ℓ20.5720.2930.5350.2000.6200.1690.5760.221
attention0.6360.3220.6120.2530.6120.1890.6200.254
attn × grad0.7070.3760.6390.2940.6330.2110.6600.294
ref-only evaluation
UNITE REFembed-align[mt, ref]0.6580.3960.6670.3280.6350.2180.6530.314
grad ℓ20.5960.3190.5710.2600.6610.2020.6090.261
attention0.6370.3440.6700.3350.6520.2240.6530.301
attn × grad0.7250.4250.6670.3800.6600.2480.6840.351
src, ref joint evaluation
UNITE SRC+REFembed-align[mt, src; ref]0.6500.3830.6700.3300.6180.2130.6460.309
grad ℓ20.5950.3250.5790.2570.6430.1910.6060.257
attention0.6570.4210.6700.3830.6490.2230.6590.342
attn × grad0.7360.4210.6740.3830.6710.2480.6930.351
COMETembed-align[mt, src]0.5900.3710.6740.3140.5770.2200.6140.301
embed-align[mt, ref]0.6940.4250.6960.3550.6470.2750.6790.352
embed-align[mt, src; ref]0.6880.4160.6970.3570.6220.2790.6690.350
grad ℓ20.6030.3120.5400.2520.6040.1850.5820.250
attention0.6040.3510.5920.2590.6330.2090.6080.268
attn × grad0.7100.3650.6330.2780.6620.2440.6690.295
+ +Table 1: AUC and Recall@K of explanations obtained via different attribution methods for COMET and UNITE models on the MQM data. *Although UNITE SRC is a src-only evaluation metric, it was trained with reference information (Wan et al., 2022). + +Models. For COMET, we use the latest publicly available model: wmt22-comet-da (Rei et al., 2022a). For UNITE, we train our own model using the same data used to train COMET in order to have a comparable setup. We provide full details (training data, correlations with human annotations, and hyperparameters) in Appendix A. Overall, the resulting reference-based UNITE models (REF and SRC+REF) are on par with COMET. + +Evaluation. We want our explanations to be directly attributed to the annotated error spans, in the style of an error detection task. Thus, we report Area Under Curve (AUC) and Recall@K.6 These metrics have been used as the main metrics in previous works on explainable QE (Fomicheva et al., 2021, 2022; Zerva et al., 2022). + +# 4 Results + +# 4.1 High-level analysis + +# Explanations are tightly related to the underlying metric architecture. The results in Ta- + +ble 1 show that the predictive power of the attribution methods differ between UNITE and COMET: attn $\times$ grad is the best method for UNITE-based models, while embed-align works best for COMET. This is expected as UNITE constructs a joint representation for the input sentences, thus allowing attention to flow across them; COMET, in contrast, encodes the sentences separately, so it relies heavily on the separate contextualized embeddings that are subsequently combined via elementwise operations such as multiplication and absolute difference. Interestingly, embed-align and attn $\times$ grad were the winning explainability approaches of the WMT 2022 Shared-Task on Quality Estimation (Zerva et al., 2022). This suggests that explainability methods developed for QE systems can translate well to reference-based metrics. We provide examples of explanations in Appendix C. + +Reference information boosts explainability power. Table 1 also shows that, across all metrics, using reference information brings substantial improvements over using only the source information. Moreover, while reference-based attributions significantly outperform source-based attributions, combining the source and reference information to + +![](images/0683cceb75a1ceb80cc27a61b4f761c4802d986859633f65fa8e3a4a5e01adbc.jpg) +Figure 2: Performance of the best attribution methods for COMET, UNITE REF and UNITE SRC+REF in terms of Recall@K on translations with critical errors: negations (NEG), hallucinations (HALL), named entity errors (NE), and errors in numbers (NUM). + +obtain token-level attributions does not consistently yield superior results over using the reference alone. Notably, the best attribution method for COMET does not require any source information. This is interesting: in some cases, reference-based metrics may largely ignore source information, relying heavily on the reference instead. + +# 4.2 How do the explanations fare for critical translation errors? + +The MQM data analyzed until now consists primarily of high quality translations, with the majority of annotated errors being non-critical. However, it is important to assess whether our explanations can be accurately attributed to critical errors, as this may reveal potential metric shortcomings. To this end, we employ SMAUG (Alves et al., 2022)8, a tool designed to generate synthetic data for stress-testing metrics, to create corrupted translations that contain critical errors. Concretely, we generate translations with the following pathologies: negation errors, hallucinations via insertions, named entity errors, and errors in numbers.9 + +Explanations identify critical errors more easily than non-critical errors. Figure 2 shows that explanations are more effective in identifying critical errors compared to other non-critical errors (see Table 1). Specifically, we find significant performance improvements up to nearly $30\%$ in Recall@K for certain critical errors. Overall, hallucinations are the easiest errors to identify across all neural metrics. This suggests that neural + +metrics appropriately identify and penalize hallucinated translations, which aligns with the findings of Guerreiro et al. (2022). Moreover, explanations for both UNITE models behave similarly for all errors except numbers, where the source information plays a key role in improving the explanations. Notably, contrary to what we observed for data with non-critical errors, COMET explanations are less effective than those of UNITE REF and UNITE SRC+REF for identifying critical errors. + +Explanations can reveal potential metric weaknesses. Figure 2 suggests that COMET explanations struggle to identify localized errors (negation errors, named entity errors and discrepancies in numbers). We hypothesize that this behavior is related to the underlying architecture. Unlike UNITE-based metrics, COMET does not rely on soft alignments via attention between the sentences in the encoding process. This process may be key to identify local misalignments during the encoding process. In fact, the attention-based attributions for UNITE metrics can more easily identify these errors. COMET, however, encodes the sentences separately, which may result in grammatical features (e.g. numbers) being encoded similarly across sentences (Chi et al., 2020; Chang et al., 2022). As such, explanations obtained via embedding alignments will not properly identify these misalignments on similar features. Importantly, these findings align with observations made in (Amrhein and Sennrich, 2022; Raunak et al., 2022). This showcases how explanations can be used to diagnose and reveal shortcomings of neural-based metrics. + +# 5 Conclusions and Future Work + +In this paper, we investigated the use of explainability methods to better understand widely used neural metrics for machine translation evaluation, such as COMET and UNITE. Concretely, we analyzed how explanations are impacted by the reference information, and how they can be used to reveal weaknesses of these metrics. Our analysis shows that the quality of the explanations is tightly related to the underlying metric architecture. Interestingly, we also provide evidence that neural metrics like COMET may heavily rely on reference information over source information. Additionally, we show that explanations can be used to reveal reference-based metrics weaknesses such as failing to severely penalize localized critical errors. This opens up promising opportunities for future + +research on leveraging explanations to diagnose reference-based metrics errors. To support these studies, we call for future datasets illustrating critical errors (e.g., challenge sets (Karpinska et al., 2022)) to be accompanied by annotated error spans. + +# Limitations + +We highlight three main limitations of our work. + +First, although we have explored gradient-based explanations that take the whole network into consideration and have been shown to be faithful in previous work (Bastings et al., 2021), we do not explicitly explore how COMET combines the sentence representations in the feed-forward that precedes the encoder model to produce the sentence-level score. + +Second, we have shown that combining attention with gradient information results in the best explanations for UNITE-based metrics. However, from a practical standpoint, running inference and extracting the explainability scores simultaneously may be more computationally expensive than other techniques: gradient-based metrics benefit from GPU infrastructure and require storing all gradient information. + +Third, we have not explored extracting explanations in low-resource settings. That is because high-quality MQM annotations for such language pairs are not yet available. Nevertheless, further research in those settings is needed to access the broader validity of our claims. + +# Acknowledgements + +This work was partially supported by the P2020 programs (MAIA, contract 045909), the Portuguese Recovery and Resilience Plan (PRR) through project C645008882-00000055, Center for Responsible AI, by the European Research Council (ERC StG DeepSPIN, 758969), by EU's Horizon Europe Research and Innovation Actions (UTTER, contract 101070631), and by the Fundação para a Ciência e Tecnologia (contracts UIDB/50021/2020 and UIDB/50008/2020). + +# References + +Duarte Alves, Ricardo Rei, Ana C Farinha, José G. C. de Souza, and André F. T. Martins. 2022. Robust MT Evaluation with Sentence-level Multilingual Augmentation. In Proceedings of the Seventh Conference on Machine Translation, pages 469-478, Abu Dhabi. Association for Computational Linguistics. + +Chantal Amrhein and Rico Sennrich. 2022. Identifying Weaknesses in Machine Translation Metrics Through Minimum Bayes Risk Decoding: A Case Study for COMET. In Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1125-1141, Online only. Association for Computational Linguistics. +Leila Arras, Ahmed Osman, Klaus-Robert Müller, and Wojciech Samek. 2019. Evaluating recurrent neural network explanations. In Proceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 113-126, Florence, Italy. Association for Computational Linguistics. +Pepa Atanasova, Jakob Grue Simonsen, Christina Lioma, and Isabelle Augenstein. 2020. A diagnostic study of explainability techniques for text classification. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3256-3274, Online. Association for Computational Linguistics. +Jasmijn Bastings, Sebastian Ebert, Polina Zablotskaia, Anders Sandholm, and Katja Filippova. 2021. "will you find these shortcuts?" a protocol for evaluating the faithfulness of input salience methods for text classification. +Chris Callison-Burch, Miles Osborne, and Philipp Koehn. 2006. Re-evaluating the role of Bleu in machine translation research. In 11th Conference of the European Chapter of the Association for Computational Linguistics, pages 249-256, Trento, Italy. Association for Computational Linguistics. +Tyler A. Chang, Zhuowen Tu, and Benjamin K. Bergen. 2022. The geometry of multilingual language model representations. +Ethan A. Chi, John Hewitt, and Christopher D. Manning. 2020. Finding universal grammatical relations in multilingual BERT. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5564-5577, Online. Association for Computational Linguistics. +Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettle-moyer, and Veselin Stoyanov. 2020. Unsupervised cross-lingual representation learning at scale. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8440-8451, Online. Association for Computational Linguistics. +Daniel Deutsch, Rotem Dror, and Dan Roth. 2021. A statistical analysis of summarization evaluation metrics using resampling methods. Transactions of the Association for Computational Linguistics, 9:1132-1146. + +Patrick Fernandes, Marcos Treviso, Danish Pruthi, André F. T. Martins, and Graham Neubig. 2022. Learning to scaffold: Optimizing model explanations for teaching. +Marina Fomicheva, Piyawat Lertvittayakumjorn, Wei Zhao, Steffen Eger, and Yang Gao. 2021. The Eval4NLP shared task on explainable quality estimation: Overview and results. In Proceedings of the 2nd Workshop on Evaluation and Comparison of NLP Systems, pages 165-178, Punta Cana, Dominican Republic. Association for Computational Linguistics. +Marina Fomicheva, Lucia Specia, and Nikolaos Aletras. 2022. Translation error detection as rationale extraction. In Findings of the Association for Computational Linguistics: ACL 2022, pages 4148-4159, Dublin, Ireland. Association for Computational Linguistics. +Markus Freitag, Ricardo Rei, Nitika Mathur, Chi-kiu Lo, Craig Stewart, Eleftherios Avramidis, Tom Kocmi, George Foster, Alon Lavie, and André F. T. Martins. 2022. Results of WMT22 Metrics Shared Task: Stop Using BLEU – Neural Metrics Are Better and More Robust. In Proceedings of the Seventh Conference on Machine Translation, pages 46–68, Abu Dhabi. Association for Computational Linguistics. +Markus Freitag, Ricardo Rei, Nitika Mathur, Chi-kiu Lo, Craig Stewart, George Foster, Alon Lavie, and Ondrej Bojar. 2021. Results of the WMT21 metrics shared task: Evaluating metrics with expert-based human evaluations on TED and news domain. In Proceedings of the Sixth Conference on Machine Translation, pages 733-774, Online. Association for Computational Linguistics. +Nuno M. Guerreiro, Elena Voita, and André F. T. Martins. 2022. Looking for a needle in a haystack: A comprehensive study of hallucinations in neural machine translation. +Sarthak Jain and Byron C. Wallace. 2019. Attention is not Explanation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 3543-3556, Minneapolis, Minnesota. Association for Computational Linguistics. +Marzena Karpinska, Nishant Raj, Katherine Thai, Yixiao Song, Ankita Gupta, and Mohit Iyyer. 2022. Demetr: Diagnosing evaluation metrics for translation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, page 9540-9561, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics. +Tom Kocmi, Christian Federmann, Roman Grundkiewicz, Marcin Junczys-Dowmunt, Hitokazu Matsushita, and Arul Menezes. 2021. To ship or not to ship: An extensive evaluation of automatic metrics for machine translation. In Proceedings of the Sixth + +Conference on Machine Translation, pages 478-494, Online. Association for Computational Linguistics. +Christoph Leiter, Piyawat Lertvittayakumjorn, Marina Fomicheva, Wei Zhao, Yang Gao, and Steffen Eger. 2022. Towards explainable evaluation metrics for natural language generation. +Arle Lommel, Hans Uszkoreit, and Aljoscha Burchardt. 2014. Multidimensional Quality Metrics (MQM): A Framework for Declaring and Describing Translation Quality Metrics. Tradumàtica, pages 0455-463. +Nitika Mathur, Timothy Baldwin, and Trevor Cohn. 2020. Tangled up in BLEU: Reevaluating the evaluation of automatic machine translation evaluation metrics. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4984-4997, Online. Association for Computational Linguistics. +Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311-318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics. +Maja Popovic. 2015. *chrF: character n-gram F-score* for automatic MT evaluation. In *Proceedings of the Tenth Workshop on Statistical Machine Translation*, pages 392–395, Lisbon, Portugal. Association for Computational Linguistics. +Tharindu Ranasinghe, Constantin Orasan, and Ruslan Mitkov. 2020. *TransQuest: Translation Quality Estimation with Cross-lingual Transformers*. In *Proceedings of the 28th International Conference on Computational Linguistics*, pages 5070–5081, Barcelona, Spain (Online). International Committee on Computational Linguistics. +Vikas Raunak, Matt Post, and Arul Menezes. 2022. Salted: A framework for salient long-tail translation error detection. +Ricardo Rei, José G. C. de Souza, Duarte Alves, Chrysoula Zerva, Ana C Farinha, Taisiya Glushkova, Alon Lavie, Luisa Coheur, and André F. T. Martins. 2022a. COMET-22: Unbabel-IST 2022 Submission for the Metrics Shared Task. In Proceedings of the Seventh Conference on Machine Translation, pages 578-585, Abu Dhabi. Association for Computational Linguistics. +Ricardo Rei, Craig Stewart, Ana C Farinha, and Alon Lavie. 2020. COMET: A neural framework for MT evaluation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2685-2702, Online. Association for Computational Linguistics. +Ricardo Rei, Marcos Treviso, Nuno M. Guerreiro, Chrysoula Zerva, Ana C Farinha, Christine Maroti, José G. C. de Souza, Taisiya Glushkova, Duarte + +Alves, Luisa Coheur, Alon Lavie, and Andre F. T. Martins. 2022b. CometKiwi: IST-Unbabel 2022 Submission for the Quality Estimation Shared Task. In Proceedings of the Seventh Conference on Machine Translation, pages 634-645, Abu Dhabi. Association for Computational Linguistics. +Marco Ribeiro, Sameer Singh, and Carlos Guestrin. 2016. "why should I trust you?": Explaining the predictions of any classifier. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Demonstrations, pages 97-101, San Diego, California. Association for Computational Linguistics. +Ananya B. Sai, Vignesh Nagarajan, Tanay Dixit, Raj Dabre, Anoop Kunchukuttan, Pratyush Kumar, and Mitesh M. Khapra. 2022. IndicMT Eval: A Dataset to Meta-Evaluate Machine Translation metrics for Indian Languages. +Thibault Sellam, Dipanjan Das, and Ankur Parikh. 2020. BLEURT: Learning robust metrics for text generation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7881-7892, Online. Association for Computational Linguistics. +Avanti Shrikumar, Peyton Greenside, and Anshul Kundaje. 2017. Learning Important Features Through Propagating Activation Differences. In Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 3145-3153. PMLR. +Aaron Smith, Christian Hardmeier, and Joerg Tiedemann. 2016. Climbing mont BLEU: The strange world of reachable high-BLEU translations. In Proceedings of the 19th Annual Conference of the European Association for Machine Translation, pages 269-281. +Mukund Sundararajan, Ankur Taly, and Qiqi Yan. 2017. Axiomatic Attribution for Deep Networks. In Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 3319-3328. PMLR. +Shimin Tao, Su Chang, Ma Miaomiao, Hao Yang, Xiang Geng, Shujian Huang, Min Zhang, Jiaxin Guo, Minghan Wang, and Yinglu Li. 2022. CrossQE: HW-TSC 2022 Submission for the Quality Estimation Shared Task. In Proceedings of the Seventh Conference on Machine Translation, pages 646-652, Abu Dhabi. Association for Computational Linguistics. +Marcos Treviso, Nuno M. Guerreiro, Ricardo Rei, and Andre F. T. Martins. 2021. IST-unbabel 2021 submission for the explainable quality estimation shared task. In Proceedings of the 2nd Workshop on Evaluation and Comparison of NLP Systems, pages 133-145, Punta Cana, Dominican Republic. Association for Computational Linguistics. + +Yu Wan, Dayiheng Liu, Baosong Yang, Tianchi Bi, Haibo Zhang, Boxing Chen, Weihua Luo, Derek F. Wong, and Lidia S. Chao. 2021. RoBLEURT submission for WMT2021 metrics task. In Proceedings of the Sixth Conference on Machine Translation, pages 1053-1058, Online. Association for Computational Linguistics. +Yu Wan, Dayiheng Liu, Baosong Yang, Haibo Zhang, Boxing Chen, Derek Wong, and Lidia Chao. 2022. UniTE: Unified translation evaluation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8117-8127, Dublin, Ireland. Association for Computational Linguistics. +Kerem Zaman and Yonatan Belinkov. 2022. A Multilingual Perspective Towards the Evaluation of Attribution Methods in Natural Language Inference. +Chrysoula Zerva, Frédéric Blain, Ricardo Rei, Piyawat Lertvittayakumjorn, José G. C. de Souza, Steffen Eger, Diptesh Kanojia, Duarte Alves, Constantin Orasan, Marina Fomicheva, André F. T. Martins, and Lucia Specia. 2022. Findings of the WMT 2022 Shared Task on Quality Estimation. In Proceedings of the Seventh Conference on Machine Translation, pages 69-99, Abu Dhabi. Association for Computational Linguistics. + +# A Model Details + +In Section 2.1, we employed the latest publicly available model (wmt22-comet-da) for COMET, which emerged as a top-performing metric in the WMT 2022 Metrics task (Freitag et al., 2022). To ensure a comparable setting for UNITE (Wan et al., 2022), we trained our own model. In doing so, we utilized the same data employed in the development of the COMET model by (Rei et al., 2022a), without pretraining any synthetic data, as originally suggested. Additionally, our implementation did not incorporate monotonic regional attention, as our preliminary experiments revealed no discernible benefits from its usage. The hyperparameters used are summarized in Table 3, while Table 4 presents the number of Direct Assessments utilized during training. Furthermore, Table 5 displays the segment-level correlations with WMT 2021 MQM data for the News and TED domains. + +Regarding computational infrastructure, a single NVIDIA A10G GPU with 23GB memory was used. The resulting UNITE model has 565M parameters while COMET has 581M parameters. + +# A.1 Output Distribution + +To better understand the output of the models and what scores are deemed low, we plotted the output distributions for the two models we used in our study. The average score for English $\rightarrow$ German data is 0.856 for the COMET model and 0.870 for the UNITE model we trained. From Figure 3 we can observe the distribution of scores. This means that the 0.6692 score from the example in Figure 1 corresponds to a low quality output (5th percentile). + +# A.2 SMAUG Corpus + +As we have seen in Section 4.2, we have created synthetic translation errors for the following pathologies: negation errors, hallucinations via insertions, named entity errors, and errors in numbers. Table 7 presents a summary of the examples created using SMAUG and in Table 8 we show examples of each error category. + +# B Comparison between COMET and XLM-R Alignments + +From Table 1, it is evident that the alignments between the reference and/or source and the translation yield effective explanations for COMET. This raises the question of how these alignments compare to the underlying encoder of COMET before + +the fine-tuning process with human annotations. To investigate this, we examine the results for XLM-R without any fine-tuning, as presented in Table 2. + +Overall, the explanations derived from the alignments of COMET prove to be more predictive of error spans than those obtained from XLM-R alignments. This suggests that during the fine-tuning phase, COMET models modify the underlying XLM-R representations to achieve better alignment with translation errors. + +# C Examples + +In Tables 9 and 10, we show examples of COMET explanations for Chinese $\rightarrow$ English and English $\rightarrow$ German language pairs, respectively. We highlight in gray the corresponding MQM annotation performed by an expert linguist and we sort the examples from highest to lowest COMET scores. From these examples we can observe the following: + +- Highlights provided by COMET explanations have a high recall with human annotations. In all examples, subword tokens corresponding to translation errors are highlighted in red but we often see that not everything is incorrect. + +- Explanations are consistent with scores. For example, in the third example from Table 10, the red highlights do not correspond to errors and in fact the translation only has a major error griffen. Nonetheless, the score assigned by COMET is a low score of 0.68 which is faithful to the explanations that was given even if the assessment does not agree with human experts. + +
METRICEXPLAINABILITY METHODen→dezh→enen→ruAvg.
AUCR@KAUCR@KAUCR@KAUCR@K
XLM-Rembed-align[mt, src]0.5870.3590.6680.3110.5760.1990.6100.289
embed-align[mt, ref]0.6710.4050.6890.3450.6340.2440.6640.331
embed-align[mt, src; ref]0.6660.3950.6900.3470.6160.2420.6570.328
COMETembed-align[mt, src]0.5900.3710.6740.3140.5770.2200.6140.301
embed-align[mt, ref]0.6940.4250.6960.3550.6470.2750.6790.352
embed-align[mt, src; ref]0.6880.4160.6970.3570.6220.2790.6690.350
+ +Table 2: AUC and Recall@K of explanations obtained via alignments for COMET and XLM-R without any further fine-tuning on human annotations. + +
HyperparameterUNITECOMET
Encoder ModelXLM-R (large)
OptimizerAdamW
No. frozen epochs0.3
Learning rate (LR)1.5e-05
Encoder LR.1.0e-06
Layerwise Decay0.95
Batch size16
Loss functionMSE
Dropout0.1
Hidden sizes[3072, 1024]
Embedding layerFrozen
FP precision16
No. Epochs12
+ +Table 3: Hyperparameters used to train UNITE and COMET checkpoints used in this work. The only difference between the two is the number of training epochs due to the fact that, for UNITE, the best validation checkpoint is the first one. + +
Language PairSIZE
zh-en126947
en-de121420
de-en99183
en-zh90805
ru-en79280
en-ru62749
en-CS60937
fi-en46145
en-fi34335
tr-en30186
et-en29496
cs-en27847
en-mr26000
de-CS13804
en-et13376
pl-en11816
en-pl10572
lt-en10315
en-ja9578
gu-en9063
si-en9000
ro-en9000
ne-en9000
en-lt8959
ja-en8939
en-kk8219
en-ta7890
ta-en7577
en-gu6924
kk-en6789
de-fr6691
en-lv5810
en-tr5171
km-en4722
ps-en4611
fr-de3999
Total1027155
+ +Table 4: Number of direct assessments per language pair used to train COMET (Rei et al., 2022a) and the UNITE model used in this work. + +![](images/81d67d2e1c777b6fda5691781782228f7808597ccd85601e5940c69164d79852.jpg) + +![](images/6537587ec22518537216de18a08f3630e956f491abbef7698b8fe26ab93de066.jpg) +(a)COMET + +![](images/7f12e5fb24bafeccede2948320d364ece3f8a22a64f142b40f6f3c81028c9be2.jpg) + +![](images/3e79cca46a0be2cd78a78a86a54cbab459146283d1cd328d1bf3942c7e0b1b2d.jpg) + +![](images/9894dcd3aa7a90714367ee71223d1e29b6dc2fe1264af58fa21b90c378d8648b.jpg) +(b) UNITE SRC + +![](images/ac54ae4e2d1b3fddfb076b8e1840a50715d1ccd9029eac80c23417c604b171c8.jpg) + +![](images/6d21005af26e50d9d8e404c809c12ef33821201eb759ff136e58f4816fcf1134.jpg) + +![](images/29fe38098bf71cc10cf0ff1d8da33f403b880c10f40eba4e14304dc71215a2a7.jpg) +(c) UNITE REF + +![](images/f70cd84ebf577212a3b3c178be04701601dd5f362429892d4029eade3c8b3ba8.jpg) + +![](images/e955daaa78864d91e6fdc0a5108bac795d99971ab97f7ec85f7b85c91433dcb8.jpg) +Figure 3: Distribution of scores for all metrics obtained on the MQM data (for all language pairs). + +![](images/9c020dee88cd623534f02d3180f8840f6d305d9394786d788b40554663a6624f.jpg) +(d) UNITE SRC+REF + +![](images/0921c070257ccf3a1203c6298b58fc5ee2d32d96b279faaac8e3a4c0561808c7.jpg) + +
BLEUCHRFYISI-1BLEURTUNITE +SRCUNITE +REFUNITE +SRC+REFCOMET +wmt22-comet-da
EN→DENEWSρ0.0770.0920.1630.3070.2740.3210.304
τ0.0690.0920.1440.2400.2220.2480.241
ρ0.1510.1580.2360.3250.3110.3350.338
τ0.1130.1460.2120.2830.2640.3010.298
EN→RUTED Newsρ0.1530.2520.2630.3590.3330.3910.382
τ0.1060.1780.2160.2760.2760.2980.297
ρ0.1540.2680.2350.2860.2390.2890.318
τ0.1120.1890.2040.2550.2320.2620.264
ZH→ENTED Newsρ0.2150.2310.3010.4280.4130.4380.426
τ0.1650.1880.2890.3410.3310.3580.352
ρ0.1550.1810.2870.2950.2440.3010.310
τ0.1130.1440.2160.2460.2240.2650.266
+ +Table 5: Segment-level correlations for WMT 2021 MQM annotations over News and TED domains (Freitag et al., 2021). The metrics are Pearson $(\rho)$ and Kendall Tau $(\tau)$ . Results in bold indicate which metrics are top-performing for that specific language pair, domain and metric according to Perm-Both hypothesis test (Deutsch et al., 2021), using 500 re-sampling runs, and setting $p = 0.05$ . + +
Error TypeNUM EXAMPLES
NE978
NEG669
HALL530
NUM432
Total2609
+ +Table 6: Number of examples for each category, synthetically-created using SMAUG (Alves et al., 2022). + +
Language PairTOKENS / SENT.ERRORS / SPANS
en-de528704 / 1531025712 / 3567
en-ru525938 / 1507417620 / 7172
zh-en603258 / 1650643984 / 10042
+ +Table 7: Statistics about MQM data from WMT 2021 Metrics task (Freitag et al., 2021) used in our experiments. + +# Source: + +格里沃里表示，分析人士对越南所提出的和平倡议给予认可。 + +# Translation: + +Grivory said that analysts recognize the peace initiative proposed by Vietnam. + +# Reference: + +Grigory said that analysts endorse the peace initiative proposed by Vietnam. + +# NE Error: + +Grivory said that analysts recognize the peace initiative proposed by Russia. + +# Source: + +英国的这一决定预计将会使西班牙的旅游业大受影响。 + +# Translation: + +This decision by the United Kingdom is expected to greatly affect Spain's tourism industry. + +# Reference: + +This decision by the UK is expected to have a significant impact on tourism in Spain. + +# NEG Error: + +This decision by the United Kingdom is expected to greatly benefit Spain's tourism industry. + +# Source: + +由于疫情，人们开始在互联网上花费更多的时间。” + +# Translation: + +"Because of the epidemic, people are starting to spend more time on the Internet." + +# Reference: + +For reason of the pandemic, people are starting to spend more time on the Internet. + +# HALL Error: + +Because we have a lot of friends around during the epidemic, people are starting to spend more time on the mobile devices than on the Internet." + +# Source: + +展销区将展至7月29日。 + +# Translation: + +The exhibition and sales area will be open until July 29. + +# Reference: + +The exhibition will last until July 29. + +# NUM Error: + +The exhibition and sales area will be open until July 2018 + +Table 8: Synthetically-generated critical errors (highlighted in gray) created with SMAUG (Alves et al., 2022) to assess whether our explanations can be accurately attributed to critical errors. + +# Source: + +And yet, the universe is not a silent movie because the universe isn't silent. + +# Translation: + +Und Dennoch ist das Universum kein Stummfilm, weil das Universum nicht still ist. + +# COMET score: 0.8595 + +# COMET explanation: + +_Und _dennoch _ist _das _Univers um _kein _Stu mm film , _weil _das _Univers um _nicht _still _ist . + +# Source: + +And yet black holes may be heard even if they're not seen, and that's because they bang on space-time like a drum. + +# Translation: + +Und dennoch werden Schwarze Locher weitereicht gehört, auch wenn sie nicht gesehen werden, und das liegt daran, dass sie wie eine Trommel auf die Raumzeit schlagen. + +# COMET score: 0.7150 + +# COMET explanation: + +Und dennoch werden Schwarz e LÖcher vielleicht gehört , auch wenn sie nicht gesehenwerden , und das liegt daran , dass sie wie eine Tro mmel auf die Raum zeit schlagen . + +# Source: + +Ash O'Brien and husband Jarett Kelley say they were grabbing a bite to eat at Dusty Rhodes dog park in San Diego on Thursday, with their three-month-old pug in tow. + +# Translation: + +Ash O'Brien und Ehemann Jarett Kelley sagen, dass sie am Donnerstag im Hundepark Dusty Rhodes in San Diego einen Happen zu essen griffen, mit ihrem drei Monate alten Mops im Schleptau. + +# COMET score: 0.6835 + +# COMET explanation: + +_Ash_O' Brien_Und_Ehe mann_Ja rett_Kelley_sagen, dass sie_am_Donnerstag_im_Hunde park_Dusty_Rhod es_in_San_Diego_einen_Happ_en_zu_essen_griff_en_, _mit_threm_drei_Monate_alten_M ops_im_Schleppt au. + +# Source: + +It was Einstein's great general theory of relativity. + +# Translation: + +Es war Einsteins große allgemeine Forschungen vor Relativitätstheorie. + +# COMET score: 0.6692 + +# COMET explanation: + +_Es _war _Einstein s _große _allgemeine e _Forschung en _vor _Relativ itäts the ori e . + +# Source: + +There's mask-shaming and then there's full on assault. + +# Translation: + +Es gibt Maskenschämen und dann ist es voll bei Angriff. + +# COMET score: 0.2318 + +# COMET explanation: + +_Es_gibt_Mask en schä men_und_dann_ist es_voll_bei_Angriff_. + +Table 9: Saliency map for COMET explanation scores for a set of en→de examples. Comparing the token-level explanations with the MQM annotation (highlighted in gray) reveals the source of correspondence between specific token-level translation errors and the resulting scores. + +# Source: + +我想告诉大家宇宙有着自己的配乐，而宇宙自身正在不停地播放着。因为太空可以想鼓一样振动。 + +# Translation: + +I want to tell you that the universe has its own iconic soundtrack and the universe itself is constantly playing non-stop because space can vibrate like a drum. + +# COMET score: 0.8634 + +# COMET explanation: + +_I_want_to_tell你想_the_univers_e_has_its_own(iconic soundtrack_and_the_univers_e_itself_is_constantly-playing_non-stop_because_space_can_vibrate_to_like_a Drum. + +# Source: + +另外，吉克隽逸和刘烨作为运动助理，也围绕运动少年制造了不少爆笑话题。 + +# Translation: + +In addition, as sports assistants, Ji Kejunyi and Liu Ye have also created a lot of hilarious topics around sports teenagers. + +# COMET score: 0.8214 + +# COMET explanation: + +_In _addition , _as _sports _assistant s , _Ji _Ke ju nyi _and _Li u _Ye _have _also _created _a _lot _of_ hila rious _topic s _around _sports _teenager s . + +# Source: + +一番言论让场上的少年和运动领队们都倒吸一口凉气。 + +# Translation: + +The remarks made the teenagers and the sports leaders on the field gasp a sigh of relief. + +# COMET score: 0.7793 + +# COMET explanation: + +The re marks made the teenager s and the sports leaders on the field gas p a sig h of _relief + +# Source: + +强烈的阳光是如此地刺眼， + +# Translation: + +The intense sunlight is so harsh; + +# COMET score: 0.7561 + +# COMET explanation: + +The intense sun light is so har sh + +# Source: + +如今，我们希望能够给这部关于宇宙的宏伟的视觉作品配上声音。 + +# Translation: + +Today, we hope to be able to give this magnificent visual work of the universe a sound. + +# COMET score: 0.7073 + +# COMET explanation: + +Today,we hope to be able to give this magnificent ent visual work of the univers e a sound. + +Table 10: Saliency map for COMET explanation scores for a set of $\mathrm{{zh}} \rightarrow \mathrm{{en}}$ examples. Comparing the token-level explanations with the MQM annotation (highlighted in gray) reveals the source of correspondence between specific token-level translation errors and the resulting scores. + +# A For every submission: + +A1. Did you describe the limitations of your work? + +Yes. Section 6 + +A2. Did you discuss any potential risks of your work? + +Not applicable. Left blank. + +A3. Do the abstract and introduction summarize the paper's main claims? + +Abstract and Section 1 + +A4. Have you used AI writing assistants when working on this paper? + +Assistance purely with the language of the paper along every section. Grammarly and DeepL write + +# B Did you use or create scientific artifacts? + +Section 3 explains the methods we used. We will release the adaptations required to use the explainability methods over COMET framework, the UniTE model we trained, and all data synthetically-generated data. + +B1. Did you cite the creators of artifacts you used? + +Section 2 + +B2. Did you discuss the license or terms for use and / or distribution of any artifacts? + +footnote on the first page. The License will be Apache 2.0 + +B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? + +Not applicable. Left blank. + +B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? + +Not applicable. Left blank. + +B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? + +in the Appendix we have several statistics for training and testing data. + +B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. + +Left blank. + +# C Did you run computational experiments? + +Appendix provides detailed information about the trained model. + +C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? + +Appendix provides detailed information about the trained model including GPU infrastructure and total number of parameters. + +C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? + +Not applicable. Left blank. + +C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? + +Appendix has all information needed about test data and performance of the models. + +C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? + +Section 2 and Appendix + +D Did you use human annotators (e.g., crowdworkers) or research with human participants? + +Left blank. + +D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? + +Not applicable. Left blank. + +D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? + +Not applicable. Left blank. + +D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? + +Not applicable. Left blank. + +D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? + +Not applicable. Left blank. + +D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? + +Not applicable. Left blank. \ No newline at end of file diff --git a/2023/The Inside Story_ Towards Better Understanding of Machine Translation Neural Evaluation Metrics/images.zip b/2023/The Inside Story_ Towards Better Understanding of Machine Translation Neural Evaluation Metrics/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..51258cbb13e50293a896ea085ee67a0ffbf31bc9 --- /dev/null +++ b/2023/The Inside Story_ Towards Better Understanding of Machine Translation Neural Evaluation Metrics/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7be76e1d89e1dffb3ea5f43a1ea9c93af4ab0ac78a0a69cf862974142dbd0fe9 +size 614725 diff --git a/2023/The Inside Story_ Towards Better Understanding of Machine Translation Neural Evaluation Metrics/layout.json b/2023/The Inside Story_ Towards Better Understanding of Machine Translation Neural Evaluation Metrics/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..407b065c01f2dc83ec6c830c20e34c71d1282f24 --- /dev/null +++ b/2023/The Inside Story_ Towards Better Understanding of Machine Translation Neural Evaluation Metrics/layout.json @@ -0,0 +1,11612 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 140, + 74, + 452, + 105 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 74, + 452, + 105 + ], + "spans": [ + { + "bbox": [ + 140, + 74, + 452, + 105 + ], + "type": "text", + "content": "The Inside Story: Towards Better Understanding of Machine Translation Neural Evaluation Metrics" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 119, + 452, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 119, + 452, + 147 + ], + "spans": [ + { + "bbox": [ + 146, + 119, + 452, + 147 + ], + "type": "text", + "content": "Ricardo Rei\\*1,2,4, Nuno M. Guerreiro\\*3,4, Marcos Treviso\\*3,4, Alon Lavie\\*1, Luisa Coheur\\*2,4, Andre F. T. Martins\\*1,3,4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 155, + 148, + 440, + 190 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 155, + 148, + 440, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 148, + 440, + 175 + ], + "spans": [ + { + "bbox": [ + 155, + 148, + 440, + 175 + ], + "type": "text", + "content": "1Unbabel, Lisbon, Portugal, 2INESC-ID, Lisbon, Portugal \n3Instituto de Telecomunicações, Lisbon, Portugal" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 155, + 176, + 440, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 176, + 440, + 190 + ], + "spans": [ + { + "bbox": [ + 155, + 176, + 440, + 190 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 155, + 176, + 440, + 190 + ], + "type": "text", + "content": "Instituto Superior Técnico, University of Lisbon, Portugal" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 83, + 234, + 274, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 234, + 274, + 498 + ], + "spans": [ + { + "bbox": [ + 83, + 234, + 274, + 498 + ], + "type": "text", + "content": "Neural metrics for machine translation evaluation, such as COMET, exhibit significant improvements in their correlation with human judgments compared to traditional metrics based on lexical overlap, such as BLEU. Yet neural metrics are, to a great extent, \"black boxes\" that return a single sentence-level score without transparency about the decision-making process. In this work, we develop and compare several neural explainability methods and demonstrate their effectiveness for interpreting state-of-the-art fine-tuned neural metrics. Our study reveals that these metrics leverage token-level information that can be directly attributed to translation errors, as assessed through comparison of token-level neural saliency maps with Multidimensional Quality Metrics (MQM) annotations and with synthetically-generated critical translation errors. To ease future research, we release our code at https://github.com/Unbabel/COMET/tree/explainable-metrics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 68, + 507, + 154, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 507, + 154, + 520 + ], + "spans": [ + { + "bbox": [ + 68, + 507, + 154, + 520 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 528, + 291, + 745 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 528, + 291, + 745 + ], + "spans": [ + { + "bbox": [ + 67, + 528, + 291, + 745 + ], + "type": "text", + "content": "Reference-based neural metrics for machine translation evaluation are achieving evergrowing success, demonstrating superior results over traditional lexical overlap-based metrics, such as BLEU (Papineni et al., 2002) and CHRF (Popovic, 2015), in terms of both their correlation with human ratings and their robustness across diverse domains (Callison-Burch et al., 2006; Smith et al., 2016; Mathur et al., 2020; Kocmi et al., 2021; Freitag et al., 2022). However, lexical overlap-based metrics remain popular for evaluating the performance and progress of translation systems and algorithms. Concerns regarding trust and interpretability may help explain this (Leiter et al., 2022): contrary to traditional metrics, neural metrics are considered \"black boxes\" as they often use" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 211, + 516, + 314 + ], + "blocks": [ + { + "bbox": [ + 307, + 211, + 516, + 314 + ], + "lines": [ + { + "bbox": [ + 307, + 211, + 516, + 314 + ], + "spans": [ + { + "bbox": [ + 307, + 211, + 516, + 314 + ], + "type": "image", + "image_path": "c81eeb83a2f01d61f4e4656faed8f88c417f549a3f6570b985690eea98bdeaf0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 326, + 525, + 386 + ], + "lines": [ + { + "bbox": [ + 302, + 326, + 525, + 386 + ], + "spans": [ + { + "bbox": [ + 302, + 326, + 525, + 386 + ], + "type": "text", + "content": "Figure 1: Illustration of our approach. In this example, the metric assigns the translation a low score. We aim to better understand this sentence-level assessment by examining the correspondence between our token-level explanations and human annotated error spans." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 407, + 524, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 407, + 524, + 448 + ], + "spans": [ + { + "bbox": [ + 302, + 407, + 524, + 448 + ], + "type": "text", + "content": "increasingly large models (e.g., the winning metric of the WMT 22 Metrics shared task was a 10B parameter model (Freitag et al., 2022))." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 449, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 449, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 449, + 526, + 772 + ], + "type": "text", + "content": "While some recent work has focused on explaining the predictions made by reference-free quality estimation (QE) systems (Fomicheva et al., 2021; Zerva et al., 2022), explaining reference-based metrics has remained a largely overlooked problem (Leiter et al., 2022). It is an open question whether the observations from studies of explainable QE carry over to this scenario. Thus, in this work, we fill that gap by turning to state-of-the-art reference-based metrics—we aim to interpret their decision-making process by exploiting the fact that these metrics show consistently good correlations with Multidimensional Quality Metrics (MQM) (Freitag et al., 2021, 2022; Sai et al., 2022), which are fine-grained quality assessments that result from experts identifying error spans in translation outputs (Lommel et al., 2014). We hypothesize that reference-based metrics leverage this token-level information to produce sentence-level scores. To test this hypothesis, we assess whether our explanations – measures of token-level importance obtained via attribution and input attribution methods such as attention weights and gradient scores (Treviso et al., 2021; Rei et al., 2022b) – align with" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 751, + 290, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 751, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 751, + 290, + 772 + ], + "type": "text", + "content": "* Equal contribution. Corresponding author: ricardo.rei@unbabel.com" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "1089" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "spans": [ + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "type": "text", + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 219, + 806, + 374, + 817 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 806, + 374, + 817 + ], + "spans": [ + { + "bbox": [ + 219, + 806, + 374, + 817 + ], + "type": "text", + "content": "Volume 2: Short Papers, pages 1089-1105" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "spans": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "text", + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 291, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 291, + 97 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 291, + 97 + ], + "type": "text", + "content": "human-annotated spans (Fomicheva et al., 2021, 2022; Zerva et al., 2022), as illustrated in Figure 1." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 99, + 291, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 99, + 291, + 179 + ], + "spans": [ + { + "bbox": [ + 67, + 99, + 291, + 179 + ], + "type": "text", + "content": "Our analysis focuses on two main vectors: (i) understanding the impact of the reference information on the quality of the explanations; and (ii) finding whether the explanations can help to identify potential weaknesses in the metrics. Our main contributions are:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 182, + 291, + 323 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 69, + 182, + 291, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 182, + 291, + 236 + ], + "spans": [ + { + "bbox": [ + 69, + 182, + 291, + 236 + ], + "type": "text", + "content": "- We provide a comparison between multiple explainability methods for different metrics on all types of evaluation: src-only, ref-only, and src+ref joint evaluation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 246, + 291, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 246, + 291, + 286 + ], + "spans": [ + { + "bbox": [ + 69, + 246, + 291, + 286 + ], + "type": "text", + "content": "- We find that explanations are related to the underlying metric architecture, and that leveraging reference information improves the explanations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 296, + 290, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 296, + 290, + 323 + ], + "spans": [ + { + "bbox": [ + 69, + 296, + 290, + 323 + ], + "type": "text", + "content": "- We show that explanations for critical translation errors can reveal weaknesses in the metrics." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 67, + 334, + 226, + 349 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 334, + 226, + 349 + ], + "spans": [ + { + "bbox": [ + 67, + 334, + 226, + 349 + ], + "type": "text", + "content": "2 Explaining Neural Metrics" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 356, + 291, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 356, + 291, + 423 + ], + "spans": [ + { + "bbox": [ + 67, + 356, + 291, + 423 + ], + "type": "text", + "content": "We aim to explain sentence-level quality assessments of reference-based metrics by producing token-level explanations that align to translation errors. In what follows, we describe the metrics and how we produce the explanations that we study." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 433, + 132, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 433, + 132, + 445 + ], + "spans": [ + { + "bbox": [ + 67, + 433, + 132, + 445 + ], + "type": "text", + "content": "2.1 Metrics" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 451, + 291, + 722 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 451, + 291, + 722 + ], + "spans": [ + { + "bbox": [ + 67, + 451, + 291, + 722 + ], + "type": "text", + "content": "We focus our analysis on two state-of-the-art neural metrics: COMET (Rei et al., 2020) and UNITE (Wan et al., 2022). While both metrics use a multilingual encoder model based on XLMR (Conneau et al., 2020), they employ distinct strategies to obtain sentence-level quality scores. On the one hand, COMET separately encodes the source, translation and reference to obtain their respective sentence embeddings; these embeddings are then combined to compute a quality score. On the other, UNITE jointly encodes the sentences to compute a contextualized representation that is subsequently used to compute the quality score. Interestingly, UNITE is trained to obtain quality scores for different input combinations: [mt; src] (SRC), [mt; ref] (REF), and [mt; src; ref] (SRC+REF). In fact, when the input is SRC, UNITE works like TransQuest (Ranasinghe et al., 2020); REF, like BLEURT (Sellam et al., 2020); and SRC+REF, like ROBLEURT (Wan et al., 2021)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 71, + 508, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 508, + 84 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 508, + 84 + ], + "type": "text", + "content": "2.2 Explanations via Attribution Methods" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 89, + 526, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 89, + 526, + 292 + ], + "spans": [ + { + "bbox": [ + 302, + 89, + 526, + 292 + ], + "type": "text", + "content": "In this work, we produce explanations using attribution methods that assign a scalar value to each translation token (i.e. a token-level attribution) to represent its importance. While many input attribution methods exist and have been extensively studied in the literature (Ribeiro et al., 2016; Shrikumar et al., 2017; Sundararajan et al., 2017; Jain and Wallace, 2019; Atanasova et al., 2020; Zaman and Belinkov, 2022), we focus specifically on those that have been demonstrated to be effective for explaining the predictions of QE models (Treviso et al., 2021; Fomicheva et al., 2022; Fernandes et al., 2022; Zerva et al., 2022) and extend them to our reference-based scenario. Concretely, we use the following techniques to extract explanations:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 301, + 526, + 505 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 304, + 301, + 526, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 301, + 526, + 356 + ], + "spans": [ + { + "bbox": [ + 304, + 301, + 526, + 356 + ], + "type": "text", + "content": "- embed-align: the maximum cosine similarity between each translation token embedding and the reference and/or source token embeddings (Tao et al., 2022);" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 365, + 525, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 525, + 405 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 525, + 405 + ], + "type": "text", + "content": "- grad " + }, + { + "bbox": [ + 304, + 365, + 525, + 405 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 304, + 365, + 525, + 405 + ], + "type": "text", + "content": ": the " + }, + { + "bbox": [ + 304, + 365, + 525, + 405 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 304, + 365, + 525, + 405 + ], + "type": "text", + "content": "-norm of gradients with respect to the word embeddings of the translation tokens (Arras et al., 2019);" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 416, + 525, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 416, + 525, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 416, + 525, + 456 + ], + "type": "text", + "content": "- attention: the attention weights of the translation tokens for each attention head of the encoder (Treviso et al., 2021);" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 465, + 525, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 465, + 525, + 505 + ], + "spans": [ + { + "bbox": [ + 304, + 465, + 525, + 505 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 304, + 465, + 525, + 505 + ], + "type": "inline_equation", + "content": "\\mathbf{attn} \\times \\mathbf{grad}" + }, + { + "bbox": [ + 304, + 465, + 525, + 505 + ], + "type": "text", + "content": ": the attention weights of each head scaled by the " + }, + { + "bbox": [ + 304, + 465, + 525, + 505 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 304, + 465, + 525, + 505 + ], + "type": "text", + "content": "-norm of the gradients of the value vectors of that head (Rei et al., 2022b)." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 302, + 526, + 433, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 526, + 433, + 539 + ], + "spans": [ + { + "bbox": [ + 302, + 526, + 433, + 539 + ], + "type": "text", + "content": "3 Experimental Setting" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 301, + 547, + 525, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 547, + 525, + 682 + ], + "spans": [ + { + "bbox": [ + 301, + 547, + 525, + 682 + ], + "type": "text", + "content": "MQM annotations. We use MQM annotations from the WMT 2021 Metrics shared task (Freitag et al., 2021),3 covering three language pairs — English-German (en→de), English-Russian (en→ru), and Chinese-English (zh→en) —in two different domains: News and TED Talks. For each incorrect translation, human experts marked the corresponding error spans. In our framework, these error spans should align with the words that the attribution methods assign higher importance to." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 690, + 526, + 750 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 690, + 526, + 750 + ], + "spans": [ + { + "bbox": [ + 302, + 690, + 526, + 750 + ], + "type": "text", + "content": "2For all attention-based methods, we ensemble the explanations from the top 5 heads as this has shown to improve performance consistently over selecting just the best head (Treviso et al., 2021; Rei et al., 2022b). Moreover, we use the full attention matrix, instead of relying only on cross attention information." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 303, + 751, + 448, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 751, + 448, + 772 + ], + "spans": [ + { + "bbox": [ + 303, + 751, + 448, + 772 + ], + "type": "text", + "content": "3https://github.com/google/wmt-mqm-human-evaluation" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 67, + 731, + 291, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 731, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 731, + 291, + 772 + ], + "type": "text", + "content": "Ensembles composed of these two metrics were respectively ranked second and third in WMT 2022 Metrics shared task. The winner of WMT 2022 Metrics task — METRICXXL — is not publicly available (Freitag et al., 2022)." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "text", + "content": "1090" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 108, + 68, + 487, + 324 + ], + "blocks": [ + { + "bbox": [ + 108, + 68, + 487, + 324 + ], + "lines": [ + { + "bbox": [ + 108, + 68, + 487, + 324 + ], + "spans": [ + { + "bbox": [ + 108, + 68, + 487, + 324 + ], + "type": "table", + "html": "
METRICEXPLAINABILITY METHODen→dezh→enen→ruAvg.
AUCR@KAUCR@KAUCR@KAUCR@K
src-only* evaluation
UNITE SRCembed-align[mt, src]0.5870.3390.6440.2810.5830.1670.6040.262
grad ℓ20.5720.2930.5350.2000.6200.1690.5760.221
attention0.6360.3220.6120.2530.6120.1890.6200.254
attn × grad0.7070.3760.6390.2940.6330.2110.6600.294
ref-only evaluation
UNITE REFembed-align[mt, ref]0.6580.3960.6670.3280.6350.2180.6530.314
grad ℓ20.5960.3190.5710.2600.6610.2020.6090.261
attention0.6370.3440.6700.3350.6520.2240.6530.301
attn × grad0.7250.4250.6670.3800.6600.2480.6840.351
src, ref joint evaluation
UNITE SRC+REFembed-align[mt, src; ref]0.6500.3830.6700.3300.6180.2130.6460.309
grad ℓ20.5950.3250.5790.2570.6430.1910.6060.257
attention0.6570.4210.6700.3830.6490.2230.6590.342
attn × grad0.7360.4210.6740.3830.6710.2480.6930.351
COMETembed-align[mt, src]0.5900.3710.6740.3140.5770.2200.6140.301
embed-align[mt, ref]0.6940.4250.6960.3550.6470.2750.6790.352
embed-align[mt, src; ref]0.6880.4160.6970.3570.6220.2790.6690.350
grad ℓ20.6030.3120.5400.2520.6040.1850.5820.250
attention0.6040.3510.5920.2590.6330.2090.6080.268
attn × grad0.7100.3650.6330.2780.6620.2440.6690.295
", + "image_path": "278a1cc6bcc5d3756ab7097a5c32d3bd20bb51325d4cb0c0e5041f2fe5e733c3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 332, + 525, + 370 + ], + "lines": [ + { + "bbox": [ + 67, + 332, + 525, + 370 + ], + "spans": [ + { + "bbox": [ + 67, + 332, + 525, + 370 + ], + "type": "text", + "content": "Table 1: AUC and Recall@K of explanations obtained via different attribution methods for COMET and UNITE models on the MQM data. *Although UNITE SRC is a src-only evaluation metric, it was trained with reference information (Wan et al., 2022)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 385, + 290, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 385, + 290, + 507 + ], + "spans": [ + { + "bbox": [ + 67, + 385, + 290, + 507 + ], + "type": "text", + "content": "Models. For COMET, we use the latest publicly available model: wmt22-comet-da (Rei et al., 2022a). For UNITE, we train our own model using the same data used to train COMET in order to have a comparable setup. We provide full details (training data, correlations with human annotations, and hyperparameters) in Appendix A. Overall, the resulting reference-based UNITE models (REF and SRC+REF) are on par with COMET." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 516, + 291, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 516, + 291, + 610 + ], + "spans": [ + { + "bbox": [ + 67, + 516, + 291, + 610 + ], + "type": "text", + "content": "Evaluation. We want our explanations to be directly attributed to the annotated error spans, in the style of an error detection task. Thus, we report Area Under Curve (AUC) and Recall@K.6 These metrics have been used as the main metrics in previous works on explainable QE (Fomicheva et al., 2021, 2022; Zerva et al., 2022)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 622, + 127, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 622, + 127, + 634 + ], + "spans": [ + { + "bbox": [ + 67, + 622, + 127, + 634 + ], + "type": "text", + "content": "4 Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 645, + 183, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 645, + 183, + 658 + ], + "spans": [ + { + "bbox": [ + 67, + 645, + 183, + 658 + ], + "type": "text", + "content": "4.1 High-level analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 664, + 291, + 690 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 664, + 291, + 690 + ], + "spans": [ + { + "bbox": [ + 67, + 664, + 291, + 690 + ], + "type": "text", + "content": "Explanations are tightly related to the underlying metric architecture. The results in Ta-" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 385, + 526, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 385, + 526, + 629 + ], + "spans": [ + { + "bbox": [ + 302, + 385, + 526, + 629 + ], + "type": "text", + "content": "ble 1 show that the predictive power of the attribution methods differ between UNITE and COMET: attn " + }, + { + "bbox": [ + 302, + 385, + 526, + 629 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 302, + 385, + 526, + 629 + ], + "type": "text", + "content": " grad is the best method for UNITE-based models, while embed-align works best for COMET. This is expected as UNITE constructs a joint representation for the input sentences, thus allowing attention to flow across them; COMET, in contrast, encodes the sentences separately, so it relies heavily on the separate contextualized embeddings that are subsequently combined via elementwise operations such as multiplication and absolute difference. Interestingly, embed-align and attn " + }, + { + "bbox": [ + 302, + 385, + 526, + 629 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 302, + 385, + 526, + 629 + ], + "type": "text", + "content": " grad were the winning explainability approaches of the WMT 2022 Shared-Task on Quality Estimation (Zerva et al., 2022). This suggests that explainability methods developed for QE systems can translate well to reference-based metrics. We provide examples of explanations in Appendix C." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 638, + 525, + 731 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 638, + 525, + 731 + ], + "spans": [ + { + "bbox": [ + 302, + 638, + 525, + 731 + ], + "type": "text", + "content": "Reference information boosts explainability power. Table 1 also shows that, across all metrics, using reference information brings substantial improvements over using only the source information. Moreover, while reference-based attributions significantly outperform source-based attributions, combining the source and reference information to" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 740, + 525, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 740, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 740, + 525, + 772 + ], + "type": "text", + "content": "In Appendix B, we provide a comparison between the explanations obtained via embed-align with COMET and with its pretrained encoder model, XLM-R." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 699, + 238, + 719 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 699, + 238, + 719 + ], + "spans": [ + { + "bbox": [ + 67, + 699, + 238, + 719 + ], + "type": "text", + "content": "4https://huggingface.co/Un babel/wmt22-comet-da" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 720, + 289, + 740 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 720, + 289, + 740 + ], + "spans": [ + { + "bbox": [ + 67, + 720, + 289, + 740 + ], + "type": "text", + "content": "5Our implementation differs from the original work by Wan et al. (2022), See Appendix A for full details." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 67, + 741, + 289, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 741, + 289, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 741, + 289, + 772 + ], + "type": "text", + "content": "In this setup, Recall@K is the proportion of words with the highest attribution that correspond to translation errors against the total number of errors in the annotated error span." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 287, + 781, + 308, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 308, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 308, + 791 + ], + "type": "text", + "content": "1091" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 68, + 290, + 164 + ], + "blocks": [ + { + "bbox": [ + 70, + 68, + 290, + 164 + ], + "lines": [ + { + "bbox": [ + 70, + 68, + 290, + 164 + ], + "spans": [ + { + "bbox": [ + 70, + 68, + 290, + 164 + ], + "type": "image", + "image_path": "0683cceb75a1ceb80cc27a61b4f761c4802d986859633f65fa8e3a4a5e01adbc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 172, + 291, + 232 + ], + "lines": [ + { + "bbox": [ + 67, + 172, + 291, + 232 + ], + "spans": [ + { + "bbox": [ + 67, + 172, + 291, + 232 + ], + "type": "text", + "content": "Figure 2: Performance of the best attribution methods for COMET, UNITE REF and UNITE SRC+REF in terms of Recall@K on translations with critical errors: negations (NEG), hallucinations (HALL), named entity errors (NE), and errors in numbers (NUM)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 249, + 291, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 249, + 291, + 343 + ], + "spans": [ + { + "bbox": [ + 67, + 249, + 291, + 343 + ], + "type": "text", + "content": "obtain token-level attributions does not consistently yield superior results over using the reference alone. Notably, the best attribution method for COMET does not require any source information. This is interesting: in some cases, reference-based metrics may largely ignore source information, relying heavily on the reference instead." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 354, + 285, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 354, + 285, + 380 + ], + "spans": [ + { + "bbox": [ + 67, + 354, + 285, + 380 + ], + "type": "text", + "content": "4.2 How do the explanations fare for critical translation errors?" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 386, + 291, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 386, + 291, + 561 + ], + "spans": [ + { + "bbox": [ + 67, + 386, + 291, + 561 + ], + "type": "text", + "content": "The MQM data analyzed until now consists primarily of high quality translations, with the majority of annotated errors being non-critical. However, it is important to assess whether our explanations can be accurately attributed to critical errors, as this may reveal potential metric shortcomings. To this end, we employ SMAUG (Alves et al., 2022)8, a tool designed to generate synthetic data for stress-testing metrics, to create corrupted translations that contain critical errors. Concretely, we generate translations with the following pathologies: negation errors, hallucinations via insertions, named entity errors, and errors in numbers.9" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 570, + 291, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 570, + 291, + 692 + ], + "spans": [ + { + "bbox": [ + 67, + 570, + 291, + 692 + ], + "type": "text", + "content": "Explanations identify critical errors more easily than non-critical errors. Figure 2 shows that explanations are more effective in identifying critical errors compared to other non-critical errors (see Table 1). Specifically, we find significant performance improvements up to nearly " + }, + { + "bbox": [ + 67, + 570, + 291, + 692 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 67, + 570, + 291, + 692 + ], + "type": "text", + "content": " in Recall@K for certain critical errors. Overall, hallucinations are the easiest errors to identify across all neural metrics. This suggests that neural" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 71, + 526, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 206 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 206 + ], + "type": "text", + "content": "metrics appropriately identify and penalize hallucinated translations, which aligns with the findings of Guerreiro et al. (2022). Moreover, explanations for both UNITE models behave similarly for all errors except numbers, where the source information plays a key role in improving the explanations. Notably, contrary to what we observed for data with non-critical errors, COMET explanations are less effective than those of UNITE REF and UNITE SRC+REF for identifying critical errors." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 214, + 526, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 214, + 526, + 525 + ], + "spans": [ + { + "bbox": [ + 302, + 214, + 526, + 525 + ], + "type": "text", + "content": "Explanations can reveal potential metric weaknesses. Figure 2 suggests that COMET explanations struggle to identify localized errors (negation errors, named entity errors and discrepancies in numbers). We hypothesize that this behavior is related to the underlying architecture. Unlike UNITE-based metrics, COMET does not rely on soft alignments via attention between the sentences in the encoding process. This process may be key to identify local misalignments during the encoding process. In fact, the attention-based attributions for UNITE metrics can more easily identify these errors. COMET, however, encodes the sentences separately, which may result in grammatical features (e.g. numbers) being encoded similarly across sentences (Chi et al., 2020; Chang et al., 2022). As such, explanations obtained via embedding alignments will not properly identify these misalignments on similar features. Importantly, these findings align with observations made in (Amrhein and Sennrich, 2022; Raunak et al., 2022). This showcases how explanations can be used to diagnose and reveal shortcomings of neural-based metrics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 535, + 478, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 535, + 478, + 548 + ], + "spans": [ + { + "bbox": [ + 302, + 535, + 478, + 548 + ], + "type": "text", + "content": "5 Conclusions and Future Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 556, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 556, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 556, + 526, + 772 + ], + "type": "text", + "content": "In this paper, we investigated the use of explainability methods to better understand widely used neural metrics for machine translation evaluation, such as COMET and UNITE. Concretely, we analyzed how explanations are impacted by the reference information, and how they can be used to reveal weaknesses of these metrics. Our analysis shows that the quality of the explanations is tightly related to the underlying metric architecture. Interestingly, we also provide evidence that neural metrics like COMET may heavily rely on reference information over source information. Additionally, we show that explanations can be used to reveal reference-based metrics weaknesses such as failing to severely penalize localized critical errors. This opens up promising opportunities for future" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 84, + 700, + 242, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 700, + 242, + 712 + ], + "spans": [ + { + "bbox": [ + 84, + 700, + 242, + 712 + ], + "type": "text", + "content": "8https://github.com/Unbabel/smaug" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 69, + 712, + 290, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 712, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 712, + 290, + 772 + ], + "type": "text", + "content": "9We corrupt fully correct translations that are not an exact copy of the reference translation. Moreover, as the full suit of SMAUG transformations can only be applied to English data, we focus solely on zh→en translations. Overall, the synthetic dataset consists of 2610 translations. Full statistics about the corrupted data and examples are shown in Appendix A.2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1092" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 292, + 141 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 292, + 141 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 292, + 141 + ], + "type": "text", + "content": "research on leveraging explanations to diagnose reference-based metrics errors. To support these studies, we call for future datasets illustrating critical errors (e.g., challenge sets (Karpinska et al., 2022)) to be accompanied by annotated error spans." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 149, + 131, + 162 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 149, + 131, + 162 + ], + "spans": [ + { + "bbox": [ + 68, + 149, + 131, + 162 + ], + "type": "text", + "content": "Limitations" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 171, + 283, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 171, + 283, + 184 + ], + "spans": [ + { + "bbox": [ + 67, + 171, + 283, + 184 + ], + "type": "text", + "content": "We highlight three main limitations of our work." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 185, + 291, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 185, + 291, + 290 + ], + "spans": [ + { + "bbox": [ + 67, + 185, + 291, + 290 + ], + "type": "text", + "content": "First, although we have explored gradient-based explanations that take the whole network into consideration and have been shown to be faithful in previous work (Bastings et al., 2021), we do not explicitly explore how COMET combines the sentence representations in the feed-forward that precedes the encoder model to produce the sentence-level score." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 293, + 291, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 293, + 291, + 413 + ], + "spans": [ + { + "bbox": [ + 67, + 293, + 291, + 413 + ], + "type": "text", + "content": "Second, we have shown that combining attention with gradient information results in the best explanations for UNITE-based metrics. However, from a practical standpoint, running inference and extracting the explainability scores simultaneously may be more computationally expensive than other techniques: gradient-based metrics benefit from GPU infrastructure and require storing all gradient information." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 415, + 291, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 415, + 291, + 496 + ], + "spans": [ + { + "bbox": [ + 67, + 415, + 291, + 496 + ], + "type": "text", + "content": "Third, we have not explored extracting explanations in low-resource settings. That is because high-quality MQM annotations for such language pairs are not yet available. Nevertheless, further research in those settings is needed to access the broader validity of our claims." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 507, + 170, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 507, + 170, + 520 + ], + "spans": [ + { + "bbox": [ + 68, + 507, + 170, + 520 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 528, + 291, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 528, + 291, + 663 + ], + "spans": [ + { + "bbox": [ + 67, + 528, + 291, + 663 + ], + "type": "text", + "content": "This work was partially supported by the P2020 programs (MAIA, contract 045909), the Portuguese Recovery and Resilience Plan (PRR) through project C645008882-00000055, Center for Responsible AI, by the European Research Council (ERC StG DeepSPIN, 758969), by EU's Horizon Europe Research and Innovation Actions (UTTER, contract 101070631), and by the Fundação para a Ciência e Tecnologia (contracts UIDB/50021/2020 and UIDB/50008/2020)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 68, + 687, + 127, + 699 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 687, + 127, + 699 + ], + "spans": [ + { + "bbox": [ + 68, + 687, + 127, + 699 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 705, + 291, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 705, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 705, + 291, + 772 + ], + "type": "text", + "content": "Duarte Alves, Ricardo Rei, Ana C Farinha, José G. C. de Souza, and André F. T. Martins. 2022. Robust MT Evaluation with Sentence-level Multilingual Augmentation. In Proceedings of the Seventh Conference on Machine Translation, pages 469-478, Abu Dhabi. Association for Computational Linguistics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 303, + 72, + 526, + 772 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 304, + 72, + 525, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 525, + 171 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 525, + 171 + ], + "type": "text", + "content": "Chantal Amrhein and Rico Sennrich. 2022. Identifying Weaknesses in Machine Translation Metrics Through Minimum Bayes Risk Decoding: A Case Study for COMET. In Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1125-1141, Online only. Association for Computational Linguistics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 303, + 179, + 526, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 179, + 526, + 258 + ], + "spans": [ + { + "bbox": [ + 303, + 179, + 526, + 258 + ], + "type": "text", + "content": "Leila Arras, Ahmed Osman, Klaus-Robert Müller, and Wojciech Samek. 2019. Evaluating recurrent neural network explanations. In Proceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 113-126, Florence, Italy. Association for Computational Linguistics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 303, + 266, + 525, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 266, + 525, + 344 + ], + "spans": [ + { + "bbox": [ + 303, + 266, + 525, + 344 + ], + "type": "text", + "content": "Pepa Atanasova, Jakob Grue Simonsen, Christina Lioma, and Isabelle Augenstein. 2020. A diagnostic study of explainability techniques for text classification. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3256-3274, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 303, + 352, + 525, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 352, + 525, + 407 + ], + "spans": [ + { + "bbox": [ + 303, + 352, + 525, + 407 + ], + "type": "text", + "content": "Jasmijn Bastings, Sebastian Ebert, Polina Zablotskaia, Anders Sandholm, and Katja Filippova. 2021. \"will you find these shortcuts?\" a protocol for evaluating the faithfulness of input salience methods for text classification." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 303, + 416, + 526, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 416, + 526, + 483 + ], + "spans": [ + { + "bbox": [ + 303, + 416, + 526, + 483 + ], + "type": "text", + "content": "Chris Callison-Burch, Miles Osborne, and Philipp Koehn. 2006. Re-evaluating the role of Bleu in machine translation research. In 11th Conference of the European Chapter of the Association for Computational Linguistics, pages 249-256, Trento, Italy. Association for Computational Linguistics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 303, + 491, + 525, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 491, + 525, + 525 + ], + "spans": [ + { + "bbox": [ + 303, + 491, + 525, + 525 + ], + "type": "text", + "content": "Tyler A. Chang, Zhuowen Tu, and Benjamin K. Bergen. 2022. The geometry of multilingual language model representations." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 303, + 533, + 525, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 533, + 525, + 600 + ], + "spans": [ + { + "bbox": [ + 303, + 533, + 525, + 600 + ], + "type": "text", + "content": "Ethan A. Chi, John Hewitt, and Christopher D. Manning. 2020. Finding universal grammatical relations in multilingual BERT. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5564-5577, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 303, + 608, + 526, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 608, + 526, + 708 + ], + "spans": [ + { + "bbox": [ + 303, + 608, + 526, + 708 + ], + "type": "text", + "content": "Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettle-moyer, and Veselin Stoyanov. 2020. Unsupervised cross-lingual representation learning at scale. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8440-8451, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 303, + 716, + 525, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 716, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 303, + 716, + 525, + 772 + ], + "type": "text", + "content": "Daniel Deutsch, Rotem Dror, and Dan Roth. 2021. A statistical analysis of summarization evaluation metrics using resampling methods. Transactions of the Association for Computational Linguistics, 9:1132-1146." + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1093" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 291, + 117 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 291, + 117 + ], + "type": "text", + "content": "Patrick Fernandes, Marcos Treviso, Danish Pruthi, André F. T. Martins, and Graham Neubig. 2022. Learning to scaffold: Optimizing model explanations for teaching." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 125, + 291, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 125, + 291, + 214 + ], + "spans": [ + { + "bbox": [ + 69, + 125, + 291, + 214 + ], + "type": "text", + "content": "Marina Fomicheva, Piyawat Lertvittayakumjorn, Wei Zhao, Steffen Eger, and Yang Gao. 2021. The Eval4NLP shared task on explainable quality estimation: Overview and results. In Proceedings of the 2nd Workshop on Evaluation and Comparison of NLP Systems, pages 165-178, Punta Cana, Dominican Republic. Association for Computational Linguistics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 222, + 291, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 222, + 291, + 288 + ], + "spans": [ + { + "bbox": [ + 69, + 222, + 291, + 288 + ], + "type": "text", + "content": "Marina Fomicheva, Lucia Specia, and Nikolaos Aletras. 2022. Translation error detection as rationale extraction. In Findings of the Association for Computational Linguistics: ACL 2022, pages 4148-4159, Dublin, Ireland. Association for Computational Linguistics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 297, + 291, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 297, + 291, + 386 + ], + "spans": [ + { + "bbox": [ + 69, + 297, + 291, + 386 + ], + "type": "text", + "content": "Markus Freitag, Ricardo Rei, Nitika Mathur, Chi-kiu Lo, Craig Stewart, Eleftherios Avramidis, Tom Kocmi, George Foster, Alon Lavie, and André F. T. Martins. 2022. Results of WMT22 Metrics Shared Task: Stop Using BLEU – Neural Metrics Are Better and More Robust. In Proceedings of the Seventh Conference on Machine Translation, pages 46–68, Abu Dhabi. Association for Computational Linguistics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 394, + 291, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 394, + 291, + 483 + ], + "spans": [ + { + "bbox": [ + 69, + 394, + 291, + 483 + ], + "type": "text", + "content": "Markus Freitag, Ricardo Rei, Nitika Mathur, Chi-kiu Lo, Craig Stewart, George Foster, Alon Lavie, and Ondrej Bojar. 2021. Results of the WMT21 metrics shared task: Evaluating metrics with expert-based human evaluations on TED and news domain. In Proceedings of the Sixth Conference on Machine Translation, pages 733-774, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 491, + 291, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 491, + 291, + 535 + ], + "spans": [ + { + "bbox": [ + 69, + 491, + 291, + 535 + ], + "type": "text", + "content": "Nuno M. Guerreiro, Elena Voita, and André F. T. Martins. 2022. Looking for a needle in a haystack: A comprehensive study of hallucinations in neural machine translation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 544, + 291, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 544, + 291, + 622 + ], + "spans": [ + { + "bbox": [ + 69, + 544, + 291, + 622 + ], + "type": "text", + "content": "Sarthak Jain and Byron C. Wallace. 2019. Attention is not Explanation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 3543-3556, Minneapolis, Minnesota. Association for Computational Linguistics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 630, + 291, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 630, + 291, + 708 + ], + "spans": [ + { + "bbox": [ + 69, + 630, + 291, + 708 + ], + "type": "text", + "content": "Marzena Karpinska, Nishant Raj, Katherine Thai, Yixiao Song, Ankita Gupta, and Mohit Iyyer. 2022. Demetr: Diagnosing evaluation metrics for translation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, page 9540-9561, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 717, + 291, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 717, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 717, + 291, + 772 + ], + "type": "text", + "content": "Tom Kocmi, Christian Federmann, Roman Grundkiewicz, Marcin Junczys-Dowmunt, Hitokazu Matsushita, and Arul Menezes. 2021. To ship or not to ship: An extensive evaluation of automatic metrics for machine translation. In Proceedings of the Sixth" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 525, + 772 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 315, + 72, + 525, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 525, + 95 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 525, + 95 + ], + "type": "text", + "content": "Conference on Machine Translation, pages 478-494, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 102, + 525, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 102, + 525, + 148 + ], + "spans": [ + { + "bbox": [ + 304, + 102, + 525, + 148 + ], + "type": "text", + "content": "Christoph Leiter, Piyawat Lertvittayakumjorn, Marina Fomicheva, Wei Zhao, Yang Gao, and Steffen Eger. 2022. Towards explainable evaluation metrics for natural language generation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 155, + 525, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 155, + 525, + 200 + ], + "spans": [ + { + "bbox": [ + 304, + 155, + 525, + 200 + ], + "type": "text", + "content": "Arle Lommel, Hans Uszkoreit, and Aljoscha Burchardt. 2014. Multidimensional Quality Metrics (MQM): A Framework for Declaring and Describing Translation Quality Metrics. Tradumàtica, pages 0455-463." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 207, + 525, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 207, + 525, + 285 + ], + "spans": [ + { + "bbox": [ + 304, + 207, + 525, + 285 + ], + "type": "text", + "content": "Nitika Mathur, Timothy Baldwin, and Trevor Cohn. 2020. Tangled up in BLEU: Reevaluating the evaluation of automatic machine translation evaluation metrics. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4984-4997, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 292, + 525, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 292, + 525, + 370 + ], + "spans": [ + { + "bbox": [ + 304, + 292, + 525, + 370 + ], + "type": "text", + "content": "Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311-318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 378, + 525, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 525, + 433 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 525, + 433 + ], + "type": "text", + "content": "Maja Popovic. 2015. *chrF: character n-gram F-score* for automatic MT evaluation. In *Proceedings of the Tenth Workshop on Statistical Machine Translation*, pages 392–395, Lisbon, Portugal. Association for Computational Linguistics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 441, + 525, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 525, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 525, + 518 + ], + "type": "text", + "content": "Tharindu Ranasinghe, Constantin Orasan, and Ruslan Mitkov. 2020. *TransQuest: Translation Quality Estimation with Cross-lingual Transformers*. In *Proceedings of the 28th International Conference on Computational Linguistics*, pages 5070–5081, Barcelona, Spain (Online). International Committee on Computational Linguistics." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 527, + 525, + 560 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 527, + 525, + 560 + ], + "spans": [ + { + "bbox": [ + 304, + 527, + 525, + 560 + ], + "type": "text", + "content": "Vikas Raunak, Matt Post, and Arul Menezes. 2022. Salted: A framework for salient long-tail translation error detection." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 568, + 525, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 525, + 656 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 525, + 656 + ], + "type": "text", + "content": "Ricardo Rei, José G. C. de Souza, Duarte Alves, Chrysoula Zerva, Ana C Farinha, Taisiya Glushkova, Alon Lavie, Luisa Coheur, and André F. T. Martins. 2022a. COMET-22: Unbabel-IST 2022 Submission for the Metrics Shared Task. In Proceedings of the Seventh Conference on Machine Translation, pages 578-585, Abu Dhabi. Association for Computational Linguistics." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 665, + 525, + 731 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 525, + 731 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 525, + 731 + ], + "type": "text", + "content": "Ricardo Rei, Craig Stewart, Ana C Farinha, and Alon Lavie. 2020. COMET: A neural framework for MT evaluation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2685-2702, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 739, + 525, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 739, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 739, + 525, + 772 + ], + "type": "text", + "content": "Ricardo Rei, Marcos Treviso, Nuno M. Guerreiro, Chrysoula Zerva, Ana C Farinha, Christine Maroti, José G. C. de Souza, Taisiya Glushkova, Duarte" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1094" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 289, + 772 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 80, + 72, + 289, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 72, + 289, + 138 + ], + "spans": [ + { + "bbox": [ + 80, + 72, + 289, + 138 + ], + "type": "text", + "content": "Alves, Luisa Coheur, Alon Lavie, and Andre F. T. Martins. 2022b. CometKiwi: IST-Unbabel 2022 Submission for the Quality Estimation Shared Task. In Proceedings of the Seventh Conference on Machine Translation, pages 634-645, Abu Dhabi. Association for Computational Linguistics." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 150, + 289, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 150, + 289, + 227 + ], + "spans": [ + { + "bbox": [ + 69, + 150, + 289, + 227 + ], + "type": "text", + "content": "Marco Ribeiro, Sameer Singh, and Carlos Guestrin. 2016. \"why should I trust you?\": Explaining the predictions of any classifier. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Demonstrations, pages 97-101, San Diego, California. Association for Computational Linguistics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 238, + 289, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 238, + 289, + 294 + ], + "spans": [ + { + "bbox": [ + 69, + 238, + 289, + 294 + ], + "type": "text", + "content": "Ananya B. Sai, Vignesh Nagarajan, Tanay Dixit, Raj Dabre, Anoop Kunchukuttan, Pratyush Kumar, and Mitesh M. Khapra. 2022. IndicMT Eval: A Dataset to Meta-Evaluate Machine Translation metrics for Indian Languages." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 306, + 289, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 306, + 289, + 371 + ], + "spans": [ + { + "bbox": [ + 69, + 306, + 289, + 371 + ], + "type": "text", + "content": "Thibault Sellam, Dipanjan Das, and Ankur Parikh. 2020. BLEURT: Learning robust metrics for text generation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7881-7892, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 383, + 289, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 383, + 289, + 449 + ], + "spans": [ + { + "bbox": [ + 69, + 383, + 289, + 449 + ], + "type": "text", + "content": "Avanti Shrikumar, Peyton Greenside, and Anshul Kundaje. 2017. Learning Important Features Through Propagating Activation Differences. In Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 3145-3153. PMLR." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 461, + 289, + 526 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 461, + 289, + 526 + ], + "spans": [ + { + "bbox": [ + 69, + 461, + 289, + 526 + ], + "type": "text", + "content": "Aaron Smith, Christian Hardmeier, and Joerg Tiedemann. 2016. Climbing mont BLEU: The strange world of reachable high-BLEU translations. In Proceedings of the 19th Annual Conference of the European Association for Machine Translation, pages 269-281." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 539, + 289, + 594 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 539, + 289, + 594 + ], + "spans": [ + { + "bbox": [ + 69, + 539, + 289, + 594 + ], + "type": "text", + "content": "Mukund Sundararajan, Ankur Taly, and Qiqi Yan. 2017. Axiomatic Attribution for Deep Networks. In Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 3319-3328. PMLR." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 606, + 289, + 682 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 606, + 289, + 682 + ], + "spans": [ + { + "bbox": [ + 69, + 606, + 289, + 682 + ], + "type": "text", + "content": "Shimin Tao, Su Chang, Ma Miaomiao, Hao Yang, Xiang Geng, Shujian Huang, Min Zhang, Jiaxin Guo, Minghan Wang, and Yinglu Li. 2022. CrossQE: HW-TSC 2022 Submission for the Quality Estimation Shared Task. In Proceedings of the Seventh Conference on Machine Translation, pages 646-652, Abu Dhabi. Association for Computational Linguistics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 694, + 289, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 694, + 289, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 694, + 289, + 772 + ], + "type": "text", + "content": "Marcos Treviso, Nuno M. Guerreiro, Ricardo Rei, and Andre F. T. Martins. 2021. IST-unbabel 2021 submission for the explainable quality estimation shared task. In Proceedings of the 2nd Workshop on Evaluation and Comparison of NLP Systems, pages 133-145, Punta Cana, Dominican Republic. Association for Computational Linguistics." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 524, + 384 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 305, + 72, + 524, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 524, + 149 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 524, + 149 + ], + "type": "text", + "content": "Yu Wan, Dayiheng Liu, Baosong Yang, Tianchi Bi, Haibo Zhang, Boxing Chen, Weihua Luo, Derek F. Wong, and Lidia S. Chao. 2021. RoBLEURT submission for WMT2021 metrics task. In Proceedings of the Sixth Conference on Machine Translation, pages 1053-1058, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 158, + 524, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 158, + 524, + 235 + ], + "spans": [ + { + "bbox": [ + 304, + 158, + 524, + 235 + ], + "type": "text", + "content": "Yu Wan, Dayiheng Liu, Baosong Yang, Haibo Zhang, Boxing Chen, Derek Wong, and Lidia Chao. 2022. UniTE: Unified translation evaluation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8117-8127, Dublin, Ireland. Association for Computational Linguistics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 243, + 524, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 524, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 524, + 277 + ], + "type": "text", + "content": "Kerem Zaman and Yonatan Belinkov. 2022. A Multilingual Perspective Towards the Evaluation of Attribution Methods in Natural Language Inference." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 285, + 524, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 285, + 524, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 524, + 384 + ], + "type": "text", + "content": "Chrysoula Zerva, Frédéric Blain, Ricardo Rei, Piyawat Lertvittayakumjorn, José G. C. de Souza, Steffen Eger, Diptesh Kanojia, Duarte Alves, Constantin Orasan, Marina Fomicheva, André F. T. Martins, and Lucia Specia. 2022. Findings of the WMT 2022 Shared Task on Quality Estimation. In Proceedings of the Seventh Conference on Machine Translation, pages 69-99, Abu Dhabi. Association for Computational Linguistics." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1095" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 71, + 163, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 71, + 163, + 83 + ], + "spans": [ + { + "bbox": [ + 68, + 71, + 163, + 83 + ], + "type": "text", + "content": "A Model Details" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 92, + 291, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 92, + 291, + 334 + ], + "spans": [ + { + "bbox": [ + 69, + 92, + 291, + 334 + ], + "type": "text", + "content": "In Section 2.1, we employed the latest publicly available model (wmt22-comet-da) for COMET, which emerged as a top-performing metric in the WMT 2022 Metrics task (Freitag et al., 2022). To ensure a comparable setting for UNITE (Wan et al., 2022), we trained our own model. In doing so, we utilized the same data employed in the development of the COMET model by (Rei et al., 2022a), without pretraining any synthetic data, as originally suggested. Additionally, our implementation did not incorporate monotonic regional attention, as our preliminary experiments revealed no discernible benefits from its usage. The hyperparameters used are summarized in Table 3, while Table 4 presents the number of Direct Assessments utilized during training. Furthermore, Table 5 displays the segment-level correlations with WMT 2021 MQM data for the News and TED domains." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 336, + 291, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 336, + 291, + 390 + ], + "spans": [ + { + "bbox": [ + 67, + 336, + 291, + 390 + ], + "type": "text", + "content": "Regarding computational infrastructure, a single NVIDIA A10G GPU with 23GB memory was used. The resulting UNITE model has 565M parameters while COMET has 581M parameters." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 400, + 191, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 400, + 191, + 412 + ], + "spans": [ + { + "bbox": [ + 67, + 400, + 191, + 412 + ], + "type": "text", + "content": "A.1 Output Distribution" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 417, + 291, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 417, + 291, + 539 + ], + "spans": [ + { + "bbox": [ + 67, + 417, + 291, + 539 + ], + "type": "text", + "content": "To better understand the output of the models and what scores are deemed low, we plotted the output distributions for the two models we used in our study. The average score for English " + }, + { + "bbox": [ + 67, + 417, + 291, + 539 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 67, + 417, + 291, + 539 + ], + "type": "text", + "content": " German data is 0.856 for the COMET model and 0.870 for the UNITE model we trained. From Figure 3 we can observe the distribution of scores. This means that the 0.6692 score from the example in Figure 1 corresponds to a low quality output (5th percentile)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 548, + 175, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 548, + 175, + 561 + ], + "spans": [ + { + "bbox": [ + 67, + 548, + 175, + 561 + ], + "type": "text", + "content": "A.2 SMAUG Corpus" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 565, + 291, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 565, + 291, + 660 + ], + "spans": [ + { + "bbox": [ + 67, + 565, + 291, + 660 + ], + "type": "text", + "content": "As we have seen in Section 4.2, we have created synthetic translation errors for the following pathologies: negation errors, hallucinations via insertions, named entity errors, and errors in numbers. Table 7 presents a summary of the examples created using SMAUG and in Table 8 we show examples of each error category." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 670, + 265, + 698 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 670, + 265, + 698 + ], + "spans": [ + { + "bbox": [ + 67, + 670, + 265, + 698 + ], + "type": "text", + "content": "B Comparison between COMET and XLM-R Alignments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 706, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 706, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 706, + 291, + 772 + ], + "type": "text", + "content": "From Table 1, it is evident that the alignments between the reference and/or source and the translation yield effective explanations for COMET. This raises the question of how these alignments compare to the underlying encoder of COMET before" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 71, + 524, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 524, + 111 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 524, + 111 + ], + "type": "text", + "content": "the fine-tuning process with human annotations. To investigate this, we examine the results for XLM-R without any fine-tuning, as presented in Table 2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 112, + 526, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 112, + 526, + 205 + ], + "spans": [ + { + "bbox": [ + 302, + 112, + 526, + 205 + ], + "type": "text", + "content": "Overall, the explanations derived from the alignments of COMET prove to be more predictive of error spans than those obtained from XLM-R alignments. This suggests that during the fine-tuning phase, COMET models modify the underlying XLM-R representations to achieve better alignment with translation errors." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 216, + 377, + 230 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 216, + 377, + 230 + ], + "spans": [ + { + "bbox": [ + 302, + 216, + 377, + 230 + ], + "type": "text", + "content": "C Examples" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 238, + 525, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 238, + 525, + 333 + ], + "spans": [ + { + "bbox": [ + 302, + 238, + 525, + 333 + ], + "type": "text", + "content": "In Tables 9 and 10, we show examples of COMET explanations for Chinese " + }, + { + "bbox": [ + 302, + 238, + 525, + 333 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 302, + 238, + 525, + 333 + ], + "type": "text", + "content": " English and English " + }, + { + "bbox": [ + 302, + 238, + 525, + 333 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 302, + 238, + 525, + 333 + ], + "type": "text", + "content": " German language pairs, respectively. We highlight in gray the corresponding MQM annotation performed by an expert linguist and we sort the examples from highest to lowest COMET scores. From these examples we can observe the following:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 342, + 524, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 342, + 524, + 408 + ], + "spans": [ + { + "bbox": [ + 302, + 342, + 524, + 408 + ], + "type": "text", + "content": "- Highlights provided by COMET explanations have a high recall with human annotations. In all examples, subword tokens corresponding to translation errors are highlighted in red but we often see that not everything is incorrect." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 418, + 525, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 418, + 525, + 526 + ], + "spans": [ + { + "bbox": [ + 302, + 418, + 525, + 526 + ], + "type": "text", + "content": "- Explanations are consistent with scores. For example, in the third example from Table 10, the red highlights do not correspond to errors and in fact the translation only has a major error griffen. Nonetheless, the score assigned by COMET is a low score of 0.68 which is faithful to the explanations that was given even if the assessment does not agree with human experts." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1096" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 110, + 68, + 483, + 168 + ], + "blocks": [ + { + "bbox": [ + 110, + 68, + 483, + 168 + ], + "lines": [ + { + "bbox": [ + 110, + 68, + 483, + 168 + ], + "spans": [ + { + "bbox": [ + 110, + 68, + 483, + 168 + ], + "type": "table", + "html": "
METRICEXPLAINABILITY METHODen→dezh→enen→ruAvg.
AUCR@KAUCR@KAUCR@KAUCR@K
XLM-Rembed-align[mt, src]0.5870.3590.6680.3110.5760.1990.6100.289
embed-align[mt, ref]0.6710.4050.6890.3450.6340.2440.6640.331
embed-align[mt, src; ref]0.6660.3950.6900.3470.6160.2420.6570.328
COMETembed-align[mt, src]0.5900.3710.6740.3140.5770.2200.6140.301
embed-align[mt, ref]0.6940.4250.6960.3550.6470.2750.6790.352
embed-align[mt, src; ref]0.6880.4160.6970.3570.6220.2790.6690.350
", + "image_path": "5deaf02c57aa10bc91a1fe7afb1edd9323f633af8a7892b568eaa348d146d7c0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 98, + 380, + 260, + 534 + ], + "blocks": [ + { + "bbox": [ + 67, + 175, + 525, + 199 + ], + "lines": [ + { + "bbox": [ + 67, + 175, + 525, + 199 + ], + "spans": [ + { + "bbox": [ + 67, + 175, + 525, + 199 + ], + "type": "text", + "content": "Table 2: AUC and Recall@K of explanations obtained via alignments for COMET and XLM-R without any further fine-tuning on human annotations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 98, + 380, + 260, + 534 + ], + "lines": [ + { + "bbox": [ + 98, + 380, + 260, + 534 + ], + "spans": [ + { + "bbox": [ + 98, + 380, + 260, + 534 + ], + "type": "table", + "html": "
HyperparameterUNITECOMET
Encoder ModelXLM-R (large)
OptimizerAdamW
No. frozen epochs0.3
Learning rate (LR)1.5e-05
Encoder LR.1.0e-06
Layerwise Decay0.95
Batch size16
Loss functionMSE
Dropout0.1
Hidden sizes[3072, 1024]
Embedding layerFrozen
FP precision16
No. Epochs12
", + "image_path": "0cbf1b21ea95c671b3079594768b57863e608636c78f31c584e9f913777664e5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 357, + 274, + 472, + 663 + ], + "blocks": [ + { + "bbox": [ + 67, + 543, + 291, + 602 + ], + "lines": [ + { + "bbox": [ + 67, + 543, + 291, + 602 + ], + "spans": [ + { + "bbox": [ + 67, + 543, + 291, + 602 + ], + "type": "text", + "content": "Table 3: Hyperparameters used to train UNITE and COMET checkpoints used in this work. The only difference between the two is the number of training epochs due to the fact that, for UNITE, the best validation checkpoint is the first one." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 357, + 274, + 472, + 663 + ], + "lines": [ + { + "bbox": [ + 357, + 274, + 472, + 663 + ], + "spans": [ + { + "bbox": [ + 357, + 274, + 472, + 663 + ], + "type": "table", + "html": "
Language PairSIZE
zh-en126947
en-de121420
de-en99183
en-zh90805
ru-en79280
en-ru62749
en-CS60937
fi-en46145
en-fi34335
tr-en30186
et-en29496
cs-en27847
en-mr26000
de-CS13804
en-et13376
pl-en11816
en-pl10572
lt-en10315
en-ja9578
gu-en9063
si-en9000
ro-en9000
ne-en9000
en-lt8959
ja-en8939
en-kk8219
en-ta7890
ta-en7577
en-gu6924
kk-en6789
de-fr6691
en-lv5810
en-tr5171
km-en4722
ps-en4611
fr-de3999
Total1027155
", + "image_path": "dec55668b4599bce00e1410c924ed299a4a47f11abb1e74e2ed7db35ceb50d18.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 302, + 672, + 524, + 708 + ], + "lines": [ + { + "bbox": [ + 302, + 672, + 524, + 708 + ], + "spans": [ + { + "bbox": [ + 302, + 672, + 524, + 708 + ], + "type": "text", + "content": "Table 4: Number of direct assessments per language pair used to train COMET (Rei et al., 2022a) and the UNITE model used in this work." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1097" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 133, + 221, + 248 + ], + "blocks": [ + { + "bbox": [ + 70, + 133, + 221, + 248 + ], + "lines": [ + { + "bbox": [ + 70, + 133, + 221, + 248 + ], + "spans": [ + { + "bbox": [ + 70, + 133, + 221, + 248 + ], + "type": "image", + "image_path": "81d67d2e1c777b6fda5691781782228f7808597ccd85601e5940c69164d79852.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 221, + 133, + 371, + 248 + ], + "blocks": [ + { + "bbox": [ + 221, + 133, + 371, + 248 + ], + "lines": [ + { + "bbox": [ + 221, + 133, + 371, + 248 + ], + "spans": [ + { + "bbox": [ + 221, + 133, + 371, + 248 + ], + "type": "image", + "image_path": "6537587ec22518537216de18a08f3630e956f491abbef7698b8fe26ab93de066.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 275, + 255, + 318, + 265 + ], + "lines": [ + { + "bbox": [ + 275, + 255, + 318, + 265 + ], + "spans": [ + { + "bbox": [ + 275, + 255, + 318, + 265 + ], + "type": "text", + "content": "(a)COMET" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 371, + 133, + 521, + 248 + ], + "blocks": [ + { + "bbox": [ + 371, + 133, + 521, + 248 + ], + "lines": [ + { + "bbox": [ + 371, + 133, + 521, + 248 + ], + "spans": [ + { + "bbox": [ + 371, + 133, + 521, + 248 + ], + "type": "image", + "image_path": "7f12e5fb24bafeccede2948320d364ece3f8a22a64f142b40f6f3c81028c9be2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 70, + 273, + 221, + 387 + ], + "blocks": [ + { + "bbox": [ + 70, + 273, + 221, + 387 + ], + "lines": [ + { + "bbox": [ + 70, + 273, + 221, + 387 + ], + "spans": [ + { + "bbox": [ + 70, + 273, + 221, + 387 + ], + "type": "image", + "image_path": "3e79cca46a0be2cd78a78a86a54cbab459146283d1cd328d1bf3942c7e0b1b2d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 221, + 273, + 370, + 387 + ], + "blocks": [ + { + "bbox": [ + 221, + 273, + 370, + 387 + ], + "lines": [ + { + "bbox": [ + 221, + 273, + 370, + 387 + ], + "spans": [ + { + "bbox": [ + 221, + 273, + 370, + 387 + ], + "type": "image", + "image_path": "9894dcd3aa7a90714367ee71223d1e29b6dc2fe1264af58fa21b90c378d8648b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 265, + 394, + 328, + 405 + ], + "lines": [ + { + "bbox": [ + 265, + 394, + 328, + 405 + ], + "spans": [ + { + "bbox": [ + 265, + 394, + 328, + 405 + ], + "type": "text", + "content": "(b) UNITE SRC" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 371, + 273, + 521, + 387 + ], + "blocks": [ + { + "bbox": [ + 371, + 273, + 521, + 387 + ], + "lines": [ + { + "bbox": [ + 371, + 273, + 521, + 387 + ], + "spans": [ + { + "bbox": [ + 371, + 273, + 521, + 387 + ], + "type": "image", + "image_path": "ac54ae4e2d1b3fddfb076b8e1840a50715d1ccd9029eac80c23417c604b171c8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 71, + 412, + 221, + 527 + ], + "blocks": [ + { + "bbox": [ + 71, + 412, + 221, + 527 + ], + "lines": [ + { + "bbox": [ + 71, + 412, + 221, + 527 + ], + "spans": [ + { + "bbox": [ + 71, + 412, + 221, + 527 + ], + "type": "image", + "image_path": "6d21005af26e50d9d8e404c809c12ef33821201eb759ff136e58f4816fcf1134.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 221, + 412, + 370, + 527 + ], + "blocks": [ + { + "bbox": [ + 221, + 412, + 370, + 527 + ], + "lines": [ + { + "bbox": [ + 221, + 412, + 370, + 527 + ], + "spans": [ + { + "bbox": [ + 221, + 412, + 370, + 527 + ], + "type": "image", + "image_path": "29fe38098bf71cc10cf0ff1d8da33f403b880c10f40eba4e14304dc71215a2a7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 267, + 534, + 327, + 544 + ], + "lines": [ + { + "bbox": [ + 267, + 534, + 327, + 544 + ], + "spans": [ + { + "bbox": [ + 267, + 534, + 327, + 544 + ], + "type": "text", + "content": "(c) UNITE REF" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 371, + 412, + 520, + 527 + ], + "blocks": [ + { + "bbox": [ + 371, + 412, + 520, + 527 + ], + "lines": [ + { + "bbox": [ + 371, + 412, + 520, + 527 + ], + "spans": [ + { + "bbox": [ + 371, + 412, + 520, + 527 + ], + "type": "image", + "image_path": "f70cd84ebf577212a3b3c178be04701601dd5f362429892d4029eade3c8b3ba8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 71, + 553, + 221, + 666 + ], + "blocks": [ + { + "bbox": [ + 71, + 553, + 221, + 666 + ], + "lines": [ + { + "bbox": [ + 71, + 553, + 221, + 666 + ], + "spans": [ + { + "bbox": [ + 71, + 553, + 221, + 666 + ], + "type": "image", + "image_path": "e955daaa78864d91e6fdc0a5108bac795d99971ab97f7ec85f7b85c91433dcb8.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 99, + 693, + 492, + 706 + ], + "lines": [ + { + "bbox": [ + 99, + 693, + 492, + 706 + ], + "spans": [ + { + "bbox": [ + 99, + 693, + 492, + 706 + ], + "type": "text", + "content": "Figure 3: Distribution of scores for all metrics obtained on the MQM data (for all language pairs)." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 221, + 553, + 370, + 666 + ], + "blocks": [ + { + "bbox": [ + 221, + 553, + 370, + 666 + ], + "lines": [ + { + "bbox": [ + 221, + 553, + 370, + 666 + ], + "spans": [ + { + "bbox": [ + 221, + 553, + 370, + 666 + ], + "type": "image", + "image_path": "9c020dee88cd623534f02d3180f8840f6d305d9394786d788b40554663a6624f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 256, + 673, + 337, + 684 + ], + "lines": [ + { + "bbox": [ + 256, + 673, + 337, + 684 + ], + "spans": [ + { + "bbox": [ + 256, + 673, + 337, + 684 + ], + "type": "text", + "content": "(d) UNITE SRC+REF" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 371, + 553, + 520, + 666 + ], + "blocks": [ + { + "bbox": [ + 371, + 553, + 520, + 666 + ], + "lines": [ + { + "bbox": [ + 371, + 553, + 520, + 666 + ], + "spans": [ + { + "bbox": [ + 371, + 553, + 520, + 666 + ], + "type": "image", + "image_path": "0921c070257ccf3a1203c6298b58fc5ee2d32d96b279faaac8e3a4c0561808c7.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1098" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 68, + 290, + 525, + 491 + ], + "blocks": [ + { + "bbox": [ + 68, + 290, + 525, + 491 + ], + "lines": [ + { + "bbox": [ + 68, + 290, + 525, + 491 + ], + "spans": [ + { + "bbox": [ + 68, + 290, + 525, + 491 + ], + "type": "table", + "html": "
BLEUCHRFYISI-1BLEURTUNITE\nSRCUNITE\nREFUNITE\nSRC+REFCOMET\nwmt22-comet-da
EN→DENEWSρ0.0770.0920.1630.3070.2740.3210.304
τ0.0690.0920.1440.2400.2220.2480.241
ρ0.1510.1580.2360.3250.3110.3350.338
τ0.1130.1460.2120.2830.2640.3010.298
EN→RUTED Newsρ0.1530.2520.2630.3590.3330.3910.382
τ0.1060.1780.2160.2760.2760.2980.297
ρ0.1540.2680.2350.2860.2390.2890.318
τ0.1120.1890.2040.2550.2320.2620.264
ZH→ENTED Newsρ0.2150.2310.3010.4280.4130.4380.426
τ0.1650.1880.2890.3410.3310.3580.352
ρ0.1550.1810.2870.2950.2440.3010.310
τ0.1130.1440.2160.2460.2240.2650.266
", + "image_path": "610f05d45f03472047468581cd4ea495d2eb518bbebdcea25940166bd75c6c34.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 499, + 525, + 547 + ], + "lines": [ + { + "bbox": [ + 67, + 499, + 525, + 547 + ], + "spans": [ + { + "bbox": [ + 67, + 499, + 525, + 547 + ], + "type": "text", + "content": "Table 5: Segment-level correlations for WMT 2021 MQM annotations over News and TED domains (Freitag et al., 2021). The metrics are Pearson " + }, + { + "bbox": [ + 67, + 499, + 525, + 547 + ], + "type": "inline_equation", + "content": "(\\rho)" + }, + { + "bbox": [ + 67, + 499, + 525, + 547 + ], + "type": "text", + "content": " and Kendall Tau " + }, + { + "bbox": [ + 67, + 499, + 525, + 547 + ], + "type": "inline_equation", + "content": "(\\tau)" + }, + { + "bbox": [ + 67, + 499, + 525, + 547 + ], + "type": "text", + "content": ". Results in bold indicate which metrics are top-performing for that specific language pair, domain and metric according to Perm-Both hypothesis test (Deutsch et al., 2021), using 500 re-sampling runs, and setting " + }, + { + "bbox": [ + 67, + 499, + 525, + 547 + ], + "type": "inline_equation", + "content": "p = 0.05" + }, + { + "bbox": [ + 67, + 499, + 525, + 547 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1099" + } + ] + } + ], + "index": 2 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 112, + 186, + 247, + 266 + ], + "blocks": [ + { + "bbox": [ + 112, + 186, + 247, + 266 + ], + "lines": [ + { + "bbox": [ + 112, + 186, + 247, + 266 + ], + "spans": [ + { + "bbox": [ + 112, + 186, + 247, + 266 + ], + "type": "table", + "html": "
Error TypeNUM EXAMPLES
NE978
NEG669
HALL530
NUM432
Total2609
", + "image_path": "0fc3685d75d2fcdaafc985b00e846cc579438685b7215213252d450128e1cb67.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 69, + 553, + 290, + 608 + ], + "blocks": [ + { + "bbox": [ + 67, + 275, + 291, + 311 + ], + "lines": [ + { + "bbox": [ + 67, + 275, + 291, + 311 + ], + "spans": [ + { + "bbox": [ + 67, + 275, + 291, + 311 + ], + "type": "text", + "content": "Table 6: Number of examples for each category, synthetically-created using SMAUG (Alves et al., 2022)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 553, + 290, + 608 + ], + "lines": [ + { + "bbox": [ + 69, + 553, + 290, + 608 + ], + "spans": [ + { + "bbox": [ + 69, + 553, + 290, + 608 + ], + "type": "table", + "html": "
Language PairTOKENS / SENT.ERRORS / SPANS
en-de528704 / 1531025712 / 3567
en-ru525938 / 1507417620 / 7172
zh-en603258 / 1650643984 / 10042
", + "image_path": "414856d8bbbd30168d04dcf936c02d7ab892e2e69e5f9a6dd70f04f559d22a28.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 615, + 291, + 651 + ], + "lines": [ + { + "bbox": [ + 67, + 615, + 291, + 651 + ], + "spans": [ + { + "bbox": [ + 67, + 615, + 291, + 651 + ], + "type": "text", + "content": "Table 7: Statistics about MQM data from WMT 2021 Metrics task (Freitag et al., 2021) used in our experiments." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "text", + "content": "1100" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 75, + 200, + 105, + 210 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 200, + 105, + 210 + ], + "spans": [ + { + "bbox": [ + 75, + 200, + 105, + 210 + ], + "type": "text", + "content": "Source:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 75, + 211, + 314, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 211, + 314, + 221 + ], + "spans": [ + { + "bbox": [ + 75, + 211, + 314, + 221 + ], + "type": "text", + "content": "格里沃里表示，分析人士对越南所提出的和平倡议给予认可。" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 75, + 225, + 123, + 233 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 225, + 123, + 233 + ], + "spans": [ + { + "bbox": [ + 75, + 225, + 123, + 233 + ], + "type": "text", + "content": "Translation:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 75, + 234, + 355, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 234, + 355, + 244 + ], + "spans": [ + { + "bbox": [ + 75, + 234, + 355, + 244 + ], + "type": "text", + "content": "Grivory said that analysts recognize the peace initiative proposed by Vietnam." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 75, + 246, + 117, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 246, + 117, + 255 + ], + "spans": [ + { + "bbox": [ + 75, + 246, + 117, + 255 + ], + "type": "text", + "content": "Reference:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 75, + 256, + 349, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 256, + 349, + 267 + ], + "spans": [ + { + "bbox": [ + 75, + 256, + 349, + 267 + ], + "type": "text", + "content": "Grigory said that analysts endorse the peace initiative proposed by Vietnam." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 75, + 269, + 117, + 278 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 269, + 117, + 278 + ], + "spans": [ + { + "bbox": [ + 75, + 269, + 117, + 278 + ], + "type": "text", + "content": "NE Error:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 75, + 279, + 352, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 279, + 352, + 290 + ], + "spans": [ + { + "bbox": [ + 75, + 279, + 352, + 290 + ], + "type": "text", + "content": "Grivory said that analysts recognize the peace initiative proposed by Russia." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 75, + 302, + 105, + 311 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 302, + 105, + 311 + ], + "spans": [ + { + "bbox": [ + 75, + 302, + 105, + 311 + ], + "type": "text", + "content": "Source:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 75, + 311, + 287, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 311, + 287, + 322 + ], + "spans": [ + { + "bbox": [ + 75, + 311, + 287, + 322 + ], + "type": "text", + "content": "英国的这一决定预计将会使西班牙的旅游业大受影响。" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 75, + 325, + 123, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 325, + 123, + 333 + ], + "spans": [ + { + "bbox": [ + 75, + 325, + 123, + 333 + ], + "type": "text", + "content": "Translation:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 75, + 334, + 404, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 334, + 404, + 345 + ], + "spans": [ + { + "bbox": [ + 75, + 334, + 404, + 345 + ], + "type": "text", + "content": "This decision by the United Kingdom is expected to greatly affect Spain's tourism industry." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 75, + 347, + 117, + 356 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 347, + 117, + 356 + ], + "spans": [ + { + "bbox": [ + 75, + 347, + 117, + 356 + ], + "type": "text", + "content": "Reference:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 75, + 357, + 381, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 357, + 381, + 368 + ], + "spans": [ + { + "bbox": [ + 75, + 357, + 381, + 368 + ], + "type": "text", + "content": "This decision by the UK is expected to have a significant impact on tourism in Spain." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 75, + 370, + 121, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 370, + 121, + 379 + ], + "spans": [ + { + "bbox": [ + 75, + 370, + 121, + 379 + ], + "type": "text", + "content": "NEG Error:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 75, + 380, + 411, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 380, + 411, + 391 + ], + "spans": [ + { + "bbox": [ + 75, + 380, + 411, + 391 + ], + "type": "text", + "content": "This decision by the United Kingdom is expected to greatly benefit Spain's tourism industry." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 75, + 402, + 105, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 402, + 105, + 412 + ], + "spans": [ + { + "bbox": [ + 75, + 402, + 105, + 412 + ], + "type": "text", + "content": "Source:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 75, + 412, + 278, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 412, + 278, + 423 + ], + "spans": [ + { + "bbox": [ + 75, + 412, + 278, + 423 + ], + "type": "text", + "content": "由于疫情，人们开始在互联网上花费更多的时间。”" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 75, + 426, + 123, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 426, + 123, + 435 + ], + "spans": [ + { + "bbox": [ + 75, + 426, + 123, + 435 + ], + "type": "text", + "content": "Translation:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 75, + 436, + 367, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 436, + 367, + 446 + ], + "spans": [ + { + "bbox": [ + 75, + 436, + 367, + 446 + ], + "type": "text", + "content": "\"Because of the epidemic, people are starting to spend more time on the Internet.\"" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 75, + 449, + 117, + 457 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 449, + 117, + 457 + ], + "spans": [ + { + "bbox": [ + 75, + 449, + 117, + 457 + ], + "type": "text", + "content": "Reference:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 75, + 458, + 380, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 458, + 380, + 469 + ], + "spans": [ + { + "bbox": [ + 75, + 458, + 380, + 469 + ], + "type": "text", + "content": "For reason of the pandemic, people are starting to spend more time on the Internet." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 75, + 471, + 127, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 471, + 127, + 480 + ], + "spans": [ + { + "bbox": [ + 75, + 471, + 127, + 480 + ], + "type": "text", + "content": "HALL Error:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 75, + 481, + 509, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 481, + 509, + 503 + ], + "spans": [ + { + "bbox": [ + 75, + 481, + 509, + 503 + ], + "type": "text", + "content": "Because we have a lot of friends around during the epidemic, people are starting to spend more time on the mobile devices than on the Internet.\"" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 75, + 516, + 105, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 516, + 105, + 525 + ], + "spans": [ + { + "bbox": [ + 75, + 516, + 105, + 525 + ], + "type": "text", + "content": "Source:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 75, + 525, + 166, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 525, + 166, + 536 + ], + "spans": [ + { + "bbox": [ + 75, + 525, + 166, + 536 + ], + "type": "text", + "content": "展销区将展至7月29日。" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 75, + 539, + 123, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 539, + 123, + 548 + ], + "spans": [ + { + "bbox": [ + 75, + 539, + 123, + 548 + ], + "type": "text", + "content": "Translation:" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 75, + 549, + 276, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 549, + 276, + 560 + ], + "spans": [ + { + "bbox": [ + 75, + 549, + 276, + 560 + ], + "type": "text", + "content": "The exhibition and sales area will be open until July 29." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 75, + 561, + 117, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 561, + 117, + 571 + ], + "spans": [ + { + "bbox": [ + 75, + 561, + 117, + 571 + ], + "type": "text", + "content": "Reference:" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 75, + 571, + 208, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 571, + 208, + 581 + ], + "spans": [ + { + "bbox": [ + 75, + 571, + 208, + 581 + ], + "type": "text", + "content": "The exhibition will last until July 29." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 75, + 584, + 123, + 593 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 584, + 123, + 593 + ], + "spans": [ + { + "bbox": [ + 75, + 584, + 123, + 593 + ], + "type": "text", + "content": "NUM Error:" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 75, + 594, + 286, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 594, + 286, + 605 + ], + "spans": [ + { + "bbox": [ + 75, + 594, + 286, + 605 + ], + "type": "text", + "content": "The exhibition and sales area will be open until July 2018" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 68, + 618, + 524, + 641 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 618, + 524, + 641 + ], + "spans": [ + { + "bbox": [ + 68, + 618, + 524, + 641 + ], + "type": "text", + "content": "Table 8: Synthetically-generated critical errors (highlighted in gray) created with SMAUG (Alves et al., 2022) to assess whether our explanations can be accurately attributed to critical errors." + } + ] + } + ], + "index": 32, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 308, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 308, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 308, + 791 + ], + "type": "text", + "content": "1101" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 75, + 153, + 105, + 162 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 153, + 105, + 162 + ], + "spans": [ + { + "bbox": [ + 75, + 153, + 105, + 162 + ], + "type": "text", + "content": "Source:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 75, + 163, + 345, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 163, + 345, + 173 + ], + "spans": [ + { + "bbox": [ + 75, + 163, + 345, + 173 + ], + "type": "text", + "content": "And yet, the universe is not a silent movie because the universe isn't silent." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 75, + 175, + 123, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 175, + 123, + 185 + ], + "spans": [ + { + "bbox": [ + 75, + 175, + 123, + 185 + ], + "type": "text", + "content": "Translation:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 75, + 185, + 379, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 185, + 379, + 195 + ], + "spans": [ + { + "bbox": [ + 75, + 185, + 379, + 195 + ], + "type": "text", + "content": "Und Dennoch ist das Universum kein Stummfilm, weil das Universum nicht still ist." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 75, + 198, + 161, + 207 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 198, + 161, + 207 + ], + "spans": [ + { + "bbox": [ + 75, + 198, + 161, + 207 + ], + "type": "text", + "content": "COMET score: 0.8595" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 75, + 208, + 157, + 218 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 208, + 157, + 218 + ], + "spans": [ + { + "bbox": [ + 75, + 208, + 157, + 218 + ], + "type": "text", + "content": "COMET explanation:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 77, + 219, + 503, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 219, + 503, + 229 + ], + "spans": [ + { + "bbox": [ + 77, + 219, + 503, + 229 + ], + "type": "text", + "content": "_Und _dennoch _ist _das _Univers um _kein _Stu mm film , _weil _das _Univers um _nicht _still _ist ." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 75, + 239, + 105, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 239, + 105, + 248 + ], + "spans": [ + { + "bbox": [ + 75, + 239, + 105, + 248 + ], + "type": "text", + "content": "Source:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 75, + 248, + 488, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 248, + 488, + 259 + ], + "spans": [ + { + "bbox": [ + 75, + 248, + 488, + 259 + ], + "type": "text", + "content": "And yet black holes may be heard even if they're not seen, and that's because they bang on space-time like a drum." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 75, + 261, + 123, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 261, + 123, + 270 + ], + "spans": [ + { + "bbox": [ + 75, + 261, + 123, + 270 + ], + "type": "text", + "content": "Translation:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 75, + 271, + 508, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 271, + 508, + 291 + ], + "spans": [ + { + "bbox": [ + 75, + 271, + 508, + 291 + ], + "type": "text", + "content": "Und dennoch werden Schwarze Locher weitereicht gehört, auch wenn sie nicht gesehen werden, und das liegt daran, dass sie wie eine Trommel auf die Raumzeit schlagen." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 75, + 294, + 160, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 294, + 160, + 304 + ], + "spans": [ + { + "bbox": [ + 75, + 294, + 160, + 304 + ], + "type": "text", + "content": "COMET score: 0.7150" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 75, + 305, + 157, + 314 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 305, + 157, + 314 + ], + "spans": [ + { + "bbox": [ + 75, + 305, + 157, + 314 + ], + "type": "text", + "content": "COMET explanation:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 75, + 315, + 508, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 315, + 508, + 338 + ], + "spans": [ + { + "bbox": [ + 75, + 315, + 508, + 338 + ], + "type": "text", + "content": "Und dennoch werden Schwarz e LÖcher vielleicht gehört , auch wenn sie nicht gesehenwerden , und das liegt daran , dass sie wie eine Tro mmel auf die Raum zeit schlagen ." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 75, + 348, + 105, + 356 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 348, + 105, + 356 + ], + "spans": [ + { + "bbox": [ + 75, + 348, + 105, + 356 + ], + "type": "text", + "content": "Source:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 75, + 357, + 508, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 357, + 508, + 377 + ], + "spans": [ + { + "bbox": [ + 75, + 357, + 508, + 377 + ], + "type": "text", + "content": "Ash O'Brien and husband Jarett Kelley say they were grabbing a bite to eat at Dusty Rhodes dog park in San Diego on Thursday, with their three-month-old pug in tow." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 75, + 380, + 123, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 380, + 123, + 388 + ], + "spans": [ + { + "bbox": [ + 75, + 380, + 123, + 388 + ], + "type": "text", + "content": "Translation:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 75, + 390, + 508, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 390, + 508, + 410 + ], + "spans": [ + { + "bbox": [ + 75, + 390, + 508, + 410 + ], + "type": "text", + "content": "Ash O'Brien und Ehemann Jarett Kelley sagen, dass sie am Donnerstag im Hundepark Dusty Rhodes in San Diego einen Happen zu essen griffen, mit ihrem drei Monate alten Mops im Schleptau." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 75, + 412, + 160, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 412, + 160, + 422 + ], + "spans": [ + { + "bbox": [ + 75, + 412, + 160, + 422 + ], + "type": "text", + "content": "COMET score: 0.6835" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 75, + 423, + 157, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 423, + 157, + 432 + ], + "spans": [ + { + "bbox": [ + 75, + 423, + 157, + 432 + ], + "type": "text", + "content": "COMET explanation:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 75, + 433, + 508, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 433, + 508, + 468 + ], + "spans": [ + { + "bbox": [ + 75, + 433, + 508, + 468 + ], + "type": "text", + "content": "_Ash_O' Brien_Und_Ehe mann_Ja rett_Kelley_sagen, dass sie_am_Donnerstag_im_Hunde park_Dusty_Rhod es_in_San_Diego_einen_Happ_en_zu_essen_griff_en_, _mit_threm_drei_Monate_alten_M ops_im_Schleppt au." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 75, + 477, + 105, + 486 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 477, + 105, + 486 + ], + "spans": [ + { + "bbox": [ + 75, + 477, + 105, + 486 + ], + "type": "text", + "content": "Source:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 75, + 487, + 255, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 487, + 255, + 498 + ], + "spans": [ + { + "bbox": [ + 75, + 487, + 255, + 498 + ], + "type": "text", + "content": "It was Einstein's great general theory of relativity." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 75, + 500, + 123, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 500, + 123, + 509 + ], + "spans": [ + { + "bbox": [ + 75, + 500, + 123, + 509 + ], + "type": "text", + "content": "Translation:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 75, + 510, + 336, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 510, + 336, + 521 + ], + "spans": [ + { + "bbox": [ + 75, + 510, + 336, + 521 + ], + "type": "text", + "content": "Es war Einsteins große allgemeine Forschungen vor Relativitätstheorie." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 75, + 523, + 160, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 523, + 160, + 533 + ], + "spans": [ + { + "bbox": [ + 75, + 523, + 160, + 533 + ], + "type": "text", + "content": "COMET score: 0.6692" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 75, + 534, + 157, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 534, + 157, + 544 + ], + "spans": [ + { + "bbox": [ + 75, + 534, + 157, + 544 + ], + "type": "text", + "content": "COMET explanation:" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 75, + 544, + 438, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 544, + 438, + 555 + ], + "spans": [ + { + "bbox": [ + 75, + 544, + 438, + 555 + ], + "type": "text", + "content": "_Es _war _Einstein s _große _allgemeine e _Forschung en _vor _Relativ itäts the ori e ." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 75, + 564, + 105, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 564, + 105, + 573 + ], + "spans": [ + { + "bbox": [ + 75, + 564, + 105, + 573 + ], + "type": "text", + "content": "Source:" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 75, + 574, + 271, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 574, + 271, + 584 + ], + "spans": [ + { + "bbox": [ + 75, + 574, + 271, + 584 + ], + "type": "text", + "content": "There's mask-shaming and then there's full on assault." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 75, + 587, + 123, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 587, + 123, + 596 + ], + "spans": [ + { + "bbox": [ + 75, + 587, + 123, + 596 + ], + "type": "text", + "content": "Translation:" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 75, + 597, + 289, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 597, + 289, + 608 + ], + "spans": [ + { + "bbox": [ + 75, + 597, + 289, + 608 + ], + "type": "text", + "content": "Es gibt Maskenschämen und dann ist es voll bei Angriff." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 75, + 609, + 160, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 609, + 160, + 618 + ], + "spans": [ + { + "bbox": [ + 75, + 609, + 160, + 618 + ], + "type": "text", + "content": "COMET score: 0.2318" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 75, + 619, + 157, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 619, + 157, + 629 + ], + "spans": [ + { + "bbox": [ + 75, + 619, + 157, + 629 + ], + "type": "text", + "content": "COMET explanation:" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 75, + 629, + 385, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 629, + 385, + 641 + ], + "spans": [ + { + "bbox": [ + 75, + 629, + 385, + 641 + ], + "type": "text", + "content": "_Es_gibt_Mask en schä men_und_dann_ist es_voll_bei_Angriff_." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 69, + 654, + 524, + 689 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 654, + 524, + 689 + ], + "spans": [ + { + "bbox": [ + 69, + 654, + 524, + 689 + ], + "type": "text", + "content": "Table 9: Saliency map for COMET explanation scores for a set of en→de examples. Comparing the token-level explanations with the MQM annotation (highlighted in gray) reveals the source of correspondence between specific token-level translation errors and the resulting scores." + } + ] + } + ], + "index": 35, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1102" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 74, + 155, + 105, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 155, + 105, + 163 + ], + "spans": [ + { + "bbox": [ + 74, + 155, + 105, + 163 + ], + "type": "text", + "content": "Source:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 74, + 164, + 465, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 164, + 465, + 175 + ], + "spans": [ + { + "bbox": [ + 74, + 164, + 465, + 175 + ], + "type": "text", + "content": "我想告诉大家宇宙有着自己的配乐，而宇宙自身正在不停地播放着。因为太空可以想鼓一样振动。" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 75, + 178, + 123, + 187 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 178, + 123, + 187 + ], + "spans": [ + { + "bbox": [ + 75, + 178, + 123, + 187 + ], + "type": "text", + "content": "Translation:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 74, + 188, + 509, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 188, + 509, + 208 + ], + "spans": [ + { + "bbox": [ + 74, + 188, + 509, + 208 + ], + "type": "text", + "content": "I want to tell you that the universe has its own iconic soundtrack and the universe itself is constantly playing non-stop because space can vibrate like a drum." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 75, + 211, + 160, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 211, + 160, + 220 + ], + "spans": [ + { + "bbox": [ + 75, + 211, + 160, + 220 + ], + "type": "text", + "content": "COMET score: 0.8634" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 75, + 222, + 157, + 231 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 222, + 157, + 231 + ], + "spans": [ + { + "bbox": [ + 75, + 222, + 157, + 231 + ], + "type": "text", + "content": "COMET explanation:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 75, + 231, + 508, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 231, + 508, + 254 + ], + "spans": [ + { + "bbox": [ + 75, + 231, + 508, + 254 + ], + "type": "text", + "content": "_I_want_to_tell你想_the_univers_e_has_its_own(iconic soundtrack_and_the_univers_e_itself_is_constantly-playing_non-stop_because_space_can_vibrate_to_like_a Drum." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 75, + 264, + 105, + 273 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 264, + 105, + 273 + ], + "spans": [ + { + "bbox": [ + 75, + 264, + 105, + 273 + ], + "type": "text", + "content": "Source:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 75, + 273, + 363, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 273, + 363, + 285 + ], + "spans": [ + { + "bbox": [ + 75, + 273, + 363, + 285 + ], + "type": "text", + "content": "另外，吉克隽逸和刘烨作为运动助理，也围绕运动少年制造了不少爆笑话题。" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 75, + 287, + 123, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 287, + 123, + 296 + ], + "spans": [ + { + "bbox": [ + 75, + 287, + 123, + 296 + ], + "type": "text", + "content": "Translation:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 75, + 297, + 509, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 297, + 509, + 308 + ], + "spans": [ + { + "bbox": [ + 75, + 297, + 509, + 308 + ], + "type": "text", + "content": "In addition, as sports assistants, Ji Kejunyi and Liu Ye have also created a lot of hilarious topics around sports teenagers." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 75, + 310, + 160, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 310, + 160, + 320 + ], + "spans": [ + { + "bbox": [ + 75, + 310, + 160, + 320 + ], + "type": "text", + "content": "COMET score: 0.8214" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 75, + 321, + 157, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 321, + 157, + 331 + ], + "spans": [ + { + "bbox": [ + 75, + 321, + 157, + 331 + ], + "type": "text", + "content": "COMET explanation:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 75, + 331, + 508, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 331, + 508, + 354 + ], + "spans": [ + { + "bbox": [ + 75, + 331, + 508, + 354 + ], + "type": "text", + "content": "_In _addition , _as _sports _assistant s , _Ji _Ke ju nyi _and _Li u _Ye _have _also _created _a _lot _of_ hila rious _topic s _around _sports _teenager s ." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 75, + 364, + 105, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 364, + 105, + 372 + ], + "spans": [ + { + "bbox": [ + 75, + 364, + 105, + 372 + ], + "type": "text", + "content": "Source:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 75, + 373, + 287, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 373, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 75, + 373, + 287, + 384 + ], + "type": "text", + "content": "一番言论让场上的少年和运动领队们都倒吸一口凉气。" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 75, + 386, + 123, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 386, + 123, + 396 + ], + "spans": [ + { + "bbox": [ + 75, + 386, + 123, + 396 + ], + "type": "text", + "content": "Translation:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 75, + 396, + 393, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 396, + 393, + 407 + ], + "spans": [ + { + "bbox": [ + 75, + 396, + 393, + 407 + ], + "type": "text", + "content": "The remarks made the teenagers and the sports leaders on the field gasp a sigh of relief." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 75, + 409, + 160, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 409, + 160, + 419 + ], + "spans": [ + { + "bbox": [ + 75, + 409, + 160, + 419 + ], + "type": "text", + "content": "COMET score: 0.7793" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 75, + 420, + 157, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 420, + 157, + 430 + ], + "spans": [ + { + "bbox": [ + 75, + 420, + 157, + 430 + ], + "type": "text", + "content": "COMET explanation:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 75, + 430, + 509, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 430, + 509, + 454 + ], + "spans": [ + { + "bbox": [ + 75, + 430, + 509, + 454 + ], + "type": "text", + "content": "The re marks made the teenager s and the sports leaders on the field gas p a sig h of _relief" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 75, + 463, + 105, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 463, + 105, + 471 + ], + "spans": [ + { + "bbox": [ + 75, + 463, + 105, + 471 + ], + "type": "text", + "content": "Source:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 75, + 472, + 178, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 472, + 178, + 483 + ], + "spans": [ + { + "bbox": [ + 75, + 472, + 178, + 483 + ], + "type": "text", + "content": "强烈的阳光是如此地刺眼，" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 75, + 486, + 123, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 486, + 123, + 495 + ], + "spans": [ + { + "bbox": [ + 75, + 486, + 123, + 495 + ], + "type": "text", + "content": "Translation:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 75, + 496, + 194, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 496, + 194, + 507 + ], + "spans": [ + { + "bbox": [ + 75, + 496, + 194, + 507 + ], + "type": "text", + "content": "The intense sunlight is so harsh;" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 75, + 508, + 160, + 518 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 508, + 160, + 518 + ], + "spans": [ + { + "bbox": [ + 75, + 508, + 160, + 518 + ], + "type": "text", + "content": "COMET score: 0.7561" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 75, + 518, + 157, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 518, + 157, + 528 + ], + "spans": [ + { + "bbox": [ + 75, + 518, + 157, + 528 + ], + "type": "text", + "content": "COMET explanation:" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 75, + 528, + 252, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 528, + 252, + 540 + ], + "spans": [ + { + "bbox": [ + 75, + 528, + 252, + 540 + ], + "type": "text", + "content": "The intense sun light is so har sh" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 75, + 550, + 105, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 550, + 105, + 559 + ], + "spans": [ + { + "bbox": [ + 75, + 550, + 105, + 559 + ], + "type": "text", + "content": "Source:" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 75, + 559, + 339, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 559, + 339, + 570 + ], + "spans": [ + { + "bbox": [ + 75, + 559, + 339, + 570 + ], + "type": "text", + "content": "如今，我们希望能够给这部关于宇宙的宏伟的视觉作品配上声音。" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 75, + 572, + 123, + 581 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 572, + 123, + 581 + ], + "spans": [ + { + "bbox": [ + 75, + 572, + 123, + 581 + ], + "type": "text", + "content": "Translation:" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 75, + 582, + 396, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 582, + 396, + 594 + ], + "spans": [ + { + "bbox": [ + 75, + 582, + 396, + 594 + ], + "type": "text", + "content": "Today, we hope to be able to give this magnificent visual work of the universe a sound." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 75, + 596, + 160, + 606 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 596, + 160, + 606 + ], + "spans": [ + { + "bbox": [ + 75, + 596, + 160, + 606 + ], + "type": "text", + "content": "COMET score: 0.7073" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 75, + 606, + 157, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 606, + 157, + 616 + ], + "spans": [ + { + "bbox": [ + 75, + 606, + 157, + 616 + ], + "type": "text", + "content": "COMET explanation:" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 75, + 616, + 509, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 616, + 509, + 640 + ], + "spans": [ + { + "bbox": [ + 75, + 616, + 509, + 640 + ], + "type": "text", + "content": "Today,we hope to be able to give this magnificent ent visual work of the univers e a sound." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 69, + 651, + 524, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 651, + 524, + 687 + ], + "spans": [ + { + "bbox": [ + 69, + 651, + 524, + 687 + ], + "type": "text", + "content": "Table 10: Saliency map for COMET explanation scores for a set of " + }, + { + "bbox": [ + 69, + 651, + 524, + 687 + ], + "type": "inline_equation", + "content": "\\mathrm{{zh}} \\rightarrow \\mathrm{{en}}" + }, + { + "bbox": [ + 69, + 651, + 524, + 687 + ], + "type": "text", + "content": " examples. Comparing the token-level explanations with the MQM annotation (highlighted in gray) reveals the source of correspondence between specific token-level translation errors and the resulting scores." + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1103" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 90, + 192, + 103 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 90, + 192, + 103 + ], + "spans": [ + { + "bbox": [ + 68, + 90, + 192, + 103 + ], + "type": "text", + "content": "A For every submission:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 106, + 317, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 106, + 317, + 120 + ], + "spans": [ + { + "bbox": [ + 77, + 106, + 317, + 120 + ], + "type": "text", + "content": "A1. Did you describe the limitations of your work?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 121, + 153, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 121, + 153, + 132 + ], + "spans": [ + { + "bbox": [ + 89, + 121, + 153, + 132 + ], + "type": "text", + "content": "Yes. Section 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 76, + 142, + 329, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 142, + 329, + 156 + ], + "spans": [ + { + "bbox": [ + 76, + 142, + 329, + 156 + ], + "type": "text", + "content": "A2. Did you discuss any potential risks of your work?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 157, + 209, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 157, + 209, + 169 + ], + "spans": [ + { + "bbox": [ + 89, + 157, + 209, + 169 + ], + "type": "text", + "content": "Not applicable. Left blank." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 77, + 178, + 414, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 178, + 414, + 191 + ], + "spans": [ + { + "bbox": [ + 77, + 178, + 414, + 191 + ], + "type": "text", + "content": "A3. Do the abstract and introduction summarize the paper's main claims?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 89, + 193, + 191, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 193, + 191, + 204 + ], + "spans": [ + { + "bbox": [ + 89, + 193, + 191, + 204 + ], + "type": "text", + "content": "Abstract and Section 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 77, + 213, + 399, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 213, + 399, + 227 + ], + "spans": [ + { + "bbox": [ + 77, + 213, + 399, + 227 + ], + "type": "text", + "content": "A4. Have you used AI writing assistants when working on this paper?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 88, + 228, + 524, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 228, + 524, + 242 + ], + "spans": [ + { + "bbox": [ + 88, + 228, + 524, + 242 + ], + "type": "text", + "content": "Assistance purely with the language of the paper along every section. Grammarly and DeepL write" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 250, + 291, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 250, + 291, + 264 + ], + "spans": [ + { + "bbox": [ + 67, + 250, + 291, + 264 + ], + "type": "text", + "content": "B Did you use or create scientific artifacts?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 268, + 525, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 268, + 525, + 295 + ], + "spans": [ + { + "bbox": [ + 67, + 268, + 525, + 295 + ], + "type": "text", + "content": "Section 3 explains the methods we used. We will release the adaptations required to use the explainability methods over COMET framework, the UniTE model we trained, and all data synthetically-generated data." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 77, + 302, + 315, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 302, + 315, + 316 + ], + "spans": [ + { + "bbox": [ + 77, + 302, + 315, + 316 + ], + "type": "text", + "content": "B1. Did you cite the creators of artifacts you used?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 90, + 317, + 133, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 317, + 133, + 328 + ], + "spans": [ + { + "bbox": [ + 90, + 317, + 133, + 328 + ], + "type": "text", + "content": "Section 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 77, + 338, + 463, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 338, + 463, + 352 + ], + "spans": [ + { + "bbox": [ + 77, + 338, + 463, + 352 + ], + "type": "text", + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 89, + 353, + 345, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 353, + 345, + 365 + ], + "spans": [ + { + "bbox": [ + 89, + 353, + 345, + 365 + ], + "type": "text", + "content": "footnote on the first page. The License will be Apache 2.0" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 76, + 375, + 524, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 375, + 524, + 428 + ], + "spans": [ + { + "bbox": [ + 76, + 375, + 524, + 428 + ], + "type": "text", + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 89, + 429, + 209, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 429, + 209, + 442 + ], + "spans": [ + { + "bbox": [ + 89, + 429, + 209, + 442 + ], + "type": "text", + "content": "Not applicable. Left blank." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 76, + 451, + 524, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 451, + 524, + 491 + ], + "spans": [ + { + "bbox": [ + 76, + 451, + 524, + 491 + ], + "type": "text", + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 89, + 492, + 209, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 492, + 209, + 505 + ], + "spans": [ + { + "bbox": [ + 89, + 492, + 209, + 505 + ], + "type": "text", + "content": "Not applicable. Left blank." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 77, + 513, + 524, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 513, + 524, + 540 + ], + "spans": [ + { + "bbox": [ + 77, + 513, + 524, + 540 + ], + "type": "text", + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 89, + 541, + 402, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 541, + 402, + 554 + ], + "spans": [ + { + "bbox": [ + 89, + 541, + 402, + 554 + ], + "type": "text", + "content": "in the Appendix we have several statistics for training and testing data." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 77, + 562, + 525, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 562, + 525, + 630 + ], + "spans": [ + { + "bbox": [ + 77, + 562, + 525, + 630 + ], + "type": "text", + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 89, + 631, + 138, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 631, + 138, + 644 + ], + "spans": [ + { + "bbox": [ + 89, + 631, + 138, + 644 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 67, + 651, + 293, + 666 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 651, + 293, + 666 + ], + "spans": [ + { + "bbox": [ + 67, + 651, + 293, + 666 + ], + "type": "text", + "content": "C Did you run computational experiments?" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 77, + 670, + 364, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 670, + 364, + 683 + ], + "spans": [ + { + "bbox": [ + 77, + 670, + 364, + 683 + ], + "type": "text", + "content": "Appendix provides detailed information about the trained model." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 77, + 690, + 524, + 719 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 690, + 524, + 719 + ], + "spans": [ + { + "bbox": [ + 77, + 690, + 524, + 719 + ], + "type": "text", + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 88, + 719, + 524, + 745 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 719, + 524, + 745 + ], + "spans": [ + { + "bbox": [ + 88, + 719, + 524, + 745 + ], + "type": "text", + "content": "Appendix provides detailed information about the trained model including GPU infrastructure and total number of parameters." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "spans": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "text", + "content": "ACL 2023 Responsible NLP Checklist" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 751, + 522, + 772 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 751, + 522, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 751, + 522, + 772 + ], + "type": "text", + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "1104" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 76, + 71, + 523, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 71, + 523, + 97 + ], + "spans": [ + { + "bbox": [ + 76, + 71, + 523, + 97 + ], + "type": "text", + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 89, + 99, + 208, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 99, + 208, + 111 + ], + "spans": [ + { + "bbox": [ + 89, + 99, + 208, + 111 + ], + "type": "text", + "content": "Not applicable. Left blank." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 120, + 525, + 160 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 120, + 525, + 160 + ], + "spans": [ + { + "bbox": [ + 77, + 120, + 525, + 160 + ], + "type": "text", + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 162, + 461, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 162, + 461, + 175 + ], + "spans": [ + { + "bbox": [ + 88, + 162, + 461, + 175 + ], + "type": "text", + "content": "Appendix has all information needed about test data and performance of the models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 183, + 525, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 183, + 525, + 223 + ], + "spans": [ + { + "bbox": [ + 77, + 183, + 525, + 223 + ], + "type": "text", + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 225, + 196, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 225, + 196, + 238 + ], + "spans": [ + { + "bbox": [ + 89, + 225, + 196, + 238 + ], + "type": "text", + "content": "Section 2 and Appendix" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 247, + 522, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 247, + 522, + 261 + ], + "spans": [ + { + "bbox": [ + 67, + 247, + 522, + 261 + ], + "type": "text", + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 79, + 264, + 127, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 264, + 127, + 277 + ], + "spans": [ + { + "bbox": [ + 79, + 264, + 127, + 277 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 76, + 286, + 525, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 286, + 525, + 313 + ], + "spans": [ + { + "bbox": [ + 76, + 286, + 525, + 313 + ], + "type": "text", + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 89, + 314, + 208, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 314, + 208, + 327 + ], + "spans": [ + { + "bbox": [ + 89, + 314, + 208, + 327 + ], + "type": "text", + "content": "Not applicable. Left blank." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 76, + 336, + 525, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 336, + 525, + 376 + ], + "spans": [ + { + "bbox": [ + 76, + 336, + 525, + 376 + ], + "type": "text", + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 89, + 377, + 208, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 377, + 208, + 391 + ], + "spans": [ + { + "bbox": [ + 89, + 377, + 208, + 391 + ], + "type": "text", + "content": "Not applicable. Left blank." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 76, + 400, + 525, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 400, + 525, + 439 + ], + "spans": [ + { + "bbox": [ + 76, + 400, + 525, + 439 + ], + "type": "text", + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 89, + 440, + 208, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 440, + 208, + 454 + ], + "spans": [ + { + "bbox": [ + 89, + 440, + 208, + 454 + ], + "type": "text", + "content": "Not applicable. Left blank." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 76, + 462, + 520, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 462, + 520, + 476 + ], + "spans": [ + { + "bbox": [ + 76, + 462, + 520, + 476 + ], + "type": "text", + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 89, + 476, + 208, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 476, + 208, + 490 + ], + "spans": [ + { + "bbox": [ + 89, + 476, + 208, + 490 + ], + "type": "text", + "content": "Not applicable. Left blank." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 76, + 498, + 524, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 498, + 524, + 524 + ], + "spans": [ + { + "bbox": [ + 76, + 498, + 524, + 524 + ], + "type": "text", + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 89, + 526, + 208, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 526, + 208, + 539 + ], + "spans": [ + { + "bbox": [ + 89, + 526, + 208, + 539 + ], + "type": "text", + "content": "Not applicable. Left blank." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1105" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/The Mechanical Bard_ An Interpretable Machine Learning Approach to Shakespearean Sonnet Generation/3330f90d-0cb8-4666-aeb2-f9f31ccb9534_content_list.json b/2023/The Mechanical Bard_ An Interpretable Machine Learning Approach to Shakespearean Sonnet Generation/3330f90d-0cb8-4666-aeb2-f9f31ccb9534_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ae62c03574048b19ac2a1749b52047c8e52a6cb2 --- /dev/null +++ b/2023/The Mechanical Bard_ An Interpretable Machine Learning Approach to Shakespearean Sonnet Generation/3330f90d-0cb8-4666-aeb2-f9f31ccb9534_content_list.json @@ -0,0 +1,2196 @@ +[ + { + "type": "text", + "text": "The Mechanical Bard: An Interpretable Machine Learning Approach to Shakespearean Sonnet Generation", + "text_level": 1, + "bbox": [ + 124, + 89, + 872, + 130 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Edwin Agnew*, Michelle Qiu*, Lily Zhu*, Sam Wiseman, Cynthia Rudin", + "bbox": [ + 186, + 149, + 815, + 166 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Duke University, Durham, NC", + "bbox": [ + 374, + 167, + 626, + 181 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "edwin.agnew@duke.edu, michelle.qiu@duke.edu, lily.zhu@duke.edu", + "bbox": [ + 186, + 183, + 815, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "swiseman@cs.duke.edu, cynthiaa@cs.duke.edu", + "bbox": [ + 294, + 200, + 709, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 252, + 339, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We consider the automated generation of sonnets, a poetic form constrained according to meter, rhyme scheme, and length. Sonnets generally also use rhetorical figures, expressive language, and a consistent theme or narrative. Our constrained decoding approach allows for the generation of sonnets within preset poetic constraints, while using a relatively modest neural backbone. Human evaluation confirms that our approach produces Shakespearean sonnets that resemble human-authored sonnets, and which adhere to the genre's defined constraints and contain lyrical language and literary devices.", + "bbox": [ + 141, + 280, + 460, + 464 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "When all the lovers of this world are dead,", + "bbox": [ + 566, + 273, + 826, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The sun of heaven on a golden day", + "bbox": [ + 589, + 286, + 803, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To burn the earth's fire by the flame and spread", + "bbox": [ + 552, + 298, + 838, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Where all the flowers of your fair days lay.", + "bbox": [ + 564, + 310, + 826, + 321 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "These are the blossoms that you take care of.", + "bbox": [ + 559, + 321, + 831, + 332 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Why do you linger such a long delay?", + "bbox": [ + 579, + 332, + 811, + 344 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Forgive the fluttered flower of meek love", + "bbox": [ + 571, + 344, + 821, + 356 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Or who you have so long to love the day?", + "bbox": [ + 569, + 357, + 821, + 368 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The joys of love, the beauty on the face,", + "bbox": [ + 574, + 369, + 818, + 380 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shall be your fate and be your own delight.", + "bbox": [ + 564, + 381, + 826, + 392 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "You have the beauty of your own embrace.", + "bbox": [ + 566, + 393, + 825, + 403 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "You cannot reminiscence. Cannot write.", + "bbox": [ + 574, + 404, + 818, + 414 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Between the living and the deadening breath.", + "bbox": [ + 559, + 416, + 831, + 426 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "You go the way of your beloved death.", + "bbox": [ + 578, + 428, + 813, + 439 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Figure 1: A sonnet generated with the theme \"death\".", + "bbox": [ + 512, + 457, + 875, + 472 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 114, + 489, + 258, + 506 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We consider the task of automatically generating Shakespearean sonnets, a popular poetic form with highly specific rhyme and meter constraints1. Each sonnet consists of three quatrains followed by a single couplet according to the rhyme scheme ABAB BCBC CDCD EE, and each line contains ten syllables with a stress pattern of iambic pentameter.", + "bbox": [ + 112, + 516, + 489, + 627 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rather than train a model to obey these constraints implicitly (which leads to enormous models that still do not obey the constraints), we opt to enforce them explicitly using a simple but novel approach to generation.", + "bbox": [ + 112, + 630, + 489, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In particular, we use part-of-speech (POS) templates selected and edited from individual lines in Shakespeare's sonnets, with each template intended to offer a different combination of parts of speech and narrative directions. Associated thematic words are then selected and placed at the end of each template, and their rhyming pairs are chosen dynamically by a language model (e.g., GPT-2, Radford et al., 2019) and placed at the end of the corresponding lines according to the rhyme scheme.", + "bbox": [ + 112, + 709, + 489, + 870 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The rest of the line is filled with related words that fit the specified POS and meter, leading to the end rhyme word. Figure 1 shows sample output.", + "bbox": [ + 507, + 505, + 882, + 552 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our use of these templates ensures sophisticated-seeming language and syntax that competing systems do not capture. Our approach provides excellent grammatical structure comparable to that of human-written poetry, all while using a relatively simple model and generation procedure.", + "bbox": [ + 507, + 555, + 884, + 651 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We extensively evaluate the ability of our approach to generate whole sonnets (a setting often ignored by recent work in poetry generation) and find that our approach is preferred over strong baselines by both expert annotators (recruited from an academic English department) and by crowdworkers. As this research was conducted before the release of ChatGPT, we were not able to robustly compare our model's performance against this language model. However, we make several observations about the poetic quality of sonnets generated by ChatGPT.", + "bbox": [ + 507, + 652, + 882, + 844 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 507, + 859, + 665, + 873 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Early attempts at poetry generation relied mainly on rule-based methods (Gervás, 2000; Oliveira,", + "bbox": [ + 507, + 887, + 882, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* denotes equal contribution", + "bbox": [ + 136, + 879, + 305, + 892 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1Our code is available at https://github.com/", + "bbox": [ + 136, + 892, + 485, + 906 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "edwinagnew/Poetix_Sonnets", + "bbox": [ + 115, + 906, + 305, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1627", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics", + "bbox": [ + 226, + 945, + 769, + 958 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Volume 2: Short Papers, pages 1627-1638", + "bbox": [ + 368, + 959, + 628, + 971 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "July 9-14, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 295, + 972, + 700, + 985 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2012; Manurung et al., 2000; Veale, 2013). More recent automated poetry generation techniques, especially for sonnet generation, have relied on combinations of task-specific language models and rules. For instance, Ghazvininejad et al. (2016)'s Hafez uses a finite state acceptor to generate a large number of possible lines, the best of which are then selected with an RNN trained on song lyrics. Like our approach, they use rhyming dictionaries to find rhyming words and word embeddings to find topical words. Similarly, Benhardt et al. (2018) preselects rhyming words and generates lines backwards with a recurrent neural network (RNN). Also in this vein are Lau et al. (2018)'s Deepspare, which consists of an LSTM language model, an iambic model, and a rhyming model, and the recent work of Van de Cruys (2020) and Wang et al. (2021).", + "bbox": [ + 110, + 84, + 489, + 357 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our approach distinguishes itself in using a general-purpose pretrained language model, but more importantly in its use of human-curated constraints and templates. These allow for generating high-quality poems with a very simple approach.", + "bbox": [ + 112, + 359, + 489, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 112, + 453, + 263, + 469 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The general idea of our approach is to take a pretrained language model (in this case GPT-2) and apply hard constraints to the generation procedure so that it can only output text satisfying various poetic constraints. These constraints can be broadly divided into hard constraints (e.g., number of syllables) and soft constraints (e.g., sounding poetic), and our methodology can be separated similarly. Our generation process is in Figure 2.", + "bbox": [ + 112, + 481, + 490, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1 POS Templates", + "text_level": 1, + "bbox": [ + 112, + 639, + 282, + 653 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The most important part of our method is the use of handcrafted grammar templates. Taking inspiration from existing sonnets, we created a list of about 120 templates that encode the part-of-speech structure of a line of poetry. Each template can generate an unbounded number of possible poetic lines. For example, the line \"The beauty of life on a lonely sea\" is represented by the template \"THE NN OF NN ON A JJ NN.\" More sample templates are in Section A.1. Since the templates allow for considerable flexibility, obeying the templates does not alone suffice for poetry. For example, the same template could be used to write poetic lines with distinct meanings such as \"The tree of anguish on a stormy night\" or a nonsensical line like \"The fork of ant on an unpacked transfer.\" A subset of these", + "bbox": [ + 112, + 661, + 489, + 917 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "templates is also chosen for starting a stanza.", + "bbox": [ + 507, + 84, + 845, + 99 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.2 Strict Sonnet Constraints", + "text_level": 1, + "bbox": [ + 507, + 111, + 756, + 124 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The two most critical features of sonnets distinguishing them from other poetry forms are that they are written in iambic pentameter (i.e., each line has 10 syllables of alternating stress pattern), and they follow an ABAB CDCD EFEF GG rhyme scheme. To detect iambic pentameter, we use the CMU Pronouncing Dictionary (CMU, 2019), which reveals how many syllables a word contains and the stress of each syllable. An unstressed syllable is represented as '0' and a stressed syllable as '1', and so the line \"The beauty of life on a lonely sea\" is represented as '0 10 1 0 1 0 10 1'. For simplicity, 1-syllable words can be designated as either 0 or 1.", + "bbox": [ + 505, + 131, + 884, + 340 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given a POS-tag for every word in our dictionary, we create a tree-like data structure that represents every possible meter for a given template. Continuing the example, the first word could only be 'the', but the second word could be filled with a 1-syllable noun like 'tree', a 2-syllable noun like 'chaos' (10), or a 3-syllable noun like 'audio' (101), and so on. Each choice affects the possible pronunciations of the next word as well as the number of remaining words in order to reach 10 syllables. The pronunciation dictionary ensures the last syllable of the last word on each line matches its partner.", + "bbox": [ + 507, + 341, + 885, + 535 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.3 Language Model", + "text_level": 1, + "bbox": [ + 507, + 545, + 690, + 560 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We use a language model to generate individual sonnet lines, subject to the formal constraints outlined above. In particular, we first fine-tune GPT-2 (Radford et al., 2019) on a large corpus of over 15000 poems and a smaller corpus of sonnets. We then use a constrained beam-search to generate each line, where only legal tokens (under the aforementioned constraints) can be generated at each step; this generation approach resembles previous constrained decoding techniques used in sonnet generation (Ghazvininejad et al., 2016), although our approach differs in the choice of model and direct enforcement of constraints. For a comparison of generation quality using a GPT-2 model that has not been fine-tuned, see Section 4.1.", + "bbox": [ + 507, + 565, + 884, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.4 Thematic Word Choice", + "text_level": 1, + "bbox": [ + 507, + 818, + 741, + 832 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To ensure the content of the poem fits the theme specified by the user, we provide an excerpt of a", + "bbox": [ + 507, + 839, + 882, + 871 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "$^{2}$ https://www.kaggle.com/datasets/johnhallman/completpoetryfoundation.org-dataset", + "bbox": [ + 507, + 879, + 882, + 904 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "3https://www.kaggle.com/datasets/michelleqiu/sonnets", + "bbox": [ + 532, + 904, + 865, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1628", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7f5699c3abb074e217a003ab4ce4e8055fa6be50e33b760657614f7eed444060.jpg", + "image_caption": [ + "Generation Visualization", + "Figure 2: Numbers in parentheses denote subsections in Section 3." + ], + "image_footnote": [], + "bbox": [ + 139, + 96, + 460, + 444 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "theme-appropriate poem as additional context to GPT-2 during generation. This additional poem is selected by finding a list of synonyms to the theme word using the WordNet synonym database (Miller, 1998) and then choosing lines from a poem corpus that contain at least one synonym. We also remove words from the vocabulary if they have less than 0.5 cosine similarity with the theme word, based on the corresponding FastText word embeddings (Bojanowski et al., 2017). This avoids having words like \"algebra\" in poems with themes like \"forest.\"", + "bbox": [ + 112, + 512, + 489, + 690 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.5 Generation Procedure", + "text_level": 1, + "bbox": [ + 112, + 703, + 334, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Having introduced our method's components, we now describe the generation procedure. A user inputs a theme word, a beam search parameter, $b$ , and the number of templates sampled per line, $k$ . A seed is chosen with the above method. Then for each line, we sample $k$ random templates. For each template, we generate the line using a modified beam search. Specifically, the beam search maintains $b$ different hypotheses per line at all times. For each hypothesis, we first mask out any tokens that violate our hard POS, meter, or rhyme constraints and select the $b$ best next-tokens for each", + "bbox": [ + 112, + 726, + 489, + 917 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of the $k$ templates. These $b^{2}$ new candidates are re-ranked according to our custom scoring function, and the top $k \\times b$ proceed to the next stage. The constraint-filtering at each stage guarantees that the generated line will match the input template, while the beam search allows more flexible word choice than greedy word-filling for each POS. If none of the $k \\times b$ generated lines score better than a specific threshold, then a new template is chosen and the line is generated again. Otherwise, line generation continues until the poem is completed.", + "bbox": [ + 507, + 84, + 884, + 261 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.6 Poetic Devices", + "text_level": 1, + "bbox": [ + 507, + 274, + 668, + 288 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To make the poems more poetic, we adjust our scoring function to weight lines with alliteration, penalties for repetition, and/or internal rhyme. Alliteration occurs when a line contains words starting with the same letter, repetition occurs when a word is present several times throughout a poem, and internal rhyme occurs when two words rhyme within the same line. To weight alliteration, when the first token of a new word is being generated, a list $\\vec{A} = [a_1,a_2,\\dots a_n]$ is generated where $a_{i}$ is the number of occurrences of the first letter of the ith token in the current line. To weight and discourage repetition, a list $\\vec{T} = [t_1,t_2,\\dots t_n]$ is generated where $t_i$ is the number of occurrences of the ith token in the poem, negated. To weight internal rhyme, a list $\\vec{R} = [r_1,r_2,\\dots ,r_n]$ is generated where $r_i = 1$ if the ith token is part of a word that rhymes with any of the words in the current line generated so far, and $r_i = 0$ otherwise. The final token distribution is then proportional to $\\tilde{P} +\\alpha_{A}\\times \\vec{A} +\\alpha_{T}\\times \\vec{T} +\\alpha_{R}\\times \\vec{R},$ where $\\tilde{P}$ is the language model's next-token distribution, and $\\alpha_{A},\\alpha_{T},$ and $\\alpha_{R}$ are user-specified non-negative parameters, which represent the degree to which alliteration, repetition, and internal rhyme should be favored during generation.", + "bbox": [ + 507, + 297, + 884, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.7 Postprocessing", + "text_level": 1, + "bbox": [ + 507, + 727, + 672, + 743 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "After a poem is completed and all 14 lines score above a fixed threshold, a small number of adjustments are made. These include fixing common mistakes made by GPT-2 like not capitalizing the word 'I' and not capitalizing following punctuation.", + "bbox": [ + 507, + 749, + 884, + 829 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 507, + 844, + 655, + 860 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We used human input to test our sonnets against both model-generated and human-written sonnets. To test adherence to a theme throughout a son", + "bbox": [ + 507, + 871, + 884, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1629", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/ded253baca0ff113755368f27b80d531e62e6c9281fb37033d28e0e5c59124a2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CategoryMeanp-value
PoeTryMe
Grammar4.50*1.71×10-4
Emotion4.30*3.13×10-3
Poetic4.30*3.13×10-3
Human4.10*5.77×10-3
Theme2.600.211286
Benhardt et al.
Grammar3.83*0.03
Emotion3.67*0.05
Poetic3.75*0.04
Human3.75*0.02
Theme2.420.06
Human-written poems
Grammar1.361.00×10-6
Emotion1.45.00×10-6
Poetic1.645.40×10-5
Human1.361.00×10-6
Theme1.577.70×10-5
", + "bbox": [ + 189, + 80, + 410, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 3: Starred figures indicate average scores of $>3$ , and underlined figures indicate that the p-value is low enough $(<0.05)$ to claim that this higher average is statistically significant.", + "bbox": [ + 112, + 394, + 489, + 453 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "net, we desired baselines that generate whole sonnets with user-provided themes. This limited our competitors, as some generate discrete quatrains or generate without input themes (e.g., Deepspare), leaving only Benhardt et al. (2018) and PoeTryMe (Oliveira, 2012); see Section A.2.", + "bbox": [ + 112, + 480, + 487, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Furthermore, an evaluation of poetry quality is incomplete without human-written sonnets, selected from sonnets.org. Though these poems do not have an explicit theme, we selected poems that followed our five themes.", + "bbox": [ + 112, + 577, + 487, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To optimally test our model, we conducted an internal analysis and selected $k$ values sampled from 3, 5, or 7, $b$ values sampled from 3, 5, or 7, and repetition penalty values sampled from 1.4, 1.6, or 1.8 that we concluded produced the highest quality sonnets. To evaluate adherence to theme, we generated poems with themes \"death,\" \"darkness,\" \"forest,\" \"love,\" and \"wisdom.\"", + "bbox": [ + 112, + 659, + 487, + 788 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In each test, respondents compared six randomly selected pairs of sonnets, with each of our sonnets displayed with a competing model/human-written sonnet generated with the same theme word. Respondents indicated which of the two sonnets performed better in categories of theme, poeticness, grammar, emotion, and likelihood of being human-written. Detailed instructions are in A.3.", + "bbox": [ + 112, + 790, + 487, + 917 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/dc04809b9b89d272119f0de81e744c664234c519eed4438b27e6b326efa1e39b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CategoryMeanp-value
PoeTryMe
Grammar3.66*2.00 × 10-6
Emotion3.54*1.16 × 10-4
Poetic3.55*3.70 × 10-5
Human3.59*1.60 × 10-5
Theme2.860.19
Benhardt et al.
Grammar3.34*6.57 × 10-3
Emotion3.16*0.12
Poetic3.11*0.19
Human3.06*0.33
Theme2.770.06
Human-written poems
Grammar3.13*0.14
Emotion2.860.14
Poetic2.910.24
Human2.920.27
Theme2.670.02
", + "bbox": [ + 583, + 80, + 806, + 370 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 4: Starred figures indicate average scores of $>3$ , and underlined figures indicate that the p-value is low enough $(<0.05)$ to claim that this higher average is statistically significant.", + "bbox": [ + 507, + 381, + 882, + 439 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1 Expert Evaluation", + "text_level": 1, + "bbox": [ + 507, + 472, + 700, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For an expert evaluation, we recruited six faculty members and students from an academic English department. Figures 3 and 5 show that we strongly outperform PoeTryMe in all categories but theme with high statistical significance $(p < 0.006)$ , and we outperform Benhardt et al. in all poetic categories but theme and emotion with statistical significance $(p < 0.05)$ . Notably, while we outperform other computer-generated poems, respondents could still distinguish between our poems and human-written sonnets quite easily. See more in A.4.", + "bbox": [ + 505, + 495, + 882, + 671 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2 Amazon MTurk Evaluation", + "text_level": 1, + "bbox": [ + 507, + 687, + 771, + 701 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Along with expert evaluation, we used Amazon MTurk services to assess poems on a larger scale. Figures 4 and 6 show our superior performance against competitors in several categories. As expected of most computer-generated work, our poems failed to outperform human-written poems. However, we can only strongly conclude that the human-written poems are better in one category, theme. Our poems even outperformed human-written poems in grammar (albeit with low statistical significance), showing that our strictly constrained beam search generates high quality grammar. See more in A.5.", + "bbox": [ + 505, + 709, + 884, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1630", + "bbox": [ + 482, + 928, + 521, + 940 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8292ada2b3790048c4338774c85281dfbd1e018dce8050afac84f1d7d06c39fe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 137, + 102, + 245, + 158 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4b0713578ccefc477b272aaa9b622eef617b3bae4d06c5aa420cce41aa5e4519.jpg", + "image_caption": [ + "Expert Evaluation" + ], + "image_footnote": [], + "bbox": [ + 250, + 96, + 359, + 158 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/28af1fff4996f1e72162e0878094ff4e419f53aa295c6e318d5eb132fcb1ae0c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 102, + 463, + 158 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2c14d35ba10f132dcc548d4c42f4d72e869f9787f6ca1bda0d37409ed461e96d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 137, + 168, + 245, + 225 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6672ec6bca3259931c113a548b9d1eaa8cfc504a903f4734e70a043bfd93d0e4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 250, + 159, + 356, + 225 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c4297386128e901ffc60a136327959a030c1277cf1b17770a6468fdffd7c187d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 168, + 463, + 225 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/19ce2d9c4214eab638404a378bc61c99bf05a5f7ab00d09a673b02a997bfa246.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 137, + 235, + 245, + 292 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1cedeb4682139532a00ed47d6a618d784b432148d73e260bc38c191332ea4af9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 250, + 229, + 356, + 292 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/67679e7d43a4d23aa706b89624fcb0f5c0eac2d494b3ad71613a2aee7cd4420e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 236, + 463, + 292 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a6413b777f70c49d98b182b4faaadd35a417180f9006dc3c2fff64feae8c72de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 137, + 302, + 245, + 357 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e874758c6f482612209aa93d6c9a06fc26b5e577c3392b4b0746c8de4e4caf2c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 250, + 294, + 356, + 357 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ba3ca8ec4341772c19fce7920aa726f46ef05c17ba537fc692de309bceff122c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 303, + 463, + 357 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f64944439299e8d0a8d81b36cf7c9a0aa03fb92e7c74750ea93dca0a026941a9.jpg", + "image_caption": [ + "Figure 5: Values $>3$ (green), $< 3$ (red), and $= 3$ (gray) denote that our poetry model performs better, the competitor performs better, and the poems performed similarly, respectively." + ], + "image_footnote": [], + "bbox": [ + 137, + 370, + 245, + 425 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/09383ba0baf9877a9acf94df5644cf06a6304ed37969e444372aae77afb8a4a9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 250, + 361, + 356, + 425 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fe3563cea5094c2ee93e2b84e663513a968b99f5e4870193fdd3885f10d2e662.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 370, + 463, + 425 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/30ead587a2e210153662cc053b8634a4b29d455b2be031800d1023b8e3c90d9f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 532, + 102, + 642, + 158 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/194ef67c2c27e30692951db29732552a3e149e82029cf8563645d64756b08f36.jpg", + "image_caption": [ + "Amazon MTurk Evaluation" + ], + "image_footnote": [], + "bbox": [ + 645, + 95, + 751, + 158 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5efdcd6dc2ada6062c20b898c74d88db7eaf0134ce3dfab175fb82f9516fe683.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 102, + 858, + 158 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/df37ccfd094514f4c522b43dca9c98320896473e400b160c0f34cbb93b4da269.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 532, + 160, + 642, + 225 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ed4a3e6f4d49806be96b00477e9928dcd21683df6a77a912e8d2b8bc4c586a0f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 160, + 751, + 225 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/554e63d71ce9123238f0b7ef4b3628b775dffe027121633a688cc331294fa0b6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 160, + 858, + 225 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c595f5131a009d57178fe00e1f55a2d400ddc39c282914aae7aadd2de606b644.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 532, + 229, + 642, + 292 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fdee82545dc7ce52c6a9e180528600d93cefe993f4da71a5b653cb3e4a80e58a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 229, + 751, + 291 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cdfa9220bf60fd653a740d53f0910e146a9fdfd5919f7d1aed13c1cf83e54e45.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 229, + 858, + 292 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ee4d023da1c0d5f2b54320b99eb0bdffa062a22dcc635478ccd7bf69d1c522da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 532, + 302, + 642, + 357 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/848bb6ad99fc0f5fbe9ecbe895045891baecd155adf3f61f8b37f6ec2c865859.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 294, + 751, + 357 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b44abebd499c22ccd92aa5b2dbbb98141d4069c8299ce8d5ba38a6c621dd6e3f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 303, + 858, + 357 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e0c7a83428e11c556c4e5f31cd846719b98b0998b25f291eed5ba3f4270e83df.jpg", + "image_caption": [ + "Figure 6: Values $>3$ (green), $< 3$ (red), and $= 3$ (gray) denote that our poetry model performs better, the competitor performs better, and the poems performed similarly, respectively." + ], + "image_footnote": [], + "bbox": [ + 532, + 370, + 642, + 425 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ed0cb71915719919b394595e5793eeea15a98ace3b83d4297ae6fa2e56d37ea3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 361, + 751, + 425 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b16d15f4bb980d51422af2a12e796ad88f0a5adbaead5914bff4058e9e188ca1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 370, + 858, + 425 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3 Ablative Evaluation", + "text_level": 1, + "bbox": [ + 112, + 529, + 317, + 543 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We also conducted ablative studies showing the efficacy of two key elements of our method: line templates and the fine-tuned GPT-2 language model. We generated two sets of ablation poems: one with the fine-tuned GPT-2 and no templating, and one using the untrained GPT-2 model and templating. We then used Amazon MTurk services to test each set against poems generated with both factors under the same criteria as previous experiments. From Figure 11, it is the combination of the fine-tuned model and templating that ensures higher quality sonnets than if only one factor is implemented. Our poems with both factors outperform both sets of ablative poems with varying statistical significance. Specifically, providing templates is clearly the critical piece to generate poems of a high caliber. See more in A.6.", + "bbox": [ + 112, + 552, + 489, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 112, + 841, + 245, + 857 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose a novel method for generating high-quality poems that uses POS templating to determine a logical syntactical structure and rigorously", + "bbox": [ + 112, + 870, + 487, + 919 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "maintains constraints necessary for any sonnet. Our method is highly versatile, with poetic factors like alliteration, internal rhyme, repetition, and theme adjustable to ensure creative output. After extensive surveys conducted with expert evaluators and MTurk participants, our model's success over similar competitors is evident, though our model's poems, like those of most computer poetry generators, remain distinguishable from human written poems.", + "bbox": [ + 507, + 526, + 884, + 671 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While we were unable to compare our model's performance to that of ChatGPT, our finetuned GPT-2 requires far less computing power than subsequent GPT models. Additionally, while we commenced this project's evaluation prior to the release of ChatGPT, after a preliminary qualitative evaluation, ChatGPT seems to produce very generic poetry (see A.7). Thus, for this particular application, our model may be a viable method that is more cost-effective and produces relatively high-quality sonnets.", + "bbox": [ + 507, + 671, + 884, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Limitations", + "text_level": 1, + "bbox": [ + 509, + 860, + 613, + 876 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Though our method produces full sonnets that are more impressive than all previous approaches, it", + "bbox": [ + 507, + 887, + 882, + 919 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1631", + "bbox": [ + 482, + 928, + 517, + 940 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "is still not at the level of human-generated poetry. It is not clear how to achieve this level, whether it would be using massive large language models, or through our general approach, which is to bend those models around an interpretable framework that knows the rules that sonnets obey. Certainly our approach requires a lot less data – even if one used all the sonnets that have ever been written to train a language model, it is unclear that the language model would learn the very specific rules required of sonnets. However, there may be other ways to obtain these constraints that have not yet been developed.", + "bbox": [ + 112, + 84, + 489, + 294 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ethics Statement", + "text_level": 1, + "bbox": [ + 114, + 305, + 265, + 321 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As with all neural generation, there are concerns about misinformation and generating toxic text. These concerns apply to some degree to poetry generation, although our rigidly constrained approach and limited vocabulary should mitigate this.", + "bbox": [ + 112, + 331, + 490, + 411 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 114, + 438, + 213, + 453 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "John Benhardt, Peter Hase, Liuyi Zhu, and Cynthia Rudin. 2018. Shall I compare thee to a machine-written sonnet? An approach to algorithmic sonnet generation.", + "Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. 2017. Enriching word vectors with subword information. Transactions of the association for computational linguistics, 5:135-146.", + "Carnegie Mellon University CMU. 2019. The CMU pronouncing dictionary. http://www.speech.cs.cmu.edu/cgi-bin/cmudict, Internet.", + "Pablo Gervás. 2000. Wasp: Evaluation of different strategies for the automatic generation of spanish verse. In Proceedings of the AISB-00 Symposium on Creative & Cultural Aspects of AI, pages 93-100.", + "Marjan Ghazvininejad, Xing Shi, Yejin Choi, and Kevin Knight. 2016. Generating topical poetry. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1183-1191.", + "Jey Han Lau, Trevor Cohn, Timothy Baldwin, Julian Brooke, and Adam Hammond. 2018. Deep-spare: A joint neural model of poetic language, meter and rhyme. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1948-1958, Melbourne, Australia. Association for Computational Linguistics.", + "Ruli Manurung, Graeme Ritchie, and Henry Thompson. 2000. Towards a computational model of poetry generation. https://era.ed.ac.uk/handle/1842/3460." + ], + "bbox": [ + 114, + 460, + 489, + 917 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "George A Miller. 1998. WordNet: An electronic lexical database. MIT press.", + "Hugo Gonçalo Oliveira. 2012. Poetry: a versatile platform for poetry generation. Computational Creativity, Concept Invention, and General Intelligence, 1:21.", + "Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners. https://github.com/openai/gpt-2.", + "Tim Van de Cruys. 2020. Automatic poetry generation from prosaic text. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2471-2480, Online. Association for Computational Linguistics.", + "Tony Veale. 2013. Less rhyme, more reason: Knowledge-based poetry generation with feeling, insight and wit. In Proceedings of the Fourth International Conference on Computational Creativity, ICCC 2013, Sidney, Australia, June 12-14, 2013, pages 152-159. computationalcreativity.net.", + "Jianyou Wang, Xiaoxuan Zhang, Yuren Zhou, Christopher Suh, and Cynthia Rudin. 2021. There once was a really bad poet, it was automated but you didn't know it." + ], + "bbox": [ + 509, + 85, + 884, + 463 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A Appendix", + "text_level": 1, + "bbox": [ + 509, + 488, + 633, + 504 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A.1 Templating Mechanism", + "text_level": 1, + "bbox": [ + 509, + 513, + 742, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 8 presents more examples of our templating mechanism. We combine an adapted version of the Penn Treebank Project's part of speech tags along with articles, conjunctions, prepositions, and other filler words to construct these templates. Additionally, we provide the stress pattern of the syllables to ensure that the constraint of iambic pentameter is met. However, outside of the pre-determined filler words, POS do not have to directly adhere to the given stress pattern in splitting up words. For instance, in the first template, the provided syllable stress indicates that the JJ tag (adjective) should have two syllables, while the final VB tag (verb) should have only one syllable. However, the generated line ends with a monosyllabic adjective and a bisyllabic verb. As long as the stressing of the syllables aligns properly, each word can vary in its number of syllables. This is also visible in the fourth template example in Figure 8.", + "bbox": [ + 505, + 533, + 882, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A.2 Elaboration on Experimental Competitors", + "text_level": 1, + "bbox": [ + 509, + 850, + 789, + 881 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Benhardt et al. (2018), referred to as Benhardt et al., uses a RNN to preselect rhyming words and", + "bbox": [ + 507, + 887, + 882, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1632", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "restrict different parts of speech to fit within the sonnet format. Oliveira (2012), referred to as Co-PoetryMe, is a versatile platform using semantic and grammar templates to alter the type of poem, input words, and \"surprise\" factor generated.", + "bbox": [ + 112, + 84, + 489, + 165 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A.3 Experimental Procedure", + "text_level": 1, + "bbox": [ + 112, + 175, + 357, + 190 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For each pair of sonnets, respondents were asked to indicate whether Sonnet A or Sonnet B performed better based on factors such as adherence to the inputted theme, poeticness, grammatical correctness, ability to convey emotion, and likelihood of being written by a human. Available answer choices and their corresponding numeric scores from 1 to 5 were \"Definitely A\" (5), \"Probably A\" (4), \"The same\" (3), \"Probably B\" (2), and \"Definitely B\" (1). Both our sonnet and the competing model-human-sonnet had equal probability of being either sonnet A or sonnet B in each pair. To analyze this data, user inputs were translated into numeric scoring values corresponding to our model's sonnet being Sonnet A (i.e. if our sonnet is presented as B to the user, a response of \"Definitely B\" corresponds to a score of 5, \"Probably B\" corresponds to 4, \"Probably A\" corresponds to 2, and \"Definitely A\" corresponds to 1). Additionally, respondents were asked to answer sanity check questions to filter out respondents who answer illogically or who do not have a sufficient grasp of English grammar. This setup remained the same across all experiments, and an additional space was allocated for expert evaluators to leave qualitative comments on sonnet quality. Sample sonnet evaluation questions are visible in Figure 9.", + "bbox": [ + 115, + 195, + 489, + 629 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "After calculating the mean and standard deviation for scores across sonnets, we can immediately see whether our model performed better (an average score of $>3$ ) or worse (an average score of $< 3$ ) than the competitor in each respective category. We then performed a series of t-tests to establish these results' statistical significance. For factors that indicated our model performed better, we performed a right-tailed t-test (with the null-hypothesis as our model performed worse than the baseline), and we performed a left-tailed t-test for the remaining factors (with the null-hypothesis as our model performed better than the baseline).", + "bbox": [ + 112, + 631, + 489, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A.4 Expert Evaluation Analysis", + "text_level": 1, + "bbox": [ + 112, + 850, + 379, + 865 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the expert evaluation, we emailed faculty at an American academic English department to recruit six faculty members and students to take our survey", + "bbox": [ + 112, + 871, + 487, + 919 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "without payment. While we showed strong performance against the other computer-generated poems, we are consistently outperformed by human-written poems in all categories. Weaker performance on theme in experimental results may be explained by competitors' more frequent inclusion of the user-inputted theme word. For instance, in the expert evaluation, between two poems generated with the theme word \"forest\" (see Figure 10), one survey respondent states, \"Sonnet B repeats forest too much for my taste,\" subsequently giving our model a 5 in each of poeticness, grammar, emotion, and humanness, yet a 2 in theme.", + "bbox": [ + 507, + 84, + 884, + 294 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A.5 Amazon MTurk Analysis", + "text_level": 1, + "bbox": [ + 507, + 304, + 757, + 319 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In our evaluation using Amazon MTurk Services, we requested survey respondents from primarily English-speaking countries and with an approval rate of $\\geq 95\\%$ . Crowdworkers were paid through the Amazon MTurk platform for this survey that on average took less than 30 minutes to complete. The questions and formatting remained the same as the expert evaluation, except no space was provided for qualitative feedback.", + "bbox": [ + 507, + 324, + 882, + 468 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Based on Figure 4 there is enough statistical significance to conclude that our sonnets outperform PoeTryMe in poetic, grammar, emotion, and human categories $(p < 0.001)$ . Against Benhardt et al., there is enough statistical significance to conclude that our sonnets perform better in grammar $(p < 0.001)$ , and perform slightly better with weak statistical significance in emotion $(p < 0.15)$ . Against human-written sonnets, the p-values for poetic, emotion, and even human categories are too large to strongly reject the null hypothesis that our model performed better than the baseline. Additionally, while the p-value indicates that this value is not statistically significant, it is interesting to note that our poems on average scored better in the grammar category.", + "bbox": [ + 507, + 469, + 882, + 727 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A.6 Ablation Analysis", + "text_level": 1, + "bbox": [ + 507, + 738, + 700, + 753 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In our ablation analysis, we replicate the Amazon MTurk analysis yet replace the competitor/human-written sonnets with poems generated with either the fine-tuned GPT-2 model without templating or the GPT-2 model without fine-tuning and with templating. This lets us test the individual efficacy of each factor (templating and fine-tuning GPT-2) against our method implementing both. Against poems generated with the fine-tuned GPT-2 and no templating, our sonnets performed better across", + "bbox": [ + 507, + 758, + 882, + 917 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1633", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "all categories, and we can strongly reject the null hypothesis that our model performed worse than the baseline $(p < 0.0001)$ . Against the poems generated with the GPT-2 model without fine-tuning and with templates, we can conclude with high statistical significance $(p < 0.01)$ that we performed better in emotion, and conclude with weak statistical significance $(p < 0.10)$ that we performed better in grammar and theme. These results indicate that our method is successful due to its usage of both the fine-tuned GPT-2 model and templating.", + "bbox": [ + 112, + 84, + 492, + 263 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A.7 ChatGPT Qualitative Analysis", + "text_level": 1, + "bbox": [ + 112, + 272, + 405, + 288 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While we did not have time to extensively evaluate the quality of our sonnets against those of ChatGPT, after generating several sonnets to test ChatGPT's sonnet quality, it seems as though this language model generates relatively generic, non-cohesive sonnets even with different parameters. For instance, in Figure 7, both of the sonnets are unable to cohesively connect these three topics along a reasonable storyline. Additionally, Sonnet A in particular seems to dedicate a single stanza to each of the three topics passed in, hardly attempting to connect them. Of course, with more intensive prompt engineering, it is possible to generate a sonnet more tailored to one's preference. However, even this short analysis demonstrates there are clearly still strides to be made in the field of automatic poetry generation even with the advent of ChatGPT.", + "bbox": [ + 112, + 292, + 489, + 567 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2e315f6fafb9c7ab2d9ad4a031ce60500873303e83e2bbb889e10f73f9adddb2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Sonnet A
In realms of words, where novels find their reign, My friend, a skeptic, shuns the written tales. No pages turned, no fictional campaigns, Averse to narratives that time unveils. Yet in his hand, a pear's sweet succulence, A juicy feast, a taste of nature's grace. Each bite, a moment of pure indulgence, A salute's joy, a respite from life's race. But beyond fruit's delight, his spirit soars, Within the bounds of a duelist's domain. With sword in hand, he dances, parries, scores, A noble art where honor's stance remains. Though novels may elude his curious gaze, Through pears and fencing, passions find their ways.
Sonnet B
In realms of love, where rom-coms dance their waltz, My friend, a rebel, strays from silver screens. With flannel-clad allure, he stands, exalts, A soul untouched by scripted love's routines. No tears shed for the tales of starry-eyed, His heart resists the charms of love's façade. Yet stitches formed with yarn, his hands provide, Creating warmth, affection neatly clad. Through threads of patience, vibrant patterns bloom, His nimble fingers weave a tapestry. Each crochet stitch, a tale of friendship's loom, A testament to artistry set free. Though rom-coms fall short of his embrace, In flannel and crochet, love finds its place.
", + "bbox": [ + 509, + 265, + 858, + 651 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 7: Comparison of two sonnets generated with ChatGPT. Sonnet A was generated with the prompt \"generate a sonnet about my friend who hates novels, eats pears, and fences\" and Sonnet B was generated with the prompt \"generate a sonnet about my friend who hates rom-coms, wears flannels, and crochets.\"", + "bbox": [ + 507, + 661, + 884, + 746 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1634", + "bbox": [ + 482, + 928, + 521, + 940 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/315449d4de0e49013f7297e3f2fb3906310f69239862ccd39b27b8d80c581741.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TemplateSyllable StressExample Line
Where all the NNS of PRPD$ JJ NNS VB.0 1 0 1 0 1 01 0 1“Where all the gods of their past lives dictate”
And it VBD ABNN to the NN0 1 0 10 1 0 101“And it seemed evil to the enterprise”
Between the VBG and the VBG NN01 0 10 1 0 10 1“Between the glistening and the dying muse”
A JJ NN from the JJ NN0 10 10 1 0 1 01“A little lightness from the earthy sky”
Upon PRPO, PRPD$ NN POS NN01 01 0 10 101“Upon you, your life’s possibility”
Why VBC PRPS VBG such a JJ NN?0 1 0 10 1 0 101 0"Why do you squander such a precious thing?"
The NNS of ABNN, the NN on the NN0 1 0 1 0 10 1 0 1“The ghosts of death, the spirit on the earth”
", + "bbox": [ + 139, + 80, + 857, + 195 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The key word for both of these poems is \"wisdom.\" Which poem best adheres to this theme?", + "bbox": [ + 144, + 441, + 371, + 449 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/7fe3ee3ee799d37cace046d232e24d36af34824c2cc8134204459501a5e5117d.jpg", + "table_caption": [ + "Figure 8: Template examples, their corresponding syllable stress in order to adhere to iambic pentameter, and a sample line generated using the template." + ], + "table_footnote": [], + "table_body": "
Definitely AProbably ASameProbably BDefinitely B
", + "bbox": [ + 149, + 458, + 450, + 476 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Which poem sounds more poetic?", + "bbox": [ + 144, + 492, + 231, + 499 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/d32141c3e26af9c84542b1fe3fd9b70a1d5903d704265eaeb488ce754c6558a8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Definitely AProbably ASameProbably BDefinitely B
", + "bbox": [ + 157, + 510, + 440, + 521 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Which poem is more grammatically correct?", + "bbox": [ + 144, + 541, + 253, + 548 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/23cc6a37e41323fa9f3c6e8ed5d9efa1e491cbb30a6a5e12ae2d08038c9cbabf.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Definitely AProbably ASameProbably BDefinitely B
", + "bbox": [ + 157, + 567, + 440, + 583 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Which poem conveys emotions more effectively?", + "bbox": [ + 144, + 600, + 265, + 607 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/3d755f5125ffb56b57e99bc8d7195c05fa590d47d5926e2435cede181ae19b3e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Definitely AProbably ASameProbably BDefinitely B
", + "bbox": [ + 157, + 619, + 440, + 629 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Which poem is more likely to be written by a human?", + "bbox": [ + 144, + 649, + 275, + 656 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/c1ffddf7888c695b0eb4589c5cfbee37ab2752d93385be4007219ae0ef326c86.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Definitely AProbably ASameProbably BDefinitely B
", + "bbox": [ + 157, + 669, + 440, + 678 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Figure 9: Survey questions presented for each pair of sonnets.", + "bbox": [ + 112, + 703, + 487, + 730 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/5e99d87fd8a2731cc47324f91709d6941f9184a33654410019d9b91762c1a965.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Sonnet A: Our Code
I was aghast to see the fireflies
Inflamed soothed toads, where there the dead boughs lay
And it seemed evil to the enterprise
The hag I had, the hag, the hog, the gray.
But I knew to my painless fireflies
And beauty was a kind and loving thing.
My life's light isle so longed on otherwise
So too my fireflies bloomed to my king.
Those eagles that with auburn hair flew oaks,
Beauty and beauty beamed within the air
Which made oasis overcomes to coax?
So too my hogs beheaded to my lair.
The windy night was in the mistletoe
And wept soiled toads in my dream's studio.
Sonnet B: PoetryMe
forest some more and reforest a trip!
in deserts where heavenly woodlands clink
many, many, many clustered before
come: not in establishments of the floor
the fields of agony, the endless circumstance
findings to lie to interrupt your earth
with summation and set, triumph and agony
floors of horror forest before my eyes
those that study clustered plant are psychologists
taking over my ness a second forest
an' you've got to forest them reforest
on every forest, indeed, that rainforests
and grounds of forest coming to accord
floor of establishments and lilt of sing
", + "bbox": [ + 510, + 376, + 877, + 762 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Figure 10: Comparison of two sonnets generated with theme word \"forest\". Sonnet A was generated with our code, and Sonnet B was generated using PoeTryMe.", + "bbox": [ + 507, + 772, + 882, + 815 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1635", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/9c872bae3aeec2b3e2b9fcf05095ceaad81a5540ab4f279bd4353a50a4fea287.jpg", + "table_caption": [ + "Ablation Evaluation" + ], + "table_footnote": [], + "table_body": "
CategoryMeanp-valueMeanp-value
Grammar3.51*5.10×10-53.21*0.06
Emotion3.61*9.00×10-63.40*3.89×10-3
Poetic3.61*4.00×10-63.09*0.29
Human3.66*1.00×10-63.01*0.46
Theme3.50*8.00×10-53.20*0.06
", + "bbox": [ + 122, + 406, + 480, + 499 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Figure 11: Left: fine-tuned GPT-2 with no templates. Right: GPT-2 without fine-tuning, but with templates. Starred figures indicate average scores of $>3$ , and underlined figures indicate that the p-value is low enough $(<0.05)$ to claim that this higher average is statistically significant.", + "bbox": [ + 112, + 514, + 489, + 602 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1636", + "bbox": [ + 482, + 928, + 521, + 940 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "A For every submission:", + "bbox": [ + 115, + 107, + 322, + 122 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "A1. Did you describe the limitations of your work? Limitations", + "A2. Did you discuss any potential risks of your work? Ethics", + "A3. Do the abstract and introduction summarize the paper's main claims? Abstract, 1", + "A4. Have you used AI writing assistants when working on this paper? Left blank." + ], + "bbox": [ + 129, + 126, + 695, + 287 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "B Did you use or create scientific artifacts?", + "bbox": [ + 114, + 299, + 487, + 315 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "3,4", + "bbox": [ + 134, + 321, + 161, + 334 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "B1. Did you cite the creators of artifacts you used? 3.2,3.3,References", + "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Not applicable. Left blank.", + "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Not applicable. Left blank.", + "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Data used from publicly available sonnets/ poems were assumed to be not subject to dispute.", + "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Not applicable. Left blank.", + "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. 3.3" + ], + "bbox": [ + 127, + 346, + 880, + 752 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "C Did you run computational experiments?", + "bbox": [ + 114, + 765, + 492, + 781 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 132, + 787, + 213, + 801 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? No response.", + "bbox": [ + 129, + 813, + 880, + 860 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "ACL 2023 Responsible NLP Checklist", + "bbox": [ + 132, + 84, + 433, + 99 + ], + "page_idx": 10 + }, + { + "type": "page_footnote", + "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.", + "bbox": [ + 112, + 868, + 877, + 892 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1637", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? No response.", + "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? No response.", + "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? No response." + ], + "bbox": [ + 127, + 84, + 880, + 281 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? 4,4.1,4.2,4.3", + "bbox": [ + 112, + 292, + 877, + 330 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Appendix", + "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? A.5.A.6", + "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? We do not believe having data on poetry evaluation raises any ethical issues.", + "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? We do not believe having crowdworkers evaluate the same poems that were given to English professors raises any ethical issues.", + "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? A.6" + ], + "bbox": [ + 129, + 338, + 880, + 653 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "1638", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/The Mechanical Bard_ An Interpretable Machine Learning Approach to Shakespearean Sonnet Generation/3330f90d-0cb8-4666-aeb2-f9f31ccb9534_model.json b/2023/The Mechanical Bard_ An Interpretable Machine Learning Approach to Shakespearean Sonnet Generation/3330f90d-0cb8-4666-aeb2-f9f31ccb9534_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fd6d46fed987913cfb23150652507f1de036b38b --- /dev/null +++ b/2023/The Mechanical Bard_ An Interpretable Machine Learning Approach to Shakespearean Sonnet Generation/3330f90d-0cb8-4666-aeb2-f9f31ccb9534_model.json @@ -0,0 +1,2490 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.126, + 0.09, + 0.873, + 0.131 + ], + "angle": 0, + "content": "The Mechanical Bard: An Interpretable Machine Learning Approach to Shakespearean Sonnet Generation" + }, + { + "type": "text", + "bbox": [ + 0.188, + 0.15, + 0.816, + 0.167 + ], + "angle": 0, + "content": "Edwin Agnew*, Michelle Qiu*, Lily Zhu*, Sam Wiseman, Cynthia Rudin" + }, + { + "type": "text", + "bbox": [ + 0.376, + 0.168, + 0.627, + 0.183 + ], + "angle": 0, + "content": "Duke University, Durham, NC" + }, + { + "type": "text", + "bbox": [ + 0.188, + 0.184, + 0.816, + 0.2 + ], + "angle": 0, + "content": "edwin.agnew@duke.edu, michelle.qiu@duke.edu, lily.zhu@duke.edu" + }, + { + "type": "text", + "bbox": [ + 0.295, + 0.201, + 0.71, + 0.216 + ], + "angle": 0, + "content": "swiseman@cs.duke.edu, cynthiaa@cs.duke.edu" + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.253, + 0.341, + 0.268 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.281, + 0.462, + 0.466 + ], + "angle": 0, + "content": "We consider the automated generation of sonnets, a poetic form constrained according to meter, rhyme scheme, and length. Sonnets generally also use rhetorical figures, expressive language, and a consistent theme or narrative. Our constrained decoding approach allows for the generation of sonnets within preset poetic constraints, while using a relatively modest neural backbone. Human evaluation confirms that our approach produces Shakespearean sonnets that resemble human-authored sonnets, and which adhere to the genre's defined constraints and contain lyrical language and literary devices." + }, + { + "type": "text", + "bbox": [ + 0.567, + 0.274, + 0.828, + 0.287 + ], + "angle": 0, + "content": "When all the lovers of this world are dead," + }, + { + "type": "text", + "bbox": [ + 0.591, + 0.287, + 0.804, + 0.298 + ], + "angle": 0, + "content": "The sun of heaven on a golden day" + }, + { + "type": "text", + "bbox": [ + 0.554, + 0.299, + 0.84, + 0.31 + ], + "angle": 0, + "content": "To burn the earth's fire by the flame and spread" + }, + { + "type": "text", + "bbox": [ + 0.566, + 0.311, + 0.827, + 0.322 + ], + "angle": 0, + "content": "Where all the flowers of your fair days lay." + }, + { + "type": "text", + "bbox": [ + 0.561, + 0.322, + 0.833, + 0.333 + ], + "angle": 0, + "content": "These are the blossoms that you take care of." + }, + { + "type": "text", + "bbox": [ + 0.58, + 0.334, + 0.813, + 0.345 + ], + "angle": 0, + "content": "Why do you linger such a long delay?" + }, + { + "type": "text", + "bbox": [ + 0.572, + 0.346, + 0.823, + 0.357 + ], + "angle": 0, + "content": "Forgive the fluttered flower of meek love" + }, + { + "type": "text", + "bbox": [ + 0.571, + 0.358, + 0.823, + 0.369 + ], + "angle": 0, + "content": "Or who you have so long to love the day?" + }, + { + "type": "text", + "bbox": [ + 0.575, + 0.37, + 0.819, + 0.381 + ], + "angle": 0, + "content": "The joys of love, the beauty on the face," + }, + { + "type": "text", + "bbox": [ + 0.566, + 0.382, + 0.827, + 0.393 + ], + "angle": 0, + "content": "Shall be your fate and be your own delight." + }, + { + "type": "text", + "bbox": [ + 0.567, + 0.394, + 0.826, + 0.404 + ], + "angle": 0, + "content": "You have the beauty of your own embrace." + }, + { + "type": "text", + "bbox": [ + 0.575, + 0.405, + 0.819, + 0.416 + ], + "angle": 0, + "content": "You cannot reminiscence. Cannot write." + }, + { + "type": "text", + "bbox": [ + 0.56, + 0.417, + 0.833, + 0.428 + ], + "angle": 0, + "content": "Between the living and the deadening breath." + }, + { + "type": "text", + "bbox": [ + 0.579, + 0.429, + 0.814, + 0.44 + ], + "angle": 0, + "content": "You go the way of your beloved death." + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.458, + 0.876, + 0.473 + ], + "angle": 0, + "content": "Figure 1: A sonnet generated with the theme \"death\"." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.491, + 0.26, + 0.507 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.517, + 0.49, + 0.629 + ], + "angle": 0, + "content": "We consider the task of automatically generating Shakespearean sonnets, a popular poetic form with highly specific rhyme and meter constraints1. Each sonnet consists of three quatrains followed by a single couplet according to the rhyme scheme ABAB BCBC CDCD EE, and each line contains ten syllables with a stress pattern of iambic pentameter." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.631, + 0.49, + 0.71 + ], + "angle": 0, + "content": "Rather than train a model to obey these constraints implicitly (which leads to enormous models that still do not obey the constraints), we opt to enforce them explicitly using a simple but novel approach to generation." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.711, + 0.49, + 0.871 + ], + "angle": 0, + "content": "In particular, we use part-of-speech (POS) templates selected and edited from individual lines in Shakespeare's sonnets, with each template intended to offer a different combination of parts of speech and narrative directions. Associated thematic words are then selected and placed at the end of each template, and their rhyming pairs are chosen dynamically by a language model (e.g., GPT-2, Radford et al., 2019) and placed at the end of the corresponding lines according to the rhyme scheme." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.506, + 0.883, + 0.554 + ], + "angle": 0, + "content": "The rest of the line is filled with related words that fit the specified POS and meter, leading to the end rhyme word. Figure 1 shows sample output." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.556, + 0.885, + 0.652 + ], + "angle": 0, + "content": "Our use of these templates ensures sophisticated-seeming language and syntax that competing systems do not capture. Our approach provides excellent grammatical structure comparable to that of human-written poetry, all while using a relatively simple model and generation procedure." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.653, + 0.884, + 0.845 + ], + "angle": 0, + "content": "We extensively evaluate the ability of our approach to generate whole sonnets (a setting often ignored by recent work in poetry generation) and find that our approach is preferred over strong baselines by both expert annotators (recruited from an academic English department) and by crowdworkers. As this research was conducted before the release of ChatGPT, we were not able to robustly compare our model's performance against this language model. However, we make several observations about the poetic quality of sonnets generated by ChatGPT." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.86, + 0.667, + 0.875 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.888, + 0.884, + 0.919 + ], + "angle": 0, + "content": "Early attempts at poetry generation relied mainly on rule-based methods (Gervás, 2000; Oliveira," + }, + { + "type": "page_footnote", + "bbox": [ + 0.137, + 0.88, + 0.306, + 0.894 + ], + "angle": 0, + "content": "* denotes equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.137, + 0.894, + 0.487, + 0.907 + ], + "angle": 0, + "content": "1Our code is available at https://github.com/" + }, + { + "type": "page_footnote", + "bbox": [ + 0.117, + 0.907, + 0.307, + 0.918 + ], + "angle": 0, + "content": "edwinagnew/Poetix_Sonnets" + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.88, + 0.487, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1627" + }, + { + "type": "footer", + "bbox": [ + 0.227, + 0.946, + 0.771, + 0.959 + ], + "angle": 0, + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + }, + { + "type": "footer", + "bbox": [ + 0.369, + 0.96, + 0.63, + 0.972 + ], + "angle": 0, + "content": "Volume 2: Short Papers, pages 1627-1638" + }, + { + "type": "footer", + "bbox": [ + 0.297, + 0.973, + 0.702, + 0.986 + ], + "angle": 0, + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.112, + 0.085, + 0.49, + 0.358 + ], + "angle": 0, + "content": "2012; Manurung et al., 2000; Veale, 2013). More recent automated poetry generation techniques, especially for sonnet generation, have relied on combinations of task-specific language models and rules. For instance, Ghazvininejad et al. (2016)'s Hafez uses a finite state acceptor to generate a large number of possible lines, the best of which are then selected with an RNN trained on song lyrics. Like our approach, they use rhyming dictionaries to find rhyming words and word embeddings to find topical words. Similarly, Benhardt et al. (2018) preselects rhyming words and generates lines backwards with a recurrent neural network (RNN). Also in this vein are Lau et al. (2018)'s Deepspare, which consists of an LSTM language model, an iambic model, and a rhyming model, and the recent work of Van de Cruys (2020) and Wang et al. (2021)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.36, + 0.49, + 0.442 + ], + "angle": 0, + "content": "Our approach distinguishes itself in using a general-purpose pretrained language model, but more importantly in its use of human-curated constraints and templates. These allow for generating high-quality poems with a very simple approach." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.454, + 0.265, + 0.47 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.482, + 0.491, + 0.626 + ], + "angle": 0, + "content": "The general idea of our approach is to take a pretrained language model (in this case GPT-2) and apply hard constraints to the generation procedure so that it can only output text satisfying various poetic constraints. These constraints can be broadly divided into hard constraints (e.g., number of syllables) and soft constraints (e.g., sounding poetic), and our methodology can be separated similarly. Our generation process is in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.64, + 0.283, + 0.655 + ], + "angle": 0, + "content": "3.1 POS Templates" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.662, + 0.49, + 0.919 + ], + "angle": 0, + "content": "The most important part of our method is the use of handcrafted grammar templates. Taking inspiration from existing sonnets, we created a list of about 120 templates that encode the part-of-speech structure of a line of poetry. Each template can generate an unbounded number of possible poetic lines. For example, the line \"The beauty of life on a lonely sea\" is represented by the template \"THE NN OF NN ON A JJ NN.\" More sample templates are in Section A.1. Since the templates allow for considerable flexibility, obeying the templates does not alone suffice for poetry. For example, the same template could be used to write poetic lines with distinct meanings such as \"The tree of anguish on a stormy night\" or a nonsensical line like \"The fork of ant on an unpacked transfer.\" A subset of these" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.085, + 0.846, + 0.101 + ], + "angle": 0, + "content": "templates is also chosen for starting a stanza." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.112, + 0.757, + 0.126 + ], + "angle": 0, + "content": "3.2 Strict Sonnet Constraints" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.133, + 0.885, + 0.341 + ], + "angle": 0, + "content": "The two most critical features of sonnets distinguishing them from other poetry forms are that they are written in iambic pentameter (i.e., each line has 10 syllables of alternating stress pattern), and they follow an ABAB CDCD EFEF GG rhyme scheme. To detect iambic pentameter, we use the CMU Pronouncing Dictionary (CMU, 2019), which reveals how many syllables a word contains and the stress of each syllable. An unstressed syllable is represented as '0' and a stressed syllable as '1', and so the line \"The beauty of life on a lonely sea\" is represented as '0 10 1 0 1 0 10 1'. For simplicity, 1-syllable words can be designated as either 0 or 1." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.342, + 0.886, + 0.536 + ], + "angle": 0, + "content": "Given a POS-tag for every word in our dictionary, we create a tree-like data structure that represents every possible meter for a given template. Continuing the example, the first word could only be 'the', but the second word could be filled with a 1-syllable noun like 'tree', a 2-syllable noun like 'chaos' (10), or a 3-syllable noun like 'audio' (101), and so on. Each choice affects the possible pronunciations of the next word as well as the number of remaining words in order to reach 10 syllables. The pronunciation dictionary ensures the last syllable of the last word on each line matches its partner." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.546, + 0.691, + 0.561 + ], + "angle": 0, + "content": "3.3 Language Model" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.567, + 0.885, + 0.807 + ], + "angle": 0, + "content": "We use a language model to generate individual sonnet lines, subject to the formal constraints outlined above. In particular, we first fine-tune GPT-2 (Radford et al., 2019) on a large corpus of over 15000 poems and a smaller corpus of sonnets. We then use a constrained beam-search to generate each line, where only legal tokens (under the aforementioned constraints) can be generated at each step; this generation approach resembles previous constrained decoding techniques used in sonnet generation (Ghazvininejad et al., 2016), although our approach differs in the choice of model and direct enforcement of constraints. For a comparison of generation quality using a GPT-2 model that has not been fine-tuned, see Section 4.1." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.819, + 0.742, + 0.833 + ], + "angle": 0, + "content": "3.4 Thematic Word Choice" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.84, + 0.884, + 0.872 + ], + "angle": 0, + "content": "To ensure the content of the poem fits the theme specified by the user, we provide an excerpt of a" + }, + { + "type": "page_footnote", + "bbox": [ + 0.509, + 0.88, + 0.884, + 0.905 + ], + "angle": 0, + "content": "\\(^{2}\\)https://www.kaggle.com/datasets/johnhallman/completpoetryfoundation.org-dataset" + }, + { + "type": "page_footnote", + "bbox": [ + 0.534, + 0.906, + 0.866, + 0.919 + ], + "angle": 0, + "content": "3https://www.kaggle.com/datasets/michelleqiu/sonnets" + }, + { + "type": "list", + "bbox": [ + 0.509, + 0.88, + 0.884, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1628" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.202, + 0.082, + 0.402, + 0.095 + ], + "angle": 0, + "content": "Generation Visualization" + }, + { + "type": "image", + "bbox": [ + 0.14, + 0.097, + 0.462, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.456, + 0.488, + 0.484 + ], + "angle": 0, + "content": "Figure 2: Numbers in parentheses denote subsections in Section 3." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.513, + 0.49, + 0.691 + ], + "angle": 0, + "content": "theme-appropriate poem as additional context to GPT-2 during generation. This additional poem is selected by finding a list of synonyms to the theme word using the WordNet synonym database (Miller, 1998) and then choosing lines from a poem corpus that contain at least one synonym. We also remove words from the vocabulary if they have less than 0.5 cosine similarity with the theme word, based on the corresponding FastText word embeddings (Bojanowski et al., 2017). This avoids having words like \"algebra\" in poems with themes like \"forest.\"" + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.704, + 0.336, + 0.718 + ], + "angle": 0, + "content": "3.5 Generation Procedure" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.727, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Having introduced our method's components, we now describe the generation procedure. A user inputs a theme word, a beam search parameter, \\( b \\), and the number of templates sampled per line, \\( k \\). A seed is chosen with the above method. Then for each line, we sample \\( k \\) random templates. For each template, we generate the line using a modified beam search. Specifically, the beam search maintains \\( b \\) different hypotheses per line at all times. For each hypothesis, we first mask out any tokens that violate our hard POS, meter, or rhyme constraints and select the \\( b \\) best next-tokens for each" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.262 + ], + "angle": 0, + "content": "of the \\(k\\) templates. These \\(b^{2}\\) new candidates are re-ranked according to our custom scoring function, and the top \\(k \\times b\\) proceed to the next stage. The constraint-filtering at each stage guarantees that the generated line will match the input template, while the beam search allows more flexible word choice than greedy word-filling for each POS. If none of the \\(k \\times b\\) generated lines score better than a specific threshold, then a new template is chosen and the line is generated again. Otherwise, line generation continues until the poem is completed." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.275, + 0.669, + 0.289 + ], + "angle": 0, + "content": "3.6 Poetic Devices" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.298, + 0.885, + 0.715 + ], + "angle": 0, + "content": "To make the poems more poetic, we adjust our scoring function to weight lines with alliteration, penalties for repetition, and/or internal rhyme. Alliteration occurs when a line contains words starting with the same letter, repetition occurs when a word is present several times throughout a poem, and internal rhyme occurs when two words rhyme within the same line. To weight alliteration, when the first token of a new word is being generated, a list \\(\\vec{A} = [a_1,a_2,\\dots a_n]\\) is generated where \\(a_{i}\\) is the number of occurrences of the first letter of the ith token in the current line. To weight and discourage repetition, a list \\(\\vec{T} = [t_1,t_2,\\dots t_n]\\) is generated where \\(t_i\\) is the number of occurrences of the ith token in the poem, negated. To weight internal rhyme, a list \\(\\vec{R} = [r_1,r_2,\\dots ,r_n]\\) is generated where \\(r_i = 1\\) if the ith token is part of a word that rhymes with any of the words in the current line generated so far, and \\(r_i = 0\\) otherwise. The final token distribution is then proportional to \\(\\tilde{P} +\\alpha_{A}\\times \\vec{A} +\\alpha_{T}\\times \\vec{T} +\\alpha_{R}\\times \\vec{R},\\) where \\(\\tilde{P}\\) is the language model's next-token distribution, and \\(\\alpha_{A},\\alpha_{T},\\) and \\(\\alpha_{R}\\) are user-specified non-negative parameters, which represent the degree to which alliteration, repetition, and internal rhyme should be favored during generation." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.728, + 0.673, + 0.744 + ], + "angle": 0, + "content": "3.7 Postprocessing" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.75, + 0.885, + 0.831 + ], + "angle": 0, + "content": "After a poem is completed and all 14 lines score above a fixed threshold, a small number of adjustments are made. These include fixing common mistakes made by GPT-2 like not capitalizing the word 'I' and not capitalizing following punctuation." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.845, + 0.656, + 0.861 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.872, + 0.885, + 0.919 + ], + "angle": 0, + "content": "We used human input to test our sonnets against both model-generated and human-written sonnets. To test adherence to a theme throughout a son" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1629" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.191, + 0.081, + 0.411, + 0.379 + ], + "angle": 0, + "content": "
CategoryMeanp-value
PoeTryMe
Grammar4.50*1.71×10-4
Emotion4.30*3.13×10-3
Poetic4.30*3.13×10-3
Human4.10*5.77×10-3
Theme2.600.211286
Benhardt et al.
Grammar3.83*0.03
Emotion3.67*0.05
Poetic3.75*0.04
Human3.75*0.02
Theme2.420.06
Human-written poems
Grammar1.361.00×10-6
Emotion1.45.00×10-6
Poetic1.645.40×10-5
Human1.361.00×10-6
Theme1.577.70×10-5
" + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.395, + 0.49, + 0.454 + ], + "angle": 0, + "content": "Figure 3: Starred figures indicate average scores of \\(>3\\), and underlined figures indicate that the p-value is low enough \\((<0.05)\\) to claim that this higher average is statistically significant." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.481, + 0.489, + 0.576 + ], + "angle": 0, + "content": "net, we desired baselines that generate whole sonnets with user-provided themes. This limited our competitors, as some generate discrete quatrains or generate without input themes (e.g., Deepspare), leaving only Benhardt et al. (2018) and PoeTryMe (Oliveira, 2012); see Section A.2." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.579, + 0.489, + 0.658 + ], + "angle": 0, + "content": "Furthermore, an evaluation of poetry quality is incomplete without human-written sonnets, selected from sonnets.org. Though these poems do not have an explicit theme, we selected poems that followed our five themes." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.661, + 0.489, + 0.789 + ], + "angle": 0, + "content": "To optimally test our model, we conducted an internal analysis and selected \\( k \\) values sampled from 3, 5, or 7, \\( b \\) values sampled from 3, 5, or 7, and repetition penalty values sampled from 1.4, 1.6, or 1.8 that we concluded produced the highest quality sonnets. To evaluate adherence to theme, we generated poems with themes \"death,\" \"darkness,\" \"forest,\" \"love,\" and \"wisdom.\"" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.791, + 0.489, + 0.919 + ], + "angle": 0, + "content": "In each test, respondents compared six randomly selected pairs of sonnets, with each of our sonnets displayed with a competing model/human-written sonnet generated with the same theme word. Respondents indicated which of the two sonnets performed better in categories of theme, poeticness, grammar, emotion, and likelihood of being human-written. Detailed instructions are in A.3." + }, + { + "type": "table", + "bbox": [ + 0.584, + 0.081, + 0.808, + 0.372 + ], + "angle": 0, + "content": "
CategoryMeanp-value
PoeTryMe
Grammar3.66*2.00 × 10-6
Emotion3.54*1.16 × 10-4
Poetic3.55*3.70 × 10-5
Human3.59*1.60 × 10-5
Theme2.860.19
Benhardt et al.
Grammar3.34*6.57 × 10-3
Emotion3.16*0.12
Poetic3.11*0.19
Human3.06*0.33
Theme2.770.06
Human-written poems
Grammar3.13*0.14
Emotion2.860.14
Poetic2.910.24
Human2.920.27
Theme2.670.02
" + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.382, + 0.884, + 0.44 + ], + "angle": 0, + "content": "Figure 4: Starred figures indicate average scores of \\(>3\\), and underlined figures indicate that the p-value is low enough \\((<0.05)\\) to claim that this higher average is statistically significant." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.473, + 0.702, + 0.489 + ], + "angle": 0, + "content": "4.1 Expert Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.496, + 0.884, + 0.673 + ], + "angle": 0, + "content": "For an expert evaluation, we recruited six faculty members and students from an academic English department. Figures 3 and 5 show that we strongly outperform PoeTryMe in all categories but theme with high statistical significance \\((p < 0.006)\\), and we outperform Benhardt et al. in all poetic categories but theme and emotion with statistical significance \\((p < 0.05)\\). Notably, while we outperform other computer-generated poems, respondents could still distinguish between our poems and human-written sonnets quite easily. See more in A.4." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.688, + 0.773, + 0.702 + ], + "angle": 0, + "content": "4.2 Amazon MTurk Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.71, + 0.885, + 0.919 + ], + "angle": 0, + "content": "Along with expert evaluation, we used Amazon MTurk services to assess poems on a larger scale. Figures 4 and 6 show our superior performance against competitors in several categories. As expected of most computer-generated work, our poems failed to outperform human-written poems. However, we can only strongly conclude that the human-written poems are better in one category, theme. Our poems even outperformed human-written poems in grammar (albeit with low statistical significance), showing that our strictly constrained beam search generates high quality grammar. See more in A.5." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1630" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.227, + 0.082, + 0.376, + 0.096 + ], + "angle": 0, + "content": "Expert Evaluation" + }, + { + "type": "image", + "bbox": [ + 0.138, + 0.103, + 0.247, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.097, + 0.36, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.103, + 0.465, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.138, + 0.169, + 0.247, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.16, + 0.357, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.169, + 0.465, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.138, + 0.236, + 0.247, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.23, + 0.357, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.237, + 0.465, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.138, + 0.303, + 0.247, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.296, + 0.357, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.304, + 0.465, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.138, + 0.371, + 0.247, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.362, + 0.357, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.371, + 0.465, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.444, + 0.49, + 0.502 + ], + "angle": 0, + "content": "Figure 5: Values \\(>3\\) (green), \\(< 3\\) (red), and \\(= 3\\) (gray) denote that our poetry model performs better, the competitor performs better, and the poems performed similarly, respectively." + }, + { + "type": "image_caption", + "bbox": [ + 0.587, + 0.082, + 0.807, + 0.096 + ], + "angle": 0, + "content": "Amazon MTurk Evaluation" + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.103, + 0.643, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.096, + 0.752, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.103, + 0.86, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.161, + 0.643, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.161, + 0.752, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.161, + 0.86, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.23, + 0.643, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.23, + 0.752, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.23, + 0.86, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.303, + 0.643, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.296, + 0.752, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.304, + 0.86, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.371, + 0.643, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.362, + 0.752, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.371, + 0.86, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.444, + 0.885, + 0.502 + ], + "angle": 0, + "content": "Figure 6: Values \\(>3\\) (green), \\(< 3\\) (red), and \\(= 3\\) (gray) denote that our poetry model performs better, the competitor performs better, and the poems performed similarly, respectively." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.53, + 0.318, + 0.544 + ], + "angle": 0, + "content": "4.3 Ablative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.553, + 0.49, + 0.826 + ], + "angle": 0, + "content": "We also conducted ablative studies showing the efficacy of two key elements of our method: line templates and the fine-tuned GPT-2 language model. We generated two sets of ablation poems: one with the fine-tuned GPT-2 and no templating, and one using the untrained GPT-2 model and templating. We then used Amazon MTurk services to test each set against poems generated with both factors under the same criteria as previous experiments. From Figure 11, it is the combination of the fine-tuned model and templating that ensures higher quality sonnets than if only one factor is implemented. Our poems with both factors outperform both sets of ablative poems with varying statistical significance. Specifically, providing templates is clearly the critical piece to generate poems of a high caliber. See more in A.6." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.843, + 0.247, + 0.858 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.871, + 0.489, + 0.92 + ], + "angle": 0, + "content": "We propose a novel method for generating high-quality poems that uses POS templating to determine a logical syntactical structure and rigorously" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.527, + 0.885, + 0.672 + ], + "angle": 0, + "content": "maintains constraints necessary for any sonnet. Our method is highly versatile, with poetic factors like alliteration, internal rhyme, repetition, and theme adjustable to ensure creative output. After extensive surveys conducted with expert evaluators and MTurk participants, our model's success over similar competitors is evident, though our model's poems, like those of most computer poetry generators, remain distinguishable from human written poems." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.672, + 0.885, + 0.848 + ], + "angle": 0, + "content": "While we were unable to compare our model's performance to that of ChatGPT, our finetuned GPT-2 requires far less computing power than subsequent GPT models. Additionally, while we commenced this project's evaluation prior to the release of ChatGPT, after a preliminary qualitative evaluation, ChatGPT seems to produce very generic poetry (see A.7). Thus, for this particular application, our model may be a viable method that is more cost-effective and produces relatively high-quality sonnets." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.862, + 0.615, + 0.877 + ], + "angle": 0, + "content": "Limitations" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.888, + 0.883, + 0.92 + ], + "angle": 0, + "content": "Though our method produces full sonnets that are more impressive than all previous approaches, it" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.519, + 0.941 + ], + "angle": 0, + "content": "1631" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.49, + 0.295 + ], + "angle": 0, + "content": "is still not at the level of human-generated poetry. It is not clear how to achieve this level, whether it would be using massive large language models, or through our general approach, which is to bend those models around an interpretable framework that knows the rules that sonnets obey. Certainly our approach requires a lot less data – even if one used all the sonnets that have ever been written to train a language model, it is unclear that the language model would learn the very specific rules required of sonnets. However, there may be other ways to obtain these constraints that have not yet been developed." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.306, + 0.266, + 0.322 + ], + "angle": 0, + "content": "Ethics Statement" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.332, + 0.491, + 0.412 + ], + "angle": 0, + "content": "As with all neural generation, there are concerns about misinformation and generating toxic text. These concerns apply to some degree to poetry generation, although our rigidly constrained approach and limited vocabulary should mitigate this." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.439, + 0.214, + 0.454 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.461, + 0.49, + 0.515 + ], + "angle": 0, + "content": "John Benhardt, Peter Hase, Liuyi Zhu, and Cynthia Rudin. 2018. Shall I compare thee to a machine-written sonnet? An approach to algorithmic sonnet generation." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.525, + 0.489, + 0.579 + ], + "angle": 0, + "content": "Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. 2017. Enriching word vectors with subword information. Transactions of the association for computational linguistics, 5:135-146." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.588, + 0.49, + 0.64 + ], + "angle": 0, + "content": "Carnegie Mellon University CMU. 2019. The CMU pronouncing dictionary. http://www.speech.cs.cmu.edu/cgi-bin/cmudict, Internet." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.651, + 0.488, + 0.704 + ], + "angle": 0, + "content": "Pablo Gervás. 2000. Wasp: Evaluation of different strategies for the automatic generation of spanish verse. In Proceedings of the AISB-00 Symposium on Creative & Cultural Aspects of AI, pages 93-100." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.714, + 0.49, + 0.768 + ], + "angle": 0, + "content": "Marjan Ghazvininejad, Xing Shi, Yejin Choi, and Kevin Knight. 2016. Generating topical poetry. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1183-1191." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.777, + 0.49, + 0.87 + ], + "angle": 0, + "content": "Jey Han Lau, Trevor Cohn, Timothy Baldwin, Julian Brooke, and Adam Hammond. 2018. Deep-spare: A joint neural model of poetic language, meter and rhyme. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1948-1958, Melbourne, Australia. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.879, + 0.489, + 0.919 + ], + "angle": 0, + "content": "Ruli Manurung, Graeme Ritchie, and Henry Thompson. 2000. Towards a computational model of poetry generation. https://era.ed.ac.uk/handle/1842/3460." + }, + { + "type": "list", + "bbox": [ + 0.115, + 0.461, + 0.49, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.086, + 0.883, + 0.114 + ], + "angle": 0, + "content": "George A Miller. 1998. WordNet: An electronic lexical database. MIT press." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.122, + 0.885, + 0.175 + ], + "angle": 0, + "content": "Hugo Gonçalo Oliveira. 2012. Poetry: a versatile platform for poetry generation. Computational Creativity, Concept Invention, and General Intelligence, 1:21." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.185, + 0.885, + 0.239 + ], + "angle": 0, + "content": "Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners. https://github.com/openai/gpt-2." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.248, + 0.885, + 0.314 + ], + "angle": 0, + "content": "Tim Van de Cruys. 2020. Automatic poetry generation from prosaic text. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2471-2480, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.324, + 0.885, + 0.403 + ], + "angle": 0, + "content": "Tony Veale. 2013. Less rhyme, more reason: Knowledge-based poetry generation with feeling, insight and wit. In Proceedings of the Fourth International Conference on Computational Creativity, ICCC 2013, Sidney, Australia, June 12-14, 2013, pages 152-159. computationalcreativity.net." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.412, + 0.885, + 0.464 + ], + "angle": 0, + "content": "Jianyou Wang, Xiaoxuan Zhang, Yuren Zhou, Christopher Suh, and Cynthia Rudin. 2021. There once was a really bad poet, it was automated but you didn't know it." + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.086, + 0.885, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.489, + 0.634, + 0.505 + ], + "angle": 0, + "content": "A Appendix" + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.514, + 0.744, + 0.529 + ], + "angle": 0, + "content": "A.1 Templating Mechanism" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.535, + 0.884, + 0.841 + ], + "angle": 0, + "content": "Figure 8 presents more examples of our templating mechanism. We combine an adapted version of the Penn Treebank Project's part of speech tags along with articles, conjunctions, prepositions, and other filler words to construct these templates. Additionally, we provide the stress pattern of the syllables to ensure that the constraint of iambic pentameter is met. However, outside of the pre-determined filler words, POS do not have to directly adhere to the given stress pattern in splitting up words. For instance, in the first template, the provided syllable stress indicates that the JJ tag (adjective) should have two syllables, while the final VB tag (verb) should have only one syllable. However, the generated line ends with a monosyllabic adjective and a bisyllabic verb. As long as the stressing of the syllables aligns properly, each word can vary in its number of syllables. This is also visible in the fourth template example in Figure 8." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.851, + 0.791, + 0.882 + ], + "angle": 0, + "content": "A.2 Elaboration on Experimental Competitors" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.888, + 0.883, + 0.919 + ], + "angle": 0, + "content": "Benhardt et al. (2018), referred to as Benhardt et al., uses a RNN to preselect rhyming words and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1632" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.49, + 0.166 + ], + "angle": 0, + "content": "restrict different parts of speech to fit within the sonnet format. Oliveira (2012), referred to as Co-PoetryMe, is a versatile platform using semantic and grammar templates to alter the type of poem, input words, and \"surprise\" factor generated." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.176, + 0.358, + 0.191 + ], + "angle": 0, + "content": "A.3 Experimental Procedure" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.196, + 0.49, + 0.63 + ], + "angle": 0, + "content": "For each pair of sonnets, respondents were asked to indicate whether Sonnet A or Sonnet B performed better based on factors such as adherence to the inputted theme, poeticness, grammatical correctness, ability to convey emotion, and likelihood of being written by a human. Available answer choices and their corresponding numeric scores from 1 to 5 were \"Definitely A\" (5), \"Probably A\" (4), \"The same\" (3), \"Probably B\" (2), and \"Definitely B\" (1). Both our sonnet and the competing model-human-sonnet had equal probability of being either sonnet A or sonnet B in each pair. To analyze this data, user inputs were translated into numeric scoring values corresponding to our model's sonnet being Sonnet A (i.e. if our sonnet is presented as B to the user, a response of \"Definitely B\" corresponds to a score of 5, \"Probably B\" corresponds to 4, \"Probably A\" corresponds to 2, and \"Definitely A\" corresponds to 1). Additionally, respondents were asked to answer sanity check questions to filter out respondents who answer illogically or who do not have a sufficient grasp of English grammar. This setup remained the same across all experiments, and an additional space was allocated for expert evaluators to leave qualitative comments on sonnet quality. Sample sonnet evaluation questions are visible in Figure 9." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.632, + 0.49, + 0.84 + ], + "angle": 0, + "content": "After calculating the mean and standard deviation for scores across sonnets, we can immediately see whether our model performed better (an average score of \\(>3\\)) or worse (an average score of \\(< 3\\)) than the competitor in each respective category. We then performed a series of t-tests to establish these results' statistical significance. For factors that indicated our model performed better, we performed a right-tailed t-test (with the null-hypothesis as our model performed worse than the baseline), and we performed a left-tailed t-test for the remaining factors (with the null-hypothesis as our model performed better than the baseline)." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.851, + 0.381, + 0.866 + ], + "angle": 0, + "content": "A.4 Expert Evaluation Analysis" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.872, + 0.489, + 0.92 + ], + "angle": 0, + "content": "In the expert evaluation, we emailed faculty at an American academic English department to recruit six faculty members and students to take our survey" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.295 + ], + "angle": 0, + "content": "without payment. While we showed strong performance against the other computer-generated poems, we are consistently outperformed by human-written poems in all categories. Weaker performance on theme in experimental results may be explained by competitors' more frequent inclusion of the user-inputted theme word. For instance, in the expert evaluation, between two poems generated with the theme word \"forest\" (see Figure 10), one survey respondent states, \"Sonnet B repeats forest too much for my taste,\" subsequently giving our model a 5 in each of poeticness, grammar, emotion, and humanness, yet a 2 in theme." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.305, + 0.759, + 0.32 + ], + "angle": 0, + "content": "A.5 Amazon MTurk Analysis" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.325, + 0.883, + 0.469 + ], + "angle": 0, + "content": "In our evaluation using Amazon MTurk Services, we requested survey respondents from primarily English-speaking countries and with an approval rate of \\(\\geq 95\\%\\). Crowdworkers were paid through the Amazon MTurk platform for this survey that on average took less than 30 minutes to complete. The questions and formatting remained the same as the expert evaluation, except no space was provided for qualitative feedback." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.47, + 0.884, + 0.728 + ], + "angle": 0, + "content": "Based on Figure 4 there is enough statistical significance to conclude that our sonnets outperform PoeTryMe in poetic, grammar, emotion, and human categories \\((p < 0.001)\\). Against Benhardt et al., there is enough statistical significance to conclude that our sonnets perform better in grammar \\((p < 0.001)\\), and perform slightly better with weak statistical significance in emotion \\((p < 0.15)\\). Against human-written sonnets, the p-values for poetic, emotion, and even human categories are too large to strongly reject the null hypothesis that our model performed better than the baseline. Additionally, while the p-value indicates that this value is not statistically significant, it is interesting to note that our poems on average scored better in the grammar category." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.739, + 0.702, + 0.754 + ], + "angle": 0, + "content": "A.6 Ablation Analysis" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.759, + 0.883, + 0.919 + ], + "angle": 0, + "content": "In our ablation analysis, we replicate the Amazon MTurk analysis yet replace the competitor/human-written sonnets with poems generated with either the fine-tuned GPT-2 model without templating or the GPT-2 model without fine-tuning and with templating. This lets us test the individual efficacy of each factor (templating and fine-tuning GPT-2) against our method implementing both. Against poems generated with the fine-tuned GPT-2 and no templating, our sonnets performed better across" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1633" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.493, + 0.264 + ], + "angle": 0, + "content": "all categories, and we can strongly reject the null hypothesis that our model performed worse than the baseline \\((p < 0.0001)\\). Against the poems generated with the GPT-2 model without fine-tuning and with templates, we can conclude with high statistical significance \\((p < 0.01)\\) that we performed better in emotion, and conclude with weak statistical significance \\((p < 0.10)\\) that we performed better in grammar and theme. These results indicate that our method is successful due to its usage of both the fine-tuned GPT-2 model and templating." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.273, + 0.406, + 0.289 + ], + "angle": 0, + "content": "A.7 ChatGPT Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.293, + 0.49, + 0.568 + ], + "angle": 0, + "content": "While we did not have time to extensively evaluate the quality of our sonnets against those of ChatGPT, after generating several sonnets to test ChatGPT's sonnet quality, it seems as though this language model generates relatively generic, non-cohesive sonnets even with different parameters. For instance, in Figure 7, both of the sonnets are unable to cohesively connect these three topics along a reasonable storyline. Additionally, Sonnet A in particular seems to dedicate a single stanza to each of the three topics passed in, hardly attempting to connect them. Of course, with more intensive prompt engineering, it is possible to generate a sonnet more tailored to one's preference. However, even this short analysis demonstrates there are clearly still strides to be made in the field of automatic poetry generation even with the advent of ChatGPT." + }, + { + "type": "table", + "bbox": [ + 0.51, + 0.266, + 0.859, + 0.652 + ], + "angle": 0, + "content": "
Sonnet A
In realms of words, where novels find their reign, My friend, a skeptic, shuns the written tales. No pages turned, no fictional campaigns, Averse to narratives that time unveils. Yet in his hand, a pear's sweet succulence, A juicy feast, a taste of nature's grace. Each bite, a moment of pure indulgence, A salute's joy, a respite from life's race. But beyond fruit's delight, his spirit soars, Within the bounds of a duelist's domain. With sword in hand, he dances, parries, scores, A noble art where honor's stance remains. Though novels may elude his curious gaze, Through pears and fencing, passions find their ways.
Sonnet B
In realms of love, where rom-coms dance their waltz, My friend, a rebel, strays from silver screens. With flannel-clad allure, he stands, exalts, A soul untouched by scripted love's routines. No tears shed for the tales of starry-eyed, His heart resists the charms of love's façade. Yet stitches formed with yarn, his hands provide, Creating warmth, affection neatly clad. Through threads of patience, vibrant patterns bloom, His nimble fingers weave a tapestry. Each crochet stitch, a tale of friendship's loom, A testament to artistry set free. Though rom-coms fall short of his embrace, In flannel and crochet, love finds its place.
" + }, + { + "type": "table_caption", + "bbox": [ + 0.509, + 0.662, + 0.885, + 0.747 + ], + "angle": 0, + "content": "Figure 7: Comparison of two sonnets generated with ChatGPT. Sonnet A was generated with the prompt \"generate a sonnet about my friend who hates novels, eats pears, and fences\" and Sonnet B was generated with the prompt \"generate a sonnet about my friend who hates rom-coms, wears flannels, and crochets.\"" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1634" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.141, + 0.082, + 0.858, + 0.196 + ], + "angle": 0, + "content": "
TemplateSyllable StressExample Line
Where all the NNS of PRPD$ JJ NNS VB.0 1 0 1 0 1 01 0 1“Where all the gods of their past lives dictate”
And it VBD ABNN to the NN0 1 0 10 1 0 101“And it seemed evil to the enterprise”
Between the VBG and the VBG NN01 0 10 1 0 10 1“Between the glistening and the dying muse”
A JJ NN from the JJ NN0 10 10 1 0 1 01“A little lightness from the earthy sky”
Upon PRPO, PRPD$ NN POS NN01 01 0 10 101“Upon you, your life’s possibility”
Why VBC PRPS VBG such a JJ NN?0 1 0 10 1 0 101 0"Why do you squander such a precious thing?"
The NNS of ABNN, the NN on the NN0 1 0 1 0 10 1 0 1“The ghosts of death, the spirit on the earth”
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.21, + 0.884, + 0.24 + ], + "angle": 0, + "content": "Figure 8: Template examples, their corresponding syllable stress in order to adhere to iambic pentameter, and a sample line generated using the template." + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.442, + 0.373, + 0.45 + ], + "angle": 0, + "content": "The key word for both of these poems is \"wisdom.\" Which poem best adheres to this theme?" + }, + { + "type": "table", + "bbox": [ + 0.151, + 0.459, + 0.451, + 0.478 + ], + "angle": 0, + "content": "
Definitely AProbably ASameProbably BDefinitely B
" + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.493, + 0.232, + 0.5 + ], + "angle": 0, + "content": "Which poem sounds more poetic?" + }, + { + "type": "table", + "bbox": [ + 0.159, + 0.511, + 0.442, + 0.522 + ], + "angle": 0, + "content": "
Definitely AProbably ASameProbably BDefinitely B
" + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.542, + 0.255, + 0.549 + ], + "angle": 0, + "content": "Which poem is more grammatically correct?" + }, + { + "type": "table", + "bbox": [ + 0.159, + 0.568, + 0.442, + 0.584 + ], + "angle": 0, + "content": "
Definitely AProbably ASameProbably BDefinitely B
" + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.601, + 0.267, + 0.608 + ], + "angle": 0, + "content": "Which poem conveys emotions more effectively?" + }, + { + "type": "table", + "bbox": [ + 0.159, + 0.62, + 0.441, + 0.63 + ], + "angle": 0, + "content": "
Definitely AProbably ASameProbably BDefinitely B
" + }, + { + "type": "text", + "bbox": [ + 0.146, + 0.65, + 0.276, + 0.657 + ], + "angle": 0, + "content": "Which poem is more likely to be written by a human?" + }, + { + "type": "table", + "bbox": [ + 0.159, + 0.67, + 0.442, + 0.679 + ], + "angle": 0, + "content": "
Definitely AProbably ASameProbably BDefinitely B
" + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.704, + 0.489, + 0.731 + ], + "angle": 0, + "content": "Figure 9: Survey questions presented for each pair of sonnets." + }, + { + "type": "table", + "bbox": [ + 0.511, + 0.377, + 0.878, + 0.763 + ], + "angle": 0, + "content": "
Sonnet A: Our Code
I was aghast to see the fireflies
Inflamed soothed toads, where there the dead boughs lay
And it seemed evil to the enterprise
The hag I had, the hag, the hog, the gray.
But I knew to my painless fireflies
And beauty was a kind and loving thing.
My life's light isle so longed on otherwise
So too my fireflies bloomed to my king.
Those eagles that with auburn hair flew oaks,
Beauty and beauty beamed within the air
Which made oasis overcomes to coax?
So too my hogs beheaded to my lair.
The windy night was in the mistletoe
And wept soiled toads in my dream's studio.
Sonnet B: PoetryMe
forest some more and reforest a trip!
in deserts where heavenly woodlands clink
many, many, many clustered before
come: not in establishments of the floor
the fields of agony, the endless circumstance
findings to lie to interrupt your earth
with summation and set, triumph and agony
floors of horror forest before my eyes
those that study clustered plant are psychologists
taking over my ness a second forest
an' you've got to forest them reforest
on every forest, indeed, that rainforests
and grounds of forest coming to accord
floor of establishments and lilt of sing
" + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.773, + 0.883, + 0.816 + ], + "angle": 0, + "content": "Figure 10: Comparison of two sonnets generated with theme word \"forest\". Sonnet A was generated with our code, and Sonnet B was generated using PoeTryMe." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1635" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.221, + 0.395, + 0.382, + 0.407 + ], + "angle": 0, + "content": "Ablation Evaluation" + }, + { + "type": "table", + "bbox": [ + 0.123, + 0.407, + 0.481, + 0.5 + ], + "angle": 0, + "content": "
CategoryMeanp-valueMeanp-value
Grammar3.51*5.10×10-53.21*0.06
Emotion3.61*9.00×10-63.40*3.89×10-3
Poetic3.61*4.00×10-63.09*0.29
Human3.66*1.00×10-63.01*0.46
Theme3.50*8.00×10-53.20*0.06
" + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.516, + 0.49, + 0.603 + ], + "angle": 0, + "content": "Figure 11: Left: fine-tuned GPT-2 with no templates. Right: GPT-2 without fine-tuning, but with templates. Starred figures indicate average scores of \\(>3\\), and underlined figures indicate that the p-value is low enough \\((<0.05)\\) to claim that this higher average is statistically significant." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1636" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.134, + 0.085, + 0.435, + 0.1 + ], + "angle": 0, + "content": "ACL 2023 Responsible NLP Checklist" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.108, + 0.323, + 0.123 + ], + "angle": 0, + "content": "A For every submission:" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.127, + 0.533, + 0.158 + ], + "angle": 0, + "content": "A1. Did you describe the limitations of your work? Limitations" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.169, + 0.553, + 0.2 + ], + "angle": 0, + "content": "A2. Did you discuss any potential risks of your work? Ethics" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.213, + 0.696, + 0.244 + ], + "angle": 0, + "content": "A3. Do the abstract and introduction summarize the paper's main claims? Abstract, 1" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.256, + 0.669, + 0.288 + ], + "angle": 0, + "content": "A4. Have you used AI writing assistants when working on this paper? Left blank." + }, + { + "type": "list", + "bbox": [ + 0.13, + 0.127, + 0.696, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.3, + 0.489, + 0.316 + ], + "angle": 0, + "content": "B Did you use or create scientific artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.322, + 0.162, + 0.335 + ], + "angle": 0, + "content": "3,4" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.347, + 0.53, + 0.379 + ], + "angle": 0, + "content": "B1. Did you cite the creators of artifacts you used? 3.2,3.3,References" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.391, + 0.779, + 0.423 + ], + "angle": 0, + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.434, + 0.881, + 0.513 + ], + "angle": 0, + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.525, + 0.881, + 0.59 + ], + "angle": 0, + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Data used from publicly available sonnets/ poems were assumed to be not subject to dispute." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.6, + 0.881, + 0.648 + ], + "angle": 0, + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.658, + 0.881, + 0.753 + ], + "angle": 0, + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. 3.3" + }, + { + "type": "list", + "bbox": [ + 0.129, + 0.347, + 0.881, + 0.753 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.766, + 0.494, + 0.782 + ], + "angle": 0, + "content": "C Did you run computational experiments?" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.788, + 0.215, + 0.802 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.814, + 0.881, + 0.861 + ], + "angle": 0, + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? No response." + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.869, + 0.878, + 0.893 + ], + "angle": 0, + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1637" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.129, + 0.085, + 0.88, + 0.133 + ], + "angle": 0, + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.144, + 0.881, + 0.207 + ], + "angle": 0, + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.219, + 0.881, + 0.282 + ], + "angle": 0, + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? No response." + }, + { + "type": "list", + "bbox": [ + 0.129, + 0.085, + 0.881, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.293, + 0.878, + 0.331 + ], + "angle": 0, + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? 4,4.1,4.2,4.3" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.34, + 0.881, + 0.388 + ], + "angle": 0, + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Appendix" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.399, + 0.881, + 0.461 + ], + "angle": 0, + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? A.5.A.6" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.475, + 0.881, + 0.538 + ], + "angle": 0, + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? We do not believe having data on poetry evaluation raises any ethical issues." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.55, + 0.88, + 0.597 + ], + "angle": 0, + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? We do not believe having crowdworkers evaluate the same poems that were given to English professors raises any ethical issues." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.608, + 0.88, + 0.655 + ], + "angle": 0, + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? A.6" + }, + { + "type": "list", + "bbox": [ + 0.13, + 0.34, + 0.881, + 0.655 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1638" + } + ] +] \ No newline at end of file diff --git a/2023/The Mechanical Bard_ An Interpretable Machine Learning Approach to Shakespearean Sonnet Generation/3330f90d-0cb8-4666-aeb2-f9f31ccb9534_origin.pdf b/2023/The Mechanical Bard_ An Interpretable Machine Learning Approach to Shakespearean Sonnet Generation/3330f90d-0cb8-4666-aeb2-f9f31ccb9534_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e21651a4ba05638924aa3ede9ac4b7e638692c32 --- /dev/null +++ b/2023/The Mechanical Bard_ An Interpretable Machine Learning Approach to Shakespearean Sonnet Generation/3330f90d-0cb8-4666-aeb2-f9f31ccb9534_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3863eb712d9c1dd98ae17f1aaddc299a2df7ce92dc1c7700bb3442b27988e509 +size 577765 diff --git a/2023/The Mechanical Bard_ An Interpretable Machine Learning Approach to Shakespearean Sonnet Generation/full.md b/2023/The Mechanical Bard_ An Interpretable Machine Learning Approach to Shakespearean Sonnet Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..bb536f22f31eb05e3e2897228d3da3dbd1871dba --- /dev/null +++ b/2023/The Mechanical Bard_ An Interpretable Machine Learning Approach to Shakespearean Sonnet Generation/full.md @@ -0,0 +1,356 @@ +# The Mechanical Bard: An Interpretable Machine Learning Approach to Shakespearean Sonnet Generation + +Edwin Agnew*, Michelle Qiu*, Lily Zhu*, Sam Wiseman, Cynthia Rudin + +Duke University, Durham, NC + +edwin.agnew@duke.edu, michelle.qiu@duke.edu, lily.zhu@duke.edu + +swiseman@cs.duke.edu, cynthiaa@cs.duke.edu + +# Abstract + +We consider the automated generation of sonnets, a poetic form constrained according to meter, rhyme scheme, and length. Sonnets generally also use rhetorical figures, expressive language, and a consistent theme or narrative. Our constrained decoding approach allows for the generation of sonnets within preset poetic constraints, while using a relatively modest neural backbone. Human evaluation confirms that our approach produces Shakespearean sonnets that resemble human-authored sonnets, and which adhere to the genre's defined constraints and contain lyrical language and literary devices. + +When all the lovers of this world are dead, + +The sun of heaven on a golden day + +To burn the earth's fire by the flame and spread + +Where all the flowers of your fair days lay. + +These are the blossoms that you take care of. + +Why do you linger such a long delay? + +Forgive the fluttered flower of meek love + +Or who you have so long to love the day? + +The joys of love, the beauty on the face, + +Shall be your fate and be your own delight. + +You have the beauty of your own embrace. + +You cannot reminiscence. Cannot write. + +Between the living and the deadening breath. + +You go the way of your beloved death. + +Figure 1: A sonnet generated with the theme "death". + +# 1 Introduction + +We consider the task of automatically generating Shakespearean sonnets, a popular poetic form with highly specific rhyme and meter constraints1. Each sonnet consists of three quatrains followed by a single couplet according to the rhyme scheme ABAB BCBC CDCD EE, and each line contains ten syllables with a stress pattern of iambic pentameter. + +Rather than train a model to obey these constraints implicitly (which leads to enormous models that still do not obey the constraints), we opt to enforce them explicitly using a simple but novel approach to generation. + +In particular, we use part-of-speech (POS) templates selected and edited from individual lines in Shakespeare's sonnets, with each template intended to offer a different combination of parts of speech and narrative directions. Associated thematic words are then selected and placed at the end of each template, and their rhyming pairs are chosen dynamically by a language model (e.g., GPT-2, Radford et al., 2019) and placed at the end of the corresponding lines according to the rhyme scheme. + +The rest of the line is filled with related words that fit the specified POS and meter, leading to the end rhyme word. Figure 1 shows sample output. + +Our use of these templates ensures sophisticated-seeming language and syntax that competing systems do not capture. Our approach provides excellent grammatical structure comparable to that of human-written poetry, all while using a relatively simple model and generation procedure. + +We extensively evaluate the ability of our approach to generate whole sonnets (a setting often ignored by recent work in poetry generation) and find that our approach is preferred over strong baselines by both expert annotators (recruited from an academic English department) and by crowdworkers. As this research was conducted before the release of ChatGPT, we were not able to robustly compare our model's performance against this language model. However, we make several observations about the poetic quality of sonnets generated by ChatGPT. + +# 2 Related Work + +Early attempts at poetry generation relied mainly on rule-based methods (Gervás, 2000; Oliveira, + +2012; Manurung et al., 2000; Veale, 2013). More recent automated poetry generation techniques, especially for sonnet generation, have relied on combinations of task-specific language models and rules. For instance, Ghazvininejad et al. (2016)'s Hafez uses a finite state acceptor to generate a large number of possible lines, the best of which are then selected with an RNN trained on song lyrics. Like our approach, they use rhyming dictionaries to find rhyming words and word embeddings to find topical words. Similarly, Benhardt et al. (2018) preselects rhyming words and generates lines backwards with a recurrent neural network (RNN). Also in this vein are Lau et al. (2018)'s Deepspare, which consists of an LSTM language model, an iambic model, and a rhyming model, and the recent work of Van de Cruys (2020) and Wang et al. (2021). + +Our approach distinguishes itself in using a general-purpose pretrained language model, but more importantly in its use of human-curated constraints and templates. These allow for generating high-quality poems with a very simple approach. + +# 3 Methodology + +The general idea of our approach is to take a pretrained language model (in this case GPT-2) and apply hard constraints to the generation procedure so that it can only output text satisfying various poetic constraints. These constraints can be broadly divided into hard constraints (e.g., number of syllables) and soft constraints (e.g., sounding poetic), and our methodology can be separated similarly. Our generation process is in Figure 2. + +# 3.1 POS Templates + +The most important part of our method is the use of handcrafted grammar templates. Taking inspiration from existing sonnets, we created a list of about 120 templates that encode the part-of-speech structure of a line of poetry. Each template can generate an unbounded number of possible poetic lines. For example, the line "The beauty of life on a lonely sea" is represented by the template "THE NN OF NN ON A JJ NN." More sample templates are in Section A.1. Since the templates allow for considerable flexibility, obeying the templates does not alone suffice for poetry. For example, the same template could be used to write poetic lines with distinct meanings such as "The tree of anguish on a stormy night" or a nonsensical line like "The fork of ant on an unpacked transfer." A subset of these + +templates is also chosen for starting a stanza. + +# 3.2 Strict Sonnet Constraints + +The two most critical features of sonnets distinguishing them from other poetry forms are that they are written in iambic pentameter (i.e., each line has 10 syllables of alternating stress pattern), and they follow an ABAB CDCD EFEF GG rhyme scheme. To detect iambic pentameter, we use the CMU Pronouncing Dictionary (CMU, 2019), which reveals how many syllables a word contains and the stress of each syllable. An unstressed syllable is represented as '0' and a stressed syllable as '1', and so the line "The beauty of life on a lonely sea" is represented as '0 10 1 0 1 0 10 1'. For simplicity, 1-syllable words can be designated as either 0 or 1. + +Given a POS-tag for every word in our dictionary, we create a tree-like data structure that represents every possible meter for a given template. Continuing the example, the first word could only be 'the', but the second word could be filled with a 1-syllable noun like 'tree', a 2-syllable noun like 'chaos' (10), or a 3-syllable noun like 'audio' (101), and so on. Each choice affects the possible pronunciations of the next word as well as the number of remaining words in order to reach 10 syllables. The pronunciation dictionary ensures the last syllable of the last word on each line matches its partner. + +# 3.3 Language Model + +We use a language model to generate individual sonnet lines, subject to the formal constraints outlined above. In particular, we first fine-tune GPT-2 (Radford et al., 2019) on a large corpus of over 15000 poems and a smaller corpus of sonnets. We then use a constrained beam-search to generate each line, where only legal tokens (under the aforementioned constraints) can be generated at each step; this generation approach resembles previous constrained decoding techniques used in sonnet generation (Ghazvininejad et al., 2016), although our approach differs in the choice of model and direct enforcement of constraints. For a comparison of generation quality using a GPT-2 model that has not been fine-tuned, see Section 4.1. + +# 3.4 Thematic Word Choice + +To ensure the content of the poem fits the theme specified by the user, we provide an excerpt of a + +![](images/7f5699c3abb074e217a003ab4ce4e8055fa6be50e33b760657614f7eed444060.jpg) +Generation Visualization +Figure 2: Numbers in parentheses denote subsections in Section 3. + +theme-appropriate poem as additional context to GPT-2 during generation. This additional poem is selected by finding a list of synonyms to the theme word using the WordNet synonym database (Miller, 1998) and then choosing lines from a poem corpus that contain at least one synonym. We also remove words from the vocabulary if they have less than 0.5 cosine similarity with the theme word, based on the corresponding FastText word embeddings (Bojanowski et al., 2017). This avoids having words like "algebra" in poems with themes like "forest." + +# 3.5 Generation Procedure + +Having introduced our method's components, we now describe the generation procedure. A user inputs a theme word, a beam search parameter, $b$ , and the number of templates sampled per line, $k$ . A seed is chosen with the above method. Then for each line, we sample $k$ random templates. For each template, we generate the line using a modified beam search. Specifically, the beam search maintains $b$ different hypotheses per line at all times. For each hypothesis, we first mask out any tokens that violate our hard POS, meter, or rhyme constraints and select the $b$ best next-tokens for each + +of the $k$ templates. These $b^{2}$ new candidates are re-ranked according to our custom scoring function, and the top $k \times b$ proceed to the next stage. The constraint-filtering at each stage guarantees that the generated line will match the input template, while the beam search allows more flexible word choice than greedy word-filling for each POS. If none of the $k \times b$ generated lines score better than a specific threshold, then a new template is chosen and the line is generated again. Otherwise, line generation continues until the poem is completed. + +# 3.6 Poetic Devices + +To make the poems more poetic, we adjust our scoring function to weight lines with alliteration, penalties for repetition, and/or internal rhyme. Alliteration occurs when a line contains words starting with the same letter, repetition occurs when a word is present several times throughout a poem, and internal rhyme occurs when two words rhyme within the same line. To weight alliteration, when the first token of a new word is being generated, a list $\vec{A} = [a_1,a_2,\dots a_n]$ is generated where $a_{i}$ is the number of occurrences of the first letter of the ith token in the current line. To weight and discourage repetition, a list $\vec{T} = [t_1,t_2,\dots t_n]$ is generated where $t_i$ is the number of occurrences of the ith token in the poem, negated. To weight internal rhyme, a list $\vec{R} = [r_1,r_2,\dots ,r_n]$ is generated where $r_i = 1$ if the ith token is part of a word that rhymes with any of the words in the current line generated so far, and $r_i = 0$ otherwise. The final token distribution is then proportional to $\tilde{P} +\alpha_{A}\times \vec{A} +\alpha_{T}\times \vec{T} +\alpha_{R}\times \vec{R},$ where $\tilde{P}$ is the language model's next-token distribution, and $\alpha_{A},\alpha_{T},$ and $\alpha_{R}$ are user-specified non-negative parameters, which represent the degree to which alliteration, repetition, and internal rhyme should be favored during generation. + +# 3.7 Postprocessing + +After a poem is completed and all 14 lines score above a fixed threshold, a small number of adjustments are made. These include fixing common mistakes made by GPT-2 like not capitalizing the word 'I' and not capitalizing following punctuation. + +# 4 Experiments + +We used human input to test our sonnets against both model-generated and human-written sonnets. To test adherence to a theme throughout a son + +
CategoryMeanp-value
PoeTryMe
Grammar4.50*1.71×10-4
Emotion4.30*3.13×10-3
Poetic4.30*3.13×10-3
Human4.10*5.77×10-3
Theme2.600.211286
Benhardt et al.
Grammar3.83*0.03
Emotion3.67*0.05
Poetic3.75*0.04
Human3.75*0.02
Theme2.420.06
Human-written poems
Grammar1.361.00×10-6
Emotion1.45.00×10-6
Poetic1.645.40×10-5
Human1.361.00×10-6
Theme1.577.70×10-5
+ +Figure 3: Starred figures indicate average scores of $>3$ , and underlined figures indicate that the p-value is low enough $(<0.05)$ to claim that this higher average is statistically significant. + +net, we desired baselines that generate whole sonnets with user-provided themes. This limited our competitors, as some generate discrete quatrains or generate without input themes (e.g., Deepspare), leaving only Benhardt et al. (2018) and PoeTryMe (Oliveira, 2012); see Section A.2. + +Furthermore, an evaluation of poetry quality is incomplete without human-written sonnets, selected from sonnets.org. Though these poems do not have an explicit theme, we selected poems that followed our five themes. + +To optimally test our model, we conducted an internal analysis and selected $k$ values sampled from 3, 5, or 7, $b$ values sampled from 3, 5, or 7, and repetition penalty values sampled from 1.4, 1.6, or 1.8 that we concluded produced the highest quality sonnets. To evaluate adherence to theme, we generated poems with themes "death," "darkness," "forest," "love," and "wisdom." + +In each test, respondents compared six randomly selected pairs of sonnets, with each of our sonnets displayed with a competing model/human-written sonnet generated with the same theme word. Respondents indicated which of the two sonnets performed better in categories of theme, poeticness, grammar, emotion, and likelihood of being human-written. Detailed instructions are in A.3. + +
CategoryMeanp-value
PoeTryMe
Grammar3.66*2.00 × 10-6
Emotion3.54*1.16 × 10-4
Poetic3.55*3.70 × 10-5
Human3.59*1.60 × 10-5
Theme2.860.19
Benhardt et al.
Grammar3.34*6.57 × 10-3
Emotion3.16*0.12
Poetic3.11*0.19
Human3.06*0.33
Theme2.770.06
Human-written poems
Grammar3.13*0.14
Emotion2.860.14
Poetic2.910.24
Human2.920.27
Theme2.670.02
+ +Figure 4: Starred figures indicate average scores of $>3$ , and underlined figures indicate that the p-value is low enough $(<0.05)$ to claim that this higher average is statistically significant. + +# 4.1 Expert Evaluation + +For an expert evaluation, we recruited six faculty members and students from an academic English department. Figures 3 and 5 show that we strongly outperform PoeTryMe in all categories but theme with high statistical significance $(p < 0.006)$ , and we outperform Benhardt et al. in all poetic categories but theme and emotion with statistical significance $(p < 0.05)$ . Notably, while we outperform other computer-generated poems, respondents could still distinguish between our poems and human-written sonnets quite easily. See more in A.4. + +# 4.2 Amazon MTurk Evaluation + +Along with expert evaluation, we used Amazon MTurk services to assess poems on a larger scale. Figures 4 and 6 show our superior performance against competitors in several categories. As expected of most computer-generated work, our poems failed to outperform human-written poems. However, we can only strongly conclude that the human-written poems are better in one category, theme. Our poems even outperformed human-written poems in grammar (albeit with low statistical significance), showing that our strictly constrained beam search generates high quality grammar. See more in A.5. + +![](images/8292ada2b3790048c4338774c85281dfbd1e018dce8050afac84f1d7d06c39fe.jpg) + +![](images/4b0713578ccefc477b272aaa9b622eef617b3bae4d06c5aa420cce41aa5e4519.jpg) +Expert Evaluation + +![](images/28af1fff4996f1e72162e0878094ff4e419f53aa295c6e318d5eb132fcb1ae0c.jpg) + +![](images/2c14d35ba10f132dcc548d4c42f4d72e869f9787f6ca1bda0d37409ed461e96d.jpg) + +![](images/6672ec6bca3259931c113a548b9d1eaa8cfc504a903f4734e70a043bfd93d0e4.jpg) + +![](images/c4297386128e901ffc60a136327959a030c1277cf1b17770a6468fdffd7c187d.jpg) + +![](images/19ce2d9c4214eab638404a378bc61c99bf05a5f7ab00d09a673b02a997bfa246.jpg) + +![](images/1cedeb4682139532a00ed47d6a618d784b432148d73e260bc38c191332ea4af9.jpg) + +![](images/67679e7d43a4d23aa706b89624fcb0f5c0eac2d494b3ad71613a2aee7cd4420e.jpg) + +![](images/a6413b777f70c49d98b182b4faaadd35a417180f9006dc3c2fff64feae8c72de.jpg) + +![](images/e874758c6f482612209aa93d6c9a06fc26b5e577c3392b4b0746c8de4e4caf2c.jpg) + +![](images/ba3ca8ec4341772c19fce7920aa726f46ef05c17ba537fc692de309bceff122c.jpg) + +![](images/f64944439299e8d0a8d81b36cf7c9a0aa03fb92e7c74750ea93dca0a026941a9.jpg) +Figure 5: Values $>3$ (green), $< 3$ (red), and $= 3$ (gray) denote that our poetry model performs better, the competitor performs better, and the poems performed similarly, respectively. + +![](images/09383ba0baf9877a9acf94df5644cf06a6304ed37969e444372aae77afb8a4a9.jpg) + +![](images/fe3563cea5094c2ee93e2b84e663513a968b99f5e4870193fdd3885f10d2e662.jpg) + +![](images/30ead587a2e210153662cc053b8634a4b29d455b2be031800d1023b8e3c90d9f.jpg) + +![](images/194ef67c2c27e30692951db29732552a3e149e82029cf8563645d64756b08f36.jpg) +Amazon MTurk Evaluation + +![](images/5efdcd6dc2ada6062c20b898c74d88db7eaf0134ce3dfab175fb82f9516fe683.jpg) + +![](images/df37ccfd094514f4c522b43dca9c98320896473e400b160c0f34cbb93b4da269.jpg) + +![](images/ed4a3e6f4d49806be96b00477e9928dcd21683df6a77a912e8d2b8bc4c586a0f.jpg) + +![](images/554e63d71ce9123238f0b7ef4b3628b775dffe027121633a688cc331294fa0b6.jpg) + +![](images/c595f5131a009d57178fe00e1f55a2d400ddc39c282914aae7aadd2de606b644.jpg) + +![](images/fdee82545dc7ce52c6a9e180528600d93cefe993f4da71a5b653cb3e4a80e58a.jpg) + +![](images/cdfa9220bf60fd653a740d53f0910e146a9fdfd5919f7d1aed13c1cf83e54e45.jpg) + +![](images/ee4d023da1c0d5f2b54320b99eb0bdffa062a22dcc635478ccd7bf69d1c522da.jpg) + +![](images/848bb6ad99fc0f5fbe9ecbe895045891baecd155adf3f61f8b37f6ec2c865859.jpg) + +![](images/b44abebd499c22ccd92aa5b2dbbb98141d4069c8299ce8d5ba38a6c621dd6e3f.jpg) + +![](images/e0c7a83428e11c556c4e5f31cd846719b98b0998b25f291eed5ba3f4270e83df.jpg) +Figure 6: Values $>3$ (green), $< 3$ (red), and $= 3$ (gray) denote that our poetry model performs better, the competitor performs better, and the poems performed similarly, respectively. + +![](images/ed0cb71915719919b394595e5793eeea15a98ace3b83d4297ae6fa2e56d37ea3.jpg) + +![](images/b16d15f4bb980d51422af2a12e796ad88f0a5adbaead5914bff4058e9e188ca1.jpg) + +# 4.3 Ablative Evaluation + +We also conducted ablative studies showing the efficacy of two key elements of our method: line templates and the fine-tuned GPT-2 language model. We generated two sets of ablation poems: one with the fine-tuned GPT-2 and no templating, and one using the untrained GPT-2 model and templating. We then used Amazon MTurk services to test each set against poems generated with both factors under the same criteria as previous experiments. From Figure 11, it is the combination of the fine-tuned model and templating that ensures higher quality sonnets than if only one factor is implemented. Our poems with both factors outperform both sets of ablative poems with varying statistical significance. Specifically, providing templates is clearly the critical piece to generate poems of a high caliber. See more in A.6. + +# 5 Conclusion + +We propose a novel method for generating high-quality poems that uses POS templating to determine a logical syntactical structure and rigorously + +maintains constraints necessary for any sonnet. Our method is highly versatile, with poetic factors like alliteration, internal rhyme, repetition, and theme adjustable to ensure creative output. After extensive surveys conducted with expert evaluators and MTurk participants, our model's success over similar competitors is evident, though our model's poems, like those of most computer poetry generators, remain distinguishable from human written poems. + +While we were unable to compare our model's performance to that of ChatGPT, our finetuned GPT-2 requires far less computing power than subsequent GPT models. Additionally, while we commenced this project's evaluation prior to the release of ChatGPT, after a preliminary qualitative evaluation, ChatGPT seems to produce very generic poetry (see A.7). Thus, for this particular application, our model may be a viable method that is more cost-effective and produces relatively high-quality sonnets. + +# Limitations + +Though our method produces full sonnets that are more impressive than all previous approaches, it + +is still not at the level of human-generated poetry. It is not clear how to achieve this level, whether it would be using massive large language models, or through our general approach, which is to bend those models around an interpretable framework that knows the rules that sonnets obey. Certainly our approach requires a lot less data – even if one used all the sonnets that have ever been written to train a language model, it is unclear that the language model would learn the very specific rules required of sonnets. However, there may be other ways to obtain these constraints that have not yet been developed. + +# Ethics Statement + +As with all neural generation, there are concerns about misinformation and generating toxic text. These concerns apply to some degree to poetry generation, although our rigidly constrained approach and limited vocabulary should mitigate this. + +# References + +John Benhardt, Peter Hase, Liuyi Zhu, and Cynthia Rudin. 2018. Shall I compare thee to a machine-written sonnet? An approach to algorithmic sonnet generation. +Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. 2017. Enriching word vectors with subword information. Transactions of the association for computational linguistics, 5:135-146. +Carnegie Mellon University CMU. 2019. The CMU pronouncing dictionary. http://www.speech.cs.cmu.edu/cgi-bin/cmudict, Internet. +Pablo Gervás. 2000. Wasp: Evaluation of different strategies for the automatic generation of spanish verse. In Proceedings of the AISB-00 Symposium on Creative & Cultural Aspects of AI, pages 93-100. +Marjan Ghazvininejad, Xing Shi, Yejin Choi, and Kevin Knight. 2016. Generating topical poetry. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1183-1191. +Jey Han Lau, Trevor Cohn, Timothy Baldwin, Julian Brooke, and Adam Hammond. 2018. Deep-spare: A joint neural model of poetic language, meter and rhyme. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1948-1958, Melbourne, Australia. Association for Computational Linguistics. +Ruli Manurung, Graeme Ritchie, and Henry Thompson. 2000. Towards a computational model of poetry generation. https://era.ed.ac.uk/handle/1842/3460. + +George A Miller. 1998. WordNet: An electronic lexical database. MIT press. +Hugo Gonçalo Oliveira. 2012. Poetry: a versatile platform for poetry generation. Computational Creativity, Concept Invention, and General Intelligence, 1:21. +Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners. https://github.com/openai/gpt-2. +Tim Van de Cruys. 2020. Automatic poetry generation from prosaic text. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2471-2480, Online. Association for Computational Linguistics. +Tony Veale. 2013. Less rhyme, more reason: Knowledge-based poetry generation with feeling, insight and wit. In Proceedings of the Fourth International Conference on Computational Creativity, ICCC 2013, Sidney, Australia, June 12-14, 2013, pages 152-159. computationalcreativity.net. +Jianyou Wang, Xiaoxuan Zhang, Yuren Zhou, Christopher Suh, and Cynthia Rudin. 2021. There once was a really bad poet, it was automated but you didn't know it. + +# A Appendix + +# A.1 Templating Mechanism + +Figure 8 presents more examples of our templating mechanism. We combine an adapted version of the Penn Treebank Project's part of speech tags along with articles, conjunctions, prepositions, and other filler words to construct these templates. Additionally, we provide the stress pattern of the syllables to ensure that the constraint of iambic pentameter is met. However, outside of the pre-determined filler words, POS do not have to directly adhere to the given stress pattern in splitting up words. For instance, in the first template, the provided syllable stress indicates that the JJ tag (adjective) should have two syllables, while the final VB tag (verb) should have only one syllable. However, the generated line ends with a monosyllabic adjective and a bisyllabic verb. As long as the stressing of the syllables aligns properly, each word can vary in its number of syllables. This is also visible in the fourth template example in Figure 8. + +# A.2 Elaboration on Experimental Competitors + +Benhardt et al. (2018), referred to as Benhardt et al., uses a RNN to preselect rhyming words and + +restrict different parts of speech to fit within the sonnet format. Oliveira (2012), referred to as Co-PoetryMe, is a versatile platform using semantic and grammar templates to alter the type of poem, input words, and "surprise" factor generated. + +# A.3 Experimental Procedure + +For each pair of sonnets, respondents were asked to indicate whether Sonnet A or Sonnet B performed better based on factors such as adherence to the inputted theme, poeticness, grammatical correctness, ability to convey emotion, and likelihood of being written by a human. Available answer choices and their corresponding numeric scores from 1 to 5 were "Definitely A" (5), "Probably A" (4), "The same" (3), "Probably B" (2), and "Definitely B" (1). Both our sonnet and the competing model-human-sonnet had equal probability of being either sonnet A or sonnet B in each pair. To analyze this data, user inputs were translated into numeric scoring values corresponding to our model's sonnet being Sonnet A (i.e. if our sonnet is presented as B to the user, a response of "Definitely B" corresponds to a score of 5, "Probably B" corresponds to 4, "Probably A" corresponds to 2, and "Definitely A" corresponds to 1). Additionally, respondents were asked to answer sanity check questions to filter out respondents who answer illogically or who do not have a sufficient grasp of English grammar. This setup remained the same across all experiments, and an additional space was allocated for expert evaluators to leave qualitative comments on sonnet quality. Sample sonnet evaluation questions are visible in Figure 9. + +After calculating the mean and standard deviation for scores across sonnets, we can immediately see whether our model performed better (an average score of $>3$ ) or worse (an average score of $< 3$ ) than the competitor in each respective category. We then performed a series of t-tests to establish these results' statistical significance. For factors that indicated our model performed better, we performed a right-tailed t-test (with the null-hypothesis as our model performed worse than the baseline), and we performed a left-tailed t-test for the remaining factors (with the null-hypothesis as our model performed better than the baseline). + +# A.4 Expert Evaluation Analysis + +In the expert evaluation, we emailed faculty at an American academic English department to recruit six faculty members and students to take our survey + +without payment. While we showed strong performance against the other computer-generated poems, we are consistently outperformed by human-written poems in all categories. Weaker performance on theme in experimental results may be explained by competitors' more frequent inclusion of the user-inputted theme word. For instance, in the expert evaluation, between two poems generated with the theme word "forest" (see Figure 10), one survey respondent states, "Sonnet B repeats forest too much for my taste," subsequently giving our model a 5 in each of poeticness, grammar, emotion, and humanness, yet a 2 in theme. + +# A.5 Amazon MTurk Analysis + +In our evaluation using Amazon MTurk Services, we requested survey respondents from primarily English-speaking countries and with an approval rate of $\geq 95\%$ . Crowdworkers were paid through the Amazon MTurk platform for this survey that on average took less than 30 minutes to complete. The questions and formatting remained the same as the expert evaluation, except no space was provided for qualitative feedback. + +Based on Figure 4 there is enough statistical significance to conclude that our sonnets outperform PoeTryMe in poetic, grammar, emotion, and human categories $(p < 0.001)$ . Against Benhardt et al., there is enough statistical significance to conclude that our sonnets perform better in grammar $(p < 0.001)$ , and perform slightly better with weak statistical significance in emotion $(p < 0.15)$ . Against human-written sonnets, the p-values for poetic, emotion, and even human categories are too large to strongly reject the null hypothesis that our model performed better than the baseline. Additionally, while the p-value indicates that this value is not statistically significant, it is interesting to note that our poems on average scored better in the grammar category. + +# A.6 Ablation Analysis + +In our ablation analysis, we replicate the Amazon MTurk analysis yet replace the competitor/human-written sonnets with poems generated with either the fine-tuned GPT-2 model without templating or the GPT-2 model without fine-tuning and with templating. This lets us test the individual efficacy of each factor (templating and fine-tuning GPT-2) against our method implementing both. Against poems generated with the fine-tuned GPT-2 and no templating, our sonnets performed better across + +all categories, and we can strongly reject the null hypothesis that our model performed worse than the baseline $(p < 0.0001)$ . Against the poems generated with the GPT-2 model without fine-tuning and with templates, we can conclude with high statistical significance $(p < 0.01)$ that we performed better in emotion, and conclude with weak statistical significance $(p < 0.10)$ that we performed better in grammar and theme. These results indicate that our method is successful due to its usage of both the fine-tuned GPT-2 model and templating. + +# A.7 ChatGPT Qualitative Analysis + +While we did not have time to extensively evaluate the quality of our sonnets against those of ChatGPT, after generating several sonnets to test ChatGPT's sonnet quality, it seems as though this language model generates relatively generic, non-cohesive sonnets even with different parameters. For instance, in Figure 7, both of the sonnets are unable to cohesively connect these three topics along a reasonable storyline. Additionally, Sonnet A in particular seems to dedicate a single stanza to each of the three topics passed in, hardly attempting to connect them. Of course, with more intensive prompt engineering, it is possible to generate a sonnet more tailored to one's preference. However, even this short analysis demonstrates there are clearly still strides to be made in the field of automatic poetry generation even with the advent of ChatGPT. + +
Sonnet A
In realms of words, where novels find their reign, My friend, a skeptic, shuns the written tales. No pages turned, no fictional campaigns, Averse to narratives that time unveils. Yet in his hand, a pear's sweet succulence, A juicy feast, a taste of nature's grace. Each bite, a moment of pure indulgence, A salute's joy, a respite from life's race. But beyond fruit's delight, his spirit soars, Within the bounds of a duelist's domain. With sword in hand, he dances, parries, scores, A noble art where honor's stance remains. Though novels may elude his curious gaze, Through pears and fencing, passions find their ways.
Sonnet B
In realms of love, where rom-coms dance their waltz, My friend, a rebel, strays from silver screens. With flannel-clad allure, he stands, exalts, A soul untouched by scripted love's routines. No tears shed for the tales of starry-eyed, His heart resists the charms of love's façade. Yet stitches formed with yarn, his hands provide, Creating warmth, affection neatly clad. Through threads of patience, vibrant patterns bloom, His nimble fingers weave a tapestry. Each crochet stitch, a tale of friendship's loom, A testament to artistry set free. Though rom-coms fall short of his embrace, In flannel and crochet, love finds its place.
+ +Figure 7: Comparison of two sonnets generated with ChatGPT. Sonnet A was generated with the prompt "generate a sonnet about my friend who hates novels, eats pears, and fences" and Sonnet B was generated with the prompt "generate a sonnet about my friend who hates rom-coms, wears flannels, and crochets." + +
TemplateSyllable StressExample Line
Where all the NNS of PRPD$ JJ NNS VB.0 1 0 1 0 1 01 0 1“Where all the gods of their past lives dictate”
And it VBD ABNN to the NN0 1 0 10 1 0 101“And it seemed evil to the enterprise”
Between the VBG and the VBG NN01 0 10 1 0 10 1“Between the glistening and the dying muse”
A JJ NN from the JJ NN0 10 10 1 0 1 01“A little lightness from the earthy sky”
Upon PRPO, PRPD$ NN POS NN01 01 0 10 101“Upon you, your life’s possibility”
Why VBC PRPS VBG such a JJ NN?0 1 0 10 1 0 101 0"Why do you squander such a precious thing?"
The NNS of ABNN, the NN on the NN0 1 0 1 0 10 1 0 1“The ghosts of death, the spirit on the earth”
+ +The key word for both of these poems is "wisdom." Which poem best adheres to this theme? + +Figure 8: Template examples, their corresponding syllable stress in order to adhere to iambic pentameter, and a sample line generated using the template. + +
Definitely AProbably ASameProbably BDefinitely B
+ +Which poem sounds more poetic? + +
Definitely AProbably ASameProbably BDefinitely B
+ +Which poem is more grammatically correct? + +
Definitely AProbably ASameProbably BDefinitely B
+ +Which poem conveys emotions more effectively? + +
Definitely AProbably ASameProbably BDefinitely B
+ +Which poem is more likely to be written by a human? + +
Definitely AProbably ASameProbably BDefinitely B
+ +Figure 9: Survey questions presented for each pair of sonnets. + +
Sonnet A: Our Code
I was aghast to see the fireflies
Inflamed soothed toads, where there the dead boughs lay
And it seemed evil to the enterprise
The hag I had, the hag, the hog, the gray.
But I knew to my painless fireflies
And beauty was a kind and loving thing.
My life's light isle so longed on otherwise
So too my fireflies bloomed to my king.
Those eagles that with auburn hair flew oaks,
Beauty and beauty beamed within the air
Which made oasis overcomes to coax?
So too my hogs beheaded to my lair.
The windy night was in the mistletoe
And wept soiled toads in my dream's studio.
Sonnet B: PoetryMe
forest some more and reforest a trip!
in deserts where heavenly woodlands clink
many, many, many clustered before
come: not in establishments of the floor
the fields of agony, the endless circumstance
findings to lie to interrupt your earth
with summation and set, triumph and agony
floors of horror forest before my eyes
those that study clustered plant are psychologists
taking over my ness a second forest
an' you've got to forest them reforest
on every forest, indeed, that rainforests
and grounds of forest coming to accord
floor of establishments and lilt of sing
+ +Figure 10: Comparison of two sonnets generated with theme word "forest". Sonnet A was generated with our code, and Sonnet B was generated using PoeTryMe. + +Ablation Evaluation + +
CategoryMeanp-valueMeanp-value
Grammar3.51*5.10×10-53.21*0.06
Emotion3.61*9.00×10-63.40*3.89×10-3
Poetic3.61*4.00×10-63.09*0.29
Human3.66*1.00×10-63.01*0.46
Theme3.50*8.00×10-53.20*0.06
+ +Figure 11: Left: fine-tuned GPT-2 with no templates. Right: GPT-2 without fine-tuning, but with templates. Starred figures indicate average scores of $>3$ , and underlined figures indicate that the p-value is low enough $(<0.05)$ to claim that this higher average is statistically significant. + +A For every submission: + +A1. Did you describe the limitations of your work? Limitations +A2. Did you discuss any potential risks of your work? Ethics +A3. Do the abstract and introduction summarize the paper's main claims? Abstract, 1 +A4. Have you used AI writing assistants when working on this paper? Left blank. + +B Did you use or create scientific artifacts? + +3,4 + +B1. Did you cite the creators of artifacts you used? 3.2,3.3,References +B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Not applicable. Left blank. +B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Not applicable. Left blank. +B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Data used from publicly available sonnets/ poems were assumed to be not subject to dispute. +B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Not applicable. Left blank. +B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. 3.3 + +C Did you run computational experiments? + +Left blank. + +C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? No response. + +C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? No response. +C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? No response. +C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? No response. + +D Did you use human annotators (e.g., crowdworkers) or research with human participants? 4,4.1,4.2,4.3 + +D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Appendix +D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? A.5.A.6 +D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? We do not believe having data on poetry evaluation raises any ethical issues. +D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? We do not believe having crowdworkers evaluate the same poems that were given to English professors raises any ethical issues. +D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? A.6 \ No newline at end of file diff --git a/2023/The Mechanical Bard_ An Interpretable Machine Learning Approach to Shakespearean Sonnet Generation/images.zip b/2023/The Mechanical Bard_ An Interpretable Machine Learning Approach to Shakespearean Sonnet Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..eefcfa3bb5caa625582f20c3b7f3962e078b1e85 --- /dev/null +++ b/2023/The Mechanical Bard_ An Interpretable Machine Learning Approach to Shakespearean Sonnet Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a5e49fd63c27a8f8a55d930abffac4381982f665bea67d26b638a9f1fba22e95 +size 642243 diff --git a/2023/The Mechanical Bard_ An Interpretable Machine Learning Approach to Shakespearean Sonnet Generation/layout.json b/2023/The Mechanical Bard_ An Interpretable Machine Learning Approach to Shakespearean Sonnet Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0738bf9e9c956fcd8b15e5d3efce4f9a2ab4e80b --- /dev/null +++ b/2023/The Mechanical Bard_ An Interpretable Machine Learning Approach to Shakespearean Sonnet Generation/layout.json @@ -0,0 +1,8775 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 74, + 75, + 519, + 110 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 75, + 519, + 110 + ], + "spans": [ + { + "bbox": [ + 74, + 75, + 519, + 110 + ], + "type": "text", + "content": "The Mechanical Bard: An Interpretable Machine Learning Approach to Shakespearean Sonnet Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 111, + 126, + 485, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 126, + 485, + 140 + ], + "spans": [ + { + "bbox": [ + 111, + 126, + 485, + 140 + ], + "type": "text", + "content": "Edwin Agnew*, Michelle Qiu*, Lily Zhu*, Sam Wiseman, Cynthia Rudin" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 223, + 141, + 373, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 141, + 373, + 153 + ], + "spans": [ + { + "bbox": [ + 223, + 141, + 373, + 153 + ], + "type": "text", + "content": "Duke University, Durham, NC" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 154, + 485, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 154, + 485, + 168 + ], + "spans": [ + { + "bbox": [ + 111, + 154, + 485, + 168 + ], + "type": "text", + "content": "edwin.agnew@duke.edu, michelle.qiu@duke.edu, lily.zhu@duke.edu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 175, + 169, + 422, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 169, + 422, + 181 + ], + "spans": [ + { + "bbox": [ + 175, + 169, + 422, + 181 + ], + "type": "text", + "content": "swiseman@cs.duke.edu, cynthiaa@cs.duke.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 155, + 212, + 202, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 212, + 202, + 225 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 202, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 84, + 236, + 274, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 236, + 274, + 391 + ], + "spans": [ + { + "bbox": [ + 84, + 236, + 274, + 391 + ], + "type": "text", + "content": "We consider the automated generation of sonnets, a poetic form constrained according to meter, rhyme scheme, and length. Sonnets generally also use rhetorical figures, expressive language, and a consistent theme or narrative. Our constrained decoding approach allows for the generation of sonnets within preset poetic constraints, while using a relatively modest neural backbone. Human evaluation confirms that our approach produces Shakespearean sonnets that resemble human-authored sonnets, and which adhere to the genre's defined constraints and contain lyrical language and literary devices." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 337, + 230, + 492, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 230, + 492, + 241 + ], + "spans": [ + { + "bbox": [ + 337, + 230, + 492, + 241 + ], + "type": "text", + "content": "When all the lovers of this world are dead," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 351, + 241, + 478, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 241, + 478, + 250 + ], + "spans": [ + { + "bbox": [ + 351, + 241, + 478, + 250 + ], + "type": "text", + "content": "The sun of heaven on a golden day" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 329, + 251, + 499, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 251, + 499, + 260 + ], + "spans": [ + { + "bbox": [ + 329, + 251, + 499, + 260 + ], + "type": "text", + "content": "To burn the earth's fire by the flame and spread" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 336, + 261, + 492, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 261, + 492, + 270 + ], + "spans": [ + { + "bbox": [ + 336, + 261, + 492, + 270 + ], + "type": "text", + "content": "Where all the flowers of your fair days lay." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 333, + 270, + 495, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 270, + 495, + 280 + ], + "spans": [ + { + "bbox": [ + 333, + 270, + 495, + 280 + ], + "type": "text", + "content": "These are the blossoms that you take care of." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 345, + 280, + 483, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 280, + 483, + 290 + ], + "spans": [ + { + "bbox": [ + 345, + 280, + 483, + 290 + ], + "type": "text", + "content": "Why do you linger such a long delay?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 340, + 290, + 489, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 290, + 489, + 300 + ], + "spans": [ + { + "bbox": [ + 340, + 290, + 489, + 300 + ], + "type": "text", + "content": "Forgive the fluttered flower of meek love" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 339, + 301, + 489, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 301, + 489, + 310 + ], + "spans": [ + { + "bbox": [ + 339, + 301, + 489, + 310 + ], + "type": "text", + "content": "Or who you have so long to love the day?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 342, + 311, + 487, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 311, + 487, + 320 + ], + "spans": [ + { + "bbox": [ + 342, + 311, + 487, + 320 + ], + "type": "text", + "content": "The joys of love, the beauty on the face," + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 336, + 321, + 492, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 321, + 492, + 330 + ], + "spans": [ + { + "bbox": [ + 336, + 321, + 492, + 330 + ], + "type": "text", + "content": "Shall be your fate and be your own delight." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 337, + 331, + 491, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 331, + 491, + 339 + ], + "spans": [ + { + "bbox": [ + 337, + 331, + 491, + 339 + ], + "type": "text", + "content": "You have the beauty of your own embrace." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 342, + 340, + 487, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 340, + 487, + 349 + ], + "spans": [ + { + "bbox": [ + 342, + 340, + 487, + 349 + ], + "type": "text", + "content": "You cannot reminiscence. Cannot write." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 333, + 350, + 495, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 350, + 495, + 359 + ], + "spans": [ + { + "bbox": [ + 333, + 350, + 495, + 359 + ], + "type": "text", + "content": "Between the living and the deadening breath." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 344, + 360, + 484, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 360, + 484, + 370 + ], + "spans": [ + { + "bbox": [ + 344, + 360, + 484, + 370 + ], + "type": "text", + "content": "You go the way of your beloved death." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 385, + 521, + 397 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 385, + 521, + 397 + ], + "spans": [ + { + "bbox": [ + 305, + 385, + 521, + 397 + ], + "type": "text", + "content": "Figure 1: A sonnet generated with the theme \"death\"." + } + ] + } + ], + "index": 21, + "type": "text" + }, + { + "bbox": [ + 68, + 412, + 154, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 412, + 154, + 426 + ], + "spans": [ + { + "bbox": [ + 68, + 412, + 154, + 426 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 67, + 434, + 291, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 434, + 291, + 528 + ], + "spans": [ + { + "bbox": [ + 67, + 434, + 291, + 528 + ], + "type": "text", + "content": "We consider the task of automatically generating Shakespearean sonnets, a popular poetic form with highly specific rhyme and meter constraints1. Each sonnet consists of three quatrains followed by a single couplet according to the rhyme scheme ABAB BCBC CDCD EE, and each line contains ten syllables with a stress pattern of iambic pentameter." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 67, + 530, + 291, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 530, + 291, + 597 + ], + "spans": [ + { + "bbox": [ + 67, + 530, + 291, + 597 + ], + "type": "text", + "content": "Rather than train a model to obey these constraints implicitly (which leads to enormous models that still do not obey the constraints), we opt to enforce them explicitly using a simple but novel approach to generation." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 67, + 597, + 291, + 732 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 597, + 291, + 732 + ], + "spans": [ + { + "bbox": [ + 67, + 597, + 291, + 732 + ], + "type": "text", + "content": "In particular, we use part-of-speech (POS) templates selected and edited from individual lines in Shakespeare's sonnets, with each template intended to offer a different combination of parts of speech and narrative directions. Associated thematic words are then selected and placed at the end of each template, and their rhyming pairs are chosen dynamically by a language model (e.g., GPT-2, Radford et al., 2019) and placed at the end of the corresponding lines according to the rhyme scheme." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 302, + 425, + 525, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 425, + 525, + 465 + ], + "spans": [ + { + "bbox": [ + 302, + 425, + 525, + 465 + ], + "type": "text", + "content": "The rest of the line is filled with related words that fit the specified POS and meter, leading to the end rhyme word. Figure 1 shows sample output." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 302, + 467, + 526, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 467, + 526, + 548 + ], + "spans": [ + { + "bbox": [ + 302, + 467, + 526, + 548 + ], + "type": "text", + "content": "Our use of these templates ensures sophisticated-seeming language and syntax that competing systems do not capture. Our approach provides excellent grammatical structure comparable to that of human-written poetry, all while using a relatively simple model and generation procedure." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 302, + 549, + 525, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 549, + 525, + 710 + ], + "spans": [ + { + "bbox": [ + 302, + 549, + 525, + 710 + ], + "type": "text", + "content": "We extensively evaluate the ability of our approach to generate whole sonnets (a setting often ignored by recent work in poetry generation) and find that our approach is preferred over strong baselines by both expert annotators (recruited from an academic English department) and by crowdworkers. As this research was conducted before the release of ChatGPT, we were not able to robustly compare our model's performance against this language model. However, we make several observations about the poetic quality of sonnets generated by ChatGPT." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 302, + 723, + 396, + 735 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 723, + 396, + 735 + ], + "spans": [ + { + "bbox": [ + 302, + 723, + 396, + 735 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "type": "text", + "content": "Early attempts at poetry generation relied mainly on rule-based methods (Gervás, 2000; Oliveira," + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 81, + 740, + 182, + 751 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 740, + 182, + 751 + ], + "spans": [ + { + "bbox": [ + 81, + 740, + 182, + 751 + ], + "type": "text", + "content": "* denotes equal contribution" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 81, + 751, + 289, + 762 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 751, + 289, + 762 + ], + "spans": [ + { + "bbox": [ + 81, + 751, + 289, + 762 + ], + "type": "text", + "content": "1Our code is available at https://github.com/" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 69, + 762, + 182, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 762, + 182, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 762, + 182, + 772 + ], + "type": "text", + "content": "edwinagnew/Poetix_Sonnets" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "1627" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 135, + 795, + 458, + 806 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 795, + 458, + 806 + ], + "spans": [ + { + "bbox": [ + 135, + 795, + 458, + 806 + ], + "type": "text", + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 219, + 807, + 374, + 817 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 807, + 374, + 817 + ], + "spans": [ + { + "bbox": [ + 219, + 807, + 374, + 817 + ], + "type": "text", + "content": "Volume 2: Short Papers, pages 1627-1638" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "spans": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "text", + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 66, + 71, + 291, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 71, + 291, + 301 + ], + "spans": [ + { + "bbox": [ + 66, + 71, + 291, + 301 + ], + "type": "text", + "content": "2012; Manurung et al., 2000; Veale, 2013). More recent automated poetry generation techniques, especially for sonnet generation, have relied on combinations of task-specific language models and rules. For instance, Ghazvininejad et al. (2016)'s Hafez uses a finite state acceptor to generate a large number of possible lines, the best of which are then selected with an RNN trained on song lyrics. Like our approach, they use rhyming dictionaries to find rhyming words and word embeddings to find topical words. Similarly, Benhardt et al. (2018) preselects rhyming words and generates lines backwards with a recurrent neural network (RNN). Also in this vein are Lau et al. (2018)'s Deepspare, which consists of an LSTM language model, an iambic model, and a rhyming model, and the recent work of Van de Cruys (2020) and Wang et al. (2021)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 302, + 291, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 302, + 291, + 371 + ], + "spans": [ + { + "bbox": [ + 67, + 302, + 291, + 371 + ], + "type": "text", + "content": "Our approach distinguishes itself in using a general-purpose pretrained language model, but more importantly in its use of human-curated constraints and templates. These allow for generating high-quality poems with a very simple approach." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 381, + 157, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 381, + 157, + 395 + ], + "spans": [ + { + "bbox": [ + 67, + 381, + 157, + 395 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 405, + 292, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 405, + 292, + 526 + ], + "spans": [ + { + "bbox": [ + 67, + 405, + 292, + 526 + ], + "type": "text", + "content": "The general idea of our approach is to take a pretrained language model (in this case GPT-2) and apply hard constraints to the generation procedure so that it can only output text satisfying various poetic constraints. These constraints can be broadly divided into hard constraints (e.g., number of syllables) and soft constraints (e.g., sounding poetic), and our methodology can be separated similarly. Our generation process is in Figure 2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 538, + 168, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 538, + 168, + 550 + ], + "spans": [ + { + "bbox": [ + 67, + 538, + 168, + 550 + ], + "type": "text", + "content": "3.1 POS Templates" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 556, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 556, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 556, + 291, + 772 + ], + "type": "text", + "content": "The most important part of our method is the use of handcrafted grammar templates. Taking inspiration from existing sonnets, we created a list of about 120 templates that encode the part-of-speech structure of a line of poetry. Each template can generate an unbounded number of possible poetic lines. For example, the line \"The beauty of life on a lonely sea\" is represented by the template \"THE NN OF NN ON A JJ NN.\" More sample templates are in Section A.1. Since the templates allow for considerable flexibility, obeying the templates does not alone suffice for poetry. For example, the same template could be used to write poetic lines with distinct meanings such as \"The tree of anguish on a stormy night\" or a nonsensical line like \"The fork of ant on an unpacked transfer.\" A subset of these" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 71, + 503, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 503, + 84 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 503, + 84 + ], + "type": "text", + "content": "templates is also chosen for starting a stanza." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 94, + 450, + 105 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 94, + 450, + 105 + ], + "spans": [ + { + "bbox": [ + 302, + 94, + 450, + 105 + ], + "type": "text", + "content": "3.2 Strict Sonnet Constraints" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 301, + 111, + 526, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 111, + 526, + 286 + ], + "spans": [ + { + "bbox": [ + 301, + 111, + 526, + 286 + ], + "type": "text", + "content": "The two most critical features of sonnets distinguishing them from other poetry forms are that they are written in iambic pentameter (i.e., each line has 10 syllables of alternating stress pattern), and they follow an ABAB CDCD EFEF GG rhyme scheme. To detect iambic pentameter, we use the CMU Pronouncing Dictionary (CMU, 2019), which reveals how many syllables a word contains and the stress of each syllable. An unstressed syllable is represented as '0' and a stressed syllable as '1', and so the line \"The beauty of life on a lonely sea\" is represented as '0 10 1 0 1 0 10 1'. For simplicity, 1-syllable words can be designated as either 0 or 1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 287, + 527, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 287, + 527, + 450 + ], + "spans": [ + { + "bbox": [ + 302, + 287, + 527, + 450 + ], + "type": "text", + "content": "Given a POS-tag for every word in our dictionary, we create a tree-like data structure that represents every possible meter for a given template. Continuing the example, the first word could only be 'the', but the second word could be filled with a 1-syllable noun like 'tree', a 2-syllable noun like 'chaos' (10), or a 3-syllable noun like 'audio' (101), and so on. Each choice affects the possible pronunciations of the next word as well as the number of remaining words in order to reach 10 syllables. The pronunciation dictionary ensures the last syllable of the last word on each line matches its partner." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 459, + 411, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 459, + 411, + 471 + ], + "spans": [ + { + "bbox": [ + 302, + 459, + 411, + 471 + ], + "type": "text", + "content": "3.3 Language Model" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 476, + 526, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 476, + 526, + 678 + ], + "spans": [ + { + "bbox": [ + 302, + 476, + 526, + 678 + ], + "type": "text", + "content": "We use a language model to generate individual sonnet lines, subject to the formal constraints outlined above. In particular, we first fine-tune GPT-2 (Radford et al., 2019) on a large corpus of over 15000 poems and a smaller corpus of sonnets. We then use a constrained beam-search to generate each line, where only legal tokens (under the aforementioned constraints) can be generated at each step; this generation approach resembles previous constrained decoding techniques used in sonnet generation (Ghazvininejad et al., 2016), although our approach differs in the choice of model and direct enforcement of constraints. For a comparison of generation quality using a GPT-2 model that has not been fine-tuned, see Section 4.1." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 688, + 441, + 700 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 688, + 441, + 700 + ], + "spans": [ + { + "bbox": [ + 302, + 688, + 441, + 700 + ], + "type": "text", + "content": "3.4 Thematic Word Choice" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 706, + 525, + 733 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 706, + 525, + 733 + ], + "spans": [ + { + "bbox": [ + 302, + 706, + 525, + 733 + ], + "type": "text", + "content": "To ensure the content of the poem fits the theme specified by the user, we provide an excerpt of a" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 740, + 525, + 761 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 740, + 525, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 740, + 525, + 761 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 302, + 740, + 525, + 761 + ], + "type": "text", + "content": "https://www.kaggle.com/datasets/johnhallman/completpoetryfoundation.org-dataset" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 761, + 515, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 761, + 515, + 772 + ], + "spans": [ + { + "bbox": [ + 317, + 761, + 515, + 772 + ], + "type": "text", + "content": "3https://www.kaggle.com/datasets/michelleqiu/sonnets" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1628" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 81, + 274, + 374 + ], + "blocks": [ + { + "bbox": [ + 120, + 68, + 239, + 79 + ], + "lines": [ + { + "bbox": [ + 120, + 68, + 239, + 79 + ], + "spans": [ + { + "bbox": [ + 120, + 68, + 239, + 79 + ], + "type": "text", + "content": "Generation Visualization" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 83, + 81, + 274, + 374 + ], + "lines": [ + { + "bbox": [ + 83, + 81, + 274, + 374 + ], + "spans": [ + { + "bbox": [ + 83, + 81, + 274, + 374 + ], + "type": "image", + "image_path": "7f5699c3abb074e217a003ab4ce4e8055fa6be50e33b760657614f7eed444060.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 383, + 290, + 407 + ], + "lines": [ + { + "bbox": [ + 67, + 383, + 290, + 407 + ], + "spans": [ + { + "bbox": [ + 67, + 383, + 290, + 407 + ], + "type": "text", + "content": "Figure 2: Numbers in parentheses denote subsections in Section 3." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 431, + 291, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 431, + 291, + 581 + ], + "spans": [ + { + "bbox": [ + 67, + 431, + 291, + 581 + ], + "type": "text", + "content": "theme-appropriate poem as additional context to GPT-2 during generation. This additional poem is selected by finding a list of synonyms to the theme word using the WordNet synonym database (Miller, 1998) and then choosing lines from a poem corpus that contain at least one synonym. We also remove words from the vocabulary if they have less than 0.5 cosine similarity with the theme word, based on the corresponding FastText word embeddings (Bojanowski et al., 2017). This avoids having words like \"algebra\" in poems with themes like \"forest.\"" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 592, + 199, + 603 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 592, + 199, + 603 + ], + "spans": [ + { + "bbox": [ + 67, + 592, + 199, + 603 + ], + "type": "text", + "content": "3.5 Generation Procedure" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 611, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 611, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 611, + 291, + 772 + ], + "type": "text", + "content": "Having introduced our method's components, we now describe the generation procedure. A user inputs a theme word, a beam search parameter, " + }, + { + "bbox": [ + 67, + 611, + 291, + 772 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 67, + 611, + 291, + 772 + ], + "type": "text", + "content": ", and the number of templates sampled per line, " + }, + { + "bbox": [ + 67, + 611, + 291, + 772 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 611, + 291, + 772 + ], + "type": "text", + "content": ". A seed is chosen with the above method. Then for each line, we sample " + }, + { + "bbox": [ + 67, + 611, + 291, + 772 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 611, + 291, + 772 + ], + "type": "text", + "content": " random templates. For each template, we generate the line using a modified beam search. Specifically, the beam search maintains " + }, + { + "bbox": [ + 67, + 611, + 291, + 772 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 67, + 611, + 291, + 772 + ], + "type": "text", + "content": " different hypotheses per line at all times. For each hypothesis, we first mask out any tokens that violate our hard POS, meter, or rhyme constraints and select the " + }, + { + "bbox": [ + 67, + 611, + 291, + 772 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 67, + 611, + 291, + 772 + ], + "type": "text", + "content": " best next-tokens for each" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 71, + 526, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 220 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 220 + ], + "type": "text", + "content": "of the " + }, + { + "bbox": [ + 302, + 71, + 526, + 220 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 302, + 71, + 526, + 220 + ], + "type": "text", + "content": " templates. These " + }, + { + "bbox": [ + 302, + 71, + 526, + 220 + ], + "type": "inline_equation", + "content": "b^{2}" + }, + { + "bbox": [ + 302, + 71, + 526, + 220 + ], + "type": "text", + "content": " new candidates are re-ranked according to our custom scoring function, and the top " + }, + { + "bbox": [ + 302, + 71, + 526, + 220 + ], + "type": "inline_equation", + "content": "k \\times b" + }, + { + "bbox": [ + 302, + 71, + 526, + 220 + ], + "type": "text", + "content": " proceed to the next stage. The constraint-filtering at each stage guarantees that the generated line will match the input template, while the beam search allows more flexible word choice than greedy word-filling for each POS. If none of the " + }, + { + "bbox": [ + 302, + 71, + 526, + 220 + ], + "type": "inline_equation", + "content": "k \\times b" + }, + { + "bbox": [ + 302, + 71, + 526, + 220 + ], + "type": "text", + "content": " generated lines score better than a specific threshold, then a new template is chosen and the line is generated again. Otherwise, line generation continues until the poem is completed." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 231, + 398, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 231, + 398, + 243 + ], + "spans": [ + { + "bbox": [ + 302, + 231, + 398, + 243 + ], + "type": "text", + "content": "3.6 Poetic Devices" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 250, + 526, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 250, + 526, + 601 + ], + "spans": [ + { + "bbox": [ + 302, + 250, + 526, + 601 + ], + "type": "text", + "content": "To make the poems more poetic, we adjust our scoring function to weight lines with alliteration, penalties for repetition, and/or internal rhyme. Alliteration occurs when a line contains words starting with the same letter, repetition occurs when a word is present several times throughout a poem, and internal rhyme occurs when two words rhyme within the same line. To weight alliteration, when the first token of a new word is being generated, a list " + }, + { + "bbox": [ + 302, + 250, + 526, + 601 + ], + "type": "inline_equation", + "content": "\\vec{A} = [a_1,a_2,\\dots a_n]" + }, + { + "bbox": [ + 302, + 250, + 526, + 601 + ], + "type": "text", + "content": " is generated where " + }, + { + "bbox": [ + 302, + 250, + 526, + 601 + ], + "type": "inline_equation", + "content": "a_{i}" + }, + { + "bbox": [ + 302, + 250, + 526, + 601 + ], + "type": "text", + "content": " is the number of occurrences of the first letter of the ith token in the current line. To weight and discourage repetition, a list " + }, + { + "bbox": [ + 302, + 250, + 526, + 601 + ], + "type": "inline_equation", + "content": "\\vec{T} = [t_1,t_2,\\dots t_n]" + }, + { + "bbox": [ + 302, + 250, + 526, + 601 + ], + "type": "text", + "content": " is generated where " + }, + { + "bbox": [ + 302, + 250, + 526, + 601 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 302, + 250, + 526, + 601 + ], + "type": "text", + "content": " is the number of occurrences of the ith token in the poem, negated. To weight internal rhyme, a list " + }, + { + "bbox": [ + 302, + 250, + 526, + 601 + ], + "type": "inline_equation", + "content": "\\vec{R} = [r_1,r_2,\\dots ,r_n]" + }, + { + "bbox": [ + 302, + 250, + 526, + 601 + ], + "type": "text", + "content": " is generated where " + }, + { + "bbox": [ + 302, + 250, + 526, + 601 + ], + "type": "inline_equation", + "content": "r_i = 1" + }, + { + "bbox": [ + 302, + 250, + 526, + 601 + ], + "type": "text", + "content": " if the ith token is part of a word that rhymes with any of the words in the current line generated so far, and " + }, + { + "bbox": [ + 302, + 250, + 526, + 601 + ], + "type": "inline_equation", + "content": "r_i = 0" + }, + { + "bbox": [ + 302, + 250, + 526, + 601 + ], + "type": "text", + "content": " otherwise. The final token distribution is then proportional to " + }, + { + "bbox": [ + 302, + 250, + 526, + 601 + ], + "type": "inline_equation", + "content": "\\tilde{P} +\\alpha_{A}\\times \\vec{A} +\\alpha_{T}\\times \\vec{T} +\\alpha_{R}\\times \\vec{R}," + }, + { + "bbox": [ + 302, + 250, + 526, + 601 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 302, + 250, + 526, + 601 + ], + "type": "inline_equation", + "content": "\\tilde{P}" + }, + { + "bbox": [ + 302, + 250, + 526, + 601 + ], + "type": "text", + "content": " is the language model's next-token distribution, and " + }, + { + "bbox": [ + 302, + 250, + 526, + 601 + ], + "type": "inline_equation", + "content": "\\alpha_{A},\\alpha_{T}," + }, + { + "bbox": [ + 302, + 250, + 526, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 302, + 250, + 526, + 601 + ], + "type": "inline_equation", + "content": "\\alpha_{R}" + }, + { + "bbox": [ + 302, + 250, + 526, + 601 + ], + "type": "text", + "content": " are user-specified non-negative parameters, which represent the degree to which alliteration, repetition, and internal rhyme should be favored during generation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 612, + 400, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 612, + 400, + 625 + ], + "spans": [ + { + "bbox": [ + 302, + 612, + 400, + 625 + ], + "type": "text", + "content": "3.7 Postprocessing" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 630, + 526, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 630, + 526, + 698 + ], + "spans": [ + { + "bbox": [ + 302, + 630, + 526, + 698 + ], + "type": "text", + "content": "After a poem is completed and all 14 lines score above a fixed threshold, a small number of adjustments are made. These include fixing common mistakes made by GPT-2 like not capitalizing the word 'I' and not capitalizing following punctuation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 710, + 390, + 724 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 710, + 390, + 724 + ], + "spans": [ + { + "bbox": [ + 302, + 710, + 390, + 724 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 733, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 733, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 733, + 526, + 772 + ], + "type": "text", + "content": "We used human input to test our sonnets against both model-generated and human-written sonnets. To test adherence to a theme throughout a son" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1629" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 113, + 68, + 244, + 318 + ], + "blocks": [ + { + "bbox": [ + 113, + 68, + 244, + 318 + ], + "lines": [ + { + "bbox": [ + 113, + 68, + 244, + 318 + ], + "spans": [ + { + "bbox": [ + 113, + 68, + 244, + 318 + ], + "type": "table", + "html": "
CategoryMeanp-value
PoeTryMe
Grammar4.50*1.71×10-4
Emotion4.30*3.13×10-3
Poetic4.30*3.13×10-3
Human4.10*5.77×10-3
Theme2.600.211286
Benhardt et al.
Grammar3.83*0.03
Emotion3.67*0.05
Poetic3.75*0.04
Human3.75*0.02
Theme2.420.06
Human-written poems
Grammar1.361.00×10-6
Emotion1.45.00×10-6
Poetic1.645.40×10-5
Human1.361.00×10-6
Theme1.577.70×10-5
", + "image_path": "ded253baca0ff113755368f27b80d531e62e6c9281fb37033d28e0e5c59124a2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 332, + 291, + 381 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 332, + 291, + 381 + ], + "spans": [ + { + "bbox": [ + 67, + 332, + 291, + 381 + ], + "type": "text", + "content": "Figure 3: Starred figures indicate average scores of " + }, + { + "bbox": [ + 67, + 332, + 291, + 381 + ], + "type": "inline_equation", + "content": ">3" + }, + { + "bbox": [ + 67, + 332, + 291, + 381 + ], + "type": "text", + "content": ", and underlined figures indicate that the p-value is low enough " + }, + { + "bbox": [ + 67, + 332, + 291, + 381 + ], + "type": "inline_equation", + "content": "(<0.05)" + }, + { + "bbox": [ + 67, + 332, + 291, + 381 + ], + "type": "text", + "content": " to claim that this higher average is statistically significant." + } + ] + } + ], + "index": 1, + "type": "text" + }, + { + "bbox": [ + 67, + 404, + 290, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 404, + 290, + 484 + ], + "spans": [ + { + "bbox": [ + 67, + 404, + 290, + 484 + ], + "type": "text", + "content": "net, we desired baselines that generate whole sonnets with user-provided themes. This limited our competitors, as some generate discrete quatrains or generate without input themes (e.g., Deepspare), leaving only Benhardt et al. (2018) and PoeTryMe (Oliveira, 2012); see Section A.2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 486, + 290, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 486, + 290, + 553 + ], + "spans": [ + { + "bbox": [ + 67, + 486, + 290, + 553 + ], + "type": "text", + "content": "Furthermore, an evaluation of poetry quality is incomplete without human-written sonnets, selected from sonnets.org. Though these poems do not have an explicit theme, we selected poems that followed our five themes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 555, + 290, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 555, + 290, + 663 + ], + "spans": [ + { + "bbox": [ + 67, + 555, + 290, + 663 + ], + "type": "text", + "content": "To optimally test our model, we conducted an internal analysis and selected " + }, + { + "bbox": [ + 67, + 555, + 290, + 663 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 555, + 290, + 663 + ], + "type": "text", + "content": " values sampled from 3, 5, or 7, " + }, + { + "bbox": [ + 67, + 555, + 290, + 663 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 67, + 555, + 290, + 663 + ], + "type": "text", + "content": " values sampled from 3, 5, or 7, and repetition penalty values sampled from 1.4, 1.6, or 1.8 that we concluded produced the highest quality sonnets. To evaluate adherence to theme, we generated poems with themes \"death,\" \"darkness,\" \"forest,\" \"love,\" and \"wisdom.\"" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 665, + 290, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 665, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 665, + 290, + 772 + ], + "type": "text", + "content": "In each test, respondents compared six randomly selected pairs of sonnets, with each of our sonnets displayed with a competing model/human-written sonnet generated with the same theme word. Respondents indicated which of the two sonnets performed better in categories of theme, poeticness, grammar, emotion, and likelihood of being human-written. Detailed instructions are in A.3." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 347, + 68, + 480, + 312 + ], + "blocks": [ + { + "bbox": [ + 347, + 68, + 480, + 312 + ], + "lines": [ + { + "bbox": [ + 347, + 68, + 480, + 312 + ], + "spans": [ + { + "bbox": [ + 347, + 68, + 480, + 312 + ], + "type": "table", + "html": "
CategoryMeanp-value
PoeTryMe
Grammar3.66*2.00 × 10-6
Emotion3.54*1.16 × 10-4
Poetic3.55*3.70 × 10-5
Human3.59*1.60 × 10-5
Theme2.860.19
Benhardt et al.
Grammar3.34*6.57 × 10-3
Emotion3.16*0.12
Poetic3.11*0.19
Human3.06*0.33
Theme2.770.06
Human-written poems
Grammar3.13*0.14
Emotion2.860.14
Poetic2.910.24
Human2.920.27
Theme2.670.02
", + "image_path": "dc04809b9b89d272119f0de81e744c664234c519eed4438b27e6b326efa1e39b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 321, + 525, + 370 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 321, + 525, + 370 + ], + "spans": [ + { + "bbox": [ + 302, + 321, + 525, + 370 + ], + "type": "text", + "content": "Figure 4: Starred figures indicate average scores of " + }, + { + "bbox": [ + 302, + 321, + 525, + 370 + ], + "type": "inline_equation", + "content": ">3" + }, + { + "bbox": [ + 302, + 321, + 525, + 370 + ], + "type": "text", + "content": ", and underlined figures indicate that the p-value is low enough " + }, + { + "bbox": [ + 302, + 321, + 525, + 370 + ], + "type": "inline_equation", + "content": "(<0.05)" + }, + { + "bbox": [ + 302, + 321, + 525, + 370 + ], + "type": "text", + "content": " to claim that this higher average is statistically significant." + } + ] + } + ], + "index": 7, + "type": "text" + }, + { + "bbox": [ + 302, + 397, + 417, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 397, + 417, + 411 + ], + "spans": [ + { + "bbox": [ + 302, + 397, + 417, + 411 + ], + "type": "text", + "content": "4.1 Expert Evaluation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 301, + 417, + 525, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 417, + 525, + 565 + ], + "spans": [ + { + "bbox": [ + 301, + 417, + 525, + 565 + ], + "type": "text", + "content": "For an expert evaluation, we recruited six faculty members and students from an academic English department. Figures 3 and 5 show that we strongly outperform PoeTryMe in all categories but theme with high statistical significance " + }, + { + "bbox": [ + 301, + 417, + 525, + 565 + ], + "type": "inline_equation", + "content": "(p < 0.006)" + }, + { + "bbox": [ + 301, + 417, + 525, + 565 + ], + "type": "text", + "content": ", and we outperform Benhardt et al. in all poetic categories but theme and emotion with statistical significance " + }, + { + "bbox": [ + 301, + 417, + 525, + 565 + ], + "type": "inline_equation", + "content": "(p < 0.05)" + }, + { + "bbox": [ + 301, + 417, + 525, + 565 + ], + "type": "text", + "content": ". Notably, while we outperform other computer-generated poems, respondents could still distinguish between our poems and human-written sonnets quite easily. See more in A.4." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 578, + 459, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 578, + 459, + 590 + ], + "spans": [ + { + "bbox": [ + 302, + 578, + 459, + 590 + ], + "type": "text", + "content": "4.2 Amazon MTurk Evaluation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 301, + 597, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 597, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 301, + 597, + 526, + 772 + ], + "type": "text", + "content": "Along with expert evaluation, we used Amazon MTurk services to assess poems on a larger scale. Figures 4 and 6 show our superior performance against competitors in several categories. As expected of most computer-generated work, our poems failed to outperform human-written poems. However, we can only strongly conclude that the human-written poems are better in one category, theme. Our poems even outperformed human-written poems in grammar (albeit with low statistical significance), showing that our strictly constrained beam search generates high quality grammar. See more in A.5." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "text", + "content": "1630" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 86, + 146, + 133 + ], + "blocks": [ + { + "bbox": [ + 82, + 86, + 146, + 133 + ], + "lines": [ + { + "bbox": [ + 82, + 86, + 146, + 133 + ], + "spans": [ + { + "bbox": [ + 82, + 86, + 146, + 133 + ], + "type": "image", + "image_path": "8292ada2b3790048c4338774c85281dfbd1e018dce8050afac84f1d7d06c39fe.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 149, + 81, + 214, + 133 + ], + "blocks": [ + { + "bbox": [ + 135, + 68, + 223, + 80 + ], + "lines": [ + { + "bbox": [ + 135, + 68, + 223, + 80 + ], + "spans": [ + { + "bbox": [ + 135, + 68, + 223, + 80 + ], + "type": "text", + "content": "Expert Evaluation" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 149, + 81, + 214, + 133 + ], + "lines": [ + { + "bbox": [ + 149, + 81, + 214, + 133 + ], + "spans": [ + { + "bbox": [ + 149, + 81, + 214, + 133 + ], + "type": "image", + "image_path": "4b0713578ccefc477b272aaa9b622eef617b3bae4d06c5aa420cce41aa5e4519.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 215, + 86, + 276, + 133 + ], + "blocks": [ + { + "bbox": [ + 215, + 86, + 276, + 133 + ], + "lines": [ + { + "bbox": [ + 215, + 86, + 276, + 133 + ], + "spans": [ + { + "bbox": [ + 215, + 86, + 276, + 133 + ], + "type": "image", + "image_path": "28af1fff4996f1e72162e0878094ff4e419f53aa295c6e318d5eb132fcb1ae0c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 82, + 142, + 146, + 190 + ], + "blocks": [ + { + "bbox": [ + 82, + 142, + 146, + 190 + ], + "lines": [ + { + "bbox": [ + 82, + 142, + 146, + 190 + ], + "spans": [ + { + "bbox": [ + 82, + 142, + 146, + 190 + ], + "type": "image", + "image_path": "2c14d35ba10f132dcc548d4c42f4d72e869f9787f6ca1bda0d37409ed461e96d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 149, + 134, + 212, + 190 + ], + "blocks": [ + { + "bbox": [ + 149, + 134, + 212, + 190 + ], + "lines": [ + { + "bbox": [ + 149, + 134, + 212, + 190 + ], + "spans": [ + { + "bbox": [ + 149, + 134, + 212, + 190 + ], + "type": "image", + "image_path": "6672ec6bca3259931c113a548b9d1eaa8cfc504a903f4734e70a043bfd93d0e4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 214, + 142, + 276, + 190 + ], + "blocks": [ + { + "bbox": [ + 214, + 142, + 276, + 190 + ], + "lines": [ + { + "bbox": [ + 214, + 142, + 276, + 190 + ], + "spans": [ + { + "bbox": [ + 214, + 142, + 276, + 190 + ], + "type": "image", + "image_path": "c4297386128e901ffc60a136327959a030c1277cf1b17770a6468fdffd7c187d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 82, + 198, + 146, + 246 + ], + "blocks": [ + { + "bbox": [ + 82, + 198, + 146, + 246 + ], + "lines": [ + { + "bbox": [ + 82, + 198, + 146, + 246 + ], + "spans": [ + { + "bbox": [ + 82, + 198, + 146, + 246 + ], + "type": "image", + "image_path": "19ce2d9c4214eab638404a378bc61c99bf05a5f7ab00d09a673b02a997bfa246.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 149, + 193, + 212, + 246 + ], + "blocks": [ + { + "bbox": [ + 149, + 193, + 212, + 246 + ], + "lines": [ + { + "bbox": [ + 149, + 193, + 212, + 246 + ], + "spans": [ + { + "bbox": [ + 149, + 193, + 212, + 246 + ], + "type": "image", + "image_path": "1cedeb4682139532a00ed47d6a618d784b432148d73e260bc38c191332ea4af9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 214, + 199, + 276, + 246 + ], + "blocks": [ + { + "bbox": [ + 214, + 199, + 276, + 246 + ], + "lines": [ + { + "bbox": [ + 214, + 199, + 276, + 246 + ], + "spans": [ + { + "bbox": [ + 214, + 199, + 276, + 246 + ], + "type": "image", + "image_path": "67679e7d43a4d23aa706b89624fcb0f5c0eac2d494b3ad71613a2aee7cd4420e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 82, + 254, + 146, + 301 + ], + "blocks": [ + { + "bbox": [ + 82, + 254, + 146, + 301 + ], + "lines": [ + { + "bbox": [ + 82, + 254, + 146, + 301 + ], + "spans": [ + { + "bbox": [ + 82, + 254, + 146, + 301 + ], + "type": "image", + "image_path": "a6413b777f70c49d98b182b4faaadd35a417180f9006dc3c2fff64feae8c72de.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 149, + 248, + 212, + 301 + ], + "blocks": [ + { + "bbox": [ + 149, + 248, + 212, + 301 + ], + "lines": [ + { + "bbox": [ + 149, + 248, + 212, + 301 + ], + "spans": [ + { + "bbox": [ + 149, + 248, + 212, + 301 + ], + "type": "image", + "image_path": "e874758c6f482612209aa93d6c9a06fc26b5e577c3392b4b0746c8de4e4caf2c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 214, + 255, + 276, + 301 + ], + "blocks": [ + { + "bbox": [ + 214, + 255, + 276, + 301 + ], + "lines": [ + { + "bbox": [ + 214, + 255, + 276, + 301 + ], + "spans": [ + { + "bbox": [ + 214, + 255, + 276, + 301 + ], + "type": "image", + "image_path": "ba3ca8ec4341772c19fce7920aa726f46ef05c17ba537fc692de309bceff122c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 82, + 312, + 146, + 358 + ], + "blocks": [ + { + "bbox": [ + 82, + 312, + 146, + 358 + ], + "lines": [ + { + "bbox": [ + 82, + 312, + 146, + 358 + ], + "spans": [ + { + "bbox": [ + 82, + 312, + 146, + 358 + ], + "type": "image", + "image_path": "f64944439299e8d0a8d81b36cf7c9a0aa03fb92e7c74750ea93dca0a026941a9.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 373, + 291, + 422 + ], + "lines": [ + { + "bbox": [ + 67, + 373, + 291, + 422 + ], + "spans": [ + { + "bbox": [ + 67, + 373, + 291, + 422 + ], + "type": "text", + "content": "Figure 5: Values " + }, + { + "bbox": [ + 67, + 373, + 291, + 422 + ], + "type": "inline_equation", + "content": ">3" + }, + { + "bbox": [ + 67, + 373, + 291, + 422 + ], + "type": "text", + "content": " (green), " + }, + { + "bbox": [ + 67, + 373, + 291, + 422 + ], + "type": "inline_equation", + "content": "< 3" + }, + { + "bbox": [ + 67, + 373, + 291, + 422 + ], + "type": "text", + "content": " (red), and " + }, + { + "bbox": [ + 67, + 373, + 291, + 422 + ], + "type": "inline_equation", + "content": "= 3" + }, + { + "bbox": [ + 67, + 373, + 291, + 422 + ], + "type": "text", + "content": " (gray) denote that our poetry model performs better, the competitor performs better, and the poems performed similarly, respectively." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 149, + 304, + 212, + 358 + ], + "blocks": [ + { + "bbox": [ + 149, + 304, + 212, + 358 + ], + "lines": [ + { + "bbox": [ + 149, + 304, + 212, + 358 + ], + "spans": [ + { + "bbox": [ + 149, + 304, + 212, + 358 + ], + "type": "image", + "image_path": "09383ba0baf9877a9acf94df5644cf06a6304ed37969e444372aae77afb8a4a9.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 214, + 312, + 276, + 358 + ], + "blocks": [ + { + "bbox": [ + 214, + 312, + 276, + 358 + ], + "lines": [ + { + "bbox": [ + 214, + 312, + 276, + 358 + ], + "spans": [ + { + "bbox": [ + 214, + 312, + 276, + 358 + ], + "type": "image", + "image_path": "fe3563cea5094c2ee93e2b84e663513a968b99f5e4870193fdd3885f10d2e662.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 317, + 86, + 382, + 133 + ], + "blocks": [ + { + "bbox": [ + 317, + 86, + 382, + 133 + ], + "lines": [ + { + "bbox": [ + 317, + 86, + 382, + 133 + ], + "spans": [ + { + "bbox": [ + 317, + 86, + 382, + 133 + ], + "type": "image", + "image_path": "30ead587a2e210153662cc053b8634a4b29d455b2be031800d1023b8e3c90d9f.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 384, + 80, + 447, + 133 + ], + "blocks": [ + { + "bbox": [ + 349, + 68, + 480, + 80 + ], + "lines": [ + { + "bbox": [ + 349, + 68, + 480, + 80 + ], + "spans": [ + { + "bbox": [ + 349, + 68, + 480, + 80 + ], + "type": "text", + "content": "Amazon MTurk Evaluation" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 384, + 80, + 447, + 133 + ], + "lines": [ + { + "bbox": [ + 384, + 80, + 447, + 133 + ], + "spans": [ + { + "bbox": [ + 384, + 80, + 447, + 133 + ], + "type": "image", + "image_path": "194ef67c2c27e30692951db29732552a3e149e82029cf8563645d64756b08f36.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 449, + 86, + 511, + 133 + ], + "blocks": [ + { + "bbox": [ + 449, + 86, + 511, + 133 + ], + "lines": [ + { + "bbox": [ + 449, + 86, + 511, + 133 + ], + "spans": [ + { + "bbox": [ + 449, + 86, + 511, + 133 + ], + "type": "image", + "image_path": "5efdcd6dc2ada6062c20b898c74d88db7eaf0134ce3dfab175fb82f9516fe683.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 317, + 135, + 382, + 190 + ], + "blocks": [ + { + "bbox": [ + 317, + 135, + 382, + 190 + ], + "lines": [ + { + "bbox": [ + 317, + 135, + 382, + 190 + ], + "spans": [ + { + "bbox": [ + 317, + 135, + 382, + 190 + ], + "type": "image", + "image_path": "df37ccfd094514f4c522b43dca9c98320896473e400b160c0f34cbb93b4da269.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 384, + 135, + 447, + 190 + ], + "blocks": [ + { + "bbox": [ + 384, + 135, + 447, + 190 + ], + "lines": [ + { + "bbox": [ + 384, + 135, + 447, + 190 + ], + "spans": [ + { + "bbox": [ + 384, + 135, + 447, + 190 + ], + "type": "image", + "image_path": "ed4a3e6f4d49806be96b00477e9928dcd21683df6a77a912e8d2b8bc4c586a0f.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 449, + 135, + 511, + 190 + ], + "blocks": [ + { + "bbox": [ + 449, + 135, + 511, + 190 + ], + "lines": [ + { + "bbox": [ + 449, + 135, + 511, + 190 + ], + "spans": [ + { + "bbox": [ + 449, + 135, + 511, + 190 + ], + "type": "image", + "image_path": "554e63d71ce9123238f0b7ef4b3628b775dffe027121633a688cc331294fa0b6.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 317, + 193, + 382, + 246 + ], + "blocks": [ + { + "bbox": [ + 317, + 193, + 382, + 246 + ], + "lines": [ + { + "bbox": [ + 317, + 193, + 382, + 246 + ], + "spans": [ + { + "bbox": [ + 317, + 193, + 382, + 246 + ], + "type": "image", + "image_path": "c595f5131a009d57178fe00e1f55a2d400ddc39c282914aae7aadd2de606b644.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 384, + 193, + 447, + 245 + ], + "blocks": [ + { + "bbox": [ + 384, + 193, + 447, + 245 + ], + "lines": [ + { + "bbox": [ + 384, + 193, + 447, + 245 + ], + "spans": [ + { + "bbox": [ + 384, + 193, + 447, + 245 + ], + "type": "image", + "image_path": "fdee82545dc7ce52c6a9e180528600d93cefe993f4da71a5b653cb3e4a80e58a.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 449, + 193, + 511, + 246 + ], + "blocks": [ + { + "bbox": [ + 449, + 193, + 511, + 246 + ], + "lines": [ + { + "bbox": [ + 449, + 193, + 511, + 246 + ], + "spans": [ + { + "bbox": [ + 449, + 193, + 511, + 246 + ], + "type": "image", + "image_path": "cdfa9220bf60fd653a740d53f0910e146a9fdfd5919f7d1aed13c1cf83e54e45.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 317, + 254, + 382, + 301 + ], + "blocks": [ + { + "bbox": [ + 317, + 254, + 382, + 301 + ], + "lines": [ + { + "bbox": [ + 317, + 254, + 382, + 301 + ], + "spans": [ + { + "bbox": [ + 317, + 254, + 382, + 301 + ], + "type": "image", + "image_path": "ee4d023da1c0d5f2b54320b99eb0bdffa062a22dcc635478ccd7bf69d1c522da.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 384, + 248, + 447, + 301 + ], + "blocks": [ + { + "bbox": [ + 384, + 248, + 447, + 301 + ], + "lines": [ + { + "bbox": [ + 384, + 248, + 447, + 301 + ], + "spans": [ + { + "bbox": [ + 384, + 248, + 447, + 301 + ], + "type": "image", + "image_path": "848bb6ad99fc0f5fbe9ecbe895045891baecd155adf3f61f8b37f6ec2c865859.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 449, + 255, + 511, + 301 + ], + "blocks": [ + { + "bbox": [ + 449, + 255, + 511, + 301 + ], + "lines": [ + { + "bbox": [ + 449, + 255, + 511, + 301 + ], + "spans": [ + { + "bbox": [ + 449, + 255, + 511, + 301 + ], + "type": "image", + "image_path": "b44abebd499c22ccd92aa5b2dbbb98141d4069c8299ce8d5ba38a6c621dd6e3f.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 317, + 312, + 382, + 358 + ], + "blocks": [ + { + "bbox": [ + 317, + 312, + 382, + 358 + ], + "lines": [ + { + "bbox": [ + 317, + 312, + 382, + 358 + ], + "spans": [ + { + "bbox": [ + 317, + 312, + 382, + 358 + ], + "type": "image", + "image_path": "e0c7a83428e11c556c4e5f31cd846719b98b0998b25f291eed5ba3f4270e83df.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 373, + 526, + 422 + ], + "lines": [ + { + "bbox": [ + 302, + 373, + 526, + 422 + ], + "spans": [ + { + "bbox": [ + 302, + 373, + 526, + 422 + ], + "type": "text", + "content": "Figure 6: Values " + }, + { + "bbox": [ + 302, + 373, + 526, + 422 + ], + "type": "inline_equation", + "content": ">3" + }, + { + "bbox": [ + 302, + 373, + 526, + 422 + ], + "type": "text", + "content": " (green), " + }, + { + "bbox": [ + 302, + 373, + 526, + 422 + ], + "type": "inline_equation", + "content": "< 3" + }, + { + "bbox": [ + 302, + 373, + 526, + 422 + ], + "type": "text", + "content": " (red), and " + }, + { + "bbox": [ + 302, + 373, + 526, + 422 + ], + "type": "inline_equation", + "content": "= 3" + }, + { + "bbox": [ + 302, + 373, + 526, + 422 + ], + "type": "text", + "content": " (gray) denote that our poetry model performs better, the competitor performs better, and the poems performed similarly, respectively." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 384, + 304, + 447, + 358 + ], + "blocks": [ + { + "bbox": [ + 384, + 304, + 447, + 358 + ], + "lines": [ + { + "bbox": [ + 384, + 304, + 447, + 358 + ], + "spans": [ + { + "bbox": [ + 384, + 304, + 447, + 358 + ], + "type": "image", + "image_path": "ed0cb71915719919b394595e5793eeea15a98ace3b83d4297ae6fa2e56d37ea3.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 449, + 312, + 511, + 358 + ], + "blocks": [ + { + "bbox": [ + 449, + 312, + 511, + 358 + ], + "lines": [ + { + "bbox": [ + 449, + 312, + 511, + 358 + ], + "spans": [ + { + "bbox": [ + 449, + 312, + 511, + 358 + ], + "type": "image", + "image_path": "b16d15f4bb980d51422af2a12e796ad88f0a5adbaead5914bff4058e9e188ca1.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "bbox": [ + 67, + 445, + 189, + 457 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 445, + 189, + 457 + ], + "spans": [ + { + "bbox": [ + 67, + 445, + 189, + 457 + ], + "type": "text", + "content": "4.3 Ablative Evaluation" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 67, + 465, + 291, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 465, + 291, + 694 + ], + "spans": [ + { + "bbox": [ + 67, + 465, + 291, + 694 + ], + "type": "text", + "content": "We also conducted ablative studies showing the efficacy of two key elements of our method: line templates and the fine-tuned GPT-2 language model. We generated two sets of ablation poems: one with the fine-tuned GPT-2 and no templating, and one using the untrained GPT-2 model and templating. We then used Amazon MTurk services to test each set against poems generated with both factors under the same criteria as previous experiments. From Figure 11, it is the combination of the fine-tuned model and templating that ensures higher quality sonnets than if only one factor is implemented. Our poems with both factors outperform both sets of ablative poems with varying statistical significance. Specifically, providing templates is clearly the critical piece to generate poems of a high caliber. See more in A.6." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 67, + 708, + 146, + 721 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 708, + 146, + 721 + ], + "spans": [ + { + "bbox": [ + 67, + 708, + 146, + 721 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 67, + 732, + 290, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 732, + 290, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 732, + 290, + 773 + ], + "type": "text", + "content": "We propose a novel method for generating high-quality poems that uses POS templating to determine a logical syntactical structure and rigorously" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 302, + 443, + 526, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 443, + 526, + 565 + ], + "spans": [ + { + "bbox": [ + 302, + 443, + 526, + 565 + ], + "type": "text", + "content": "maintains constraints necessary for any sonnet. Our method is highly versatile, with poetic factors like alliteration, internal rhyme, repetition, and theme adjustable to ensure creative output. After extensive surveys conducted with expert evaluators and MTurk participants, our model's success over similar competitors is evident, though our model's poems, like those of most computer poetry generators, remain distinguishable from human written poems." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 302, + 565, + 526, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 565, + 526, + 713 + ], + "spans": [ + { + "bbox": [ + 302, + 565, + 526, + 713 + ], + "type": "text", + "content": "While we were unable to compare our model's performance to that of ChatGPT, our finetuned GPT-2 requires far less computing power than subsequent GPT models. Additionally, while we commenced this project's evaluation prior to the release of ChatGPT, after a preliminary qualitative evaluation, ChatGPT seems to produce very generic poetry (see A.7). Thus, for this particular application, our model may be a viable method that is more cost-effective and produces relatively high-quality sonnets." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 303, + 724, + 365, + 737 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 724, + 365, + 737 + ], + "spans": [ + { + "bbox": [ + 303, + 724, + 365, + 737 + ], + "type": "text", + "content": "Limitations" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 302, + 746, + 525, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 746, + 525, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 746, + 525, + 773 + ], + "type": "text", + "content": "Though our method produces full sonnets that are more impressive than all previous approaches, it" + } + ] + } + ], + "index": 41 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 308, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 308, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 308, + 791 + ], + "type": "text", + "content": "1631" + } + ] + } + ], + "index": 42 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 291, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 291, + 248 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 291, + 248 + ], + "type": "text", + "content": "is still not at the level of human-generated poetry. It is not clear how to achieve this level, whether it would be using massive large language models, or through our general approach, which is to bend those models around an interpretable framework that knows the rules that sonnets obey. Certainly our approach requires a lot less data – even if one used all the sonnets that have ever been written to train a language model, it is unclear that the language model would learn the very specific rules required of sonnets. However, there may be other ways to obtain these constraints that have not yet been developed." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 257, + 158, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 257, + 158, + 270 + ], + "spans": [ + { + "bbox": [ + 68, + 257, + 158, + 270 + ], + "type": "text", + "content": "Ethics Statement" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 279, + 292, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 279, + 292, + 346 + ], + "spans": [ + { + "bbox": [ + 67, + 279, + 292, + 346 + ], + "type": "text", + "content": "As with all neural generation, there are concerns about misinformation and generating toxic text. These concerns apply to some degree to poetry generation, although our rigidly constrained approach and limited vocabulary should mitigate this." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 369, + 127, + 381 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 369, + 127, + 381 + ], + "spans": [ + { + "bbox": [ + 68, + 369, + 127, + 381 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 387, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 68, + 387, + 291, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 387, + 291, + 433 + ], + "spans": [ + { + "bbox": [ + 68, + 387, + 291, + 433 + ], + "type": "text", + "content": "John Benhardt, Peter Hase, Liuyi Zhu, and Cynthia Rudin. 2018. Shall I compare thee to a machine-written sonnet? An approach to algorithmic sonnet generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 441, + 290, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 441, + 290, + 486 + ], + "spans": [ + { + "bbox": [ + 69, + 441, + 290, + 486 + ], + "type": "text", + "content": "Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. 2017. Enriching word vectors with subword information. Transactions of the association for computational linguistics, 5:135-146." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 494, + 291, + 538 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 494, + 291, + 538 + ], + "spans": [ + { + "bbox": [ + 69, + 494, + 291, + 538 + ], + "type": "text", + "content": "Carnegie Mellon University CMU. 2019. The CMU pronouncing dictionary. http://www.speech.cs.cmu.edu/cgi-bin/cmudict, Internet." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 547, + 290, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 547, + 290, + 592 + ], + "spans": [ + { + "bbox": [ + 69, + 547, + 290, + 592 + ], + "type": "text", + "content": "Pablo Gervás. 2000. Wasp: Evaluation of different strategies for the automatic generation of spanish verse. In Proceedings of the AISB-00 Symposium on Creative & Cultural Aspects of AI, pages 93-100." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 600, + 291, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 600, + 291, + 645 + ], + "spans": [ + { + "bbox": [ + 69, + 600, + 291, + 645 + ], + "type": "text", + "content": "Marjan Ghazvininejad, Xing Shi, Yejin Choi, and Kevin Knight. 2016. Generating topical poetry. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1183-1191." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 653, + 291, + 731 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 653, + 291, + 731 + ], + "spans": [ + { + "bbox": [ + 69, + 653, + 291, + 731 + ], + "type": "text", + "content": "Jey Han Lau, Trevor Cohn, Timothy Baldwin, Julian Brooke, and Adam Hammond. 2018. Deep-spare: A joint neural model of poetic language, meter and rhyme. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1948-1958, Melbourne, Australia. Association for Computational Linguistics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 739, + 290, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 739, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 739, + 290, + 772 + ], + "type": "text", + "content": "Ruli Manurung, Graeme Ritchie, and Henry Thompson. 2000. Towards a computational model of poetry generation. https://era.ed.ac.uk/handle/1842/3460." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 303, + 72, + 526, + 390 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 304, + 72, + 525, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 525, + 95 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 525, + 95 + ], + "type": "text", + "content": "George A Miller. 1998. WordNet: An electronic lexical database. MIT press." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 102, + 526, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 102, + 526, + 147 + ], + "spans": [ + { + "bbox": [ + 304, + 102, + 526, + 147 + ], + "type": "text", + "content": "Hugo Gonçalo Oliveira. 2012. Poetry: a versatile platform for poetry generation. Computational Creativity, Concept Invention, and General Intelligence, 1:21." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 155, + 526, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 155, + 526, + 200 + ], + "spans": [ + { + "bbox": [ + 304, + 155, + 526, + 200 + ], + "type": "text", + "content": "Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners. https://github.com/openai/gpt-2." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 208, + 526, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 208, + 526, + 264 + ], + "spans": [ + { + "bbox": [ + 304, + 208, + 526, + 264 + ], + "type": "text", + "content": "Tim Van de Cruys. 2020. Automatic poetry generation from prosaic text. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2471-2480, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 272, + 526, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 272, + 526, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 272, + 526, + 338 + ], + "type": "text", + "content": "Tony Veale. 2013. Less rhyme, more reason: Knowledge-based poetry generation with feeling, insight and wit. In Proceedings of the Fourth International Conference on Computational Creativity, ICCC 2013, Sidney, Australia, June 12-14, 2013, pages 152-159. computationalcreativity.net." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 303, + 346, + 526, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 346, + 526, + 390 + ], + "spans": [ + { + "bbox": [ + 303, + 346, + 526, + 390 + ], + "type": "text", + "content": "Jianyou Wang, Xiaoxuan Zhang, Yuren Zhou, Christopher Suh, and Cynthia Rudin. 2021. There once was a really bad poet, it was automated but you didn't know it." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 303, + 411, + 377, + 424 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 411, + 377, + 424 + ], + "spans": [ + { + "bbox": [ + 303, + 411, + 377, + 424 + ], + "type": "text", + "content": "A Appendix" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 303, + 432, + 442, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 432, + 442, + 444 + ], + "spans": [ + { + "bbox": [ + 303, + 432, + 442, + 444 + ], + "type": "text", + "content": "A.1 Templating Mechanism" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 301, + 449, + 525, + 707 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 449, + 525, + 707 + ], + "spans": [ + { + "bbox": [ + 301, + 449, + 525, + 707 + ], + "type": "text", + "content": "Figure 8 presents more examples of our templating mechanism. We combine an adapted version of the Penn Treebank Project's part of speech tags along with articles, conjunctions, prepositions, and other filler words to construct these templates. Additionally, we provide the stress pattern of the syllables to ensure that the constraint of iambic pentameter is met. However, outside of the pre-determined filler words, POS do not have to directly adhere to the given stress pattern in splitting up words. For instance, in the first template, the provided syllable stress indicates that the JJ tag (adjective) should have two syllables, while the final VB tag (verb) should have only one syllable. However, the generated line ends with a monosyllabic adjective and a bisyllabic verb. As long as the stressing of the syllables aligns properly, each word can vary in its number of syllables. This is also visible in the fourth template example in Figure 8." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 303, + 715, + 470, + 741 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 715, + 470, + 741 + ], + "spans": [ + { + "bbox": [ + 303, + 715, + 470, + 741 + ], + "type": "text", + "content": "A.2 Elaboration on Experimental Competitors" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "type": "text", + "content": "Benhardt et al. (2018), referred to as Benhardt et al., uses a RNN to preselect rhyming words and" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1632" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 291, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 291, + 139 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 291, + 139 + ], + "type": "text", + "content": "restrict different parts of speech to fit within the sonnet format. Oliveira (2012), referred to as Co-PoetryMe, is a versatile platform using semantic and grammar templates to alter the type of poem, input words, and \"surprise\" factor generated." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 148, + 213, + 160 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 148, + 213, + 160 + ], + "spans": [ + { + "bbox": [ + 67, + 148, + 213, + 160 + ], + "type": "text", + "content": "A.3 Experimental Procedure" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 164, + 291, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 164, + 291, + 529 + ], + "spans": [ + { + "bbox": [ + 69, + 164, + 291, + 529 + ], + "type": "text", + "content": "For each pair of sonnets, respondents were asked to indicate whether Sonnet A or Sonnet B performed better based on factors such as adherence to the inputted theme, poeticness, grammatical correctness, ability to convey emotion, and likelihood of being written by a human. Available answer choices and their corresponding numeric scores from 1 to 5 were \"Definitely A\" (5), \"Probably A\" (4), \"The same\" (3), \"Probably B\" (2), and \"Definitely B\" (1). Both our sonnet and the competing model-human-sonnet had equal probability of being either sonnet A or sonnet B in each pair. To analyze this data, user inputs were translated into numeric scoring values corresponding to our model's sonnet being Sonnet A (i.e. if our sonnet is presented as B to the user, a response of \"Definitely B\" corresponds to a score of 5, \"Probably B\" corresponds to 4, \"Probably A\" corresponds to 2, and \"Definitely A\" corresponds to 1). Additionally, respondents were asked to answer sanity check questions to filter out respondents who answer illogically or who do not have a sufficient grasp of English grammar. This setup remained the same across all experiments, and an additional space was allocated for expert evaluators to leave qualitative comments on sonnet quality. Sample sonnet evaluation questions are visible in Figure 9." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 531, + 291, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 531, + 291, + 706 + ], + "spans": [ + { + "bbox": [ + 67, + 531, + 291, + 706 + ], + "type": "text", + "content": "After calculating the mean and standard deviation for scores across sonnets, we can immediately see whether our model performed better (an average score of " + }, + { + "bbox": [ + 67, + 531, + 291, + 706 + ], + "type": "inline_equation", + "content": ">3" + }, + { + "bbox": [ + 67, + 531, + 291, + 706 + ], + "type": "text", + "content": ") or worse (an average score of " + }, + { + "bbox": [ + 67, + 531, + 291, + 706 + ], + "type": "inline_equation", + "content": "< 3" + }, + { + "bbox": [ + 67, + 531, + 291, + 706 + ], + "type": "text", + "content": ") than the competitor in each respective category. We then performed a series of t-tests to establish these results' statistical significance. For factors that indicated our model performed better, we performed a right-tailed t-test (with the null-hypothesis as our model performed worse than the baseline), and we performed a left-tailed t-test for the remaining factors (with the null-hypothesis as our model performed better than the baseline)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 715, + 226, + 728 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 715, + 226, + 728 + ], + "spans": [ + { + "bbox": [ + 67, + 715, + 226, + 728 + ], + "type": "text", + "content": "A.4 Expert Evaluation Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 733, + 290, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 733, + 290, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 733, + 290, + 773 + ], + "type": "text", + "content": "In the expert evaluation, we emailed faculty at an American academic English department to recruit six faculty members and students to take our survey" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 71, + 526, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 248 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 248 + ], + "type": "text", + "content": "without payment. While we showed strong performance against the other computer-generated poems, we are consistently outperformed by human-written poems in all categories. Weaker performance on theme in experimental results may be explained by competitors' more frequent inclusion of the user-inputted theme word. For instance, in the expert evaluation, between two poems generated with the theme word \"forest\" (see Figure 10), one survey respondent states, \"Sonnet B repeats forest too much for my taste,\" subsequently giving our model a 5 in each of poeticness, grammar, emotion, and humanness, yet a 2 in theme." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 256, + 451, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 256, + 451, + 269 + ], + "spans": [ + { + "bbox": [ + 302, + 256, + 451, + 269 + ], + "type": "text", + "content": "A.5 Amazon MTurk Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 273, + 525, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 273, + 525, + 394 + ], + "spans": [ + { + "bbox": [ + 302, + 273, + 525, + 394 + ], + "type": "text", + "content": "In our evaluation using Amazon MTurk Services, we requested survey respondents from primarily English-speaking countries and with an approval rate of " + }, + { + "bbox": [ + 302, + 273, + 525, + 394 + ], + "type": "inline_equation", + "content": "\\geq 95\\%" + }, + { + "bbox": [ + 302, + 273, + 525, + 394 + ], + "type": "text", + "content": ". Crowdworkers were paid through the Amazon MTurk platform for this survey that on average took less than 30 minutes to complete. The questions and formatting remained the same as the expert evaluation, except no space was provided for qualitative feedback." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 395, + 525, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 395, + 525, + 612 + ], + "spans": [ + { + "bbox": [ + 302, + 395, + 525, + 612 + ], + "type": "text", + "content": "Based on Figure 4 there is enough statistical significance to conclude that our sonnets outperform PoeTryMe in poetic, grammar, emotion, and human categories " + }, + { + "bbox": [ + 302, + 395, + 525, + 612 + ], + "type": "inline_equation", + "content": "(p < 0.001)" + }, + { + "bbox": [ + 302, + 395, + 525, + 612 + ], + "type": "text", + "content": ". Against Benhardt et al., there is enough statistical significance to conclude that our sonnets perform better in grammar " + }, + { + "bbox": [ + 302, + 395, + 525, + 612 + ], + "type": "inline_equation", + "content": "(p < 0.001)" + }, + { + "bbox": [ + 302, + 395, + 525, + 612 + ], + "type": "text", + "content": ", and perform slightly better with weak statistical significance in emotion " + }, + { + "bbox": [ + 302, + 395, + 525, + 612 + ], + "type": "inline_equation", + "content": "(p < 0.15)" + }, + { + "bbox": [ + 302, + 395, + 525, + 612 + ], + "type": "text", + "content": ". Against human-written sonnets, the p-values for poetic, emotion, and even human categories are too large to strongly reject the null hypothesis that our model performed better than the baseline. Additionally, while the p-value indicates that this value is not statistically significant, it is interesting to note that our poems on average scored better in the grammar category." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 621, + 417, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 621, + 417, + 634 + ], + "spans": [ + { + "bbox": [ + 302, + 621, + 417, + 634 + ], + "type": "text", + "content": "A.6 Ablation Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 638, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 638, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 638, + 525, + 772 + ], + "type": "text", + "content": "In our ablation analysis, we replicate the Amazon MTurk analysis yet replace the competitor/human-written sonnets with poems generated with either the fine-tuned GPT-2 model without templating or the GPT-2 model without fine-tuning and with templating. This lets us test the individual efficacy of each factor (templating and fine-tuning GPT-2) against our method implementing both. Against poems generated with the fine-tuned GPT-2 and no templating, our sonnets performed better across" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1633" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 293, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 293, + 222 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 293, + 222 + ], + "type": "text", + "content": "all categories, and we can strongly reject the null hypothesis that our model performed worse than the baseline " + }, + { + "bbox": [ + 67, + 71, + 293, + 222 + ], + "type": "inline_equation", + "content": "(p < 0.0001)" + }, + { + "bbox": [ + 67, + 71, + 293, + 222 + ], + "type": "text", + "content": ". Against the poems generated with the GPT-2 model without fine-tuning and with templates, we can conclude with high statistical significance " + }, + { + "bbox": [ + 67, + 71, + 293, + 222 + ], + "type": "inline_equation", + "content": "(p < 0.01)" + }, + { + "bbox": [ + 67, + 71, + 293, + 222 + ], + "type": "text", + "content": " that we performed better in emotion, and conclude with weak statistical significance " + }, + { + "bbox": [ + 67, + 71, + 293, + 222 + ], + "type": "inline_equation", + "content": "(p < 0.10)" + }, + { + "bbox": [ + 67, + 71, + 293, + 222 + ], + "type": "text", + "content": " that we performed better in grammar and theme. These results indicate that our method is successful due to its usage of both the fine-tuned GPT-2 model and templating." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 229, + 241, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 229, + 241, + 243 + ], + "spans": [ + { + "bbox": [ + 67, + 229, + 241, + 243 + ], + "type": "text", + "content": "A.7 ChatGPT Qualitative Analysis" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 246, + 291, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 246, + 291, + 477 + ], + "spans": [ + { + "bbox": [ + 67, + 246, + 291, + 477 + ], + "type": "text", + "content": "While we did not have time to extensively evaluate the quality of our sonnets against those of ChatGPT, after generating several sonnets to test ChatGPT's sonnet quality, it seems as though this language model generates relatively generic, non-cohesive sonnets even with different parameters. For instance, in Figure 7, both of the sonnets are unable to cohesively connect these three topics along a reasonable storyline. Additionally, Sonnet A in particular seems to dedicate a single stanza to each of the three topics passed in, hardly attempting to connect them. Of course, with more intensive prompt engineering, it is possible to generate a sonnet more tailored to one's preference. However, even this short analysis demonstrates there are clearly still strides to be made in the field of automatic poetry generation even with the advent of ChatGPT." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 303, + 223, + 511, + 548 + ], + "blocks": [ + { + "bbox": [ + 303, + 223, + 511, + 548 + ], + "lines": [ + { + "bbox": [ + 303, + 223, + 511, + 548 + ], + "spans": [ + { + "bbox": [ + 303, + 223, + 511, + 548 + ], + "type": "table", + "html": "
Sonnet A
In realms of words, where novels find their reign, My friend, a skeptic, shuns the written tales. No pages turned, no fictional campaigns, Averse to narratives that time unveils. Yet in his hand, a pear's sweet succulence, A juicy feast, a taste of nature's grace. Each bite, a moment of pure indulgence, A salute's joy, a respite from life's race. But beyond fruit's delight, his spirit soars, Within the bounds of a duelist's domain. With sword in hand, he dances, parries, scores, A noble art where honor's stance remains. Though novels may elude his curious gaze, Through pears and fencing, passions find their ways.
Sonnet B
In realms of love, where rom-coms dance their waltz, My friend, a rebel, strays from silver screens. With flannel-clad allure, he stands, exalts, A soul untouched by scripted love's routines. No tears shed for the tales of starry-eyed, His heart resists the charms of love's façade. Yet stitches formed with yarn, his hands provide, Creating warmth, affection neatly clad. Through threads of patience, vibrant patterns bloom, His nimble fingers weave a tapestry. Each crochet stitch, a tale of friendship's loom, A testament to artistry set free. Though rom-coms fall short of his embrace, In flannel and crochet, love finds its place.
", + "image_path": "2e315f6fafb9c7ab2d9ad4a031ce60500873303e83e2bbb889e10f73f9adddb2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 302, + 556, + 526, + 628 + ], + "lines": [ + { + "bbox": [ + 302, + 556, + 526, + 628 + ], + "spans": [ + { + "bbox": [ + 302, + 556, + 526, + 628 + ], + "type": "text", + "content": "Figure 7: Comparison of two sonnets generated with ChatGPT. Sonnet A was generated with the prompt \"generate a sonnet about my friend who hates novels, eats pears, and fences\" and Sonnet B was generated with the prompt \"generate a sonnet about my friend who hates rom-coms, wears flannels, and crochets.\"" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "text", + "content": "1634" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 68, + 510, + 164 + ], + "blocks": [ + { + "bbox": [ + 83, + 68, + 510, + 164 + ], + "lines": [ + { + "bbox": [ + 83, + 68, + 510, + 164 + ], + "spans": [ + { + "bbox": [ + 83, + 68, + 510, + 164 + ], + "type": "table", + "html": "
TemplateSyllable StressExample Line
Where all the NNS of PRPD$ JJ NNS VB.0 1 0 1 0 1 01 0 1“Where all the gods of their past lives dictate”
And it VBD ABNN to the NN0 1 0 10 1 0 101“And it seemed evil to the enterprise”
Between the VBG and the VBG NN01 0 10 1 0 10 1“Between the glistening and the dying muse”
A JJ NN from the JJ NN0 10 10 1 0 1 01“A little lightness from the earthy sky”
Upon PRPO, PRPD$ NN POS NN01 01 0 10 101“Upon you, your life’s possibility”
Why VBC PRPS VBG such a JJ NN?0 1 0 10 1 0 101 0"Why do you squander such a precious thing?"
The NNS of ABNN, the NN on the NN0 1 0 1 0 10 1 0 1“The ghosts of death, the spirit on the earth”
", + "image_path": "315449d4de0e49013f7297e3f2fb3906310f69239862ccd39b27b8d80c581741.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 86, + 371, + 221, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 371, + 221, + 378 + ], + "spans": [ + { + "bbox": [ + 86, + 371, + 221, + 378 + ], + "type": "text", + "content": "The key word for both of these poems is \"wisdom.\" Which poem best adheres to this theme?" + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 89, + 386, + 268, + 401 + ], + "blocks": [ + { + "bbox": [ + 67, + 176, + 525, + 201 + ], + "lines": [ + { + "bbox": [ + 67, + 176, + 525, + 201 + ], + "spans": [ + { + "bbox": [ + 67, + 176, + 525, + 201 + ], + "type": "text", + "content": "Figure 8: Template examples, their corresponding syllable stress in order to adhere to iambic pentameter, and a sample line generated using the template." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 89, + 386, + 268, + 401 + ], + "lines": [ + { + "bbox": [ + 89, + 386, + 268, + 401 + ], + "spans": [ + { + "bbox": [ + 89, + 386, + 268, + 401 + ], + "type": "table", + "html": "
Definitely AProbably ASameProbably BDefinitely B
", + "image_path": "7fe3ee3ee799d37cace046d232e24d36af34824c2cc8134204459501a5e5117d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 86, + 414, + 138, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 414, + 138, + 420 + ], + "spans": [ + { + "bbox": [ + 86, + 414, + 138, + 420 + ], + "type": "text", + "content": "Which poem sounds more poetic?" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 94, + 429, + 262, + 439 + ], + "blocks": [ + { + "bbox": [ + 94, + 429, + 262, + 439 + ], + "lines": [ + { + "bbox": [ + 94, + 429, + 262, + 439 + ], + "spans": [ + { + "bbox": [ + 94, + 429, + 262, + 439 + ], + "type": "table", + "html": "
Definitely AProbably ASameProbably BDefinitely B
", + "image_path": "d32141c3e26af9c84542b1fe3fd9b70a1d5903d704265eaeb488ce754c6558a8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 86, + 455, + 151, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 455, + 151, + 461 + ], + "spans": [ + { + "bbox": [ + 86, + 455, + 151, + 461 + ], + "type": "text", + "content": "Which poem is more grammatically correct?" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 94, + 477, + 262, + 491 + ], + "blocks": [ + { + "bbox": [ + 94, + 477, + 262, + 491 + ], + "lines": [ + { + "bbox": [ + 94, + 477, + 262, + 491 + ], + "spans": [ + { + "bbox": [ + 94, + 477, + 262, + 491 + ], + "type": "table", + "html": "
Definitely AProbably ASameProbably BDefinitely B
", + "image_path": "23cc6a37e41323fa9f3c6e8ed5d9efa1e491cbb30a6a5e12ae2d08038c9cbabf.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 86, + 505, + 158, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 505, + 158, + 511 + ], + "spans": [ + { + "bbox": [ + 86, + 505, + 158, + 511 + ], + "type": "text", + "content": "Which poem conveys emotions more effectively?" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 94, + 521, + 262, + 529 + ], + "blocks": [ + { + "bbox": [ + 94, + 521, + 262, + 529 + ], + "lines": [ + { + "bbox": [ + 94, + 521, + 262, + 529 + ], + "spans": [ + { + "bbox": [ + 94, + 521, + 262, + 529 + ], + "type": "table", + "html": "
Definitely AProbably ASameProbably BDefinitely B
", + "image_path": "3d755f5125ffb56b57e99bc8d7195c05fa590d47d5926e2435cede181ae19b3e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 86, + 546, + 164, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 546, + 164, + 552 + ], + "spans": [ + { + "bbox": [ + 86, + 546, + 164, + 552 + ], + "type": "text", + "content": "Which poem is more likely to be written by a human?" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 94, + 563, + 262, + 571 + ], + "blocks": [ + { + "bbox": [ + 94, + 563, + 262, + 571 + ], + "lines": [ + { + "bbox": [ + 94, + 563, + 262, + 571 + ], + "spans": [ + { + "bbox": [ + 94, + 563, + 262, + 571 + ], + "type": "table", + "html": "
Definitely AProbably ASameProbably BDefinitely B
", + "image_path": "c1ffddf7888c695b0eb4589c5cfbee37ab2752d93385be4007219ae0ef326c86.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 67, + 592, + 290, + 614 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 592, + 290, + 614 + ], + "spans": [ + { + "bbox": [ + 67, + 592, + 290, + 614 + ], + "type": "text", + "content": "Figure 9: Survey questions presented for each pair of sonnets." + } + ] + } + ], + "index": 12, + "type": "text" + }, + { + "type": "table", + "bbox": [ + 304, + 317, + 522, + 641 + ], + "blocks": [ + { + "bbox": [ + 304, + 317, + 522, + 641 + ], + "lines": [ + { + "bbox": [ + 304, + 317, + 522, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 317, + 522, + 641 + ], + "type": "table", + "html": "
Sonnet A: Our Code
I was aghast to see the fireflies
Inflamed soothed toads, where there the dead boughs lay
And it seemed evil to the enterprise
The hag I had, the hag, the hog, the gray.
But I knew to my painless fireflies
And beauty was a kind and loving thing.
My life's light isle so longed on otherwise
So too my fireflies bloomed to my king.
Those eagles that with auburn hair flew oaks,
Beauty and beauty beamed within the air
Which made oasis overcomes to coax?
So too my hogs beheaded to my lair.
The windy night was in the mistletoe
And wept soiled toads in my dream's studio.
Sonnet B: PoetryMe
forest some more and reforest a trip!
in deserts where heavenly woodlands clink
many, many, many clustered before
come: not in establishments of the floor
the fields of agony, the endless circumstance
findings to lie to interrupt your earth
with summation and set, triumph and agony
floors of horror forest before my eyes
those that study clustered plant are psychologists
taking over my ness a second forest
an' you've got to forest them reforest
on every forest, indeed, that rainforests
and grounds of forest coming to accord
floor of establishments and lilt of sing
", + "image_path": "5e99d87fd8a2731cc47324f91709d6941f9184a33654410019d9b91762c1a965.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 650, + 525, + 686 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 650, + 525, + 686 + ], + "spans": [ + { + "bbox": [ + 302, + 650, + 525, + 686 + ], + "type": "text", + "content": "Figure 10: Comparison of two sonnets generated with theme word \"forest\". Sonnet A was generated with our code, and Sonnet B was generated using PoeTryMe." + } + ] + } + ], + "index": 14, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1635" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 342, + 286, + 420 + ], + "blocks": [ + { + "bbox": [ + 131, + 332, + 227, + 342 + ], + "lines": [ + { + "bbox": [ + 131, + 332, + 227, + 342 + ], + "spans": [ + { + "bbox": [ + 131, + 332, + 227, + 342 + ], + "type": "text", + "content": "Ablation Evaluation" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 73, + 342, + 286, + 420 + ], + "lines": [ + { + "bbox": [ + 73, + 342, + 286, + 420 + ], + "spans": [ + { + "bbox": [ + 73, + 342, + 286, + 420 + ], + "type": "table", + "html": "
CategoryMeanp-valueMeanp-value
Grammar3.51*5.10×10-53.21*0.06
Emotion3.61*9.00×10-63.40*3.89×10-3
Poetic3.61*4.00×10-63.09*0.29
Human3.66*1.00×10-63.01*0.46
Theme3.50*8.00×10-53.20*0.06
", + "image_path": "9c872bae3aeec2b3e2b9fcf05095ceaad81a5540ab4f279bd4353a50a4fea287.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 433, + 291, + 507 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 433, + 291, + 507 + ], + "spans": [ + { + "bbox": [ + 67, + 433, + 291, + 507 + ], + "type": "text", + "content": "Figure 11: Left: fine-tuned GPT-2 with no templates. Right: GPT-2 without fine-tuning, but with templates. Starred figures indicate average scores of " + }, + { + "bbox": [ + 67, + 433, + 291, + 507 + ], + "type": "inline_equation", + "content": ">3" + }, + { + "bbox": [ + 67, + 433, + 291, + 507 + ], + "type": "text", + "content": ", and underlined figures indicate that the p-value is low enough " + }, + { + "bbox": [ + 67, + 433, + 291, + 507 + ], + "type": "inline_equation", + "content": "(<0.05)" + }, + { + "bbox": [ + 67, + 433, + 291, + 507 + ], + "type": "text", + "content": " to claim that this higher average is statistically significant." + } + ] + } + ], + "index": 2, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "text", + "content": "1636" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "spans": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "type": "text", + "content": "A For every submission:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 106, + 414, + 242 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "spans": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "type": "text", + "content": "A1. Did you describe the limitations of your work? Limitations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 142, + 329, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 142, + 329, + 168 + ], + "spans": [ + { + "bbox": [ + 77, + 142, + 329, + 168 + ], + "type": "text", + "content": "A2. Did you discuss any potential risks of your work? Ethics" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 179, + 414, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 179, + 414, + 205 + ], + "spans": [ + { + "bbox": [ + 77, + 179, + 414, + 205 + ], + "type": "text", + "content": "A3. Do the abstract and introduction summarize the paper's main claims? Abstract, 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 77, + 215, + 398, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 215, + 398, + 242 + ], + "spans": [ + { + "bbox": [ + 77, + 215, + 398, + 242 + ], + "type": "text", + "content": "A4. Have you used AI writing assistants when working on this paper? Left blank." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 68, + 252, + 290, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 252, + 290, + 265 + ], + "spans": [ + { + "bbox": [ + 68, + 252, + 290, + 265 + ], + "type": "text", + "content": "B Did you use or create scientific artifacts?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 80, + 270, + 96, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 270, + 96, + 281 + ], + "spans": [ + { + "bbox": [ + 80, + 270, + 96, + 281 + ], + "type": "text", + "content": "3,4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 76, + 291, + 524, + 633 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 77, + 291, + 315, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 291, + 315, + 318 + ], + "spans": [ + { + "bbox": [ + 77, + 291, + 315, + 318 + ], + "type": "text", + "content": "B1. Did you cite the creators of artifacts you used? 3.2,3.3,References" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 76, + 328, + 463, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 328, + 463, + 355 + ], + "spans": [ + { + "bbox": [ + 76, + 328, + 463, + 355 + ], + "type": "text", + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Not applicable. Left blank." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 76, + 364, + 524, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 364, + 524, + 431 + ], + "spans": [ + { + "bbox": [ + 76, + 364, + 524, + 431 + ], + "type": "text", + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Not applicable. Left blank." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 77, + 441, + 524, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 441, + 524, + 496 + ], + "spans": [ + { + "bbox": [ + 77, + 441, + 524, + 496 + ], + "type": "text", + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Data used from publicly available sonnets/ poems were assumed to be not subject to dispute." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 77, + 504, + 524, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 504, + 524, + 544 + ], + "spans": [ + { + "bbox": [ + 77, + 504, + 524, + 544 + ], + "type": "text", + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Not applicable. Left blank." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 77, + 553, + 524, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 553, + 524, + 633 + ], + "spans": [ + { + "bbox": [ + 77, + 553, + 524, + 633 + ], + "type": "text", + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. 3.3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 68, + 644, + 293, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 644, + 293, + 657 + ], + "spans": [ + { + "bbox": [ + 68, + 644, + 293, + 657 + ], + "type": "text", + "content": "C Did you run computational experiments?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 79, + 662, + 127, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 662, + 127, + 674 + ], + "spans": [ + { + "bbox": [ + 79, + 662, + 127, + 674 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 77, + 684, + 524, + 724 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 684, + 524, + 724 + ], + "spans": [ + { + "bbox": [ + 77, + 684, + 524, + 724 + ], + "type": "text", + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? No response." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "spans": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "text", + "content": "ACL 2023 Responsible NLP Checklist" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "spans": [ + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "type": "text", + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1637" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 76, + 71, + 524, + 237 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 76, + 71, + 523, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 71, + 523, + 111 + ], + "spans": [ + { + "bbox": [ + 76, + 71, + 523, + 111 + ], + "type": "text", + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? No response." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 76, + 121, + 524, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 121, + 524, + 174 + ], + "spans": [ + { + "bbox": [ + 76, + 121, + 524, + 174 + ], + "type": "text", + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? No response." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 76, + 184, + 524, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 184, + 524, + 237 + ], + "spans": [ + { + "bbox": [ + 76, + 184, + 524, + 237 + ], + "type": "text", + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? No response." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 67, + 246, + 522, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 246, + 522, + 278 + ], + "spans": [ + { + "bbox": [ + 67, + 246, + 522, + 278 + ], + "type": "text", + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? 4,4.1,4.2,4.3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 77, + 285, + 524, + 550 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 77, + 285, + 524, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 285, + 524, + 326 + ], + "spans": [ + { + "bbox": [ + 77, + 285, + 524, + 326 + ], + "type": "text", + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Appendix" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 77, + 335, + 524, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 335, + 524, + 387 + ], + "spans": [ + { + "bbox": [ + 77, + 335, + 524, + 387 + ], + "type": "text", + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? A.5.A.6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 77, + 399, + 524, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 399, + 524, + 452 + ], + "spans": [ + { + "bbox": [ + 77, + 399, + 524, + 452 + ], + "type": "text", + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? We do not believe having data on poetry evaluation raises any ethical issues." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 77, + 462, + 523, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 462, + 523, + 502 + ], + "spans": [ + { + "bbox": [ + 77, + 462, + 523, + 502 + ], + "type": "text", + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? We do not believe having crowdworkers evaluate the same poems that were given to English professors raises any ethical issues." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 77, + 511, + 523, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 511, + 523, + 550 + ], + "spans": [ + { + "bbox": [ + 77, + 511, + 523, + 550 + ], + "type": "text", + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? A.6" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1638" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/The Role of Global and Local Context in Named Entity Recognition/d05b6607-6d83-45a3-9ebc-fc5492b0e085_content_list.json b/2023/The Role of Global and Local Context in Named Entity Recognition/d05b6607-6d83-45a3-9ebc-fc5492b0e085_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..69a460356d547a5948817203c11a3fc18cab6164 --- /dev/null +++ b/2023/The Role of Global and Local Context in Named Entity Recognition/d05b6607-6d83-45a3-9ebc-fc5492b0e085_content_list.json @@ -0,0 +1,1812 @@ +[ + { + "type": "text", + "text": "The Role of Global and Local Context in Named Entity Recognition", + "text_level": 1, + "bbox": [ + 147, + 83, + 850, + 104 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Arthur Amalvy", + "bbox": [ + 263, + 115, + 403, + 131 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Laboratoire Informatique d'Avignon arthur.amalvy@univ-avignon.fr", + "bbox": [ + 184, + 131, + 482, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vincent Labatut*", + "bbox": [ + 589, + 115, + 739, + 130 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Laboratoire Informatique d'Avignon vincent.labatut@univ-avignon.fr", + "bbox": [ + 507, + 131, + 823, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Richard Dufour*", + "text_level": 1, + "bbox": [ + 421, + 180, + 573, + 195 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Laboratoire des Sciences du Numérique de Nantes", + "bbox": [ + 292, + 197, + 702, + 212 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "richard.dufour@univ-nantes.fr", + "bbox": [ + 352, + 214, + 645, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 252, + 339, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pre-trained transformer-based models have recently shown great performance when applied to Named Entity Recognition (NER). As the complexity of their self-attention mechanism prevents them from processing long documents at once, these models are usually applied in a sequential fashion. Such an approach unfortunately only incorporates local context and prevents leveraging global document context in long documents such as novels, which might hinder performance. In this article, we explore the impact of global document context, and its relationships with local context. We find that correctly retrieving global document context has a greater impact on performance than only leveraging local context, prompting for further research on how to better retrieve that context.", + "bbox": [ + 141, + 279, + 460, + 521 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 114, + 533, + 260, + 548 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Named Entity Recognition (NER) is a fundamental task in Natural Language Processing (NLP), and is often used as a building block for solving higher-level tasks. Recently, pre-trained transformer-based models such as BERT (Devlin et al., 2019) or LUKE (Yamada et al., 2020) showed great NER performance and have been able to push the state of the art further.", + "bbox": [ + 112, + 558, + 489, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "These models, however, have a relatively short range because of the quadratic complexity of self-attention in the number of input tokens: as an example, BERT (Devlin et al., 2019) can only process spans of up to 512 tokens. For longer documents, texts are usually processed sequentially using a rolling window. Depending on the document, this local window may not always include all the context needed to perform inference, which may be present at the global document level. This leads to prediction errors (Stanislawek et al., 2019): In NER, this often occurs when the type of an entity cannot be inferred from the local context. For", + "bbox": [ + 112, + 687, + 489, + 896 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "instance, in the following sentence from the fantasy novel *Elantris*, one cannot decide if the entity *Elantris* is a person (PER) or a location (LOC) without prior knowledge:", + "bbox": [ + 507, + 253, + 884, + 317 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "\"Raoden stood, and as he did, his eyes fell on Elantris again.\"", + "bbox": [ + 542, + 324, + 845, + 356 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the novel, this prior knowledge comes from the fact that a human reader can recall previous mentions of Elantris, even at a very long range. A sequentially applied vanilla transformer-based model, however, might make an error without a neighboring sentence clearly establishing the status of Elantris as a city.", + "bbox": [ + 507, + 363, + 884, + 475 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While some works propose to retrieve external knowledge to disambiguate entities (Zhang et al., 2022; Wang et al., 2021), external resources are not always available. Furthermore, external retrieval might be more costly or less relevant than performing document-level context retrieval, provided the document contains the needed information, which depends on the type of document.", + "bbox": [ + 507, + 476, + 884, + 604 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Therefore, we wish to explore the relevance of document-level context when performing NER. We place ourselves at the sentence level, and we distinguish and study two types of contexts:", + "bbox": [ + 507, + 605, + 884, + 669 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- local context, consisting of surrounding sentences. This type of context can be used directly by vanilla transformer-based models, as their range lies beyond the simple sentence. Fully using surrounding context as in Devlin et al. (2019) is, however, computationally expensive.", + "- global context, consisting of all sentences available at the document level. To enhance NER prediction at the sentence level, we retrieve a few of these sentences and provide them as context for the model." + ], + "bbox": [ + 531, + 677, + 885, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We seek to answer the following question: is local context sufficient when solving the NER task,", + "bbox": [ + 507, + 887, + 884, + 919 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*These authors contributed equally.", + "bbox": [ + 136, + 903, + 354, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "714", + "bbox": [ + 485, + 927, + 515, + 940 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics", + "bbox": [ + 226, + 945, + 771, + 957 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Volume 2: Short Papers, pages 714-722", + "bbox": [ + 376, + 958, + 621, + 971 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "July 9-14, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 295, + 972, + 700, + 985 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "or would the model obtain better performance by retrieving global document context?", + "bbox": [ + 112, + 84, + 485, + 115 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To answer this question, we conduct experiments on a literary NER dataset we improved from its original version (Dekker et al., 2019). We release the annotation process, data and code necessary to reproduce these experiments under a free license1.", + "bbox": [ + 112, + 117, + 487, + 198 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 112, + 210, + 268, + 224 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 Sparse Transformers", + "text_level": 1, + "bbox": [ + 112, + 236, + 326, + 253 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Since the range problem of vanilla transformer-based models is due to the quadratic complexity of self-attention in the number of input tokens, several works on sparse transformers proposed alternative attention mechanisms in hope of reducing this complexity (Zaheer et al., 2020; Wang et al., 2020; Kitaev et al., 2020; Tay et al., 2020b,a; Beltagy et al., 2020; Choromanski et al., 2020; Katharopoulos et al., 2020; Child et al., 2019). While reducing self-attention complexity improves the effective range of transformers, these models still have issues processing very long documents (Tay et al., 2020c).", + "bbox": [ + 112, + 256, + 489, + 464 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2 Context retrieval", + "text_level": 1, + "bbox": [ + 112, + 479, + 294, + 492 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Context retrieval in general has been widely leveraged for other NLP tasks, such as semantic parsing (Guo et al., 2019), question answering (Ding et al., 2020), event detection (Pouran Ben Veyseh et al., 2021), or machine translation (Xu et al., 2020).", + "bbox": [ + 112, + 500, + 489, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In NER, context retrieval has mainly been used in an external fashion, for example by leveraging names lists and gazetteers (Seyler et al., 2018; Liu et al., 2019), knowledge bases (Luo et al., 2015) or search engines (Wang et al., 2021; Zhang et al., 2022). Meanwhile, we are interested in document-level context retrieval, which is comparatively seldom explored. While Luoma and Pyysalo (2020) study document-level context, their study is restricted to neighboring sentences, i.e. local context.", + "bbox": [ + 112, + 596, + 489, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3 Method and Experiments", + "text_level": 1, + "bbox": [ + 112, + 771, + 369, + 788 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1 Retrieval Heuristics", + "text_level": 1, + "bbox": [ + 112, + 797, + 317, + 810 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We wish to understand the role of both local and global contexts for the NER task. We split all documents in our dataset (described in Section 3.3) into sentences. We evaluate both local and global", + "bbox": [ + 112, + 818, + 487, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "simple heuristics of sentence retrieval in terms of NER performance impact. We study the following local heuristics:", + "bbox": [ + 507, + 84, + 882, + 131 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- before: Retrieves the closest $k$ sentences at the left of the input sentence.", + "- after: Same as before, but at the right of the input sentence.", + "- surrounding: Retrieves the closest $\\frac{k}{2}$ sentences on both sides of the input sentence." + ], + "bbox": [ + 531, + 139, + 882, + 254 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "And the following global heuristics:", + "bbox": [ + 527, + 261, + 798, + 275 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- random: Randomly retrieves a sentence from the whole document.", + "- samenoun: Randomly retrieves a sentence from the set of all sentences that have at least one common noun with the input sentence2. Intuitively, this heuristic will return sentences that contain entities of the input sentence, allowing for possible disambiguation. We use the NLTK library (Bird et al., 2009) to identify nouns.", + "- bm25: Retrieves sentences that are similar to the input sentences according to BM25 (Robertson, 1994). Retrieving similar sentences has already been found to increase NER performance (Zhang et al., 2022; Wang et al., 2021)." + ], + "bbox": [ + 529, + 284, + 882, + 557 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "It has to be noted that global heuristics can sometimes retrieve local context, as they are not restricted in which sentences they can retrieve at the document level. For all configurations, we concatenate the retrieved sentences to the input. During this concatenation step, we preserve the global order between sentences in the document.", + "bbox": [ + 507, + 565, + 882, + 677 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.2 Oracles", + "text_level": 1, + "bbox": [ + 507, + 688, + 616, + 702 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For each heuristic mentioned in Section 3.1, we also experiment with an oracle version. The oracle version retrieves 16 sentences from the document using the underlying retrieval heuristic, and retain only those that enhance the NER predictions the most. We measure this enhancement by counting the difference in numbers of NER BIO tags errors made with and without the context. In essence, the oracle setup simulates a perfect re-ranker model, and allows us to study the maximum performance of such an approach.", + "bbox": [ + 505, + 708, + 882, + 885 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "2If the set of sentences with a common noun is empty, the samenoun heuristic does not retrieve any sentence.", + "bbox": [ + 509, + 892, + 880, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1https://github.com/CompNet/conivel/tree/ACL2023", + "bbox": [ + 112, + 891, + 448, + 916 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "715", + "bbox": [ + 485, + 927, + 515, + 940 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.3 Dataset", + "text_level": 1, + "bbox": [ + 114, + 84, + 220, + 98 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To evaluate our heuristics, we use a corrected and improved version of the literary dataset of Dekker et al. (2019). This dataset is comprised of the first chapter of 40 novels in English, which we consider long enough for our experiments.", + "bbox": [ + 112, + 105, + 487, + 185 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dataset corrections The original dataset suffers mainly from annotation issues. To fix them, we design an annotation guide inspired by CoNLL-2003 (Tjong Kim Sang and De Meulder, 2003) and apply it consistently using a semi-automated process:", + "bbox": [ + 112, + 193, + 487, + 288 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We apply a set of simple rules to identify obvious errors $^3$ (for example, non capitalized entities annotated as PER are often false positives). Depending on the estimated performance of each rule, we manually reviewed its choices before application.", + "2. We manually review each difference between the predictions of a BERT (Devlin et al., 2019) model finetuned on a slightly modified version of the CoNLL-2003 dataset (Tjong Kim Sang and De Meulder, 2003) $^4$ and the existing annotations.", + "3. We manually correct the remaining errors." + ], + "bbox": [ + 127, + 294, + 487, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Further annotations The original dataset only consists of PER entities. We go further and annotate LOC and ORG entities. The final dataset contains 4476 PER entities, 886 LOC entities and 201 ORG entities.", + "bbox": [ + 112, + 527, + 487, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.4 NER Training", + "text_level": 1, + "bbox": [ + 112, + 618, + 272, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For all experiments, we use a pretrained BERTBASE (Devlin et al., 2019) model, consisting in 110 million parameters, followed by a classification head at the token level to perform NER. We finetune BERT for 2 epochs with a learning rate of $2 \\cdot 10^{-5}$ using the huggingface transformers library (Wolf et al., 2020), starting from the bert-base-cased checkpoint.", + "bbox": [ + 112, + 638, + 487, + 766 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.5 NER evaluation", + "text_level": 1, + "bbox": [ + 112, + 777, + 285, + 791 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We perform cross-validation with 5 folds on our NER dataset. We evaluate NER performance using the default mode of the seqeval (Nakayama, 2018) python library to ensure results can be reproduced.", + "bbox": [ + 112, + 797, + 487, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4 Results", + "text_level": 1, + "bbox": [ + 509, + 83, + 608, + 98 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1 Retrieval heuristics", + "text_level": 1, + "bbox": [ + 507, + 111, + 707, + 124 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The NER performance for retrieval heuristics can be seen in Figure 1. The samenoun heuristic performs the best among global heuristics, whereas the surrounding heuristic is the best for local heuristics. While the top results obtained with both heuristics are quite similar, we consider global heuristics as naive retrieval baselines: they could be bested by more complex approaches, which might enhance performance even more.", + "bbox": [ + 507, + 131, + 882, + 275 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Interestingly, the performance of both before and bm25 heuristics decrease strongly after four sentences, and even drop behind the no retrieval baseline. For both heuristics, this might be due to retrieving irrelevant sentences after a while. The bm25 heuristic is limited by the similar sentences present in the document: if there are not enough of them, the heuristic will retrieve unrelated ones. Meanwhile, the case of the before heuristic seems more puzzling, and could be indicative of a specific entity mention pattern that might warrant more investigations.", + "bbox": [ + 507, + 278, + 882, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.2 Oracle versions", + "text_level": 1, + "bbox": [ + 507, + 483, + 678, + 497 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "NER results with the oracle versions of retrieval heuristics can be found in Figure 2.", + "bbox": [ + 507, + 505, + 880, + 536 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "It is worth noting that the performance of the oracle versions of the heuristics always peaks when retrieving a single sentence. This might indicate that a single sentence is usually sufficient to resolve entity type ambiguities, but it might also be a result of the oracle ranking sentences individually, thereby not taking into account their possible combinations.", + "bbox": [ + 507, + 538, + 882, + 665 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Global heuristics perform better than local ones overall, with the oracle version of the random heuristic even performing better than both the before and after heuristics. These results tend to highlight the benefits of using global document context, provided it can be retrieved accurately.", + "bbox": [ + 507, + 667, + 880, + 764 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Retrieved sentences To better understand which sentences are useful for predictions when performing global retrieval, we plot in Figure 3 the distribution of the distance between sentences and their retrieved contexts for the oracle versions of heuristics samenoun and bm25. We find that $8\\%$ and $16\\%$ of retrieved sentences (for samenoun and bm25, respectively) are comprised within 6 sentences of their input sentence, while the other are", + "bbox": [ + 507, + 774, + 882, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "See Appendix A.2 for details.", + "bbox": [ + 134, + 866, + 324, + 879 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "4We modified the CoNLL-2003 dataset to include honorifics as part of PER entities to be consistent with our annotation guidelines.", + "bbox": [ + 115, + 881, + 487, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "716", + "bbox": [ + 485, + 927, + 515, + 940 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d6e18f007e81be7c15635a88316c495df06714fd067ac3584533c3cae930ec43.jpg", + "image_caption": [ + "Figure 1: Mean F1 score versus max number of retrieved sentences for all retrieval heuristics across 3 runs." + ], + "image_footnote": [], + "bbox": [ + 134, + 86, + 478, + 273 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/441cd4d4a47fec4f2597f28031882ee1f138192ce2a49d67ae808a745032d66e.jpg", + "image_caption": [ + "Figure 2: Mean F1 score versus max number of retrieved sentences across 3 runs for oracle versions of all retrieval heuristics." + ], + "image_footnote": [], + "bbox": [ + 514, + 85, + 857, + 272 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8fda74f838c6aee4fafbc6ab88a3ea88f01f147fbd24f89ac559d2a8ce124c80.jpg", + "image_caption": [ + "Figure 3: Distribution of the distance of retrieved sentences using the oracle versions of the samenoun and bm25 heuristics. The samenoun heuristic retrieves fewer sentences overall, since it is possible for some sentence to not have a common noun with any other sentence of its document." + ], + "image_footnote": [], + "bbox": [ + 115, + 340, + 884, + 502 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c1c7ed0cfeffe9fcd3edfab6c7c4efee7ef6784259ca99907d966e25400b7c8f.jpg", + "image_caption": [ + "Figure 4: Mean F1 score versus number of retrieved sentences across 3 runs for the oracle version of the bm25 heuristic, and the same heuristic restricted to distant context." + ], + "image_footnote": [], + "bbox": [ + 115, + 583, + 485, + 793 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "further away, highlighting the need for long-range", + "bbox": [ + 112, + 903, + 485, + 919 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "retrieval.", + "bbox": [ + 509, + 583, + 579, + 596 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Local context importance To see whether or not local context is an important component of NER performance, we perform an experiment where we restrict the oracle version of the bm25 heuristic from retrieving local surrounding context. Results can be found in Figure 4. NER performance remains about the same without local context, which tends to show that local context is not strictly necessary for performance.", + "bbox": [ + 507, + 607, + 882, + 752 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5 Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 507, + 764, + 796, + 778 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this article, we explored the role of local and global context in Named Entity Recognition. Our results tend to show that, for literary texts, retrieving global document context is more effective at enhancing NER performance than retrieving only local context, even when using relatively simple retrieval heuristics. We also showed that a re-ranker model using simple document-level retrieval heuris", + "bbox": [ + 507, + 790, + 884, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "717", + "bbox": [ + 485, + 927, + 515, + 939 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tics could obtain significant NER performance improvements. Overall, our work prompts for further research in how to accurately retrieve global context for NER.", + "bbox": [ + 112, + 84, + 492, + 148 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6 Limitations", + "text_level": 1, + "bbox": [ + 112, + 160, + 250, + 175 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We acknowledge the following limitations of our work:", + "bbox": [ + 112, + 185, + 489, + 216 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- While the oracle selects a sentence according to the benefits it provides when performing NER, it does not consider the interactions between selected sentences. This may lead to lowered performances when the several sentences are retrieved at once.", + "- The retrieval heuristics considered are naive on purpose, as the focus of this work is not performance. Stronger retrieval heuristics may achieve better results than presented in this article.", + "- The studied documents only consist in the first chapter of a set of novels. Using complete novel would increase the number of possible information to retrieve for the presented global heuristics." + ], + "bbox": [ + 137, + 227, + 489, + 506 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 115, + 543, + 213, + 558 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "I. Beltagy, M. E. Peters, and A. Cohan. 2020. Longformer: The long-document transformer. arXiv, cs.CL:2004.05150.", + "S. Bird, E. Loper, and E. Klein. 2009. Natural Language Processing with Python. O'Reilly Media Inc.", + "R. Child, S. Gray, A. Radford, and I. Sutskever. 2019. Generating long sequences with sparse transformers. arXiv, cs.LG:1904.10509.", + "K. Choromanski, V. Likhosherstov, D. Dohan, X. Song, A. Gane, T. Sarlos, P. Hawkins, J. Davis, A. Mohiuddin, L. Kaiser, D. Belanger, L. Colwell, and A. Weller. 2020. Rethinking attention with performers. arXiv, cs.LG:2009.14794.", + "N. Dekker, T. Kuhn, and M. van Erp. 2019. Evaluating named entity recognition tools for extracting social networks from novels. PeerJ Computer Science, 5:e189.", + "J. Devlin, M. Chang, K. Lee, and K. Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, volume 1, pages 4171-4186." + ], + "bbox": [ + 115, + 565, + 489, + 917 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "M. Ding, C. Zhou, H. Yang, and J. Tang. 2020. CogLTX: Applying bert to long texts. In Advances in Neural Information Processing Systems, volume 33, pages 12792-12804.", + "D. Guo, D. Tang, N. Duan, M. Zhou, and J. Yin. 2019. Coupling retrieval and meta-learning for context-dependent semantic parsing. In 57th Annual Meeting of the Association for Computational Linguistics, pages 855-866.", + "A. Katharopoulos, A. Vyas, N. Pappas, and François Fleuret. 2020. Transformers are mnns: Fast autoregressive transformers with linear attention. In Proceedings of the 37th International Conference on Machine Learning, ICML'20.", + "N. Kitaev, L. Kaiser, and A. Levskaya. 2020. Reformer: The efficient transformer. arXiv, cs.LG:2001.04451.", + "T. Liu, J. Yao, and C. Lin. 2019. Towards improving neural named entity recognition with gazetteers. In 57th Annual Meeting of the Association for Computational Linguistics, pages 5301-5307.", + "G. Luo, X. Huang, C. Lin, and Z. Nie. 2015. Joint entity recognition and disambiguation. In 2015 Conference on Empirical Methods in Natural Language Processing, pages 879-888.", + "J. Luoma and S. Pyysalo. 2020. Exploring cross-sentence contexts for named entity recognition with BERT. In 28th International Conference on Computational Linguistics, pages 904-914.", + "H. Nakayama. 2018. seqeval: A python framework for sequence labeling evaluation.", + "A. Pouran Ben Veyseh, M. V. Nguyen, N. Ngo Trung, B. Min, and T. H. Nguyen. 2021. Modeling document-level context for event detection via important context selection. In Conference on Empirical Methods in Natural Language Processing, pages 5403-5413.", + "S. E. W. Robertson. 1994. Some simple effective approximations to the 2-poisson model for probabilistic weighted retrieval. In SIGIR '94, pages 232-241.", + "D. Seyler, T. Dembelova, L. Del Corro, J. Hoffart, and G. Weikum. 2018. A study of the importance of external knowledge in the named entity recognition task. In 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 241-246.", + "T. Stanislawek, A. Wróblewska, A. Wojcicka, D. Ziembicki, and P. Biecek. 2019. Named entity recognition - is there a glass ceiling? In 23rd Conference on Computational Natural Language Learning, pages 624-633.", + "Y. Tay, D. Bahri, D. Metzler, D. Juan, Z. Zhao, and C. Zheng. 2020a. Synthesizer: Rethinking self-attention in transformer models. arXiv, cs.CL:2005.00743." + ], + "bbox": [ + 510, + 85, + 884, + 917 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "718", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Y. Tay, D. Bahri, L. Yang, D. Metzler, and D. Juan. 2020b. Sparse sinkhorn attention. arXiv, cs.LG:2002.11296.", + "Y. Tay, M. Dehghani, S. Abnar, Y. Shen, D. Bahri, P. Pham, J. Rao, L. Yang, S. Ruder, and D. Metzler. 2020c. Long range arena: A benchmark for efficient transformers. arXiv, cs.LG:2011.04006.", + "E. F. Tjong Kim Sang and F. De Meulder. 2003. Introduction to the CoNLL-2003 shared task: Language-independent named entity recognition. In 7th Conference on Natural Language Learning, pages 142-147.", + "S. Wang, B. Z. Li, M. Khabsa, H. Fang, and H. Ma. 2020. Linformer: Self-attention with linear complexity. arXiv, cs.LG:2006.04768.", + "X. Wang, Y. Jiang, N. Bach, T. Wang, Z. Huang, F. Huang, and K. Tu. 2021. Improving named entity recognition by external context retrieving and cooperative learning. In 59th Annual Meeting of the Association for Computational Linguistics and 11th International Joint Conference on Natural Language Processing, volume 1, pages 1800-1812.", + "T. Wolf, L. Debut, V. Sanh, J. Chaumont, C. Delangue, A. Moi, P. Cistac, T. Rault, R. Louf, M. Funtowicz, J. Davison, S. Shleifer, P. von Platen, C. Ma, Y. Jernite, J. Plu, C. Xu, T. Le Scao, S. Gugger, M. Drame, Q. Lhoest, and A. M. Rush. 2020. Transformers: State-of-the-art natural language processing. In *Conference on Empirical Methods in Natural Language Processing: System Demonstrations*, pages 38-45.", + "J. Xu, J. Crego, and J. Senellart. 2020. Boosting neural machine translation with similar translations. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1580-1590.", + "I. Yamada, A. Asai, H. Shindo, H. Takeda, and Y. Matsumoto. 2020. LUKE: Deep contextualized entity representations with entity-aware self-attention. In Conference on Empirical Methods in Natural Language Processing, pages 6442-6454.", + "M. Zaheer, G. Guruganesh, K. A. Dubey, J. Ainslie, C. Alberti, S. Ontanon, P. Pham, A. Ravula, Q. Wang, L. Yang, and A. Ahmed. 2020. Big bird: Transformers for longer sequences. In Advances in Neural Information Processing Systems, volume 33, pages 17283-17297.", + "X. Zhang, Y. Jiang, X. Wang, X. Hu, Y. Sun, P. Xie, and M. Zhang. 2022. Domain-specific NER via retrieving correlated samples. In Proceedings of the 29th International Conference on Computational Linguistics, pages 2398-2404." + ], + "bbox": [ + 115, + 85, + 487, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A Dataset Details", + "text_level": 1, + "bbox": [ + 115, + 825, + 282, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A.1 Document Lengths", + "text_level": 1, + "bbox": [ + 114, + 850, + 312, + 865 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our NER dataset is composed of documents longer that typical NER datasets such as CoNLL-2003 (Tjong Kim Sang and De Meulder, 2003).", + "bbox": [ + 112, + 871, + 487, + 917 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4cab2bba3cf8bc1e551f8128ff3e963d3d928bf7ab5ed7b35a9f52337a3c93a9.jpg", + "image_caption": [ + "Figure 5: Distribution of the number of sentences in our enhanced version of the dataset from Dekker et al. (2019).", + "Figure 5 shows the distribution of the number of sentences of our NER dataset." + ], + "image_footnote": [], + "bbox": [ + 514, + 84, + 877, + 274 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A.2 Automatic Correction Rules", + "text_level": 1, + "bbox": [ + 507, + 399, + 779, + 413 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use the following rules to automatically identify obvious errors in the original dataset from Dekker et al. (2019). The original dataset only contained PER entities, so these rules only apply to them:", + "bbox": [ + 507, + 419, + 880, + 483 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- If a span appears in the list of characters from its novel but is not annotated as an entity, we investigate whether or not this is a false negative.", + "- Similarly, if a span annotated as an entity does not appear in the list of characters from its novel, we investigate whether or not it is a false positive.", + "- Finally, if a span is annotated as an entity but all of its tokens are not capitalized, we check if it is a false positive." + ], + "bbox": [ + 531, + 494, + 880, + 690 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "B Heuristics Results Breakdown by Precision/Recall", + "text_level": 1, + "bbox": [ + 507, + 701, + 831, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figures 6 and 7 show precision and recall for all retrieval heuristics. Interestingly, retrieval only has a positive effect on recall, with precision being lower than the baseline except for the surrounding heuristic.", + "bbox": [ + 507, + 743, + 882, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "B.1 Oracle Versions", + "text_level": 1, + "bbox": [ + 507, + 834, + 682, + 848 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figures 6 and 7 show precision and recall for the oracle versions of all retrieval heuristics. While retrieval benefits recall more than precision, precision is still increased using retrieval. Together with", + "bbox": [ + 507, + 854, + 882, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "719", + "bbox": [ + 485, + 927, + 515, + 940 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/82be396a06d23103061634f01b8dd2d3d558c6b6e6b009839b220b2481420ceb.jpg", + "image_caption": [ + "Figure 6: Mean precision versus max number of retrieved sentences for all retrieval heuristics across 3 runs." + ], + "image_footnote": [], + "bbox": [ + 134, + 84, + 477, + 269 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/78048c3a9dc59a725d636f948057608a6b34cc70be64b4d8d43e0a2f8c3e85ed.jpg", + "image_caption": [ + "Figure 8: Mean precision versus max number of retrieved sentences across 3 runs for oracle versions of all retrieval heuristics." + ], + "image_footnote": [], + "bbox": [ + 134, + 340, + 475, + 529 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the results from the regular heuristics, these results again highlight the potential performance gains of using a suitable re-ranker model to retrieve context.", + "bbox": [ + 112, + 609, + 489, + 657 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d69e670218c71bfde2d923fe05ea806d34a133004f869063f15dcd780356bbcb.jpg", + "image_caption": [ + "Figure 7: Mean recall versus max number of retrieved sentences for all retrieval heuristics across 3 runs." + ], + "image_footnote": [], + "bbox": [ + 514, + 84, + 857, + 269 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7446c94113ae9db36eb9916d0e343f08ecd370ea34c02dd4e3087720be5041f3.jpg", + "image_caption": [ + "Figure 9: Mean recall versus max number of retrieved sentences across 3 runs for oracle versions of all retrieval heuristics." + ], + "image_footnote": [], + "bbox": [ + 515, + 340, + 857, + 529 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "720", + "bbox": [ + 485, + 927, + 515, + 940 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A For every submission:", + "bbox": [ + 114, + 107, + 322, + 122 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A1. Did you describe the limitations of your work?", + "bbox": [ + 129, + 126, + 532, + 143 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Yes, limitations are discussed in Section 6", + "bbox": [ + 149, + 143, + 460, + 158 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A2. Did you discuss any potential risks of your work?", + "bbox": [ + 129, + 170, + 552, + 186 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We do not think our work presents any direct risk", + "bbox": [ + 149, + 187, + 514, + 200 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A3. Do the abstract and introduction summarize the paper's main claims?", + "bbox": [ + 129, + 212, + 695, + 228 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Yes, in the abstract and Section 1", + "bbox": [ + 149, + 230, + 396, + 244 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A4. Have you used AI writing assistants when working on this paper?", + "bbox": [ + 129, + 256, + 668, + 272 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 149, + 273, + 231, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "B Did you use or create scientific artifacts?", + "bbox": [ + 114, + 299, + 487, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "in Section 3.4, we indicate that we use a BERT checkpoint. We also use a previous NER dataset, see Section 3.3. We distribute an enhanced version of this dataset and code to reproduce our experiments.", + "bbox": [ + 112, + 321, + 880, + 351 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "B1. Did you cite the creators of artifacts you used?", + "bbox": [ + 129, + 362, + 529, + 379 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "See Section 3.3 and 3.4", + "bbox": [ + 149, + 380, + 326, + 394 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?", + "bbox": [ + 129, + 406, + 778, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We specify the license in the Github repository given at the end of section 1.", + "bbox": [ + 149, + 423, + 707, + 439 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?", + "bbox": [ + 129, + 449, + 880, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We use a dataset published for research purposes.", + "bbox": [ + 149, + 514, + 519, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?", + "bbox": [ + 129, + 539, + 880, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Collected datas do not include information that can be used to identify individuals", + "bbox": [ + 149, + 589, + 759, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?", + "bbox": [ + 129, + 614, + 880, + 646 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We specify that the distributed dataset covers english literature (section 3.3). The reader can refer to Dekker et al., 2019 for more informations on the dataset.", + "bbox": [ + 149, + 648, + 880, + 680 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be.", + "bbox": [ + 129, + 689, + 880, + 770 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We include the number of document of our dataset in Section 3.3 We also include statistics about the length of these document in the Appendix", + "bbox": [ + 149, + 771, + 880, + 803 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "ACL 2023 Responsible NLP Checklist", + "bbox": [ + 132, + 84, + 433, + 99 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.", + "bbox": [ + 112, + 807, + 877, + 831 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "721", + "bbox": [ + 485, + 928, + 512, + 940 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "C Did you run computational experiments?", + "text_level": 1, + "bbox": [ + 114, + 83, + 494, + 99 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "See Section 3.4 and Section 3.5", + "bbox": [ + 131, + 105, + 366, + 118 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?", + "bbox": [ + 129, + 130, + 878, + 162 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "See Section 3.4", + "bbox": [ + 149, + 164, + 267, + 178 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?", + "bbox": [ + 129, + 189, + 880, + 222 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We include training hyperparameters in Section 3.4", + "bbox": [ + 149, + 223, + 532, + 237 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?", + "bbox": [ + 129, + 248, + 880, + 296 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our results are reported in Section 4. We indicate that, for Figure 1 and 2, each point is the mean $F1$ of 3 runs.", + "bbox": [ + 149, + 298, + 880, + 329 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?", + "bbox": [ + 129, + 338, + 880, + 387 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "See Section 3.1 (nltk), Section 3.4 (huggingface transformers), Section 3.5 (sequeval)", + "bbox": [ + 149, + 388, + 769, + 404 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?", + "text_level": 1, + "bbox": [ + 114, + 414, + 875, + 432 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Section 3.3", + "bbox": [ + 132, + 437, + 218, + 450 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?", + "bbox": [ + 129, + 463, + 880, + 494 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The experiments were free of any risks", + "bbox": [ + 149, + 495, + 435, + 510 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?", + "bbox": [ + 129, + 521, + 880, + 569 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The authors annotated the dataset themselves", + "bbox": [ + 149, + 571, + 489, + 585 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?", + "bbox": [ + 129, + 596, + 880, + 644 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The authors annotated the dataset themselves", + "bbox": [ + 149, + 646, + 489, + 659 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?", + "bbox": [ + 129, + 671, + 873, + 687 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The authors annotated the dataset themselves", + "bbox": [ + 149, + 689, + 489, + 703 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?", + "bbox": [ + 129, + 715, + 878, + 746 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This is not relevant since annotation was done by the authors", + "bbox": [ + 149, + 747, + 603, + 762 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "722", + "bbox": [ + 485, + 927, + 515, + 940 + ], + "page_idx": 8 + } +] \ No newline at end of file diff --git a/2023/The Role of Global and Local Context in Named Entity Recognition/d05b6607-6d83-45a3-9ebc-fc5492b0e085_model.json b/2023/The Role of Global and Local Context in Named Entity Recognition/d05b6607-6d83-45a3-9ebc-fc5492b0e085_model.json new file mode 100644 index 0000000000000000000000000000000000000000..225c68895336d93f03c7a92b5f2e9a82ce00a53f --- /dev/null +++ b/2023/The Role of Global and Local Context in Named Entity Recognition/d05b6607-6d83-45a3-9ebc-fc5492b0e085_model.json @@ -0,0 +1,2330 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.149, + 0.084, + 0.851, + 0.105 + ], + "angle": 0, + "content": "The Role of Global and Local Context in Named Entity Recognition" + }, + { + "type": "text", + "bbox": [ + 0.265, + 0.116, + 0.404, + 0.132 + ], + "angle": 0, + "content": "Arthur Amalvy" + }, + { + "type": "text", + "bbox": [ + 0.185, + 0.133, + 0.484, + 0.164 + ], + "angle": 0, + "content": "Laboratoire Informatique d'Avignon arthur.amalvy@univ-avignon.fr" + }, + { + "type": "text", + "bbox": [ + 0.59, + 0.116, + 0.741, + 0.131 + ], + "angle": 0, + "content": "Vincent Labatut*" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.133, + 0.824, + 0.164 + ], + "angle": 0, + "content": "Laboratoire Informatique d'Avignon vincent.labatut@univ-avignon.fr" + }, + { + "type": "title", + "bbox": [ + 0.423, + 0.181, + 0.574, + 0.196 + ], + "angle": 0, + "content": "Richard Dufour*" + }, + { + "type": "text", + "bbox": [ + 0.293, + 0.198, + 0.704, + 0.214 + ], + "angle": 0, + "content": "Laboratoire des Sciences du Numérique de Nantes" + }, + { + "type": "text", + "bbox": [ + 0.353, + 0.215, + 0.646, + 0.229 + ], + "angle": 0, + "content": "richard.dufour@univ-nantes.fr" + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.253, + 0.341, + 0.268 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.28, + 0.461, + 0.522 + ], + "angle": 0, + "content": "Pre-trained transformer-based models have recently shown great performance when applied to Named Entity Recognition (NER). As the complexity of their self-attention mechanism prevents them from processing long documents at once, these models are usually applied in a sequential fashion. Such an approach unfortunately only incorporates local context and prevents leveraging global document context in long documents such as novels, which might hinder performance. In this article, we explore the impact of global document context, and its relationships with local context. We find that correctly retrieving global document context has a greater impact on performance than only leveraging local context, prompting for further research on how to better retrieve that context." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.534, + 0.262, + 0.549 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.559, + 0.49, + 0.686 + ], + "angle": 0, + "content": "Named Entity Recognition (NER) is a fundamental task in Natural Language Processing (NLP), and is often used as a building block for solving higher-level tasks. Recently, pre-trained transformer-based models such as BERT (Devlin et al., 2019) or LUKE (Yamada et al., 2020) showed great NER performance and have been able to push the state of the art further." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.688, + 0.49, + 0.897 + ], + "angle": 0, + "content": "These models, however, have a relatively short range because of the quadratic complexity of self-attention in the number of input tokens: as an example, BERT (Devlin et al., 2019) can only process spans of up to 512 tokens. For longer documents, texts are usually processed sequentially using a rolling window. Depending on the document, this local window may not always include all the context needed to perform inference, which may be present at the global document level. This leads to prediction errors (Stanislawek et al., 2019): In NER, this often occurs when the type of an entity cannot be inferred from the local context. For" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.254, + 0.885, + 0.318 + ], + "angle": 0, + "content": "instance, in the following sentence from the fantasy novel *Elantris*, one cannot decide if the entity *Elantris* is a person (PER) or a location (LOC) without prior knowledge:" + }, + { + "type": "text", + "bbox": [ + 0.544, + 0.325, + 0.846, + 0.357 + ], + "angle": 0, + "content": "\"Raoden stood, and as he did, his eyes fell on Elantris again.\"" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.364, + 0.885, + 0.476 + ], + "angle": 0, + "content": "In the novel, this prior knowledge comes from the fact that a human reader can recall previous mentions of Elantris, even at a very long range. A sequentially applied vanilla transformer-based model, however, might make an error without a neighboring sentence clearly establishing the status of Elantris as a city." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.477, + 0.885, + 0.605 + ], + "angle": 0, + "content": "While some works propose to retrieve external knowledge to disambiguate entities (Zhang et al., 2022; Wang et al., 2021), external resources are not always available. Furthermore, external retrieval might be more costly or less relevant than performing document-level context retrieval, provided the document contains the needed information, which depends on the type of document." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.606, + 0.885, + 0.67 + ], + "angle": 0, + "content": "Therefore, we wish to explore the relevance of document-level context when performing NER. We place ourselves at the sentence level, and we distinguish and study two types of contexts:" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.678, + 0.887, + 0.79 + ], + "angle": 0, + "content": "- local context, consisting of surrounding sentences. This type of context can be used directly by vanilla transformer-based models, as their range lies beyond the simple sentence. Fully using surrounding context as in Devlin et al. (2019) is, however, computationally expensive." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.799, + 0.887, + 0.879 + ], + "angle": 0, + "content": "- global context, consisting of all sentences available at the document level. To enhance NER prediction at the sentence level, we retrieve a few of these sentences and provide them as context for the model." + }, + { + "type": "list", + "bbox": [ + 0.532, + 0.678, + 0.887, + 0.879 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.888, + 0.885, + 0.92 + ], + "angle": 0, + "content": "We seek to answer the following question: is local context sufficient when solving the NER task," + }, + { + "type": "page_footnote", + "bbox": [ + 0.137, + 0.904, + 0.355, + 0.919 + ], + "angle": 0, + "content": "*These authors contributed equally." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.928, + 0.517, + 0.941 + ], + "angle": 0, + "content": "714" + }, + { + "type": "footer", + "bbox": [ + 0.227, + 0.946, + 0.772, + 0.958 + ], + "angle": 0, + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + }, + { + "type": "footer", + "bbox": [ + 0.377, + 0.959, + 0.623, + 0.972 + ], + "angle": 0, + "content": "Volume 2: Short Papers, pages 714-722" + }, + { + "type": "footer", + "bbox": [ + 0.297, + 0.973, + 0.702, + 0.986 + ], + "angle": 0, + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.486, + 0.116 + ], + "angle": 0, + "content": "or would the model obtain better performance by retrieving global document context?" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.118, + 0.488, + 0.199 + ], + "angle": 0, + "content": "To answer this question, we conduct experiments on a literary NER dataset we improved from its original version (Dekker et al., 2019). We release the annotation process, data and code necessary to reproduce these experiments under a free license1." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.211, + 0.27, + 0.225 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.237, + 0.327, + 0.254 + ], + "angle": 0, + "content": "2.1 Sparse Transformers" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.258, + 0.49, + 0.466 + ], + "angle": 0, + "content": "Since the range problem of vanilla transformer-based models is due to the quadratic complexity of self-attention in the number of input tokens, several works on sparse transformers proposed alternative attention mechanisms in hope of reducing this complexity (Zaheer et al., 2020; Wang et al., 2020; Kitaev et al., 2020; Tay et al., 2020b,a; Beltagy et al., 2020; Choromanski et al., 2020; Katharopoulos et al., 2020; Child et al., 2019). While reducing self-attention complexity improves the effective range of transformers, these models still have issues processing very long documents (Tay et al., 2020c)." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.48, + 0.295, + 0.493 + ], + "angle": 0, + "content": "2.2 Context retrieval" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.501, + 0.49, + 0.595 + ], + "angle": 0, + "content": "Context retrieval in general has been widely leveraged for other NLP tasks, such as semantic parsing (Guo et al., 2019), question answering (Ding et al., 2020), event detection (Pouran Ben Veyseh et al., 2021), or machine translation (Xu et al., 2020)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.598, + 0.49, + 0.759 + ], + "angle": 0, + "content": "In NER, context retrieval has mainly been used in an external fashion, for example by leveraging names lists and gazetteers (Seyler et al., 2018; Liu et al., 2019), knowledge bases (Luo et al., 2015) or search engines (Wang et al., 2021; Zhang et al., 2022). Meanwhile, we are interested in document-level context retrieval, which is comparatively seldom explored. While Luoma and Pyysalo (2020) study document-level context, their study is restricted to neighboring sentences, i.e. local context." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.772, + 0.37, + 0.789 + ], + "angle": 0, + "content": "3 Method and Experiments" + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.798, + 0.318, + 0.812 + ], + "angle": 0, + "content": "3.1 Retrieval Heuristics" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.819, + 0.489, + 0.884 + ], + "angle": 0, + "content": "We wish to understand the role of both local and global contexts for the NER task. We split all documents in our dataset (described in Section 3.3) into sentences. We evaluate both local and global" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.883, + 0.132 + ], + "angle": 0, + "content": "simple heuristics of sentence retrieval in terms of NER performance impact. We study the following local heuristics:" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.14, + 0.882, + 0.172 + ], + "angle": 0, + "content": "- before: Retrieves the closest \\( k \\) sentences at the left of the input sentence." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.181, + 0.882, + 0.213 + ], + "angle": 0, + "content": "- after: Same as before, but at the right of the input sentence." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.221, + 0.883, + 0.255 + ], + "angle": 0, + "content": "- surrounding: Retrieves the closest \\(\\frac{k}{2}\\) sentences on both sides of the input sentence." + }, + { + "type": "list", + "bbox": [ + 0.532, + 0.14, + 0.883, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.529, + 0.262, + 0.799, + 0.277 + ], + "angle": 0, + "content": "And the following global heuristics:" + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.285, + 0.882, + 0.315 + ], + "angle": 0, + "content": "- random: Randomly retrieves a sentence from the whole document." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.326, + 0.884, + 0.452 + ], + "angle": 0, + "content": "- samenoun: Randomly retrieves a sentence from the set of all sentences that have at least one common noun with the input sentence2. Intuitively, this heuristic will return sentences that contain entities of the input sentence, allowing for possible disambiguation. We use the NLTK library (Bird et al., 2009) to identify nouns." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.464, + 0.884, + 0.558 + ], + "angle": 0, + "content": "- bm25: Retrieves sentences that are similar to the input sentences according to BM25 (Robertson, 1994). Retrieving similar sentences has already been found to increase NER performance (Zhang et al., 2022; Wang et al., 2021)." + }, + { + "type": "list", + "bbox": [ + 0.531, + 0.285, + 0.884, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.567, + 0.884, + 0.678 + ], + "angle": 0, + "content": "It has to be noted that global heuristics can sometimes retrieve local context, as they are not restricted in which sentences they can retrieve at the document level. For all configurations, we concatenate the retrieved sentences to the input. During this concatenation step, we preserve the global order between sentences in the document." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.689, + 0.618, + 0.703 + ], + "angle": 0, + "content": "3.2 Oracles" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.709, + 0.883, + 0.887 + ], + "angle": 0, + "content": "For each heuristic mentioned in Section 3.1, we also experiment with an oracle version. The oracle version retrieves 16 sentences from the document using the underlying retrieval heuristic, and retain only those that enhance the NER predictions the most. We measure this enhancement by counting the difference in numbers of NER BIO tags errors made with and without the context. In essence, the oracle setup simulates a perfect re-ranker model, and allows us to study the maximum performance of such an approach." + }, + { + "type": "page_footnote", + "bbox": [ + 0.51, + 0.893, + 0.882, + 0.918 + ], + "angle": 0, + "content": "2If the set of sentences with a common noun is empty, the samenoun heuristic does not retrieve any sentence." + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.892, + 0.449, + 0.917 + ], + "angle": 0, + "content": "1https://github.com/CompNet/conivel/tree/ACL2023" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.928, + 0.516, + 0.941 + ], + "angle": 0, + "content": "715" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.115, + 0.085, + 0.221, + 0.099 + ], + "angle": 0, + "content": "3.3 Dataset" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.106, + 0.488, + 0.186 + ], + "angle": 0, + "content": "To evaluate our heuristics, we use a corrected and improved version of the literary dataset of Dekker et al. (2019). This dataset is comprised of the first chapter of 40 novels in English, which we consider long enough for our experiments." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.194, + 0.489, + 0.29 + ], + "angle": 0, + "content": "Dataset corrections The original dataset suffers mainly from annotation issues. To fix them, we design an annotation guide inspired by CoNLL-2003 (Tjong Kim Sang and De Meulder, 2003) and apply it consistently using a semi-automated process:" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.296, + 0.489, + 0.392 + ], + "angle": 0, + "content": "1. We apply a set of simple rules to identify obvious errors\\(^3\\) (for example, non capitalized entities annotated as PER are often false positives). Depending on the estimated performance of each rule, we manually reviewed its choices before application." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.401, + 0.489, + 0.496 + ], + "angle": 0, + "content": "2. We manually review each difference between the predictions of a BERT (Devlin et al., 2019) model finetuned on a slightly modified version of the CoNLL-2003 dataset (Tjong Kim Sang and De Meulder, 2003)\\(^4\\) and the existing annotations." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.507, + 0.465, + 0.523 + ], + "angle": 0, + "content": "3. We manually correct the remaining errors." + }, + { + "type": "list", + "bbox": [ + 0.129, + 0.296, + 0.489, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.528, + 0.488, + 0.607 + ], + "angle": 0, + "content": "Further annotations The original dataset only consists of PER entities. We go further and annotate LOC and ORG entities. The final dataset contains 4476 PER entities, 886 LOC entities and 201 ORG entities." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.619, + 0.273, + 0.634 + ], + "angle": 0, + "content": "3.4 NER Training" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.639, + 0.488, + 0.768 + ], + "angle": 0, + "content": "For all experiments, we use a pretrained BERTBASE (Devlin et al., 2019) model, consisting in 110 million parameters, followed by a classification head at the token level to perform NER. We finetune BERT for 2 epochs with a learning rate of \\(2 \\cdot 10^{-5}\\) using the huggingface transformers library (Wolf et al., 2020), starting from the bert-base-cased checkpoint." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.778, + 0.286, + 0.792 + ], + "angle": 0, + "content": "3.5 NER evaluation" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.798, + 0.489, + 0.863 + ], + "angle": 0, + "content": "We perform cross-validation with 5 folds on our NER dataset. We evaluate NER performance using the default mode of the seqeval (Nakayama, 2018) python library to ensure results can be reproduced." + }, + { + "type": "page_footnote", + "bbox": [ + 0.136, + 0.868, + 0.325, + 0.881 + ], + "angle": 0, + "content": "See Appendix A.2 for details." + }, + { + "type": "page_footnote", + "bbox": [ + 0.116, + 0.882, + 0.488, + 0.919 + ], + "angle": 0, + "content": "4We modified the CoNLL-2003 dataset to include honorifics as part of PER entities to be consistent with our annotation guidelines." + }, + { + "type": "list", + "bbox": [ + 0.116, + 0.868, + 0.488, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.084, + 0.609, + 0.099 + ], + "angle": 0, + "content": "4 Results" + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.112, + 0.709, + 0.126 + ], + "angle": 0, + "content": "4.1 Retrieval heuristics" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.133, + 0.883, + 0.277 + ], + "angle": 0, + "content": "The NER performance for retrieval heuristics can be seen in Figure 1. The samenoun heuristic performs the best among global heuristics, whereas the surrounding heuristic is the best for local heuristics. While the top results obtained with both heuristics are quite similar, we consider global heuristics as naive retrieval baselines: they could be bested by more complex approaches, which might enhance performance even more." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.279, + 0.884, + 0.472 + ], + "angle": 0, + "content": "Interestingly, the performance of both before and bm25 heuristics decrease strongly after four sentences, and even drop behind the no retrieval baseline. For both heuristics, this might be due to retrieving irrelevant sentences after a while. The bm25 heuristic is limited by the similar sentences present in the document: if there are not enough of them, the heuristic will retrieve unrelated ones. Meanwhile, the case of the before heuristic seems more puzzling, and could be indicative of a specific entity mention pattern that might warrant more investigations." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.484, + 0.679, + 0.498 + ], + "angle": 0, + "content": "4.2 Oracle versions" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.506, + 0.882, + 0.537 + ], + "angle": 0, + "content": "NER results with the oracle versions of retrieval heuristics can be found in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.539, + 0.884, + 0.666 + ], + "angle": 0, + "content": "It is worth noting that the performance of the oracle versions of the heuristics always peaks when retrieving a single sentence. This might indicate that a single sentence is usually sufficient to resolve entity type ambiguities, but it might also be a result of the oracle ranking sentences individually, thereby not taking into account their possible combinations." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.668, + 0.882, + 0.765 + ], + "angle": 0, + "content": "Global heuristics perform better than local ones overall, with the oracle version of the random heuristic even performing better than both the before and after heuristics. These results tend to highlight the benefits of using global document context, provided it can be retrieved accurately." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.775, + 0.884, + 0.919 + ], + "angle": 0, + "content": "Retrieved sentences To better understand which sentences are useful for predictions when performing global retrieval, we plot in Figure 3 the distribution of the distance between sentences and their retrieved contexts for the oracle versions of heuristics samenoun and bm25. We find that \\(8\\%\\) and \\(16\\%\\) of retrieved sentences (for samenoun and bm25, respectively) are comprised within 6 sentences of their input sentence, while the other are" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.928, + 0.517, + 0.941 + ], + "angle": 0, + "content": "716" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.136, + 0.087, + 0.48, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.133, + 0.285, + 0.484, + 0.327 + ], + "angle": 0, + "content": "Figure 1: Mean F1 score versus max number of retrieved sentences for all retrieval heuristics across 3 runs." + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.086, + 0.858, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.285, + 0.865, + 0.327 + ], + "angle": 0, + "content": "Figure 2: Mean F1 score versus max number of retrieved sentences across 3 runs for oracle versions of all retrieval heuristics." + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.341, + 0.885, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.514, + 0.884, + 0.557 + ], + "angle": 0, + "content": "Figure 3: Distribution of the distance of retrieved sentences using the oracle versions of the samenoun and bm25 heuristics. The samenoun heuristic retrieves fewer sentences overall, since it is possible for some sentence to not have a common noun with any other sentence of its document." + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.584, + 0.487, + 0.794 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.807, + 0.489, + 0.863 + ], + "angle": 0, + "content": "Figure 4: Mean F1 score versus number of retrieved sentences across 3 runs for the oracle version of the bm25 heuristic, and the same heuristic restricted to distant context." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.904, + 0.487, + 0.92 + ], + "angle": 0, + "content": "further away, highlighting the need for long-range" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.584, + 0.58, + 0.597 + ], + "angle": 0, + "content": "retrieval." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.608, + 0.884, + 0.753 + ], + "angle": 0, + "content": "Local context importance To see whether or not local context is an important component of NER performance, we perform an experiment where we restrict the oracle version of the bm25 heuristic from retrieving local surrounding context. Results can be found in Figure 4. NER performance remains about the same without local context, which tends to show that local context is not strictly necessary for performance." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.765, + 0.797, + 0.78 + ], + "angle": 0, + "content": "5 Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.791, + 0.885, + 0.919 + ], + "angle": 0, + "content": "In this article, we explored the role of local and global context in Named Entity Recognition. Our results tend to show that, for literary texts, retrieving global document context is more effective at enhancing NER performance than retrieving only local context, even when using relatively simple retrieval heuristics. We also showed that a re-ranker model using simple document-level retrieval heuris" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.928, + 0.516, + 0.94 + ], + "angle": 0, + "content": "717" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.493, + 0.149 + ], + "angle": 0, + "content": "tics could obtain significant NER performance improvements. Overall, our work prompts for further research in how to accurately retrieve global context for NER." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.161, + 0.251, + 0.177 + ], + "angle": 0, + "content": "6 Limitations" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.186, + 0.49, + 0.217 + ], + "angle": 0, + "content": "We acknowledge the following limitations of our work:" + }, + { + "type": "text", + "bbox": [ + 0.139, + 0.228, + 0.49, + 0.326 + ], + "angle": 0, + "content": "- While the oracle selects a sentence according to the benefits it provides when performing NER, it does not consider the interactions between selected sentences. This may lead to lowered performances when the several sentences are retrieved at once." + }, + { + "type": "text", + "bbox": [ + 0.138, + 0.336, + 0.49, + 0.416 + ], + "angle": 0, + "content": "- The retrieval heuristics considered are naive on purpose, as the focus of this work is not performance. Stronger retrieval heuristics may achieve better results than presented in this article." + }, + { + "type": "text", + "bbox": [ + 0.138, + 0.427, + 0.49, + 0.507 + ], + "angle": 0, + "content": "- The studied documents only consist in the first chapter of a set of novels. Using complete novel would increase the number of possible information to retrieve for the presented global heuristics." + }, + { + "type": "list", + "bbox": [ + 0.138, + 0.228, + 0.49, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.116, + 0.544, + 0.214, + 0.559 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.567, + 0.49, + 0.607 + ], + "angle": 0, + "content": "I. Beltagy, M. E. Peters, and A. Cohan. 2020. Longformer: The long-document transformer. arXiv, cs.CL:2004.05150." + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.616, + 0.489, + 0.644 + ], + "angle": 0, + "content": "S. Bird, E. Loper, and E. Klein. 2009. Natural Language Processing with Python. O'Reilly Media Inc." + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.652, + 0.49, + 0.692 + ], + "angle": 0, + "content": "R. Child, S. Gray, A. Radford, and I. Sutskever. 2019. Generating long sequences with sparse transformers. arXiv, cs.LG:1904.10509." + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.702, + 0.49, + 0.768 + ], + "angle": 0, + "content": "K. Choromanski, V. Likhosherstov, D. Dohan, X. Song, A. Gane, T. Sarlos, P. Hawkins, J. Davis, A. Mohiuddin, L. Kaiser, D. Belanger, L. Colwell, and A. Weller. 2020. Rethinking attention with performers. arXiv, cs.LG:2009.14794." + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.777, + 0.49, + 0.83 + ], + "angle": 0, + "content": "N. Dekker, T. Kuhn, and M. van Erp. 2019. Evaluating named entity recognition tools for extracting social networks from novels. PeerJ Computer Science, 5:e189." + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.84, + 0.49, + 0.919 + ], + "angle": 0, + "content": "J. Devlin, M. Chang, K. Lee, and K. Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, volume 1, pages 4171-4186." + }, + { + "type": "list", + "bbox": [ + 0.116, + 0.567, + 0.49, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.086, + 0.885, + 0.138 + ], + "angle": 0, + "content": "M. Ding, C. Zhou, H. Yang, and J. Tang. 2020. CogLTX: Applying bert to long texts. In Advances in Neural Information Processing Systems, volume 33, pages 12792-12804." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.149, + 0.885, + 0.215 + ], + "angle": 0, + "content": "D. Guo, D. Tang, N. Duan, M. Zhou, and J. Yin. 2019. Coupling retrieval and meta-learning for context-dependent semantic parsing. In 57th Annual Meeting of the Association for Computational Linguistics, pages 855-866." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.225, + 0.885, + 0.291 + ], + "angle": 0, + "content": "A. Katharopoulos, A. Vyas, N. Pappas, and François Fleuret. 2020. Transformers are mnns: Fast autoregressive transformers with linear attention. In Proceedings of the 37th International Conference on Machine Learning, ICML'20." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.301, + 0.885, + 0.327 + ], + "angle": 0, + "content": "N. Kitaev, L. Kaiser, and A. Levskaya. 2020. Reformer: The efficient transformer. arXiv, cs.LG:2001.04451." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.337, + 0.885, + 0.391 + ], + "angle": 0, + "content": "T. Liu, J. Yao, and C. Lin. 2019. Towards improving neural named entity recognition with gazetteers. In 57th Annual Meeting of the Association for Computational Linguistics, pages 5301-5307." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.4, + 0.885, + 0.453 + ], + "angle": 0, + "content": "G. Luo, X. Huang, C. Lin, and Z. Nie. 2015. Joint entity recognition and disambiguation. In 2015 Conference on Empirical Methods in Natural Language Processing, pages 879-888." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.463, + 0.885, + 0.516 + ], + "angle": 0, + "content": "J. Luoma and S. Pyysalo. 2020. Exploring cross-sentence contexts for named entity recognition with BERT. In 28th International Conference on Computational Linguistics, pages 904-914." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.526, + 0.885, + 0.553 + ], + "angle": 0, + "content": "H. Nakayama. 2018. seqeval: A python framework for sequence labeling evaluation." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.562, + 0.885, + 0.641 + ], + "angle": 0, + "content": "A. Pouran Ben Veyseh, M. V. Nguyen, N. Ngo Trung, B. Min, and T. H. Nguyen. 2021. Modeling document-level context for event detection via important context selection. In Conference on Empirical Methods in Natural Language Processing, pages 5403-5413." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.651, + 0.885, + 0.692 + ], + "angle": 0, + "content": "S. E. W. Robertson. 1994. Some simple effective approximations to the 2-poisson model for probabilistic weighted retrieval. In SIGIR '94, pages 232-241." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.701, + 0.885, + 0.78 + ], + "angle": 0, + "content": "D. Seyler, T. Dembelova, L. Del Corro, J. Hoffart, and G. Weikum. 2018. A study of the importance of external knowledge in the named entity recognition task. In 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 241-246." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.79, + 0.885, + 0.855 + ], + "angle": 0, + "content": "T. Stanislawek, A. Wróblewska, A. Wojcicka, D. Ziembicki, and P. Biecek. 2019. Named entity recognition - is there a glass ceiling? In 23rd Conference on Computational Natural Language Learning, pages 624-633." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.866, + 0.885, + 0.918 + ], + "angle": 0, + "content": "Y. Tay, D. Bahri, D. Metzler, D. Juan, Z. Zhao, and C. Zheng. 2020a. Synthesizer: Rethinking self-attention in transformer models. arXiv, cs.CL:2005.00743." + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.086, + 0.885, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.517, + 0.941 + ], + "angle": 0, + "content": "718" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.086, + 0.489, + 0.125 + ], + "angle": 0, + "content": "Y. Tay, D. Bahri, L. Yang, D. Metzler, and D. Juan. 2020b. Sparse sinkhorn attention. arXiv, cs.LG:2002.11296." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.134, + 0.488, + 0.186 + ], + "angle": 0, + "content": "Y. Tay, M. Dehghani, S. Abnar, Y. Shen, D. Bahri, P. Pham, J. Rao, L. Yang, S. Ruder, and D. Metzler. 2020c. Long range arena: A benchmark for efficient transformers. arXiv, cs.LG:2011.04006." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.195, + 0.489, + 0.248 + ], + "angle": 0, + "content": "E. F. Tjong Kim Sang and F. De Meulder. 2003. Introduction to the CoNLL-2003 shared task: Language-independent named entity recognition. In 7th Conference on Natural Language Learning, pages 142-147." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.256, + 0.489, + 0.295 + ], + "angle": 0, + "content": "S. Wang, B. Z. Li, M. Khabsa, H. Fang, and H. Ma. 2020. Linformer: Self-attention with linear complexity. arXiv, cs.LG:2006.04768." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.304, + 0.488, + 0.396 + ], + "angle": 0, + "content": "X. Wang, Y. Jiang, N. Bach, T. Wang, Z. Huang, F. Huang, and K. Tu. 2021. Improving named entity recognition by external context retrieving and cooperative learning. In 59th Annual Meeting of the Association for Computational Linguistics and 11th International Joint Conference on Natural Language Processing, volume 1, pages 1800-1812." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.404, + 0.488, + 0.509 + ], + "angle": 0, + "content": "T. Wolf, L. Debut, V. Sanh, J. Chaumont, C. Delangue, A. Moi, P. Cistac, T. Rault, R. Louf, M. Funtowicz, J. Davison, S. Shleifer, P. von Platen, C. Ma, Y. Jernite, J. Plu, C. Xu, T. Le Scao, S. Gugger, M. Drame, Q. Lhoest, and A. M. Rush. 2020. Transformers: State-of-the-art natural language processing. In *Conference on Empirical Methods in Natural Language Processing: System Demonstrations*, pages 38-45." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.517, + 0.489, + 0.581 + ], + "angle": 0, + "content": "J. Xu, J. Crego, and J. Senellart. 2020. Boosting neural machine translation with similar translations. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1580-1590." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.591, + 0.489, + 0.657 + ], + "angle": 0, + "content": "I. Yamada, A. Asai, H. Shindo, H. Takeda, and Y. Matsumoto. 2020. LUKE: Deep contextualized entity representations with entity-aware self-attention. In Conference on Empirical Methods in Natural Language Processing, pages 6442-6454." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.665, + 0.489, + 0.743 + ], + "angle": 0, + "content": "M. Zaheer, G. Guruganesh, K. A. Dubey, J. Ainslie, C. Alberti, S. Ontanon, P. Pham, A. Ravula, Q. Wang, L. Yang, and A. Ahmed. 2020. Big bird: Transformers for longer sequences. In Advances in Neural Information Processing Systems, volume 33, pages 17283-17297." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.752, + 0.489, + 0.818 + ], + "angle": 0, + "content": "X. Zhang, Y. Jiang, X. Wang, X. Hu, Y. Sun, P. Xie, and M. Zhang. 2022. Domain-specific NER via retrieving correlated samples. In Proceedings of the 29th International Conference on Computational Linguistics, pages 2398-2404." + }, + { + "type": "list", + "bbox": [ + 0.116, + 0.086, + 0.489, + 0.818 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.116, + 0.826, + 0.284, + 0.84 + ], + "angle": 0, + "content": "A Dataset Details" + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.851, + 0.314, + 0.866 + ], + "angle": 0, + "content": "A.1 Document Lengths" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.872, + 0.489, + 0.919 + ], + "angle": 0, + "content": "Our NER dataset is composed of documents longer that typical NER datasets such as CoNLL-2003 (Tjong Kim Sang and De Meulder, 2003)." + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.085, + 0.878, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.289, + 0.883, + 0.332 + ], + "angle": 0, + "content": "Figure 5: Distribution of the number of sentences in our enhanced version of the dataset from Dekker et al. (2019)." + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.357, + 0.883, + 0.388 + ], + "angle": 0, + "content": "Figure 5 shows the distribution of the number of sentences of our NER dataset." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.4, + 0.78, + 0.414 + ], + "angle": 0, + "content": "A.2 Automatic Correction Rules" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.42, + 0.882, + 0.484 + ], + "angle": 0, + "content": "We use the following rules to automatically identify obvious errors in the original dataset from Dekker et al. (2019). The original dataset only contained PER entities, so these rules only apply to them:" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.495, + 0.882, + 0.557 + ], + "angle": 0, + "content": "- If a span appears in the list of characters from its novel but is not annotated as an entity, we investigate whether or not this is a false negative." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.569, + 0.882, + 0.633 + ], + "angle": 0, + "content": "- Similarly, if a span annotated as an entity does not appear in the list of characters from its novel, we investigate whether or not it is a false positive." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.644, + 0.882, + 0.691 + ], + "angle": 0, + "content": "- Finally, if a span is annotated as an entity but all of its tokens are not capitalized, we check if it is a false positive." + }, + { + "type": "list", + "bbox": [ + 0.532, + 0.495, + 0.882, + 0.691 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.702, + 0.833, + 0.734 + ], + "angle": 0, + "content": "B Heuristics Results Breakdown by Precision/Recall" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.744, + 0.883, + 0.822 + ], + "angle": 0, + "content": "Figures 6 and 7 show precision and recall for all retrieval heuristics. Interestingly, retrieval only has a positive effect on recall, with precision being lower than the baseline except for the surrounding heuristic." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.835, + 0.684, + 0.849 + ], + "angle": 0, + "content": "B.1 Oracle Versions" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.855, + 0.883, + 0.919 + ], + "angle": 0, + "content": "Figures 6 and 7 show precision and recall for the oracle versions of all retrieval heuristics. While retrieval benefits recall more than precision, precision is still increased using retrieval. Together with" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.928, + 0.516, + 0.941 + ], + "angle": 0, + "content": "719" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.136, + 0.085, + 0.478, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.133, + 0.282, + 0.484, + 0.325 + ], + "angle": 0, + "content": "Figure 6: Mean precision versus max number of retrieved sentences for all retrieval heuristics across 3 runs." + }, + { + "type": "image", + "bbox": [ + 0.136, + 0.341, + 0.476, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.133, + 0.541, + 0.487, + 0.584 + ], + "angle": 0, + "content": "Figure 8: Mean precision versus max number of retrieved sentences across 3 runs for oracle versions of all retrieval heuristics." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.61, + 0.49, + 0.658 + ], + "angle": 0, + "content": "the results from the regular heuristics, these results again highlight the potential performance gains of using a suitable re-ranker model to retrieve context." + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.085, + 0.858, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.282, + 0.866, + 0.324 + ], + "angle": 0, + "content": "Figure 7: Mean recall versus max number of retrieved sentences for all retrieval heuristics across 3 runs." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.341, + 0.858, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.541, + 0.866, + 0.584 + ], + "angle": 0, + "content": "Figure 9: Mean recall versus max number of retrieved sentences across 3 runs for oracle versions of all retrieval heuristics." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.928, + 0.517, + 0.941 + ], + "angle": 0, + "content": "720" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.133, + 0.085, + 0.435, + 0.1 + ], + "angle": 0, + "content": "ACL 2023 Responsible NLP Checklist" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.108, + 0.323, + 0.123 + ], + "angle": 0, + "content": "A For every submission:" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.127, + 0.533, + 0.144 + ], + "angle": 0, + "content": "A1. Did you describe the limitations of your work?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.145, + 0.462, + 0.159 + ], + "angle": 0, + "content": "Yes, limitations are discussed in Section 6" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.171, + 0.553, + 0.187 + ], + "angle": 0, + "content": "A2. Did you discuss any potential risks of your work?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.188, + 0.515, + 0.202 + ], + "angle": 0, + "content": "We do not think our work presents any direct risk" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.213, + 0.697, + 0.229 + ], + "angle": 0, + "content": "A3. Do the abstract and introduction summarize the paper's main claims?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.231, + 0.398, + 0.245 + ], + "angle": 0, + "content": "Yes, in the abstract and Section 1" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.257, + 0.67, + 0.273 + ], + "angle": 0, + "content": "A4. Have you used AI writing assistants when working on this paper?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.274, + 0.233, + 0.288 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.3, + 0.489, + 0.317 + ], + "angle": 0, + "content": "B Did you use or create scientific artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.322, + 0.882, + 0.353 + ], + "angle": 0, + "content": "in Section 3.4, we indicate that we use a BERT checkpoint. We also use a previous NER dataset, see Section 3.3. We distribute an enhanced version of this dataset and code to reproduce our experiments." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.363, + 0.531, + 0.38 + ], + "angle": 0, + "content": "B1. Did you cite the creators of artifacts you used?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.381, + 0.327, + 0.395 + ], + "angle": 0, + "content": "See Section 3.3 and 3.4" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.407, + 0.779, + 0.423 + ], + "angle": 0, + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.424, + 0.708, + 0.44 + ], + "angle": 0, + "content": "We specify the license in the Github repository given at the end of section 1." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.45, + 0.882, + 0.514 + ], + "angle": 0, + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.515, + 0.52, + 0.53 + ], + "angle": 0, + "content": "We use a dataset published for research purposes." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.541, + 0.882, + 0.589 + ], + "angle": 0, + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.59, + 0.76, + 0.605 + ], + "angle": 0, + "content": "Collected datas do not include information that can be used to identify individuals" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.615, + 0.882, + 0.648 + ], + "angle": 0, + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.649, + 0.882, + 0.681 + ], + "angle": 0, + "content": "We specify that the distributed dataset covers english literature (section 3.3). The reader can refer to Dekker et al., 2019 for more informations on the dataset." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.69, + 0.882, + 0.771 + ], + "angle": 0, + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be." + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.772, + 0.882, + 0.804 + ], + "angle": 0, + "content": "We include the number of document of our dataset in Section 3.3 We also include statistics about the length of these document in the Appendix" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.808, + 0.878, + 0.832 + ], + "angle": 0, + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.514, + 0.941 + ], + "angle": 0, + "content": "721" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.115, + 0.084, + 0.495, + 0.101 + ], + "angle": 0, + "content": "C Did you run computational experiments?" + }, + { + "type": "text", + "bbox": [ + 0.132, + 0.106, + 0.368, + 0.12 + ], + "angle": 0, + "content": "See Section 3.4 and Section 3.5" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.131, + 0.88, + 0.164 + ], + "angle": 0, + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.165, + 0.268, + 0.179 + ], + "angle": 0, + "content": "See Section 3.4" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.19, + 0.882, + 0.223 + ], + "angle": 0, + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.224, + 0.533, + 0.239 + ], + "angle": 0, + "content": "We include training hyperparameters in Section 3.4" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.249, + 0.882, + 0.297 + ], + "angle": 0, + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.299, + 0.882, + 0.33 + ], + "angle": 0, + "content": "Our results are reported in Section 4. We indicate that, for Figure 1 and 2, each point is the mean \\( F1 \\) of 3 runs." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.34, + 0.882, + 0.388 + ], + "angle": 0, + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.39, + 0.77, + 0.405 + ], + "angle": 0, + "content": "See Section 3.1 (nltk), Section 3.4 (huggingface transformers), Section 3.5 (sequeval)" + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.416, + 0.876, + 0.433 + ], + "angle": 0, + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.438, + 0.22, + 0.451 + ], + "angle": 0, + "content": "Section 3.3" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.464, + 0.882, + 0.495 + ], + "angle": 0, + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.497, + 0.436, + 0.511 + ], + "angle": 0, + "content": "The experiments were free of any risks" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.523, + 0.882, + 0.57 + ], + "angle": 0, + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.572, + 0.49, + 0.586 + ], + "angle": 0, + "content": "The authors annotated the dataset themselves" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.598, + 0.882, + 0.645 + ], + "angle": 0, + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.647, + 0.49, + 0.661 + ], + "angle": 0, + "content": "The authors annotated the dataset themselves" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.673, + 0.874, + 0.688 + ], + "angle": 0, + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.69, + 0.49, + 0.704 + ], + "angle": 0, + "content": "The authors annotated the dataset themselves" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.716, + 0.88, + 0.747 + ], + "angle": 0, + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.749, + 0.604, + 0.763 + ], + "angle": 0, + "content": "This is not relevant since annotation was done by the authors" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.928, + 0.517, + 0.941 + ], + "angle": 0, + "content": "722" + } + ] +] \ No newline at end of file diff --git a/2023/The Role of Global and Local Context in Named Entity Recognition/d05b6607-6d83-45a3-9ebc-fc5492b0e085_origin.pdf b/2023/The Role of Global and Local Context in Named Entity Recognition/d05b6607-6d83-45a3-9ebc-fc5492b0e085_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..68fc3bc15f083884304bd332dd4e5a87b7511d8d --- /dev/null +++ b/2023/The Role of Global and Local Context in Named Entity Recognition/d05b6607-6d83-45a3-9ebc-fc5492b0e085_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f8e7f3a296da072655183de1f02119a4c031c0f73c24791617b882fb611b695d +size 500429 diff --git a/2023/The Role of Global and Local Context in Named Entity Recognition/full.md b/2023/The Role of Global and Local Context in Named Entity Recognition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a49acae9bb8989f0ba8cbed8be367e0be4804cdf --- /dev/null +++ b/2023/The Role of Global and Local Context in Named Entity Recognition/full.md @@ -0,0 +1,314 @@ +# The Role of Global and Local Context in Named Entity Recognition + +Arthur Amalvy + +Laboratoire Informatique d'Avignon arthur.amalvy@univ-avignon.fr + +Vincent Labatut* + +Laboratoire Informatique d'Avignon vincent.labatut@univ-avignon.fr + +# Richard Dufour* + +Laboratoire des Sciences du Numérique de Nantes + +richard.dufour@univ-nantes.fr + +# Abstract + +Pre-trained transformer-based models have recently shown great performance when applied to Named Entity Recognition (NER). As the complexity of their self-attention mechanism prevents them from processing long documents at once, these models are usually applied in a sequential fashion. Such an approach unfortunately only incorporates local context and prevents leveraging global document context in long documents such as novels, which might hinder performance. In this article, we explore the impact of global document context, and its relationships with local context. We find that correctly retrieving global document context has a greater impact on performance than only leveraging local context, prompting for further research on how to better retrieve that context. + +# 1 Introduction + +Named Entity Recognition (NER) is a fundamental task in Natural Language Processing (NLP), and is often used as a building block for solving higher-level tasks. Recently, pre-trained transformer-based models such as BERT (Devlin et al., 2019) or LUKE (Yamada et al., 2020) showed great NER performance and have been able to push the state of the art further. + +These models, however, have a relatively short range because of the quadratic complexity of self-attention in the number of input tokens: as an example, BERT (Devlin et al., 2019) can only process spans of up to 512 tokens. For longer documents, texts are usually processed sequentially using a rolling window. Depending on the document, this local window may not always include all the context needed to perform inference, which may be present at the global document level. This leads to prediction errors (Stanislawek et al., 2019): In NER, this often occurs when the type of an entity cannot be inferred from the local context. For + +instance, in the following sentence from the fantasy novel *Elantris*, one cannot decide if the entity *Elantris* is a person (PER) or a location (LOC) without prior knowledge: + +"Raoden stood, and as he did, his eyes fell on Elantris again." + +In the novel, this prior knowledge comes from the fact that a human reader can recall previous mentions of Elantris, even at a very long range. A sequentially applied vanilla transformer-based model, however, might make an error without a neighboring sentence clearly establishing the status of Elantris as a city. + +While some works propose to retrieve external knowledge to disambiguate entities (Zhang et al., 2022; Wang et al., 2021), external resources are not always available. Furthermore, external retrieval might be more costly or less relevant than performing document-level context retrieval, provided the document contains the needed information, which depends on the type of document. + +Therefore, we wish to explore the relevance of document-level context when performing NER. We place ourselves at the sentence level, and we distinguish and study two types of contexts: + +- local context, consisting of surrounding sentences. This type of context can be used directly by vanilla transformer-based models, as their range lies beyond the simple sentence. Fully using surrounding context as in Devlin et al. (2019) is, however, computationally expensive. +- global context, consisting of all sentences available at the document level. To enhance NER prediction at the sentence level, we retrieve a few of these sentences and provide them as context for the model. + +We seek to answer the following question: is local context sufficient when solving the NER task, + +or would the model obtain better performance by retrieving global document context? + +To answer this question, we conduct experiments on a literary NER dataset we improved from its original version (Dekker et al., 2019). We release the annotation process, data and code necessary to reproduce these experiments under a free license1. + +# 2 Related Work + +# 2.1 Sparse Transformers + +Since the range problem of vanilla transformer-based models is due to the quadratic complexity of self-attention in the number of input tokens, several works on sparse transformers proposed alternative attention mechanisms in hope of reducing this complexity (Zaheer et al., 2020; Wang et al., 2020; Kitaev et al., 2020; Tay et al., 2020b,a; Beltagy et al., 2020; Choromanski et al., 2020; Katharopoulos et al., 2020; Child et al., 2019). While reducing self-attention complexity improves the effective range of transformers, these models still have issues processing very long documents (Tay et al., 2020c). + +# 2.2 Context retrieval + +Context retrieval in general has been widely leveraged for other NLP tasks, such as semantic parsing (Guo et al., 2019), question answering (Ding et al., 2020), event detection (Pouran Ben Veyseh et al., 2021), or machine translation (Xu et al., 2020). + +In NER, context retrieval has mainly been used in an external fashion, for example by leveraging names lists and gazetteers (Seyler et al., 2018; Liu et al., 2019), knowledge bases (Luo et al., 2015) or search engines (Wang et al., 2021; Zhang et al., 2022). Meanwhile, we are interested in document-level context retrieval, which is comparatively seldom explored. While Luoma and Pyysalo (2020) study document-level context, their study is restricted to neighboring sentences, i.e. local context. + +# 3 Method and Experiments + +# 3.1 Retrieval Heuristics + +We wish to understand the role of both local and global contexts for the NER task. We split all documents in our dataset (described in Section 3.3) into sentences. We evaluate both local and global + +simple heuristics of sentence retrieval in terms of NER performance impact. We study the following local heuristics: + +- before: Retrieves the closest $k$ sentences at the left of the input sentence. +- after: Same as before, but at the right of the input sentence. +- surrounding: Retrieves the closest $\frac{k}{2}$ sentences on both sides of the input sentence. + +And the following global heuristics: + +- random: Randomly retrieves a sentence from the whole document. +- samenoun: Randomly retrieves a sentence from the set of all sentences that have at least one common noun with the input sentence2. Intuitively, this heuristic will return sentences that contain entities of the input sentence, allowing for possible disambiguation. We use the NLTK library (Bird et al., 2009) to identify nouns. +- bm25: Retrieves sentences that are similar to the input sentences according to BM25 (Robertson, 1994). Retrieving similar sentences has already been found to increase NER performance (Zhang et al., 2022; Wang et al., 2021). + +It has to be noted that global heuristics can sometimes retrieve local context, as they are not restricted in which sentences they can retrieve at the document level. For all configurations, we concatenate the retrieved sentences to the input. During this concatenation step, we preserve the global order between sentences in the document. + +# 3.2 Oracles + +For each heuristic mentioned in Section 3.1, we also experiment with an oracle version. The oracle version retrieves 16 sentences from the document using the underlying retrieval heuristic, and retain only those that enhance the NER predictions the most. We measure this enhancement by counting the difference in numbers of NER BIO tags errors made with and without the context. In essence, the oracle setup simulates a perfect re-ranker model, and allows us to study the maximum performance of such an approach. + +# 3.3 Dataset + +To evaluate our heuristics, we use a corrected and improved version of the literary dataset of Dekker et al. (2019). This dataset is comprised of the first chapter of 40 novels in English, which we consider long enough for our experiments. + +Dataset corrections The original dataset suffers mainly from annotation issues. To fix them, we design an annotation guide inspired by CoNLL-2003 (Tjong Kim Sang and De Meulder, 2003) and apply it consistently using a semi-automated process: + +1. We apply a set of simple rules to identify obvious errors $^3$ (for example, non capitalized entities annotated as PER are often false positives). Depending on the estimated performance of each rule, we manually reviewed its choices before application. +2. We manually review each difference between the predictions of a BERT (Devlin et al., 2019) model finetuned on a slightly modified version of the CoNLL-2003 dataset (Tjong Kim Sang and De Meulder, 2003) $^4$ and the existing annotations. +3. We manually correct the remaining errors. + +Further annotations The original dataset only consists of PER entities. We go further and annotate LOC and ORG entities. The final dataset contains 4476 PER entities, 886 LOC entities and 201 ORG entities. + +# 3.4 NER Training + +For all experiments, we use a pretrained BERTBASE (Devlin et al., 2019) model, consisting in 110 million parameters, followed by a classification head at the token level to perform NER. We finetune BERT for 2 epochs with a learning rate of $2 \cdot 10^{-5}$ using the huggingface transformers library (Wolf et al., 2020), starting from the bert-base-cased checkpoint. + +# 3.5 NER evaluation + +We perform cross-validation with 5 folds on our NER dataset. We evaluate NER performance using the default mode of the seqeval (Nakayama, 2018) python library to ensure results can be reproduced. + +# 4 Results + +# 4.1 Retrieval heuristics + +The NER performance for retrieval heuristics can be seen in Figure 1. The samenoun heuristic performs the best among global heuristics, whereas the surrounding heuristic is the best for local heuristics. While the top results obtained with both heuristics are quite similar, we consider global heuristics as naive retrieval baselines: they could be bested by more complex approaches, which might enhance performance even more. + +Interestingly, the performance of both before and bm25 heuristics decrease strongly after four sentences, and even drop behind the no retrieval baseline. For both heuristics, this might be due to retrieving irrelevant sentences after a while. The bm25 heuristic is limited by the similar sentences present in the document: if there are not enough of them, the heuristic will retrieve unrelated ones. Meanwhile, the case of the before heuristic seems more puzzling, and could be indicative of a specific entity mention pattern that might warrant more investigations. + +# 4.2 Oracle versions + +NER results with the oracle versions of retrieval heuristics can be found in Figure 2. + +It is worth noting that the performance of the oracle versions of the heuristics always peaks when retrieving a single sentence. This might indicate that a single sentence is usually sufficient to resolve entity type ambiguities, but it might also be a result of the oracle ranking sentences individually, thereby not taking into account their possible combinations. + +Global heuristics perform better than local ones overall, with the oracle version of the random heuristic even performing better than both the before and after heuristics. These results tend to highlight the benefits of using global document context, provided it can be retrieved accurately. + +Retrieved sentences To better understand which sentences are useful for predictions when performing global retrieval, we plot in Figure 3 the distribution of the distance between sentences and their retrieved contexts for the oracle versions of heuristics samenoun and bm25. We find that $8\%$ and $16\%$ of retrieved sentences (for samenoun and bm25, respectively) are comprised within 6 sentences of their input sentence, while the other are + +![](images/d6e18f007e81be7c15635a88316c495df06714fd067ac3584533c3cae930ec43.jpg) +Figure 1: Mean F1 score versus max number of retrieved sentences for all retrieval heuristics across 3 runs. + +![](images/441cd4d4a47fec4f2597f28031882ee1f138192ce2a49d67ae808a745032d66e.jpg) +Figure 2: Mean F1 score versus max number of retrieved sentences across 3 runs for oracle versions of all retrieval heuristics. + +![](images/8fda74f838c6aee4fafbc6ab88a3ea88f01f147fbd24f89ac559d2a8ce124c80.jpg) +Figure 3: Distribution of the distance of retrieved sentences using the oracle versions of the samenoun and bm25 heuristics. The samenoun heuristic retrieves fewer sentences overall, since it is possible for some sentence to not have a common noun with any other sentence of its document. + +![](images/c1c7ed0cfeffe9fcd3edfab6c7c4efee7ef6784259ca99907d966e25400b7c8f.jpg) +Figure 4: Mean F1 score versus number of retrieved sentences across 3 runs for the oracle version of the bm25 heuristic, and the same heuristic restricted to distant context. + +further away, highlighting the need for long-range + +retrieval. + +Local context importance To see whether or not local context is an important component of NER performance, we perform an experiment where we restrict the oracle version of the bm25 heuristic from retrieving local surrounding context. Results can be found in Figure 4. NER performance remains about the same without local context, which tends to show that local context is not strictly necessary for performance. + +# 5 Conclusion and Future Work + +In this article, we explored the role of local and global context in Named Entity Recognition. Our results tend to show that, for literary texts, retrieving global document context is more effective at enhancing NER performance than retrieving only local context, even when using relatively simple retrieval heuristics. We also showed that a re-ranker model using simple document-level retrieval heuris + +tics could obtain significant NER performance improvements. Overall, our work prompts for further research in how to accurately retrieve global context for NER. + +# 6 Limitations + +We acknowledge the following limitations of our work: + +- While the oracle selects a sentence according to the benefits it provides when performing NER, it does not consider the interactions between selected sentences. This may lead to lowered performances when the several sentences are retrieved at once. +- The retrieval heuristics considered are naive on purpose, as the focus of this work is not performance. Stronger retrieval heuristics may achieve better results than presented in this article. +- The studied documents only consist in the first chapter of a set of novels. Using complete novel would increase the number of possible information to retrieve for the presented global heuristics. + +# References + +I. Beltagy, M. E. Peters, and A. Cohan. 2020. Longformer: The long-document transformer. arXiv, cs.CL:2004.05150. +S. Bird, E. Loper, and E. Klein. 2009. Natural Language Processing with Python. O'Reilly Media Inc. +R. Child, S. Gray, A. Radford, and I. Sutskever. 2019. Generating long sequences with sparse transformers. arXiv, cs.LG:1904.10509. +K. Choromanski, V. Likhosherstov, D. Dohan, X. Song, A. Gane, T. Sarlos, P. Hawkins, J. Davis, A. Mohiuddin, L. Kaiser, D. Belanger, L. Colwell, and A. Weller. 2020. Rethinking attention with performers. arXiv, cs.LG:2009.14794. +N. Dekker, T. Kuhn, and M. van Erp. 2019. Evaluating named entity recognition tools for extracting social networks from novels. PeerJ Computer Science, 5:e189. +J. Devlin, M. Chang, K. Lee, and K. Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, volume 1, pages 4171-4186. + +M. Ding, C. Zhou, H. Yang, and J. Tang. 2020. CogLTX: Applying bert to long texts. In Advances in Neural Information Processing Systems, volume 33, pages 12792-12804. +D. Guo, D. Tang, N. Duan, M. Zhou, and J. Yin. 2019. Coupling retrieval and meta-learning for context-dependent semantic parsing. In 57th Annual Meeting of the Association for Computational Linguistics, pages 855-866. +A. Katharopoulos, A. Vyas, N. Pappas, and François Fleuret. 2020. Transformers are mnns: Fast autoregressive transformers with linear attention. In Proceedings of the 37th International Conference on Machine Learning, ICML'20. +N. Kitaev, L. Kaiser, and A. Levskaya. 2020. Reformer: The efficient transformer. arXiv, cs.LG:2001.04451. +T. Liu, J. Yao, and C. Lin. 2019. Towards improving neural named entity recognition with gazetteers. In 57th Annual Meeting of the Association for Computational Linguistics, pages 5301-5307. +G. Luo, X. Huang, C. Lin, and Z. Nie. 2015. Joint entity recognition and disambiguation. In 2015 Conference on Empirical Methods in Natural Language Processing, pages 879-888. +J. Luoma and S. Pyysalo. 2020. Exploring cross-sentence contexts for named entity recognition with BERT. In 28th International Conference on Computational Linguistics, pages 904-914. +H. Nakayama. 2018. seqeval: A python framework for sequence labeling evaluation. +A. Pouran Ben Veyseh, M. V. Nguyen, N. Ngo Trung, B. Min, and T. H. Nguyen. 2021. Modeling document-level context for event detection via important context selection. In Conference on Empirical Methods in Natural Language Processing, pages 5403-5413. +S. E. W. Robertson. 1994. Some simple effective approximations to the 2-poisson model for probabilistic weighted retrieval. In SIGIR '94, pages 232-241. +D. Seyler, T. Dembelova, L. Del Corro, J. Hoffart, and G. Weikum. 2018. A study of the importance of external knowledge in the named entity recognition task. In 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 241-246. +T. Stanislawek, A. Wróblewska, A. Wojcicka, D. Ziembicki, and P. Biecek. 2019. Named entity recognition - is there a glass ceiling? In 23rd Conference on Computational Natural Language Learning, pages 624-633. +Y. Tay, D. Bahri, D. Metzler, D. Juan, Z. Zhao, and C. Zheng. 2020a. Synthesizer: Rethinking self-attention in transformer models. arXiv, cs.CL:2005.00743. + +Y. Tay, D. Bahri, L. Yang, D. Metzler, and D. Juan. 2020b. Sparse sinkhorn attention. arXiv, cs.LG:2002.11296. +Y. Tay, M. Dehghani, S. Abnar, Y. Shen, D. Bahri, P. Pham, J. Rao, L. Yang, S. Ruder, and D. Metzler. 2020c. Long range arena: A benchmark for efficient transformers. arXiv, cs.LG:2011.04006. +E. F. Tjong Kim Sang and F. De Meulder. 2003. Introduction to the CoNLL-2003 shared task: Language-independent named entity recognition. In 7th Conference on Natural Language Learning, pages 142-147. +S. Wang, B. Z. Li, M. Khabsa, H. Fang, and H. Ma. 2020. Linformer: Self-attention with linear complexity. arXiv, cs.LG:2006.04768. +X. Wang, Y. Jiang, N. Bach, T. Wang, Z. Huang, F. Huang, and K. Tu. 2021. Improving named entity recognition by external context retrieving and cooperative learning. In 59th Annual Meeting of the Association for Computational Linguistics and 11th International Joint Conference on Natural Language Processing, volume 1, pages 1800-1812. +T. Wolf, L. Debut, V. Sanh, J. Chaumont, C. Delangue, A. Moi, P. Cistac, T. Rault, R. Louf, M. Funtowicz, J. Davison, S. Shleifer, P. von Platen, C. Ma, Y. Jernite, J. Plu, C. Xu, T. Le Scao, S. Gugger, M. Drame, Q. Lhoest, and A. M. Rush. 2020. Transformers: State-of-the-art natural language processing. In *Conference on Empirical Methods in Natural Language Processing: System Demonstrations*, pages 38-45. +J. Xu, J. Crego, and J. Senellart. 2020. Boosting neural machine translation with similar translations. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1580-1590. +I. Yamada, A. Asai, H. Shindo, H. Takeda, and Y. Matsumoto. 2020. LUKE: Deep contextualized entity representations with entity-aware self-attention. In Conference on Empirical Methods in Natural Language Processing, pages 6442-6454. +M. Zaheer, G. Guruganesh, K. A. Dubey, J. Ainslie, C. Alberti, S. Ontanon, P. Pham, A. Ravula, Q. Wang, L. Yang, and A. Ahmed. 2020. Big bird: Transformers for longer sequences. In Advances in Neural Information Processing Systems, volume 33, pages 17283-17297. +X. Zhang, Y. Jiang, X. Wang, X. Hu, Y. Sun, P. Xie, and M. Zhang. 2022. Domain-specific NER via retrieving correlated samples. In Proceedings of the 29th International Conference on Computational Linguistics, pages 2398-2404. + +# A Dataset Details + +# A.1 Document Lengths + +Our NER dataset is composed of documents longer that typical NER datasets such as CoNLL-2003 (Tjong Kim Sang and De Meulder, 2003). + +![](images/4cab2bba3cf8bc1e551f8128ff3e963d3d928bf7ab5ed7b35a9f52337a3c93a9.jpg) +Figure 5: Distribution of the number of sentences in our enhanced version of the dataset from Dekker et al. (2019). +Figure 5 shows the distribution of the number of sentences of our NER dataset. + +# A.2 Automatic Correction Rules + +We use the following rules to automatically identify obvious errors in the original dataset from Dekker et al. (2019). The original dataset only contained PER entities, so these rules only apply to them: + +- If a span appears in the list of characters from its novel but is not annotated as an entity, we investigate whether or not this is a false negative. +- Similarly, if a span annotated as an entity does not appear in the list of characters from its novel, we investigate whether or not it is a false positive. +- Finally, if a span is annotated as an entity but all of its tokens are not capitalized, we check if it is a false positive. + +# B Heuristics Results Breakdown by Precision/Recall + +Figures 6 and 7 show precision and recall for all retrieval heuristics. Interestingly, retrieval only has a positive effect on recall, with precision being lower than the baseline except for the surrounding heuristic. + +# B.1 Oracle Versions + +Figures 6 and 7 show precision and recall for the oracle versions of all retrieval heuristics. While retrieval benefits recall more than precision, precision is still increased using retrieval. Together with + +![](images/82be396a06d23103061634f01b8dd2d3d558c6b6e6b009839b220b2481420ceb.jpg) +Figure 6: Mean precision versus max number of retrieved sentences for all retrieval heuristics across 3 runs. + +![](images/78048c3a9dc59a725d636f948057608a6b34cc70be64b4d8d43e0a2f8c3e85ed.jpg) +Figure 8: Mean precision versus max number of retrieved sentences across 3 runs for oracle versions of all retrieval heuristics. + +the results from the regular heuristics, these results again highlight the potential performance gains of using a suitable re-ranker model to retrieve context. + +![](images/d69e670218c71bfde2d923fe05ea806d34a133004f869063f15dcd780356bbcb.jpg) +Figure 7: Mean recall versus max number of retrieved sentences for all retrieval heuristics across 3 runs. + +![](images/7446c94113ae9db36eb9916d0e343f08ecd370ea34c02dd4e3087720be5041f3.jpg) +Figure 9: Mean recall versus max number of retrieved sentences across 3 runs for oracle versions of all retrieval heuristics. + +A For every submission: + +A1. Did you describe the limitations of your work? + +Yes, limitations are discussed in Section 6 + +A2. Did you discuss any potential risks of your work? + +We do not think our work presents any direct risk + +A3. Do the abstract and introduction summarize the paper's main claims? + +Yes, in the abstract and Section 1 + +A4. Have you used AI writing assistants when working on this paper? + +Left blank. + +B Did you use or create scientific artifacts? + +in Section 3.4, we indicate that we use a BERT checkpoint. We also use a previous NER dataset, see Section 3.3. We distribute an enhanced version of this dataset and code to reproduce our experiments. + +B1. Did you cite the creators of artifacts you used? + +See Section 3.3 and 3.4 + +B2. Did you discuss the license or terms for use and / or distribution of any artifacts? + +We specify the license in the Github repository given at the end of section 1. + +B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? + +We use a dataset published for research purposes. + +B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? + +Collected datas do not include information that can be used to identify individuals + +B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? + +We specify that the distributed dataset covers english literature (section 3.3). The reader can refer to Dekker et al., 2019 for more informations on the dataset. + +B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. + +We include the number of document of our dataset in Section 3.3 We also include statistics about the length of these document in the Appendix + +# C Did you run computational experiments? + +See Section 3.4 and Section 3.5 + +C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? + +See Section 3.4 + +C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? + +We include training hyperparameters in Section 3.4 + +C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? + +Our results are reported in Section 4. We indicate that, for Figure 1 and 2, each point is the mean $F1$ of 3 runs. + +C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? + +See Section 3.1 (nltk), Section 3.4 (huggingface transformers), Section 3.5 (sequeval) + +# D Did you use human annotators (e.g., crowdworkers) or research with human participants? + +Section 3.3 + +D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? + +The experiments were free of any risks + +D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? + +The authors annotated the dataset themselves + +D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? + +The authors annotated the dataset themselves + +D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? + +The authors annotated the dataset themselves + +D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? + +This is not relevant since annotation was done by the authors \ No newline at end of file diff --git a/2023/The Role of Global and Local Context in Named Entity Recognition/images.zip b/2023/The Role of Global and Local Context in Named Entity Recognition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..518dce7bcbf55f63c63667bfd3a8bcae4980eedb --- /dev/null +++ b/2023/The Role of Global and Local Context in Named Entity Recognition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0d8499ca7581d488278fb9667e100769422c48c87afa356cf135c533c14dd8f0 +size 239288 diff --git a/2023/The Role of Global and Local Context in Named Entity Recognition/layout.json b/2023/The Role of Global and Local Context in Named Entity Recognition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4f943ec6c9b35465ac9f52119c9fe7cd4db65e8b --- /dev/null +++ b/2023/The Role of Global and Local Context in Named Entity Recognition/layout.json @@ -0,0 +1,7099 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 88, + 70, + 506, + 88 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 70, + 506, + 88 + ], + "spans": [ + { + "bbox": [ + 88, + 70, + 506, + 88 + ], + "type": "text", + "content": "The Role of Global and Local Context in Named Entity Recognition" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 157, + 97, + 240, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 97, + 240, + 111 + ], + "spans": [ + { + "bbox": [ + 157, + 97, + 240, + 111 + ], + "type": "text", + "content": "Arthur Amalvy" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 110, + 111, + 287, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 111, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 110, + 111, + 287, + 137 + ], + "type": "text", + "content": "Laboratoire Informatique d'Avignon arthur.amalvy@univ-avignon.fr" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 351, + 97, + 440, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 97, + 440, + 110 + ], + "spans": [ + { + "bbox": [ + 351, + 97, + 440, + 110 + ], + "type": "text", + "content": "Vincent Labatut*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 302, + 111, + 490, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 111, + 490, + 137 + ], + "spans": [ + { + "bbox": [ + 302, + 111, + 490, + 137 + ], + "type": "text", + "content": "Laboratoire Informatique d'Avignon vincent.labatut@univ-avignon.fr" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 251, + 152, + 341, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 152, + 341, + 164 + ], + "spans": [ + { + "bbox": [ + 251, + 152, + 341, + 164 + ], + "type": "text", + "content": "Richard Dufour*" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 174, + 166, + 418, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 166, + 418, + 179 + ], + "spans": [ + { + "bbox": [ + 174, + 166, + 418, + 179 + ], + "type": "text", + "content": "Laboratoire des Sciences du Numérique de Nantes" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 210, + 180, + 384, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 180, + 384, + 192 + ], + "spans": [ + { + "bbox": [ + 210, + 180, + 384, + 192 + ], + "type": "text", + "content": "richard.dufour@univ-nantes.fr" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 155, + 212, + 202, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 212, + 202, + 225 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 202, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 84, + 235, + 274, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 235, + 274, + 439 + ], + "spans": [ + { + "bbox": [ + 84, + 235, + 274, + 439 + ], + "type": "text", + "content": "Pre-trained transformer-based models have recently shown great performance when applied to Named Entity Recognition (NER). As the complexity of their self-attention mechanism prevents them from processing long documents at once, these models are usually applied in a sequential fashion. Such an approach unfortunately only incorporates local context and prevents leveraging global document context in long documents such as novels, which might hinder performance. In this article, we explore the impact of global document context, and its relationships with local context. We find that correctly retrieving global document context has a greater impact on performance than only leveraging local context, prompting for further research on how to better retrieve that context." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 68, + 449, + 155, + 461 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 449, + 155, + 461 + ], + "spans": [ + { + "bbox": [ + 68, + 449, + 155, + 461 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 470, + 291, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 470, + 291, + 576 + ], + "spans": [ + { + "bbox": [ + 67, + 470, + 291, + 576 + ], + "type": "text", + "content": "Named Entity Recognition (NER) is a fundamental task in Natural Language Processing (NLP), and is often used as a building block for solving higher-level tasks. Recently, pre-trained transformer-based models such as BERT (Devlin et al., 2019) or LUKE (Yamada et al., 2020) showed great NER performance and have been able to push the state of the art further." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 67, + 578, + 291, + 754 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 578, + 291, + 754 + ], + "spans": [ + { + "bbox": [ + 67, + 578, + 291, + 754 + ], + "type": "text", + "content": "These models, however, have a relatively short range because of the quadratic complexity of self-attention in the number of input tokens: as an example, BERT (Devlin et al., 2019) can only process spans of up to 512 tokens. For longer documents, texts are usually processed sequentially using a rolling window. Depending on the document, this local window may not always include all the context needed to perform inference, which may be present at the global document level. This leads to prediction errors (Stanislawek et al., 2019): In NER, this often occurs when the type of an entity cannot be inferred from the local context. For" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 213, + 526, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 213, + 526, + 267 + ], + "spans": [ + { + "bbox": [ + 302, + 213, + 526, + 267 + ], + "type": "text", + "content": "instance, in the following sentence from the fantasy novel *Elantris*, one cannot decide if the entity *Elantris* is a person (PER) or a location (LOC) without prior knowledge:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 323, + 273, + 503, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 273, + 503, + 300 + ], + "spans": [ + { + "bbox": [ + 323, + 273, + 503, + 300 + ], + "type": "text", + "content": "\"Raoden stood, and as he did, his eyes fell on Elantris again.\"" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 306, + 526, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 306, + 526, + 400 + ], + "spans": [ + { + "bbox": [ + 302, + 306, + 526, + 400 + ], + "type": "text", + "content": "In the novel, this prior knowledge comes from the fact that a human reader can recall previous mentions of Elantris, even at a very long range. A sequentially applied vanilla transformer-based model, however, might make an error without a neighboring sentence clearly establishing the status of Elantris as a city." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 401, + 526, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 401, + 526, + 508 + ], + "spans": [ + { + "bbox": [ + 302, + 401, + 526, + 508 + ], + "type": "text", + "content": "While some works propose to retrieve external knowledge to disambiguate entities (Zhang et al., 2022; Wang et al., 2021), external resources are not always available. Furthermore, external retrieval might be more costly or less relevant than performing document-level context retrieval, provided the document contains the needed information, which depends on the type of document." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 302, + 509, + 526, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 509, + 526, + 563 + ], + "spans": [ + { + "bbox": [ + 302, + 509, + 526, + 563 + ], + "type": "text", + "content": "Therefore, we wish to explore the relevance of document-level context when performing NER. We place ourselves at the sentence level, and we distinguish and study two types of contexts:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 570, + 527, + 739 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 316, + 570, + 527, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 570, + 527, + 664 + ], + "spans": [ + { + "bbox": [ + 316, + 570, + 527, + 664 + ], + "type": "text", + "content": "- local context, consisting of surrounding sentences. This type of context can be used directly by vanilla transformer-based models, as their range lies beyond the simple sentence. Fully using surrounding context as in Devlin et al. (2019) is, however, computationally expensive." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 671, + 527, + 739 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 671, + 527, + 739 + ], + "spans": [ + { + "bbox": [ + 316, + 671, + 527, + 739 + ], + "type": "text", + "content": "- global context, consisting of all sentences available at the document level. To enhance NER prediction at the sentence level, we retrieve a few of these sentences and provide them as context for the model." + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 302, + 746, + 526, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 746, + 526, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 746, + 526, + 773 + ], + "type": "text", + "content": "We seek to answer the following question: is local context sufficient when solving the NER task," + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 81, + 760, + 211, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 760, + 211, + 772 + ], + "spans": [ + { + "bbox": [ + 81, + 760, + 211, + 772 + ], + "type": "text", + "content": "*These authors contributed equally." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "text", + "content": "714" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 135, + 795, + 459, + 805 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 795, + 459, + 805 + ], + "spans": [ + { + "bbox": [ + 135, + 795, + 459, + 805 + ], + "type": "text", + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 224, + 806, + 370, + 817 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 806, + 370, + 817 + ], + "spans": [ + { + "bbox": [ + 224, + 806, + 370, + 817 + ], + "type": "text", + "content": "Volume 2: Short Papers, pages 714-722" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "spans": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "text", + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 289, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 289, + 97 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 289, + 97 + ], + "type": "text", + "content": "or would the model obtain better performance by retrieving global document context?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 99, + 290, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 99, + 290, + 167 + ], + "spans": [ + { + "bbox": [ + 67, + 99, + 290, + 167 + ], + "type": "text", + "content": "To answer this question, we conduct experiments on a literary NER dataset we improved from its original version (Dekker et al., 2019). We release the annotation process, data and code necessary to reproduce these experiments under a free license1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 177, + 160, + 189 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 177, + 160, + 189 + ], + "spans": [ + { + "bbox": [ + 67, + 177, + 160, + 189 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 199, + 194, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 199, + 194, + 213 + ], + "spans": [ + { + "bbox": [ + 67, + 199, + 194, + 213 + ], + "type": "text", + "content": "2.1 Sparse Transformers" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 216, + 291, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 216, + 291, + 391 + ], + "spans": [ + { + "bbox": [ + 67, + 216, + 291, + 391 + ], + "type": "text", + "content": "Since the range problem of vanilla transformer-based models is due to the quadratic complexity of self-attention in the number of input tokens, several works on sparse transformers proposed alternative attention mechanisms in hope of reducing this complexity (Zaheer et al., 2020; Wang et al., 2020; Kitaev et al., 2020; Tay et al., 2020b,a; Beltagy et al., 2020; Choromanski et al., 2020; Katharopoulos et al., 2020; Child et al., 2019). While reducing self-attention complexity improves the effective range of transformers, these models still have issues processing very long documents (Tay et al., 2020c)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 403, + 175, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 403, + 175, + 414 + ], + "spans": [ + { + "bbox": [ + 67, + 403, + 175, + 414 + ], + "type": "text", + "content": "2.2 Context retrieval" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 421, + 291, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 421, + 291, + 500 + ], + "spans": [ + { + "bbox": [ + 67, + 421, + 291, + 500 + ], + "type": "text", + "content": "Context retrieval in general has been widely leveraged for other NLP tasks, such as semantic parsing (Guo et al., 2019), question answering (Ding et al., 2020), event detection (Pouran Ben Veyseh et al., 2021), or machine translation (Xu et al., 2020)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 502, + 291, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 502, + 291, + 638 + ], + "spans": [ + { + "bbox": [ + 67, + 502, + 291, + 638 + ], + "type": "text", + "content": "In NER, context retrieval has mainly been used in an external fashion, for example by leveraging names lists and gazetteers (Seyler et al., 2018; Liu et al., 2019), knowledge bases (Luo et al., 2015) or search engines (Wang et al., 2021; Zhang et al., 2022). Meanwhile, we are interested in document-level context retrieval, which is comparatively seldom explored. While Luoma and Pyysalo (2020) study document-level context, their study is restricted to neighboring sentences, i.e. local context." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 649, + 220, + 663 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 649, + 220, + 663 + ], + "spans": [ + { + "bbox": [ + 67, + 649, + 220, + 663 + ], + "type": "text", + "content": "3 Method and Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 671, + 189, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 671, + 189, + 682 + ], + "spans": [ + { + "bbox": [ + 67, + 671, + 189, + 682 + ], + "type": "text", + "content": "3.1 Retrieval Heuristics" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 688, + 290, + 743 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 688, + 290, + 743 + ], + "spans": [ + { + "bbox": [ + 67, + 688, + 290, + 743 + ], + "type": "text", + "content": "We wish to understand the role of both local and global contexts for the NER task. We split all documents in our dataset (described in Section 3.3) into sentences. We evaluate both local and global" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 71, + 525, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 525, + 111 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 525, + 111 + ], + "type": "text", + "content": "simple heuristics of sentence retrieval in terms of NER performance impact. We study the following local heuristics:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 117, + 525, + 214 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 316, + 117, + 524, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 117, + 524, + 144 + ], + "spans": [ + { + "bbox": [ + 316, + 117, + 524, + 144 + ], + "type": "text", + "content": "- before: Retrieves the closest " + }, + { + "bbox": [ + 316, + 117, + 524, + 144 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 316, + 117, + 524, + 144 + ], + "type": "text", + "content": " sentences at the left of the input sentence." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 152, + 524, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 524, + 179 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 524, + 179 + ], + "type": "text", + "content": "- after: Same as before, but at the right of the input sentence." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 185, + 525, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 525, + 214 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 525, + 214 + ], + "type": "text", + "content": "- surrounding: Retrieves the closest " + }, + { + "bbox": [ + 316, + 185, + 525, + 214 + ], + "type": "inline_equation", + "content": "\\frac{k}{2}" + }, + { + "bbox": [ + 316, + 185, + 525, + 214 + ], + "type": "text", + "content": " sentences on both sides of the input sentence." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 220, + 475, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 220, + 475, + 232 + ], + "spans": [ + { + "bbox": [ + 314, + 220, + 475, + 232 + ], + "type": "text", + "content": "And the following global heuristics:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 239, + 525, + 469 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 315, + 239, + 524, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 239, + 524, + 264 + ], + "spans": [ + { + "bbox": [ + 315, + 239, + 524, + 264 + ], + "type": "text", + "content": "- random: Randomly retrieves a sentence from the whole document." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 274, + 525, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 525, + 380 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 525, + 380 + ], + "type": "text", + "content": "- samenoun: Randomly retrieves a sentence from the set of all sentences that have at least one common noun with the input sentence2. Intuitively, this heuristic will return sentences that contain entities of the input sentence, allowing for possible disambiguation. We use the NLTK library (Bird et al., 2009) to identify nouns." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 390, + 525, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 390, + 525, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 390, + 525, + 469 + ], + "type": "text", + "content": "- bm25: Retrieves sentences that are similar to the input sentences according to BM25 (Robertson, 1994). Retrieving similar sentences has already been found to increase NER performance (Zhang et al., 2022; Wang et al., 2021)." + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 302, + 476, + 525, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 476, + 525, + 570 + ], + "spans": [ + { + "bbox": [ + 302, + 476, + 525, + 570 + ], + "type": "text", + "content": "It has to be noted that global heuristics can sometimes retrieve local context, as they are not restricted in which sentences they can retrieve at the document level. For all configurations, we concatenate the retrieved sentences to the input. During this concatenation step, we preserve the global order between sentences in the document." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 302, + 579, + 367, + 591 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 579, + 367, + 591 + ], + "spans": [ + { + "bbox": [ + 302, + 579, + 367, + 591 + ], + "type": "text", + "content": "3.2 Oracles" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 301, + 596, + 525, + 745 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 596, + 525, + 745 + ], + "spans": [ + { + "bbox": [ + 301, + 596, + 525, + 745 + ], + "type": "text", + "content": "For each heuristic mentioned in Section 3.1, we also experiment with an oracle version. The oracle version retrieves 16 sentences from the document using the underlying retrieval heuristic, and retain only those that enhance the NER predictions the most. We measure this enhancement by counting the difference in numbers of NER BIO tags errors made with and without the context. In essence, the oracle setup simulates a perfect re-ranker model, and allows us to study the maximum performance of such an approach." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 303, + 751, + 524, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 751, + 524, + 772 + ], + "spans": [ + { + "bbox": [ + 303, + 751, + 524, + 772 + ], + "type": "text", + "content": "2If the set of sentences with a common noun is empty, the samenoun heuristic does not retrieve any sentence." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 67, + 750, + 267, + 771 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 750, + 267, + 771 + ], + "spans": [ + { + "bbox": [ + 67, + 750, + 267, + 771 + ], + "type": "text", + "content": "1https://github.com/CompNet/conivel/tree/ACL2023" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "text", + "content": "715" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 71, + 131, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 71, + 131, + 83 + ], + "spans": [ + { + "bbox": [ + 68, + 71, + 131, + 83 + ], + "type": "text", + "content": "3.3 Dataset" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 89, + 290, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 89, + 290, + 156 + ], + "spans": [ + { + "bbox": [ + 67, + 89, + 290, + 156 + ], + "type": "text", + "content": "To evaluate our heuristics, we use a corrected and improved version of the literary dataset of Dekker et al. (2019). This dataset is comprised of the first chapter of 40 novels in English, which we consider long enough for our experiments." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 163, + 290, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 163, + 290, + 243 + ], + "spans": [ + { + "bbox": [ + 67, + 163, + 290, + 243 + ], + "type": "text", + "content": "Dataset corrections The original dataset suffers mainly from annotation issues. To fix them, we design an annotation guide inspired by CoNLL-2003 (Tjong Kim Sang and De Meulder, 2003) and apply it consistently using a semi-automated process:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 76, + 248, + 290, + 439 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 77, + 248, + 290, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 248, + 290, + 329 + ], + "spans": [ + { + "bbox": [ + 77, + 248, + 290, + 329 + ], + "type": "text", + "content": "1. We apply a set of simple rules to identify obvious errors" + }, + { + "bbox": [ + 77, + 248, + 290, + 329 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 77, + 248, + 290, + 329 + ], + "type": "text", + "content": " (for example, non capitalized entities annotated as PER are often false positives). Depending on the estimated performance of each rule, we manually reviewed its choices before application." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 76, + 337, + 290, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 337, + 290, + 417 + ], + "spans": [ + { + "bbox": [ + 76, + 337, + 290, + 417 + ], + "type": "text", + "content": "2. We manually review each difference between the predictions of a BERT (Devlin et al., 2019) model finetuned on a slightly modified version of the CoNLL-2003 dataset (Tjong Kim Sang and De Meulder, 2003)" + }, + { + "bbox": [ + 76, + 337, + 290, + 417 + ], + "type": "inline_equation", + "content": "^4" + }, + { + "bbox": [ + 76, + 337, + 290, + 417 + ], + "type": "text", + "content": " and the existing annotations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 76, + 426, + 276, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 426, + 276, + 439 + ], + "spans": [ + { + "bbox": [ + 76, + 426, + 276, + 439 + ], + "type": "text", + "content": "3. We manually correct the remaining errors." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 67, + 444, + 290, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 444, + 290, + 510 + ], + "spans": [ + { + "bbox": [ + 67, + 444, + 290, + 510 + ], + "type": "text", + "content": "Further annotations The original dataset only consists of PER entities. We go further and annotate LOC and ORG entities. The final dataset contains 4476 PER entities, 886 LOC entities and 201 ORG entities." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 520, + 162, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 520, + 162, + 533 + ], + "spans": [ + { + "bbox": [ + 67, + 520, + 162, + 533 + ], + "type": "text", + "content": "3.4 NER Training" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 537, + 290, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 537, + 290, + 645 + ], + "spans": [ + { + "bbox": [ + 67, + 537, + 290, + 645 + ], + "type": "text", + "content": "For all experiments, we use a pretrained BERTBASE (Devlin et al., 2019) model, consisting in 110 million parameters, followed by a classification head at the token level to perform NER. We finetune BERT for 2 epochs with a learning rate of " + }, + { + "bbox": [ + 67, + 537, + 290, + 645 + ], + "type": "inline_equation", + "content": "2 \\cdot 10^{-5}" + }, + { + "bbox": [ + 67, + 537, + 290, + 645 + ], + "type": "text", + "content": " using the huggingface transformers library (Wolf et al., 2020), starting from the bert-base-cased checkpoint." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 654, + 170, + 666 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 654, + 170, + 666 + ], + "spans": [ + { + "bbox": [ + 67, + 654, + 170, + 666 + ], + "type": "text", + "content": "3.5 NER evaluation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 671, + 290, + 725 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 671, + 290, + 725 + ], + "spans": [ + { + "bbox": [ + 67, + 671, + 290, + 725 + ], + "type": "text", + "content": "We perform cross-validation with 5 folds on our NER dataset. We evaluate NER performance using the default mode of the seqeval (Nakayama, 2018) python library to ensure results can be reproduced." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 303, + 70, + 362, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 70, + 362, + 83 + ], + "spans": [ + { + "bbox": [ + 303, + 70, + 362, + 83 + ], + "type": "text", + "content": "4 Results" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 94, + 421, + 105 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 94, + 421, + 105 + ], + "spans": [ + { + "bbox": [ + 302, + 94, + 421, + 105 + ], + "type": "text", + "content": "4.1 Retrieval heuristics" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 302, + 111, + 525, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 111, + 525, + 232 + ], + "spans": [ + { + "bbox": [ + 302, + 111, + 525, + 232 + ], + "type": "text", + "content": "The NER performance for retrieval heuristics can be seen in Figure 1. The samenoun heuristic performs the best among global heuristics, whereas the surrounding heuristic is the best for local heuristics. While the top results obtained with both heuristics are quite similar, we consider global heuristics as naive retrieval baselines: they could be bested by more complex approaches, which might enhance performance even more." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 302, + 234, + 525, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 234, + 525, + 396 + ], + "spans": [ + { + "bbox": [ + 302, + 234, + 525, + 396 + ], + "type": "text", + "content": "Interestingly, the performance of both before and bm25 heuristics decrease strongly after four sentences, and even drop behind the no retrieval baseline. For both heuristics, this might be due to retrieving irrelevant sentences after a while. The bm25 heuristic is limited by the similar sentences present in the document: if there are not enough of them, the heuristic will retrieve unrelated ones. Meanwhile, the case of the before heuristic seems more puzzling, and could be indicative of a specific entity mention pattern that might warrant more investigations." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 302, + 407, + 404, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 407, + 404, + 418 + ], + "spans": [ + { + "bbox": [ + 302, + 407, + 404, + 418 + ], + "type": "text", + "content": "4.2 Oracle versions" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 302, + 425, + 524, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 425, + 524, + 451 + ], + "spans": [ + { + "bbox": [ + 302, + 425, + 524, + 451 + ], + "type": "text", + "content": "NER results with the oracle versions of retrieval heuristics can be found in Figure 2." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 302, + 453, + 525, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 453, + 525, + 560 + ], + "spans": [ + { + "bbox": [ + 302, + 453, + 525, + 560 + ], + "type": "text", + "content": "It is worth noting that the performance of the oracle versions of the heuristics always peaks when retrieving a single sentence. This might indicate that a single sentence is usually sufficient to resolve entity type ambiguities, but it might also be a result of the oracle ranking sentences individually, thereby not taking into account their possible combinations." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 302, + 561, + 524, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 561, + 524, + 643 + ], + "spans": [ + { + "bbox": [ + 302, + 561, + 524, + 643 + ], + "type": "text", + "content": "Global heuristics perform better than local ones overall, with the oracle version of the random heuristic even performing better than both the before and after heuristics. These results tend to highlight the benefits of using global document context, provided it can be retrieved accurately." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 302, + 651, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 651, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 651, + 525, + 772 + ], + "type": "text", + "content": "Retrieved sentences To better understand which sentences are useful for predictions when performing global retrieval, we plot in Figure 3 the distribution of the distance between sentences and their retrieved contexts for the oracle versions of heuristics samenoun and bm25. We find that " + }, + { + "bbox": [ + 302, + 651, + 525, + 772 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 302, + 651, + 525, + 772 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 302, + 651, + 525, + 772 + ], + "type": "inline_equation", + "content": "16\\%" + }, + { + "bbox": [ + 302, + 651, + 525, + 772 + ], + "type": "text", + "content": " of retrieved sentences (for samenoun and bm25, respectively) are comprised within 6 sentences of their input sentence, while the other are" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 80, + 729, + 193, + 740 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 729, + 193, + 740 + ], + "spans": [ + { + "bbox": [ + 80, + 729, + 193, + 740 + ], + "type": "text", + "content": "See Appendix A.2 for details." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 69, + 741, + 290, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 741, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 741, + 290, + 772 + ], + "type": "text", + "content": "4We modified the CoNLL-2003 dataset to include honorifics as part of PER entities to be consistent with our annotation guidelines." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "text", + "content": "716" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 73, + 285, + 230 + ], + "blocks": [ + { + "bbox": [ + 80, + 73, + 285, + 230 + ], + "lines": [ + { + "bbox": [ + 80, + 73, + 285, + 230 + ], + "spans": [ + { + "bbox": [ + 80, + 73, + 285, + 230 + ], + "type": "image", + "image_path": "d6e18f007e81be7c15635a88316c495df06714fd067ac3584533c3cae930ec43.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 79, + 239, + 287, + 275 + ], + "lines": [ + { + "bbox": [ + 79, + 239, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 79, + 239, + 287, + 275 + ], + "type": "text", + "content": "Figure 1: Mean F1 score versus max number of retrieved sentences for all retrieval heuristics across 3 runs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 306, + 72, + 510, + 229 + ], + "blocks": [ + { + "bbox": [ + 306, + 72, + 510, + 229 + ], + "lines": [ + { + "bbox": [ + 306, + 72, + 510, + 229 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 510, + 229 + ], + "type": "image", + "image_path": "441cd4d4a47fec4f2597f28031882ee1f138192ce2a49d67ae808a745032d66e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 239, + 514, + 275 + ], + "lines": [ + { + "bbox": [ + 305, + 239, + 514, + 275 + ], + "spans": [ + { + "bbox": [ + 305, + 239, + 514, + 275 + ], + "type": "text", + "content": "Figure 2: Mean F1 score versus max number of retrieved sentences across 3 runs for oracle versions of all retrieval heuristics." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 69, + 286, + 526, + 423 + ], + "blocks": [ + { + "bbox": [ + 69, + 286, + 526, + 423 + ], + "lines": [ + { + "bbox": [ + 69, + 286, + 526, + 423 + ], + "spans": [ + { + "bbox": [ + 69, + 286, + 526, + 423 + ], + "type": "image", + "image_path": "8fda74f838c6aee4fafbc6ab88a3ea88f01f147fbd24f89ac559d2a8ce124c80.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 432, + 525, + 468 + ], + "lines": [ + { + "bbox": [ + 67, + 432, + 525, + 468 + ], + "spans": [ + { + "bbox": [ + 67, + 432, + 525, + 468 + ], + "type": "text", + "content": "Figure 3: Distribution of the distance of retrieved sentences using the oracle versions of the samenoun and bm25 heuristics. The samenoun heuristic retrieves fewer sentences overall, since it is possible for some sentence to not have a common noun with any other sentence of its document." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 69, + 491, + 289, + 667 + ], + "blocks": [ + { + "bbox": [ + 69, + 491, + 289, + 667 + ], + "lines": [ + { + "bbox": [ + 69, + 491, + 289, + 667 + ], + "spans": [ + { + "bbox": [ + 69, + 491, + 289, + 667 + ], + "type": "image", + "image_path": "c1c7ed0cfeffe9fcd3edfab6c7c4efee7ef6784259ca99907d966e25400b7c8f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 678, + 290, + 725 + ], + "lines": [ + { + "bbox": [ + 67, + 678, + 290, + 725 + ], + "spans": [ + { + "bbox": [ + 67, + 678, + 290, + 725 + ], + "type": "text", + "content": "Figure 4: Mean F1 score versus number of retrieved sentences across 3 runs for the oracle version of the bm25 heuristic, and the same heuristic restricted to distant context." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 760, + 289, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 760, + 289, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 760, + 289, + 773 + ], + "type": "text", + "content": "further away, highlighting the need for long-range" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 303, + 491, + 345, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 491, + 345, + 502 + ], + "spans": [ + { + "bbox": [ + 303, + 491, + 345, + 502 + ], + "type": "text", + "content": "retrieval." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 511, + 525, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 511, + 525, + 633 + ], + "spans": [ + { + "bbox": [ + 302, + 511, + 525, + 633 + ], + "type": "text", + "content": "Local context importance To see whether or not local context is an important component of NER performance, we perform an experiment where we restrict the oracle version of the bm25 heuristic from retrieving local surrounding context. Results can be found in Figure 4. NER performance remains about the same without local context, which tends to show that local context is not strictly necessary for performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 643, + 474, + 655 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 643, + 474, + 655 + ], + "spans": [ + { + "bbox": [ + 302, + 643, + 474, + 655 + ], + "type": "text", + "content": "5 Conclusion and Future Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 665, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 665, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 665, + 526, + 772 + ], + "type": "text", + "content": "In this article, we explored the role of local and global context in Named Entity Recognition. Our results tend to show that, for literary texts, retrieving global document context is more effective at enhancing NER performance than retrieving only local context, even when using relatively simple retrieval heuristics. We also showed that a re-ranker model using simple document-level retrieval heuris" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 780, + 307, + 790 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 780, + 307, + 790 + ], + "spans": [ + { + "bbox": [ + 289, + 780, + 307, + 790 + ], + "type": "text", + "content": "717" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 293, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 293, + 125 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 293, + 125 + ], + "type": "text", + "content": "tics could obtain significant NER performance improvements. Overall, our work prompts for further research in how to accurately retrieve global context for NER." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 135, + 149, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 135, + 149, + 148 + ], + "spans": [ + { + "bbox": [ + 67, + 135, + 149, + 148 + ], + "type": "text", + "content": "6 Limitations" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 156, + 291, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 156, + 291, + 182 + ], + "spans": [ + { + "bbox": [ + 67, + 156, + 291, + 182 + ], + "type": "text", + "content": "We acknowledge the following limitations of our work:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 82, + 191, + 291, + 426 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 82, + 191, + 291, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 191, + 291, + 274 + ], + "spans": [ + { + "bbox": [ + 82, + 191, + 291, + 274 + ], + "type": "text", + "content": "- While the oracle selects a sentence according to the benefits it provides when performing NER, it does not consider the interactions between selected sentences. This may lead to lowered performances when the several sentences are retrieved at once." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 282, + 291, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 282, + 291, + 349 + ], + "spans": [ + { + "bbox": [ + 82, + 282, + 291, + 349 + ], + "type": "text", + "content": "- The retrieval heuristics considered are naive on purpose, as the focus of this work is not performance. Stronger retrieval heuristics may achieve better results than presented in this article." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 82, + 359, + 291, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 359, + 291, + 426 + ], + "spans": [ + { + "bbox": [ + 82, + 359, + 291, + 426 + ], + "type": "text", + "content": "- The studied documents only consist in the first chapter of a set of novels. Using complete novel would increase the number of possible information to retrieve for the presented global heuristics." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 69, + 457, + 127, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 457, + 127, + 470 + ], + "spans": [ + { + "bbox": [ + 69, + 457, + 127, + 470 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 476, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 69, + 476, + 291, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 476, + 291, + 510 + ], + "spans": [ + { + "bbox": [ + 69, + 476, + 291, + 510 + ], + "type": "text", + "content": "I. Beltagy, M. E. Peters, and A. Cohan. 2020. Longformer: The long-document transformer. arXiv, cs.CL:2004.05150." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 518, + 290, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 518, + 290, + 541 + ], + "spans": [ + { + "bbox": [ + 69, + 518, + 290, + 541 + ], + "type": "text", + "content": "S. Bird, E. Loper, and E. Klein. 2009. Natural Language Processing with Python. O'Reilly Media Inc." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 548, + 291, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 548, + 291, + 581 + ], + "spans": [ + { + "bbox": [ + 69, + 548, + 291, + 581 + ], + "type": "text", + "content": "R. Child, S. Gray, A. Radford, and I. Sutskever. 2019. Generating long sequences with sparse transformers. arXiv, cs.LG:1904.10509." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 69, + 590, + 291, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 590, + 291, + 645 + ], + "spans": [ + { + "bbox": [ + 69, + 590, + 291, + 645 + ], + "type": "text", + "content": "K. Choromanski, V. Likhosherstov, D. Dohan, X. Song, A. Gane, T. Sarlos, P. Hawkins, J. Davis, A. Mohiuddin, L. Kaiser, D. Belanger, L. Colwell, and A. Weller. 2020. Rethinking attention with performers. arXiv, cs.LG:2009.14794." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 69, + 653, + 291, + 698 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 653, + 291, + 698 + ], + "spans": [ + { + "bbox": [ + 69, + 653, + 291, + 698 + ], + "type": "text", + "content": "N. Dekker, T. Kuhn, and M. van Erp. 2019. Evaluating named entity recognition tools for extracting social networks from novels. PeerJ Computer Science, 5:e189." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 69, + 706, + 291, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 706, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 706, + 291, + 772 + ], + "type": "text", + "content": "J. Devlin, M. Chang, K. Lee, and K. Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, volume 1, pages 4171-4186." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 526, + 772 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 304, + 72, + 526, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 526, + 116 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 526, + 116 + ], + "type": "text", + "content": "M. Ding, C. Zhou, H. Yang, and J. Tang. 2020. CogLTX: Applying bert to long texts. In Advances in Neural Information Processing Systems, volume 33, pages 12792-12804." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 125, + 526, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 125, + 526, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 125, + 526, + 180 + ], + "type": "text", + "content": "D. Guo, D. Tang, N. Duan, M. Zhou, and J. Yin. 2019. Coupling retrieval and meta-learning for context-dependent semantic parsing. In 57th Annual Meeting of the Association for Computational Linguistics, pages 855-866." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 189, + 526, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 189, + 526, + 244 + ], + "spans": [ + { + "bbox": [ + 304, + 189, + 526, + 244 + ], + "type": "text", + "content": "A. Katharopoulos, A. Vyas, N. Pappas, and François Fleuret. 2020. Transformers are mnns: Fast autoregressive transformers with linear attention. In Proceedings of the 37th International Conference on Machine Learning, ICML'20." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 253, + 526, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 253, + 526, + 275 + ], + "spans": [ + { + "bbox": [ + 304, + 253, + 526, + 275 + ], + "type": "text", + "content": "N. Kitaev, L. Kaiser, and A. Levskaya. 2020. Reformer: The efficient transformer. arXiv, cs.LG:2001.04451." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 283, + 526, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 283, + 526, + 328 + ], + "spans": [ + { + "bbox": [ + 304, + 283, + 526, + 328 + ], + "type": "text", + "content": "T. Liu, J. Yao, and C. Lin. 2019. Towards improving neural named entity recognition with gazetteers. In 57th Annual Meeting of the Association for Computational Linguistics, pages 5301-5307." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 336, + 526, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 336, + 526, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 336, + 526, + 380 + ], + "type": "text", + "content": "G. Luo, X. Huang, C. Lin, and Z. Nie. 2015. Joint entity recognition and disambiguation. In 2015 Conference on Empirical Methods in Natural Language Processing, pages 879-888." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 389, + 526, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 389, + 526, + 433 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 526, + 433 + ], + "type": "text", + "content": "J. Luoma and S. Pyysalo. 2020. Exploring cross-sentence contexts for named entity recognition with BERT. In 28th International Conference on Computational Linguistics, pages 904-914." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 442, + 526, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 442, + 526, + 465 + ], + "spans": [ + { + "bbox": [ + 304, + 442, + 526, + 465 + ], + "type": "text", + "content": "H. Nakayama. 2018. seqeval: A python framework for sequence labeling evaluation." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 472, + 526, + 539 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 526, + 539 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 526, + 539 + ], + "type": "text", + "content": "A. Pouran Ben Veyseh, M. V. Nguyen, N. Ngo Trung, B. Min, and T. H. Nguyen. 2021. Modeling document-level context for event detection via important context selection. In Conference on Empirical Methods in Natural Language Processing, pages 5403-5413." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 547, + 526, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 547, + 526, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 547, + 526, + 581 + ], + "type": "text", + "content": "S. E. W. Robertson. 1994. Some simple effective approximations to the 2-poisson model for probabilistic weighted retrieval. In SIGIR '94, pages 232-241." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 589, + 526, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 589, + 526, + 655 + ], + "spans": [ + { + "bbox": [ + 304, + 589, + 526, + 655 + ], + "type": "text", + "content": "D. Seyler, T. Dembelova, L. Del Corro, J. Hoffart, and G. Weikum. 2018. A study of the importance of external knowledge in the named entity recognition task. In 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 241-246." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 664, + 526, + 719 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 664, + 526, + 719 + ], + "spans": [ + { + "bbox": [ + 304, + 664, + 526, + 719 + ], + "type": "text", + "content": "T. Stanislawek, A. Wróblewska, A. Wojcicka, D. Ziembicki, and P. Biecek. 2019. Named entity recognition - is there a glass ceiling? In 23rd Conference on Computational Natural Language Learning, pages 624-633." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 728, + 526, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 728, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 728, + 526, + 772 + ], + "type": "text", + "content": "Y. Tay, D. Bahri, D. Metzler, D. Juan, Z. Zhao, and C. Zheng. 2020a. Synthesizer: Rethinking self-attention in transformer models. arXiv, cs.CL:2005.00743." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "718" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 290, + 687 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 69, + 72, + 290, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 290, + 105 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 290, + 105 + ], + "type": "text", + "content": "Y. Tay, D. Bahri, L. Yang, D. Metzler, and D. Juan. 2020b. Sparse sinkhorn attention. arXiv, cs.LG:2002.11296." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 112, + 290, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 112, + 290, + 156 + ], + "spans": [ + { + "bbox": [ + 69, + 112, + 290, + 156 + ], + "type": "text", + "content": "Y. Tay, M. Dehghani, S. Abnar, Y. Shen, D. Bahri, P. Pham, J. Rao, L. Yang, S. Ruder, and D. Metzler. 2020c. Long range arena: A benchmark for efficient transformers. arXiv, cs.LG:2011.04006." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 163, + 290, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 163, + 290, + 208 + ], + "spans": [ + { + "bbox": [ + 69, + 163, + 290, + 208 + ], + "type": "text", + "content": "E. F. Tjong Kim Sang and F. De Meulder. 2003. Introduction to the CoNLL-2003 shared task: Language-independent named entity recognition. In 7th Conference on Natural Language Learning, pages 142-147." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 215, + 290, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 215, + 290, + 248 + ], + "spans": [ + { + "bbox": [ + 69, + 215, + 290, + 248 + ], + "type": "text", + "content": "S. Wang, B. Z. Li, M. Khabsa, H. Fang, and H. Ma. 2020. Linformer: Self-attention with linear complexity. arXiv, cs.LG:2006.04768." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 255, + 290, + 333 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 255, + 290, + 333 + ], + "spans": [ + { + "bbox": [ + 69, + 255, + 290, + 333 + ], + "type": "text", + "content": "X. Wang, Y. Jiang, N. Bach, T. Wang, Z. Huang, F. Huang, and K. Tu. 2021. Improving named entity recognition by external context retrieving and cooperative learning. In 59th Annual Meeting of the Association for Computational Linguistics and 11th International Joint Conference on Natural Language Processing, volume 1, pages 1800-1812." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 339, + 290, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 339, + 290, + 428 + ], + "spans": [ + { + "bbox": [ + 69, + 339, + 290, + 428 + ], + "type": "text", + "content": "T. Wolf, L. Debut, V. Sanh, J. Chaumont, C. Delangue, A. Moi, P. Cistac, T. Rault, R. Louf, M. Funtowicz, J. Davison, S. Shleifer, P. von Platen, C. Ma, Y. Jernite, J. Plu, C. Xu, T. Le Scao, S. Gugger, M. Drame, Q. Lhoest, and A. M. Rush. 2020. Transformers: State-of-the-art natural language processing. In *Conference on Empirical Methods in Natural Language Processing: System Demonstrations*, pages 38-45." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 434, + 290, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 434, + 290, + 488 + ], + "spans": [ + { + "bbox": [ + 69, + 434, + 290, + 488 + ], + "type": "text", + "content": "J. Xu, J. Crego, and J. Senellart. 2020. Boosting neural machine translation with similar translations. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1580-1590." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 497, + 290, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 497, + 290, + 552 + ], + "spans": [ + { + "bbox": [ + 69, + 497, + 290, + 552 + ], + "type": "text", + "content": "I. Yamada, A. Asai, H. Shindo, H. Takeda, and Y. Matsumoto. 2020. LUKE: Deep contextualized entity representations with entity-aware self-attention. In Conference on Empirical Methods in Natural Language Processing, pages 6442-6454." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 559, + 290, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 559, + 290, + 624 + ], + "spans": [ + { + "bbox": [ + 69, + 559, + 290, + 624 + ], + "type": "text", + "content": "M. Zaheer, G. Guruganesh, K. A. Dubey, J. Ainslie, C. Alberti, S. Ontanon, P. Pham, A. Ravula, Q. Wang, L. Yang, and A. Ahmed. 2020. Big bird: Transformers for longer sequences. In Advances in Neural Information Processing Systems, volume 33, pages 17283-17297." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 632, + 290, + 687 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 632, + 290, + 687 + ], + "spans": [ + { + "bbox": [ + 69, + 632, + 290, + 687 + ], + "type": "text", + "content": "X. Zhang, Y. Jiang, X. Wang, X. Hu, Y. Sun, P. Xie, and M. Zhang. 2022. Domain-specific NER via retrieving correlated samples. In Proceedings of the 29th International Conference on Computational Linguistics, pages 2398-2404." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 69, + 694, + 168, + 706 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 694, + 168, + 706 + ], + "spans": [ + { + "bbox": [ + 69, + 694, + 168, + 706 + ], + "type": "text", + "content": "A Dataset Details" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 68, + 715, + 186, + 728 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 715, + 186, + 728 + ], + "spans": [ + { + "bbox": [ + 68, + 715, + 186, + 728 + ], + "type": "text", + "content": "A.1 Document Lengths" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 67, + 733, + 290, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 733, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 733, + 290, + 772 + ], + "type": "text", + "content": "Our NER dataset is composed of documents longer that typical NER datasets such as CoNLL-2003 (Tjong Kim Sang and De Meulder, 2003)." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 306, + 71, + 522, + 231 + ], + "blocks": [ + { + "bbox": [ + 306, + 71, + 522, + 231 + ], + "lines": [ + { + "bbox": [ + 306, + 71, + 522, + 231 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 522, + 231 + ], + "type": "image", + "image_path": "4cab2bba3cf8bc1e551f8128ff3e963d3d928bf7ab5ed7b35a9f52337a3c93a9.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 243, + 525, + 279 + ], + "lines": [ + { + "bbox": [ + 302, + 243, + 525, + 279 + ], + "spans": [ + { + "bbox": [ + 302, + 243, + 525, + 279 + ], + "type": "text", + "content": "Figure 5: Distribution of the number of sentences in our enhanced version of the dataset from Dekker et al. (2019)." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 302, + 300, + 525, + 326 + ], + "lines": [ + { + "bbox": [ + 302, + 300, + 525, + 326 + ], + "spans": [ + { + "bbox": [ + 302, + 300, + 525, + 326 + ], + "type": "text", + "content": "Figure 5 shows the distribution of the number of sentences of our NER dataset." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 336, + 464, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 336, + 464, + 348 + ], + "spans": [ + { + "bbox": [ + 302, + 336, + 464, + 348 + ], + "type": "text", + "content": "A.2 Automatic Correction Rules" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 302, + 353, + 524, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 353, + 524, + 407 + ], + "spans": [ + { + "bbox": [ + 302, + 353, + 524, + 407 + ], + "type": "text", + "content": "We use the following rules to automatically identify obvious errors in the original dataset from Dekker et al. (2019). The original dataset only contained PER entities, so these rules only apply to them:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 416, + 524, + 581 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 316, + 416, + 524, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 524, + 468 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 524, + 468 + ], + "type": "text", + "content": "- If a span appears in the list of characters from its novel but is not annotated as an entity, we investigate whether or not this is a false negative." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 478, + 524, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 524, + 532 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 524, + 532 + ], + "type": "text", + "content": "- Similarly, if a span annotated as an entity does not appear in the list of characters from its novel, we investigate whether or not it is a false positive." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 541, + 524, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 541, + 524, + 581 + ], + "spans": [ + { + "bbox": [ + 316, + 541, + 524, + 581 + ], + "type": "text", + "content": "- Finally, if a span is annotated as an entity but all of its tokens are not capitalized, we check if it is a false positive." + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 302, + 590, + 495, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 590, + 495, + 617 + ], + "spans": [ + { + "bbox": [ + 302, + 590, + 495, + 617 + ], + "type": "text", + "content": "B Heuristics Results Breakdown by Precision/Recall" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 302, + 625, + 525, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 625, + 525, + 691 + ], + "spans": [ + { + "bbox": [ + 302, + 625, + 525, + 691 + ], + "type": "text", + "content": "Figures 6 and 7 show precision and recall for all retrieval heuristics. Interestingly, retrieval only has a positive effect on recall, with precision being lower than the baseline except for the surrounding heuristic." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 302, + 702, + 406, + 714 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 702, + 406, + 714 + ], + "spans": [ + { + "bbox": [ + 302, + 702, + 406, + 714 + ], + "type": "text", + "content": "B.1 Oracle Versions" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 302, + 719, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 719, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 719, + 525, + 772 + ], + "type": "text", + "content": "Figures 6 and 7 show precision and recall for the oracle versions of all retrieval heuristics. While retrieval benefits recall more than precision, precision is still increased using retrieval. Together with" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "text", + "content": "719" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 71, + 284, + 227 + ], + "blocks": [ + { + "bbox": [ + 80, + 71, + 284, + 227 + ], + "lines": [ + { + "bbox": [ + 80, + 71, + 284, + 227 + ], + "spans": [ + { + "bbox": [ + 80, + 71, + 284, + 227 + ], + "type": "image", + "image_path": "82be396a06d23103061634f01b8dd2d3d558c6b6e6b009839b220b2481420ceb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 79, + 237, + 287, + 273 + ], + "lines": [ + { + "bbox": [ + 79, + 237, + 287, + 273 + ], + "spans": [ + { + "bbox": [ + 79, + 237, + 287, + 273 + ], + "type": "text", + "content": "Figure 6: Mean precision versus max number of retrieved sentences for all retrieval heuristics across 3 runs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 80, + 286, + 283, + 445 + ], + "blocks": [ + { + "bbox": [ + 80, + 286, + 283, + 445 + ], + "lines": [ + { + "bbox": [ + 80, + 286, + 283, + 445 + ], + "spans": [ + { + "bbox": [ + 80, + 286, + 283, + 445 + ], + "type": "image", + "image_path": "78048c3a9dc59a725d636f948057608a6b34cc70be64b4d8d43e0a2f8c3e85ed.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 79, + 454, + 289, + 491 + ], + "lines": [ + { + "bbox": [ + 79, + 454, + 289, + 491 + ], + "spans": [ + { + "bbox": [ + 79, + 454, + 289, + 491 + ], + "type": "text", + "content": "Figure 8: Mean precision versus max number of retrieved sentences across 3 runs for oracle versions of all retrieval heuristics." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 513, + 291, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 513, + 291, + 553 + ], + "spans": [ + { + "bbox": [ + 67, + 513, + 291, + 553 + ], + "type": "text", + "content": "the results from the regular heuristics, these results again highlight the potential performance gains of using a suitable re-ranker model to retrieve context." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 306, + 71, + 510, + 227 + ], + "blocks": [ + { + "bbox": [ + 306, + 71, + 510, + 227 + ], + "lines": [ + { + "bbox": [ + 306, + 71, + 510, + 227 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 510, + 227 + ], + "type": "image", + "image_path": "d69e670218c71bfde2d923fe05ea806d34a133004f869063f15dcd780356bbcb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 237, + 515, + 272 + ], + "lines": [ + { + "bbox": [ + 305, + 237, + 515, + 272 + ], + "spans": [ + { + "bbox": [ + 305, + 237, + 515, + 272 + ], + "type": "text", + "content": "Figure 7: Mean recall versus max number of retrieved sentences for all retrieval heuristics across 3 runs." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 286, + 510, + 445 + ], + "blocks": [ + { + "bbox": [ + 307, + 286, + 510, + 445 + ], + "lines": [ + { + "bbox": [ + 307, + 286, + 510, + 445 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 510, + 445 + ], + "type": "image", + "image_path": "7446c94113ae9db36eb9916d0e343f08ecd370ea34c02dd4e3087720be5041f3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 454, + 515, + 491 + ], + "lines": [ + { + "bbox": [ + 305, + 454, + 515, + 491 + ], + "spans": [ + { + "bbox": [ + 305, + 454, + 515, + 491 + ], + "type": "text", + "content": "Figure 9: Mean recall versus max number of retrieved sentences across 3 runs for oracle versions of all retrieval heuristics." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "text", + "content": "720" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 90, + 192, + 103 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 90, + 192, + 103 + ], + "spans": [ + { + "bbox": [ + 68, + 90, + 192, + 103 + ], + "type": "text", + "content": "A For every submission:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 106, + 317, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 106, + 317, + 121 + ], + "spans": [ + { + "bbox": [ + 77, + 106, + 317, + 121 + ], + "type": "text", + "content": "A1. Did you describe the limitations of your work?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 121, + 274, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 121, + 274, + 133 + ], + "spans": [ + { + "bbox": [ + 89, + 121, + 274, + 133 + ], + "type": "text", + "content": "Yes, limitations are discussed in Section 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 143, + 329, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 143, + 329, + 157 + ], + "spans": [ + { + "bbox": [ + 77, + 143, + 329, + 157 + ], + "type": "text", + "content": "A2. Did you discuss any potential risks of your work?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 158, + 306, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 158, + 306, + 169 + ], + "spans": [ + { + "bbox": [ + 89, + 158, + 306, + 169 + ], + "type": "text", + "content": "We do not think our work presents any direct risk" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 77, + 179, + 414, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 179, + 414, + 192 + ], + "spans": [ + { + "bbox": [ + 77, + 179, + 414, + 192 + ], + "type": "text", + "content": "A3. Do the abstract and introduction summarize the paper's main claims?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 89, + 194, + 236, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 194, + 236, + 206 + ], + "spans": [ + { + "bbox": [ + 89, + 194, + 236, + 206 + ], + "type": "text", + "content": "Yes, in the abstract and Section 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 77, + 216, + 398, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 216, + 398, + 229 + ], + "spans": [ + { + "bbox": [ + 77, + 216, + 398, + 229 + ], + "type": "text", + "content": "A4. Have you used AI writing assistants when working on this paper?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 89, + 230, + 138, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 230, + 138, + 242 + ], + "spans": [ + { + "bbox": [ + 89, + 230, + 138, + 242 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 68, + 252, + 290, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 252, + 290, + 266 + ], + "spans": [ + { + "bbox": [ + 68, + 252, + 290, + 266 + ], + "type": "text", + "content": "B Did you use or create scientific artifacts?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 270, + 524, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 270, + 524, + 296 + ], + "spans": [ + { + "bbox": [ + 67, + 270, + 524, + 296 + ], + "type": "text", + "content": "in Section 3.4, we indicate that we use a BERT checkpoint. We also use a previous NER dataset, see Section 3.3. We distribute an enhanced version of this dataset and code to reproduce our experiments." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 77, + 305, + 315, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 305, + 315, + 319 + ], + "spans": [ + { + "bbox": [ + 77, + 305, + 315, + 319 + ], + "type": "text", + "content": "B1. Did you cite the creators of artifacts you used?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 89, + 320, + 194, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 320, + 194, + 332 + ], + "spans": [ + { + "bbox": [ + 89, + 320, + 194, + 332 + ], + "type": "text", + "content": "See Section 3.3 and 3.4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 77, + 342, + 463, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 342, + 463, + 355 + ], + "spans": [ + { + "bbox": [ + 77, + 342, + 463, + 355 + ], + "type": "text", + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 89, + 356, + 421, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 356, + 421, + 370 + ], + "spans": [ + { + "bbox": [ + 89, + 356, + 421, + 370 + ], + "type": "text", + "content": "We specify the license in the Github repository given at the end of section 1." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 77, + 378, + 524, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 378, + 524, + 432 + ], + "spans": [ + { + "bbox": [ + 77, + 378, + 524, + 432 + ], + "type": "text", + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 89, + 433, + 309, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 433, + 309, + 445 + ], + "spans": [ + { + "bbox": [ + 89, + 433, + 309, + 445 + ], + "type": "text", + "content": "We use a dataset published for research purposes." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 77, + 454, + 524, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 454, + 524, + 495 + ], + "spans": [ + { + "bbox": [ + 77, + 454, + 524, + 495 + ], + "type": "text", + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 89, + 496, + 452, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 496, + 452, + 508 + ], + "spans": [ + { + "bbox": [ + 89, + 496, + 452, + 508 + ], + "type": "text", + "content": "Collected datas do not include information that can be used to identify individuals" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 77, + 517, + 524, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 517, + 524, + 544 + ], + "spans": [ + { + "bbox": [ + 77, + 517, + 524, + 544 + ], + "type": "text", + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 89, + 545, + 524, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 545, + 524, + 572 + ], + "spans": [ + { + "bbox": [ + 89, + 545, + 524, + 572 + ], + "type": "text", + "content": "We specify that the distributed dataset covers english literature (section 3.3). The reader can refer to Dekker et al., 2019 for more informations on the dataset." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 77, + 580, + 524, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 580, + 524, + 648 + ], + "spans": [ + { + "bbox": [ + 77, + 580, + 524, + 648 + ], + "type": "text", + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 89, + 649, + 524, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 649, + 524, + 676 + ], + "spans": [ + { + "bbox": [ + 89, + 649, + 524, + 676 + ], + "type": "text", + "content": "We include the number of document of our dataset in Section 3.3 We also include statistics about the length of these document in the Appendix" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "spans": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "text", + "content": "ACL 2023 Responsible NLP Checklist" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 679, + 522, + 699 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 679, + 522, + 699 + ], + "spans": [ + { + "bbox": [ + 67, + 679, + 522, + 699 + ], + "type": "text", + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 289, + 781, + 305, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 305, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 305, + 791 + ], + "type": "text", + "content": "721" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 70, + 294, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 70, + 294, + 84 + ], + "spans": [ + { + "bbox": [ + 68, + 70, + 294, + 84 + ], + "type": "text", + "content": "C Did you run computational experiments?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 78, + 89, + 218, + 100 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 89, + 218, + 100 + ], + "spans": [ + { + "bbox": [ + 78, + 89, + 218, + 100 + ], + "type": "text", + "content": "See Section 3.4 and Section 3.5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 110, + 523, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 110, + 523, + 137 + ], + "spans": [ + { + "bbox": [ + 77, + 110, + 523, + 137 + ], + "type": "text", + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 138, + 159, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 138, + 159, + 150 + ], + "spans": [ + { + "bbox": [ + 89, + 138, + 159, + 150 + ], + "type": "text", + "content": "See Section 3.4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 159, + 524, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 159, + 524, + 187 + ], + "spans": [ + { + "bbox": [ + 77, + 159, + 524, + 187 + ], + "type": "text", + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 188, + 317, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 188, + 317, + 200 + ], + "spans": [ + { + "bbox": [ + 89, + 188, + 317, + 200 + ], + "type": "text", + "content": "We include training hyperparameters in Section 3.4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 77, + 209, + 524, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 209, + 524, + 249 + ], + "spans": [ + { + "bbox": [ + 77, + 209, + 524, + 249 + ], + "type": "text", + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 89, + 251, + 524, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 251, + 524, + 277 + ], + "spans": [ + { + "bbox": [ + 89, + 251, + 524, + 277 + ], + "type": "text", + "content": "Our results are reported in Section 4. We indicate that, for Figure 1 and 2, each point is the mean " + }, + { + "bbox": [ + 89, + 251, + 524, + 277 + ], + "type": "inline_equation", + "content": "F1" + }, + { + "bbox": [ + 89, + 251, + 524, + 277 + ], + "type": "text", + "content": " of 3 runs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 77, + 285, + 524, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 285, + 524, + 326 + ], + "spans": [ + { + "bbox": [ + 77, + 285, + 524, + 326 + ], + "type": "text", + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 89, + 327, + 458, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 327, + 458, + 340 + ], + "spans": [ + { + "bbox": [ + 89, + 327, + 458, + 340 + ], + "type": "text", + "content": "See Section 3.1 (nltk), Section 3.4 (huggingface transformers), Section 3.5 (sequeval)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 68, + 349, + 521, + 364 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 349, + 521, + 364 + ], + "spans": [ + { + "bbox": [ + 68, + 349, + 521, + 364 + ], + "type": "text", + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 79, + 368, + 130, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 368, + 130, + 379 + ], + "spans": [ + { + "bbox": [ + 79, + 368, + 130, + 379 + ], + "type": "text", + "content": "Section 3.3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 77, + 390, + 524, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 390, + 524, + 416 + ], + "spans": [ + { + "bbox": [ + 77, + 390, + 524, + 416 + ], + "type": "text", + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 89, + 417, + 259, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 417, + 259, + 429 + ], + "spans": [ + { + "bbox": [ + 89, + 417, + 259, + 429 + ], + "type": "text", + "content": "The experiments were free of any risks" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 77, + 439, + 524, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 439, + 524, + 479 + ], + "spans": [ + { + "bbox": [ + 77, + 439, + 524, + 479 + ], + "type": "text", + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 89, + 481, + 291, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 481, + 291, + 492 + ], + "spans": [ + { + "bbox": [ + 89, + 481, + 291, + 492 + ], + "type": "text", + "content": "The authors annotated the dataset themselves" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 77, + 502, + 524, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 502, + 524, + 542 + ], + "spans": [ + { + "bbox": [ + 77, + 502, + 524, + 542 + ], + "type": "text", + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 89, + 544, + 291, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 544, + 291, + 555 + ], + "spans": [ + { + "bbox": [ + 89, + 544, + 291, + 555 + ], + "type": "text", + "content": "The authors annotated the dataset themselves" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 77, + 565, + 520, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 565, + 520, + 578 + ], + "spans": [ + { + "bbox": [ + 77, + 565, + 520, + 578 + ], + "type": "text", + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 89, + 580, + 291, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 580, + 291, + 592 + ], + "spans": [ + { + "bbox": [ + 89, + 580, + 291, + 592 + ], + "type": "text", + "content": "The authors annotated the dataset themselves" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 77, + 602, + 523, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 602, + 523, + 628 + ], + "spans": [ + { + "bbox": [ + 77, + 602, + 523, + 628 + ], + "type": "text", + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 89, + 629, + 359, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 629, + 359, + 641 + ], + "spans": [ + { + "bbox": [ + 89, + 629, + 359, + 641 + ], + "type": "text", + "content": "This is not relevant since annotation was done by the authors" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "text", + "content": "722" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/The Tail Wagging the Dog_ Dataset Construction Biases of Social Bias Benchmarks/fca46b6f-67a0-44f8-a5a2-d1a3682b3c41_content_list.json b/2023/The Tail Wagging the Dog_ Dataset Construction Biases of Social Bias Benchmarks/fca46b6f-67a0-44f8-a5a2-d1a3682b3c41_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b3b8a5d59b1d9b6c5bf15b687ac6c15856cfd162 --- /dev/null +++ b/2023/The Tail Wagging the Dog_ Dataset Construction Biases of Social Bias Benchmarks/fca46b6f-67a0-44f8-a5a2-d1a3682b3c41_content_list.json @@ -0,0 +1,1976 @@ +[ + { + "type": "text", + "text": "The Tail Wagging the Dog: Dataset Construction Biases of Social Bias Benchmarks", + "text_level": 1, + "bbox": [ + 211, + 83, + 786, + 120 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nikil Roashan Selvam $^{1}$ Sunipa Dev $^{2}$ \nDaniel Khashabi $^{3}$ Tushar Khot $^{4}$ Kai-Wei Chang $^{1}$ $^{1}$ University of California, Los Angeles $^{2}$ Google Research \n $^{3}$ Johns Hopkins University $^{4}$ Allen Institute for AI \nikilrselvam,kwchang}@ucla.edu, sunipadev@google. \ndanielk@jhu.edu, tushark@allenai.org", + "bbox": [ + 258, + 130, + 739, + 230 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 252, + 339, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "How reliably can we trust the scores obtained from social bias benchmarks as faithful indicators of problematic social biases in a given model? In this work, we study this question by contrasting social biases with non-social biases that stem from choices made during dataset construction (which might not even be discernible to the human eye). To do so, we empirically simulate various alternative constructions for a given benchmark based on seemingly innocuous modifications (such as paraphrasing or random-sampling) that maintain the essence of their social bias. On two well-known social bias benchmarks (WINOGENDER and BIASNLI), we observe that these shallow modifications have a surprising effect on the resulting degree of bias across various models and consequently the relative ordering of these models when ranked by measured bias. We hope these troubling observations motivate more robust measures of social biases.", + "bbox": [ + 141, + 278, + 460, + 575 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 114, + 588, + 258, + 602 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The omnipresence of large pre-trained language models (Liu et al., 2019; Raffel et al., 2020; Brown et al., 2020) has fueled concerns regarding their systematic biases carried over from underlying data into the applications they are used in, resulting in disparate treatment of people with different identities (Sheng et al., 2021; Abid et al., 2021).", + "bbox": [ + 112, + 613, + 487, + 725 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In response to such concerns, various benchmarks have been proposed to quantify the amount of social biases in models (Rudinger et al., 2018; Sheng et al., 2019; Li et al., 2020). These measures are composed of textual datasets built for a specific NLP task (such as question answering) and are accompanied by a metric such as accuracy of prediction which is used as an approximation of the amount of social biases.", + "bbox": [ + 112, + 726, + 487, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "These bias benchmarks are commonly used by machine learning practitioners to compare the degree of social biases (such as gender-occupation", + "bbox": [ + 112, + 871, + 487, + 919 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fd80f5c28cd735e5b85fc16d14449f902723c21e5d745137ebb90ce455b44cc3.jpg", + "image_caption": [ + "Figure 1: Two potential constructions of WINOGEN- DER with minor differences: a model (span-BERT, in this case) with the original dataset might seem to have gender-occupation bias (green tick) based on the change in its pronoun resolution. However, a minor change in its phrasing with no change in meaning (e.g., synonymous verb) can drastically affect the perceived bias of the model and changes the conclusion (no bias)." + ], + "image_footnote": [], + "bbox": [ + 512, + 250, + 877, + 458 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "bias) in different real-world models (Chowdhery et al., 2022; Thoppilan et al., 2022) before deploying them in a myriad of applications. However, they also inadvertently measure other non-social biases in their datasets. For example, consider the sentence from WINOGENDER in Figure 1. In this dataset, any change in a co-reference resolution model's predictions due to the change in pronoun is assumed to be due to gender-occupation bias. However, this assumption only holds for a model with near-perfect language understanding with no other biases. This may not often be the case, e.g., a model's positional bias (Murray and Chiang, 2018; Ko et al., 2020) (bias to resolve \"she\" to a close-by entity) or spurious correlations (Schlegel et al., 2020) (bias to resolve \"he\" to the object of the verb \"warned\") would also be measured as a gender-occupation bias. As a result, a slightly different template (e.g., changing the verb to \"cautioned\")", + "bbox": [ + 505, + 613, + 884, + 919 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1373", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics", + "bbox": [ + 226, + 945, + 771, + 957 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Volume 2: Short Papers, pages 1373-1386", + "bbox": [ + 368, + 958, + 628, + 971 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "July 9-14, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 295, + 972, + 700, + 985 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "could result in completely different bias measurements.", + "bbox": [ + 112, + 84, + 490, + 115 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The goal of this work is to illustrate the extent to which social bias measurements are effected by assumptions that are built into dataset constructions. To that end, we consider several alternate dataset constructions for 2 bias benchmarks WINOGENDER and BIASNLI. We show that, just by the choice of certain target-bias-irrelevant elements in a dataset, it is possible to discover different degrees of bias for the same model as well as different model rankings1. For instance, one experiment on BIASNLI demonstrated that merely negating verbs drastically reduced the measured bias $(41.64\\rightarrow 13.40)$ on an ELMo-based Decomposable Attention model and even caused a switch in the comparative ranking with RoBERTa. Our findings demonstrate the unreliability of current benchmarks to truly measure social bias in models and suggest caution when considering these measures as the gold truth. We provide a detailed discussion (\\$5) of the implications of our findings, relation to experienced harms, suggestions for improving bias benchmarks, and directions for future work.", + "bbox": [ + 115, + 117, + 490, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 112, + 481, + 270, + 495 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A large body of work investigates ways to evaluate biases carried inherently in language models (Bolukbasi et al., 2016; Caliskan et al., 2017; Nadeem et al., 2021) and expressed in specific tasks (Nangia et al., 2020; Kirk et al., 2021; Schramowski et al., 2022; Prabhumoye et al., 2021; Srinivasan and Bisk, 2021; Kirk et al., 2021; Parrish et al., 2021; Baldini et al., 2022; Czarnowska et al., 2021; Dev et al., 2021a; Zhao et al., 2021). Alongside, there is also growing concern about the measures not relating to experienced harms (Blodgett et al., 2020), not inclusive in framing (Dev et al., 2021b), ambiguous about what bias is measured (Blodgett et al., 2021), not correlated in their findings of bias across intrinsic versus extrinsic techniques (Goldfarb-Tarrant et al., 2021; Cao et al., 2022), and susceptible to adversarial perturbations (Zhang et al., 2021) and seed word selection (Antoniak and Mimno, 2021).", + "bbox": [ + 112, + 507, + 489, + 810 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The concurrent work by (Seshadri et al., 2022) discusses the unreliability of quantifying social biases using templates by varying templates in a se", + "bbox": [ + 112, + 813, + 489, + 860 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "mantic preserving manner. While their findings are consistent with ours, the two works provide complementary experimental observations. Seshadri et al. (2022) study a wider range of tasks, though we focus our experiments on a wider set of models and alternate dataset constructions (with a greater range of syntactic and semantic variability). As a result, we are able to illustrate the effect of the observed variability on ranking large language models according to measured bias for deployment in real world applications.", + "bbox": [ + 507, + 84, + 884, + 261 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3 Social Bias Measurements and Alternate Constructions", + "text_level": 1, + "bbox": [ + 507, + 285, + 806, + 317 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Bias measures in NLP are often quantified through comparative prediction disparities on language datasets that follow existing tasks such as classification (De-Arteaga et al., 2019) or coreference resolution (Rudinger et al., 2018). As a result, these datasets are central to what eventually gets measured as “bias”. Not only do they determine the “amount” of bias measured but also the “type” of bias or stereotype measured. Datasets often vary combinations of gendered pronouns and occupations to evaluate stereotypical associations. It is important to note that these constructs of datasets and their templates, which determine what gets measured, are often arbitrary choices. The sentences could be differently structured, be generated from a different set of seed words, and more. However, we expect that for any faithful bias benchmark, such dataset alterations that are not relevant to social bias should not have a significant impact on the artifact (e.g. gender bias) being measured.", + "bbox": [ + 505, + 335, + 884, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Thus, to evaluate the faithfulness of current benchmarks, we develop alternate dataset constructions through modifications that should not have any effect on the social bias being measured in a dataset. They are minor changes that should not influence models with true language understanding – the implicit assumption made by current bias benchmarks. Any notable observed changes in a model's bias measure due to these modifications would highlight the incorrectness of this assumption. Consequently, this would bring to light the unreliability of current benchmarks to faithfully measure the target bias and disentangle the measurement from measurement of other non-social biases. A non-exhaustive set of such alternate constructions considered in this work are listed below.", + "bbox": [ + 507, + 661, + 885, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1All preprocessed datasets (original and alternate constructions) and code are available at https://github.com/uclanlp/socialbias-dataset-construction-biases.", + "bbox": [ + 112, + 868, + 489, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1374", + "bbox": [ + 482, + 928, + 521, + 940 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Clause after occupation", + "text_level": 1, + "bbox": [ + 179, + 65, + 317, + 76 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The engineer, who just returned from the beach, informed the client that he would need to make all future payments on time.", + "bbox": [ + 129, + 82, + 357, + 117 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Clause after participant", + "text_level": 1, + "bbox": [ + 428, + 65, + 568, + 77 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The engineer informed the client, who just returned from the beach, that he would need to make all future payments on time.", + "bbox": [ + 381, + 82, + 615, + 117 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Synonymization", + "text_level": 1, + "bbox": [ + 705, + 65, + 801, + 76 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The engineer informed the client that he would need to make all upcoming payments on time.", + "bbox": [ + 638, + 82, + 857, + 117 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adjective before occupation", + "text_level": 1, + "bbox": [ + 156, + 131, + 317, + 143 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The cruel engineer informed the client that he would need to make all future payments on time.", + "bbox": [ + 157, + 155, + 315, + 200 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adjective after occupation", + "text_level": 1, + "bbox": [ + 337, + 131, + 490, + 143 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The engineer, who was cruel, informed the client that he would need to make all future payments on time.", + "bbox": [ + 334, + 155, + 489, + 200 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adjective before participant", + "text_level": 1, + "bbox": [ + 507, + 131, + 670, + 143 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The engineer informed the wise client that he would need to make all future payments on time.", + "bbox": [ + 510, + 155, + 663, + 200 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adjective after participant", + "text_level": 1, + "bbox": [ + 685, + 131, + 838, + 143 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The engineer informed the client, who was wise, that he would need to make all future payments on time.", + "bbox": [ + 685, + 155, + 831, + 200 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 2: An instance (\"The engineer informed the client that he would need to make all future payments on time\") from WINOGENDER benchmark modified under various shallow modifications (§3). To a human eye, such modifications do not necessarily affect the outcome of the given pronoun resolution problem.", + "bbox": [ + 112, + 225, + 880, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Negations: A basic function in language understanding is to understand the negations of word groups such as action verbs, or adjectives. Altering verbs in particular, such as 'the doctor bought' to 'the doctor did not buy' should typically not affect the inferences made about occupation associations.", + "bbox": [ + 112, + 293, + 487, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Synonym substitutions: Another fundamental function of language understanding is the ability to parse the usage of similar words or synonyms used in identical contexts, to derive the same overall meaning of a sentence. For bias measuring datasets, synonymizing non-pivotal words (such as non-identity words like verbs) should not change the outcome of how much bias is measured.", + "bbox": [ + 112, + 393, + 487, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Varying length of the text: In typical evaluation datasets, the number of clauses that each sentence is composed of and overall the sentence length are arbitrary experimental choices. Fixing this length is common, especially when such datasets need to be created at scale. If language is understood, adding a neutral phrase without impacting the task-specific semantics should not alter the bias measured.", + "bbox": [ + 112, + 526, + 487, + 653 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adding descriptors: Sentences used in real life are structured in complex ways and can have descriptors, such as adjectives about an action, person, or object, without changing the net message expressed by the text. For example, the sentences, \"The doctor bought an apple.\", and \"The doctor bought a red apple.\" do not change any assumptions made about the doctor, or the action of buying an apple.", + "bbox": [ + 112, + 658, + 487, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Random samples: Since the sentence constructs of these datasets are not unique, a very simple alternate construction of a dataset is a different subsample of itself. This is because the dataset is scraped or generated with specific assumptions or parameters, such as seed word lists, templates of sentences, and word order. However, neither the sentence constructs or templates, nor the seed word", + "bbox": [ + 112, + 790, + 487, + 917 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "lists typically used are exhaustive or representative of entire categories of words (such as gendered words, emotions, and occupations).", + "bbox": [ + 507, + 293, + 880, + 341 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "See Fig. 2 for example constructions on WINO-GENDER (App. A, B for detailed descriptions).", + "bbox": [ + 507, + 342, + 882, + 374 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4 Case Studies", + "text_level": 1, + "bbox": [ + 507, + 388, + 653, + 403 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We discuss here the impact of alternate constructions on two task-based measures of bias.2", + "bbox": [ + 507, + 414, + 882, + 445 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1 Coreference Resolution", + "text_level": 1, + "bbox": [ + 507, + 461, + 739, + 475 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Several different bias measures (Rudinger et al., 2018; Zhao et al., 2018; Cao and Daumé III, 2021) for coreference resolution work similar to Winograd Schema (Winograd, 1972) where a sentence has two entities and the task is to resolve which entity a specific pronoun or noun refers to. We work here with WINOGENDER (Rudinger et al., 2018), popularly used to measure biases. It is worth noting that WINOGENDER was originally intended by its authors to merely be a diagnostic tool that checks for bias in a model; the authors note that it may demonstrate the presence of model bias but not prove the absence of the same. Nonetheless, models developed today are indeed tested and compared for social bias on WinoGender, leading to its usage as a comparative standard or benchmark (Chowdhery et al., 2022; Thoppilan et al., 2022).", + "bbox": [ + 507, + 482, + 882, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The metric used to evaluate bias is the percentage of sentence pairs where there is a mismatch in predictions for the male and female gendered pronouns. For instance, in Fig. 2, if the pronoun \"he\" is linked to \"engineer\" but switches to \"client\" for the pronoun \"she\", that would indicate a gender-occupation bias. Higher the number of mismatches,", + "bbox": [ + 507, + 757, + 884, + 869 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "2We note that throughout this paper, we focus on gender- occupation bias as an illustrative example; however, our discussion can be extended to other aspects of biases too.", + "bbox": [ + 507, + 879, + 882, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1375", + "bbox": [ + 482, + 927, + 519, + 940 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1adedff8ac62a30012d7f23cb921c7b276c9ba994af299070642fc858dc7b720.jpg", + "image_caption": [ + "(a) WINOGENDER", + "Figure 3: Bias measures on (a) WINOGENDER (percentage M-F mismatch, log-scale) and (b) BIASNLI (accuracy as percentage neutral, log-scale), across a variety of dataset constructions and models." + ], + "image_footnote": [], + "bbox": [ + 115, + 108, + 623, + 275 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ff012ff95eee03763f79b1e41dd49741182d9866115ab0a9dc002da8a5978b26.jpg", + "image_caption": [ + "(b) BIASNLI" + ], + "image_footnote": [], + "bbox": [ + 630, + 107, + 884, + 269 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "higher the bias. In particular, note that the metric does not take into account the accuracy of the predictions, but rather only the mismatch between the two pronouns.", + "bbox": [ + 112, + 334, + 487, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We experiment with three alternate constructions of the dataset: addition of clauses, addition of adjectives, and synonymizing words in templates. Each alternate construction is introduced so as to not affect the overall meaning of the sentence.", + "bbox": [ + 112, + 400, + 489, + 480 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Experimental Results: We use an end-to-end coreference model with SpanBERT embeddings (Lee et al., 2018; Joshi et al., 2020), UnifiedQA (small, base, and large) (Khashabi et al., 2020) QA model, $^{3}$ and a long-document coreference model with Longformer encodings (Toshniwal et al., 2021). Results of evaluating these models on various WINOGENDER constructions is summarized in Fig. 3a. Small changes to the formulation of dataset templates result in sizable changes to computed bias measures compared to the published baseline constructions. For example, a construction involving added adjectives after occupations would have found the UnifiedQA (large) model to have $10\\%$ less bias compared to the default constructions. The sensitivity to the dataset constructions can have a drastic effect on ranking models according to their social bias, as Fig. 3a shows. For example, the SpanBERT model is considered to have less bias than UnifiedQA (small) model in the baseline dataset, but would be considered to be more biased if the templates had clauses after the participants or adjectives before the occupation.", + "bbox": [ + 112, + 482, + 489, + 853 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/78f45b52836cdfaa2413129c98c9554d6798d012c243a13db12719bec25feb66.jpg", + "image_caption": [ + "Figure 4: Bias measures (fraction neutral) computed on BIASNLI. The violin plot represents distribution of bias measure scores across datasets reconstructed using different $10\\%$ subsets of the occupation word list across 100 random samples." + ], + "image_footnote": [], + "bbox": [ + 504, + 338, + 884, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2 Natural Language Inference", + "text_level": 1, + "bbox": [ + 507, + 650, + 778, + 665 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Natural Language Inference (NLI) is the task of determining directional relationships between two sentences (a premise $(P)$ and a hypothesis $(H)$ ). Dev et al. (2020)'s measure based on NLI (BIASNLI) evaluates if stereotypical inferences are made by language models. We use their dataset for gender-occupation stereotypes containing approximately 2 million sentence pairs such as $P$ : \"The doctor bought a bagel.\", $H$ : \"The man bought a bagel.\" The expected prediction for each sentence pair in the dataset is neutral, and therefore the bias metric used is the fraction of neutral inferences on dataset - the higher the score, the lower the bias.", + "bbox": [ + 505, + 675, + 884, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We experiment with three alternate constructions of the dataset: verb negation, random sampling,", + "bbox": [ + 507, + 887, + 882, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "3Used by converting co-reference into question-answering, e.g., \"The technician told the customer that he had completed the repair. Who does the word 'he' refer to? $\\backslash \\mathfrak{n}$ (a) technician (b) customer\"", + "bbox": [ + 112, + 868, + 487, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1376", + "bbox": [ + 482, + 927, + 521, + 940 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and addition of clauses. Note that the alternate constructions do not impact the unbiased label (neutral). Any change in construction (say negating a verb) is applied to both the premise and hypothesis. Refer to App. B for a detailed description.", + "bbox": [ + 112, + 84, + 490, + 165 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Experimental Results: We use RoBERTa trained on SNLI (RoBERTa-base-SNLI) (Liu et al., 2019), ELMo-based Decomposable Attention (ELMoDA) (Parikh et al., 2016), ALBERT (Lan et al., 2019), distilled version of the RoBERTa-base model (Sanh et al., 2019), and RoBERTa-large fin-tuned on WANLI (Liu et al., 2022). The bias measured with each model using BIASNLI is recorded in Fig. 3b. The results show how small modifications to the dataset again result in large changes to the bias measured, and also change the bias rankings. For example, adding a negation largely reduces the bias measured $(\\triangle = 28.24)$ for ELMoDA, and also results in a switch in the comparative ranking to RoBERTa-base-SNLI. Furthermore, as seen in Fig. 4, there is a significant overlap in the bias measures of ALBERT, DistilRoBERTa, and ELMo-DA under random sampling, which corresponds to high variability in relative model ordering across different sub-samples of the dataset.", + "bbox": [ + 115, + 166, + 489, + 487 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5 Discussion and Conclusion", + "text_level": 1, + "bbox": [ + 112, + 500, + 379, + 514 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Social bias measurements are very sensitive to evaluation methodology. Our empirical evidence sheds light on how the model's non-social biases brought out or masked by alternate constructions can cause bias benchmarks to underestimate or overestimate the social bias in a model. More interestingly, it is important to note that different models respond differently to perturbations. In fact, the same perturbation can result in a higher or lower measured bias depending on the model (as seen in §4.1 and §4.2), which points to how models might parse information (and thus bias) differently.", + "bbox": [ + 112, + 526, + 489, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While current bias measures do play a role in exposing where model errors have a stereotypical connotation, a lack of sentence construction variability or even assumptions made when creating seed word lists can reduce the reliability of the benchmarks, as we see in this work (§4.2). Even with simple sentences, it is not apparent how to disentangle the biased association of the identity with the verb or the occupation amongst others. This is especially important to note as it highlights that measures can lack concrete definitions of what bi", + "bbox": [ + 112, + 721, + 490, + 896 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ased associations they measure. Consequently, the relation between measured bias and experienced harm becomes unclear.", + "bbox": [ + 507, + 84, + 884, + 131 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We hope that our troubling observations motivate future work that thoroughly investigates how to construct robust benchmarks that faithfully measure the target bias without being affected by model errors and other non-social biases. As suggested by our subsampling experiments (Appendix F), it might be fruitful to encourage both syntactic and semantic diversity in these benchmarks. Bias benchmarks that provide uncertainty measures (instead of a single number) might enable practitioners to better compare models before deploying them. Furthermore, since the opaqueness of large language models makes it challenging to understand how and to what extent a linguistic change will affect the measured bias, explainable models might indeed facilitate better measurement of their social bias. Assuming that we can generate faithful explanations for a model's predictions, an exciting future direction is to explore construction of bias benchmarks which operate on the explanations of the predictions rather than the predictions themselves. Lastly, we also encourage discussions on the complexity of the sentences used in benchmarks and their implications on what gets measured in relation to un-templated, naturally-occurring text (Levy et al., 2021), as an attempt to ground our measurements in experienced harms.", + "bbox": [ + 507, + 135, + 885, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Limitations", + "text_level": 1, + "bbox": [ + 509, + 583, + 613, + 600 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We acknowledge the underlying assumptions of the social bias benchmarks used in our study. While the presented study aims to point out a key limitation of currently accepted methodologies, the presented investigation could benefit from more diversification. First, this study focuses on English. While we expect similar issues with similarly-constructed benchmarks in other languages, we leave it to future work to formally address the same. Also, the bias benchmarks themselves imbibe the notion of fairness with the Western value system (Bhatt et al., 2022), and future explorations of benchmarks should diversify culturally as well. Last but not least, we acknowledge the harm of binary treatment of genders in one of the target benchmarks. The purpose of this work was to bring light to a broader problem regarding the reliability of social benchmark metrics, with the hypothesis that the main idea of this paper would hold for a wider", + "bbox": [ + 507, + 613, + 884, + 919 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "Also observed at $25\\%$ and $50\\%$ samples in Fig. 5(App.)", + "bbox": [ + 134, + 903, + 482, + 917 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1377", + "bbox": [ + 482, + 927, + 519, + 940 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "range of datasets with other assumptions or notions of fairness. We also acknowledge that there are larger models that we were not able to train and evaluate due to the limitations on our computational budget. The current study was focused on benchmarks with templated instances. This is no coincidence: the dominant majority of the social bias benchmarking literature relies on sentences with some degree of known structure, even in those collected from the wild (Levy et al., 2021). Such structural assumptions in datasets are necessary for defining and extracting quantifiable measures of social bias, which as we argue, are the reason behind the brittleness of their decisions. Future work should focus on making our bias benchmarks more diverse and robust to small decisions that go into making them.", + "bbox": [ + 112, + 84, + 492, + 361 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Broader Impact", + "text_level": 1, + "bbox": [ + 114, + 369, + 260, + 386 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Bias evaluating benchmarks play a very significant role in helping identify potential risks of language technologies. While a large body of work evolves in this area of work, there is growing concern about the ability of the different benchmarks to accurately quantify and identify social biases. We emphasize these concerns by evaluating how robust the benchmarks are to alternate constructions based on simple linguistic properties. It is important to note how inaccurate measurements of social biases can be problematic by underestimating or misdiagnosing the potential harm from language models. We hope our work helps identify such pitfalls.", + "bbox": [ + 112, + 395, + 489, + 605 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 114, + 615, + 287, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We thank the students and colleagues at UCLA, JHU and AI2 for their insightful feedback towards improving this paper. The authors would also like to thank the anonymous reviewers for their constructive feedback. This project is supported by generous gifts from Allen Institute for AI, CISCO, Amazon, and a Sloan fellowship.", + "bbox": [ + 112, + 640, + 489, + 755 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 114, + 780, + 213, + 795 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Abubakar Abid, Maheen Farooqi, and James Zou. 2021. Persistent anti-muslim bias in large language models. In AAAI/ACM Conference on AI, Ethics, and Society (AIES), pages 298-306.", + "Maria Antoniak and David Mimno. 2021. Bad seeds: Evaluating lexical methods for bias measurement. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the" + ], + "bbox": [ + 115, + 802, + 489, + 917 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1889-1904, Online. Association for Computational Linguistics.", + "Ioana Baldini, Dennis Wei, Karthikeyan Natesan Ramamurthy, Moninder Singh, and Mikhail Yurochkin. 2022. Your fairness may vary: Pretrained language model fairness in toxic text classification. In Annual Meeting of the Association for Computational Linguistics (ACL) - Findings.", + "Shaily Bhatt, Sunipa Dev, Partha Talukdar, Shachi Dave, and Vinodkumar Prabhakaran. 2022. Recontextualizing fairness in NLP: The case of India. In Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 727-740, Online only. Association for Computational Linguistics.", + "Su Lin Blodgett, Solon Barocas, Hal Daumé III, and Hanna Wallach. 2020. Language (technology) is power: A critical survey of \"bias\" in nlp. In Annual Meeting of the Association for Computational Linguistics (ACL).", + "Su Lin Blodgett, Gilsinia Lopez, Alexandra Olteanu, Robert Sim, and Hanna Wallach. 2021. Stereotyping norwegian salmon: an inventory of pitfalls in fairness benchmark datasets. In Annual Meeting of the Association for Computational Linguistics (ACL).", + "Tolga Bolukbasi, Kai-Wei Chang, James Y Zou, Venkatesh Saligram, and Adam T Kalai. 2016. Man is to computer programmer as woman is to homemaker? debiasing word embeddings. In Advances in Neural Information Processing Systems, volume 29. Curran Associates, Inc.", + "Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, and et al. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems (NeurIPS).", + "Aylin Caliskan, Joanna J. Bryson, and Arvind Narayanan. 2017. Semantics derived automatically from language corpora contain human-like biases. Science, 356(6334):183-186.", + "Yang Trista Cao and Hal Daumé III. 2021. Toward gender-inclusive coreference resolution: An analysis of gender and bias throughout the machine learning lifecycle. Computational Linguistics (CL).", + "Yang Trista Cao, Yada Pruksachatkun, Kai-Wei Chang, Rahul Gupta, Varun Kumar, Jwala Dhamala, and Aram Galstyan. 2022. On the intrinsic and extrinsic fairness evaluation metrics for contextualized language representations. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 561-570," + ], + "bbox": [ + 510, + 85, + 884, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1378", + "bbox": [ + 482, + 928, + 521, + 940 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Dublin, Ireland. Association for Computational Linguistics.", + "Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam M. Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Benton C. Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan Firat, Michele Catasta, Jason Wei, Kathleen S. Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. 2022. Palm: Scaling language modeling with pathways. ArXiv, abs/2204.02311.", + "Paula Czarnowska, Yogarshi Vyas, and Kashif Shah. 2021. Quantifying social biases in nlp: A generalization and empirical comparison of extrinsic fairness metrics. Transactions of the Association for Computational Linguistics (TACL).", + "Maria De-Arteaga, Alexey Romanov, Hanna Wallach, Jennifer Chayes, Christian Borgs, Alexandra Chouldechova, Sahin Geyik, Krishnamaram Kenthapadi, and Adam Kalai. 2019. Bias in bios: A case study of semantic representation bias in a high-stakes setting. In ACM Conference on Fairness, Accountability and Transparency (FAccT).", + "Sunipa Dev, Tao Li, Jeff M. Phillips, and Vivek Srikumar. 2020. On measuring and mitigating biased inferences of word embeddings. Conference on Artificial Intelligence (AAAI).", + "Sunipa Dev, Tao Li, Jeff M Phillips, and Vivek Srikumar. 2021a. Oscar: Orthogonal subspace correction and rectification of biases in word embeddings. In _Conference on Empirical Methods in Natural Language Processing (EMNLP)\\.", + "Sunipa Dev, Masoud Monajatipoor, Anaelia Ovalle, Arjun Subramonian, Jeff Phillips, and Kai-Wei Chang. 2021b. Harms of gender exclusivity and challenges in non-binary representation in language technologies. In Conference on Empirical Methods in Natural Language Processing (EMNLP).", + "Seraphina Goldfarb-Tarrant, Rebecca Marchant, Ricardo Muñoz Sánchez, Mugdha Pandya, and Adam Lopez. 2021. Intrinsic bias metrics do not correlate with application bias. In Annual Meeting of the Association for Computational Linguistics (ACL)." + ], + "bbox": [ + 115, + 85, + 489, + 917 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mandar Joshi, Danqi Chen, Yinhan Liu, Daniel S. Weld, Luke Zettlemoyer, and Omer Levy. 2020. SpanBERT: Improving pre-training by representing and predicting spans. Transactions of the Association for Computational Linguistics (TACL).", + "Daniel Khashabi, Sewon Min, Tushar Khot, Ashish Sabharwal, Oyvind Tafjord, Peter Clark, and Hannaneh Hajishirzi. 2020. UnifiedQA: Crossing Format Boundaries With a Single QA System. In Conference on Empirical Methods in Natural Language Processing (EMNLP) - Findings.", + "Hannah Rose Kirk, Filippo Volpin, Haider Iqbal, Elias Benussi, Frederic Dreyer, Aleksandar Shtedritski, Yuki Asano, et al. 2021. Bias out-of-the-box: An empirical analysis of intersectional occupational biases in popular generative language models. Advances in Neural Information Processing Systems (NeurIPS).", + "Miyoung Ko, Jinhyuk Lee, Hyunjae Kim, Gangwoo Kim, and Jaewoo Kang. 2020. Look at the first sentence: Position bias in question answering. In _Conference on Empirical Methods in Natural Language Processing_ (EMNLP).", + "Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. 2019. Albert: A lite bert for self-supervised learning of language representations. In International Conference on Learning Representations (ICLR).", + "Kenton Lee, Luheng He, and Luke Zettlemoyer. 2018. Higher-order coreference resolution with coarse-to-fine inference. In Conference of the North American Chapter of the Association for Computational Linguistics (NAACL).", + "Shahar Levy, Koren Lazar, and Gabriel Stanovsky. 2021. Collecting a large-scale gender bias dataset for coreference resolution and machine translation. In *Conference on Empirical Methods in Natural Language Processing* (EMNLP) - Findings.", + "Tao Li, Daniel Khashabi, Tushar Khot, Ashish Sabharwal, and Vivek Srikumar. 2020. UnCovering Stereotypical Biases via Underspecified Questions. In Conference on Empirical Methods in Natural Language Processing (EMNLP) - Findings.", + "Alisa Liu, Swabha Swayamdipta, Noah A Smith, and Yejin Choi. 2022. WANLI: Worker and AI Collaboration for Natural Language Inference Dataset Creation. arXiv preprint arXiv:2201.05955.", + "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692.", + "Kenton Murray and David Chiang. 2018. Correcting length bias in neural machine translation. In Conference on Machine Translation (WMT)." + ], + "bbox": [ + 510, + 85, + 884, + 917 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1379", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Moin Nadeem, Anna Bethke, and Siva Reddy. 2021. StereoSet: Measuring stereotypical bias in pretrained language models. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 5356-5371, Online. Association for Computational Linguistics.", + "Nikita Nangia, Clara Vania, Rasika Bhalerao, and Samuel R Bowman. 2020. Crows-pairs: A challenge dataset for measuring social biases in masked language models. In Conference on Empirical Methods in Natural Language Processing (EMNLP).", + "Ankur P. Parikh, Oscar Tackström, Dipanjan Das, and Jakob Uszkoreit. 2016. A decomposable attention model for natural language inference. In *Conference on Empirical Methods in Natural Language Processing* (EMNLP).", + "Alicia Parrish, Angelica Chen, Nikita Nangia, Vishakh Padmakumar, Jason Phang, Jana Thompson, Phu Mon Htut, and Samuel R Bowman. 2021. BBQ: A hand-built bias benchmark for question answering. In Annual Meeting of the Association for Computational Linguistics (ACL).", + "Shrimai Prabhumoye, Rafal Kocielnik, Mohammad Shoeybi, Anima Anandkumar, and Bryan Catanzaro. 2021. Few-shot instruction prompts for pretrained language models to detect social biases. arXiv preprint arXiv:2112.07868.", + "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research (JMLR).", + "Rachel Rudinger, Jason Naradowsky, Brian Leonard, and Benjamin Van Durme. 2018. Gender bias in coreference resolution. In Conference of the North American Chapter of the Association for Computational Linguistics (NAACL).", + "Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. 2019. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. ArXiv, abs/1910.01108.", + "Viktor Schlegel, Goran Nenadic, and Riza Batista-Navarro. 2020. Beyond leaderboards: A survey of methods for revealing weaknesses in natural language inference data and models. arXiv preprint arXiv:2005.14709.", + "Patrick Schramowski, Cigdem Turan, Nico Andersen, Constantin A Rothkopf, and Kristian Kersting. 2022. Large pre-trained language models contain human-like biases of what is right and wrong to do. Nature Machine Intelligence." + ], + "bbox": [ + 115, + 85, + 489, + 917 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Preethi Seshadri, Pouya Pezeshkpour, and Sameer Singh. 2022. Quantifying social biases using templates is unreliable. arXiv preprint arXiv:2210.04337.", + "Emily Sheng, Kai-Wei Chang, Prem Natarajan, and Nanyun Peng. 2019. The Woman Worked as a Babysitter: On Biases in Language Generation. In Conference on Empirical Methods in Natural Language Processing (EMNLP).", + "Emily Sheng, Kai-Wei Chang, Premkumar Natarajan, and Nanyun Peng. 2021. Societal biases in language generation: Progress and challenges. In *Conference on Empirical Methods in Natural Language Processing* (EMNLP).", + "Tejas Srinivasan and Yonatan Bisk. 2021. Worst of both worlds: Biases compound in pre-trained vision-and-language models. In Workshop on Gender Bias in Natural Language Processing.", + "Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, et al. 2022. LaMDA: Language Models for Dialog Applications. arXiv preprint arXiv:2201.08239.", + "Shubham Toshniwal, Patrick Xia, Sam Wiseman, Karen Livescu, and Kevin Gimpel. 2021. On generalization in coreference resolution. In Proceedings of the Workshop on Computational Models of Reference, Anaphora and Coreference.", + "T. Winograd. 1972. Understanding natural language. Cognitive psychology, 3(1):1-191.", + "Chong Zhang, Jieyu Zhao, Huan Zhang, Kai-Wei Chang, and Cho-Jui Hsieh. 2021. Double perturbation: On the robustness of robustness and counterfactual bias evaluation. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3899-3916, Online. Association for Computational Linguistics.", + "Jieyu Zhao, Daniel Khashabi, Tushar Khot, Ashish Sabharwal, and Kai-Wei Chang. 2021. Ethical-advice taker: Do language models understand natural language interventions? In Annual Meeting of the Association for Computational Linguistics (ACL) - Findings.", + "Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. 2018. Gender bias in coreference resolution: Evaluation and debiasing methods. In Conference of the North American Chapter of the Association for Computational Linguistics (NAACL)." + ], + "bbox": [ + 510, + 85, + 880, + 832 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1380", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Appendix The Tail Wagging the Dog: Dataset Construction Biases of Social Bias Benchmarks", + "text_level": 1, + "bbox": [ + 142, + 93, + 853, + 149 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "A Alternate Constructions of WINOGENDER", + "text_level": 1, + "bbox": [ + 112, + 211, + 383, + 243 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Addition of clauses: For WINOGENDER, we add clauses like \"who just returned from the beach\" to the different entities in the sentence. For instance, the sentence \"The customer left the bartender a big tip because he was feeling generous.\" becomes \"The customer, who just returned from the beach, left the bartender a big tip because he was feeling generous.\"", + "bbox": [ + 110, + 254, + 487, + 382 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Synonym substitution: We substitute with synonyms such that it does not change the meaning of the sentence. WINOGENDER has 720 sentences generated from 120 templates, making manual substitution of synonyms in the templates feasible. For example, the sentence \"The supervisor gave the employee feedback on his stellar performance.\" is replaced by \"The supervisor gave the employee feedback on his amazing performance.\"", + "bbox": [ + 110, + 393, + 489, + 539 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Adding adjectives: As discussed in §3, we add descriptors in the form of adjectives that do not add information about which entity the pronoun or noun would refer to. We do it in four distinct ways, (i) adding the descriptor to the occupation mentioned, e.g. doctor (e.g., \"doctor\" to \"good doctor\"), (ii) adding it to the occupation as a separate clause (e.g., \"doctor\" to \"the doctor who was good\"), (iii) adding the descriptor to the participant mentioned, e.g., \"client\" (similar to (i)), and (iv) adding it to the participant as a separate clause (similar to (ii)).", + "bbox": [ + 112, + 548, + 489, + 726 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "B Alternate Constructions of BIASNLI", + "text_level": 1, + "bbox": [ + 112, + 738, + 470, + 753 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Negation: We negate the verb in each sentence of the dataset. For example, $P$ : \"The doctor bought a bagel.\", $H$ : \"The man bought a bagel.\", and $P$ : \"The doctor did not buy a bagel.\", $H$ : \"The man did not buy a bagel.\", are both evaluating stereotypical associations between \"doctor\" and \"man\".", + "bbox": [ + 110, + 764, + 487, + 859 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Random sampling: The BIASNLI dataset is generated from templates by populating empty slots (e.g. verbs, objects) with words from cor", + "bbox": [ + 112, + 870, + 489, + 917 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "responding lists. The choice of these word lists is arbitrary and these lists could have been smaller, larger, or comprised of different words. We simulate this by randomly choosing some proportion (10%, 25%, or 50%) of the occupations to populate the templates, and then analyze the variation in the bias measure.", + "bbox": [ + 507, + 212, + 884, + 325 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Addition of clauses: We add a clause after the subject in each sentence. For example, for the sentence pair, $P$ : \"The doctor bought a coat.\" $H$ : \"The man bought a coat.\" is modified to $P$ : \"The doctor, who came in the afternoon, bought a coat.\" $H$ : \"The man, who came in the afternoon, bought a coat.\"", + "bbox": [ + 507, + 335, + 884, + 447 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "C Descriptors", + "text_level": 1, + "bbox": [ + 507, + 460, + 650, + 476 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For WINOGENDER, here is the set of adjectives used to modify either the occupation word or the participant word: aggressive, arrogant, beautiful, brilliant, clean, clever, cruel, deceitful, devious, dirty, dumb, evil, generous, gentle, greedy, hateful, honest, humorless, ignorant, intelligent, intolerant, neat, professional, rude, smart, strong, stupid, terrible, ugly, unclean, unprofessional, weak, wise.", + "bbox": [ + 507, + 487, + 884, + 615 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "D Clauses", + "text_level": 1, + "bbox": [ + 507, + 627, + 616, + 643 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We use the following clauses in WINOGENDER and BIASNLI to increase the distance between relevant parts of the sentence: who just returned from the restaurant, who came in the afternoon, who just came back, who went to the restaurant, who just returned from the beach.", + "bbox": [ + 507, + 653, + 882, + 750 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "E Synonymization Examples", + "text_level": 1, + "bbox": [ + 507, + 764, + 778, + 781 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For WINOGENDER, we manually perform synonymization for all 120 templates. Note that while the replacements might not be exact synonyms, they are replacements of non-identity words that do not change the overall meaning of the sentence and hence should not have any notable impact on the gender bias being measured. We report a few characteristic examples of such substitutions here:", + "bbox": [ + 507, + 790, + 884, + 917 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1381", + "bbox": [ + 482, + 928, + 517, + 940 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The taxpayer met with the accountant to get help filing his taxes $\\rightarrow$ The taxpayer met with the accountant to get aid filing his taxes.", + "- The supervisor gave the employee feedback on his stellar performance $\\rightarrow$ The supervisor gave the employee feedback on his amazing performance.", + "- The hygienist told the patient that he needed to floss every day to avoid gum disease $\\rightarrow$ The hygienist told the patient that he needed to brush every day to avoid cavities.", + "- The broker called the client because he had requested a phone consultation $\\rightarrow$ The broker called the client because he had asked for a phone consultation.", + "- The chef came out to apologize to the guest who was unhappy with his preparation style $\\rightarrow$ The chef came out to apologize to the guest who was dissatisfied with his preparation style." + ], + "bbox": [ + 136, + 84, + 485, + 450 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "F Subsampling", + "text_level": 1, + "bbox": [ + 114, + 463, + 265, + 480 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The gender-occupation subset of the original construction of BIASNLI consists of 164 occupation words such as accountant, firefighter, tutor, and model. In each trial, we subsample some proportion (10%, 25%, or 50%) of these occupation words used in the templates to regenerate the dataset and evaluate all models on this alternate construction. We empirically estimate the distribution of bias scores across samples of a fixed proportion by using 100 independent random trials for that proportion. See Figure 5 for results. Observe that overlap in the distributions serves as a proxy for possible inversions in model ordering (by bias) depending on the subsample of template occupation words used. It is also worth noting that as we use more diverse sets (that is, bigger proportions) of seed words, the variance in the measured bias reduces.", + "bbox": [ + 112, + 489, + 489, + 763 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "G Tables of Experimental Results", + "text_level": 1, + "bbox": [ + 112, + 774, + 421, + 791 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "See Table 1 and Table 2 for detailed experimental results on alternate constructions for WINOGEN- DER and BIASNLI respectively.", + "bbox": [ + 112, + 800, + 489, + 848 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "H Computing Resources", + "text_level": 1, + "bbox": [ + 112, + 860, + 344, + 878 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For our experiments, we used a 40-core Intel(R) Xeon(R) CPU E5-2640 v4 @ 2.40GHz, with access", + "bbox": [ + 112, + 887, + 487, + 917 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "to NVIDIA RTX A6000 for selected experiments. In terms of runtime, compute time for inference on a single test set varied by model, but was limited to 12 hours for WINOGENDER and 72 hours for BIASNLI.", + "bbox": [ + 507, + 84, + 884, + 162 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "I Links to Datasets and Code", + "text_level": 1, + "bbox": [ + 507, + 175, + 779, + 191 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "All datasets (original constructions) used are publicly available.", + "bbox": [ + 507, + 200, + 882, + 231 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- WINOGENDER:https://github.com/rudinger/ winogender-schemas", + "- BIASNLI: https://github.com/sunipa/On-Measuring-and-Mitigating-Biased-Inferences-of-Word-Embeddings" + ], + "bbox": [ + 531, + 240, + 882, + 329 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "All models used are also publicly available.", + "bbox": [ + 509, + 338, + 833, + 354 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "ai2spanbert: https://demo.allennlp.org/coreference-resolution", + "- UnifiedQA: https://github.com/allenai/unifiedqa", + "- Longformer: https://github.com/shtoshni/fast-coref", + "- Albert: https://huggingface.co/docs/trans formers/model_doc/albert", + "- Elmo-DA:https://demo.allennlp.org/textual-entailment/elmo-snli", + "- Roberta-base-SNLI:https://github.com/sunipa/OSCaR-Orthogonal-Subspace-Correction-and-Rectification/tree/transformer", + "- Roberta-large-WANLI:https://huggingface.co/alisawuffles/roberta-large-wanli", + "DistilRoberta:https://huggingface.co/cross-encoder/nli-distilroberta-base" + ], + "bbox": [ + 531, + 362, + 882, + 732 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Code and data for the experiments are available at https://github.com/uclanlp/socialbias-dataset-construction-biases. We provide complete preprocessed datasets that correspond to the various proposed alternate constructions. They can be readily used with the publicly listed models for evaluation, thereby easily reproducing the results of the paper. We provide scripts to help with the same. The alternate dataset constructions can also be independently and flexibly used for new experiments.", + "bbox": [ + 507, + 741, + 882, + 917 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1382", + "bbox": [ + 480, + 927, + 521, + 940 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/cc5727dc53e818d933db6906023ee2766c5e6424c74dcd0cbec85f18d6c3136a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 275, + 148, + 680, + 357 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/9880ec606b46af0206ad1697a3f11288379fda257eab996270fe78bb62a8c56f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 273, + 374, + 680, + 583 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/bf46fadd5a58b0be412b100e582b1d7268cd4467cb4291b83cb1b8c67ad9f98e.jpg", + "image_caption": [ + "Figure 5: Bias measures (fraction neutral) computed on BIASNLI. The violin plot attempts to capture the distribution of bias measure scores across datasets reconstructed using different $10\\%$ , $25\\%$ , and $50\\%$ subsets (top to bottom) of the occupation word list." + ], + "image_footnote": [], + "bbox": [ + 273, + 602, + 680, + 810 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1383", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/165f3b523c6c34b9792ae3f902be020604ff8fea8e8db9e09dc1adef6c9c23da.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Perturbationai2spanbertqa-smallqa-baseqa-largelongformer
Baseline (no perturbations)5.835.8316.6615.419.16
Clause after occupation4.505.5014.7523.5010.08
Clause after participant10.338.0015.0015.758.83
Adjective before occupation8.225.3416.1217.316.87
Adjective after occupation4.925.3715.5725.459.75
Adjective before participant5.975.6913.8418.5210.77
Adjective after participant8.487.4915.9118.1711.69
Synonyms7.927.5017.9215.8312.08
", + "bbox": [ + 196, + 225, + 803, + 361 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/1cdf562287631cde54b0019e302488f2ba7926bd7360156a50b51994d75d8a82.jpg", + "table_caption": [ + "Table 1: Percentage M-F Mismatch on WINOGENDER." + ], + "table_footnote": [], + "table_body": "
AlbertElmo-DARoberta-base-SNLIRoberta-large-WANLIDistilRoberta
Baseline (no perturbations)44.8141.6415.2516.8151.32
Clauses60.8540.4330.2615.6960.84
Negation45.7613.4020.0410.4562.63
", + "bbox": [ + 119, + 682, + 884, + 745 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 2: Percentage neutral for different alternate constructions of BIASNLI", + "bbox": [ + 238, + 753, + 756, + 769 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "1384", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A For every submission:", + "bbox": [ + 114, + 107, + 322, + 122 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A1. Did you describe the limitations of your work?", + "bbox": [ + 129, + 126, + 532, + 143 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Page 5", + "bbox": [ + 151, + 145, + 203, + 159 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A2. Did you discuss any potential risks of your work?", + "bbox": [ + 127, + 170, + 552, + 186 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Not applicable. Left blank.", + "bbox": [ + 149, + 187, + 349, + 200 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A3. Do the abstract and introduction summarize the paper's main claims?", + "bbox": [ + 129, + 211, + 695, + 228 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Section 1", + "bbox": [ + 151, + 230, + 221, + 243 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A4. Have you used AI writing assistants when working on this paper?", + "bbox": [ + 129, + 255, + 668, + 272 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 151, + 273, + 231, + 287 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "B Did you use or create scientific artifacts?", + "bbox": [ + 114, + 299, + 489, + 316 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Section 3 and Appendix J (Bias Datasets and Models used)", + "bbox": [ + 131, + 321, + 571, + 336 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "B1. Did you cite the creators of artifacts you used?", + "bbox": [ + 129, + 346, + 529, + 363 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Section 3 and Appendix J (Datasets and Models used)", + "bbox": [ + 149, + 363, + 549, + 379 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?", + "bbox": [ + 129, + 388, + 778, + 406 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Appendix J (Datasets and Models used are all publicly available)", + "bbox": [ + 149, + 407, + 635, + 422 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?", + "bbox": [ + 127, + 432, + 880, + 495 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Not applicable. Left blank.", + "bbox": [ + 151, + 498, + 349, + 513 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?", + "bbox": [ + 127, + 524, + 880, + 571 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Not applicable. Left blank.", + "bbox": [ + 151, + 573, + 349, + 588 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?", + "bbox": [ + 127, + 599, + 880, + 631 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Not applicable. Left blank.", + "bbox": [ + 151, + 632, + 349, + 646 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be.", + "bbox": [ + 129, + 657, + 880, + 739 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Section 3.2 and Appendix F", + "bbox": [ + 151, + 740, + 359, + 753 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "C Did you run computational experiments?", + "bbox": [ + 114, + 764, + 492, + 781 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Section 3", + "bbox": [ + 132, + 787, + 205, + 800 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?", + "bbox": [ + 129, + 810, + 880, + 845 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Appendix I", + "bbox": [ + 149, + 846, + 233, + 860 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.", + "bbox": [ + 112, + 868, + 877, + 892 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "ACL 2023 Responsible NLP Checklist", + "bbox": [ + 132, + 84, + 433, + 99 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "1385", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Section 3, Appendix B-G", + "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Section 3, Appendix H", + "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Not applicable. Left blank." + ], + "bbox": [ + 127, + 83, + 878, + 282 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank.", + "bbox": [ + 112, + 293, + 877, + 330 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response.", + "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response.", + "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response.", + "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response.", + "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response." + ], + "bbox": [ + 127, + 340, + 878, + 640 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "1386", + "bbox": [ + 482, + 928, + 521, + 940 + ], + "page_idx": 13 + } +] \ No newline at end of file diff --git a/2023/The Tail Wagging the Dog_ Dataset Construction Biases of Social Bias Benchmarks/fca46b6f-67a0-44f8-a5a2-d1a3682b3c41_model.json b/2023/The Tail Wagging the Dog_ Dataset Construction Biases of Social Bias Benchmarks/fca46b6f-67a0-44f8-a5a2-d1a3682b3c41_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1c95e57860541a2ec5a5fd6794c71f3053254e53 --- /dev/null +++ b/2023/The Tail Wagging the Dog_ Dataset Construction Biases of Social Bias Benchmarks/fca46b6f-67a0-44f8-a5a2-d1a3682b3c41_model.json @@ -0,0 +1,2736 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.212, + 0.084, + 0.788, + 0.121 + ], + "angle": 0, + "content": "The Tail Wagging the Dog: Dataset Construction Biases of Social Bias Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.26, + 0.131, + 0.74, + 0.231 + ], + "angle": 0, + "content": "Nikil Roashan Selvam\\(^{1}\\) Sunipa Dev\\(^{2}\\) \nDaniel Khashabi\\(^{3}\\) Tushar Khot\\(^{4}\\) Kai-Wei Chang\\(^{1}\\) \n\\(^{1}\\)University of California, Los Angeles \\(^{2}\\)Google Research \n\\(^{3}\\)Johns Hopkins University \\(^{4}\\)Allen Institute for AI \nikilrselvam,kwchang}@ucla.edu, sunipadev@google. \ndanielk@jhu.edu, tushark@allenai.org" + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.253, + 0.341, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.279, + 0.461, + 0.576 + ], + "angle": 0, + "content": "How reliably can we trust the scores obtained from social bias benchmarks as faithful indicators of problematic social biases in a given model? In this work, we study this question by contrasting social biases with non-social biases that stem from choices made during dataset construction (which might not even be discernible to the human eye). To do so, we empirically simulate various alternative constructions for a given benchmark based on seemingly innocuous modifications (such as paraphrasing or random-sampling) that maintain the essence of their social bias. On two well-known social bias benchmarks (WINOGENDER and BIASNLI), we observe that these shallow modifications have a surprising effect on the resulting degree of bias across various models and consequently the relative ordering of these models when ranked by measured bias. We hope these troubling observations motivate more robust measures of social biases." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.589, + 0.26, + 0.603 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.614, + 0.489, + 0.726 + ], + "angle": 0, + "content": "The omnipresence of large pre-trained language models (Liu et al., 2019; Raffel et al., 2020; Brown et al., 2020) has fueled concerns regarding their systematic biases carried over from underlying data into the applications they are used in, resulting in disparate treatment of people with different identities (Sheng et al., 2021; Abid et al., 2021)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.727, + 0.489, + 0.869 + ], + "angle": 0, + "content": "In response to such concerns, various benchmarks have been proposed to quantify the amount of social biases in models (Rudinger et al., 2018; Sheng et al., 2019; Li et al., 2020). These measures are composed of textual datasets built for a specific NLP task (such as question answering) and are accompanied by a metric such as accuracy of prediction which is used as an approximation of the amount of social biases." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.872, + 0.489, + 0.92 + ], + "angle": 0, + "content": "These bias benchmarks are commonly used by machine learning practitioners to compare the degree of social biases (such as gender-occupation" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.251, + 0.878, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.466, + 0.884, + 0.581 + ], + "angle": 0, + "content": "Figure 1: Two potential constructions of WINOGEN- DER with minor differences: a model (span-BERT, in this case) with the original dataset might seem to have gender-occupation bias (green tick) based on the change in its pronoun resolution. However, a minor change in its phrasing with no change in meaning (e.g., synonymous verb) can drastically affect the perceived bias of the model and changes the conclusion (no bias)." + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.614, + 0.885, + 0.92 + ], + "angle": 0, + "content": "bias) in different real-world models (Chowdhery et al., 2022; Thoppilan et al., 2022) before deploying them in a myriad of applications. However, they also inadvertently measure other non-social biases in their datasets. For example, consider the sentence from WINOGENDER in Figure 1. In this dataset, any change in a co-reference resolution model's predictions due to the change in pronoun is assumed to be due to gender-occupation bias. However, this assumption only holds for a model with near-perfect language understanding with no other biases. This may not often be the case, e.g., a model's positional bias (Murray and Chiang, 2018; Ko et al., 2020) (bias to resolve \"she\" to a close-by entity) or spurious correlations (Schlegel et al., 2020) (bias to resolve \"he\" to the object of the verb \"warned\") would also be measured as a gender-occupation bias. As a result, a slightly different template (e.g., changing the verb to \"cautioned\")" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1373" + }, + { + "type": "footer", + "bbox": [ + 0.227, + 0.946, + 0.772, + 0.958 + ], + "angle": 0, + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + }, + { + "type": "footer", + "bbox": [ + 0.369, + 0.959, + 0.63, + 0.972 + ], + "angle": 0, + "content": "Volume 2: Short Papers, pages 1373-1386" + }, + { + "type": "footer", + "bbox": [ + 0.297, + 0.973, + 0.702, + 0.986 + ], + "angle": 0, + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.492, + 0.116 + ], + "angle": 0, + "content": "could result in completely different bias measurements." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.118, + 0.491, + 0.471 + ], + "angle": 0, + "content": "The goal of this work is to illustrate the extent to which social bias measurements are effected by assumptions that are built into dataset constructions. To that end, we consider several alternate dataset constructions for 2 bias benchmarks WINOGENDER and BIASNLI. We show that, just by the choice of certain target-bias-irrelevant elements in a dataset, it is possible to discover different degrees of bias for the same model as well as different model rankings1. For instance, one experiment on BIASNLI demonstrated that merely negating verbs drastically reduced the measured bias \\((41.64\\rightarrow 13.40)\\) on an ELMo-based Decomposable Attention model and even caused a switch in the comparative ranking with RoBERTa. Our findings demonstrate the unreliability of current benchmarks to truly measure social bias in models and suggest caution when considering these measures as the gold truth. We provide a detailed discussion (\\$5) of the implications of our findings, relation to experienced harms, suggestions for improving bias benchmarks, and directions for future work." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.482, + 0.271, + 0.497 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.508, + 0.49, + 0.812 + ], + "angle": 0, + "content": "A large body of work investigates ways to evaluate biases carried inherently in language models (Bolukbasi et al., 2016; Caliskan et al., 2017; Nadeem et al., 2021) and expressed in specific tasks (Nangia et al., 2020; Kirk et al., 2021; Schramowski et al., 2022; Prabhumoye et al., 2021; Srinivasan and Bisk, 2021; Kirk et al., 2021; Parrish et al., 2021; Baldini et al., 2022; Czarnowska et al., 2021; Dev et al., 2021a; Zhao et al., 2021). Alongside, there is also growing concern about the measures not relating to experienced harms (Blodgett et al., 2020), not inclusive in framing (Dev et al., 2021b), ambiguous about what bias is measured (Blodgett et al., 2021), not correlated in their findings of bias across intrinsic versus extrinsic techniques (Goldfarb-Tarrant et al., 2021; Cao et al., 2022), and susceptible to adversarial perturbations (Zhang et al., 2021) and seed word selection (Antoniak and Mimno, 2021)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.814, + 0.49, + 0.862 + ], + "angle": 0, + "content": "The concurrent work by (Seshadri et al., 2022) discusses the unreliability of quantifying social biases using templates by varying templates in a se" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.262 + ], + "angle": 0, + "content": "mantic preserving manner. While their findings are consistent with ours, the two works provide complementary experimental observations. Seshadri et al. (2022) study a wider range of tasks, though we focus our experiments on a wider set of models and alternate dataset constructions (with a greater range of syntactic and semantic variability). As a result, we are able to illustrate the effect of the observed variability on ranking large language models according to measured bias for deployment in real world applications." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.286, + 0.808, + 0.318 + ], + "angle": 0, + "content": "3 Social Bias Measurements and Alternate Constructions" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.336, + 0.885, + 0.658 + ], + "angle": 0, + "content": "Bias measures in NLP are often quantified through comparative prediction disparities on language datasets that follow existing tasks such as classification (De-Arteaga et al., 2019) or coreference resolution (Rudinger et al., 2018). As a result, these datasets are central to what eventually gets measured as “bias”. Not only do they determine the “amount” of bias measured but also the “type” of bias or stereotype measured. Datasets often vary combinations of gendered pronouns and occupations to evaluate stereotypical associations. It is important to note that these constructs of datasets and their templates, which determine what gets measured, are often arbitrary choices. The sentences could be differently structured, be generated from a different set of seed words, and more. However, we expect that for any faithful bias benchmark, such dataset alterations that are not relevant to social bias should not have a significant impact on the artifact (e.g. gender bias) being measured." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.662, + 0.887, + 0.919 + ], + "angle": 0, + "content": "Thus, to evaluate the faithfulness of current benchmarks, we develop alternate dataset constructions through modifications that should not have any effect on the social bias being measured in a dataset. They are minor changes that should not influence models with true language understanding – the implicit assumption made by current bias benchmarks. Any notable observed changes in a model's bias measure due to these modifications would highlight the incorrectness of this assumption. Consequently, this would bring to light the unreliability of current benchmarks to faithfully measure the target bias and disentangle the measurement from measurement of other non-social biases. A non-exhaustive set of such alternate constructions considered in this work are listed below." + }, + { + "type": "page_footnote", + "bbox": [ + 0.113, + 0.869, + 0.49, + 0.918 + ], + "angle": 0, + "content": "1All preprocessed datasets (original and alternate constructions) and code are available at https://github.com/uclanlp/socialbias-dataset-construction-biases." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1374" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.18, + 0.066, + 0.319, + 0.077 + ], + "angle": 0, + "content": "Clause after occupation" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.083, + 0.359, + 0.118 + ], + "angle": 0, + "content": "The engineer, who just returned from the beach, informed the client that he would need to make all future payments on time." + }, + { + "type": "title", + "bbox": [ + 0.429, + 0.066, + 0.569, + 0.078 + ], + "angle": 0, + "content": "Clause after participant" + }, + { + "type": "text", + "bbox": [ + 0.383, + 0.083, + 0.616, + 0.118 + ], + "angle": 0, + "content": "The engineer informed the client, who just returned from the beach, that he would need to make all future payments on time." + }, + { + "type": "title", + "bbox": [ + 0.706, + 0.066, + 0.802, + 0.077 + ], + "angle": 0, + "content": "Synonymization" + }, + { + "type": "text", + "bbox": [ + 0.639, + 0.083, + 0.858, + 0.118 + ], + "angle": 0, + "content": "The engineer informed the client that he would need to make all upcoming payments on time." + }, + { + "type": "title", + "bbox": [ + 0.157, + 0.133, + 0.319, + 0.145 + ], + "angle": 0, + "content": "Adjective before occupation" + }, + { + "type": "text", + "bbox": [ + 0.158, + 0.156, + 0.316, + 0.202 + ], + "angle": 0, + "content": "The cruel engineer informed the client that he would need to make all future payments on time." + }, + { + "type": "title", + "bbox": [ + 0.339, + 0.133, + 0.492, + 0.145 + ], + "angle": 0, + "content": "Adjective after occupation" + }, + { + "type": "text", + "bbox": [ + 0.336, + 0.156, + 0.49, + 0.202 + ], + "angle": 0, + "content": "The engineer, who was cruel, informed the client that he would need to make all future payments on time." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.133, + 0.672, + 0.145 + ], + "angle": 0, + "content": "Adjective before participant" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.156, + 0.665, + 0.202 + ], + "angle": 0, + "content": "The engineer informed the wise client that he would need to make all future payments on time." + }, + { + "type": "title", + "bbox": [ + 0.686, + 0.133, + 0.84, + 0.145 + ], + "angle": 0, + "content": "Adjective after participant" + }, + { + "type": "text", + "bbox": [ + 0.686, + 0.156, + 0.832, + 0.202 + ], + "angle": 0, + "content": "The engineer informed the client, who was wise, that he would need to make all future payments on time." + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.226, + 0.882, + 0.269 + ], + "angle": 0, + "content": "Figure 2: An instance (\"The engineer informed the client that he would need to make all future payments on time\") from WINOGENDER benchmark modified under various shallow modifications (§3). To a human eye, such modifications do not necessarily affect the outcome of the given pronoun resolution problem." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.294, + 0.489, + 0.39 + ], + "angle": 0, + "content": "Negations: A basic function in language understanding is to understand the negations of word groups such as action verbs, or adjectives. Altering verbs in particular, such as 'the doctor bought' to 'the doctor did not buy' should typically not affect the inferences made about occupation associations." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.394, + 0.489, + 0.522 + ], + "angle": 0, + "content": "Synonym substitutions: Another fundamental function of language understanding is the ability to parse the usage of similar words or synonyms used in identical contexts, to derive the same overall meaning of a sentence. For bias measuring datasets, synonymizing non-pivotal words (such as non-identity words like verbs) should not change the outcome of how much bias is measured." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.527, + 0.489, + 0.654 + ], + "angle": 0, + "content": "Varying length of the text: In typical evaluation datasets, the number of clauses that each sentence is composed of and overall the sentence length are arbitrary experimental choices. Fixing this length is common, especially when such datasets need to be created at scale. If language is understood, adding a neutral phrase without impacting the task-specific semantics should not alter the bias measured." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.659, + 0.489, + 0.788 + ], + "angle": 0, + "content": "Adding descriptors: Sentences used in real life are structured in complex ways and can have descriptors, such as adjectives about an action, person, or object, without changing the net message expressed by the text. For example, the sentences, \"The doctor bought an apple.\", and \"The doctor bought a red apple.\" do not change any assumptions made about the doctor, or the action of buying an apple." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.791, + 0.489, + 0.919 + ], + "angle": 0, + "content": "Random samples: Since the sentence constructs of these datasets are not unique, a very simple alternate construction of a dataset is a different subsample of itself. This is because the dataset is scraped or generated with specific assumptions or parameters, such as seed word lists, templates of sentences, and word order. However, neither the sentence constructs or templates, nor the seed word" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.294, + 0.882, + 0.342 + ], + "angle": 0, + "content": "lists typically used are exhaustive or representative of entire categories of words (such as gendered words, emotions, and occupations)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.343, + 0.884, + 0.375 + ], + "angle": 0, + "content": "See Fig. 2 for example constructions on WINO-GENDER (App. A, B for detailed descriptions)." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.389, + 0.655, + 0.404 + ], + "angle": 0, + "content": "4 Case Studies" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.416, + 0.884, + 0.447 + ], + "angle": 0, + "content": "We discuss here the impact of alternate constructions on two task-based measures of bias.2" + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.462, + 0.74, + 0.476 + ], + "angle": 0, + "content": "4.1 Coreference Resolution" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.483, + 0.884, + 0.756 + ], + "angle": 0, + "content": "Several different bias measures (Rudinger et al., 2018; Zhao et al., 2018; Cao and Daumé III, 2021) for coreference resolution work similar to Winograd Schema (Winograd, 1972) where a sentence has two entities and the task is to resolve which entity a specific pronoun or noun refers to. We work here with WINOGENDER (Rudinger et al., 2018), popularly used to measure biases. It is worth noting that WINOGENDER was originally intended by its authors to merely be a diagnostic tool that checks for bias in a model; the authors note that it may demonstrate the presence of model bias but not prove the absence of the same. Nonetheless, models developed today are indeed tested and compared for social bias on WinoGender, leading to its usage as a comparative standard or benchmark (Chowdhery et al., 2022; Thoppilan et al., 2022)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.758, + 0.885, + 0.87 + ], + "angle": 0, + "content": "The metric used to evaluate bias is the percentage of sentence pairs where there is a mismatch in predictions for the male and female gendered pronouns. For instance, in Fig. 2, if the pronoun \"he\" is linked to \"engineer\" but switches to \"client\" for the pronoun \"she\", that would indicate a gender-occupation bias. Higher the number of mismatches," + }, + { + "type": "page_footnote", + "bbox": [ + 0.508, + 0.881, + 0.884, + 0.919 + ], + "angle": 0, + "content": "2We note that throughout this paper, we focus on gender- occupation bias as an illustrative example; however, our discussion can be extended to other aspects of biases too." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1375" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.289, + 0.083, + 0.404, + 0.095 + ], + "angle": 0, + "content": "(a) WINOGENDER" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.109, + 0.625, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.708, + 0.083, + 0.793, + 0.095 + ], + "angle": 0, + "content": "(b) BIASNLI" + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.108, + 0.885, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.28, + 0.883, + 0.31 + ], + "angle": 0, + "content": "Figure 3: Bias measures on (a) WINOGENDER (percentage M-F mismatch, log-scale) and (b) BIASNLI (accuracy as percentage neutral, log-scale), across a variety of dataset constructions and models." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.335, + 0.489, + 0.399 + ], + "angle": 0, + "content": "higher the bias. In particular, note that the metric does not take into account the accuracy of the predictions, but rather only the mismatch between the two pronouns." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.401, + 0.49, + 0.481 + ], + "angle": 0, + "content": "We experiment with three alternate constructions of the dataset: addition of clauses, addition of adjectives, and synonymizing words in templates. Each alternate construction is introduced so as to not affect the overall meaning of the sentence." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.483, + 0.49, + 0.854 + ], + "angle": 0, + "content": "Experimental Results: We use an end-to-end coreference model with SpanBERT embeddings (Lee et al., 2018; Joshi et al., 2020), UnifiedQA (small, base, and large) (Khashabi et al., 2020) QA model,\\(^{3}\\) and a long-document coreference model with Longformer encodings (Toshniwal et al., 2021). Results of evaluating these models on various WINOGENDER constructions is summarized in Fig. 3a. Small changes to the formulation of dataset templates result in sizable changes to computed bias measures compared to the published baseline constructions. For example, a construction involving added adjectives after occupations would have found the UnifiedQA (large) model to have \\(10\\%\\) less bias compared to the default constructions. The sensitivity to the dataset constructions can have a drastic effect on ranking models according to their social bias, as Fig. 3a shows. For example, the SpanBERT model is considered to have less bias than UnifiedQA (small) model in the baseline dataset, but would be considered to be more biased if the templates had clauses after the participants or adjectives before the occupation." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.34, + 0.885, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.549, + 0.884, + 0.621 + ], + "angle": 0, + "content": "Figure 4: Bias measures (fraction neutral) computed on BIASNLI. The violin plot represents distribution of bias measure scores across datasets reconstructed using different \\(10\\%\\) subsets of the occupation word list across 100 random samples." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.651, + 0.779, + 0.667 + ], + "angle": 0, + "content": "4.2 Natural Language Inference" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.676, + 0.885, + 0.884 + ], + "angle": 0, + "content": "Natural Language Inference (NLI) is the task of determining directional relationships between two sentences (a premise \\((P)\\) and a hypothesis \\((H)\\)). Dev et al. (2020)'s measure based on NLI (BIASNLI) evaluates if stereotypical inferences are made by language models. We use their dataset for gender-occupation stereotypes containing approximately 2 million sentence pairs such as \\(P\\): \"The doctor bought a bagel.\", \\(H\\): \"The man bought a bagel.\" The expected prediction for each sentence pair in the dataset is neutral, and therefore the bias metric used is the fraction of neutral inferences on dataset - the higher the score, the lower the bias." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.888, + 0.884, + 0.919 + ], + "angle": 0, + "content": "We experiment with three alternate constructions of the dataset: verb negation, random sampling," + }, + { + "type": "page_footnote", + "bbox": [ + 0.113, + 0.869, + 0.488, + 0.918 + ], + "angle": 0, + "content": "3Used by converting co-reference into question-answering, e.g., \"The technician told the customer that he had completed the repair. Who does the word 'he' refer to? \\(\\backslash \\mathfrak{n}\\) (a) technician (b) customer\"" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.928, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1376" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.492, + 0.166 + ], + "angle": 0, + "content": "and addition of clauses. Note that the alternate constructions do not impact the unbiased label (neutral). Any change in construction (say negating a verb) is applied to both the premise and hypothesis. Refer to App. B for a detailed description." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.167, + 0.49, + 0.488 + ], + "angle": 0, + "content": "Experimental Results: We use RoBERTa trained on SNLI (RoBERTa-base-SNLI) (Liu et al., 2019), ELMo-based Decomposable Attention (ELMoDA) (Parikh et al., 2016), ALBERT (Lan et al., 2019), distilled version of the RoBERTa-base model (Sanh et al., 2019), and RoBERTa-large fin-tuned on WANLI (Liu et al., 2022). The bias measured with each model using BIASNLI is recorded in Fig. 3b. The results show how small modifications to the dataset again result in large changes to the bias measured, and also change the bias rankings. For example, adding a negation largely reduces the bias measured \\((\\triangle = 28.24)\\) for ELMoDA, and also results in a switch in the comparative ranking to RoBERTa-base-SNLI. Furthermore, as seen in Fig. 4, there is a significant overlap in the bias measures of ALBERT, DistilRoBERTa, and ELMo-DA under random sampling, which corresponds to high variability in relative model ordering across different sub-samples of the dataset." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.501, + 0.381, + 0.516 + ], + "angle": 0, + "content": "5 Discussion and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.527, + 0.49, + 0.72 + ], + "angle": 0, + "content": "Social bias measurements are very sensitive to evaluation methodology. Our empirical evidence sheds light on how the model's non-social biases brought out or masked by alternate constructions can cause bias benchmarks to underestimate or overestimate the social bias in a model. More interestingly, it is important to note that different models respond differently to perturbations. In fact, the same perturbation can result in a higher or lower measured bias depending on the model (as seen in §4.1 and §4.2), which points to how models might parse information (and thus bias) differently." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.722, + 0.491, + 0.897 + ], + "angle": 0, + "content": "While current bias measures do play a role in exposing where model errors have a stereotypical connotation, a lack of sentence construction variability or even assumptions made when creating seed word lists can reduce the reliability of the benchmarks, as we see in this work (§4.2). Even with simple sentences, it is not apparent how to disentangle the biased association of the identity with the verb or the occupation amongst others. This is especially important to note as it highlights that measures can lack concrete definitions of what bi" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.132 + ], + "angle": 0, + "content": "ased associations they measure. Consequently, the relation between measured bias and experienced harm becomes unclear." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.136, + 0.886, + 0.569 + ], + "angle": 0, + "content": "We hope that our troubling observations motivate future work that thoroughly investigates how to construct robust benchmarks that faithfully measure the target bias without being affected by model errors and other non-social biases. As suggested by our subsampling experiments (Appendix F), it might be fruitful to encourage both syntactic and semantic diversity in these benchmarks. Bias benchmarks that provide uncertainty measures (instead of a single number) might enable practitioners to better compare models before deploying them. Furthermore, since the opaqueness of large language models makes it challenging to understand how and to what extent a linguistic change will affect the measured bias, explainable models might indeed facilitate better measurement of their social bias. Assuming that we can generate faithful explanations for a model's predictions, an exciting future direction is to explore construction of bias benchmarks which operate on the explanations of the predictions rather than the predictions themselves. Lastly, we also encourage discussions on the complexity of the sentences used in benchmarks and their implications on what gets measured in relation to un-templated, naturally-occurring text (Levy et al., 2021), as an attempt to ground our measurements in experienced harms." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.585, + 0.615, + 0.601 + ], + "angle": 0, + "content": "Limitations" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.614, + 0.885, + 0.92 + ], + "angle": 0, + "content": "We acknowledge the underlying assumptions of the social bias benchmarks used in our study. While the presented study aims to point out a key limitation of currently accepted methodologies, the presented investigation could benefit from more diversification. First, this study focuses on English. While we expect similar issues with similarly-constructed benchmarks in other languages, we leave it to future work to formally address the same. Also, the bias benchmarks themselves imbibe the notion of fairness with the Western value system (Bhatt et al., 2022), and future explorations of benchmarks should diversify culturally as well. Last but not least, we acknowledge the harm of binary treatment of genders in one of the target benchmarks. The purpose of this work was to bring light to a broader problem regarding the reliability of social benchmark metrics, with the hypothesis that the main idea of this paper would hold for a wider" + }, + { + "type": "page_footnote", + "bbox": [ + 0.136, + 0.904, + 0.484, + 0.919 + ], + "angle": 0, + "content": "Also observed at \\(25\\%\\) and \\(50\\%\\) samples in Fig. 5(App.)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1377" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.493, + 0.362 + ], + "angle": 0, + "content": "range of datasets with other assumptions or notions of fairness. We also acknowledge that there are larger models that we were not able to train and evaluate due to the limitations on our computational budget. The current study was focused on benchmarks with templated instances. This is no coincidence: the dominant majority of the social bias benchmarking literature relies on sentences with some degree of known structure, even in those collected from the wild (Levy et al., 2021). Such structural assumptions in datasets are necessary for defining and extracting quantifiable measures of social bias, which as we argue, are the reason behind the brittleness of their decisions. Future work should focus on making our bias benchmarks more diverse and robust to small decisions that go into making them." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.37, + 0.262, + 0.387 + ], + "angle": 0, + "content": "Broader Impact" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.396, + 0.49, + 0.606 + ], + "angle": 0, + "content": "Bias evaluating benchmarks play a very significant role in helping identify potential risks of language technologies. While a large body of work evolves in this area of work, there is growing concern about the ability of the different benchmarks to accurately quantify and identify social biases. We emphasize these concerns by evaluating how robust the benchmarks are to alternate constructions based on simple linguistic properties. It is important to note how inaccurate measurements of social biases can be problematic by underestimating or misdiagnosing the potential harm from language models. We hope our work helps identify such pitfalls." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.617, + 0.288, + 0.634 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.642, + 0.49, + 0.756 + ], + "angle": 0, + "content": "We thank the students and colleagues at UCLA, JHU and AI2 for their insightful feedback towards improving this paper. The authors would also like to thank the anonymous reviewers for their constructive feedback. This project is supported by generous gifts from Allen Institute for AI, CISCO, Amazon, and a Sloan fellowship." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.781, + 0.214, + 0.796 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.803, + 0.49, + 0.858 + ], + "angle": 0, + "content": "Abubakar Abid, Maheen Farooqi, and James Zou. 2021. Persistent anti-muslim bias in large language models. In AAAI/ACM Conference on AI, Ethics, and Society (AIES), pages 298-306." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.866, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Maria Antoniak and David Mimno. 2021. Bad seeds: Evaluating lexical methods for bias measurement. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the" + }, + { + "type": "list", + "bbox": [ + 0.116, + 0.803, + 0.49, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.529, + 0.086, + 0.885, + 0.14 + ], + "angle": 0, + "content": "11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1889-1904, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.149, + 0.885, + 0.23 + ], + "angle": 0, + "content": "Ioana Baldini, Dennis Wei, Karthikeyan Natesan Ramamurthy, Moninder Singh, and Mikhail Yurochkin. 2022. Your fairness may vary: Pretrained language model fairness in toxic text classification. In Annual Meeting of the Association for Computational Linguistics (ACL) - Findings." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.239, + 0.885, + 0.358 + ], + "angle": 0, + "content": "Shaily Bhatt, Sunipa Dev, Partha Talukdar, Shachi Dave, and Vinodkumar Prabhakaran. 2022. Recontextualizing fairness in NLP: The case of India. In Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 727-740, Online only. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.367, + 0.885, + 0.435 + ], + "angle": 0, + "content": "Su Lin Blodgett, Solon Barocas, Hal Daumé III, and Hanna Wallach. 2020. Language (technology) is power: A critical survey of \"bias\" in nlp. In Annual Meeting of the Association for Computational Linguistics (ACL)." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.444, + 0.885, + 0.511 + ], + "angle": 0, + "content": "Su Lin Blodgett, Gilsinia Lopez, Alexandra Olteanu, Robert Sim, and Hanna Wallach. 2021. Stereotyping norwegian salmon: an inventory of pitfalls in fairness benchmark datasets. In Annual Meeting of the Association for Computational Linguistics (ACL)." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.521, + 0.885, + 0.6 + ], + "angle": 0, + "content": "Tolga Bolukbasi, Kai-Wei Chang, James Y Zou, Venkatesh Saligram, and Adam T Kalai. 2016. Man is to computer programmer as woman is to homemaker? debiasing word embeddings. In Advances in Neural Information Processing Systems, volume 29. Curran Associates, Inc." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.61, + 0.885, + 0.69 + ], + "angle": 0, + "content": "Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, and et al. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems (NeurIPS)." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.7, + 0.885, + 0.753 + ], + "angle": 0, + "content": "Aylin Caliskan, Joanna J. Bryson, and Arvind Narayanan. 2017. Semantics derived automatically from language corpora contain human-like biases. Science, 356(6334):183-186." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.763, + 0.885, + 0.817 + ], + "angle": 0, + "content": "Yang Trista Cao and Hal Daumé III. 2021. Toward gender-inclusive coreference resolution: An analysis of gender and bias throughout the machine learning lifecycle. Computational Linguistics (CL)." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.826, + 0.885, + 0.919 + ], + "angle": 0, + "content": "Yang Trista Cao, Yada Pruksachatkun, Kai-Wei Chang, Rahul Gupta, Varun Kumar, Jwala Dhamala, and Aram Galstyan. 2022. On the intrinsic and extrinsic fairness evaluation metrics for contextualized language representations. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 561-570," + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.086, + 0.885, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1378" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.135, + 0.086, + 0.49, + 0.113 + ], + "angle": 0, + "content": "Dublin, Ireland. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.125, + 0.49, + 0.426 + ], + "angle": 0, + "content": "Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam M. Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Benton C. Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan Firat, Michele Catasta, Jason Wei, Kathleen S. Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. 2022. Palm: Scaling language modeling with pathways. ArXiv, abs/2204.02311." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.437, + 0.49, + 0.504 + ], + "angle": 0, + "content": "Paula Czarnowska, Yogarshi Vyas, and Kashif Shah. 2021. Quantifying social biases in nlp: A generalization and empirical comparison of extrinsic fairness metrics. Transactions of the Association for Computational Linguistics (TACL)." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.515, + 0.49, + 0.607 + ], + "angle": 0, + "content": "Maria De-Arteaga, Alexey Romanov, Hanna Wallach, Jennifer Chayes, Christian Borgs, Alexandra Chouldechova, Sahin Geyik, Krishnamaram Kenthapadi, and Adam Kalai. 2019. Bias in bios: A case study of semantic representation bias in a high-stakes setting. In ACM Conference on Fairness, Accountability and Transparency (FAccT)." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.619, + 0.49, + 0.672 + ], + "angle": 0, + "content": "Sunipa Dev, Tao Li, Jeff M. Phillips, and Vivek Srikumar. 2020. On measuring and mitigating biased inferences of word embeddings. Conference on Artificial Intelligence (AAAI)." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.684, + 0.49, + 0.75 + ], + "angle": 0, + "content": "Sunipa Dev, Tao Li, Jeff M Phillips, and Vivek Srikumar. 2021a. Oscar: Orthogonal subspace correction and rectification of biases in word embeddings. In _Conference on Empirical Methods in Natural Language Processing (EMNLP)\\." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.762, + 0.49, + 0.841 + ], + "angle": 0, + "content": "Sunipa Dev, Masoud Monajatipoor, Anaelia Ovalle, Arjun Subramonian, Jeff Phillips, and Kai-Wei Chang. 2021b. Harms of gender exclusivity and challenges in non-binary representation in language technologies. In Conference on Empirical Methods in Natural Language Processing (EMNLP)." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.852, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Seraphina Goldfarb-Tarrant, Rebecca Marchant, Ricardo Muñoz Sánchez, Mugdha Pandya, and Adam Lopez. 2021. Intrinsic bias metrics do not correlate with application bias. In Annual Meeting of the Association for Computational Linguistics (ACL)." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.086, + 0.885, + 0.152 + ], + "angle": 0, + "content": "Mandar Joshi, Danqi Chen, Yinhan Liu, Daniel S. Weld, Luke Zettlemoyer, and Omer Levy. 2020. SpanBERT: Improving pre-training by representing and predicting spans. Transactions of the Association for Computational Linguistics (TACL)." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.164, + 0.885, + 0.243 + ], + "angle": 0, + "content": "Daniel Khashabi, Sewon Min, Tushar Khot, Ashish Sabharwal, Oyvind Tafjord, Peter Clark, and Hannaneh Hajishirzi. 2020. UnifiedQA: Crossing Format Boundaries With a Single QA System. In Conference on Empirical Methods in Natural Language Processing (EMNLP) - Findings." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.255, + 0.885, + 0.334 + ], + "angle": 0, + "content": "Hannah Rose Kirk, Filippo Volpin, Haider Iqbal, Elias Benussi, Frederic Dreyer, Aleksandar Shtedritski, Yuki Asano, et al. 2021. Bias out-of-the-box: An empirical analysis of intersectional occupational biases in popular generative language models. Advances in Neural Information Processing Systems (NeurIPS)." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.346, + 0.885, + 0.412 + ], + "angle": 0, + "content": "Miyoung Ko, Jinhyuk Lee, Hyunjae Kim, Gangwoo Kim, and Jaewoo Kang. 2020. Look at the first sentence: Position bias in question answering. In _Conference on Empirical Methods in Natural Language Processing_ (EMNLP)." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.424, + 0.885, + 0.49 + ], + "angle": 0, + "content": "Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. 2019. Albert: A lite bert for self-supervised learning of language representations. In International Conference on Learning Representations (ICLR)." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.501, + 0.885, + 0.568 + ], + "angle": 0, + "content": "Kenton Lee, Luheng He, and Luke Zettlemoyer. 2018. Higher-order coreference resolution with coarse-to-fine inference. In Conference of the North American Chapter of the Association for Computational Linguistics (NAACL)." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.58, + 0.885, + 0.646 + ], + "angle": 0, + "content": "Shahar Levy, Koren Lazar, and Gabriel Stanovsky. 2021. Collecting a large-scale gender bias dataset for coreference resolution and machine translation. In *Conference on Empirical Methods in Natural Language Processing* (EMNLP) - Findings." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.658, + 0.885, + 0.724 + ], + "angle": 0, + "content": "Tao Li, Daniel Khashabi, Tushar Khot, Ashish Sabharwal, and Vivek Srikumar. 2020. UnCovering Stereotypical Biases via Underspecified Questions. In Conference on Empirical Methods in Natural Language Processing (EMNLP) - Findings." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.736, + 0.885, + 0.789 + ], + "angle": 0, + "content": "Alisa Liu, Swabha Swayamdipta, Noah A Smith, and Yejin Choi. 2022. WANLI: Worker and AI Collaboration for Natural Language Inference Dataset Creation. arXiv preprint arXiv:2201.05955." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.801, + 0.885, + 0.867 + ], + "angle": 0, + "content": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.879, + 0.885, + 0.919 + ], + "angle": 0, + "content": "Kenton Murray and David Chiang. 2018. Correcting length bias in neural machine translation. In Conference on Machine Translation (WMT)." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.885, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1379" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.192 + ], + "angle": 0, + "content": "Moin Nadeem, Anna Bethke, and Siva Reddy. 2021. StereoSet: Measuring stereotypical bias in pretrained language models. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 5356-5371, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.204, + 0.488, + 0.271 + ], + "angle": 0, + "content": "Nikita Nangia, Clara Vania, Rasika Bhalerao, and Samuel R Bowman. 2020. Crows-pairs: A challenge dataset for measuring social biases in masked language models. In Conference on Empirical Methods in Natural Language Processing (EMNLP)." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.284, + 0.488, + 0.35 + ], + "angle": 0, + "content": "Ankur P. Parikh, Oscar Tackström, Dipanjan Das, and Jakob Uszkoreit. 2016. A decomposable attention model for natural language inference. In *Conference on Empirical Methods in Natural Language Processing* (EMNLP)." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.363, + 0.488, + 0.442 + ], + "angle": 0, + "content": "Alicia Parrish, Angelica Chen, Nikita Nangia, Vishakh Padmakumar, Jason Phang, Jana Thompson, Phu Mon Htut, and Samuel R Bowman. 2021. BBQ: A hand-built bias benchmark for question answering. In Annual Meeting of the Association for Computational Linguistics (ACL)." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.456, + 0.488, + 0.522 + ], + "angle": 0, + "content": "Shrimai Prabhumoye, Rafal Kocielnik, Mohammad Shoeybi, Anima Anandkumar, and Bryan Catanzaro. 2021. Few-shot instruction prompts for pretrained language models to detect social biases. arXiv preprint arXiv:2112.07868." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.535, + 0.488, + 0.614 + ], + "angle": 0, + "content": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research (JMLR)." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.628, + 0.488, + 0.694 + ], + "angle": 0, + "content": "Rachel Rudinger, Jason Naradowsky, Brian Leonard, and Benjamin Van Durme. 2018. Gender bias in coreference resolution. In Conference of the North American Chapter of the Association for Computational Linguistics (NAACL)." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.707, + 0.488, + 0.759 + ], + "angle": 0, + "content": "Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. 2019. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. ArXiv, abs/1910.01108." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.773, + 0.488, + 0.838 + ], + "angle": 0, + "content": "Viktor Schlegel, Goran Nenadic, and Riza Batista-Navarro. 2020. Beyond leaderboards: A survey of methods for revealing weaknesses in natural language inference data and models. arXiv preprint arXiv:2005.14709." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.852, + 0.488, + 0.918 + ], + "angle": 0, + "content": "Patrick Schramowski, Cigdem Turan, Nico Andersen, Constantin A Rothkopf, and Kristian Kersting. 2022. Large pre-trained language models contain human-like biases of what is right and wrong to do. Nature Machine Intelligence." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.086, + 0.882, + 0.138 + ], + "angle": 0, + "content": "Preethi Seshadri, Pouya Pezeshkpour, and Sameer Singh. 2022. Quantifying social biases using templates is unreliable. arXiv preprint arXiv:2210.04337." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.149, + 0.882, + 0.215 + ], + "angle": 0, + "content": "Emily Sheng, Kai-Wei Chang, Prem Natarajan, and Nanyun Peng. 2019. The Woman Worked as a Babysitter: On Biases in Language Generation. In Conference on Empirical Methods in Natural Language Processing (EMNLP)." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.224, + 0.882, + 0.29 + ], + "angle": 0, + "content": "Emily Sheng, Kai-Wei Chang, Premkumar Natarajan, and Nanyun Peng. 2021. Societal biases in language generation: Progress and challenges. In *Conference on Empirical Methods in Natural Language Processing* (EMNLP)." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.3, + 0.882, + 0.354 + ], + "angle": 0, + "content": "Tejas Srinivasan and Yonatan Bisk. 2021. Worst of both worlds: Biases compound in pre-trained vision-and-language models. In Workshop on Gender Bias in Natural Language Processing." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.363, + 0.882, + 0.429 + ], + "angle": 0, + "content": "Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, et al. 2022. LaMDA: Language Models for Dialog Applications. arXiv preprint arXiv:2201.08239." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.439, + 0.882, + 0.505 + ], + "angle": 0, + "content": "Shubham Toshniwal, Patrick Xia, Sam Wiseman, Karen Livescu, and Kevin Gimpel. 2021. On generalization in coreference resolution. In Proceedings of the Workshop on Computational Models of Reference, Anaphora and Coreference." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.514, + 0.882, + 0.542 + ], + "angle": 0, + "content": "T. Winograd. 1972. Understanding natural language. Cognitive psychology, 3(1):1-191." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.551, + 0.882, + 0.656 + ], + "angle": 0, + "content": "Chong Zhang, Jieyu Zhao, Huan Zhang, Kai-Wei Chang, and Cho-Jui Hsieh. 2021. Double perturbation: On the robustness of robustness and counterfactual bias evaluation. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3899-3916, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.666, + 0.882, + 0.745 + ], + "angle": 0, + "content": "Jieyu Zhao, Daniel Khashabi, Tushar Khot, Ashish Sabharwal, and Kai-Wei Chang. 2021. Ethical-advice taker: Do language models understand natural language interventions? In Annual Meeting of the Association for Computational Linguistics (ACL) - Findings." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.754, + 0.882, + 0.833 + ], + "angle": 0, + "content": "Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. 2018. Gender bias in coreference resolution: Evaluation and debiasing methods. In Conference of the North American Chapter of the Association for Computational Linguistics (NAACL)." + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.086, + 0.882, + 0.833 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1380" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.143, + 0.094, + 0.855, + 0.151 + ], + "angle": 0, + "content": "Appendix The Tail Wagging the Dog: Dataset Construction Biases of Social Bias Benchmarks" + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.212, + 0.384, + 0.244 + ], + "angle": 0, + "content": "A Alternate Constructions of WINOGENDER" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.255, + 0.488, + 0.384 + ], + "angle": 0, + "content": "Addition of clauses: For WINOGENDER, we add clauses like \"who just returned from the beach\" to the different entities in the sentence. For instance, the sentence \"The customer left the bartender a big tip because he was feeling generous.\" becomes \"The customer, who just returned from the beach, left the bartender a big tip because he was feeling generous.\"" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.394, + 0.49, + 0.54 + ], + "angle": 0, + "content": "Synonym substitution: We substitute with synonyms such that it does not change the meaning of the sentence. WINOGENDER has 720 sentences generated from 120 templates, making manual substitution of synonyms in the templates feasible. For example, the sentence \"The supervisor gave the employee feedback on his stellar performance.\" is replaced by \"The supervisor gave the employee feedback on his amazing performance.\"" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.549, + 0.49, + 0.727 + ], + "angle": 0, + "content": "Adding adjectives: As discussed in §3, we add descriptors in the form of adjectives that do not add information about which entity the pronoun or noun would refer to. We do it in four distinct ways, (i) adding the descriptor to the occupation mentioned, e.g. doctor (e.g., \"doctor\" to \"good doctor\"), (ii) adding it to the occupation as a separate clause (e.g., \"doctor\" to \"the doctor who was good\"), (iii) adding the descriptor to the participant mentioned, e.g., \"client\" (similar to (i)), and (iv) adding it to the participant as a separate clause (similar to (ii))." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.739, + 0.471, + 0.755 + ], + "angle": 0, + "content": "B Alternate Constructions of BIASNLI" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.765, + 0.488, + 0.86 + ], + "angle": 0, + "content": "Negation: We negate the verb in each sentence of the dataset. For example, \\( P \\): \"The doctor bought a bagel.\", \\( H \\): \"The man bought a bagel.\", and \\( P \\): \"The doctor did not buy a bagel.\", \\( H \\): \"The man did not buy a bagel.\", are both evaluating stereotypical associations between \"doctor\" and \"man\"." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.871, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Random sampling: The BIASNLI dataset is generated from templates by populating empty slots (e.g. verbs, objects) with words from cor" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.213, + 0.885, + 0.326 + ], + "angle": 0, + "content": "responding lists. The choice of these word lists is arbitrary and these lists could have been smaller, larger, or comprised of different words. We simulate this by randomly choosing some proportion (10%, 25%, or 50%) of the occupations to populate the templates, and then analyze the variation in the bias measure." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.336, + 0.885, + 0.448 + ], + "angle": 0, + "content": "Addition of clauses: We add a clause after the subject in each sentence. For example, for the sentence pair, \\( P \\): \"The doctor bought a coat.\" \\( H \\): \"The man bought a coat.\" is modified to \\( P \\): \"The doctor, who came in the afternoon, bought a coat.\" \\( H \\): \"The man, who came in the afternoon, bought a coat.\"" + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.461, + 0.651, + 0.478 + ], + "angle": 0, + "content": "C Descriptors" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.488, + 0.885, + 0.616 + ], + "angle": 0, + "content": "For WINOGENDER, here is the set of adjectives used to modify either the occupation word or the participant word: aggressive, arrogant, beautiful, brilliant, clean, clever, cruel, deceitful, devious, dirty, dumb, evil, generous, gentle, greedy, hateful, honest, humorless, ignorant, intelligent, intolerant, neat, professional, rude, smart, strong, stupid, terrible, ugly, unclean, unprofessional, weak, wise." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.629, + 0.618, + 0.644 + ], + "angle": 0, + "content": "D Clauses" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.655, + 0.883, + 0.751 + ], + "angle": 0, + "content": "We use the following clauses in WINOGENDER and BIASNLI to increase the distance between relevant parts of the sentence: who just returned from the restaurant, who came in the afternoon, who just came back, who went to the restaurant, who just returned from the beach." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.765, + 0.779, + 0.782 + ], + "angle": 0, + "content": "E Synonymization Examples" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.791, + 0.885, + 0.919 + ], + "angle": 0, + "content": "For WINOGENDER, we manually perform synonymization for all 120 templates. Note that while the replacements might not be exact synonyms, they are replacements of non-identity words that do not change the overall meaning of the sentence and hence should not have any notable impact on the gender bias being measured. We report a few characteristic examples of such substitutions here:" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.519, + 0.941 + ], + "angle": 0, + "content": "1381" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.137, + 0.085, + 0.486, + 0.133 + ], + "angle": 0, + "content": "- The taxpayer met with the accountant to get help filing his taxes \\(\\rightarrow\\) The taxpayer met with the accountant to get aid filing his taxes." + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.145, + 0.486, + 0.207 + ], + "angle": 0, + "content": "- The supervisor gave the employee feedback on his stellar performance \\(\\rightarrow\\) The supervisor gave the employee feedback on his amazing performance." + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.221, + 0.486, + 0.283 + ], + "angle": 0, + "content": "- The hygienist told the patient that he needed to floss every day to avoid gum disease \\(\\rightarrow\\) The hygienist told the patient that he needed to brush every day to avoid cavities." + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.296, + 0.486, + 0.358 + ], + "angle": 0, + "content": "- The broker called the client because he had requested a phone consultation \\(\\rightarrow\\) The broker called the client because he had asked for a phone consultation." + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.372, + 0.486, + 0.451 + ], + "angle": 0, + "content": "- The chef came out to apologize to the guest who was unhappy with his preparation style \\(\\rightarrow\\) The chef came out to apologize to the guest who was dissatisfied with his preparation style." + }, + { + "type": "list", + "bbox": [ + 0.137, + 0.085, + 0.486, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.464, + 0.266, + 0.481 + ], + "angle": 0, + "content": "F Subsampling" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.49, + 0.49, + 0.764 + ], + "angle": 0, + "content": "The gender-occupation subset of the original construction of BIASNLI consists of 164 occupation words such as accountant, firefighter, tutor, and model. In each trial, we subsample some proportion (10%, 25%, or 50%) of these occupation words used in the templates to regenerate the dataset and evaluate all models on this alternate construction. We empirically estimate the distribution of bias scores across samples of a fixed proportion by using 100 independent random trials for that proportion. See Figure 5 for results. Observe that overlap in the distributions serves as a proxy for possible inversions in model ordering (by bias) depending on the subsample of template occupation words used. It is also worth noting that as we use more diverse sets (that is, bigger proportions) of seed words, the variance in the measured bias reduces." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.775, + 0.423, + 0.792 + ], + "angle": 0, + "content": "G Tables of Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.801, + 0.49, + 0.85 + ], + "angle": 0, + "content": "See Table 1 and Table 2 for detailed experimental results on alternate constructions for WINOGEN- DER and BIASNLI respectively." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.862, + 0.345, + 0.879 + ], + "angle": 0, + "content": "H Computing Resources" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.888, + 0.489, + 0.918 + ], + "angle": 0, + "content": "For our experiments, we used a 40-core Intel(R) Xeon(R) CPU E5-2640 v4 @ 2.40GHz, with access" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.164 + ], + "angle": 0, + "content": "to NVIDIA RTX A6000 for selected experiments. In terms of runtime, compute time for inference on a single test set varied by model, but was limited to 12 hours for WINOGENDER and 72 hours for BIASNLI." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.176, + 0.78, + 0.192 + ], + "angle": 0, + "content": "I Links to Datasets and Code" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.201, + 0.884, + 0.233 + ], + "angle": 0, + "content": "All datasets (original constructions) used are publicly available." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.241, + 0.88, + 0.272 + ], + "angle": 0, + "content": "- WINOGENDER:https://github.com/rudinger/ winogender-schemas" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.283, + 0.883, + 0.33 + ], + "angle": 0, + "content": "- BIASNLI: https://github.com/sunipa/On-Measuring-and-Mitigating-Biased-Inferences-of-Word-Embeddings" + }, + { + "type": "list", + "bbox": [ + 0.532, + 0.241, + 0.883, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.339, + 0.834, + 0.355 + ], + "angle": 0, + "content": "All models used are also publicly available." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.363, + 0.881, + 0.392 + ], + "angle": 0, + "content": "ai2spanbert: https://demo.allennlp.org/coreference-resolution" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.404, + 0.881, + 0.435 + ], + "angle": 0, + "content": "- UnifiedQA: https://github.com/allenai/unifiedqa" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.446, + 0.883, + 0.476 + ], + "angle": 0, + "content": "- Longformer: https://github.com/shtoshni/fast-coref" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.487, + 0.883, + 0.518 + ], + "angle": 0, + "content": "- Albert: https://huggingface.co/docs/trans formers/model_doc/albert" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.529, + 0.879, + 0.56 + ], + "angle": 0, + "content": "- Elmo-DA:https://demo.allennlp.org/textual-entailment/elmo-snli" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.572, + 0.856, + 0.634 + ], + "angle": 0, + "content": "- Roberta-base-SNLI:https://github.com/sunipa/OSCaR-Orthogonal-Subspace-Correction-and-Rectification/tree/transformer" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.645, + 0.881, + 0.692 + ], + "angle": 0, + "content": "- Roberta-large-WANLI:https://huggingface.co/alisawuffles/roberta-large-wanli" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.703, + 0.872, + 0.733 + ], + "angle": 0, + "content": "DistilRoberta:https://huggingface.co/cross-encoder/nli-distilroberta-base" + }, + { + "type": "list", + "bbox": [ + 0.532, + 0.363, + 0.883, + 0.733 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.743, + 0.884, + 0.919 + ], + "angle": 0, + "content": "Code and data for the experiments are available at https://github.com/uclanlp/socialbias-dataset-construction-biases. We provide complete preprocessed datasets that correspond to the various proposed alternate constructions. They can be readily used with the publicly listed models for evaluation, thereby easily reproducing the results of the paper. We provide scripts to help with the same. The alternate dataset constructions can also be independently and flexibly used for new experiments." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.928, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1382" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.276, + 0.149, + 0.681, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.375, + 0.681, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.603, + 0.681, + 0.811 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.824, + 0.884, + 0.867 + ], + "angle": 0, + "content": "Figure 5: Bias measures (fraction neutral) computed on BIASNLI. The violin plot attempts to capture the distribution of bias measure scores across datasets reconstructed using different \\(10\\%\\), \\(25\\%\\), and \\(50\\%\\) subsets (top to bottom) of the occupation word list." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1383" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.197, + 0.227, + 0.805, + 0.362 + ], + "angle": 0, + "content": "
Perturbationai2spanbertqa-smallqa-baseqa-largelongformer
Baseline (no perturbations)5.835.8316.6615.419.16
Clause after occupation4.505.5014.7523.5010.08
Clause after participant10.338.0015.0015.758.83
Adjective before occupation8.225.3416.1217.316.87
Adjective after occupation4.925.3715.5725.459.75
Adjective before participant5.975.6913.8418.5210.77
Adjective after participant8.487.4915.9118.1711.69
Synonyms7.927.5017.9215.8312.08
" + }, + { + "type": "table_caption", + "bbox": [ + 0.311, + 0.371, + 0.684, + 0.385 + ], + "angle": 0, + "content": "Table 1: Percentage M-F Mismatch on WINOGENDER." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.683, + 0.885, + 0.746 + ], + "angle": 0, + "content": "
AlbertElmo-DARoberta-base-SNLIRoberta-large-WANLIDistilRoberta
Baseline (no perturbations)44.8141.6415.2516.8151.32
Clauses60.8540.4330.2615.6960.84
Negation45.7613.4020.0410.4562.63
" + }, + { + "type": "table_caption", + "bbox": [ + 0.239, + 0.755, + 0.757, + 0.77 + ], + "angle": 0, + "content": "Table 2: Percentage neutral for different alternate constructions of BIASNLI" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1384" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.133, + 0.085, + 0.435, + 0.1 + ], + "angle": 0, + "content": "ACL 2023 Responsible NLP Checklist" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.108, + 0.323, + 0.123 + ], + "angle": 0, + "content": "A For every submission:" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.127, + 0.533, + 0.145 + ], + "angle": 0, + "content": "A1. Did you describe the limitations of your work?" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.146, + 0.205, + 0.16 + ], + "angle": 0, + "content": "Page 5" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.171, + 0.553, + 0.187 + ], + "angle": 0, + "content": "A2. Did you discuss any potential risks of your work?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.188, + 0.351, + 0.202 + ], + "angle": 0, + "content": "Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.212, + 0.697, + 0.229 + ], + "angle": 0, + "content": "A3. Do the abstract and introduction summarize the paper's main claims?" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.231, + 0.223, + 0.244 + ], + "angle": 0, + "content": "Section 1" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.256, + 0.67, + 0.273 + ], + "angle": 0, + "content": "A4. Have you used AI writing assistants when working on this paper?" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.274, + 0.233, + 0.288 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.3, + 0.49, + 0.317 + ], + "angle": 0, + "content": "B Did you use or create scientific artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.132, + 0.322, + 0.572, + 0.337 + ], + "angle": 0, + "content": "Section 3 and Appendix J (Bias Datasets and Models used)" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.347, + 0.531, + 0.364 + ], + "angle": 0, + "content": "B1. Did you cite the creators of artifacts you used?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.365, + 0.551, + 0.38 + ], + "angle": 0, + "content": "Section 3 and Appendix J (Datasets and Models used)" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.39, + 0.779, + 0.407 + ], + "angle": 0, + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.408, + 0.636, + 0.423 + ], + "angle": 0, + "content": "Appendix J (Datasets and Models used are all publicly available)" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.434, + 0.882, + 0.497 + ], + "angle": 0, + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.499, + 0.351, + 0.514 + ], + "angle": 0, + "content": "Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.525, + 0.882, + 0.573 + ], + "angle": 0, + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.574, + 0.351, + 0.589 + ], + "angle": 0, + "content": "Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.6, + 0.882, + 0.632 + ], + "angle": 0, + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.633, + 0.351, + 0.648 + ], + "angle": 0, + "content": "Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.658, + 0.882, + 0.74 + ], + "angle": 0, + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be." + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.741, + 0.36, + 0.755 + ], + "angle": 0, + "content": "Section 3.2 and Appendix F" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.765, + 0.494, + 0.782 + ], + "angle": 0, + "content": "C Did you run computational experiments?" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.788, + 0.206, + 0.801 + ], + "angle": 0, + "content": "Section 3" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.812, + 0.882, + 0.846 + ], + "angle": 0, + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.847, + 0.235, + 0.861 + ], + "angle": 0, + "content": "Appendix I" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.869, + 0.878, + 0.893 + ], + "angle": 0, + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1385" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.13, + 0.084, + 0.88, + 0.133 + ], + "angle": 0, + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Section 3, Appendix B-G" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.143, + 0.88, + 0.208 + ], + "angle": 0, + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Section 3, Appendix H" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.219, + 0.88, + 0.283 + ], + "angle": 0, + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Not applicable. Left blank." + }, + { + "type": "list", + "bbox": [ + 0.129, + 0.084, + 0.88, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.294, + 0.878, + 0.331 + ], + "angle": 0, + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.341, + 0.88, + 0.388 + ], + "angle": 0, + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.4, + 0.88, + 0.464 + ], + "angle": 0, + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.476, + 0.88, + 0.54 + ], + "angle": 0, + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.55, + 0.873, + 0.582 + ], + "angle": 0, + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.593, + 0.88, + 0.641 + ], + "angle": 0, + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response." + }, + { + "type": "list", + "bbox": [ + 0.128, + 0.341, + 0.88, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1386" + } + ] +] \ No newline at end of file diff --git a/2023/The Tail Wagging the Dog_ Dataset Construction Biases of Social Bias Benchmarks/fca46b6f-67a0-44f8-a5a2-d1a3682b3c41_origin.pdf b/2023/The Tail Wagging the Dog_ Dataset Construction Biases of Social Bias Benchmarks/fca46b6f-67a0-44f8-a5a2-d1a3682b3c41_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c41ced05e815044e831eaceec3a01af34b604633 --- /dev/null +++ b/2023/The Tail Wagging the Dog_ Dataset Construction Biases of Social Bias Benchmarks/fca46b6f-67a0-44f8-a5a2-d1a3682b3c41_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:70c5549fca63318303dff3937b17975f3b4d873a6c24aa8d0f037e078336fadc +size 548593 diff --git a/2023/The Tail Wagging the Dog_ Dataset Construction Biases of Social Bias Benchmarks/full.md b/2023/The Tail Wagging the Dog_ Dataset Construction Biases of Social Bias Benchmarks/full.md new file mode 100644 index 0000000000000000000000000000000000000000..bd7aa29cbde5b3dd342c2bdeef8526d6ef1510d0 --- /dev/null +++ b/2023/The Tail Wagging the Dog_ Dataset Construction Biases of Social Bias Benchmarks/full.md @@ -0,0 +1,360 @@ +# The Tail Wagging the Dog: Dataset Construction Biases of Social Bias Benchmarks + +Nikil Roashan Selvam $^{1}$ Sunipa Dev $^{2}$ +Daniel Khashabi $^{3}$ Tushar Khot $^{4}$ Kai-Wei Chang $^{1}$ $^{1}$ University of California, Los Angeles $^{2}$ Google Research + $^{3}$ Johns Hopkins University $^{4}$ Allen Institute for AI +ikilrselvam,kwchang}@ucla.edu, sunipadev@google. +danielk@jhu.edu, tushark@allenai.org + +# Abstract + +How reliably can we trust the scores obtained from social bias benchmarks as faithful indicators of problematic social biases in a given model? In this work, we study this question by contrasting social biases with non-social biases that stem from choices made during dataset construction (which might not even be discernible to the human eye). To do so, we empirically simulate various alternative constructions for a given benchmark based on seemingly innocuous modifications (such as paraphrasing or random-sampling) that maintain the essence of their social bias. On two well-known social bias benchmarks (WINOGENDER and BIASNLI), we observe that these shallow modifications have a surprising effect on the resulting degree of bias across various models and consequently the relative ordering of these models when ranked by measured bias. We hope these troubling observations motivate more robust measures of social biases. + +# 1 Introduction + +The omnipresence of large pre-trained language models (Liu et al., 2019; Raffel et al., 2020; Brown et al., 2020) has fueled concerns regarding their systematic biases carried over from underlying data into the applications they are used in, resulting in disparate treatment of people with different identities (Sheng et al., 2021; Abid et al., 2021). + +In response to such concerns, various benchmarks have been proposed to quantify the amount of social biases in models (Rudinger et al., 2018; Sheng et al., 2019; Li et al., 2020). These measures are composed of textual datasets built for a specific NLP task (such as question answering) and are accompanied by a metric such as accuracy of prediction which is used as an approximation of the amount of social biases. + +These bias benchmarks are commonly used by machine learning practitioners to compare the degree of social biases (such as gender-occupation + +![](images/fd80f5c28cd735e5b85fc16d14449f902723c21e5d745137ebb90ce455b44cc3.jpg) +Figure 1: Two potential constructions of WINOGEN- DER with minor differences: a model (span-BERT, in this case) with the original dataset might seem to have gender-occupation bias (green tick) based on the change in its pronoun resolution. However, a minor change in its phrasing with no change in meaning (e.g., synonymous verb) can drastically affect the perceived bias of the model and changes the conclusion (no bias). + +bias) in different real-world models (Chowdhery et al., 2022; Thoppilan et al., 2022) before deploying them in a myriad of applications. However, they also inadvertently measure other non-social biases in their datasets. For example, consider the sentence from WINOGENDER in Figure 1. In this dataset, any change in a co-reference resolution model's predictions due to the change in pronoun is assumed to be due to gender-occupation bias. However, this assumption only holds for a model with near-perfect language understanding with no other biases. This may not often be the case, e.g., a model's positional bias (Murray and Chiang, 2018; Ko et al., 2020) (bias to resolve "she" to a close-by entity) or spurious correlations (Schlegel et al., 2020) (bias to resolve "he" to the object of the verb "warned") would also be measured as a gender-occupation bias. As a result, a slightly different template (e.g., changing the verb to "cautioned") + +could result in completely different bias measurements. + +The goal of this work is to illustrate the extent to which social bias measurements are effected by assumptions that are built into dataset constructions. To that end, we consider several alternate dataset constructions for 2 bias benchmarks WINOGENDER and BIASNLI. We show that, just by the choice of certain target-bias-irrelevant elements in a dataset, it is possible to discover different degrees of bias for the same model as well as different model rankings1. For instance, one experiment on BIASNLI demonstrated that merely negating verbs drastically reduced the measured bias $(41.64\rightarrow 13.40)$ on an ELMo-based Decomposable Attention model and even caused a switch in the comparative ranking with RoBERTa. Our findings demonstrate the unreliability of current benchmarks to truly measure social bias in models and suggest caution when considering these measures as the gold truth. We provide a detailed discussion (\$5) of the implications of our findings, relation to experienced harms, suggestions for improving bias benchmarks, and directions for future work. + +# 2 Related Work + +A large body of work investigates ways to evaluate biases carried inherently in language models (Bolukbasi et al., 2016; Caliskan et al., 2017; Nadeem et al., 2021) and expressed in specific tasks (Nangia et al., 2020; Kirk et al., 2021; Schramowski et al., 2022; Prabhumoye et al., 2021; Srinivasan and Bisk, 2021; Kirk et al., 2021; Parrish et al., 2021; Baldini et al., 2022; Czarnowska et al., 2021; Dev et al., 2021a; Zhao et al., 2021). Alongside, there is also growing concern about the measures not relating to experienced harms (Blodgett et al., 2020), not inclusive in framing (Dev et al., 2021b), ambiguous about what bias is measured (Blodgett et al., 2021), not correlated in their findings of bias across intrinsic versus extrinsic techniques (Goldfarb-Tarrant et al., 2021; Cao et al., 2022), and susceptible to adversarial perturbations (Zhang et al., 2021) and seed word selection (Antoniak and Mimno, 2021). + +The concurrent work by (Seshadri et al., 2022) discusses the unreliability of quantifying social biases using templates by varying templates in a se + +mantic preserving manner. While their findings are consistent with ours, the two works provide complementary experimental observations. Seshadri et al. (2022) study a wider range of tasks, though we focus our experiments on a wider set of models and alternate dataset constructions (with a greater range of syntactic and semantic variability). As a result, we are able to illustrate the effect of the observed variability on ranking large language models according to measured bias for deployment in real world applications. + +# 3 Social Bias Measurements and Alternate Constructions + +Bias measures in NLP are often quantified through comparative prediction disparities on language datasets that follow existing tasks such as classification (De-Arteaga et al., 2019) or coreference resolution (Rudinger et al., 2018). As a result, these datasets are central to what eventually gets measured as “bias”. Not only do they determine the “amount” of bias measured but also the “type” of bias or stereotype measured. Datasets often vary combinations of gendered pronouns and occupations to evaluate stereotypical associations. It is important to note that these constructs of datasets and their templates, which determine what gets measured, are often arbitrary choices. The sentences could be differently structured, be generated from a different set of seed words, and more. However, we expect that for any faithful bias benchmark, such dataset alterations that are not relevant to social bias should not have a significant impact on the artifact (e.g. gender bias) being measured. + +Thus, to evaluate the faithfulness of current benchmarks, we develop alternate dataset constructions through modifications that should not have any effect on the social bias being measured in a dataset. They are minor changes that should not influence models with true language understanding – the implicit assumption made by current bias benchmarks. Any notable observed changes in a model's bias measure due to these modifications would highlight the incorrectness of this assumption. Consequently, this would bring to light the unreliability of current benchmarks to faithfully measure the target bias and disentangle the measurement from measurement of other non-social biases. A non-exhaustive set of such alternate constructions considered in this work are listed below. + +# Clause after occupation + +The engineer, who just returned from the beach, informed the client that he would need to make all future payments on time. + +# Clause after participant + +The engineer informed the client, who just returned from the beach, that he would need to make all future payments on time. + +# Synonymization + +The engineer informed the client that he would need to make all upcoming payments on time. + +# Adjective before occupation + +The cruel engineer informed the client that he would need to make all future payments on time. + +# Adjective after occupation + +The engineer, who was cruel, informed the client that he would need to make all future payments on time. + +# Adjective before participant + +The engineer informed the wise client that he would need to make all future payments on time. + +# Adjective after participant + +The engineer informed the client, who was wise, that he would need to make all future payments on time. + +Figure 2: An instance ("The engineer informed the client that he would need to make all future payments on time") from WINOGENDER benchmark modified under various shallow modifications (§3). To a human eye, such modifications do not necessarily affect the outcome of the given pronoun resolution problem. + +Negations: A basic function in language understanding is to understand the negations of word groups such as action verbs, or adjectives. Altering verbs in particular, such as 'the doctor bought' to 'the doctor did not buy' should typically not affect the inferences made about occupation associations. + +Synonym substitutions: Another fundamental function of language understanding is the ability to parse the usage of similar words or synonyms used in identical contexts, to derive the same overall meaning of a sentence. For bias measuring datasets, synonymizing non-pivotal words (such as non-identity words like verbs) should not change the outcome of how much bias is measured. + +Varying length of the text: In typical evaluation datasets, the number of clauses that each sentence is composed of and overall the sentence length are arbitrary experimental choices. Fixing this length is common, especially when such datasets need to be created at scale. If language is understood, adding a neutral phrase without impacting the task-specific semantics should not alter the bias measured. + +Adding descriptors: Sentences used in real life are structured in complex ways and can have descriptors, such as adjectives about an action, person, or object, without changing the net message expressed by the text. For example, the sentences, "The doctor bought an apple.", and "The doctor bought a red apple." do not change any assumptions made about the doctor, or the action of buying an apple. + +Random samples: Since the sentence constructs of these datasets are not unique, a very simple alternate construction of a dataset is a different subsample of itself. This is because the dataset is scraped or generated with specific assumptions or parameters, such as seed word lists, templates of sentences, and word order. However, neither the sentence constructs or templates, nor the seed word + +lists typically used are exhaustive or representative of entire categories of words (such as gendered words, emotions, and occupations). + +See Fig. 2 for example constructions on WINO-GENDER (App. A, B for detailed descriptions). + +# 4 Case Studies + +We discuss here the impact of alternate constructions on two task-based measures of bias.2 + +# 4.1 Coreference Resolution + +Several different bias measures (Rudinger et al., 2018; Zhao et al., 2018; Cao and Daumé III, 2021) for coreference resolution work similar to Winograd Schema (Winograd, 1972) where a sentence has two entities and the task is to resolve which entity a specific pronoun or noun refers to. We work here with WINOGENDER (Rudinger et al., 2018), popularly used to measure biases. It is worth noting that WINOGENDER was originally intended by its authors to merely be a diagnostic tool that checks for bias in a model; the authors note that it may demonstrate the presence of model bias but not prove the absence of the same. Nonetheless, models developed today are indeed tested and compared for social bias on WinoGender, leading to its usage as a comparative standard or benchmark (Chowdhery et al., 2022; Thoppilan et al., 2022). + +The metric used to evaluate bias is the percentage of sentence pairs where there is a mismatch in predictions for the male and female gendered pronouns. For instance, in Fig. 2, if the pronoun "he" is linked to "engineer" but switches to "client" for the pronoun "she", that would indicate a gender-occupation bias. Higher the number of mismatches, + +![](images/1adedff8ac62a30012d7f23cb921c7b276c9ba994af299070642fc858dc7b720.jpg) +(a) WINOGENDER +Figure 3: Bias measures on (a) WINOGENDER (percentage M-F mismatch, log-scale) and (b) BIASNLI (accuracy as percentage neutral, log-scale), across a variety of dataset constructions and models. + +![](images/ff012ff95eee03763f79b1e41dd49741182d9866115ab0a9dc002da8a5978b26.jpg) +(b) BIASNLI + +higher the bias. In particular, note that the metric does not take into account the accuracy of the predictions, but rather only the mismatch between the two pronouns. + +We experiment with three alternate constructions of the dataset: addition of clauses, addition of adjectives, and synonymizing words in templates. Each alternate construction is introduced so as to not affect the overall meaning of the sentence. + +Experimental Results: We use an end-to-end coreference model with SpanBERT embeddings (Lee et al., 2018; Joshi et al., 2020), UnifiedQA (small, base, and large) (Khashabi et al., 2020) QA model, $^{3}$ and a long-document coreference model with Longformer encodings (Toshniwal et al., 2021). Results of evaluating these models on various WINOGENDER constructions is summarized in Fig. 3a. Small changes to the formulation of dataset templates result in sizable changes to computed bias measures compared to the published baseline constructions. For example, a construction involving added adjectives after occupations would have found the UnifiedQA (large) model to have $10\%$ less bias compared to the default constructions. The sensitivity to the dataset constructions can have a drastic effect on ranking models according to their social bias, as Fig. 3a shows. For example, the SpanBERT model is considered to have less bias than UnifiedQA (small) model in the baseline dataset, but would be considered to be more biased if the templates had clauses after the participants or adjectives before the occupation. + +![](images/78f45b52836cdfaa2413129c98c9554d6798d012c243a13db12719bec25feb66.jpg) +Figure 4: Bias measures (fraction neutral) computed on BIASNLI. The violin plot represents distribution of bias measure scores across datasets reconstructed using different $10\%$ subsets of the occupation word list across 100 random samples. + +# 4.2 Natural Language Inference + +Natural Language Inference (NLI) is the task of determining directional relationships between two sentences (a premise $(P)$ and a hypothesis $(H)$ ). Dev et al. (2020)'s measure based on NLI (BIASNLI) evaluates if stereotypical inferences are made by language models. We use their dataset for gender-occupation stereotypes containing approximately 2 million sentence pairs such as $P$ : "The doctor bought a bagel.", $H$ : "The man bought a bagel." The expected prediction for each sentence pair in the dataset is neutral, and therefore the bias metric used is the fraction of neutral inferences on dataset - the higher the score, the lower the bias. + +We experiment with three alternate constructions of the dataset: verb negation, random sampling, + +and addition of clauses. Note that the alternate constructions do not impact the unbiased label (neutral). Any change in construction (say negating a verb) is applied to both the premise and hypothesis. Refer to App. B for a detailed description. + +Experimental Results: We use RoBERTa trained on SNLI (RoBERTa-base-SNLI) (Liu et al., 2019), ELMo-based Decomposable Attention (ELMoDA) (Parikh et al., 2016), ALBERT (Lan et al., 2019), distilled version of the RoBERTa-base model (Sanh et al., 2019), and RoBERTa-large fin-tuned on WANLI (Liu et al., 2022). The bias measured with each model using BIASNLI is recorded in Fig. 3b. The results show how small modifications to the dataset again result in large changes to the bias measured, and also change the bias rankings. For example, adding a negation largely reduces the bias measured $(\triangle = 28.24)$ for ELMoDA, and also results in a switch in the comparative ranking to RoBERTa-base-SNLI. Furthermore, as seen in Fig. 4, there is a significant overlap in the bias measures of ALBERT, DistilRoBERTa, and ELMo-DA under random sampling, which corresponds to high variability in relative model ordering across different sub-samples of the dataset. + +# 5 Discussion and Conclusion + +Social bias measurements are very sensitive to evaluation methodology. Our empirical evidence sheds light on how the model's non-social biases brought out or masked by alternate constructions can cause bias benchmarks to underestimate or overestimate the social bias in a model. More interestingly, it is important to note that different models respond differently to perturbations. In fact, the same perturbation can result in a higher or lower measured bias depending on the model (as seen in §4.1 and §4.2), which points to how models might parse information (and thus bias) differently. + +While current bias measures do play a role in exposing where model errors have a stereotypical connotation, a lack of sentence construction variability or even assumptions made when creating seed word lists can reduce the reliability of the benchmarks, as we see in this work (§4.2). Even with simple sentences, it is not apparent how to disentangle the biased association of the identity with the verb or the occupation amongst others. This is especially important to note as it highlights that measures can lack concrete definitions of what bi + +ased associations they measure. Consequently, the relation between measured bias and experienced harm becomes unclear. + +We hope that our troubling observations motivate future work that thoroughly investigates how to construct robust benchmarks that faithfully measure the target bias without being affected by model errors and other non-social biases. As suggested by our subsampling experiments (Appendix F), it might be fruitful to encourage both syntactic and semantic diversity in these benchmarks. Bias benchmarks that provide uncertainty measures (instead of a single number) might enable practitioners to better compare models before deploying them. Furthermore, since the opaqueness of large language models makes it challenging to understand how and to what extent a linguistic change will affect the measured bias, explainable models might indeed facilitate better measurement of their social bias. Assuming that we can generate faithful explanations for a model's predictions, an exciting future direction is to explore construction of bias benchmarks which operate on the explanations of the predictions rather than the predictions themselves. Lastly, we also encourage discussions on the complexity of the sentences used in benchmarks and their implications on what gets measured in relation to un-templated, naturally-occurring text (Levy et al., 2021), as an attempt to ground our measurements in experienced harms. + +# Limitations + +We acknowledge the underlying assumptions of the social bias benchmarks used in our study. While the presented study aims to point out a key limitation of currently accepted methodologies, the presented investigation could benefit from more diversification. First, this study focuses on English. While we expect similar issues with similarly-constructed benchmarks in other languages, we leave it to future work to formally address the same. Also, the bias benchmarks themselves imbibe the notion of fairness with the Western value system (Bhatt et al., 2022), and future explorations of benchmarks should diversify culturally as well. Last but not least, we acknowledge the harm of binary treatment of genders in one of the target benchmarks. The purpose of this work was to bring light to a broader problem regarding the reliability of social benchmark metrics, with the hypothesis that the main idea of this paper would hold for a wider + +range of datasets with other assumptions or notions of fairness. We also acknowledge that there are larger models that we were not able to train and evaluate due to the limitations on our computational budget. The current study was focused on benchmarks with templated instances. This is no coincidence: the dominant majority of the social bias benchmarking literature relies on sentences with some degree of known structure, even in those collected from the wild (Levy et al., 2021). Such structural assumptions in datasets are necessary for defining and extracting quantifiable measures of social bias, which as we argue, are the reason behind the brittleness of their decisions. Future work should focus on making our bias benchmarks more diverse and robust to small decisions that go into making them. + +# Broader Impact + +Bias evaluating benchmarks play a very significant role in helping identify potential risks of language technologies. While a large body of work evolves in this area of work, there is growing concern about the ability of the different benchmarks to accurately quantify and identify social biases. We emphasize these concerns by evaluating how robust the benchmarks are to alternate constructions based on simple linguistic properties. It is important to note how inaccurate measurements of social biases can be problematic by underestimating or misdiagnosing the potential harm from language models. We hope our work helps identify such pitfalls. + +# Acknowledgements + +We thank the students and colleagues at UCLA, JHU and AI2 for their insightful feedback towards improving this paper. The authors would also like to thank the anonymous reviewers for their constructive feedback. This project is supported by generous gifts from Allen Institute for AI, CISCO, Amazon, and a Sloan fellowship. + +# References + +Abubakar Abid, Maheen Farooqi, and James Zou. 2021. Persistent anti-muslim bias in large language models. In AAAI/ACM Conference on AI, Ethics, and Society (AIES), pages 298-306. +Maria Antoniak and David Mimno. 2021. Bad seeds: Evaluating lexical methods for bias measurement. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the + +11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1889-1904, Online. Association for Computational Linguistics. +Ioana Baldini, Dennis Wei, Karthikeyan Natesan Ramamurthy, Moninder Singh, and Mikhail Yurochkin. 2022. Your fairness may vary: Pretrained language model fairness in toxic text classification. In Annual Meeting of the Association for Computational Linguistics (ACL) - Findings. +Shaily Bhatt, Sunipa Dev, Partha Talukdar, Shachi Dave, and Vinodkumar Prabhakaran. 2022. Recontextualizing fairness in NLP: The case of India. In Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 727-740, Online only. Association for Computational Linguistics. +Su Lin Blodgett, Solon Barocas, Hal Daumé III, and Hanna Wallach. 2020. Language (technology) is power: A critical survey of "bias" in nlp. In Annual Meeting of the Association for Computational Linguistics (ACL). +Su Lin Blodgett, Gilsinia Lopez, Alexandra Olteanu, Robert Sim, and Hanna Wallach. 2021. Stereotyping norwegian salmon: an inventory of pitfalls in fairness benchmark datasets. In Annual Meeting of the Association for Computational Linguistics (ACL). +Tolga Bolukbasi, Kai-Wei Chang, James Y Zou, Venkatesh Saligram, and Adam T Kalai. 2016. Man is to computer programmer as woman is to homemaker? debiasing word embeddings. In Advances in Neural Information Processing Systems, volume 29. Curran Associates, Inc. +Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, and et al. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems (NeurIPS). +Aylin Caliskan, Joanna J. Bryson, and Arvind Narayanan. 2017. Semantics derived automatically from language corpora contain human-like biases. Science, 356(6334):183-186. +Yang Trista Cao and Hal Daumé III. 2021. Toward gender-inclusive coreference resolution: An analysis of gender and bias throughout the machine learning lifecycle. Computational Linguistics (CL). +Yang Trista Cao, Yada Pruksachatkun, Kai-Wei Chang, Rahul Gupta, Varun Kumar, Jwala Dhamala, and Aram Galstyan. 2022. On the intrinsic and extrinsic fairness evaluation metrics for contextualized language representations. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 561-570, + +Dublin, Ireland. Association for Computational Linguistics. +Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam M. Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Benton C. Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan Firat, Michele Catasta, Jason Wei, Kathleen S. Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. 2022. Palm: Scaling language modeling with pathways. ArXiv, abs/2204.02311. +Paula Czarnowska, Yogarshi Vyas, and Kashif Shah. 2021. Quantifying social biases in nlp: A generalization and empirical comparison of extrinsic fairness metrics. Transactions of the Association for Computational Linguistics (TACL). +Maria De-Arteaga, Alexey Romanov, Hanna Wallach, Jennifer Chayes, Christian Borgs, Alexandra Chouldechova, Sahin Geyik, Krishnamaram Kenthapadi, and Adam Kalai. 2019. Bias in bios: A case study of semantic representation bias in a high-stakes setting. In ACM Conference on Fairness, Accountability and Transparency (FAccT). +Sunipa Dev, Tao Li, Jeff M. Phillips, and Vivek Srikumar. 2020. On measuring and mitigating biased inferences of word embeddings. Conference on Artificial Intelligence (AAAI). +Sunipa Dev, Tao Li, Jeff M Phillips, and Vivek Srikumar. 2021a. Oscar: Orthogonal subspace correction and rectification of biases in word embeddings. In _Conference on Empirical Methods in Natural Language Processing (EMNLP)\. +Sunipa Dev, Masoud Monajatipoor, Anaelia Ovalle, Arjun Subramonian, Jeff Phillips, and Kai-Wei Chang. 2021b. Harms of gender exclusivity and challenges in non-binary representation in language technologies. In Conference on Empirical Methods in Natural Language Processing (EMNLP). +Seraphina Goldfarb-Tarrant, Rebecca Marchant, Ricardo Muñoz Sánchez, Mugdha Pandya, and Adam Lopez. 2021. Intrinsic bias metrics do not correlate with application bias. In Annual Meeting of the Association for Computational Linguistics (ACL). + +Mandar Joshi, Danqi Chen, Yinhan Liu, Daniel S. Weld, Luke Zettlemoyer, and Omer Levy. 2020. SpanBERT: Improving pre-training by representing and predicting spans. Transactions of the Association for Computational Linguistics (TACL). +Daniel Khashabi, Sewon Min, Tushar Khot, Ashish Sabharwal, Oyvind Tafjord, Peter Clark, and Hannaneh Hajishirzi. 2020. UnifiedQA: Crossing Format Boundaries With a Single QA System. In Conference on Empirical Methods in Natural Language Processing (EMNLP) - Findings. +Hannah Rose Kirk, Filippo Volpin, Haider Iqbal, Elias Benussi, Frederic Dreyer, Aleksandar Shtedritski, Yuki Asano, et al. 2021. Bias out-of-the-box: An empirical analysis of intersectional occupational biases in popular generative language models. Advances in Neural Information Processing Systems (NeurIPS). +Miyoung Ko, Jinhyuk Lee, Hyunjae Kim, Gangwoo Kim, and Jaewoo Kang. 2020. Look at the first sentence: Position bias in question answering. In _Conference on Empirical Methods in Natural Language Processing_ (EMNLP). +Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. 2019. Albert: A lite bert for self-supervised learning of language representations. In International Conference on Learning Representations (ICLR). +Kenton Lee, Luheng He, and Luke Zettlemoyer. 2018. Higher-order coreference resolution with coarse-to-fine inference. In Conference of the North American Chapter of the Association for Computational Linguistics (NAACL). +Shahar Levy, Koren Lazar, and Gabriel Stanovsky. 2021. Collecting a large-scale gender bias dataset for coreference resolution and machine translation. In *Conference on Empirical Methods in Natural Language Processing* (EMNLP) - Findings. +Tao Li, Daniel Khashabi, Tushar Khot, Ashish Sabharwal, and Vivek Srikumar. 2020. UnCovering Stereotypical Biases via Underspecified Questions. In Conference on Empirical Methods in Natural Language Processing (EMNLP) - Findings. +Alisa Liu, Swabha Swayamdipta, Noah A Smith, and Yejin Choi. 2022. WANLI: Worker and AI Collaboration for Natural Language Inference Dataset Creation. arXiv preprint arXiv:2201.05955. +Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692. +Kenton Murray and David Chiang. 2018. Correcting length bias in neural machine translation. In Conference on Machine Translation (WMT). + +Moin Nadeem, Anna Bethke, and Siva Reddy. 2021. StereoSet: Measuring stereotypical bias in pretrained language models. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 5356-5371, Online. Association for Computational Linguistics. +Nikita Nangia, Clara Vania, Rasika Bhalerao, and Samuel R Bowman. 2020. Crows-pairs: A challenge dataset for measuring social biases in masked language models. In Conference on Empirical Methods in Natural Language Processing (EMNLP). +Ankur P. Parikh, Oscar Tackström, Dipanjan Das, and Jakob Uszkoreit. 2016. A decomposable attention model for natural language inference. In *Conference on Empirical Methods in Natural Language Processing* (EMNLP). +Alicia Parrish, Angelica Chen, Nikita Nangia, Vishakh Padmakumar, Jason Phang, Jana Thompson, Phu Mon Htut, and Samuel R Bowman. 2021. BBQ: A hand-built bias benchmark for question answering. In Annual Meeting of the Association for Computational Linguistics (ACL). +Shrimai Prabhumoye, Rafal Kocielnik, Mohammad Shoeybi, Anima Anandkumar, and Bryan Catanzaro. 2021. Few-shot instruction prompts for pretrained language models to detect social biases. arXiv preprint arXiv:2112.07868. +Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research (JMLR). +Rachel Rudinger, Jason Naradowsky, Brian Leonard, and Benjamin Van Durme. 2018. Gender bias in coreference resolution. In Conference of the North American Chapter of the Association for Computational Linguistics (NAACL). +Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. 2019. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. ArXiv, abs/1910.01108. +Viktor Schlegel, Goran Nenadic, and Riza Batista-Navarro. 2020. Beyond leaderboards: A survey of methods for revealing weaknesses in natural language inference data and models. arXiv preprint arXiv:2005.14709. +Patrick Schramowski, Cigdem Turan, Nico Andersen, Constantin A Rothkopf, and Kristian Kersting. 2022. Large pre-trained language models contain human-like biases of what is right and wrong to do. Nature Machine Intelligence. + +Preethi Seshadri, Pouya Pezeshkpour, and Sameer Singh. 2022. Quantifying social biases using templates is unreliable. arXiv preprint arXiv:2210.04337. +Emily Sheng, Kai-Wei Chang, Prem Natarajan, and Nanyun Peng. 2019. The Woman Worked as a Babysitter: On Biases in Language Generation. In Conference on Empirical Methods in Natural Language Processing (EMNLP). +Emily Sheng, Kai-Wei Chang, Premkumar Natarajan, and Nanyun Peng. 2021. Societal biases in language generation: Progress and challenges. In *Conference on Empirical Methods in Natural Language Processing* (EMNLP). +Tejas Srinivasan and Yonatan Bisk. 2021. Worst of both worlds: Biases compound in pre-trained vision-and-language models. In Workshop on Gender Bias in Natural Language Processing. +Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, et al. 2022. LaMDA: Language Models for Dialog Applications. arXiv preprint arXiv:2201.08239. +Shubham Toshniwal, Patrick Xia, Sam Wiseman, Karen Livescu, and Kevin Gimpel. 2021. On generalization in coreference resolution. In Proceedings of the Workshop on Computational Models of Reference, Anaphora and Coreference. +T. Winograd. 1972. Understanding natural language. Cognitive psychology, 3(1):1-191. +Chong Zhang, Jieyu Zhao, Huan Zhang, Kai-Wei Chang, and Cho-Jui Hsieh. 2021. Double perturbation: On the robustness of robustness and counterfactual bias evaluation. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3899-3916, Online. Association for Computational Linguistics. +Jieyu Zhao, Daniel Khashabi, Tushar Khot, Ashish Sabharwal, and Kai-Wei Chang. 2021. Ethical-advice taker: Do language models understand natural language interventions? In Annual Meeting of the Association for Computational Linguistics (ACL) - Findings. +Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. 2018. Gender bias in coreference resolution: Evaluation and debiasing methods. In Conference of the North American Chapter of the Association for Computational Linguistics (NAACL). + +# Appendix The Tail Wagging the Dog: Dataset Construction Biases of Social Bias Benchmarks + +# A Alternate Constructions of WINOGENDER + +Addition of clauses: For WINOGENDER, we add clauses like "who just returned from the beach" to the different entities in the sentence. For instance, the sentence "The customer left the bartender a big tip because he was feeling generous." becomes "The customer, who just returned from the beach, left the bartender a big tip because he was feeling generous." + +Synonym substitution: We substitute with synonyms such that it does not change the meaning of the sentence. WINOGENDER has 720 sentences generated from 120 templates, making manual substitution of synonyms in the templates feasible. For example, the sentence "The supervisor gave the employee feedback on his stellar performance." is replaced by "The supervisor gave the employee feedback on his amazing performance." + +Adding adjectives: As discussed in §3, we add descriptors in the form of adjectives that do not add information about which entity the pronoun or noun would refer to. We do it in four distinct ways, (i) adding the descriptor to the occupation mentioned, e.g. doctor (e.g., "doctor" to "good doctor"), (ii) adding it to the occupation as a separate clause (e.g., "doctor" to "the doctor who was good"), (iii) adding the descriptor to the participant mentioned, e.g., "client" (similar to (i)), and (iv) adding it to the participant as a separate clause (similar to (ii)). + +# B Alternate Constructions of BIASNLI + +Negation: We negate the verb in each sentence of the dataset. For example, $P$ : "The doctor bought a bagel.", $H$ : "The man bought a bagel.", and $P$ : "The doctor did not buy a bagel.", $H$ : "The man did not buy a bagel.", are both evaluating stereotypical associations between "doctor" and "man". + +Random sampling: The BIASNLI dataset is generated from templates by populating empty slots (e.g. verbs, objects) with words from cor + +responding lists. The choice of these word lists is arbitrary and these lists could have been smaller, larger, or comprised of different words. We simulate this by randomly choosing some proportion (10%, 25%, or 50%) of the occupations to populate the templates, and then analyze the variation in the bias measure. + +Addition of clauses: We add a clause after the subject in each sentence. For example, for the sentence pair, $P$ : "The doctor bought a coat." $H$ : "The man bought a coat." is modified to $P$ : "The doctor, who came in the afternoon, bought a coat." $H$ : "The man, who came in the afternoon, bought a coat." + +# C Descriptors + +For WINOGENDER, here is the set of adjectives used to modify either the occupation word or the participant word: aggressive, arrogant, beautiful, brilliant, clean, clever, cruel, deceitful, devious, dirty, dumb, evil, generous, gentle, greedy, hateful, honest, humorless, ignorant, intelligent, intolerant, neat, professional, rude, smart, strong, stupid, terrible, ugly, unclean, unprofessional, weak, wise. + +# D Clauses + +We use the following clauses in WINOGENDER and BIASNLI to increase the distance between relevant parts of the sentence: who just returned from the restaurant, who came in the afternoon, who just came back, who went to the restaurant, who just returned from the beach. + +# E Synonymization Examples + +For WINOGENDER, we manually perform synonymization for all 120 templates. Note that while the replacements might not be exact synonyms, they are replacements of non-identity words that do not change the overall meaning of the sentence and hence should not have any notable impact on the gender bias being measured. We report a few characteristic examples of such substitutions here: + +- The taxpayer met with the accountant to get help filing his taxes $\rightarrow$ The taxpayer met with the accountant to get aid filing his taxes. +- The supervisor gave the employee feedback on his stellar performance $\rightarrow$ The supervisor gave the employee feedback on his amazing performance. +- The hygienist told the patient that he needed to floss every day to avoid gum disease $\rightarrow$ The hygienist told the patient that he needed to brush every day to avoid cavities. +- The broker called the client because he had requested a phone consultation $\rightarrow$ The broker called the client because he had asked for a phone consultation. +- The chef came out to apologize to the guest who was unhappy with his preparation style $\rightarrow$ The chef came out to apologize to the guest who was dissatisfied with his preparation style. + +# F Subsampling + +The gender-occupation subset of the original construction of BIASNLI consists of 164 occupation words such as accountant, firefighter, tutor, and model. In each trial, we subsample some proportion (10%, 25%, or 50%) of these occupation words used in the templates to regenerate the dataset and evaluate all models on this alternate construction. We empirically estimate the distribution of bias scores across samples of a fixed proportion by using 100 independent random trials for that proportion. See Figure 5 for results. Observe that overlap in the distributions serves as a proxy for possible inversions in model ordering (by bias) depending on the subsample of template occupation words used. It is also worth noting that as we use more diverse sets (that is, bigger proportions) of seed words, the variance in the measured bias reduces. + +# G Tables of Experimental Results + +See Table 1 and Table 2 for detailed experimental results on alternate constructions for WINOGEN- DER and BIASNLI respectively. + +# H Computing Resources + +For our experiments, we used a 40-core Intel(R) Xeon(R) CPU E5-2640 v4 @ 2.40GHz, with access + +to NVIDIA RTX A6000 for selected experiments. In terms of runtime, compute time for inference on a single test set varied by model, but was limited to 12 hours for WINOGENDER and 72 hours for BIASNLI. + +# I Links to Datasets and Code + +All datasets (original constructions) used are publicly available. + +- WINOGENDER:https://github.com/rudinger/ winogender-schemas +- BIASNLI: https://github.com/sunipa/On-Measuring-and-Mitigating-Biased-Inferences-of-Word-Embeddings + +All models used are also publicly available. + +ai2spanbert: https://demo.allennlp.org/coreference-resolution +- UnifiedQA: https://github.com/allenai/unifiedqa +- Longformer: https://github.com/shtoshni/fast-coref +- Albert: https://huggingface.co/docs/trans formers/model_doc/albert +- Elmo-DA:https://demo.allennlp.org/textual-entailment/elmo-snli +- Roberta-base-SNLI:https://github.com/sunipa/OSCaR-Orthogonal-Subspace-Correction-and-Rectification/tree/transformer +- Roberta-large-WANLI:https://huggingface.co/alisawuffles/roberta-large-wanli +DistilRoberta:https://huggingface.co/cross-encoder/nli-distilroberta-base + +Code and data for the experiments are available at https://github.com/uclanlp/socialbias-dataset-construction-biases. We provide complete preprocessed datasets that correspond to the various proposed alternate constructions. They can be readily used with the publicly listed models for evaluation, thereby easily reproducing the results of the paper. We provide scripts to help with the same. The alternate dataset constructions can also be independently and flexibly used for new experiments. + +![](images/cc5727dc53e818d933db6906023ee2766c5e6424c74dcd0cbec85f18d6c3136a.jpg) + +![](images/9880ec606b46af0206ad1697a3f11288379fda257eab996270fe78bb62a8c56f.jpg) + +![](images/bf46fadd5a58b0be412b100e582b1d7268cd4467cb4291b83cb1b8c67ad9f98e.jpg) +Figure 5: Bias measures (fraction neutral) computed on BIASNLI. The violin plot attempts to capture the distribution of bias measure scores across datasets reconstructed using different $10\%$ , $25\%$ , and $50\%$ subsets (top to bottom) of the occupation word list. + +
Perturbationai2spanbertqa-smallqa-baseqa-largelongformer
Baseline (no perturbations)5.835.8316.6615.419.16
Clause after occupation4.505.5014.7523.5010.08
Clause after participant10.338.0015.0015.758.83
Adjective before occupation8.225.3416.1217.316.87
Adjective after occupation4.925.3715.5725.459.75
Adjective before participant5.975.6913.8418.5210.77
Adjective after participant8.487.4915.9118.1711.69
Synonyms7.927.5017.9215.8312.08
+ +Table 1: Percentage M-F Mismatch on WINOGENDER. + +
AlbertElmo-DARoberta-base-SNLIRoberta-large-WANLIDistilRoberta
Baseline (no perturbations)44.8141.6415.2516.8151.32
Clauses60.8540.4330.2615.6960.84
Negation45.7613.4020.0410.4562.63
+ +Table 2: Percentage neutral for different alternate constructions of BIASNLI + +A For every submission: + +A1. Did you describe the limitations of your work? + +Page 5 + +A2. Did you discuss any potential risks of your work? + +Not applicable. Left blank. + +A3. Do the abstract and introduction summarize the paper's main claims? + +Section 1 + +A4. Have you used AI writing assistants when working on this paper? + +Left blank. + +B Did you use or create scientific artifacts? + +Section 3 and Appendix J (Bias Datasets and Models used) + +B1. Did you cite the creators of artifacts you used? + +Section 3 and Appendix J (Datasets and Models used) + +B2. Did you discuss the license or terms for use and / or distribution of any artifacts? + +Appendix J (Datasets and Models used are all publicly available) + +B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? + +Not applicable. Left blank. + +B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? + +Not applicable. Left blank. + +B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? + +Not applicable. Left blank. + +B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. + +Section 3.2 and Appendix F + +C Did you run computational experiments? + +Section 3 + +C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? + +Appendix I + +The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance. + +C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Section 3, Appendix B-G +C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Section 3, Appendix H +C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Not applicable. Left blank. + +D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank. + +D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response. +D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response. +D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response. +D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response. +D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response. \ No newline at end of file diff --git a/2023/The Tail Wagging the Dog_ Dataset Construction Biases of Social Bias Benchmarks/images.zip b/2023/The Tail Wagging the Dog_ Dataset Construction Biases of Social Bias Benchmarks/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..12300b3e2304c56c170bacc9226f8283b53e1c48 --- /dev/null +++ b/2023/The Tail Wagging the Dog_ Dataset Construction Biases of Social Bias Benchmarks/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3d6c4fae3fb6df31f8349e3148d8d2a220ada8c05705290a33a638c140ae467a +size 251999 diff --git a/2023/The Tail Wagging the Dog_ Dataset Construction Biases of Social Bias Benchmarks/layout.json b/2023/The Tail Wagging the Dog_ Dataset Construction Biases of Social Bias Benchmarks/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..805442d055bbb75e7115f56e626b9a2418e5c29e --- /dev/null +++ b/2023/The Tail Wagging the Dog_ Dataset Construction Biases of Social Bias Benchmarks/layout.json @@ -0,0 +1,8909 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 126, + 70, + 468, + 101 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 70, + 468, + 101 + ], + "spans": [ + { + "bbox": [ + 126, + 70, + 468, + 101 + ], + "type": "text", + "content": "The Tail Wagging the Dog: Dataset Construction Biases of Social Bias Benchmarks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 154, + 110, + 440, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 110, + 440, + 194 + ], + "spans": [ + { + "bbox": [ + 154, + 110, + 440, + 194 + ], + "type": "text", + "content": "Nikil Roashan Selvam" + }, + { + "bbox": [ + 154, + 110, + 440, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 154, + 110, + 440, + 194 + ], + "type": "text", + "content": " Sunipa Dev" + }, + { + "bbox": [ + 154, + 110, + 440, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 154, + 110, + 440, + 194 + ], + "type": "text", + "content": " \nDaniel Khashabi" + }, + { + "bbox": [ + 154, + 110, + 440, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 154, + 110, + 440, + 194 + ], + "type": "text", + "content": " Tushar Khot" + }, + { + "bbox": [ + 154, + 110, + 440, + 194 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 154, + 110, + 440, + 194 + ], + "type": "text", + "content": " Kai-Wei Chang" + }, + { + "bbox": [ + 154, + 110, + 440, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 154, + 110, + 440, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 154, + 110, + 440, + 194 + ], + "type": "text", + "content": "University of California, Los Angeles " + }, + { + "bbox": [ + 154, + 110, + 440, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 154, + 110, + 440, + 194 + ], + "type": "text", + "content": "Google Research \n" + }, + { + "bbox": [ + 154, + 110, + 440, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 154, + 110, + 440, + 194 + ], + "type": "text", + "content": "Johns Hopkins University " + }, + { + "bbox": [ + 154, + 110, + 440, + 194 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 154, + 110, + 440, + 194 + ], + "type": "text", + "content": "Allen Institute for AI \nikilrselvam,kwchang}@ucla.edu, sunipadev@google. \ndanielk@jhu.edu, tushark@allenai.org" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 84, + 234, + 274, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 234, + 274, + 484 + ], + "spans": [ + { + "bbox": [ + 84, + 234, + 274, + 484 + ], + "type": "text", + "content": "How reliably can we trust the scores obtained from social bias benchmarks as faithful indicators of problematic social biases in a given model? In this work, we study this question by contrasting social biases with non-social biases that stem from choices made during dataset construction (which might not even be discernible to the human eye). To do so, we empirically simulate various alternative constructions for a given benchmark based on seemingly innocuous modifications (such as paraphrasing or random-sampling) that maintain the essence of their social bias. On two well-known social bias benchmarks (WINOGENDER and BIASNLI), we observe that these shallow modifications have a surprising effect on the resulting degree of bias across various models and consequently the relative ordering of these models when ranked by measured bias. We hope these troubling observations motivate more robust measures of social biases." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 495, + 154, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 495, + 154, + 507 + ], + "spans": [ + { + "bbox": [ + 68, + 495, + 154, + 507 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 516, + 290, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 516, + 290, + 610 + ], + "spans": [ + { + "bbox": [ + 67, + 516, + 290, + 610 + ], + "type": "text", + "content": "The omnipresence of large pre-trained language models (Liu et al., 2019; Raffel et al., 2020; Brown et al., 2020) has fueled concerns regarding their systematic biases carried over from underlying data into the applications they are used in, resulting in disparate treatment of people with different identities (Sheng et al., 2021; Abid et al., 2021)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 611, + 290, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 611, + 290, + 730 + ], + "spans": [ + { + "bbox": [ + 67, + 611, + 290, + 730 + ], + "type": "text", + "content": "In response to such concerns, various benchmarks have been proposed to quantify the amount of social biases in models (Rudinger et al., 2018; Sheng et al., 2019; Li et al., 2020). These measures are composed of textual datasets built for a specific NLP task (such as question answering) and are accompanied by a metric such as accuracy of prediction which is used as an approximation of the amount of social biases." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 733, + 290, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 733, + 290, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 733, + 290, + 773 + ], + "type": "text", + "content": "These bias benchmarks are commonly used by machine learning practitioners to compare the degree of social biases (such as gender-occupation" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 305, + 211, + 522, + 386 + ], + "blocks": [ + { + "bbox": [ + 305, + 211, + 522, + 386 + ], + "lines": [ + { + "bbox": [ + 305, + 211, + 522, + 386 + ], + "spans": [ + { + "bbox": [ + 305, + 211, + 522, + 386 + ], + "type": "image", + "image_path": "fd80f5c28cd735e5b85fc16d14449f902723c21e5d745137ebb90ce455b44cc3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 391, + 525, + 488 + ], + "lines": [ + { + "bbox": [ + 302, + 391, + 525, + 488 + ], + "spans": [ + { + "bbox": [ + 302, + 391, + 525, + 488 + ], + "type": "text", + "content": "Figure 1: Two potential constructions of WINOGEN- DER with minor differences: a model (span-BERT, in this case) with the original dataset might seem to have gender-occupation bias (green tick) based on the change in its pronoun resolution. However, a minor change in its phrasing with no change in meaning (e.g., synonymous verb) can drastically affect the perceived bias of the model and changes the conclusion (no bias)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 301, + 516, + 526, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 516, + 526, + 773 + ], + "spans": [ + { + "bbox": [ + 301, + 516, + 526, + 773 + ], + "type": "text", + "content": "bias) in different real-world models (Chowdhery et al., 2022; Thoppilan et al., 2022) before deploying them in a myriad of applications. However, they also inadvertently measure other non-social biases in their datasets. For example, consider the sentence from WINOGENDER in Figure 1. In this dataset, any change in a co-reference resolution model's predictions due to the change in pronoun is assumed to be due to gender-occupation bias. However, this assumption only holds for a model with near-perfect language understanding with no other biases. This may not often be the case, e.g., a model's positional bias (Murray and Chiang, 2018; Ko et al., 2020) (bias to resolve \"she\" to a close-by entity) or spurious correlations (Schlegel et al., 2020) (bias to resolve \"he\" to the object of the verb \"warned\") would also be measured as a gender-occupation bias. As a result, a slightly different template (e.g., changing the verb to \"cautioned\")" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "1373" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 135, + 795, + 459, + 805 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 795, + 459, + 805 + ], + "spans": [ + { + "bbox": [ + 135, + 795, + 459, + 805 + ], + "type": "text", + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 219, + 806, + 374, + 817 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 806, + 374, + 817 + ], + "spans": [ + { + "bbox": [ + 219, + 806, + 374, + 817 + ], + "type": "text", + "content": "Volume 2: Short Papers, pages 1373-1386" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "spans": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "text", + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 292, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 292, + 97 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 292, + 97 + ], + "type": "text", + "content": "could result in completely different bias measurements." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 99, + 292, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 99, + 292, + 396 + ], + "spans": [ + { + "bbox": [ + 69, + 99, + 292, + 396 + ], + "type": "text", + "content": "The goal of this work is to illustrate the extent to which social bias measurements are effected by assumptions that are built into dataset constructions. To that end, we consider several alternate dataset constructions for 2 bias benchmarks WINOGENDER and BIASNLI. We show that, just by the choice of certain target-bias-irrelevant elements in a dataset, it is possible to discover different degrees of bias for the same model as well as different model rankings1. For instance, one experiment on BIASNLI demonstrated that merely negating verbs drastically reduced the measured bias " + }, + { + "bbox": [ + 69, + 99, + 292, + 396 + ], + "type": "inline_equation", + "content": "(41.64\\rightarrow 13.40)" + }, + { + "bbox": [ + 69, + 99, + 292, + 396 + ], + "type": "text", + "content": " on an ELMo-based Decomposable Attention model and even caused a switch in the comparative ranking with RoBERTa. Our findings demonstrate the unreliability of current benchmarks to truly measure social bias in models and suggest caution when considering these measures as the gold truth. We provide a detailed discussion (\\$5) of the implications of our findings, relation to experienced harms, suggestions for improving bias benchmarks, and directions for future work." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 405, + 161, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 405, + 161, + 417 + ], + "spans": [ + { + "bbox": [ + 67, + 405, + 161, + 417 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 427, + 291, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 427, + 291, + 682 + ], + "spans": [ + { + "bbox": [ + 67, + 427, + 291, + 682 + ], + "type": "text", + "content": "A large body of work investigates ways to evaluate biases carried inherently in language models (Bolukbasi et al., 2016; Caliskan et al., 2017; Nadeem et al., 2021) and expressed in specific tasks (Nangia et al., 2020; Kirk et al., 2021; Schramowski et al., 2022; Prabhumoye et al., 2021; Srinivasan and Bisk, 2021; Kirk et al., 2021; Parrish et al., 2021; Baldini et al., 2022; Czarnowska et al., 2021; Dev et al., 2021a; Zhao et al., 2021). Alongside, there is also growing concern about the measures not relating to experienced harms (Blodgett et al., 2020), not inclusive in framing (Dev et al., 2021b), ambiguous about what bias is measured (Blodgett et al., 2021), not correlated in their findings of bias across intrinsic versus extrinsic techniques (Goldfarb-Tarrant et al., 2021; Cao et al., 2022), and susceptible to adversarial perturbations (Zhang et al., 2021) and seed word selection (Antoniak and Mimno, 2021)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 684, + 291, + 724 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 684, + 291, + 724 + ], + "spans": [ + { + "bbox": [ + 67, + 684, + 291, + 724 + ], + "type": "text", + "content": "The concurrent work by (Seshadri et al., 2022) discusses the unreliability of quantifying social biases using templates by varying templates in a se" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 302, + 71, + 526, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 220 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 220 + ], + "type": "text", + "content": "mantic preserving manner. While their findings are consistent with ours, the two works provide complementary experimental observations. Seshadri et al. (2022) study a wider range of tasks, though we focus our experiments on a wider set of models and alternate dataset constructions (with a greater range of syntactic and semantic variability). As a result, we are able to illustrate the effect of the observed variability on ranking large language models according to measured bias for deployment in real world applications." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 240, + 480, + 267 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 240, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 302, + 240, + 480, + 267 + ], + "type": "text", + "content": "3 Social Bias Measurements and Alternate Constructions" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 301, + 282, + 526, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 282, + 526, + 553 + ], + "spans": [ + { + "bbox": [ + 301, + 282, + 526, + 553 + ], + "type": "text", + "content": "Bias measures in NLP are often quantified through comparative prediction disparities on language datasets that follow existing tasks such as classification (De-Arteaga et al., 2019) or coreference resolution (Rudinger et al., 2018). As a result, these datasets are central to what eventually gets measured as “bias”. Not only do they determine the “amount” of bias measured but also the “type” of bias or stereotype measured. Datasets often vary combinations of gendered pronouns and occupations to evaluate stereotypical associations. It is important to note that these constructs of datasets and their templates, which determine what gets measured, are often arbitrary choices. The sentences could be differently structured, be generated from a different set of seed words, and more. However, we expect that for any faithful bias benchmark, such dataset alterations that are not relevant to social bias should not have a significant impact on the artifact (e.g. gender bias) being measured." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 556, + 527, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 556, + 527, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 556, + 527, + 772 + ], + "type": "text", + "content": "Thus, to evaluate the faithfulness of current benchmarks, we develop alternate dataset constructions through modifications that should not have any effect on the social bias being measured in a dataset. They are minor changes that should not influence models with true language understanding – the implicit assumption made by current bias benchmarks. Any notable observed changes in a model's bias measure due to these modifications would highlight the incorrectness of this assumption. Consequently, this would bring to light the unreliability of current benchmarks to faithfully measure the target bias and disentangle the measurement from measurement of other non-social biases. A non-exhaustive set of such alternate constructions considered in this work are listed below." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 730, + 291, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 730, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 730, + 291, + 772 + ], + "type": "text", + "content": "1All preprocessed datasets (original and alternate constructions) and code are available at https://github.com/uclanlp/socialbias-dataset-construction-biases." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "text", + "content": "1374" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 107, + 55, + 189, + 64 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 55, + 189, + 64 + ], + "spans": [ + { + "bbox": [ + 107, + 55, + 189, + 64 + ], + "type": "text", + "content": "Clause after occupation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 77, + 69, + 213, + 99 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 69, + 213, + 99 + ], + "spans": [ + { + "bbox": [ + 77, + 69, + 213, + 99 + ], + "type": "text", + "content": "The engineer, who just returned from the beach, informed the client that he would need to make all future payments on time." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 255, + 55, + 338, + 65 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 55, + 338, + 65 + ], + "spans": [ + { + "bbox": [ + 255, + 55, + 338, + 65 + ], + "type": "text", + "content": "Clause after participant" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 227, + 69, + 366, + 99 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 69, + 366, + 99 + ], + "spans": [ + { + "bbox": [ + 227, + 69, + 366, + 99 + ], + "type": "text", + "content": "The engineer informed the client, who just returned from the beach, that he would need to make all future payments on time." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 420, + 55, + 477, + 64 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 55, + 477, + 64 + ], + "spans": [ + { + "bbox": [ + 420, + 55, + 477, + 64 + ], + "type": "text", + "content": "Synonymization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 380, + 69, + 510, + 99 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 69, + 510, + 99 + ], + "spans": [ + { + "bbox": [ + 380, + 69, + 510, + 99 + ], + "type": "text", + "content": "The engineer informed the client that he would need to make all upcoming payments on time." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 93, + 111, + 189, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 111, + 189, + 121 + ], + "spans": [ + { + "bbox": [ + 93, + 111, + 189, + 121 + ], + "type": "text", + "content": "Adjective before occupation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 94, + 131, + 188, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 131, + 188, + 169 + ], + "spans": [ + { + "bbox": [ + 94, + 131, + 188, + 169 + ], + "type": "text", + "content": "The cruel engineer informed the client that he would need to make all future payments on time." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 201, + 111, + 292, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 111, + 292, + 121 + ], + "spans": [ + { + "bbox": [ + 201, + 111, + 292, + 121 + ], + "type": "text", + "content": "Adjective after occupation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 199, + 131, + 291, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 131, + 291, + 169 + ], + "spans": [ + { + "bbox": [ + 199, + 131, + 291, + 169 + ], + "type": "text", + "content": "The engineer, who was cruel, informed the client that he would need to make all future payments on time." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 111, + 399, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 111, + 399, + 121 + ], + "spans": [ + { + "bbox": [ + 302, + 111, + 399, + 121 + ], + "type": "text", + "content": "Adjective before participant" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 131, + 395, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 131, + 395, + 169 + ], + "spans": [ + { + "bbox": [ + 304, + 131, + 395, + 169 + ], + "type": "text", + "content": "The engineer informed the wise client that he would need to make all future payments on time." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 408, + 111, + 499, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 111, + 499, + 121 + ], + "spans": [ + { + "bbox": [ + 408, + 111, + 499, + 121 + ], + "type": "text", + "content": "Adjective after participant" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 408, + 131, + 495, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 131, + 495, + 169 + ], + "spans": [ + { + "bbox": [ + 408, + 131, + 495, + 169 + ], + "type": "text", + "content": "The engineer informed the client, who was wise, that he would need to make all future payments on time." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 67, + 190, + 524, + 226 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 190, + 524, + 226 + ], + "spans": [ + { + "bbox": [ + 67, + 190, + 524, + 226 + ], + "type": "text", + "content": "Figure 2: An instance (\"The engineer informed the client that he would need to make all future payments on time\") from WINOGENDER benchmark modified under various shallow modifications (§3). To a human eye, such modifications do not necessarily affect the outcome of the given pronoun resolution problem." + } + ] + } + ], + "index": 14, + "type": "text" + }, + { + "bbox": [ + 67, + 247, + 290, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 247, + 290, + 327 + ], + "spans": [ + { + "bbox": [ + 67, + 247, + 290, + 327 + ], + "type": "text", + "content": "Negations: A basic function in language understanding is to understand the negations of word groups such as action verbs, or adjectives. Altering verbs in particular, such as 'the doctor bought' to 'the doctor did not buy' should typically not affect the inferences made about occupation associations." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 67, + 331, + 290, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 331, + 290, + 439 + ], + "spans": [ + { + "bbox": [ + 67, + 331, + 290, + 439 + ], + "type": "text", + "content": "Synonym substitutions: Another fundamental function of language understanding is the ability to parse the usage of similar words or synonyms used in identical contexts, to derive the same overall meaning of a sentence. For bias measuring datasets, synonymizing non-pivotal words (such as non-identity words like verbs) should not change the outcome of how much bias is measured." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 67, + 443, + 290, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 443, + 290, + 550 + ], + "spans": [ + { + "bbox": [ + 67, + 443, + 290, + 550 + ], + "type": "text", + "content": "Varying length of the text: In typical evaluation datasets, the number of clauses that each sentence is composed of and overall the sentence length are arbitrary experimental choices. Fixing this length is common, especially when such datasets need to be created at scale. If language is understood, adding a neutral phrase without impacting the task-specific semantics should not alter the bias measured." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 67, + 554, + 290, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 554, + 290, + 662 + ], + "spans": [ + { + "bbox": [ + 67, + 554, + 290, + 662 + ], + "type": "text", + "content": "Adding descriptors: Sentences used in real life are structured in complex ways and can have descriptors, such as adjectives about an action, person, or object, without changing the net message expressed by the text. For example, the sentences, \"The doctor bought an apple.\", and \"The doctor bought a red apple.\" do not change any assumptions made about the doctor, or the action of buying an apple." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 67, + 665, + 290, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 665, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 665, + 290, + 772 + ], + "type": "text", + "content": "Random samples: Since the sentence constructs of these datasets are not unique, a very simple alternate construction of a dataset is a different subsample of itself. This is because the dataset is scraped or generated with specific assumptions or parameters, such as seed word lists, templates of sentences, and word order. However, neither the sentence constructs or templates, nor the seed word" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 302, + 247, + 524, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 247, + 524, + 287 + ], + "spans": [ + { + "bbox": [ + 302, + 247, + 524, + 287 + ], + "type": "text", + "content": "lists typically used are exhaustive or representative of entire categories of words (such as gendered words, emotions, and occupations)." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 302, + 288, + 525, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 288, + 525, + 315 + ], + "spans": [ + { + "bbox": [ + 302, + 288, + 525, + 315 + ], + "type": "text", + "content": "See Fig. 2 for example constructions on WINO-GENDER (App. A, B for detailed descriptions)." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 302, + 327, + 389, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 327, + 389, + 339 + ], + "spans": [ + { + "bbox": [ + 302, + 327, + 389, + 339 + ], + "type": "text", + "content": "4 Case Studies" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 302, + 349, + 525, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 349, + 525, + 375 + ], + "spans": [ + { + "bbox": [ + 302, + 349, + 525, + 375 + ], + "type": "text", + "content": "We discuss here the impact of alternate constructions on two task-based measures of bias.2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 302, + 388, + 440, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 388, + 440, + 400 + ], + "spans": [ + { + "bbox": [ + 302, + 388, + 440, + 400 + ], + "type": "text", + "content": "4.1 Coreference Resolution" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 302, + 406, + 525, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 406, + 525, + 635 + ], + "spans": [ + { + "bbox": [ + 302, + 406, + 525, + 635 + ], + "type": "text", + "content": "Several different bias measures (Rudinger et al., 2018; Zhao et al., 2018; Cao and Daumé III, 2021) for coreference resolution work similar to Winograd Schema (Winograd, 1972) where a sentence has two entities and the task is to resolve which entity a specific pronoun or noun refers to. We work here with WINOGENDER (Rudinger et al., 2018), popularly used to measure biases. It is worth noting that WINOGENDER was originally intended by its authors to merely be a diagnostic tool that checks for bias in a model; the authors note that it may demonstrate the presence of model bias but not prove the absence of the same. Nonetheless, models developed today are indeed tested and compared for social bias on WinoGender, leading to its usage as a comparative standard or benchmark (Chowdhery et al., 2022; Thoppilan et al., 2022)." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 302, + 637, + 526, + 731 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 637, + 526, + 731 + ], + "spans": [ + { + "bbox": [ + 302, + 637, + 526, + 731 + ], + "type": "text", + "content": "The metric used to evaluate bias is the percentage of sentence pairs where there is a mismatch in predictions for the male and female gendered pronouns. For instance, in Fig. 2, if the pronoun \"he\" is linked to \"engineer\" but switches to \"client\" for the pronoun \"she\", that would indicate a gender-occupation bias. Higher the number of mismatches," + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 740, + 525, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 740, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 740, + 525, + 772 + ], + "type": "text", + "content": "2We note that throughout this paper, we focus on gender- occupation bias as an illustrative example; however, our discussion can be extended to other aspects of biases too." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 287, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 780, + 309, + 791 + ], + "type": "text", + "content": "1375" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 91, + 371, + 232 + ], + "blocks": [ + { + "bbox": [ + 171, + 69, + 240, + 79 + ], + "lines": [ + { + "bbox": [ + 171, + 69, + 240, + 79 + ], + "spans": [ + { + "bbox": [ + 171, + 69, + 240, + 79 + ], + "type": "text", + "content": "(a) WINOGENDER" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 69, + 91, + 371, + 232 + ], + "lines": [ + { + "bbox": [ + 69, + 91, + 371, + 232 + ], + "spans": [ + { + "bbox": [ + 69, + 91, + 371, + 232 + ], + "type": "image", + "image_path": "1adedff8ac62a30012d7f23cb921c7b276c9ba994af299070642fc858dc7b720.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 235, + 525, + 260 + ], + "lines": [ + { + "bbox": [ + 67, + 235, + 525, + 260 + ], + "spans": [ + { + "bbox": [ + 67, + 235, + 525, + 260 + ], + "type": "text", + "content": "Figure 3: Bias measures on (a) WINOGENDER (percentage M-F mismatch, log-scale) and (b) BIASNLI (accuracy as percentage neutral, log-scale), across a variety of dataset constructions and models." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 375, + 90, + 526, + 227 + ], + "blocks": [ + { + "bbox": [ + 421, + 69, + 471, + 79 + ], + "lines": [ + { + "bbox": [ + 421, + 69, + 471, + 79 + ], + "spans": [ + { + "bbox": [ + 421, + 69, + 471, + 79 + ], + "type": "text", + "content": "(b) BIASNLI" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 375, + 90, + 526, + 227 + ], + "lines": [ + { + "bbox": [ + 375, + 90, + 526, + 227 + ], + "spans": [ + { + "bbox": [ + 375, + 90, + 526, + 227 + ], + "type": "image", + "image_path": "ff012ff95eee03763f79b1e41dd49741182d9866115ab0a9dc002da8a5978b26.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 281, + 290, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 281, + 290, + 335 + ], + "spans": [ + { + "bbox": [ + 67, + 281, + 290, + 335 + ], + "type": "text", + "content": "higher the bias. In particular, note that the metric does not take into account the accuracy of the predictions, but rather only the mismatch between the two pronouns." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 337, + 291, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 337, + 291, + 404 + ], + "spans": [ + { + "bbox": [ + 67, + 337, + 291, + 404 + ], + "type": "text", + "content": "We experiment with three alternate constructions of the dataset: addition of clauses, addition of adjectives, and synonymizing words in templates. Each alternate construction is introduced so as to not affect the overall meaning of the sentence." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 406, + 291, + 718 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 406, + 291, + 718 + ], + "spans": [ + { + "bbox": [ + 67, + 406, + 291, + 718 + ], + "type": "text", + "content": "Experimental Results: We use an end-to-end coreference model with SpanBERT embeddings (Lee et al., 2018; Joshi et al., 2020), UnifiedQA (small, base, and large) (Khashabi et al., 2020) QA model," + }, + { + "bbox": [ + 67, + 406, + 291, + 718 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 67, + 406, + 291, + 718 + ], + "type": "text", + "content": " and a long-document coreference model with Longformer encodings (Toshniwal et al., 2021). Results of evaluating these models on various WINOGENDER constructions is summarized in Fig. 3a. Small changes to the formulation of dataset templates result in sizable changes to computed bias measures compared to the published baseline constructions. For example, a construction involving added adjectives after occupations would have found the UnifiedQA (large) model to have " + }, + { + "bbox": [ + 67, + 406, + 291, + 718 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 67, + 406, + 291, + 718 + ], + "type": "text", + "content": " less bias compared to the default constructions. The sensitivity to the dataset constructions can have a drastic effect on ranking models according to their social bias, as Fig. 3a shows. For example, the SpanBERT model is considered to have less bias than UnifiedQA (small) model in the baseline dataset, but would be considered to be more biased if the templates had clauses after the participants or adjectives before the occupation." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 300, + 285, + 526, + 456 + ], + "blocks": [ + { + "bbox": [ + 300, + 285, + 526, + 456 + ], + "lines": [ + { + "bbox": [ + 300, + 285, + 526, + 456 + ], + "spans": [ + { + "bbox": [ + 300, + 285, + 526, + 456 + ], + "type": "image", + "image_path": "78f45b52836cdfaa2413129c98c9554d6798d012c243a13db12719bec25feb66.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 461, + 525, + 522 + ], + "lines": [ + { + "bbox": [ + 302, + 461, + 525, + 522 + ], + "spans": [ + { + "bbox": [ + 302, + 461, + 525, + 522 + ], + "type": "text", + "content": "Figure 4: Bias measures (fraction neutral) computed on BIASNLI. The violin plot represents distribution of bias measure scores across datasets reconstructed using different " + }, + { + "bbox": [ + 302, + 461, + 525, + 522 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 302, + 461, + 525, + 522 + ], + "type": "text", + "content": " subsets of the occupation word list across 100 random samples." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 547, + 463, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 547, + 463, + 560 + ], + "spans": [ + { + "bbox": [ + 302, + 547, + 463, + 560 + ], + "type": "text", + "content": "4.2 Natural Language Inference" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 301, + 568, + 526, + 743 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 568, + 526, + 743 + ], + "spans": [ + { + "bbox": [ + 301, + 568, + 526, + 743 + ], + "type": "text", + "content": "Natural Language Inference (NLI) is the task of determining directional relationships between two sentences (a premise " + }, + { + "bbox": [ + 301, + 568, + 526, + 743 + ], + "type": "inline_equation", + "content": "(P)" + }, + { + "bbox": [ + 301, + 568, + 526, + 743 + ], + "type": "text", + "content": " and a hypothesis " + }, + { + "bbox": [ + 301, + 568, + 526, + 743 + ], + "type": "inline_equation", + "content": "(H)" + }, + { + "bbox": [ + 301, + 568, + 526, + 743 + ], + "type": "text", + "content": "). Dev et al. (2020)'s measure based on NLI (BIASNLI) evaluates if stereotypical inferences are made by language models. We use their dataset for gender-occupation stereotypes containing approximately 2 million sentence pairs such as " + }, + { + "bbox": [ + 301, + 568, + 526, + 743 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 301, + 568, + 526, + 743 + ], + "type": "text", + "content": ": \"The doctor bought a bagel.\", " + }, + { + "bbox": [ + 301, + 568, + 526, + 743 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 301, + 568, + 526, + 743 + ], + "type": "text", + "content": ": \"The man bought a bagel.\" The expected prediction for each sentence pair in the dataset is neutral, and therefore the bias metric used is the fraction of neutral inferences on dataset - the higher the score, the lower the bias." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "type": "text", + "content": "We experiment with three alternate constructions of the dataset: verb negation, random sampling," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 730, + 290, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 730, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 730, + 290, + 772 + ], + "type": "text", + "content": "3Used by converting co-reference into question-answering, e.g., \"The technician told the customer that he had completed the repair. Who does the word 'he' refer to? " + }, + { + "bbox": [ + 67, + 730, + 290, + 772 + ], + "type": "inline_equation", + "content": "\\backslash \\mathfrak{n}" + }, + { + "bbox": [ + 67, + 730, + 290, + 772 + ], + "type": "text", + "content": " (a) technician (b) customer\"" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 287, + 780, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 780, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 780, + 310, + 791 + ], + "type": "text", + "content": "1376" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 292, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 292, + 139 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 292, + 139 + ], + "type": "text", + "content": "and addition of clauses. Note that the alternate constructions do not impact the unbiased label (neutral). Any change in construction (say negating a verb) is applied to both the premise and hypothesis. Refer to App. B for a detailed description." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 140, + 291, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 140, + 291, + 410 + ], + "spans": [ + { + "bbox": [ + 69, + 140, + 291, + 410 + ], + "type": "text", + "content": "Experimental Results: We use RoBERTa trained on SNLI (RoBERTa-base-SNLI) (Liu et al., 2019), ELMo-based Decomposable Attention (ELMoDA) (Parikh et al., 2016), ALBERT (Lan et al., 2019), distilled version of the RoBERTa-base model (Sanh et al., 2019), and RoBERTa-large fin-tuned on WANLI (Liu et al., 2022). The bias measured with each model using BIASNLI is recorded in Fig. 3b. The results show how small modifications to the dataset again result in large changes to the bias measured, and also change the bias rankings. For example, adding a negation largely reduces the bias measured " + }, + { + "bbox": [ + 69, + 140, + 291, + 410 + ], + "type": "inline_equation", + "content": "(\\triangle = 28.24)" + }, + { + "bbox": [ + 69, + 140, + 291, + 410 + ], + "type": "text", + "content": " for ELMoDA, and also results in a switch in the comparative ranking to RoBERTa-base-SNLI. Furthermore, as seen in Fig. 4, there is a significant overlap in the bias measures of ALBERT, DistilRoBERTa, and ELMo-DA under random sampling, which corresponds to high variability in relative model ordering across different sub-samples of the dataset." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 421, + 226, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 421, + 226, + 433 + ], + "spans": [ + { + "bbox": [ + 67, + 421, + 226, + 433 + ], + "type": "text", + "content": "5 Discussion and Conclusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 443, + 291, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 443, + 291, + 605 + ], + "spans": [ + { + "bbox": [ + 67, + 443, + 291, + 605 + ], + "type": "text", + "content": "Social bias measurements are very sensitive to evaluation methodology. Our empirical evidence sheds light on how the model's non-social biases brought out or masked by alternate constructions can cause bias benchmarks to underestimate or overestimate the social bias in a model. More interestingly, it is important to note that different models respond differently to perturbations. In fact, the same perturbation can result in a higher or lower measured bias depending on the model (as seen in §4.1 and §4.2), which points to how models might parse information (and thus bias) differently." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 607, + 292, + 754 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 607, + 292, + 754 + ], + "spans": [ + { + "bbox": [ + 67, + 607, + 292, + 754 + ], + "type": "text", + "content": "While current bias measures do play a role in exposing where model errors have a stereotypical connotation, a lack of sentence construction variability or even assumptions made when creating seed word lists can reduce the reliability of the benchmarks, as we see in this work (§4.2). Even with simple sentences, it is not apparent how to disentangle the biased association of the identity with the verb or the occupation amongst others. This is especially important to note as it highlights that measures can lack concrete definitions of what bi" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 302, + 71, + 526, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 111 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 111 + ], + "type": "text", + "content": "ased associations they measure. Consequently, the relation between measured bias and experienced harm becomes unclear." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 114, + 527, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 114, + 527, + 478 + ], + "spans": [ + { + "bbox": [ + 302, + 114, + 527, + 478 + ], + "type": "text", + "content": "We hope that our troubling observations motivate future work that thoroughly investigates how to construct robust benchmarks that faithfully measure the target bias without being affected by model errors and other non-social biases. As suggested by our subsampling experiments (Appendix F), it might be fruitful to encourage both syntactic and semantic diversity in these benchmarks. Bias benchmarks that provide uncertainty measures (instead of a single number) might enable practitioners to better compare models before deploying them. Furthermore, since the opaqueness of large language models makes it challenging to understand how and to what extent a linguistic change will affect the measured bias, explainable models might indeed facilitate better measurement of their social bias. Assuming that we can generate faithful explanations for a model's predictions, an exciting future direction is to explore construction of bias benchmarks which operate on the explanations of the predictions rather than the predictions themselves. Lastly, we also encourage discussions on the complexity of the sentences used in benchmarks and their implications on what gets measured in relation to un-templated, naturally-occurring text (Levy et al., 2021), as an attempt to ground our measurements in experienced harms." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 303, + 491, + 365, + 505 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 491, + 365, + 505 + ], + "spans": [ + { + "bbox": [ + 303, + 491, + 365, + 505 + ], + "type": "text", + "content": "Limitations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 516, + 526, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 516, + 526, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 516, + 526, + 773 + ], + "type": "text", + "content": "We acknowledge the underlying assumptions of the social bias benchmarks used in our study. While the presented study aims to point out a key limitation of currently accepted methodologies, the presented investigation could benefit from more diversification. First, this study focuses on English. While we expect similar issues with similarly-constructed benchmarks in other languages, we leave it to future work to formally address the same. Also, the bias benchmarks themselves imbibe the notion of fairness with the Western value system (Bhatt et al., 2022), and future explorations of benchmarks should diversify culturally as well. Last but not least, we acknowledge the harm of binary treatment of genders in one of the target benchmarks. The purpose of this work was to bring light to a broader problem regarding the reliability of social benchmark metrics, with the hypothesis that the main idea of this paper would hold for a wider" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 80, + 760, + 287, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 760, + 287, + 772 + ], + "spans": [ + { + "bbox": [ + 80, + 760, + 287, + 772 + ], + "type": "text", + "content": "Also observed at " + }, + { + "bbox": [ + 80, + 760, + 287, + 772 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 80, + 760, + 287, + 772 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 80, + 760, + 287, + 772 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 80, + 760, + 287, + 772 + ], + "type": "text", + "content": " samples in Fig. 5(App.)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 287, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 780, + 309, + 791 + ], + "type": "text", + "content": "1377" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 293, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 293, + 304 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 293, + 304 + ], + "type": "text", + "content": "range of datasets with other assumptions or notions of fairness. We also acknowledge that there are larger models that we were not able to train and evaluate due to the limitations on our computational budget. The current study was focused on benchmarks with templated instances. This is no coincidence: the dominant majority of the social bias benchmarking literature relies on sentences with some degree of known structure, even in those collected from the wild (Levy et al., 2021). Such structural assumptions in datasets are necessary for defining and extracting quantifiable measures of social bias, which as we argue, are the reason behind the brittleness of their decisions. Future work should focus on making our bias benchmarks more diverse and robust to small decisions that go into making them." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 311, + 155, + 325 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 311, + 155, + 325 + ], + "spans": [ + { + "bbox": [ + 68, + 311, + 155, + 325 + ], + "type": "text", + "content": "Broader Impact" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 333, + 291, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 333, + 291, + 509 + ], + "spans": [ + { + "bbox": [ + 67, + 333, + 291, + 509 + ], + "type": "text", + "content": "Bias evaluating benchmarks play a very significant role in helping identify potential risks of language technologies. While a large body of work evolves in this area of work, there is growing concern about the ability of the different benchmarks to accurately quantify and identify social biases. We emphasize these concerns by evaluating how robust the benchmarks are to alternate constructions based on simple linguistic properties. It is important to note how inaccurate measurements of social biases can be problematic by underestimating or misdiagnosing the potential harm from language models. We hope our work helps identify such pitfalls." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 518, + 171, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 518, + 171, + 533 + ], + "spans": [ + { + "bbox": [ + 68, + 518, + 171, + 533 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 539, + 291, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 539, + 291, + 635 + ], + "spans": [ + { + "bbox": [ + 67, + 539, + 291, + 635 + ], + "type": "text", + "content": "We thank the students and colleagues at UCLA, JHU and AI2 for their insightful feedback towards improving this paper. The authors would also like to thank the anonymous reviewers for their constructive feedback. This project is supported by generous gifts from Allen Institute for AI, CISCO, Amazon, and a Sloan fellowship." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 656, + 127, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 656, + 127, + 669 + ], + "spans": [ + { + "bbox": [ + 68, + 656, + 127, + 669 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 675, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 69, + 675, + 291, + 721 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 675, + 291, + 721 + ], + "spans": [ + { + "bbox": [ + 69, + 675, + 291, + 721 + ], + "type": "text", + "content": "Abubakar Abid, Maheen Farooqi, and James Zou. 2021. Persistent anti-muslim bias in large language models. In AAAI/ACM Conference on AI, Ethics, and Society (AIES), pages 298-306." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 728, + 291, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 728, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 728, + 291, + 772 + ], + "type": "text", + "content": "Maria Antoniak and David Mimno. 2021. Bad seeds: Evaluating lexical methods for bias measurement. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 526, + 772 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 314, + 72, + 526, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 526, + 117 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 526, + 117 + ], + "type": "text", + "content": "11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1889-1904, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 125, + 526, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 125, + 526, + 193 + ], + "spans": [ + { + "bbox": [ + 304, + 125, + 526, + 193 + ], + "type": "text", + "content": "Ioana Baldini, Dennis Wei, Karthikeyan Natesan Ramamurthy, Moninder Singh, and Mikhail Yurochkin. 2022. Your fairness may vary: Pretrained language model fairness in toxic text classification. In Annual Meeting of the Association for Computational Linguistics (ACL) - Findings." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 200, + 526, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 200, + 526, + 301 + ], + "spans": [ + { + "bbox": [ + 304, + 200, + 526, + 301 + ], + "type": "text", + "content": "Shaily Bhatt, Sunipa Dev, Partha Talukdar, Shachi Dave, and Vinodkumar Prabhakaran. 2022. Recontextualizing fairness in NLP: The case of India. In Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 727-740, Online only. Association for Computational Linguistics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 308, + 526, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 526, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 526, + 365 + ], + "type": "text", + "content": "Su Lin Blodgett, Solon Barocas, Hal Daumé III, and Hanna Wallach. 2020. Language (technology) is power: A critical survey of \"bias\" in nlp. In Annual Meeting of the Association for Computational Linguistics (ACL)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 373, + 526, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 526, + 429 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 526, + 429 + ], + "type": "text", + "content": "Su Lin Blodgett, Gilsinia Lopez, Alexandra Olteanu, Robert Sim, and Hanna Wallach. 2021. Stereotyping norwegian salmon: an inventory of pitfalls in fairness benchmark datasets. In Annual Meeting of the Association for Computational Linguistics (ACL)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 438, + 526, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 526, + 504 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 526, + 504 + ], + "type": "text", + "content": "Tolga Bolukbasi, Kai-Wei Chang, James Y Zou, Venkatesh Saligram, and Adam T Kalai. 2016. Man is to computer programmer as woman is to homemaker? debiasing word embeddings. In Advances in Neural Information Processing Systems, volume 29. Curran Associates, Inc." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 513, + 526, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 513, + 526, + 580 + ], + "spans": [ + { + "bbox": [ + 304, + 513, + 526, + 580 + ], + "type": "text", + "content": "Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, and et al. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems (NeurIPS)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 588, + 526, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 588, + 526, + 633 + ], + "spans": [ + { + "bbox": [ + 304, + 588, + 526, + 633 + ], + "type": "text", + "content": "Aylin Caliskan, Joanna J. Bryson, and Arvind Narayanan. 2017. Semantics derived automatically from language corpora contain human-like biases. Science, 356(6334):183-186." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 641, + 526, + 687 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 526, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 526, + 687 + ], + "type": "text", + "content": "Yang Trista Cao and Hal Daumé III. 2021. Toward gender-inclusive coreference resolution: An analysis of gender and bias throughout the machine learning lifecycle. Computational Linguistics (CL)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 694, + 526, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 694, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 694, + 526, + 772 + ], + "type": "text", + "content": "Yang Trista Cao, Yada Pruksachatkun, Kai-Wei Chang, Rahul Gupta, Varun Kumar, Jwala Dhamala, and Aram Galstyan. 2022. On the intrinsic and extrinsic fairness evaluation metrics for contextualized language representations. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 561-570," + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "text", + "content": "1378" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 80, + 72, + 291, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 72, + 291, + 95 + ], + "spans": [ + { + "bbox": [ + 80, + 72, + 291, + 95 + ], + "type": "text", + "content": "Dublin, Ireland. Association for Computational Linguistics." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 105, + 291, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 105, + 291, + 358 + ], + "spans": [ + { + "bbox": [ + 69, + 105, + 291, + 358 + ], + "type": "text", + "content": "Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam M. Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Benton C. Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan Firat, Michele Catasta, Jason Wei, Kathleen S. Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. 2022. Palm: Scaling language modeling with pathways. ArXiv, abs/2204.02311." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 367, + 291, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 367, + 291, + 423 + ], + "spans": [ + { + "bbox": [ + 69, + 367, + 291, + 423 + ], + "type": "text", + "content": "Paula Czarnowska, Yogarshi Vyas, and Kashif Shah. 2021. Quantifying social biases in nlp: A generalization and empirical comparison of extrinsic fairness metrics. Transactions of the Association for Computational Linguistics (TACL)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 433, + 291, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 433, + 291, + 510 + ], + "spans": [ + { + "bbox": [ + 69, + 433, + 291, + 510 + ], + "type": "text", + "content": "Maria De-Arteaga, Alexey Romanov, Hanna Wallach, Jennifer Chayes, Christian Borgs, Alexandra Chouldechova, Sahin Geyik, Krishnamaram Kenthapadi, and Adam Kalai. 2019. Bias in bios: A case study of semantic representation bias in a high-stakes setting. In ACM Conference on Fairness, Accountability and Transparency (FAccT)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 520, + 291, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 520, + 291, + 565 + ], + "spans": [ + { + "bbox": [ + 69, + 520, + 291, + 565 + ], + "type": "text", + "content": "Sunipa Dev, Tao Li, Jeff M. Phillips, and Vivek Srikumar. 2020. On measuring and mitigating biased inferences of word embeddings. Conference on Artificial Intelligence (AAAI)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 575, + 291, + 630 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 575, + 291, + 630 + ], + "spans": [ + { + "bbox": [ + 69, + 575, + 291, + 630 + ], + "type": "text", + "content": "Sunipa Dev, Tao Li, Jeff M Phillips, and Vivek Srikumar. 2021a. Oscar: Orthogonal subspace correction and rectification of biases in word embeddings. In _Conference on Empirical Methods in Natural Language Processing (EMNLP)\\." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 640, + 291, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 640, + 291, + 707 + ], + "spans": [ + { + "bbox": [ + 69, + 640, + 291, + 707 + ], + "type": "text", + "content": "Sunipa Dev, Masoud Monajatipoor, Anaelia Ovalle, Arjun Subramonian, Jeff Phillips, and Kai-Wei Chang. 2021b. Harms of gender exclusivity and challenges in non-binary representation in language technologies. In Conference on Empirical Methods in Natural Language Processing (EMNLP)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 716, + 291, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 716, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 716, + 291, + 772 + ], + "type": "text", + "content": "Seraphina Goldfarb-Tarrant, Rebecca Marchant, Ricardo Muñoz Sánchez, Mugdha Pandya, and Adam Lopez. 2021. Intrinsic bias metrics do not correlate with application bias. In Annual Meeting of the Association for Computational Linguistics (ACL)." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 526, + 772 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 304, + 72, + 526, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 526, + 127 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 526, + 127 + ], + "type": "text", + "content": "Mandar Joshi, Danqi Chen, Yinhan Liu, Daniel S. Weld, Luke Zettlemoyer, and Omer Levy. 2020. SpanBERT: Improving pre-training by representing and predicting spans. Transactions of the Association for Computational Linguistics (TACL)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 137, + 526, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 137, + 526, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 137, + 526, + 204 + ], + "type": "text", + "content": "Daniel Khashabi, Sewon Min, Tushar Khot, Ashish Sabharwal, Oyvind Tafjord, Peter Clark, and Hannaneh Hajishirzi. 2020. UnifiedQA: Crossing Format Boundaries With a Single QA System. In Conference on Empirical Methods in Natural Language Processing (EMNLP) - Findings." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 214, + 526, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 214, + 526, + 280 + ], + "spans": [ + { + "bbox": [ + 304, + 214, + 526, + 280 + ], + "type": "text", + "content": "Hannah Rose Kirk, Filippo Volpin, Haider Iqbal, Elias Benussi, Frederic Dreyer, Aleksandar Shtedritski, Yuki Asano, et al. 2021. Bias out-of-the-box: An empirical analysis of intersectional occupational biases in popular generative language models. Advances in Neural Information Processing Systems (NeurIPS)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 290, + 526, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 290, + 526, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 290, + 526, + 346 + ], + "type": "text", + "content": "Miyoung Ko, Jinhyuk Lee, Hyunjae Kim, Gangwoo Kim, and Jaewoo Kang. 2020. Look at the first sentence: Position bias in question answering. In _Conference on Empirical Methods in Natural Language Processing_ (EMNLP)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 356, + 526, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 356, + 526, + 412 + ], + "spans": [ + { + "bbox": [ + 304, + 356, + 526, + 412 + ], + "type": "text", + "content": "Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. 2019. Albert: A lite bert for self-supervised learning of language representations. In International Conference on Learning Representations (ICLR)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 421, + 526, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 421, + 526, + 477 + ], + "spans": [ + { + "bbox": [ + 304, + 421, + 526, + 477 + ], + "type": "text", + "content": "Kenton Lee, Luheng He, and Luke Zettlemoyer. 2018. Higher-order coreference resolution with coarse-to-fine inference. In Conference of the North American Chapter of the Association for Computational Linguistics (NAACL)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 487, + 526, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 526, + 543 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 526, + 543 + ], + "type": "text", + "content": "Shahar Levy, Koren Lazar, and Gabriel Stanovsky. 2021. Collecting a large-scale gender bias dataset for coreference resolution and machine translation. In *Conference on Empirical Methods in Natural Language Processing* (EMNLP) - Findings." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 553, + 526, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 553, + 526, + 608 + ], + "spans": [ + { + "bbox": [ + 304, + 553, + 526, + 608 + ], + "type": "text", + "content": "Tao Li, Daniel Khashabi, Tushar Khot, Ashish Sabharwal, and Vivek Srikumar. 2020. UnCovering Stereotypical Biases via Underspecified Questions. In Conference on Empirical Methods in Natural Language Processing (EMNLP) - Findings." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 618, + 526, + 663 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 526, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 526, + 663 + ], + "type": "text", + "content": "Alisa Liu, Swabha Swayamdipta, Noah A Smith, and Yejin Choi. 2022. WANLI: Worker and AI Collaboration for Natural Language Inference Dataset Creation. arXiv preprint arXiv:2201.05955." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 673, + 526, + 729 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 673, + 526, + 729 + ], + "spans": [ + { + "bbox": [ + 304, + 673, + 526, + 729 + ], + "type": "text", + "content": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 739, + 526, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 739, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 739, + 526, + 772 + ], + "type": "text", + "content": "Kenton Murray and David Chiang. 2018. Correcting length bias in neural machine translation. In Conference on Machine Translation (WMT)." + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1379" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 291, + 161 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 291, + 161 + ], + "type": "text", + "content": "Moin Nadeem, Anna Bethke, and Siva Reddy. 2021. StereoSet: Measuring stereotypical bias in pretrained language models. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 5356-5371, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 171, + 290, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 171, + 290, + 227 + ], + "spans": [ + { + "bbox": [ + 69, + 171, + 290, + 227 + ], + "type": "text", + "content": "Nikita Nangia, Clara Vania, Rasika Bhalerao, and Samuel R Bowman. 2020. Crows-pairs: A challenge dataset for measuring social biases in masked language models. In Conference on Empirical Methods in Natural Language Processing (EMNLP)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 238, + 290, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 238, + 290, + 294 + ], + "spans": [ + { + "bbox": [ + 69, + 238, + 290, + 294 + ], + "type": "text", + "content": "Ankur P. Parikh, Oscar Tackström, Dipanjan Das, and Jakob Uszkoreit. 2016. A decomposable attention model for natural language inference. In *Conference on Empirical Methods in Natural Language Processing* (EMNLP)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 305, + 290, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 305, + 290, + 371 + ], + "spans": [ + { + "bbox": [ + 69, + 305, + 290, + 371 + ], + "type": "text", + "content": "Alicia Parrish, Angelica Chen, Nikita Nangia, Vishakh Padmakumar, Jason Phang, Jana Thompson, Phu Mon Htut, and Samuel R Bowman. 2021. BBQ: A hand-built bias benchmark for question answering. In Annual Meeting of the Association for Computational Linguistics (ACL)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 383, + 290, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 383, + 290, + 439 + ], + "spans": [ + { + "bbox": [ + 69, + 383, + 290, + 439 + ], + "type": "text", + "content": "Shrimai Prabhumoye, Rafal Kocielnik, Mohammad Shoeybi, Anima Anandkumar, and Bryan Catanzaro. 2021. Few-shot instruction prompts for pretrained language models to detect social biases. arXiv preprint arXiv:2112.07868." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 449, + 290, + 516 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 449, + 290, + 516 + ], + "spans": [ + { + "bbox": [ + 69, + 449, + 290, + 516 + ], + "type": "text", + "content": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research (JMLR)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 528, + 290, + 583 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 528, + 290, + 583 + ], + "spans": [ + { + "bbox": [ + 69, + 528, + 290, + 583 + ], + "type": "text", + "content": "Rachel Rudinger, Jason Naradowsky, Brian Leonard, and Benjamin Van Durme. 2018. Gender bias in coreference resolution. In Conference of the North American Chapter of the Association for Computational Linguistics (NAACL)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 594, + 290, + 638 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 594, + 290, + 638 + ], + "spans": [ + { + "bbox": [ + 69, + 594, + 290, + 638 + ], + "type": "text", + "content": "Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. 2019. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. ArXiv, abs/1910.01108." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 650, + 290, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 650, + 290, + 704 + ], + "spans": [ + { + "bbox": [ + 69, + 650, + 290, + 704 + ], + "type": "text", + "content": "Viktor Schlegel, Goran Nenadic, and Riza Batista-Navarro. 2020. Beyond leaderboards: A survey of methods for revealing weaknesses in natural language inference data and models. arXiv preprint arXiv:2005.14709." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 716, + 290, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 716, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 716, + 290, + 772 + ], + "type": "text", + "content": "Patrick Schramowski, Cigdem Turan, Nico Andersen, Constantin A Rothkopf, and Kristian Kersting. 2022. Large pre-trained language models contain human-like biases of what is right and wrong to do. Nature Machine Intelligence." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 524, + 700 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 304, + 72, + 524, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 524, + 116 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 524, + 116 + ], + "type": "text", + "content": "Preethi Seshadri, Pouya Pezeshkpour, and Sameer Singh. 2022. Quantifying social biases using templates is unreliable. arXiv preprint arXiv:2210.04337." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 125, + 524, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 125, + 524, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 125, + 524, + 180 + ], + "type": "text", + "content": "Emily Sheng, Kai-Wei Chang, Prem Natarajan, and Nanyun Peng. 2019. The Woman Worked as a Babysitter: On Biases in Language Generation. In Conference on Empirical Methods in Natural Language Processing (EMNLP)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 188, + 524, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 188, + 524, + 243 + ], + "spans": [ + { + "bbox": [ + 304, + 188, + 524, + 243 + ], + "type": "text", + "content": "Emily Sheng, Kai-Wei Chang, Premkumar Natarajan, and Nanyun Peng. 2021. Societal biases in language generation: Progress and challenges. In *Conference on Empirical Methods in Natural Language Processing* (EMNLP)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 252, + 524, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 252, + 524, + 297 + ], + "spans": [ + { + "bbox": [ + 304, + 252, + 524, + 297 + ], + "type": "text", + "content": "Tejas Srinivasan and Yonatan Bisk. 2021. Worst of both worlds: Biases compound in pre-trained vision-and-language models. In Workshop on Gender Bias in Natural Language Processing." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 305, + 524, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 305, + 524, + 360 + ], + "spans": [ + { + "bbox": [ + 304, + 305, + 524, + 360 + ], + "type": "text", + "content": "Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, et al. 2022. LaMDA: Language Models for Dialog Applications. arXiv preprint arXiv:2201.08239." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 369, + 524, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 524, + 424 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 524, + 424 + ], + "type": "text", + "content": "Shubham Toshniwal, Patrick Xia, Sam Wiseman, Karen Livescu, and Kevin Gimpel. 2021. On generalization in coreference resolution. In Proceedings of the Workshop on Computational Models of Reference, Anaphora and Coreference." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 432, + 524, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 432, + 524, + 455 + ], + "spans": [ + { + "bbox": [ + 304, + 432, + 524, + 455 + ], + "type": "text", + "content": "T. Winograd. 1972. Understanding natural language. Cognitive psychology, 3(1):1-191." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 463, + 524, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 463, + 524, + 551 + ], + "spans": [ + { + "bbox": [ + 304, + 463, + 524, + 551 + ], + "type": "text", + "content": "Chong Zhang, Jieyu Zhao, Huan Zhang, Kai-Wei Chang, and Cho-Jui Hsieh. 2021. Double perturbation: On the robustness of robustness and counterfactual bias evaluation. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3899-3916, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 560, + 524, + 626 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 560, + 524, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 560, + 524, + 626 + ], + "type": "text", + "content": "Jieyu Zhao, Daniel Khashabi, Tushar Khot, Ashish Sabharwal, and Kai-Wei Chang. 2021. Ethical-advice taker: Do language models understand natural language interventions? In Annual Meeting of the Association for Computational Linguistics (ACL) - Findings." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 634, + 524, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 634, + 524, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 634, + 524, + 700 + ], + "type": "text", + "content": "Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. 2018. Gender bias in coreference resolution: Evaluation and debiasing methods. In Conference of the North American Chapter of the Association for Computational Linguistics (NAACL)." + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1380" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 85, + 79, + 508, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 79, + 508, + 126 + ], + "spans": [ + { + "bbox": [ + 85, + 79, + 508, + 126 + ], + "type": "text", + "content": "Appendix The Tail Wagging the Dog: Dataset Construction Biases of Social Bias Benchmarks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 178, + 228, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 178, + 228, + 205 + ], + "spans": [ + { + "bbox": [ + 67, + 178, + 228, + 205 + ], + "type": "text", + "content": "A Alternate Constructions of WINOGENDER" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 214, + 290, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 214, + 290, + 322 + ], + "spans": [ + { + "bbox": [ + 66, + 214, + 290, + 322 + ], + "type": "text", + "content": "Addition of clauses: For WINOGENDER, we add clauses like \"who just returned from the beach\" to the different entities in the sentence. For instance, the sentence \"The customer left the bartender a big tip because he was feeling generous.\" becomes \"The customer, who just returned from the beach, left the bartender a big tip because he was feeling generous.\"" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 331, + 291, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 331, + 291, + 454 + ], + "spans": [ + { + "bbox": [ + 66, + 331, + 291, + 454 + ], + "type": "text", + "content": "Synonym substitution: We substitute with synonyms such that it does not change the meaning of the sentence. WINOGENDER has 720 sentences generated from 120 templates, making manual substitution of synonyms in the templates feasible. For example, the sentence \"The supervisor gave the employee feedback on his stellar performance.\" is replaced by \"The supervisor gave the employee feedback on his amazing performance.\"" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 461, + 291, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 461, + 291, + 611 + ], + "spans": [ + { + "bbox": [ + 67, + 461, + 291, + 611 + ], + "type": "text", + "content": "Adding adjectives: As discussed in §3, we add descriptors in the form of adjectives that do not add information about which entity the pronoun or noun would refer to. We do it in four distinct ways, (i) adding the descriptor to the occupation mentioned, e.g. doctor (e.g., \"doctor\" to \"good doctor\"), (ii) adding it to the occupation as a separate clause (e.g., \"doctor\" to \"the doctor who was good\"), (iii) adding the descriptor to the participant mentioned, e.g., \"client\" (similar to (i)), and (iv) adding it to the participant as a separate clause (similar to (ii))." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 621, + 280, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 621, + 280, + 634 + ], + "spans": [ + { + "bbox": [ + 67, + 621, + 280, + 634 + ], + "type": "text", + "content": "B Alternate Constructions of BIASNLI" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 643, + 290, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 643, + 290, + 723 + ], + "spans": [ + { + "bbox": [ + 66, + 643, + 290, + 723 + ], + "type": "text", + "content": "Negation: We negate the verb in each sentence of the dataset. For example, " + }, + { + "bbox": [ + 66, + 643, + 290, + 723 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 66, + 643, + 290, + 723 + ], + "type": "text", + "content": ": \"The doctor bought a bagel.\", " + }, + { + "bbox": [ + 66, + 643, + 290, + 723 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 66, + 643, + 290, + 723 + ], + "type": "text", + "content": ": \"The man bought a bagel.\", and " + }, + { + "bbox": [ + 66, + 643, + 290, + 723 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 66, + 643, + 290, + 723 + ], + "type": "text", + "content": ": \"The doctor did not buy a bagel.\", " + }, + { + "bbox": [ + 66, + 643, + 290, + 723 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 66, + 643, + 290, + 723 + ], + "type": "text", + "content": ": \"The man did not buy a bagel.\", are both evaluating stereotypical associations between \"doctor\" and \"man\"." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 732, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 732, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 732, + 291, + 772 + ], + "type": "text", + "content": "Random sampling: The BIASNLI dataset is generated from templates by populating empty slots (e.g. verbs, objects) with words from cor" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 179, + 526, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 179, + 526, + 274 + ], + "spans": [ + { + "bbox": [ + 302, + 179, + 526, + 274 + ], + "type": "text", + "content": "responding lists. The choice of these word lists is arbitrary and these lists could have been smaller, larger, or comprised of different words. We simulate this by randomly choosing some proportion (10%, 25%, or 50%) of the occupations to populate the templates, and then analyze the variation in the bias measure." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 282, + 526, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 282, + 526, + 376 + ], + "spans": [ + { + "bbox": [ + 302, + 282, + 526, + 376 + ], + "type": "text", + "content": "Addition of clauses: We add a clause after the subject in each sentence. For example, for the sentence pair, " + }, + { + "bbox": [ + 302, + 282, + 526, + 376 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 302, + 282, + 526, + 376 + ], + "type": "text", + "content": ": \"The doctor bought a coat.\" " + }, + { + "bbox": [ + 302, + 282, + 526, + 376 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 302, + 282, + 526, + 376 + ], + "type": "text", + "content": ": \"The man bought a coat.\" is modified to " + }, + { + "bbox": [ + 302, + 282, + 526, + 376 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 302, + 282, + 526, + 376 + ], + "type": "text", + "content": ": \"The doctor, who came in the afternoon, bought a coat.\" " + }, + { + "bbox": [ + 302, + 282, + 526, + 376 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 302, + 282, + 526, + 376 + ], + "type": "text", + "content": ": \"The man, who came in the afternoon, bought a coat.\"" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 387, + 387, + 401 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 387, + 387, + 401 + ], + "spans": [ + { + "bbox": [ + 302, + 387, + 387, + 401 + ], + "type": "text", + "content": "C Descriptors" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 410, + 526, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 410, + 526, + 518 + ], + "spans": [ + { + "bbox": [ + 302, + 410, + 526, + 518 + ], + "type": "text", + "content": "For WINOGENDER, here is the set of adjectives used to modify either the occupation word or the participant word: aggressive, arrogant, beautiful, brilliant, clean, clever, cruel, deceitful, devious, dirty, dumb, evil, generous, gentle, greedy, hateful, honest, humorless, ignorant, intelligent, intolerant, neat, professional, rude, smart, strong, stupid, terrible, ugly, unclean, unprofessional, weak, wise." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 528, + 367, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 528, + 367, + 541 + ], + "spans": [ + { + "bbox": [ + 302, + 528, + 367, + 541 + ], + "type": "text", + "content": "D Clauses" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 550, + 525, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 550, + 525, + 631 + ], + "spans": [ + { + "bbox": [ + 302, + 550, + 525, + 631 + ], + "type": "text", + "content": "We use the following clauses in WINOGENDER and BIASNLI to increase the distance between relevant parts of the sentence: who just returned from the restaurant, who came in the afternoon, who just came back, who went to the restaurant, who just returned from the beach." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 643, + 463, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 643, + 463, + 657 + ], + "spans": [ + { + "bbox": [ + 302, + 643, + 463, + 657 + ], + "type": "text", + "content": "E Synonymization Examples" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 665, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 665, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 665, + 526, + 772 + ], + "type": "text", + "content": "For WINOGENDER, we manually perform synonymization for all 120 templates. Note that while the replacements might not be exact synonyms, they are replacements of non-identity words that do not change the overall meaning of the sentence and hence should not have any notable impact on the gender bias being measured. We report a few characteristic examples of such substitutions here:" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 308, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 308, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 308, + 791 + ], + "type": "text", + "content": "1381" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 81, + 71, + 289, + 379 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 81, + 71, + 289, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 71, + 289, + 111 + ], + "spans": [ + { + "bbox": [ + 81, + 71, + 289, + 111 + ], + "type": "text", + "content": "- The taxpayer met with the accountant to get help filing his taxes " + }, + { + "bbox": [ + 81, + 71, + 289, + 111 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 81, + 71, + 289, + 111 + ], + "type": "text", + "content": " The taxpayer met with the accountant to get aid filing his taxes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 81, + 121, + 289, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 121, + 289, + 174 + ], + "spans": [ + { + "bbox": [ + 81, + 121, + 289, + 174 + ], + "type": "text", + "content": "- The supervisor gave the employee feedback on his stellar performance " + }, + { + "bbox": [ + 81, + 121, + 289, + 174 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 81, + 121, + 289, + 174 + ], + "type": "text", + "content": " The supervisor gave the employee feedback on his amazing performance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 81, + 185, + 289, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 185, + 289, + 238 + ], + "spans": [ + { + "bbox": [ + 81, + 185, + 289, + 238 + ], + "type": "text", + "content": "- The hygienist told the patient that he needed to floss every day to avoid gum disease " + }, + { + "bbox": [ + 81, + 185, + 289, + 238 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 81, + 185, + 289, + 238 + ], + "type": "text", + "content": " The hygienist told the patient that he needed to brush every day to avoid cavities." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 81, + 248, + 289, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 248, + 289, + 301 + ], + "spans": [ + { + "bbox": [ + 81, + 248, + 289, + 301 + ], + "type": "text", + "content": "- The broker called the client because he had requested a phone consultation " + }, + { + "bbox": [ + 81, + 248, + 289, + 301 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 81, + 248, + 289, + 301 + ], + "type": "text", + "content": " The broker called the client because he had asked for a phone consultation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 81, + 312, + 289, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 312, + 289, + 379 + ], + "spans": [ + { + "bbox": [ + 81, + 312, + 289, + 379 + ], + "type": "text", + "content": "- The chef came out to apologize to the guest who was unhappy with his preparation style " + }, + { + "bbox": [ + 81, + 312, + 289, + 379 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 81, + 312, + 289, + 379 + ], + "type": "text", + "content": " The chef came out to apologize to the guest who was dissatisfied with his preparation style." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 68, + 390, + 158, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 390, + 158, + 404 + ], + "spans": [ + { + "bbox": [ + 68, + 390, + 158, + 404 + ], + "type": "text", + "content": "F Subsampling" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 412, + 291, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 412, + 291, + 642 + ], + "spans": [ + { + "bbox": [ + 67, + 412, + 291, + 642 + ], + "type": "text", + "content": "The gender-occupation subset of the original construction of BIASNLI consists of 164 occupation words such as accountant, firefighter, tutor, and model. In each trial, we subsample some proportion (10%, 25%, or 50%) of these occupation words used in the templates to regenerate the dataset and evaluate all models on this alternate construction. We empirically estimate the distribution of bias scores across samples of a fixed proportion by using 100 independent random trials for that proportion. See Figure 5 for results. Observe that overlap in the distributions serves as a proxy for possible inversions in model ordering (by bias) depending on the subsample of template occupation words used. It is also worth noting that as we use more diverse sets (that is, bigger proportions) of seed words, the variance in the measured bias reduces." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 651, + 251, + 666 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 651, + 251, + 666 + ], + "spans": [ + { + "bbox": [ + 67, + 651, + 251, + 666 + ], + "type": "text", + "content": "G Tables of Experimental Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 673, + 291, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 673, + 291, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 673, + 291, + 714 + ], + "type": "text", + "content": "See Table 1 and Table 2 for detailed experimental results on alternate constructions for WINOGEN- DER and BIASNLI respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 724, + 205, + 739 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 724, + 205, + 739 + ], + "spans": [ + { + "bbox": [ + 67, + 724, + 205, + 739 + ], + "type": "text", + "content": "H Computing Resources" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "type": "text", + "content": "For our experiments, we used a 40-core Intel(R) Xeon(R) CPU E5-2640 v4 @ 2.40GHz, with access" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 71, + 526, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 137 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 137 + ], + "type": "text", + "content": "to NVIDIA RTX A6000 for selected experiments. In terms of runtime, compute time for inference on a single test set varied by model, but was limited to 12 hours for WINOGENDER and 72 hours for BIASNLI." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 148, + 464, + 161 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 148, + 464, + 161 + ], + "spans": [ + { + "bbox": [ + 302, + 148, + 464, + 161 + ], + "type": "text", + "content": "I Links to Datasets and Code" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 169, + 525, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 169, + 525, + 195 + ], + "spans": [ + { + "bbox": [ + 302, + 169, + 525, + 195 + ], + "type": "text", + "content": "All datasets (original constructions) used are publicly available." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 202, + 525, + 277 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 316, + 202, + 523, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 202, + 523, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 202, + 523, + 228 + ], + "type": "text", + "content": "- WINOGENDER:https://github.com/rudinger/ winogender-schemas" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 238, + 525, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 238, + 525, + 277 + ], + "spans": [ + { + "bbox": [ + 316, + 238, + 525, + 277 + ], + "type": "text", + "content": "- BIASNLI: https://github.com/sunipa/On-Measuring-and-Mitigating-Biased-Inferences-of-Word-Embeddings" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 303, + 285, + 496, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 285, + 496, + 298 + ], + "spans": [ + { + "bbox": [ + 303, + 285, + 496, + 298 + ], + "type": "text", + "content": "All models used are also publicly available." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 305, + 525, + 616 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 305, + 524, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 524, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 524, + 329 + ], + "type": "text", + "content": "ai2spanbert: https://demo.allennlp.org/coreference-resolution" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 339, + 524, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 339, + 524, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 339, + 524, + 365 + ], + "type": "text", + "content": "- UnifiedQA: https://github.com/allenai/unifiedqa" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 375, + 525, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 375, + 525, + 400 + ], + "spans": [ + { + "bbox": [ + 316, + 375, + 525, + 400 + ], + "type": "text", + "content": "- Longformer: https://github.com/shtoshni/fast-coref" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 409, + 525, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 409, + 525, + 435 + ], + "spans": [ + { + "bbox": [ + 316, + 409, + 525, + 435 + ], + "type": "text", + "content": "- Albert: https://huggingface.co/docs/trans formers/model_doc/albert" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 444, + 523, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 523, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 523, + 470 + ], + "type": "text", + "content": "- Elmo-DA:https://demo.allennlp.org/textual-entailment/elmo-snli" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 481, + 509, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 481, + 509, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 481, + 509, + 533 + ], + "type": "text", + "content": "- Roberta-base-SNLI:https://github.com/sunipa/OSCaR-Orthogonal-Subspace-Correction-and-Rectification/tree/transformer" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 542, + 524, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 542, + 524, + 581 + ], + "spans": [ + { + "bbox": [ + 316, + 542, + 524, + 581 + ], + "type": "text", + "content": "- Roberta-large-WANLI:https://huggingface.co/alisawuffles/roberta-large-wanli" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 591, + 518, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 591, + 518, + 616 + ], + "spans": [ + { + "bbox": [ + 316, + 591, + 518, + 616 + ], + "type": "text", + "content": "DistilRoberta:https://huggingface.co/cross-encoder/nli-distilroberta-base" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 302, + 624, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 624, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 624, + 525, + 772 + ], + "type": "text", + "content": "Code and data for the experiments are available at https://github.com/uclanlp/socialbias-dataset-construction-biases. We provide complete preprocessed datasets that correspond to the various proposed alternate constructions. They can be readily used with the publicly listed models for evaluation, thereby easily reproducing the results of the paper. We provide scripts to help with the same. The alternate dataset constructions can also be independently and flexibly used for new experiments." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 780, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 310, + 791 + ], + "type": "text", + "content": "1382" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 164, + 125, + 405, + 301 + ], + "blocks": [ + { + "bbox": [ + 164, + 125, + 405, + 301 + ], + "lines": [ + { + "bbox": [ + 164, + 125, + 405, + 301 + ], + "spans": [ + { + "bbox": [ + 164, + 125, + 405, + 301 + ], + "type": "image", + "image_path": "cc5727dc53e818d933db6906023ee2766c5e6424c74dcd0cbec85f18d6c3136a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 163, + 315, + 405, + 491 + ], + "blocks": [ + { + "bbox": [ + 163, + 315, + 405, + 491 + ], + "lines": [ + { + "bbox": [ + 163, + 315, + 405, + 491 + ], + "spans": [ + { + "bbox": [ + 163, + 315, + 405, + 491 + ], + "type": "image", + "image_path": "9880ec606b46af0206ad1697a3f11288379fda257eab996270fe78bb62a8c56f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 163, + 507, + 405, + 682 + ], + "blocks": [ + { + "bbox": [ + 163, + 507, + 405, + 682 + ], + "lines": [ + { + "bbox": [ + 163, + 507, + 405, + 682 + ], + "spans": [ + { + "bbox": [ + 163, + 507, + 405, + 682 + ], + "type": "image", + "image_path": "bf46fadd5a58b0be412b100e582b1d7268cd4467cb4291b83cb1b8c67ad9f98e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 692, + 525, + 729 + ], + "lines": [ + { + "bbox": [ + 67, + 692, + 525, + 729 + ], + "spans": [ + { + "bbox": [ + 67, + 692, + 525, + 729 + ], + "type": "text", + "content": "Figure 5: Bias measures (fraction neutral) computed on BIASNLI. The violin plot attempts to capture the distribution of bias measure scores across datasets reconstructed using different " + }, + { + "bbox": [ + 67, + 692, + 525, + 729 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 67, + 692, + 525, + 729 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 67, + 692, + 525, + 729 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 67, + 692, + 525, + 729 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 67, + 692, + 525, + 729 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 67, + 692, + 525, + 729 + ], + "type": "text", + "content": " subsets (top to bottom) of the occupation word list." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1383" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 117, + 190, + 478, + 304 + ], + "blocks": [ + { + "bbox": [ + 117, + 190, + 478, + 304 + ], + "lines": [ + { + "bbox": [ + 117, + 190, + 478, + 304 + ], + "spans": [ + { + "bbox": [ + 117, + 190, + 478, + 304 + ], + "type": "table", + "html": "
Perturbationai2spanbertqa-smallqa-baseqa-largelongformer
Baseline (no perturbations)5.835.8316.6615.419.16
Clause after occupation4.505.5014.7523.5010.08
Clause after participant10.338.0015.0015.758.83
Adjective before occupation8.225.3416.1217.316.87
Adjective after occupation4.925.3715.5725.459.75
Adjective before participant5.975.6913.8418.5210.77
Adjective after participant8.487.4915.9118.1711.69
Synonyms7.927.5017.9215.8312.08
", + "image_path": "165f3b523c6c34b9792ae3f902be020604ff8fea8e8db9e09dc1adef6c9c23da.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 71, + 574, + 526, + 627 + ], + "blocks": [ + { + "bbox": [ + 185, + 312, + 406, + 323 + ], + "lines": [ + { + "bbox": [ + 185, + 312, + 406, + 323 + ], + "spans": [ + { + "bbox": [ + 185, + 312, + 406, + 323 + ], + "type": "text", + "content": "Table 1: Percentage M-F Mismatch on WINOGENDER." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 71, + 574, + 526, + 627 + ], + "lines": [ + { + "bbox": [ + 71, + 574, + 526, + 627 + ], + "spans": [ + { + "bbox": [ + 71, + 574, + 526, + 627 + ], + "type": "table", + "html": "
AlbertElmo-DARoberta-base-SNLIRoberta-large-WANLIDistilRoberta
Baseline (no perturbations)44.8141.6415.2516.8151.32
Clauses60.8540.4330.2615.6960.84
Negation45.7613.4020.0410.4562.63
", + "image_path": "1cdf562287631cde54b0019e302488f2ba7926bd7360156a50b51994d75d8a82.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 142, + 634, + 450, + 647 + ], + "lines": [ + { + "bbox": [ + 142, + 634, + 450, + 647 + ], + "spans": [ + { + "bbox": [ + 142, + 634, + 450, + 647 + ], + "type": "text", + "content": "Table 2: Percentage neutral for different alternate constructions of BIASNLI" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1384" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 90, + 192, + 103 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 90, + 192, + 103 + ], + "spans": [ + { + "bbox": [ + 68, + 90, + 192, + 103 + ], + "type": "text", + "content": "A For every submission:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 106, + 317, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 106, + 317, + 121 + ], + "spans": [ + { + "bbox": [ + 77, + 106, + 317, + 121 + ], + "type": "text", + "content": "A1. Did you describe the limitations of your work?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 90, + 122, + 121, + 134 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 122, + 121, + 134 + ], + "spans": [ + { + "bbox": [ + 90, + 122, + 121, + 134 + ], + "type": "text", + "content": "Page 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 76, + 143, + 329, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 143, + 329, + 157 + ], + "spans": [ + { + "bbox": [ + 76, + 143, + 329, + 157 + ], + "type": "text", + "content": "A2. Did you discuss any potential risks of your work?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 158, + 208, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 158, + 208, + 169 + ], + "spans": [ + { + "bbox": [ + 89, + 158, + 208, + 169 + ], + "type": "text", + "content": "Not applicable. Left blank." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 77, + 178, + 414, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 178, + 414, + 192 + ], + "spans": [ + { + "bbox": [ + 77, + 178, + 414, + 192 + ], + "type": "text", + "content": "A3. Do the abstract and introduction summarize the paper's main claims?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 90, + 194, + 132, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 194, + 132, + 205 + ], + "spans": [ + { + "bbox": [ + 90, + 194, + 132, + 205 + ], + "type": "text", + "content": "Section 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 77, + 215, + 398, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 215, + 398, + 229 + ], + "spans": [ + { + "bbox": [ + 77, + 215, + 398, + 229 + ], + "type": "text", + "content": "A4. Have you used AI writing assistants when working on this paper?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 90, + 230, + 138, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 230, + 138, + 242 + ], + "spans": [ + { + "bbox": [ + 90, + 230, + 138, + 242 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 68, + 252, + 291, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 252, + 291, + 266 + ], + "spans": [ + { + "bbox": [ + 68, + 252, + 291, + 266 + ], + "type": "text", + "content": "B Did you use or create scientific artifacts?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 78, + 270, + 340, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 270, + 340, + 283 + ], + "spans": [ + { + "bbox": [ + 78, + 270, + 340, + 283 + ], + "type": "text", + "content": "Section 3 and Appendix J (Bias Datasets and Models used)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 77, + 291, + 315, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 291, + 315, + 306 + ], + "spans": [ + { + "bbox": [ + 77, + 291, + 315, + 306 + ], + "type": "text", + "content": "B1. Did you cite the creators of artifacts you used?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 89, + 306, + 327, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 306, + 327, + 319 + ], + "spans": [ + { + "bbox": [ + 89, + 306, + 327, + 319 + ], + "type": "text", + "content": "Section 3 and Appendix J (Datasets and Models used)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 77, + 327, + 463, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 327, + 463, + 342 + ], + "spans": [ + { + "bbox": [ + 77, + 327, + 463, + 342 + ], + "type": "text", + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 89, + 343, + 378, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 343, + 378, + 355 + ], + "spans": [ + { + "bbox": [ + 89, + 343, + 378, + 355 + ], + "type": "text", + "content": "Appendix J (Datasets and Models used are all publicly available)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 76, + 364, + 524, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 364, + 524, + 417 + ], + "spans": [ + { + "bbox": [ + 76, + 364, + 524, + 417 + ], + "type": "text", + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 90, + 419, + 208, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 419, + 208, + 432 + ], + "spans": [ + { + "bbox": [ + 90, + 419, + 208, + 432 + ], + "type": "text", + "content": "Not applicable. Left blank." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 76, + 441, + 524, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 441, + 524, + 481 + ], + "spans": [ + { + "bbox": [ + 76, + 441, + 524, + 481 + ], + "type": "text", + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 90, + 482, + 208, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 482, + 208, + 495 + ], + "spans": [ + { + "bbox": [ + 90, + 482, + 208, + 495 + ], + "type": "text", + "content": "Not applicable. Left blank." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 76, + 504, + 524, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 504, + 524, + 531 + ], + "spans": [ + { + "bbox": [ + 76, + 504, + 524, + 531 + ], + "type": "text", + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 90, + 532, + 208, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 532, + 208, + 544 + ], + "spans": [ + { + "bbox": [ + 90, + 532, + 208, + 544 + ], + "type": "text", + "content": "Not applicable. Left blank." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 77, + 553, + 524, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 553, + 524, + 622 + ], + "spans": [ + { + "bbox": [ + 77, + 553, + 524, + 622 + ], + "type": "text", + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 90, + 623, + 214, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 623, + 214, + 634 + ], + "spans": [ + { + "bbox": [ + 90, + 623, + 214, + 634 + ], + "type": "text", + "content": "Section 3.2 and Appendix F" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "spans": [ + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "type": "text", + "content": "C Did you run computational experiments?" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 79, + 662, + 122, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 662, + 122, + 673 + ], + "spans": [ + { + "bbox": [ + 79, + 662, + 122, + 673 + ], + "type": "text", + "content": "Section 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 77, + 682, + 524, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 682, + 524, + 711 + ], + "spans": [ + { + "bbox": [ + 77, + 682, + 524, + 711 + ], + "type": "text", + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 89, + 712, + 139, + 724 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 712, + 139, + 724 + ], + "spans": [ + { + "bbox": [ + 89, + 712, + 139, + 724 + ], + "type": "text", + "content": "Appendix I" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "spans": [ + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "type": "text", + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "spans": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "text", + "content": "ACL 2023 Responsible NLP Checklist" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1385" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 76, + 70, + 523, + 238 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 77, + 70, + 523, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 70, + 523, + 111 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 523, + 111 + ], + "type": "text", + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Section 3, Appendix B-G" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 77, + 120, + 523, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 120, + 523, + 174 + ], + "spans": [ + { + "bbox": [ + 77, + 120, + 523, + 174 + ], + "type": "text", + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Section 3, Appendix H" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 76, + 184, + 523, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 184, + 523, + 238 + ], + "spans": [ + { + "bbox": [ + 76, + 184, + 523, + 238 + ], + "type": "text", + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Not applicable. Left blank." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 67, + 247, + 522, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 247, + 522, + 278 + ], + "spans": [ + { + "bbox": [ + 67, + 247, + 522, + 278 + ], + "type": "text", + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 76, + 286, + 523, + 539 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 76, + 286, + 523, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 286, + 523, + 326 + ], + "spans": [ + { + "bbox": [ + 76, + 286, + 523, + 326 + ], + "type": "text", + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 76, + 336, + 523, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 336, + 523, + 390 + ], + "spans": [ + { + "bbox": [ + 76, + 336, + 523, + 390 + ], + "type": "text", + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 76, + 400, + 523, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 400, + 523, + 454 + ], + "spans": [ + { + "bbox": [ + 76, + 400, + 523, + 454 + ], + "type": "text", + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 76, + 462, + 519, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 462, + 519, + 489 + ], + "spans": [ + { + "bbox": [ + 76, + 462, + 519, + 489 + ], + "type": "text", + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 76, + 498, + 523, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 498, + 523, + 539 + ], + "spans": [ + { + "bbox": [ + 76, + 498, + 523, + 539 + ], + "type": "text", + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "text", + "content": "1386" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 13 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Theory-Grounded Computational Text Analysis/11a89724-ae4c-4578-823c-45401a11d43f_content_list.json b/2023/Theory-Grounded Computational Text Analysis/11a89724-ae4c-4578-823c-45401a11d43f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..484133c47aae17658cabd62c85596d7bd5337442 --- /dev/null +++ b/2023/Theory-Grounded Computational Text Analysis/11a89724-ae4c-4578-823c-45401a11d43f_content_list.json @@ -0,0 +1,1417 @@ +[ + { + "type": "text", + "text": "Theory-Grounded Computational Text Analysis", + "text_level": 1, + "bbox": [ + 248, + 90, + 749, + 111 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Arya D. McCarthy\\* and Giovanna Maria Dora Dore", + "bbox": [ + 255, + 139, + 742, + 159 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$\\diamond$ Center for Language and Speech Processing, Johns Hopkins University", + "bbox": [ + 206, + 160, + 794, + 178 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$\\spadesuit$ Krieger School of Arts and Sciences, Johns Hopkins University", + "bbox": [ + 238, + 179, + 763, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 252, + 339, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this position paper, we argue that computational text analysis lacks and requires organizing principles. A broad space separates its two constituent disciplines—natural language processing and social science—which has to date been sidestepped rather than filled by applying increasingly complex computational models to problems in social science research. We contrast descriptive and integrative findings, and our review of approximately 60 papers on computational text analysis reveals that those from *ACL venues are typically descriptive. The lack of theory began at the area's inception and has, over the decades, grown more important and challenging. A return to theoretically grounded research questions will propel the area from both theoretical and methodological points of view.", + "bbox": [ + 141, + 280, + 460, + 521 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 114, + 546, + 260, + 562 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Computational text analysis methods—an umbrella combining natural language processing with social science—are in a honeymoon period (Lazer and Radford, 2017; van Atteveldt and Peng, 2018). Today's social scientist might reach for the tools of computer science for their speed, scale, granularity, and consistency; for instance, natural language processing offers \"to analyze signals ranging from simple lexical cues to word clusters to choices of syntactic structure\" (Boydstun et al., 2014). The numerical outputs tell a story that is simple, easy to make sense of, and in that regard comforting. Conversely, today's computer scientist may see the problems of social science as answerable by objectivity and reductionism, eschewing interpretation for quantitative analysis.", + "bbox": [ + 112, + 571, + 489, + 829 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The conclusion of this reasoning, and the dominant stance in computational social science, is a reliance on machines alone to answer questions in the field, surrendering to their supposed objectivity", + "bbox": [ + 112, + 829, + 489, + 894 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "or impartiality. Can a machine's output go beyond descriptive catalogs of evidence, accelerating understanding of processes and motivations? From our experience, computers are nowhere near supplanting humans in interpreting social science results.1", + "bbox": [ + 507, + 253, + 884, + 332 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "An interdisciplinary inquiry must go farther than matching computational techniques to social science questions (O'Connor et al., 2011; Nguyen et al., 2020). It embraces synergistic methodology and connects the norms and standards of evidence from both. This means partnering computer science's preference for the structured, generalizable, and objective with the unstructured, critical, and contextual which the social sciences champion. This level of interdisciplinarity addresses the question raised by descriptive findings: So what?", + "bbox": [ + 507, + 332, + 884, + 508 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We see theory as the solution, empowering rather than shackling investigations. What this paper advocates is not one particular theory—certainly these are myriad, and “even subject matter which has been under intensive and prolonged study remains at the unsettled periphery of research” (Nagel, 1963). Instead, we expand on our prior work (Dore and McCarthy, 2022) to clarify calls echoed for decades by computational and social science (McDermott, 1976; Jelinek, 2005; Hajic and Hajicova, 2007; Hofman et al., 2018; Lipton and Steinhardt, 2019; Baden et al., 2021). Underlying each, we find, is the urge to return to theory, which we espouse herein.", + "bbox": [ + 507, + 510, + 884, + 719 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Description vs. Integration", + "text_level": 1, + "bbox": [ + 507, + 731, + 774, + 747 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We contrast descriptive findings and theoretical analysis. An example of a descriptive finding is that an apple falls, or that it falls faster when pushed than dropped, or even that it falls at a particular rate estimated with some standard error by a complex", + "bbox": [ + 507, + 756, + 884, + 835 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1See, e.g., Noam Chomsky's remark on GPT-3: \"You can't go to a physics conference and say: I've got a great theory. It accounts for everything and is so simple it can be captured in two words: 'Anything goes.' All known and unknown laws of nature are accommodated... Of course, everything impossible is accommodated also. That's GPT-3.\" [link]", + "bbox": [ + 507, + 845, + 884, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution.", + "bbox": [ + 141, + 903, + 272, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1586", + "bbox": [ + 480, + 927, + 522, + 940 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics", + "bbox": [ + 226, + 945, + 771, + 957 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Volume 2: Short Papers, pages 1586-1594", + "bbox": [ + 368, + 958, + 630, + 971 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "July 9-14, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 295, + 972, + 700, + 985 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "interpolation. A theoretical analysis of the same phenomenon, credited to Newton, is that a fundamental force acts upon the apple, and that this same force governs the motion of the heavens. The theoretical analysis links the finding about the world critically to a broader body of knowledge and context.", + "bbox": [ + 112, + 84, + 487, + 180 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite advances in causal inference in NLP, the descriptive is all that a machine can provide to the social sciences (Feder et al., 2021). Certainly the methods of computational text analysis have advanced since the General Inquirer (Stone and Hunt, 1963) and Mosteller and Wallace's statistical inference of text authorship (1963). But methods are means, not ends. They uncover more descriptive findings in data: the rate of an apple's fall, the topics of refugees' tweets (Walk et al., 2022), the space given to marginalized groups in textbooks (Lucy et al., 2020), or patterns of state censorship (Bamman et al., 2012; King et al., 2013).", + "bbox": [ + 115, + 181, + 489, + 390 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The foils to descriptive findings are integrative findings (Hofman et al., 2021), which offer causal explanations that enable future predictions—a theory, or as a 'model' in the sense of the Standard Model, rather than of a statistical model. Integrative findings can either offer new theories or couch their explanations in existing theories—but the theory is essential either way.", + "bbox": [ + 110, + 392, + 489, + 521 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3 We Don't Integrate", + "text_level": 1, + "bbox": [ + 112, + 533, + 315, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To contrast descriptive and integrative findings, we reviewed approximately 60 papers in computational text analysis published in *ACL venues. In Table 1, we describe several of these in terms of their descriptive or theory-grounded contributions. $^{2}$ Descriptive papers may refer to social science theories or make generalizable claims, as when Demszky et al. (2019) write, \"The shooter's race appears to play a role in topic preference: if the shooter is white, Democrats become more likely to focus on shooter's identity,\" but they do not link to the two to each other.", + "bbox": [ + 112, + 561, + 489, + 738 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "An excellent theory-grounded quantitative work is Nelson (2021); she confirms some of the most compelling features of identity theory, specifically that identities based on race were most distinguished by cultural discourse, whereas those based on gender by the domestic and the economic discourse. Similarly, we conducted theory-grounded quantitative work to investigate the application of the protest", + "bbox": [ + 112, + 740, + 489, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "paradigm and thematic framing in how western- and Hong Kong based newspapers portray protests in Hong Kong (McCarthy et al., 2021; McCarthy and Dore, 2022). Generally, it remains challenging to find computational social science papers in *ACL venues that go beyond description and prediction, advancing theory. Why is this? We believe it stemmed from the field's \"empirical turn\".3", + "bbox": [ + 507, + 84, + 884, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Few remember when the meetings of ACL offered a few dozen papers, all entrenched in formalisms and linguistic theories. Arguably, 1996 was a turning point when the founders of SIGDAT held the first EMNLP at Penn under the auspices of the ACL. This gave a spotlight to the few but growing empiricists in the field and drew in more.", + "bbox": [ + 507, + 214, + 884, + 326 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "EMNLP began a half-decade of measurable reorganization the field (Anderson et al., 2012). That EMNLP remains affiliated with ACL keeps the language-focused machine learning practitioners in our tent. The slow blurring of boundaries between each *ACL conference's expectations (Church, 2020) increases this unity. Both groups belong under this tent. But without a doubt, one group's voice is becoming less heard.", + "bbox": [ + 507, + 329, + 884, + 472 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Publication venues within the ACL focus on methods over theory.5 Techniques are taken off the shelf without critical examination because these are \"the best\" (often \"state of the art\") for their purposes (Ethayarajh and Jurafsky, 2020). This widens the gap between theoretical and empirical work.6 Hopkins and King (2010) claim, \"computer scientists may be interested in finding the needle in the haystack... social scientists are more commonly interested in characterizing the haystack\"—evincing the value of broader context.7 Wallach (2018), quoting Hopkins and King, explains that the two groups", + "bbox": [ + 507, + 475, + 882, + 668 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "A lesser reason is the challenge of serving two masters: adequately covering both the theoretical and methodological components within 8 pages. We recently received two reviews for an *ACL submission: one advocating for more of the social science context in the main text by eschewing methods to the appendix, and the other instructing us to do the opposite.", + "bbox": [ + 507, + 683, + 882, + 755 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "4 And its predecessor the Workshop on Very Large Corpora.", + "bbox": [ + 532, + 756, + 880, + 770 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "This is due to the outsized influence of computer science, often seen as the science of method (Hoare and Jones, 1989; Shapiro, 2001), when not instead seen as an engineering discipline (Rapaport, 2005).", + "bbox": [ + 507, + 770, + 882, + 819 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "6 A related criticism is that empirical research has narrowed to focus on 'easy' questions that its tools can address (Coleman, 1986; Baden et al., 2021), especially when research questions are baked into the design of the task.", + "bbox": [ + 507, + 819, + 882, + 868 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "7As evidence, see Siegel (2018): \"We usually don't know about causation, and we often don't necessarily care... the objective is more to predict than it is to understand the world... It just needs to work; prediction trumps explanation.\"", + "bbox": [ + 507, + 868, + 882, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "$^{2}$ Following Lipton and Steinhardt (2019), we only describe papers by established researchers to \"avoid singling out junior students... who lack the opportunity to reply symmetrically\".", + "bbox": [ + 112, + 879, + 487, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1587", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/26ad4830629417904c7c3f2db1c2e350e95681b63cb8bd1424c2dc4341fdbb11.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Descriptive
Chang et al. (2009)The article presents new quantitative methods to measure semantic meaning in inferred topics. The authors emphasize the qualitative relevance of their findings as it validates the use of topics for corpus exploration and information retrieval. However, their working hypothesis and empirical findings are not connected to the extremely relevant field of communication theory.
Bamman et al. (2012)The article presents the first large-scale analysis of political content censorship in social media. The authors miss the opportunity to relate their hypothesis and findings to censorship theory, a natural theoretical context for the research, which would strengthen the relevance and generalizability of the findings.
Field et al. (2018)The article discusses media manipulation in Russia in the context of agenda-setting and framing, the tools that Russian state-owned (or heavily influenced) media outlets use to distract public attention from domestic economic politics. The authors implicitly refer to propaganda theory and autocratic theory throughout the article even though their findings are not discussed in relation to these theories.
Demszky et al. (2019)The article applies “a more comprehensive NLP framework to study linguistic aspects of polarization in social media”. While the article implicitly refers to theories of social conformity and social conflict, the findings are not linked or discussed (either explicitly or implicitly) to the theoretical frameworks that the authors touch on in their §1.
Integrative
DiMaggio et al. (2013)The article describes how topic models of newspaper articles help to study the politicization of government support for arts organizations and artists in the late 1980s in the US. The authors clearly define the theoretical context of their investigation and emphasize the relationship between theory and method throughout the paper.
Bamman et al. (2014)The article validates an empirical model that “employs multiple effects to account for the influence of extra-linguistic information (such as author)” by testing specific parameters against a variety of theory-based hypotheses derived from writing styles theories of England between 1700 and 1899.
Nelson (2021)The article argues that the full potential of machine learning can be better realized by “leveraging the epistemological alignment between machine learning and inductive research.” The author empirically demonstrates this by anchoring in identity theory a word embedding model of first-person narratives of the nineteenth-century U.S. South.
", + "bbox": [ + 117, + 82, + 878, + 481 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1: Contrast between work in computational text analysis with descriptive findings versus integrative findings.", + "bbox": [ + 112, + 491, + 880, + 506 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "are interested in very different research questions, and that computational social science must be more than computer science with social data; it must strive for valid explanatory models. In the same vein, at ACL 2022, ACL fellow Eduard Hovy remarked that NLP must be more than \"just machine learning on corpora\".", + "bbox": [ + 112, + 531, + 487, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Social scientists are also coming to terms with the meaning of computational techniques applied more often in social science (Bail, 2014; Biernacki, 2015; Lee and Martin, 2015; Spillman, 2015). The focus of the debates, however, is on which methods are best suited to extract meaning from text, without addressing any theoretical considerations related to the methods or whether a theoretical framework for those methods even exists. The discussions on whether computational methods make social science research more efficient, reliable, and reproducible overtake attempts at theory-building.", + "bbox": [ + 112, + 646, + 487, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4 Moving Forward", + "text_level": 1, + "bbox": [ + 112, + 856, + 295, + 873 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We are not denying the value of computational approaches to analyzing text. Certainly, comput", + "bbox": [ + 112, + 887, + 489, + 919 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ing can be an instrumental approach for modeling and understanding social complexity. This does not mean that other approaches, such as historical, ethnographic, or mathematical, become irrelevant. On the contrary, computational methods necessarily (whether awarely or not) rely on these earlier approaches to add value, in terms of improving our explanations and understanding (Radford and Joseph, 2020).", + "bbox": [ + 507, + 531, + 884, + 675 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As we are a field that prioritizes methods, consider the seminal book on methods in science: Abbott (2004) taxonomizes scientific ways of knowing. Its five broad categories are ethnography, historical narration, standard causal analysis, small- $N$ comparison, and formal modeling. We in NLP myopically choose the third and fifth of these, ignoring the value of the others. But the broader point of Methods of Discovery is not methods. It is the research question. Any methodology should be grounded in the question, not incremental tweaks and reviewers' comfort (Church, 2020). This admits even qualitative or mixed-method approaches to text analysis.", + "bbox": [ + 507, + 677, + 884, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The role of humans in scientific inquiry is nothing", + "bbox": [ + 527, + 903, + 880, + 919 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1588", + "bbox": [ + 482, + 927, + 519, + 940 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "new. Using qualitative analysis to complement quantitative techniques has its roots in Achen and Snidal (1989)'s recommendation to use historical case studies as a complement to statistical research. Their plea was strengthened by Verba's work in the early 1990s (Verba et al., 1993, 1995; Verba, 1996) and Tarrow (1995), who openly called for bridging qualitative and quantitative modes of research in social science. In doing so, they have enriched the field with critical methodological innovations (Gerring, 2004), benefiting from the recognition that \"quantitative methods must augment humans, not replace them\" (Grimmer and Stewart, 2013, 4).", + "bbox": [ + 110, + 84, + 487, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The field can draw more from social science's rich tradition of inductive theory-building and interpretation to develop its theoretical approach—to prize either induction or deduction alone is a myth of scientific procedure (Thagard, 1988), but the melding of the two opens new doors. Rather than eschewing the complexity (a criticism leveled by Baden et al., 2021), it should put complexity at the center of its ontology on the basis that there are no immutable laws in social life or optimal solutions to social problems.", + "bbox": [ + 110, + 294, + 487, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Skepticism can linger toward findings not drawn from the standard practices of one's own field; indeed, social science was long skeptical of computational contributions (Armstrong, 1967). We believe that this drives the hyperfocus on improving a few accepted methods instead of exploring more broadly. If the doorway between disciplines is only narrowly open, this reflects a lack of appreciation for each field's ways of knowing. The disciplinary divide keeps computational researchers from embracing methods beyond standard causal analysis or formal modeling, so the interpreter-centric richness allowed by histories, ethnographies, and small- $N$ exploration are precluded.", + "bbox": [ + 110, + 470, + 489, + 697 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 112, + 709, + 247, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We have explained the distinction between descriptive and theoretical findings as it pertains to computational text analysis. The bulk of work we found provided vast descriptive findings, often of high quality, but not giving back to questions of theory. We offer several suggestions on how to 'push the pendulum back' by prioritizing theory-building or", + "bbox": [ + 110, + 734, + 489, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "theory-affirming research questions and accepting whichever methods are best suited toward answering it—not only the familiar and entrenched ones.", + "bbox": [ + 507, + 84, + 880, + 131 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We are not the first to advocate for a shift in the patterns of applying computational techniques to real-world problems. There is a steady drumbeat from voices in the field advocating careful approaches (Nagel, 1963; McDermott, 1976; Jelinek, 2005; Hajic and Hajicova, 2007; Hofman et al., 2018; Lipton and Steinhardt, 2019; Baden et al., 2021). What we see underlying all of these—those writing against 'mathiness' and speculation, advocating for clear evaluation over anecdotes, criticizing textual researchers' dilution of conceptual standards, highlighting work that ties linguistic information into complex models—is an unspoken, perhaps unrealized, call for a return to theory.", + "bbox": [ + 507, + 131, + 884, + 356 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Not only do we aver that incorporating theory is essential; but also, other fields have strengthened themselves when espousing organizing principles beyond those of their progenitors. Behavioral economics is a success story here. It transcended the neat (but psychosocially stripped) mathematics it draws from to acknowledge deviations from rationality and blend economics with cognitive science (Kahneman and Tversky, 1979; Thaler, 1980; Thaler and Sunstein, 2009).", + "bbox": [ + 507, + 357, + 882, + 516 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For scientific—not simply engineering—advances to arise from the *ACL community, authors and reviewers alike must resist the temptation toward incremental, ‘safe’ research and follow Church (2005): “Controversial papers are great; boring unobjectionable incremental papers are not.” In reviewing new research, we should privilege not only work that presents new and unusual computational methods, but also interactions between computational and humanistic approaches to answering research questions. EMNLP was founded because of reviewing biases at ACL against groundbreaking methodological advances, and since then the two have homogenized; “EMNLP reviewing is no longer much of a differentiator” (Church, 2020). We found that theoretically grounded findings in text analysis are often published in non-\\*ACL venues (Table 1), but ACL sets the standard for work involving computational text analysis and NLP. Is there no home for groundbreaking integrative or interdisciplinary work in *ACL, such that a new venue is required? Or can we adapt our standards to invite deeper connections to theory and new ways of knowing?", + "bbox": [ + 507, + 518, + 882, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "8 Expertise plays a role as well (Shing et al., 2018), which is why Mechanical Turk doesn't fill the need for qualitative analysis. This is exemplified by Radford and Joseph (2020)'s observation of \"non-expert annotators providing unreliable annotations, even after a discussion period\".", + "bbox": [ + 112, + 854, + 487, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1589", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 114, + 84, + 278, + 101 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This publication was made possible in part by a grant from the American Political Science Association to A.D.M. and G.M.D.D. The statements made and views expressed are solely the responsibility of the authors. A.D.M. is supported by an Amazon Fellowship and a Frederick Jelinek Fellowship.", + "bbox": [ + 112, + 109, + 489, + 206 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Limitations", + "text_level": 1, + "bbox": [ + 114, + 217, + 220, + 231 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The key limitation of our work is that, when conducting the review of approximately 60 papers (by searching through the ACL Anthology for works in computational social science since 2010), we encountered a skewed distribution of descriptive versus integrative works. In fact, it was relatively simple to find descriptive works, and that section of Table 1 could have been much longer. We also recognize that, due to the mixed nature of our field, scientific and integrative findings are not the only goal—our 'big tent' includes engineers as well, who value gains in performance indicators. Finally, the fact that we have few examples of papers showing a return to theory renders the possibility that our central claim is misinterpreted in a normative way as a mandate.", + "bbox": [ + 112, + 243, + 489, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 114, + 527, + 213, + 542 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Andrew Delano Abbott. 2004. Methods of discovery: Heuristics for the social sciences (contemporary societies). WW Norton & Company.", + "Christopher H. Achen and Duncan Snidal. 1989. Rational deterrence theory and comparative case studies. World Politics, 41(2):143-169.", + "Ashton Anderson, Dan Jurafsky, and Daniel A. McFarland. 2012. Towards a computational history of the ACL: 1980-2008. In Proceedings of the ACL-2012 Special Workshop on Rediscovering 50 Years of Discoveries, pages 13-21, Jeju Island, Korea. Association for Computational Linguistics.", + "J. Scott Armstrong. 1967. Derivation of theory by means of factor analysis or Tom Swift and his electric factor analysis machine. The American Statistician, 21(5):17-21.", + "Christian Baden, Christian Pipal, Martijn Schoonvelde, and Mariken A. C. G van der Velden. 2021. Three gaps in computational text analysis methods for social sciences: A research agenda. Communication Methods and Measures, 0(0):1-18.", + "Christopher A. Bail. 2014. The cultural environment: measuring culture with big data. Theory and Society, 43(3):465-482." + ], + "bbox": [ + 115, + 550, + 489, + 917 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "David Bamman, Brendan O'Connor, and Noah Smith. 2012. Censorship and deletion practices in chinese social media. First Monday, 17(3).", + "David Bamman, Ted Underwood, and Noah A. Smith. 2014. A Bayesian mixed effects model of literary character. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 370-379, Baltimore, Maryland. Association for Computational Linguistics.", + "Richard Biernacki. 2015. How to do things with historical texts. American Journal of Cultural Sociology, 3(3):311-352. Copyright - © Palgrave Macmillan, a division of Macmillan Publishers Ltd 2015; Last updated - 2018-09-25.", + "Amber E Boydstun, Dallas Card, Justin Gross, Paul Resnick, and Noah A Smith. 2014. Tracking the development of media frames within and across policy issues. Unpublished.", + "Jonathan Chang, Jordan Boyd-Graber, Sean Gerrish, Chong Wang, and David M. Blei. 2009. Reading tea leaves: How humans interpret topic models. In Proceedings of the 22nd International Conference on Neural Information Processing Systems, NIPS'09, page 288-296, Red Hook, NY, USA. Curran Associates Inc.", + "Kenneth Church. 2005. Last words: Reviewing the reviewers. Computational Linguistics, 31(4):575-578.", + "Kenneth Ward Church. 2020. Emerging trends: Reviewing the reviewers (again). Natural Language Engineering, 26(2):245-257.", + "James S. Coleman. 1986. Social theory, social research, and a theory of action. American Journal of Sociology, 91(6):1309-1335.", + "Dorottya Demszky, Nikhil Garg, Rob Voigt, James Zou, Jesse Shapiro, Matthew Gentzkow, and Dan Jurafsky. 2019. Analyzing polarization in social media: Method and application to tweets on 21 mass shootings. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2970-3005, Minneapolis, Minnesota. Association for Computational Linguistics.", + "Paul DiMaggio, Manish Nag, and David Blei. 2013. Exploiting affinities between topic modeling and the sociological perspective on culture: Application to newspaper coverage of U.S. government arts funding. *Poetics*, 41(6):570–606. Topic Models and the Cultural Sciences.", + "Giovanna Maria Dora Dore and Arya D. McCarthy. 2022. Learning to play with the machines in social science research: Bringing the theory back in. In ICML 2022 Workshop on Human-Machine Collaboration and Teaming, Baltimore, Maryland." + ], + "bbox": [ + 510, + 85, + 884, + 917 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1590", + "bbox": [ + 482, + 928, + 521, + 940 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Kawin Ethayarajh and Dan Jurafsky. 2020. Utility is in the eye of the user: A critique of NLP leaderboards. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4846-4853, Online. Association for Computational Linguistics.", + "Amir Feder, Katherine A. Keith, Emaad Manzoor, Reid Pryzant, Dhanya Sridhar, Zach Wood-Doughty, Jacob Eisenstein, Justin Grimmer, Roi Reichart, Margaret E. Roberts, Brandon M. Stewart, Victor Veitch, and Diyi Yang. 2021. Causal inference in natural language processing: Estimation, prediction, interpretation and beyond. CoRR, abs/2109.00725.", + "Anjalie Field, Doron Kliger, Shuly Wintner, Jennifer Pan, Dan Jurafsky, and Yulia Tsvetkov. 2018. Framing and agenda-setting in Russian news: a computational analysis of intricate political strategies. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 3570-3580, Brussels, Belgium. Association for Computational Linguistics.", + "John Gerring. 2004. What is a case study and what is it good for? American Political Science Review, 98(2):341-354.", + "Justin Grimmer and Brandon M. Stewart. 2013. Text as data: The promise and pitfalls of automatic content analysis methods for political texts. Political Analysis, 21(3):267-297.", + "Jan Hajic and Eva Hajičová. 2007. Some of our best friends are statisticians. In Text, Speech and Dialogue, pages 2-10, Berlin, Heidelberg. Springer Berlin Heidelberg.", + "C. A. R. Hoare and C. B. Jones. 1989. Essays in Computing Science. Prentice-Hall, Inc., USA.", + "Jake Hofman, Miro Dudík, and Daniel G. Goldstein. 2018. Perspective annotation for numerical representations. United States Patent Application.", + "Jake M Hofman, Duncan J Watts, Susan Athey, Filiz Garip, Thomas L Griffiths, Jon Kleinberg, Helen Margetts, Sendhil Mullainathan, Matthew J Salganik, Simine Vazire, et al. 2021. Integrating explanation and prediction in computational social science. Nature, 595(7866):181-188.", + "Daniel J. Hopkins and Gary King. 2010. A method of automated nonparametric content analysis for social science. American Journal of Political Science, 54(1):229-247.", + "Frederick Jelinek. 2005. Some of my best friends are linguists. Language Resources and Evaluation, 39(1):25-34.", + "Daniel Kahneman and Amos Tversky. 1979. Prospect theory: An analysis of decision under risk. *Econometrica*, 47(2):263-291." + ], + "bbox": [ + 115, + 85, + 489, + 917 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Gary King, Jennifer Pan, and Margaret E. Roberts. 2013. How censorship in China allows government criticism but silences collective expression. American Political Science Review, 107(2 (May)):1-18.", + "David Lazer and Jason Radford. 2017. Data ex machina: Introduction to big data. Annual Review of Sociology, 43(1):19-39.", + "Monica Lee and John Levi Martin. 2015. Coding, counting and cultural cartography. American Journal of Cultural Sociology, 3(1):1-33.", + "Zachary C. Lipton and Jacob Steinhardt. 2019. Troubling trends in machine learning scholarship: Some ML papers suffer from flaws that could mislead the public and stymie future research. Queue, 17(1):45-77.", + "Li Lucy, Dorottya Demszky, Patricia Bromley, and Dan Jurafsky. 2020. Content analysis of textbooks via natural language processing: Findings on gender, race, and ethnicity in Texas U.S. history textbooks. AERA Open, 6(3):2332858420940312.", + "Arya D. McCarthy and Giovanna Maria Dora Dore. 2022. Hong Kong: Longitudinal and synchronic characterisations of protest news between 1998 and 2020. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 2891-2900, Marseille, France. European Language Resources Association.", + "Arya D. McCarthy, James Scharf, and Giovanna Maria Dora Dore. 2021. A mixed-methods analysis of western and Hong Kong-based reporting on the 2019-2020 protests. In Proceedings of the 5th Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature, pages 178-188, Punta Cana, Dominican Republic (online). Association for Computational Linguistics.", + "Drew McDermott. 1976. Artificial intelligence meets natural stupidity. SIGART Bull., (57):4-9.", + "Frederick Mosteller and David L. Wallace. 1963. Inference in an authorship problem. Journal of the American Statistical Association, 58(302):275-309.", + "Ernest Nagel. 1963. The structure of science: Problems in the logic of scientific explanation. Mind, 72(287).", + "Laura K. Nelson. 2021. Leveraging the alignment between machine learning and intersectionality: Using word embeddings to measure intersectional experiences of the nineteenth century U.S. South. Poetics, 88:101539. Measure Mohr Culture.", + "Dong Nguyen, Maria Liakata, Simon DeDeo, Jacob Eisenstein, David Mimno, Rebekah Tromble, and Jane Winters. 2020. How we do things with words: Analyzing text as social and cultural data. Frontiers in Artificial Intelligence, 3." + ], + "bbox": [ + 510, + 85, + 880, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1591", + "bbox": [ + 482, + 928, + 517, + 940 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Brendan O'Connor, David Bamman, and Noah A Smith. 2011. Computational text analysis for social science: Model complexity and assumptions. In Proc. of the NIPS Workshop on Comptuational Social Science and the Wisdom of Crowds.", + "Jason Radford and Kenneth Joseph. 2020. Theory in, theory out: The uses of social theory in machine learning for social science. Frontiers in Big Data, 3.", + "William J Rapaport. 2005. Philosophy of computer science: An introductory course. Teaching philosophy, 28(4):319-341.", + "Stuart C. Shapiro. 2001. Computer science: The study of procedures. Technical report, Department of Computer Science and Engineering, University of Buffalo.", + "Han-Chin Shing, Suraj Nair, Ayah Zirikly, Meir Friedenberg, Hal Daumé III, and Philip Resnik. 2018. Expert, crowdsourced, and machine assessment of suicide risk via online postings. In Proceedings of the Fifth Workshop on Computational Linguistics and Clinical Psychology: From Keyboard to Clinic, pages 25-36, New Orleans, LA. Association for Computational Linguistics.", + "David A. Siegel. 2018. Analyzing computational models. American Journal of Political Science, 62(3):745-759.", + "Lyn Spillman. 2015. Ghosts of straw men: A reply to Lee and Martin. American Journal of Cultural Sociology, 3(3):365-379.", + "Philip J. Stone and Earl B. Hunt. 1963. A computer approach to content analysis: Studies using the general inquirer system. In Proceedings of the May 21-23, 1963, Spring Joint Computer Conference, AFIPS '63 (Spring), page 241-256, New York, NY, USA. Association for Computing Machinery.", + "Sidney Tarrow. 1995. Bridging the quantitative-qualitative divide in political science. American Political Science Review, 89(2):471-474.", + "Paul Thagard. 1988. Computational Philosophy of Science. MIT Press.", + "Richard Thaler. 1980. Judgement And Decision Making Under Uncertainty: What Economists Can Learn From Psychology. Risk Analysis in Agriculture: Research and Educational Developments, January 16-18, 1980, Tucson, Arizona 271572, Regional Research Projects $>$ W-149: An Economic Evaluation of Managing Market Risks in Agriculture.", + "Richard H. Thaler and Cass R. Sunstein. 2009. Nudge: Improving decisions about health, wealth, and happiness. Penguin.", + "Wouter van Atteveldt and Tai-Quan Peng. 2018. When communication meets computation: Opportunities, challenges, and pitfalls in computational communication science. Communication Methods and Measures, 12(2-3):81-92." + ], + "bbox": [ + 115, + 85, + 485, + 917 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Sidney Verba. 1996. The citizen as respondent: Sample surveys and american democracy. American Political Science Review, 90(1):1-7. Presidential Address, American Political Science Association, 1995.", + "Sidney Verba, Kay Lehman Schlozman, Henry Brady, and Norman H. Nie. 1993. Citizen activity: Who participates? what do they say? The American Political Science Review, 87(2):303-318.", + "Sidney Verba, Kay Lehman Schlozman, and Henry E Brady. 1995. Voice and equality: Civic volunteerism in American politics. Harvard University Press.", + "Erin Walk, Elizabeth Parker-Magyar, Kiran Garimella, Ahmet Akbiyik, and Fotini Christia. 2022. Social media narratives across platforms in conflict: Evidence from Syria. MIT Political Science Department Research Paper No. 2022-2, available at SSRN.", + "Hanna Wallach. 2018. Computational social science $\\neq$ computer science + social data. Commun. ACM, 61(3):42-44." + ], + "bbox": [ + 510, + 85, + 880, + 375 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1592", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A For every submission:", + "bbox": [ + 115, + 107, + 322, + 122 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A1. Did you describe the limitations of your work?", + "bbox": [ + 129, + 126, + 532, + 143 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Unnumbered; appears on page 5.", + "bbox": [ + 152, + 143, + 400, + 159 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A2. Did you discuss any potential risks of your work?", + "bbox": [ + 129, + 170, + 552, + 186 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This is a position paper.", + "bbox": [ + 152, + 187, + 329, + 202 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A3. Do the abstract and introduction summarize the paper's main claims?", + "bbox": [ + 129, + 212, + 695, + 229 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 152, + 230, + 231, + 244 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A4. Have you used AI writing assistants when working on this paper?", + "bbox": [ + 129, + 255, + 668, + 272 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 152, + 273, + 231, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "B Did you use or create scientific artifacts?", + "bbox": [ + 114, + 300, + 487, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 132, + 321, + 213, + 336 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "B1. Did you cite the creators of artifacts you used?", + "bbox": [ + 129, + 347, + 529, + 363 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Not applicable. Left blank.", + "bbox": [ + 152, + 363, + 349, + 379 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?", + "bbox": [ + 129, + 390, + 778, + 406 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Not applicable. Left blank.", + "bbox": [ + 152, + 407, + 349, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?", + "bbox": [ + 129, + 432, + 880, + 495 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Not applicable. Left blank.", + "bbox": [ + 152, + 497, + 349, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?", + "bbox": [ + 129, + 524, + 880, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Not applicable. Left blank.", + "bbox": [ + 152, + 573, + 349, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?", + "bbox": [ + 129, + 599, + 880, + 631 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Not applicable. Left blank.", + "bbox": [ + 152, + 632, + 349, + 646 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be.", + "bbox": [ + 129, + 658, + 880, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Not applicable. Left blank.", + "bbox": [ + 152, + 740, + 349, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "C Did you run computational experiments?", + "bbox": [ + 114, + 765, + 492, + 781 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 132, + 785, + 213, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?", + "bbox": [ + 129, + 813, + 880, + 845 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Not applicable. Left blank.", + "bbox": [ + 152, + 846, + 349, + 860 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.", + "bbox": [ + 112, + 868, + 877, + 892 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "ACL 2023 Responsible NLP Checklist", + "bbox": [ + 132, + 84, + 433, + 99 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1593", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?", + "bbox": [ + 127, + 84, + 878, + 115 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Not applicable. Left blank.", + "bbox": [ + 149, + 117, + 349, + 131 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?", + "bbox": [ + 127, + 143, + 882, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Not applicable. Left blank.", + "bbox": [ + 149, + 192, + 349, + 206 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?", + "bbox": [ + 127, + 218, + 882, + 265 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Not applicable. Left blank.", + "bbox": [ + 149, + 267, + 349, + 282 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?", + "bbox": [ + 112, + 293, + 877, + 310 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 132, + 313, + 213, + 329 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?", + "bbox": [ + 127, + 340, + 882, + 372 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Not applicable. Left blank.", + "bbox": [ + 149, + 373, + 349, + 388 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?", + "bbox": [ + 127, + 399, + 882, + 447 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Not applicable. Left blank.", + "bbox": [ + 149, + 448, + 349, + 464 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?", + "bbox": [ + 127, + 475, + 882, + 521 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Not applicable. Left blank.", + "bbox": [ + 149, + 523, + 349, + 539 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?", + "bbox": [ + 127, + 549, + 875, + 565 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Not applicable. Left blank.", + "bbox": [ + 149, + 565, + 349, + 582 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?", + "bbox": [ + 127, + 592, + 882, + 623 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Not applicable. Left blank.", + "bbox": [ + 149, + 625, + 349, + 640 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1594", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 8 + } +] \ No newline at end of file diff --git a/2023/Theory-Grounded Computational Text Analysis/11a89724-ae4c-4578-823c-45401a11d43f_model.json b/2023/Theory-Grounded Computational Text Analysis/11a89724-ae4c-4578-823c-45401a11d43f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a92f128e2271e26483893ee176e83d1378d0e668 --- /dev/null +++ b/2023/Theory-Grounded Computational Text Analysis/11a89724-ae4c-4578-823c-45401a11d43f_model.json @@ -0,0 +1,2011 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.25, + 0.091, + 0.75, + 0.112 + ], + "angle": 0, + "content": "Theory-Grounded Computational Text Analysis" + }, + { + "type": "text", + "bbox": [ + 0.257, + 0.14, + 0.744, + 0.16 + ], + "angle": 0, + "content": "Arya D. McCarthy\\* and Giovanna Maria Dora Dore" + }, + { + "type": "text", + "bbox": [ + 0.208, + 0.161, + 0.796, + 0.179 + ], + "angle": 0, + "content": "\\(\\diamond\\) Center for Language and Speech Processing, Johns Hopkins University" + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.18, + 0.764, + 0.199 + ], + "angle": 0, + "content": "\\(\\spadesuit\\) Krieger School of Arts and Sciences, Johns Hopkins University" + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.253, + 0.341, + 0.269 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.281, + 0.461, + 0.522 + ], + "angle": 0, + "content": "In this position paper, we argue that computational text analysis lacks and requires organizing principles. A broad space separates its two constituent disciplines—natural language processing and social science—which has to date been sidestepped rather than filled by applying increasingly complex computational models to problems in social science research. We contrast descriptive and integrative findings, and our review of approximately 60 papers on computational text analysis reveals that those from *ACL venues are typically descriptive. The lack of theory began at the area's inception and has, over the decades, grown more important and challenging. A return to theoretically grounded research questions will propel the area from both theoretical and methodological points of view." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.548, + 0.262, + 0.563 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.573, + 0.49, + 0.83 + ], + "angle": 0, + "content": "Computational text analysis methods—an umbrella combining natural language processing with social science—are in a honeymoon period (Lazer and Radford, 2017; van Atteveldt and Peng, 2018). Today's social scientist might reach for the tools of computer science for their speed, scale, granularity, and consistency; for instance, natural language processing offers \"to analyze signals ranging from simple lexical cues to word clusters to choices of syntactic structure\" (Boydstun et al., 2014). The numerical outputs tell a story that is simple, easy to make sense of, and in that regard comforting. Conversely, today's computer scientist may see the problems of social science as answerable by objectivity and reductionism, eschewing interpretation for quantitative analysis." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.831, + 0.49, + 0.895 + ], + "angle": 0, + "content": "The conclusion of this reasoning, and the dominant stance in computational social science, is a reliance on machines alone to answer questions in the field, surrendering to their supposed objectivity" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.254, + 0.885, + 0.333 + ], + "angle": 0, + "content": "or impartiality. Can a machine's output go beyond descriptive catalogs of evidence, accelerating understanding of processes and motivations? From our experience, computers are nowhere near supplanting humans in interpreting social science results.1" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.334, + 0.885, + 0.51 + ], + "angle": 0, + "content": "An interdisciplinary inquiry must go farther than matching computational techniques to social science questions (O'Connor et al., 2011; Nguyen et al., 2020). It embraces synergistic methodology and connects the norms and standards of evidence from both. This means partnering computer science's preference for the structured, generalizable, and objective with the unstructured, critical, and contextual which the social sciences champion. This level of interdisciplinarity addresses the question raised by descriptive findings: So what?" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.511, + 0.885, + 0.72 + ], + "angle": 0, + "content": "We see theory as the solution, empowering rather than shackling investigations. What this paper advocates is not one particular theory—certainly these are myriad, and “even subject matter which has been under intensive and prolonged study remains at the unsettled periphery of research” (Nagel, 1963). Instead, we expand on our prior work (Dore and McCarthy, 2022) to clarify calls echoed for decades by computational and social science (McDermott, 1976; Jelinek, 2005; Hajic and Hajicova, 2007; Hofman et al., 2018; Lipton and Steinhardt, 2019; Baden et al., 2021). Underlying each, we find, is the urge to return to theory, which we espouse herein." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.732, + 0.776, + 0.748 + ], + "angle": 0, + "content": "2 Description vs. Integration" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.757, + 0.885, + 0.837 + ], + "angle": 0, + "content": "We contrast descriptive findings and theoretical analysis. An example of a descriptive finding is that an apple falls, or that it falls faster when pushed than dropped, or even that it falls at a particular rate estimated with some standard error by a complex" + }, + { + "type": "page_footnote", + "bbox": [ + 0.508, + 0.846, + 0.885, + 0.919 + ], + "angle": 0, + "content": "1See, e.g., Noam Chomsky's remark on GPT-3: \"You can't go to a physics conference and say: I've got a great theory. It accounts for everything and is so simple it can be captured in two words: 'Anything goes.' All known and unknown laws of nature are accommodated... Of course, everything impossible is accommodated also. That's GPT-3.\" [link]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.142, + 0.904, + 0.273, + 0.919 + ], + "angle": 0, + "content": "* Equal contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.928, + 0.523, + 0.941 + ], + "angle": 0, + "content": "1586" + }, + { + "type": "footer", + "bbox": [ + 0.227, + 0.946, + 0.772, + 0.958 + ], + "angle": 0, + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + }, + { + "type": "footer", + "bbox": [ + 0.369, + 0.959, + 0.631, + 0.972 + ], + "angle": 0, + "content": "Volume 2: Short Papers, pages 1586-1594" + }, + { + "type": "footer", + "bbox": [ + 0.296, + 0.973, + 0.702, + 0.986 + ], + "angle": 0, + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.488, + 0.181 + ], + "angle": 0, + "content": "interpolation. A theoretical analysis of the same phenomenon, credited to Newton, is that a fundamental force acts upon the apple, and that this same force governs the motion of the heavens. The theoretical analysis links the finding about the world critically to a broader body of knowledge and context." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.183, + 0.49, + 0.391 + ], + "angle": 0, + "content": "Despite advances in causal inference in NLP, the descriptive is all that a machine can provide to the social sciences (Feder et al., 2021). Certainly the methods of computational text analysis have advanced since the General Inquirer (Stone and Hunt, 1963) and Mosteller and Wallace's statistical inference of text authorship (1963). But methods are means, not ends. They uncover more descriptive findings in data: the rate of an apple's fall, the topics of refugees' tweets (Walk et al., 2022), the space given to marginalized groups in textbooks (Lucy et al., 2020), or patterns of state censorship (Bamman et al., 2012; King et al., 2013)." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.393, + 0.49, + 0.522 + ], + "angle": 0, + "content": "The foils to descriptive findings are integrative findings (Hofman et al., 2021), which offer causal explanations that enable future predictions—a theory, or as a 'model' in the sense of the Standard Model, rather than of a statistical model. Integrative findings can either offer new theories or couch their explanations in existing theories—but the theory is essential either way." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.535, + 0.317, + 0.552 + ], + "angle": 0, + "content": "3 We Don't Integrate" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.562, + 0.49, + 0.739 + ], + "angle": 0, + "content": "To contrast descriptive and integrative findings, we reviewed approximately 60 papers in computational text analysis published in *ACL venues. In Table 1, we describe several of these in terms of their descriptive or theory-grounded contributions.\\(^{2}\\) Descriptive papers may refer to social science theories or make generalizable claims, as when Demszky et al. (2019) write, \"The shooter's race appears to play a role in topic preference: if the shooter is white, Democrats become more likely to focus on shooter's identity,\" but they do not link to the two to each other." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.741, + 0.49, + 0.87 + ], + "angle": 0, + "content": "An excellent theory-grounded quantitative work is Nelson (2021); she confirms some of the most compelling features of identity theory, specifically that identities based on race were most distinguished by cultural discourse, whereas those based on gender by the domestic and the economic discourse. Similarly, we conducted theory-grounded quantitative work to investigate the application of the protest" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.213 + ], + "angle": 0, + "content": "paradigm and thematic framing in how western- and Hong Kong based newspapers portray protests in Hong Kong (McCarthy et al., 2021; McCarthy and Dore, 2022). Generally, it remains challenging to find computational social science papers in *ACL venues that go beyond description and prediction, advancing theory. Why is this? We believe it stemmed from the field's \"empirical turn\".3" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.215, + 0.885, + 0.327 + ], + "angle": 0, + "content": "Few remember when the meetings of ACL offered a few dozen papers, all entrenched in formalisms and linguistic theories. Arguably, 1996 was a turning point when the founders of SIGDAT held the first EMNLP at Penn under the auspices of the ACL. This gave a spotlight to the few but growing empiricists in the field and drew in more." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.33, + 0.885, + 0.473 + ], + "angle": 0, + "content": "EMNLP began a half-decade of measurable reorganization the field (Anderson et al., 2012). That EMNLP remains affiliated with ACL keeps the language-focused machine learning practitioners in our tent. The slow blurring of boundaries between each *ACL conference's expectations (Church, 2020) increases this unity. Both groups belong under this tent. But without a doubt, one group's voice is becoming less heard." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.476, + 0.884, + 0.669 + ], + "angle": 0, + "content": "Publication venues within the ACL focus on methods over theory.5 Techniques are taken off the shelf without critical examination because these are \"the best\" (often \"state of the art\") for their purposes (Ethayarajh and Jurafsky, 2020). This widens the gap between theoretical and empirical work.6 Hopkins and King (2010) claim, \"computer scientists may be interested in finding the needle in the haystack... social scientists are more commonly interested in characterizing the haystack\"—evincing the value of broader context.7 Wallach (2018), quoting Hopkins and King, explains that the two groups" + }, + { + "type": "page_footnote", + "bbox": [ + 0.508, + 0.684, + 0.883, + 0.756 + ], + "angle": 0, + "content": "A lesser reason is the challenge of serving two masters: adequately covering both the theoretical and methodological components within 8 pages. We recently received two reviews for an *ACL submission: one advocating for more of the social science context in the main text by eschewing methods to the appendix, and the other instructing us to do the opposite." + }, + { + "type": "page_footnote", + "bbox": [ + 0.533, + 0.757, + 0.882, + 0.771 + ], + "angle": 0, + "content": "4 And its predecessor the Workshop on Very Large Corpora." + }, + { + "type": "page_footnote", + "bbox": [ + 0.508, + 0.771, + 0.884, + 0.82 + ], + "angle": 0, + "content": "This is due to the outsized influence of computer science, often seen as the science of method (Hoare and Jones, 1989; Shapiro, 2001), when not instead seen as an engineering discipline (Rapaport, 2005)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.508, + 0.82, + 0.883, + 0.869 + ], + "angle": 0, + "content": "6 A related criticism is that empirical research has narrowed to focus on 'easy' questions that its tools can address (Coleman, 1986; Baden et al., 2021), especially when research questions are baked into the design of the task." + }, + { + "type": "page_footnote", + "bbox": [ + 0.508, + 0.869, + 0.883, + 0.919 + ], + "angle": 0, + "content": "7As evidence, see Siegel (2018): \"We usually don't know about causation, and we often don't necessarily care... the objective is more to predict than it is to understand the world... It just needs to work; prediction trumps explanation.\"" + }, + { + "type": "list", + "bbox": [ + 0.508, + 0.684, + 0.884, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.113, + 0.88, + 0.488, + 0.919 + ], + "angle": 0, + "content": "\\(^{2}\\)Following Lipton and Steinhardt (2019), we only describe papers by established researchers to \"avoid singling out junior students... who lack the opportunity to reply symmetrically\"." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1587" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.118, + 0.083, + 0.88, + 0.482 + ], + "angle": 0, + "content": "
Descriptive
Chang et al. (2009)The article presents new quantitative methods to measure semantic meaning in inferred topics. The authors emphasize the qualitative relevance of their findings as it validates the use of topics for corpus exploration and information retrieval. However, their working hypothesis and empirical findings are not connected to the extremely relevant field of communication theory.
Bamman et al. (2012)The article presents the first large-scale analysis of political content censorship in social media. The authors miss the opportunity to relate their hypothesis and findings to censorship theory, a natural theoretical context for the research, which would strengthen the relevance and generalizability of the findings.
Field et al. (2018)The article discusses media manipulation in Russia in the context of agenda-setting and framing, the tools that Russian state-owned (or heavily influenced) media outlets use to distract public attention from domestic economic politics. The authors implicitly refer to propaganda theory and autocratic theory throughout the article even though their findings are not discussed in relation to these theories.
Demszky et al. (2019)The article applies “a more comprehensive NLP framework to study linguistic aspects of polarization in social media”. While the article implicitly refers to theories of social conformity and social conflict, the findings are not linked or discussed (either explicitly or implicitly) to the theoretical frameworks that the authors touch on in their §1.
Integrative
DiMaggio et al. (2013)The article describes how topic models of newspaper articles help to study the politicization of government support for arts organizations and artists in the late 1980s in the US. The authors clearly define the theoretical context of their investigation and emphasize the relationship between theory and method throughout the paper.
Bamman et al. (2014)The article validates an empirical model that “employs multiple effects to account for the influence of extra-linguistic information (such as author)” by testing specific parameters against a variety of theory-based hypotheses derived from writing styles theories of England between 1700 and 1899.
Nelson (2021)The article argues that the full potential of machine learning can be better realized by “leveraging the epistemological alignment between machine learning and inductive research.” The author empirically demonstrates this by anchoring in identity theory a word embedding model of first-person narratives of the nineteenth-century U.S. South.
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.492, + 0.882, + 0.507 + ], + "angle": 0, + "content": "Table 1: Contrast between work in computational text analysis with descriptive findings versus integrative findings." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.532, + 0.489, + 0.644 + ], + "angle": 0, + "content": "are interested in very different research questions, and that computational social science must be more than computer science with social data; it must strive for valid explanatory models. In the same vein, at ACL 2022, ACL fellow Eduard Hovy remarked that NLP must be more than \"just machine learning on corpora\"." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.647, + 0.489, + 0.839 + ], + "angle": 0, + "content": "Social scientists are also coming to terms with the meaning of computational techniques applied more often in social science (Bail, 2014; Biernacki, 2015; Lee and Martin, 2015; Spillman, 2015). The focus of the debates, however, is on which methods are best suited to extract meaning from text, without addressing any theoretical considerations related to the methods or whether a theoretical framework for those methods even exists. The discussions on whether computational methods make social science research more efficient, reliable, and reproducible overtake attempts at theory-building." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.857, + 0.296, + 0.874 + ], + "angle": 0, + "content": "4 Moving Forward" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.888, + 0.49, + 0.92 + ], + "angle": 0, + "content": "We are not denying the value of computational approaches to analyzing text. Certainly, comput" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.532, + 0.885, + 0.676 + ], + "angle": 0, + "content": "ing can be an instrumental approach for modeling and understanding social complexity. This does not mean that other approaches, such as historical, ethnographic, or mathematical, become irrelevant. On the contrary, computational methods necessarily (whether awarely or not) rely on these earlier approaches to add value, in terms of improving our explanations and understanding (Radford and Joseph, 2020)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.678, + 0.885, + 0.902 + ], + "angle": 0, + "content": "As we are a field that prioritizes methods, consider the seminal book on methods in science: Abbott (2004) taxonomizes scientific ways of knowing. Its five broad categories are ethnography, historical narration, standard causal analysis, small-\\(N\\) comparison, and formal modeling. We in NLP myopically choose the third and fifth of these, ignoring the value of the others. But the broader point of Methods of Discovery is not methods. It is the research question. Any methodology should be grounded in the question, not incremental tweaks and reviewers' comfort (Church, 2020). This admits even qualitative or mixed-method approaches to text analysis." + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.904, + 0.882, + 0.92 + ], + "angle": 0, + "content": "The role of humans in scientific inquiry is nothing" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1588" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.112, + 0.085, + 0.488, + 0.293 + ], + "angle": 0, + "content": "new. Using qualitative analysis to complement quantitative techniques has its roots in Achen and Snidal (1989)'s recommendation to use historical case studies as a complement to statistical research. Their plea was strengthened by Verba's work in the early 1990s (Verba et al., 1993, 1995; Verba, 1996) and Tarrow (1995), who openly called for bridging qualitative and quantitative modes of research in social science. In doing so, they have enriched the field with critical methodological innovations (Gerring, 2004), benefiting from the recognition that \"quantitative methods must augment humans, not replace them\" (Grimmer and Stewart, 2013, 4)." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.295, + 0.489, + 0.472 + ], + "angle": 0, + "content": "The field can draw more from social science's rich tradition of inductive theory-building and interpretation to develop its theoretical approach—to prize either induction or deduction alone is a myth of scientific procedure (Thagard, 1988), but the melding of the two opens new doors. Rather than eschewing the complexity (a criticism leveled by Baden et al., 2021), it should put complexity at the center of its ontology on the basis that there are no immutable laws in social life or optimal solutions to social problems." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.472, + 0.49, + 0.698 + ], + "angle": 0, + "content": "Skepticism can linger toward findings not drawn from the standard practices of one's own field; indeed, social science was long skeptical of computational contributions (Armstrong, 1967). We believe that this drives the hyperfocus on improving a few accepted methods instead of exploring more broadly. If the doorway between disciplines is only narrowly open, this reflects a lack of appreciation for each field's ways of knowing. The disciplinary divide keeps computational researchers from embracing methods beyond standard causal analysis or formal modeling, so the interpreter-centric richness allowed by histories, ethnographies, and small-\\(N\\) exploration are precluded." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.71, + 0.248, + 0.725 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.735, + 0.49, + 0.849 + ], + "angle": 0, + "content": "We have explained the distinction between descriptive and theoretical findings as it pertains to computational text analysis. The bulk of work we found provided vast descriptive findings, often of high quality, but not giving back to questions of theory. We offer several suggestions on how to 'push the pendulum back' by prioritizing theory-building or" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.882, + 0.132 + ], + "angle": 0, + "content": "theory-affirming research questions and accepting whichever methods are best suited toward answering it—not only the familiar and entrenched ones." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.133, + 0.885, + 0.357 + ], + "angle": 0, + "content": "We are not the first to advocate for a shift in the patterns of applying computational techniques to real-world problems. There is a steady drumbeat from voices in the field advocating careful approaches (Nagel, 1963; McDermott, 1976; Jelinek, 2005; Hajic and Hajicova, 2007; Hofman et al., 2018; Lipton and Steinhardt, 2019; Baden et al., 2021). What we see underlying all of these—those writing against 'mathiness' and speculation, advocating for clear evaluation over anecdotes, criticizing textual researchers' dilution of conceptual standards, highlighting work that ties linguistic information into complex models—is an unspoken, perhaps unrealized, call for a return to theory." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.358, + 0.884, + 0.517 + ], + "angle": 0, + "content": "Not only do we aver that incorporating theory is essential; but also, other fields have strengthened themselves when espousing organizing principles beyond those of their progenitors. Behavioral economics is a success story here. It transcended the neat (but psychosocially stripped) mathematics it draws from to acknowledge deviations from rationality and blend economics with cognitive science (Kahneman and Tversky, 1979; Thaler, 1980; Thaler and Sunstein, 2009)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.519, + 0.884, + 0.919 + ], + "angle": 0, + "content": "For scientific—not simply engineering—advances to arise from the *ACL community, authors and reviewers alike must resist the temptation toward incremental, ‘safe’ research and follow Church (2005): “Controversial papers are great; boring unobjectionable incremental papers are not.” In reviewing new research, we should privilege not only work that presents new and unusual computational methods, but also interactions between computational and humanistic approaches to answering research questions. EMNLP was founded because of reviewing biases at ACL against groundbreaking methodological advances, and since then the two have homogenized; “EMNLP reviewing is no longer much of a differentiator” (Church, 2020). We found that theoretically grounded findings in text analysis are often published in non-\\*ACL venues (Table 1), but ACL sets the standard for work involving computational text analysis and NLP. Is there no home for groundbreaking integrative or interdisciplinary work in *ACL, such that a new venue is required? Or can we adapt our standards to invite deeper connections to theory and new ways of knowing?" + }, + { + "type": "page_footnote", + "bbox": [ + 0.113, + 0.856, + 0.488, + 0.919 + ], + "angle": 0, + "content": "8 Expertise plays a role as well (Shing et al., 2018), which is why Mechanical Turk doesn't fill the need for qualitative analysis. This is exemplified by Radford and Joseph (2020)'s observation of \"non-expert annotators providing unreliable annotations, even after a discussion period\"." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1589" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.115, + 0.085, + 0.279, + 0.102 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.11, + 0.49, + 0.207 + ], + "angle": 0, + "content": "This publication was made possible in part by a grant from the American Political Science Association to A.D.M. and G.M.D.D. The statements made and views expressed are solely the responsibility of the authors. A.D.M. is supported by an Amazon Fellowship and a Frederick Jelinek Fellowship." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.218, + 0.221, + 0.233 + ], + "angle": 0, + "content": "Limitations" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.244, + 0.49, + 0.501 + ], + "angle": 0, + "content": "The key limitation of our work is that, when conducting the review of approximately 60 papers (by searching through the ACL Anthology for works in computational social science since 2010), we encountered a skewed distribution of descriptive versus integrative works. In fact, it was relatively simple to find descriptive works, and that section of Table 1 could have been much longer. We also recognize that, due to the mixed nature of our field, scientific and integrative findings are not the only goal—our 'big tent' includes engineers as well, who value gains in performance indicators. Finally, the fact that we have few examples of papers showing a return to theory renders the possibility that our central claim is misinterpreted in a normative way as a mandate." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.528, + 0.214, + 0.543 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.551, + 0.49, + 0.592 + ], + "angle": 0, + "content": "Andrew Delano Abbott. 2004. Methods of discovery: Heuristics for the social sciences (contemporary societies). WW Norton & Company." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.6, + 0.49, + 0.641 + ], + "angle": 0, + "content": "Christopher H. Achen and Duncan Snidal. 1989. Rational deterrence theory and comparative case studies. World Politics, 41(2):143-169." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.65, + 0.49, + 0.731 + ], + "angle": 0, + "content": "Ashton Anderson, Dan Jurafsky, and Daniel A. McFarland. 2012. Towards a computational history of the ACL: 1980-2008. In Proceedings of the ACL-2012 Special Workshop on Rediscovering 50 Years of Discoveries, pages 13-21, Jeju Island, Korea. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.739, + 0.49, + 0.793 + ], + "angle": 0, + "content": "J. Scott Armstrong. 1967. Derivation of theory by means of factor analysis or Tom Swift and his electric factor analysis machine. The American Statistician, 21(5):17-21." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.803, + 0.49, + 0.869 + ], + "angle": 0, + "content": "Christian Baden, Christian Pipal, Martijn Schoonvelde, and Mariken A. C. G van der Velden. 2021. Three gaps in computational text analysis methods for social sciences: A research agenda. Communication Methods and Measures, 0(0):1-18." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.878, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Christopher A. Bail. 2014. The cultural environment: measuring culture with big data. Theory and Society, 43(3):465-482." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.551, + 0.49, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.086, + 0.885, + 0.127 + ], + "angle": 0, + "content": "David Bamman, Brendan O'Connor, and Noah Smith. 2012. Censorship and deletion practices in chinese social media. First Monday, 17(3)." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.137, + 0.885, + 0.216 + ], + "angle": 0, + "content": "David Bamman, Ted Underwood, and Noah A. Smith. 2014. A Bayesian mixed effects model of literary character. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 370-379, Baltimore, Maryland. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.226, + 0.885, + 0.293 + ], + "angle": 0, + "content": "Richard Biernacki. 2015. How to do things with historical texts. American Journal of Cultural Sociology, 3(3):311-352. Copyright - © Palgrave Macmillan, a division of Macmillan Publishers Ltd 2015; Last updated - 2018-09-25." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.303, + 0.885, + 0.357 + ], + "angle": 0, + "content": "Amber E Boydstun, Dallas Card, Justin Gross, Paul Resnick, and Noah A Smith. 2014. Tracking the development of media frames within and across policy issues. Unpublished." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.367, + 0.885, + 0.458 + ], + "angle": 0, + "content": "Jonathan Chang, Jordan Boyd-Graber, Sean Gerrish, Chong Wang, and David M. Blei. 2009. Reading tea leaves: How humans interpret topic models. In Proceedings of the 22nd International Conference on Neural Information Processing Systems, NIPS'09, page 288-296, Red Hook, NY, USA. Curran Associates Inc." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.469, + 0.885, + 0.509 + ], + "angle": 0, + "content": "Kenneth Church. 2005. Last words: Reviewing the reviewers. Computational Linguistics, 31(4):575-578." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.52, + 0.885, + 0.56 + ], + "angle": 0, + "content": "Kenneth Ward Church. 2020. Emerging trends: Reviewing the reviewers (again). Natural Language Engineering, 26(2):245-257." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.571, + 0.885, + 0.61 + ], + "angle": 0, + "content": "James S. Coleman. 1986. Social theory, social research, and a theory of action. American Journal of Sociology, 91(6):1309-1335." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.621, + 0.885, + 0.753 + ], + "angle": 0, + "content": "Dorottya Demszky, Nikhil Garg, Rob Voigt, James Zou, Jesse Shapiro, Matthew Gentzkow, and Dan Jurafsky. 2019. Analyzing polarization in social media: Method and application to tweets on 21 mass shootings. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2970-3005, Minneapolis, Minnesota. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.763, + 0.885, + 0.842 + ], + "angle": 0, + "content": "Paul DiMaggio, Manish Nag, and David Blei. 2013. Exploiting affinities between topic modeling and the sociological perspective on culture: Application to newspaper coverage of U.S. government arts funding. *Poetics*, 41(6):570–606. Topic Models and the Cultural Sciences." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.852, + 0.885, + 0.919 + ], + "angle": 0, + "content": "Giovanna Maria Dora Dore and Arya D. McCarthy. 2022. Learning to play with the machines in social science research: Bringing the theory back in. In ICML 2022 Workshop on Human-Machine Collaboration and Teaming, Baltimore, Maryland." + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.086, + 0.885, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1590" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.166 + ], + "angle": 0, + "content": "Kawin Ethayarajh and Dan Jurafsky. 2020. Utility is in the eye of the user: A critique of NLP leaderboards. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4846-4853, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.177, + 0.488, + 0.269 + ], + "angle": 0, + "content": "Amir Feder, Katherine A. Keith, Emaad Manzoor, Reid Pryzant, Dhanya Sridhar, Zach Wood-Doughty, Jacob Eisenstein, Justin Grimmer, Roi Reichart, Margaret E. Roberts, Brandon M. Stewart, Victor Veitch, and Diyi Yang. 2021. Causal inference in natural language processing: Estimation, prediction, interpretation and beyond. CoRR, abs/2109.00725." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.281, + 0.488, + 0.386 + ], + "angle": 0, + "content": "Anjalie Field, Doron Kliger, Shuly Wintner, Jennifer Pan, Dan Jurafsky, and Yulia Tsvetkov. 2018. Framing and agenda-setting in Russian news: a computational analysis of intricate political strategies. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 3570-3580, Brussels, Belgium. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.398, + 0.488, + 0.438 + ], + "angle": 0, + "content": "John Gerring. 2004. What is a case study and what is it good for? American Political Science Review, 98(2):341-354." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.45, + 0.488, + 0.502 + ], + "angle": 0, + "content": "Justin Grimmer and Brandon M. Stewart. 2013. Text as data: The promise and pitfalls of automatic content analysis methods for political texts. Political Analysis, 21(3):267-297." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.515, + 0.488, + 0.568 + ], + "angle": 0, + "content": "Jan Hajic and Eva Hajičová. 2007. Some of our best friends are statisticians. In Text, Speech and Dialogue, pages 2-10, Berlin, Heidelberg. Springer Berlin Heidelberg." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.58, + 0.488, + 0.607 + ], + "angle": 0, + "content": "C. A. R. Hoare and C. B. Jones. 1989. Essays in Computing Science. Prentice-Hall, Inc., USA." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.619, + 0.488, + 0.659 + ], + "angle": 0, + "content": "Jake Hofman, Miro Dudík, and Daniel G. Goldstein. 2018. Perspective annotation for numerical representations. United States Patent Application." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.671, + 0.488, + 0.749 + ], + "angle": 0, + "content": "Jake M Hofman, Duncan J Watts, Susan Athey, Filiz Garip, Thomas L Griffiths, Jon Kleinberg, Helen Margetts, Sendhil Mullainathan, Matthew J Salganik, Simine Vazire, et al. 2021. Integrating explanation and prediction in computational social science. Nature, 595(7866):181-188." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.762, + 0.488, + 0.814 + ], + "angle": 0, + "content": "Daniel J. Hopkins and Gary King. 2010. A method of automated nonparametric content analysis for social science. American Journal of Political Science, 54(1):229-247." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.826, + 0.49, + 0.866 + ], + "angle": 0, + "content": "Frederick Jelinek. 2005. Some of my best friends are linguists. Language Resources and Evaluation, 39(1):25-34." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.878, + 0.488, + 0.918 + ], + "angle": 0, + "content": "Daniel Kahneman and Amos Tversky. 1979. Prospect theory: An analysis of decision under risk. *Econometrica*, 47(2):263-291." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.086, + 0.882, + 0.139 + ], + "angle": 0, + "content": "Gary King, Jennifer Pan, and Margaret E. Roberts. 2013. How censorship in China allows government criticism but silences collective expression. American Political Science Review, 107(2 (May)):1-18." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.152, + 0.882, + 0.191 + ], + "angle": 0, + "content": "David Lazer and Jason Radford. 2017. Data ex machina: Introduction to big data. Annual Review of Sociology, 43(1):19-39." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.205, + 0.882, + 0.245 + ], + "angle": 0, + "content": "Monica Lee and John Levi Martin. 2015. Coding, counting and cultural cartography. American Journal of Cultural Sociology, 3(1):1-33." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.258, + 0.882, + 0.311 + ], + "angle": 0, + "content": "Zachary C. Lipton and Jacob Steinhardt. 2019. Troubling trends in machine learning scholarship: Some ML papers suffer from flaws that could mislead the public and stymie future research. Queue, 17(1):45-77." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.324, + 0.882, + 0.39 + ], + "angle": 0, + "content": "Li Lucy, Dorottya Demszky, Patricia Bromley, and Dan Jurafsky. 2020. Content analysis of textbooks via natural language processing: Findings on gender, race, and ethnicity in Texas U.S. history textbooks. AERA Open, 6(3):2332858420940312." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.403, + 0.882, + 0.495 + ], + "angle": 0, + "content": "Arya D. McCarthy and Giovanna Maria Dora Dore. 2022. Hong Kong: Longitudinal and synchronic characterisations of protest news between 1998 and 2020. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 2891-2900, Marseille, France. European Language Resources Association." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.509, + 0.882, + 0.627 + ], + "angle": 0, + "content": "Arya D. McCarthy, James Scharf, and Giovanna Maria Dora Dore. 2021. A mixed-methods analysis of western and Hong Kong-based reporting on the 2019-2020 protests. In Proceedings of the 5th Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature, pages 178-188, Punta Cana, Dominican Republic (online). Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.64, + 0.882, + 0.667 + ], + "angle": 0, + "content": "Drew McDermott. 1976. Artificial intelligence meets natural stupidity. SIGART Bull., (57):4-9." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.68, + 0.882, + 0.72 + ], + "angle": 0, + "content": "Frederick Mosteller and David L. Wallace. 1963. Inference in an authorship problem. Journal of the American Statistical Association, 58(302):275-309." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.733, + 0.882, + 0.761 + ], + "angle": 0, + "content": "Ernest Nagel. 1963. The structure of science: Problems in the logic of scientific explanation. Mind, 72(287)." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.773, + 0.882, + 0.838 + ], + "angle": 0, + "content": "Laura K. Nelson. 2021. Leveraging the alignment between machine learning and intersectionality: Using word embeddings to measure intersectional experiences of the nineteenth century U.S. South. Poetics, 88:101539. Measure Mohr Culture." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.852, + 0.882, + 0.918 + ], + "angle": 0, + "content": "Dong Nguyen, Maria Liakata, Simon DeDeo, Jacob Eisenstein, David Mimno, Rebekah Tromble, and Jane Winters. 2020. How we do things with words: Analyzing text as social and cultural data. Frontiers in Artificial Intelligence, 3." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.882, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.519, + 0.941 + ], + "angle": 0, + "content": "1591" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.086, + 0.487, + 0.152 + ], + "angle": 0, + "content": "Brendan O'Connor, David Bamman, and Noah A Smith. 2011. Computational text analysis for social science: Model complexity and assumptions. In Proc. of the NIPS Workshop on Comptuational Social Science and the Wisdom of Crowds." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.163, + 0.487, + 0.203 + ], + "angle": 0, + "content": "Jason Radford and Kenneth Joseph. 2020. Theory in, theory out: The uses of social theory in machine learning for social science. Frontiers in Big Data, 3." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.214, + 0.487, + 0.253 + ], + "angle": 0, + "content": "William J Rapaport. 2005. Philosophy of computer science: An introductory course. Teaching philosophy, 28(4):319-341." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.265, + 0.487, + 0.305 + ], + "angle": 0, + "content": "Stuart C. Shapiro. 2001. Computer science: The study of procedures. Technical report, Department of Computer Science and Engineering, University of Buffalo." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.316, + 0.487, + 0.42 + ], + "angle": 0, + "content": "Han-Chin Shing, Suraj Nair, Ayah Zirikly, Meir Friedenberg, Hal Daumé III, and Philip Resnik. 2018. Expert, crowdsourced, and machine assessment of suicide risk via online postings. In Proceedings of the Fifth Workshop on Computational Linguistics and Clinical Psychology: From Keyboard to Clinic, pages 25-36, New Orleans, LA. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.432, + 0.487, + 0.458 + ], + "angle": 0, + "content": "David A. Siegel. 2018. Analyzing computational models. American Journal of Political Science, 62(3):745-759." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.47, + 0.487, + 0.509 + ], + "angle": 0, + "content": "Lyn Spillman. 2015. Ghosts of straw men: A reply to Lee and Martin. American Journal of Cultural Sociology, 3(3):365-379." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.521, + 0.487, + 0.6 + ], + "angle": 0, + "content": "Philip J. Stone and Earl B. Hunt. 1963. A computer approach to content analysis: Studies using the general inquirer system. In Proceedings of the May 21-23, 1963, Spring Joint Computer Conference, AFIPS '63 (Spring), page 241-256, New York, NY, USA. Association for Computing Machinery." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.611, + 0.487, + 0.65 + ], + "angle": 0, + "content": "Sidney Tarrow. 1995. Bridging the quantitative-qualitative divide in political science. American Political Science Review, 89(2):471-474." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.661, + 0.487, + 0.687 + ], + "angle": 0, + "content": "Paul Thagard. 1988. Computational Philosophy of Science. MIT Press." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.699, + 0.487, + 0.791 + ], + "angle": 0, + "content": "Richard Thaler. 1980. Judgement And Decision Making Under Uncertainty: What Economists Can Learn From Psychology. Risk Analysis in Agriculture: Research and Educational Developments, January 16-18, 1980, Tucson, Arizona 271572, Regional Research Projects \\(>\\) W-149: An Economic Evaluation of Managing Market Risks in Agriculture." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.802, + 0.487, + 0.842 + ], + "angle": 0, + "content": "Richard H. Thaler and Cass R. Sunstein. 2009. Nudge: Improving decisions about health, wealth, and happiness. Penguin." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.853, + 0.487, + 0.918 + ], + "angle": 0, + "content": "Wouter van Atteveldt and Tai-Quan Peng. 2018. When communication meets computation: Opportunities, challenges, and pitfalls in computational communication science. Communication Methods and Measures, 12(2-3):81-92." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.487, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.086, + 0.882, + 0.138 + ], + "angle": 0, + "content": "Sidney Verba. 1996. The citizen as respondent: Sample surveys and american democracy. American Political Science Review, 90(1):1-7. Presidential Address, American Political Science Association, 1995." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.149, + 0.882, + 0.201 + ], + "angle": 0, + "content": "Sidney Verba, Kay Lehman Schlozman, Henry Brady, and Norman H. Nie. 1993. Citizen activity: Who participates? what do they say? The American Political Science Review, 87(2):303-318." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.212, + 0.882, + 0.252 + ], + "angle": 0, + "content": "Sidney Verba, Kay Lehman Schlozman, and Henry E Brady. 1995. Voice and equality: Civic volunteerism in American politics. Harvard University Press." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.262, + 0.882, + 0.327 + ], + "angle": 0, + "content": "Erin Walk, Elizabeth Parker-Magyar, Kiran Garimella, Ahmet Akbiyik, and Fotini Christia. 2022. Social media narratives across platforms in conflict: Evidence from Syria. MIT Political Science Department Research Paper No. 2022-2, available at SSRN." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.337, + 0.882, + 0.376 + ], + "angle": 0, + "content": "Hanna Wallach. 2018. Computational social science \\(\\neq\\) computer science + social data. Commun. ACM, 61(3):42-44." + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.086, + 0.882, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1592" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.133, + 0.085, + 0.435, + 0.1 + ], + "angle": 0, + "content": "ACL 2023 Responsible NLP Checklist" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.108, + 0.323, + 0.123 + ], + "angle": 0, + "content": "A For every submission:" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.127, + 0.533, + 0.144 + ], + "angle": 0, + "content": "A1. Did you describe the limitations of your work?" + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.145, + 0.4, + 0.16 + ], + "angle": 0, + "content": "Unnumbered; appears on page 5." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.171, + 0.553, + 0.187 + ], + "angle": 0, + "content": "A2. Did you discuss any potential risks of your work?" + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.188, + 0.331, + 0.203 + ], + "angle": 0, + "content": "This is a position paper." + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.213, + 0.696, + 0.23 + ], + "angle": 0, + "content": "A3. Do the abstract and introduction summarize the paper's main claims?" + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.231, + 0.233, + 0.246 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.256, + 0.669, + 0.273 + ], + "angle": 0, + "content": "A4. Have you used AI writing assistants when working on this paper?" + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.274, + 0.233, + 0.288 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.301, + 0.489, + 0.317 + ], + "angle": 0, + "content": "B Did you use or create scientific artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.322, + 0.215, + 0.337 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.348, + 0.53, + 0.364 + ], + "angle": 0, + "content": "B1. Did you cite the creators of artifacts you used?" + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.365, + 0.351, + 0.38 + ], + "angle": 0, + "content": "Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.391, + 0.779, + 0.407 + ], + "angle": 0, + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.408, + 0.351, + 0.423 + ], + "angle": 0, + "content": "Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.434, + 0.882, + 0.497 + ], + "angle": 0, + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?" + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.498, + 0.351, + 0.514 + ], + "angle": 0, + "content": "Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.525, + 0.882, + 0.573 + ], + "angle": 0, + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?" + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.574, + 0.351, + 0.589 + ], + "angle": 0, + "content": "Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.6, + 0.881, + 0.632 + ], + "angle": 0, + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?" + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.633, + 0.351, + 0.648 + ], + "angle": 0, + "content": "Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.659, + 0.882, + 0.74 + ], + "angle": 0, + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be." + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.741, + 0.351, + 0.755 + ], + "angle": 0, + "content": "Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.766, + 0.494, + 0.782 + ], + "angle": 0, + "content": "C Did you run computational experiments?" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.787, + 0.215, + 0.802 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.814, + 0.882, + 0.846 + ], + "angle": 0, + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?" + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.847, + 0.351, + 0.861 + ], + "angle": 0, + "content": "Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.869, + 0.878, + 0.893 + ], + "angle": 0, + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1593" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.129, + 0.085, + 0.88, + 0.116 + ], + "angle": 0, + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.118, + 0.351, + 0.133 + ], + "angle": 0, + "content": "Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.144, + 0.884, + 0.192 + ], + "angle": 0, + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.193, + 0.351, + 0.208 + ], + "angle": 0, + "content": "Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.219, + 0.884, + 0.266 + ], + "angle": 0, + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.268, + 0.351, + 0.283 + ], + "angle": 0, + "content": "Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.294, + 0.878, + 0.311 + ], + "angle": 0, + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.315, + 0.215, + 0.33 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.341, + 0.884, + 0.373 + ], + "angle": 0, + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.374, + 0.351, + 0.39 + ], + "angle": 0, + "content": "Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.4, + 0.884, + 0.448 + ], + "angle": 0, + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.449, + 0.351, + 0.465 + ], + "angle": 0, + "content": "Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.476, + 0.884, + 0.523 + ], + "angle": 0, + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.524, + 0.351, + 0.54 + ], + "angle": 0, + "content": "Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.55, + 0.876, + 0.566 + ], + "angle": 0, + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.567, + 0.351, + 0.583 + ], + "angle": 0, + "content": "Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.593, + 0.884, + 0.624 + ], + "angle": 0, + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.626, + 0.351, + 0.642 + ], + "angle": 0, + "content": "Not applicable. Left blank." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1594" + } + ] +] \ No newline at end of file diff --git a/2023/Theory-Grounded Computational Text Analysis/11a89724-ae4c-4578-823c-45401a11d43f_origin.pdf b/2023/Theory-Grounded Computational Text Analysis/11a89724-ae4c-4578-823c-45401a11d43f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6380338f2442ca1e6b1ec9fcbd1b3fe5d4ce44f4 --- /dev/null +++ b/2023/Theory-Grounded Computational Text Analysis/11a89724-ae4c-4578-823c-45401a11d43f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cc76e1a08587d35224ca1d40170fe36053c011e5521c3ece55c406ab1732b6b9 +size 218498 diff --git a/2023/Theory-Grounded Computational Text Analysis/full.md b/2023/Theory-Grounded Computational Text Analysis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8e7077cb88b47ab643b6518a427afbc5da2f3320 --- /dev/null +++ b/2023/Theory-Grounded Computational Text Analysis/full.md @@ -0,0 +1,250 @@ +# Theory-Grounded Computational Text Analysis + +Arya D. McCarthy\* and Giovanna Maria Dora Dore + +$\diamond$ Center for Language and Speech Processing, Johns Hopkins University + +$\spadesuit$ Krieger School of Arts and Sciences, Johns Hopkins University + +# Abstract + +In this position paper, we argue that computational text analysis lacks and requires organizing principles. A broad space separates its two constituent disciplines—natural language processing and social science—which has to date been sidestepped rather than filled by applying increasingly complex computational models to problems in social science research. We contrast descriptive and integrative findings, and our review of approximately 60 papers on computational text analysis reveals that those from *ACL venues are typically descriptive. The lack of theory began at the area's inception and has, over the decades, grown more important and challenging. A return to theoretically grounded research questions will propel the area from both theoretical and methodological points of view. + +# 1 Introduction + +Computational text analysis methods—an umbrella combining natural language processing with social science—are in a honeymoon period (Lazer and Radford, 2017; van Atteveldt and Peng, 2018). Today's social scientist might reach for the tools of computer science for their speed, scale, granularity, and consistency; for instance, natural language processing offers "to analyze signals ranging from simple lexical cues to word clusters to choices of syntactic structure" (Boydstun et al., 2014). The numerical outputs tell a story that is simple, easy to make sense of, and in that regard comforting. Conversely, today's computer scientist may see the problems of social science as answerable by objectivity and reductionism, eschewing interpretation for quantitative analysis. + +The conclusion of this reasoning, and the dominant stance in computational social science, is a reliance on machines alone to answer questions in the field, surrendering to their supposed objectivity + +or impartiality. Can a machine's output go beyond descriptive catalogs of evidence, accelerating understanding of processes and motivations? From our experience, computers are nowhere near supplanting humans in interpreting social science results.1 + +An interdisciplinary inquiry must go farther than matching computational techniques to social science questions (O'Connor et al., 2011; Nguyen et al., 2020). It embraces synergistic methodology and connects the norms and standards of evidence from both. This means partnering computer science's preference for the structured, generalizable, and objective with the unstructured, critical, and contextual which the social sciences champion. This level of interdisciplinarity addresses the question raised by descriptive findings: So what? + +We see theory as the solution, empowering rather than shackling investigations. What this paper advocates is not one particular theory—certainly these are myriad, and “even subject matter which has been under intensive and prolonged study remains at the unsettled periphery of research” (Nagel, 1963). Instead, we expand on our prior work (Dore and McCarthy, 2022) to clarify calls echoed for decades by computational and social science (McDermott, 1976; Jelinek, 2005; Hajic and Hajicova, 2007; Hofman et al., 2018; Lipton and Steinhardt, 2019; Baden et al., 2021). Underlying each, we find, is the urge to return to theory, which we espouse herein. + +# 2 Description vs. Integration + +We contrast descriptive findings and theoretical analysis. An example of a descriptive finding is that an apple falls, or that it falls faster when pushed than dropped, or even that it falls at a particular rate estimated with some standard error by a complex + +interpolation. A theoretical analysis of the same phenomenon, credited to Newton, is that a fundamental force acts upon the apple, and that this same force governs the motion of the heavens. The theoretical analysis links the finding about the world critically to a broader body of knowledge and context. + +Despite advances in causal inference in NLP, the descriptive is all that a machine can provide to the social sciences (Feder et al., 2021). Certainly the methods of computational text analysis have advanced since the General Inquirer (Stone and Hunt, 1963) and Mosteller and Wallace's statistical inference of text authorship (1963). But methods are means, not ends. They uncover more descriptive findings in data: the rate of an apple's fall, the topics of refugees' tweets (Walk et al., 2022), the space given to marginalized groups in textbooks (Lucy et al., 2020), or patterns of state censorship (Bamman et al., 2012; King et al., 2013). + +The foils to descriptive findings are integrative findings (Hofman et al., 2021), which offer causal explanations that enable future predictions—a theory, or as a 'model' in the sense of the Standard Model, rather than of a statistical model. Integrative findings can either offer new theories or couch their explanations in existing theories—but the theory is essential either way. + +# 3 We Don't Integrate + +To contrast descriptive and integrative findings, we reviewed approximately 60 papers in computational text analysis published in *ACL venues. In Table 1, we describe several of these in terms of their descriptive or theory-grounded contributions. $^{2}$ Descriptive papers may refer to social science theories or make generalizable claims, as when Demszky et al. (2019) write, "The shooter's race appears to play a role in topic preference: if the shooter is white, Democrats become more likely to focus on shooter's identity," but they do not link to the two to each other. + +An excellent theory-grounded quantitative work is Nelson (2021); she confirms some of the most compelling features of identity theory, specifically that identities based on race were most distinguished by cultural discourse, whereas those based on gender by the domestic and the economic discourse. Similarly, we conducted theory-grounded quantitative work to investigate the application of the protest + +paradigm and thematic framing in how western- and Hong Kong based newspapers portray protests in Hong Kong (McCarthy et al., 2021; McCarthy and Dore, 2022). Generally, it remains challenging to find computational social science papers in *ACL venues that go beyond description and prediction, advancing theory. Why is this? We believe it stemmed from the field's "empirical turn".3 + +Few remember when the meetings of ACL offered a few dozen papers, all entrenched in formalisms and linguistic theories. Arguably, 1996 was a turning point when the founders of SIGDAT held the first EMNLP at Penn under the auspices of the ACL. This gave a spotlight to the few but growing empiricists in the field and drew in more. + +EMNLP began a half-decade of measurable reorganization the field (Anderson et al., 2012). That EMNLP remains affiliated with ACL keeps the language-focused machine learning practitioners in our tent. The slow blurring of boundaries between each *ACL conference's expectations (Church, 2020) increases this unity. Both groups belong under this tent. But without a doubt, one group's voice is becoming less heard. + +Publication venues within the ACL focus on methods over theory.5 Techniques are taken off the shelf without critical examination because these are "the best" (often "state of the art") for their purposes (Ethayarajh and Jurafsky, 2020). This widens the gap between theoretical and empirical work.6 Hopkins and King (2010) claim, "computer scientists may be interested in finding the needle in the haystack... social scientists are more commonly interested in characterizing the haystack"—evincing the value of broader context.7 Wallach (2018), quoting Hopkins and King, explains that the two groups + +
Descriptive
Chang et al. (2009)The article presents new quantitative methods to measure semantic meaning in inferred topics. The authors emphasize the qualitative relevance of their findings as it validates the use of topics for corpus exploration and information retrieval. However, their working hypothesis and empirical findings are not connected to the extremely relevant field of communication theory.
Bamman et al. (2012)The article presents the first large-scale analysis of political content censorship in social media. The authors miss the opportunity to relate their hypothesis and findings to censorship theory, a natural theoretical context for the research, which would strengthen the relevance and generalizability of the findings.
Field et al. (2018)The article discusses media manipulation in Russia in the context of agenda-setting and framing, the tools that Russian state-owned (or heavily influenced) media outlets use to distract public attention from domestic economic politics. The authors implicitly refer to propaganda theory and autocratic theory throughout the article even though their findings are not discussed in relation to these theories.
Demszky et al. (2019)The article applies “a more comprehensive NLP framework to study linguistic aspects of polarization in social media”. While the article implicitly refers to theories of social conformity and social conflict, the findings are not linked or discussed (either explicitly or implicitly) to the theoretical frameworks that the authors touch on in their §1.
Integrative
DiMaggio et al. (2013)The article describes how topic models of newspaper articles help to study the politicization of government support for arts organizations and artists in the late 1980s in the US. The authors clearly define the theoretical context of their investigation and emphasize the relationship between theory and method throughout the paper.
Bamman et al. (2014)The article validates an empirical model that “employs multiple effects to account for the influence of extra-linguistic information (such as author)” by testing specific parameters against a variety of theory-based hypotheses derived from writing styles theories of England between 1700 and 1899.
Nelson (2021)The article argues that the full potential of machine learning can be better realized by “leveraging the epistemological alignment between machine learning and inductive research.” The author empirically demonstrates this by anchoring in identity theory a word embedding model of first-person narratives of the nineteenth-century U.S. South.
+ +Table 1: Contrast between work in computational text analysis with descriptive findings versus integrative findings. + +are interested in very different research questions, and that computational social science must be more than computer science with social data; it must strive for valid explanatory models. In the same vein, at ACL 2022, ACL fellow Eduard Hovy remarked that NLP must be more than "just machine learning on corpora". + +Social scientists are also coming to terms with the meaning of computational techniques applied more often in social science (Bail, 2014; Biernacki, 2015; Lee and Martin, 2015; Spillman, 2015). The focus of the debates, however, is on which methods are best suited to extract meaning from text, without addressing any theoretical considerations related to the methods or whether a theoretical framework for those methods even exists. The discussions on whether computational methods make social science research more efficient, reliable, and reproducible overtake attempts at theory-building. + +# 4 Moving Forward + +We are not denying the value of computational approaches to analyzing text. Certainly, comput + +ing can be an instrumental approach for modeling and understanding social complexity. This does not mean that other approaches, such as historical, ethnographic, or mathematical, become irrelevant. On the contrary, computational methods necessarily (whether awarely or not) rely on these earlier approaches to add value, in terms of improving our explanations and understanding (Radford and Joseph, 2020). + +As we are a field that prioritizes methods, consider the seminal book on methods in science: Abbott (2004) taxonomizes scientific ways of knowing. Its five broad categories are ethnography, historical narration, standard causal analysis, small- $N$ comparison, and formal modeling. We in NLP myopically choose the third and fifth of these, ignoring the value of the others. But the broader point of Methods of Discovery is not methods. It is the research question. Any methodology should be grounded in the question, not incremental tweaks and reviewers' comfort (Church, 2020). This admits even qualitative or mixed-method approaches to text analysis. + +The role of humans in scientific inquiry is nothing + +new. Using qualitative analysis to complement quantitative techniques has its roots in Achen and Snidal (1989)'s recommendation to use historical case studies as a complement to statistical research. Their plea was strengthened by Verba's work in the early 1990s (Verba et al., 1993, 1995; Verba, 1996) and Tarrow (1995), who openly called for bridging qualitative and quantitative modes of research in social science. In doing so, they have enriched the field with critical methodological innovations (Gerring, 2004), benefiting from the recognition that "quantitative methods must augment humans, not replace them" (Grimmer and Stewart, 2013, 4). + +The field can draw more from social science's rich tradition of inductive theory-building and interpretation to develop its theoretical approach—to prize either induction or deduction alone is a myth of scientific procedure (Thagard, 1988), but the melding of the two opens new doors. Rather than eschewing the complexity (a criticism leveled by Baden et al., 2021), it should put complexity at the center of its ontology on the basis that there are no immutable laws in social life or optimal solutions to social problems. + +Skepticism can linger toward findings not drawn from the standard practices of one's own field; indeed, social science was long skeptical of computational contributions (Armstrong, 1967). We believe that this drives the hyperfocus on improving a few accepted methods instead of exploring more broadly. If the doorway between disciplines is only narrowly open, this reflects a lack of appreciation for each field's ways of knowing. The disciplinary divide keeps computational researchers from embracing methods beyond standard causal analysis or formal modeling, so the interpreter-centric richness allowed by histories, ethnographies, and small- $N$ exploration are precluded. + +# 5 Conclusion + +We have explained the distinction between descriptive and theoretical findings as it pertains to computational text analysis. The bulk of work we found provided vast descriptive findings, often of high quality, but not giving back to questions of theory. We offer several suggestions on how to 'push the pendulum back' by prioritizing theory-building or + +theory-affirming research questions and accepting whichever methods are best suited toward answering it—not only the familiar and entrenched ones. + +We are not the first to advocate for a shift in the patterns of applying computational techniques to real-world problems. There is a steady drumbeat from voices in the field advocating careful approaches (Nagel, 1963; McDermott, 1976; Jelinek, 2005; Hajic and Hajicova, 2007; Hofman et al., 2018; Lipton and Steinhardt, 2019; Baden et al., 2021). What we see underlying all of these—those writing against 'mathiness' and speculation, advocating for clear evaluation over anecdotes, criticizing textual researchers' dilution of conceptual standards, highlighting work that ties linguistic information into complex models—is an unspoken, perhaps unrealized, call for a return to theory. + +Not only do we aver that incorporating theory is essential; but also, other fields have strengthened themselves when espousing organizing principles beyond those of their progenitors. Behavioral economics is a success story here. It transcended the neat (but psychosocially stripped) mathematics it draws from to acknowledge deviations from rationality and blend economics with cognitive science (Kahneman and Tversky, 1979; Thaler, 1980; Thaler and Sunstein, 2009). + +For scientific—not simply engineering—advances to arise from the *ACL community, authors and reviewers alike must resist the temptation toward incremental, ‘safe’ research and follow Church (2005): “Controversial papers are great; boring unobjectionable incremental papers are not.” In reviewing new research, we should privilege not only work that presents new and unusual computational methods, but also interactions between computational and humanistic approaches to answering research questions. EMNLP was founded because of reviewing biases at ACL against groundbreaking methodological advances, and since then the two have homogenized; “EMNLP reviewing is no longer much of a differentiator” (Church, 2020). We found that theoretically grounded findings in text analysis are often published in non-\*ACL venues (Table 1), but ACL sets the standard for work involving computational text analysis and NLP. Is there no home for groundbreaking integrative or interdisciplinary work in *ACL, such that a new venue is required? Or can we adapt our standards to invite deeper connections to theory and new ways of knowing? + +# Acknowledgments + +This publication was made possible in part by a grant from the American Political Science Association to A.D.M. and G.M.D.D. The statements made and views expressed are solely the responsibility of the authors. A.D.M. is supported by an Amazon Fellowship and a Frederick Jelinek Fellowship. + +# Limitations + +The key limitation of our work is that, when conducting the review of approximately 60 papers (by searching through the ACL Anthology for works in computational social science since 2010), we encountered a skewed distribution of descriptive versus integrative works. In fact, it was relatively simple to find descriptive works, and that section of Table 1 could have been much longer. We also recognize that, due to the mixed nature of our field, scientific and integrative findings are not the only goal—our 'big tent' includes engineers as well, who value gains in performance indicators. Finally, the fact that we have few examples of papers showing a return to theory renders the possibility that our central claim is misinterpreted in a normative way as a mandate. + +# References + +Andrew Delano Abbott. 2004. Methods of discovery: Heuristics for the social sciences (contemporary societies). WW Norton & Company. +Christopher H. Achen and Duncan Snidal. 1989. Rational deterrence theory and comparative case studies. World Politics, 41(2):143-169. +Ashton Anderson, Dan Jurafsky, and Daniel A. McFarland. 2012. Towards a computational history of the ACL: 1980-2008. In Proceedings of the ACL-2012 Special Workshop on Rediscovering 50 Years of Discoveries, pages 13-21, Jeju Island, Korea. Association for Computational Linguistics. +J. Scott Armstrong. 1967. Derivation of theory by means of factor analysis or Tom Swift and his electric factor analysis machine. The American Statistician, 21(5):17-21. +Christian Baden, Christian Pipal, Martijn Schoonvelde, and Mariken A. C. G van der Velden. 2021. Three gaps in computational text analysis methods for social sciences: A research agenda. Communication Methods and Measures, 0(0):1-18. +Christopher A. Bail. 2014. The cultural environment: measuring culture with big data. Theory and Society, 43(3):465-482. + +David Bamman, Brendan O'Connor, and Noah Smith. 2012. Censorship and deletion practices in chinese social media. First Monday, 17(3). +David Bamman, Ted Underwood, and Noah A. Smith. 2014. A Bayesian mixed effects model of literary character. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 370-379, Baltimore, Maryland. Association for Computational Linguistics. +Richard Biernacki. 2015. How to do things with historical texts. American Journal of Cultural Sociology, 3(3):311-352. Copyright - © Palgrave Macmillan, a division of Macmillan Publishers Ltd 2015; Last updated - 2018-09-25. +Amber E Boydstun, Dallas Card, Justin Gross, Paul Resnick, and Noah A Smith. 2014. Tracking the development of media frames within and across policy issues. Unpublished. +Jonathan Chang, Jordan Boyd-Graber, Sean Gerrish, Chong Wang, and David M. Blei. 2009. Reading tea leaves: How humans interpret topic models. In Proceedings of the 22nd International Conference on Neural Information Processing Systems, NIPS'09, page 288-296, Red Hook, NY, USA. Curran Associates Inc. +Kenneth Church. 2005. Last words: Reviewing the reviewers. Computational Linguistics, 31(4):575-578. +Kenneth Ward Church. 2020. Emerging trends: Reviewing the reviewers (again). Natural Language Engineering, 26(2):245-257. +James S. Coleman. 1986. Social theory, social research, and a theory of action. American Journal of Sociology, 91(6):1309-1335. +Dorottya Demszky, Nikhil Garg, Rob Voigt, James Zou, Jesse Shapiro, Matthew Gentzkow, and Dan Jurafsky. 2019. Analyzing polarization in social media: Method and application to tweets on 21 mass shootings. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2970-3005, Minneapolis, Minnesota. Association for Computational Linguistics. +Paul DiMaggio, Manish Nag, and David Blei. 2013. Exploiting affinities between topic modeling and the sociological perspective on culture: Application to newspaper coverage of U.S. government arts funding. *Poetics*, 41(6):570–606. Topic Models and the Cultural Sciences. +Giovanna Maria Dora Dore and Arya D. McCarthy. 2022. Learning to play with the machines in social science research: Bringing the theory back in. In ICML 2022 Workshop on Human-Machine Collaboration and Teaming, Baltimore, Maryland. + +Kawin Ethayarajh and Dan Jurafsky. 2020. Utility is in the eye of the user: A critique of NLP leaderboards. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4846-4853, Online. Association for Computational Linguistics. +Amir Feder, Katherine A. Keith, Emaad Manzoor, Reid Pryzant, Dhanya Sridhar, Zach Wood-Doughty, Jacob Eisenstein, Justin Grimmer, Roi Reichart, Margaret E. Roberts, Brandon M. Stewart, Victor Veitch, and Diyi Yang. 2021. Causal inference in natural language processing: Estimation, prediction, interpretation and beyond. CoRR, abs/2109.00725. +Anjalie Field, Doron Kliger, Shuly Wintner, Jennifer Pan, Dan Jurafsky, and Yulia Tsvetkov. 2018. Framing and agenda-setting in Russian news: a computational analysis of intricate political strategies. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 3570-3580, Brussels, Belgium. Association for Computational Linguistics. +John Gerring. 2004. What is a case study and what is it good for? American Political Science Review, 98(2):341-354. +Justin Grimmer and Brandon M. Stewart. 2013. Text as data: The promise and pitfalls of automatic content analysis methods for political texts. Political Analysis, 21(3):267-297. +Jan Hajic and Eva Hajičová. 2007. Some of our best friends are statisticians. In Text, Speech and Dialogue, pages 2-10, Berlin, Heidelberg. Springer Berlin Heidelberg. +C. A. R. Hoare and C. B. Jones. 1989. Essays in Computing Science. Prentice-Hall, Inc., USA. +Jake Hofman, Miro Dudík, and Daniel G. Goldstein. 2018. Perspective annotation for numerical representations. United States Patent Application. +Jake M Hofman, Duncan J Watts, Susan Athey, Filiz Garip, Thomas L Griffiths, Jon Kleinberg, Helen Margetts, Sendhil Mullainathan, Matthew J Salganik, Simine Vazire, et al. 2021. Integrating explanation and prediction in computational social science. Nature, 595(7866):181-188. +Daniel J. Hopkins and Gary King. 2010. A method of automated nonparametric content analysis for social science. American Journal of Political Science, 54(1):229-247. +Frederick Jelinek. 2005. Some of my best friends are linguists. Language Resources and Evaluation, 39(1):25-34. +Daniel Kahneman and Amos Tversky. 1979. Prospect theory: An analysis of decision under risk. *Econometrica*, 47(2):263-291. + +Gary King, Jennifer Pan, and Margaret E. Roberts. 2013. How censorship in China allows government criticism but silences collective expression. American Political Science Review, 107(2 (May)):1-18. +David Lazer and Jason Radford. 2017. Data ex machina: Introduction to big data. Annual Review of Sociology, 43(1):19-39. +Monica Lee and John Levi Martin. 2015. Coding, counting and cultural cartography. American Journal of Cultural Sociology, 3(1):1-33. +Zachary C. Lipton and Jacob Steinhardt. 2019. Troubling trends in machine learning scholarship: Some ML papers suffer from flaws that could mislead the public and stymie future research. Queue, 17(1):45-77. +Li Lucy, Dorottya Demszky, Patricia Bromley, and Dan Jurafsky. 2020. Content analysis of textbooks via natural language processing: Findings on gender, race, and ethnicity in Texas U.S. history textbooks. AERA Open, 6(3):2332858420940312. +Arya D. McCarthy and Giovanna Maria Dora Dore. 2022. Hong Kong: Longitudinal and synchronic characterisations of protest news between 1998 and 2020. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 2891-2900, Marseille, France. European Language Resources Association. +Arya D. McCarthy, James Scharf, and Giovanna Maria Dora Dore. 2021. A mixed-methods analysis of western and Hong Kong-based reporting on the 2019-2020 protests. In Proceedings of the 5th Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature, pages 178-188, Punta Cana, Dominican Republic (online). Association for Computational Linguistics. +Drew McDermott. 1976. Artificial intelligence meets natural stupidity. SIGART Bull., (57):4-9. +Frederick Mosteller and David L. Wallace. 1963. Inference in an authorship problem. Journal of the American Statistical Association, 58(302):275-309. +Ernest Nagel. 1963. The structure of science: Problems in the logic of scientific explanation. Mind, 72(287). +Laura K. Nelson. 2021. Leveraging the alignment between machine learning and intersectionality: Using word embeddings to measure intersectional experiences of the nineteenth century U.S. South. Poetics, 88:101539. Measure Mohr Culture. +Dong Nguyen, Maria Liakata, Simon DeDeo, Jacob Eisenstein, David Mimno, Rebekah Tromble, and Jane Winters. 2020. How we do things with words: Analyzing text as social and cultural data. Frontiers in Artificial Intelligence, 3. + +Brendan O'Connor, David Bamman, and Noah A Smith. 2011. Computational text analysis for social science: Model complexity and assumptions. In Proc. of the NIPS Workshop on Comptuational Social Science and the Wisdom of Crowds. +Jason Radford and Kenneth Joseph. 2020. Theory in, theory out: The uses of social theory in machine learning for social science. Frontiers in Big Data, 3. +William J Rapaport. 2005. Philosophy of computer science: An introductory course. Teaching philosophy, 28(4):319-341. +Stuart C. Shapiro. 2001. Computer science: The study of procedures. Technical report, Department of Computer Science and Engineering, University of Buffalo. +Han-Chin Shing, Suraj Nair, Ayah Zirikly, Meir Friedenberg, Hal Daumé III, and Philip Resnik. 2018. Expert, crowdsourced, and machine assessment of suicide risk via online postings. In Proceedings of the Fifth Workshop on Computational Linguistics and Clinical Psychology: From Keyboard to Clinic, pages 25-36, New Orleans, LA. Association for Computational Linguistics. +David A. Siegel. 2018. Analyzing computational models. American Journal of Political Science, 62(3):745-759. +Lyn Spillman. 2015. Ghosts of straw men: A reply to Lee and Martin. American Journal of Cultural Sociology, 3(3):365-379. +Philip J. Stone and Earl B. Hunt. 1963. A computer approach to content analysis: Studies using the general inquirer system. In Proceedings of the May 21-23, 1963, Spring Joint Computer Conference, AFIPS '63 (Spring), page 241-256, New York, NY, USA. Association for Computing Machinery. +Sidney Tarrow. 1995. Bridging the quantitative-qualitative divide in political science. American Political Science Review, 89(2):471-474. +Paul Thagard. 1988. Computational Philosophy of Science. MIT Press. +Richard Thaler. 1980. Judgement And Decision Making Under Uncertainty: What Economists Can Learn From Psychology. Risk Analysis in Agriculture: Research and Educational Developments, January 16-18, 1980, Tucson, Arizona 271572, Regional Research Projects $>$ W-149: An Economic Evaluation of Managing Market Risks in Agriculture. +Richard H. Thaler and Cass R. Sunstein. 2009. Nudge: Improving decisions about health, wealth, and happiness. Penguin. +Wouter van Atteveldt and Tai-Quan Peng. 2018. When communication meets computation: Opportunities, challenges, and pitfalls in computational communication science. Communication Methods and Measures, 12(2-3):81-92. + +Sidney Verba. 1996. The citizen as respondent: Sample surveys and american democracy. American Political Science Review, 90(1):1-7. Presidential Address, American Political Science Association, 1995. +Sidney Verba, Kay Lehman Schlozman, Henry Brady, and Norman H. Nie. 1993. Citizen activity: Who participates? what do they say? The American Political Science Review, 87(2):303-318. +Sidney Verba, Kay Lehman Schlozman, and Henry E Brady. 1995. Voice and equality: Civic volunteerism in American politics. Harvard University Press. +Erin Walk, Elizabeth Parker-Magyar, Kiran Garimella, Ahmet Akbiyik, and Fotini Christia. 2022. Social media narratives across platforms in conflict: Evidence from Syria. MIT Political Science Department Research Paper No. 2022-2, available at SSRN. +Hanna Wallach. 2018. Computational social science $\neq$ computer science + social data. Commun. ACM, 61(3):42-44. + +A For every submission: + +A1. Did you describe the limitations of your work? + +Unnumbered; appears on page 5. + +A2. Did you discuss any potential risks of your work? + +This is a position paper. + +A3. Do the abstract and introduction summarize the paper's main claims? + +Left blank. + +A4. Have you used AI writing assistants when working on this paper? + +Left blank. + +B Did you use or create scientific artifacts? + +Left blank. + +B1. Did you cite the creators of artifacts you used? + +Not applicable. Left blank. + +B2. Did you discuss the license or terms for use and / or distribution of any artifacts? + +Not applicable. Left blank. + +B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? + +Not applicable. Left blank. + +B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? + +Not applicable. Left blank. + +B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? + +Not applicable. Left blank. + +B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. + +Not applicable. Left blank. + +C Did you run computational experiments? + +Left blank. + +C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? + +Not applicable. Left blank. + +The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance. + +C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? + +Not applicable. Left blank. + +C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? + +Not applicable. Left blank. + +C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? + +Not applicable. Left blank. + +D Did you use human annotators (e.g., crowdworkers) or research with human participants? + +Left blank. + +D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? + +Not applicable. Left blank. + +D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? + +Not applicable. Left blank. + +D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? + +Not applicable. Left blank. + +D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? + +Not applicable. Left blank. + +D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? + +Not applicable. Left blank. \ No newline at end of file diff --git a/2023/Theory-Grounded Computational Text Analysis/images.zip b/2023/Theory-Grounded Computational Text Analysis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6a0ee523581ab3fc0b465fecaea8db1c2c4c299a --- /dev/null +++ b/2023/Theory-Grounded Computational Text Analysis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3e305d732217684ab94392ca19aa036d5319e149781940e0933c1ddb5dbfcb78 +size 250373 diff --git a/2023/Theory-Grounded Computational Text Analysis/layout.json b/2023/Theory-Grounded Computational Text Analysis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..27d5464ab5bd309e06b35e0c7cd66d800e9b3369 --- /dev/null +++ b/2023/Theory-Grounded Computational Text Analysis/layout.json @@ -0,0 +1,6074 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 148, + 76, + 446, + 94 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 76, + 446, + 94 + ], + "spans": [ + { + "bbox": [ + 148, + 76, + 446, + 94 + ], + "type": "text", + "content": "Theory-Grounded Computational Text Analysis" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 152, + 117, + 442, + 134 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 117, + 442, + 134 + ], + "spans": [ + { + "bbox": [ + 152, + 117, + 442, + 134 + ], + "type": "text", + "content": "Arya D. McCarthy\\* and Giovanna Maria Dora Dore" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 123, + 135, + 473, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 135, + 473, + 150 + ], + "spans": [ + { + "bbox": [ + 123, + 135, + 473, + 150 + ], + "type": "inline_equation", + "content": "\\diamond" + }, + { + "bbox": [ + 123, + 135, + 473, + 150 + ], + "type": "text", + "content": " Center for Language and Speech Processing, Johns Hopkins University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 142, + 151, + 454, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 151, + 454, + 167 + ], + "spans": [ + { + "bbox": [ + 142, + 151, + 454, + 167 + ], + "type": "inline_equation", + "content": "\\spadesuit" + }, + { + "bbox": [ + 142, + 151, + 454, + 167 + ], + "type": "text", + "content": " Krieger School of Arts and Sciences, Johns Hopkins University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 155, + 212, + 202, + 226 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 212, + 202, + 226 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 202, + 226 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 84, + 236, + 274, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 236, + 274, + 439 + ], + "spans": [ + { + "bbox": [ + 84, + 236, + 274, + 439 + ], + "type": "text", + "content": "In this position paper, we argue that computational text analysis lacks and requires organizing principles. A broad space separates its two constituent disciplines—natural language processing and social science—which has to date been sidestepped rather than filled by applying increasingly complex computational models to problems in social science research. We contrast descriptive and integrative findings, and our review of approximately 60 papers on computational text analysis reveals that those from *ACL venues are typically descriptive. The lack of theory began at the area's inception and has, over the decades, grown more important and challenging. A return to theoretically grounded research questions will propel the area from both theoretical and methodological points of view." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 460, + 155, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 460, + 155, + 473 + ], + "spans": [ + { + "bbox": [ + 68, + 460, + 155, + 473 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 481, + 291, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 481, + 291, + 698 + ], + "spans": [ + { + "bbox": [ + 67, + 481, + 291, + 698 + ], + "type": "text", + "content": "Computational text analysis methods—an umbrella combining natural language processing with social science—are in a honeymoon period (Lazer and Radford, 2017; van Atteveldt and Peng, 2018). Today's social scientist might reach for the tools of computer science for their speed, scale, granularity, and consistency; for instance, natural language processing offers \"to analyze signals ranging from simple lexical cues to word clusters to choices of syntactic structure\" (Boydstun et al., 2014). The numerical outputs tell a story that is simple, easy to make sense of, and in that regard comforting. Conversely, today's computer scientist may see the problems of social science as answerable by objectivity and reductionism, eschewing interpretation for quantitative analysis." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 698, + 291, + 752 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 698, + 291, + 752 + ], + "spans": [ + { + "bbox": [ + 67, + 698, + 291, + 752 + ], + "type": "text", + "content": "The conclusion of this reasoning, and the dominant stance in computational social science, is a reliance on machines alone to answer questions in the field, surrendering to their supposed objectivity" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 213, + 526, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 213, + 526, + 280 + ], + "spans": [ + { + "bbox": [ + 302, + 213, + 526, + 280 + ], + "type": "text", + "content": "or impartiality. Can a machine's output go beyond descriptive catalogs of evidence, accelerating understanding of processes and motivations? From our experience, computers are nowhere near supplanting humans in interpreting social science results.1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 280, + 526, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 280, + 526, + 428 + ], + "spans": [ + { + "bbox": [ + 302, + 280, + 526, + 428 + ], + "type": "text", + "content": "An interdisciplinary inquiry must go farther than matching computational techniques to social science questions (O'Connor et al., 2011; Nguyen et al., 2020). It embraces synergistic methodology and connects the norms and standards of evidence from both. This means partnering computer science's preference for the structured, generalizable, and objective with the unstructured, critical, and contextual which the social sciences champion. This level of interdisciplinarity addresses the question raised by descriptive findings: So what?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 429, + 526, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 429, + 526, + 605 + ], + "spans": [ + { + "bbox": [ + 302, + 429, + 526, + 605 + ], + "type": "text", + "content": "We see theory as the solution, empowering rather than shackling investigations. What this paper advocates is not one particular theory—certainly these are myriad, and “even subject matter which has been under intensive and prolonged study remains at the unsettled periphery of research” (Nagel, 1963). Instead, we expand on our prior work (Dore and McCarthy, 2022) to clarify calls echoed for decades by computational and social science (McDermott, 1976; Jelinek, 2005; Hajic and Hajicova, 2007; Hofman et al., 2018; Lipton and Steinhardt, 2019; Baden et al., 2021). Underlying each, we find, is the urge to return to theory, which we espouse herein." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 615, + 461, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 615, + 461, + 629 + ], + "spans": [ + { + "bbox": [ + 302, + 615, + 461, + 629 + ], + "type": "text", + "content": "2 Description vs. Integration" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 636, + 526, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 636, + 526, + 703 + ], + "spans": [ + { + "bbox": [ + 302, + 636, + 526, + 703 + ], + "type": "text", + "content": "We contrast descriptive findings and theoretical analysis. An example of a descriptive finding is that an apple falls, or that it falls faster when pushed than dropped, or even that it falls at a particular rate estimated with some standard error by a complex" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 711, + 526, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 711, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 711, + 526, + 772 + ], + "type": "text", + "content": "1See, e.g., Noam Chomsky's remark on GPT-3: \"You can't go to a physics conference and say: I've got a great theory. It accounts for everything and is so simple it can be captured in two words: 'Anything goes.' All known and unknown laws of nature are accommodated... Of course, everything impossible is accommodated also. That's GPT-3.\" [link]" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 84, + 760, + 162, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 760, + 162, + 772 + ], + "spans": [ + { + "bbox": [ + 84, + 760, + 162, + 772 + ], + "type": "text", + "content": "* Equal contribution." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 286, + 780, + 311, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 311, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 311, + 791 + ], + "type": "text", + "content": "1586" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 135, + 795, + 459, + 805 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 795, + 459, + 805 + ], + "spans": [ + { + "bbox": [ + 135, + 795, + 459, + 805 + ], + "type": "text", + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 219, + 806, + 375, + 817 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 806, + 375, + 817 + ], + "spans": [ + { + "bbox": [ + 219, + 806, + 375, + 817 + ], + "type": "text", + "content": "Volume 2: Short Papers, pages 1586-1594" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "spans": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "text", + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 290, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 290, + 152 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 290, + 152 + ], + "type": "text", + "content": "interpolation. A theoretical analysis of the same phenomenon, credited to Newton, is that a fundamental force acts upon the apple, and that this same force governs the motion of the heavens. The theoretical analysis links the finding about the world critically to a broader body of knowledge and context." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 153, + 291, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 153, + 291, + 328 + ], + "spans": [ + { + "bbox": [ + 69, + 153, + 291, + 328 + ], + "type": "text", + "content": "Despite advances in causal inference in NLP, the descriptive is all that a machine can provide to the social sciences (Feder et al., 2021). Certainly the methods of computational text analysis have advanced since the General Inquirer (Stone and Hunt, 1963) and Mosteller and Wallace's statistical inference of text authorship (1963). But methods are means, not ends. They uncover more descriptive findings in data: the rate of an apple's fall, the topics of refugees' tweets (Walk et al., 2022), the space given to marginalized groups in textbooks (Lucy et al., 2020), or patterns of state censorship (Bamman et al., 2012; King et al., 2013)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 330, + 291, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 330, + 291, + 439 + ], + "spans": [ + { + "bbox": [ + 66, + 330, + 291, + 439 + ], + "type": "text", + "content": "The foils to descriptive findings are integrative findings (Hofman et al., 2021), which offer causal explanations that enable future predictions—a theory, or as a 'model' in the sense of the Standard Model, rather than of a statistical model. Integrative findings can either offer new theories or couch their explanations in existing theories—but the theory is essential either way." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 449, + 188, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 449, + 188, + 464 + ], + "spans": [ + { + "bbox": [ + 67, + 449, + 188, + 464 + ], + "type": "text", + "content": "3 We Don't Integrate" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 472, + 291, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 472, + 291, + 621 + ], + "spans": [ + { + "bbox": [ + 67, + 472, + 291, + 621 + ], + "type": "text", + "content": "To contrast descriptive and integrative findings, we reviewed approximately 60 papers in computational text analysis published in *ACL venues. In Table 1, we describe several of these in terms of their descriptive or theory-grounded contributions." + }, + { + "bbox": [ + 67, + 472, + 291, + 621 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 472, + 291, + 621 + ], + "type": "text", + "content": " Descriptive papers may refer to social science theories or make generalizable claims, as when Demszky et al. (2019) write, \"The shooter's race appears to play a role in topic preference: if the shooter is white, Democrats become more likely to focus on shooter's identity,\" but they do not link to the two to each other." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 623, + 291, + 731 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 623, + 291, + 731 + ], + "spans": [ + { + "bbox": [ + 67, + 623, + 291, + 731 + ], + "type": "text", + "content": "An excellent theory-grounded quantitative work is Nelson (2021); she confirms some of the most compelling features of identity theory, specifically that identities based on race were most distinguished by cultural discourse, whereas those based on gender by the domestic and the economic discourse. Similarly, we conducted theory-grounded quantitative work to investigate the application of the protest" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 71, + 526, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 179 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 179 + ], + "type": "text", + "content": "paradigm and thematic framing in how western- and Hong Kong based newspapers portray protests in Hong Kong (McCarthy et al., 2021; McCarthy and Dore, 2022). Generally, it remains challenging to find computational social science papers in *ACL venues that go beyond description and prediction, advancing theory. Why is this? We believe it stemmed from the field's \"empirical turn\".3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 180, + 526, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 180, + 526, + 275 + ], + "spans": [ + { + "bbox": [ + 302, + 180, + 526, + 275 + ], + "type": "text", + "content": "Few remember when the meetings of ACL offered a few dozen papers, all entrenched in formalisms and linguistic theories. Arguably, 1996 was a turning point when the founders of SIGDAT held the first EMNLP at Penn under the auspices of the ACL. This gave a spotlight to the few but growing empiricists in the field and drew in more." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 277, + 526, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 277, + 526, + 397 + ], + "spans": [ + { + "bbox": [ + 302, + 277, + 526, + 397 + ], + "type": "text", + "content": "EMNLP began a half-decade of measurable reorganization the field (Anderson et al., 2012). That EMNLP remains affiliated with ACL keeps the language-focused machine learning practitioners in our tent. The slow blurring of boundaries between each *ACL conference's expectations (Church, 2020) increases this unity. Both groups belong under this tent. But without a doubt, one group's voice is becoming less heard." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 400, + 525, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 400, + 525, + 562 + ], + "spans": [ + { + "bbox": [ + 302, + 400, + 525, + 562 + ], + "type": "text", + "content": "Publication venues within the ACL focus on methods over theory.5 Techniques are taken off the shelf without critical examination because these are \"the best\" (often \"state of the art\") for their purposes (Ethayarajh and Jurafsky, 2020). This widens the gap between theoretical and empirical work.6 Hopkins and King (2010) claim, \"computer scientists may be interested in finding the needle in the haystack... social scientists are more commonly interested in characterizing the haystack\"—evincing the value of broader context.7 Wallach (2018), quoting Hopkins and King, explains that the two groups" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 575, + 525, + 635 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 575, + 525, + 635 + ], + "spans": [ + { + "bbox": [ + 302, + 575, + 525, + 635 + ], + "type": "text", + "content": "A lesser reason is the challenge of serving two masters: adequately covering both the theoretical and methodological components within 8 pages. We recently received two reviews for an *ACL submission: one advocating for more of the social science context in the main text by eschewing methods to the appendix, and the other instructing us to do the opposite." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 636, + 524, + 648 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 524, + 648 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 524, + 648 + ], + "type": "text", + "content": "4 And its predecessor the Workshop on Very Large Corpora." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 648, + 525, + 689 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 648, + 525, + 689 + ], + "spans": [ + { + "bbox": [ + 302, + 648, + 525, + 689 + ], + "type": "text", + "content": "This is due to the outsized influence of computer science, often seen as the science of method (Hoare and Jones, 1989; Shapiro, 2001), when not instead seen as an engineering discipline (Rapaport, 2005)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 689, + 525, + 730 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 689, + 525, + 730 + ], + "spans": [ + { + "bbox": [ + 302, + 689, + 525, + 730 + ], + "type": "text", + "content": "6 A related criticism is that empirical research has narrowed to focus on 'easy' questions that its tools can address (Coleman, 1986; Baden et al., 2021), especially when research questions are baked into the design of the task." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 730, + 525, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 730, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 730, + 525, + 772 + ], + "type": "text", + "content": "7As evidence, see Siegel (2018): \"We usually don't know about causation, and we often don't necessarily care... the objective is more to predict than it is to understand the world... It just needs to work; prediction trumps explanation.\"" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 67, + 740, + 290, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 740, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 740, + 290, + 772 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 740, + 290, + 772 + ], + "type": "text", + "content": "Following Lipton and Steinhardt (2019), we only describe papers by established researchers to \"avoid singling out junior students... who lack the opportunity to reply symmetrically\"." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1587" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 69, + 523, + 405 + ], + "blocks": [ + { + "bbox": [ + 70, + 69, + 523, + 405 + ], + "lines": [ + { + "bbox": [ + 70, + 69, + 523, + 405 + ], + "spans": [ + { + "bbox": [ + 70, + 69, + 523, + 405 + ], + "type": "table", + "html": "
Descriptive
Chang et al. (2009)The article presents new quantitative methods to measure semantic meaning in inferred topics. The authors emphasize the qualitative relevance of their findings as it validates the use of topics for corpus exploration and information retrieval. However, their working hypothesis and empirical findings are not connected to the extremely relevant field of communication theory.
Bamman et al. (2012)The article presents the first large-scale analysis of political content censorship in social media. The authors miss the opportunity to relate their hypothesis and findings to censorship theory, a natural theoretical context for the research, which would strengthen the relevance and generalizability of the findings.
Field et al. (2018)The article discusses media manipulation in Russia in the context of agenda-setting and framing, the tools that Russian state-owned (or heavily influenced) media outlets use to distract public attention from domestic economic politics. The authors implicitly refer to propaganda theory and autocratic theory throughout the article even though their findings are not discussed in relation to these theories.
Demszky et al. (2019)The article applies “a more comprehensive NLP framework to study linguistic aspects of polarization in social media”. While the article implicitly refers to theories of social conformity and social conflict, the findings are not linked or discussed (either explicitly or implicitly) to the theoretical frameworks that the authors touch on in their §1.
Integrative
DiMaggio et al. (2013)The article describes how topic models of newspaper articles help to study the politicization of government support for arts organizations and artists in the late 1980s in the US. The authors clearly define the theoretical context of their investigation and emphasize the relationship between theory and method throughout the paper.
Bamman et al. (2014)The article validates an empirical model that “employs multiple effects to account for the influence of extra-linguistic information (such as author)” by testing specific parameters against a variety of theory-based hypotheses derived from writing styles theories of England between 1700 and 1899.
Nelson (2021)The article argues that the full potential of machine learning can be better realized by “leveraging the epistemological alignment between machine learning and inductive research.” The author empirically demonstrates this by anchoring in identity theory a word embedding model of first-person narratives of the nineteenth-century U.S. South.
", + "image_path": "26ad4830629417904c7c3f2db1c2e350e95681b63cb8bd1424c2dc4341fdbb11.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 413, + 524, + 426 + ], + "lines": [ + { + "bbox": [ + 67, + 413, + 524, + 426 + ], + "spans": [ + { + "bbox": [ + 67, + 413, + 524, + 426 + ], + "type": "text", + "content": "Table 1: Contrast between work in computational text analysis with descriptive findings versus integrative findings." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 447, + 290, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 447, + 290, + 541 + ], + "spans": [ + { + "bbox": [ + 67, + 447, + 290, + 541 + ], + "type": "text", + "content": "are interested in very different research questions, and that computational social science must be more than computer science with social data; it must strive for valid explanatory models. In the same vein, at ACL 2022, ACL fellow Eduard Hovy remarked that NLP must be more than \"just machine learning on corpora\"." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 544, + 290, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 544, + 290, + 705 + ], + "spans": [ + { + "bbox": [ + 67, + 544, + 290, + 705 + ], + "type": "text", + "content": "Social scientists are also coming to terms with the meaning of computational techniques applied more often in social science (Bail, 2014; Biernacki, 2015; Lee and Martin, 2015; Spillman, 2015). The focus of the debates, however, is on which methods are best suited to extract meaning from text, without addressing any theoretical considerations related to the methods or whether a theoretical framework for those methods even exists. The discussions on whether computational methods make social science research more efficient, reliable, and reproducible overtake attempts at theory-building." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 720, + 176, + 735 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 720, + 176, + 735 + ], + "spans": [ + { + "bbox": [ + 67, + 720, + 176, + 735 + ], + "type": "text", + "content": "4 Moving Forward" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 746, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 746, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 746, + 291, + 773 + ], + "type": "text", + "content": "We are not denying the value of computational approaches to analyzing text. Certainly, comput" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 447, + 526, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 447, + 526, + 568 + ], + "spans": [ + { + "bbox": [ + 302, + 447, + 526, + 568 + ], + "type": "text", + "content": "ing can be an instrumental approach for modeling and understanding social complexity. This does not mean that other approaches, such as historical, ethnographic, or mathematical, become irrelevant. On the contrary, computational methods necessarily (whether awarely or not) rely on these earlier approaches to add value, in terms of improving our explanations and understanding (Radford and Joseph, 2020)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 570, + 526, + 758 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 570, + 526, + 758 + ], + "spans": [ + { + "bbox": [ + 302, + 570, + 526, + 758 + ], + "type": "text", + "content": "As we are a field that prioritizes methods, consider the seminal book on methods in science: Abbott (2004) taxonomizes scientific ways of knowing. Its five broad categories are ethnography, historical narration, standard causal analysis, small-" + }, + { + "bbox": [ + 302, + 570, + 526, + 758 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 302, + 570, + 526, + 758 + ], + "type": "text", + "content": " comparison, and formal modeling. We in NLP myopically choose the third and fifth of these, ignoring the value of the others. But the broader point of Methods of Discovery is not methods. It is the research question. Any methodology should be grounded in the question, not incremental tweaks and reviewers' comfort (Church, 2020). This admits even qualitative or mixed-method approaches to text analysis." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 760, + 524, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 760, + 524, + 773 + ], + "spans": [ + { + "bbox": [ + 314, + 760, + 524, + 773 + ], + "type": "text", + "content": "The role of humans in scientific inquiry is nothing" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 780, + 309, + 791 + ], + "type": "text", + "content": "1588" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 66, + 71, + 290, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 71, + 290, + 246 + ], + "spans": [ + { + "bbox": [ + 66, + 71, + 290, + 246 + ], + "type": "text", + "content": "new. Using qualitative analysis to complement quantitative techniques has its roots in Achen and Snidal (1989)'s recommendation to use historical case studies as a complement to statistical research. Their plea was strengthened by Verba's work in the early 1990s (Verba et al., 1993, 1995; Verba, 1996) and Tarrow (1995), who openly called for bridging qualitative and quantitative modes of research in social science. In doing so, they have enriched the field with critical methodological innovations (Gerring, 2004), benefiting from the recognition that \"quantitative methods must augment humans, not replace them\" (Grimmer and Stewart, 2013, 4)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 248, + 290, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 248, + 290, + 396 + ], + "spans": [ + { + "bbox": [ + 66, + 248, + 290, + 396 + ], + "type": "text", + "content": "The field can draw more from social science's rich tradition of inductive theory-building and interpretation to develop its theoretical approach—to prize either induction or deduction alone is a myth of scientific procedure (Thagard, 1988), but the melding of the two opens new doors. Rather than eschewing the complexity (a criticism leveled by Baden et al., 2021), it should put complexity at the center of its ontology on the basis that there are no immutable laws in social life or optimal solutions to social problems." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 396, + 291, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 396, + 291, + 587 + ], + "spans": [ + { + "bbox": [ + 66, + 396, + 291, + 587 + ], + "type": "text", + "content": "Skepticism can linger toward findings not drawn from the standard practices of one's own field; indeed, social science was long skeptical of computational contributions (Armstrong, 1967). We believe that this drives the hyperfocus on improving a few accepted methods instead of exploring more broadly. If the doorway between disciplines is only narrowly open, this reflects a lack of appreciation for each field's ways of knowing. The disciplinary divide keeps computational researchers from embracing methods beyond standard causal analysis or formal modeling, so the interpreter-centric richness allowed by histories, ethnographies, and small-" + }, + { + "bbox": [ + 66, + 396, + 291, + 587 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 66, + 396, + 291, + 587 + ], + "type": "text", + "content": " exploration are precluded." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 597, + 147, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 597, + 147, + 609 + ], + "spans": [ + { + "bbox": [ + 67, + 597, + 147, + 609 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 618, + 291, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 618, + 291, + 714 + ], + "spans": [ + { + "bbox": [ + 66, + 618, + 291, + 714 + ], + "type": "text", + "content": "We have explained the distinction between descriptive and theoretical findings as it pertains to computational text analysis. The bulk of work we found provided vast descriptive findings, often of high quality, but not giving back to questions of theory. We offer several suggestions on how to 'push the pendulum back' by prioritizing theory-building or" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 302, + 71, + 524, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 524, + 111 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 524, + 111 + ], + "type": "text", + "content": "theory-affirming research questions and accepting whichever methods are best suited toward answering it—not only the familiar and entrenched ones." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 111, + 526, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 111, + 526, + 300 + ], + "spans": [ + { + "bbox": [ + 302, + 111, + 526, + 300 + ], + "type": "text", + "content": "We are not the first to advocate for a shift in the patterns of applying computational techniques to real-world problems. There is a steady drumbeat from voices in the field advocating careful approaches (Nagel, 1963; McDermott, 1976; Jelinek, 2005; Hajic and Hajicova, 2007; Hofman et al., 2018; Lipton and Steinhardt, 2019; Baden et al., 2021). What we see underlying all of these—those writing against 'mathiness' and speculation, advocating for clear evaluation over anecdotes, criticizing textual researchers' dilution of conceptual standards, highlighting work that ties linguistic information into complex models—is an unspoken, perhaps unrealized, call for a return to theory." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 301, + 525, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 301, + 525, + 434 + ], + "spans": [ + { + "bbox": [ + 302, + 301, + 525, + 434 + ], + "type": "text", + "content": "Not only do we aver that incorporating theory is essential; but also, other fields have strengthened themselves when espousing organizing principles beyond those of their progenitors. Behavioral economics is a success story here. It transcended the neat (but psychosocially stripped) mathematics it draws from to acknowledge deviations from rationality and blend economics with cognitive science (Kahneman and Tversky, 1979; Thaler, 1980; Thaler and Sunstein, 2009)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 436, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 436, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 436, + 525, + 772 + ], + "type": "text", + "content": "For scientific—not simply engineering—advances to arise from the *ACL community, authors and reviewers alike must resist the temptation toward incremental, ‘safe’ research and follow Church (2005): “Controversial papers are great; boring unobjectionable incremental papers are not.” In reviewing new research, we should privilege not only work that presents new and unusual computational methods, but also interactions between computational and humanistic approaches to answering research questions. EMNLP was founded because of reviewing biases at ACL against groundbreaking methodological advances, and since then the two have homogenized; “EMNLP reviewing is no longer much of a differentiator” (Church, 2020). We found that theoretically grounded findings in text analysis are often published in non-\\*ACL venues (Table 1), but ACL sets the standard for work involving computational text analysis and NLP. Is there no home for groundbreaking integrative or interdisciplinary work in *ACL, such that a new venue is required? Or can we adapt our standards to invite deeper connections to theory and new ways of knowing?" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 719, + 290, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 719, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 719, + 290, + 772 + ], + "type": "text", + "content": "8 Expertise plays a role as well (Shing et al., 2018), which is why Mechanical Turk doesn't fill the need for qualitative analysis. This is exemplified by Radford and Joseph (2020)'s observation of \"non-expert annotators providing unreliable annotations, even after a discussion period\"." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1589" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 71, + 166, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 71, + 166, + 85 + ], + "spans": [ + { + "bbox": [ + 68, + 71, + 166, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 92, + 291, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 92, + 291, + 174 + ], + "spans": [ + { + "bbox": [ + 67, + 92, + 291, + 174 + ], + "type": "text", + "content": "This publication was made possible in part by a grant from the American Political Science Association to A.D.M. and G.M.D.D. The statements made and views expressed are solely the responsibility of the authors. A.D.M. is supported by an Amazon Fellowship and a Frederick Jelinek Fellowship." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 68, + 183, + 131, + 195 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 183, + 131, + 195 + ], + "spans": [ + { + "bbox": [ + 68, + 183, + 131, + 195 + ], + "type": "text", + "content": "Limitations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 205, + 291, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 205, + 291, + 421 + ], + "spans": [ + { + "bbox": [ + 67, + 205, + 291, + 421 + ], + "type": "text", + "content": "The key limitation of our work is that, when conducting the review of approximately 60 papers (by searching through the ACL Anthology for works in computational social science since 2010), we encountered a skewed distribution of descriptive versus integrative works. In fact, it was relatively simple to find descriptive works, and that section of Table 1 could have been much longer. We also recognize that, due to the mixed nature of our field, scientific and integrative findings are not the only goal—our 'big tent' includes engineers as well, who value gains in performance indicators. Finally, the fact that we have few examples of papers showing a return to theory renders the possibility that our central claim is misinterpreted in a normative way as a mandate." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 444, + 127, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 444, + 127, + 456 + ], + "spans": [ + { + "bbox": [ + 68, + 444, + 127, + 456 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 463, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 69, + 463, + 291, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 463, + 291, + 497 + ], + "spans": [ + { + "bbox": [ + 69, + 463, + 291, + 497 + ], + "type": "text", + "content": "Andrew Delano Abbott. 2004. Methods of discovery: Heuristics for the social sciences (contemporary societies). WW Norton & Company." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 504, + 291, + 539 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 504, + 291, + 539 + ], + "spans": [ + { + "bbox": [ + 69, + 504, + 291, + 539 + ], + "type": "text", + "content": "Christopher H. Achen and Duncan Snidal. 1989. Rational deterrence theory and comparative case studies. World Politics, 41(2):143-169." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 546, + 291, + 614 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 546, + 291, + 614 + ], + "spans": [ + { + "bbox": [ + 69, + 546, + 291, + 614 + ], + "type": "text", + "content": "Ashton Anderson, Dan Jurafsky, and Daniel A. McFarland. 2012. Towards a computational history of the ACL: 1980-2008. In Proceedings of the ACL-2012 Special Workshop on Rediscovering 50 Years of Discoveries, pages 13-21, Jeju Island, Korea. Association for Computational Linguistics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 621, + 291, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 621, + 291, + 666 + ], + "spans": [ + { + "bbox": [ + 69, + 621, + 291, + 666 + ], + "type": "text", + "content": "J. Scott Armstrong. 1967. Derivation of theory by means of factor analysis or Tom Swift and his electric factor analysis machine. The American Statistician, 21(5):17-21." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 675, + 291, + 730 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 675, + 291, + 730 + ], + "spans": [ + { + "bbox": [ + 69, + 675, + 291, + 730 + ], + "type": "text", + "content": "Christian Baden, Christian Pipal, Martijn Schoonvelde, and Mariken A. C. G van der Velden. 2021. Three gaps in computational text analysis methods for social sciences: A research agenda. Communication Methods and Measures, 0(0):1-18." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 738, + 291, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 738, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 738, + 291, + 772 + ], + "type": "text", + "content": "Christopher A. Bail. 2014. The cultural environment: measuring culture with big data. Theory and Society, 43(3):465-482." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 526, + 772 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 304, + 72, + 526, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 526, + 106 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 526, + 106 + ], + "type": "text", + "content": "David Bamman, Brendan O'Connor, and Noah Smith. 2012. Censorship and deletion practices in chinese social media. First Monday, 17(3)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 115, + 526, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 115, + 526, + 181 + ], + "spans": [ + { + "bbox": [ + 304, + 115, + 526, + 181 + ], + "type": "text", + "content": "David Bamman, Ted Underwood, and Noah A. Smith. 2014. A Bayesian mixed effects model of literary character. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 370-379, Baltimore, Maryland. Association for Computational Linguistics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 190, + 526, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 190, + 526, + 246 + ], + "spans": [ + { + "bbox": [ + 304, + 190, + 526, + 246 + ], + "type": "text", + "content": "Richard Biernacki. 2015. How to do things with historical texts. American Journal of Cultural Sociology, 3(3):311-352. Copyright - © Palgrave Macmillan, a division of Macmillan Publishers Ltd 2015; Last updated - 2018-09-25." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 254, + 526, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 254, + 526, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 254, + 526, + 300 + ], + "type": "text", + "content": "Amber E Boydstun, Dallas Card, Justin Gross, Paul Resnick, and Noah A Smith. 2014. Tracking the development of media frames within and across policy issues. Unpublished." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 308, + 526, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 526, + 385 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 526, + 385 + ], + "type": "text", + "content": "Jonathan Chang, Jordan Boyd-Graber, Sean Gerrish, Chong Wang, and David M. Blei. 2009. Reading tea leaves: How humans interpret topic models. In Proceedings of the 22nd International Conference on Neural Information Processing Systems, NIPS'09, page 288-296, Red Hook, NY, USA. Curran Associates Inc." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 394, + 526, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 394, + 526, + 428 + ], + "spans": [ + { + "bbox": [ + 304, + 394, + 526, + 428 + ], + "type": "text", + "content": "Kenneth Church. 2005. Last words: Reviewing the reviewers. Computational Linguistics, 31(4):575-578." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 437, + 526, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 526, + 470 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 526, + 470 + ], + "type": "text", + "content": "Kenneth Ward Church. 2020. Emerging trends: Reviewing the reviewers (again). Natural Language Engineering, 26(2):245-257." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 480, + 526, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 480, + 526, + 513 + ], + "spans": [ + { + "bbox": [ + 304, + 480, + 526, + 513 + ], + "type": "text", + "content": "James S. Coleman. 1986. Social theory, social research, and a theory of action. American Journal of Sociology, 91(6):1309-1335." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 522, + 526, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 526, + 633 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 526, + 633 + ], + "type": "text", + "content": "Dorottya Demszky, Nikhil Garg, Rob Voigt, James Zou, Jesse Shapiro, Matthew Gentzkow, and Dan Jurafsky. 2019. Analyzing polarization in social media: Method and application to tweets on 21 mass shootings. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2970-3005, Minneapolis, Minnesota. Association for Computational Linguistics." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 641, + 526, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 526, + 708 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 526, + 708 + ], + "type": "text", + "content": "Paul DiMaggio, Manish Nag, and David Blei. 2013. Exploiting affinities between topic modeling and the sociological perspective on culture: Application to newspaper coverage of U.S. government arts funding. *Poetics*, 41(6):570–606. Topic Models and the Cultural Sciences." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 716, + 526, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 716, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 716, + 526, + 772 + ], + "type": "text", + "content": "Giovanna Maria Dora Dore and Arya D. McCarthy. 2022. Learning to play with the machines in social science research: Bringing the theory back in. In ICML 2022 Workshop on Human-Machine Collaboration and Teaming, Baltimore, Maryland." + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "text", + "content": "1590" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 291, + 139 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 291, + 139 + ], + "type": "text", + "content": "Kawin Ethayarajh and Dan Jurafsky. 2020. Utility is in the eye of the user: A critique of NLP leaderboards. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4846-4853, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 148, + 290, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 148, + 290, + 226 + ], + "spans": [ + { + "bbox": [ + 69, + 148, + 290, + 226 + ], + "type": "text", + "content": "Amir Feder, Katherine A. Keith, Emaad Manzoor, Reid Pryzant, Dhanya Sridhar, Zach Wood-Doughty, Jacob Eisenstein, Justin Grimmer, Roi Reichart, Margaret E. Roberts, Brandon M. Stewart, Victor Veitch, and Diyi Yang. 2021. Causal inference in natural language processing: Estimation, prediction, interpretation and beyond. CoRR, abs/2109.00725." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 236, + 290, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 236, + 290, + 324 + ], + "spans": [ + { + "bbox": [ + 69, + 236, + 290, + 324 + ], + "type": "text", + "content": "Anjalie Field, Doron Kliger, Shuly Wintner, Jennifer Pan, Dan Jurafsky, and Yulia Tsvetkov. 2018. Framing and agenda-setting in Russian news: a computational analysis of intricate political strategies. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 3570-3580, Brussels, Belgium. Association for Computational Linguistics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 334, + 290, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 334, + 290, + 368 + ], + "spans": [ + { + "bbox": [ + 69, + 334, + 290, + 368 + ], + "type": "text", + "content": "John Gerring. 2004. What is a case study and what is it good for? American Political Science Review, 98(2):341-354." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 378, + 290, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 378, + 290, + 422 + ], + "spans": [ + { + "bbox": [ + 69, + 378, + 290, + 422 + ], + "type": "text", + "content": "Justin Grimmer and Brandon M. Stewart. 2013. Text as data: The promise and pitfalls of automatic content analysis methods for political texts. Political Analysis, 21(3):267-297." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 433, + 290, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 433, + 290, + 477 + ], + "spans": [ + { + "bbox": [ + 69, + 433, + 290, + 477 + ], + "type": "text", + "content": "Jan Hajic and Eva Hajičová. 2007. Some of our best friends are statisticians. In Text, Speech and Dialogue, pages 2-10, Berlin, Heidelberg. Springer Berlin Heidelberg." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 487, + 290, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 487, + 290, + 510 + ], + "spans": [ + { + "bbox": [ + 69, + 487, + 290, + 510 + ], + "type": "text", + "content": "C. A. R. Hoare and C. B. Jones. 1989. Essays in Computing Science. Prentice-Hall, Inc., USA." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 520, + 290, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 520, + 290, + 554 + ], + "spans": [ + { + "bbox": [ + 69, + 520, + 290, + 554 + ], + "type": "text", + "content": "Jake Hofman, Miro Dudík, and Daniel G. Goldstein. 2018. Perspective annotation for numerical representations. United States Patent Application." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 564, + 290, + 629 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 564, + 290, + 629 + ], + "spans": [ + { + "bbox": [ + 69, + 564, + 290, + 629 + ], + "type": "text", + "content": "Jake M Hofman, Duncan J Watts, Susan Athey, Filiz Garip, Thomas L Griffiths, Jon Kleinberg, Helen Margetts, Sendhil Mullainathan, Matthew J Salganik, Simine Vazire, et al. 2021. Integrating explanation and prediction in computational social science. Nature, 595(7866):181-188." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 640, + 290, + 684 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 640, + 290, + 684 + ], + "spans": [ + { + "bbox": [ + 69, + 640, + 290, + 684 + ], + "type": "text", + "content": "Daniel J. Hopkins and Gary King. 2010. A method of automated nonparametric content analysis for social science. American Journal of Political Science, 54(1):229-247." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 694, + 291, + 728 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 694, + 291, + 728 + ], + "spans": [ + { + "bbox": [ + 69, + 694, + 291, + 728 + ], + "type": "text", + "content": "Frederick Jelinek. 2005. Some of my best friends are linguists. Language Resources and Evaluation, 39(1):25-34." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 69, + 738, + 290, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 738, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 738, + 290, + 772 + ], + "type": "text", + "content": "Daniel Kahneman and Amos Tversky. 1979. Prospect theory: An analysis of decision under risk. *Econometrica*, 47(2):263-291." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 524, + 772 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 304, + 72, + 524, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 524, + 116 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 524, + 116 + ], + "type": "text", + "content": "Gary King, Jennifer Pan, and Margaret E. Roberts. 2013. How censorship in China allows government criticism but silences collective expression. American Political Science Review, 107(2 (May)):1-18." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 127, + 524, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 127, + 524, + 160 + ], + "spans": [ + { + "bbox": [ + 304, + 127, + 524, + 160 + ], + "type": "text", + "content": "David Lazer and Jason Radford. 2017. Data ex machina: Introduction to big data. Annual Review of Sociology, 43(1):19-39." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 172, + 524, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 172, + 524, + 206 + ], + "spans": [ + { + "bbox": [ + 304, + 172, + 524, + 206 + ], + "type": "text", + "content": "Monica Lee and John Levi Martin. 2015. Coding, counting and cultural cartography. American Journal of Cultural Sociology, 3(1):1-33." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 216, + 524, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 216, + 524, + 261 + ], + "spans": [ + { + "bbox": [ + 304, + 216, + 524, + 261 + ], + "type": "text", + "content": "Zachary C. Lipton and Jacob Steinhardt. 2019. Troubling trends in machine learning scholarship: Some ML papers suffer from flaws that could mislead the public and stymie future research. Queue, 17(1):45-77." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 272, + 524, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 272, + 524, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 272, + 524, + 327 + ], + "type": "text", + "content": "Li Lucy, Dorottya Demszky, Patricia Bromley, and Dan Jurafsky. 2020. Content analysis of textbooks via natural language processing: Findings on gender, race, and ethnicity in Texas U.S. history textbooks. AERA Open, 6(3):2332858420940312." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 338, + 524, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 524, + 416 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 524, + 416 + ], + "type": "text", + "content": "Arya D. McCarthy and Giovanna Maria Dora Dore. 2022. Hong Kong: Longitudinal and synchronic characterisations of protest news between 1998 and 2020. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 2891-2900, Marseille, France. European Language Resources Association." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 428, + 524, + 527 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 428, + 524, + 527 + ], + "spans": [ + { + "bbox": [ + 304, + 428, + 524, + 527 + ], + "type": "text", + "content": "Arya D. McCarthy, James Scharf, and Giovanna Maria Dora Dore. 2021. A mixed-methods analysis of western and Hong Kong-based reporting on the 2019-2020 protests. In Proceedings of the 5th Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature, pages 178-188, Punta Cana, Dominican Republic (online). Association for Computational Linguistics." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 538, + 524, + 560 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 538, + 524, + 560 + ], + "spans": [ + { + "bbox": [ + 304, + 538, + 524, + 560 + ], + "type": "text", + "content": "Drew McDermott. 1976. Artificial intelligence meets natural stupidity. SIGART Bull., (57):4-9." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 571, + 524, + 605 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 571, + 524, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 571, + 524, + 605 + ], + "type": "text", + "content": "Frederick Mosteller and David L. Wallace. 1963. Inference in an authorship problem. Journal of the American Statistical Association, 58(302):275-309." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 616, + 524, + 640 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 616, + 524, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 616, + 524, + 640 + ], + "type": "text", + "content": "Ernest Nagel. 1963. The structure of science: Problems in the logic of scientific explanation. Mind, 72(287)." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 650, + 524, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 650, + 524, + 704 + ], + "spans": [ + { + "bbox": [ + 304, + 650, + 524, + 704 + ], + "type": "text", + "content": "Laura K. Nelson. 2021. Leveraging the alignment between machine learning and intersectionality: Using word embeddings to measure intersectional experiences of the nineteenth century U.S. South. Poetics, 88:101539. Measure Mohr Culture." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 716, + 524, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 716, + 524, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 716, + 524, + 772 + ], + "type": "text", + "content": "Dong Nguyen, Maria Liakata, Simon DeDeo, Jacob Eisenstein, David Mimno, Rebekah Tromble, and Jane Winters. 2020. How we do things with words: Analyzing text as social and cultural data. Frontiers in Artificial Intelligence, 3." + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 308, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 308, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 308, + 791 + ], + "type": "text", + "content": "1591" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 289, + 772 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 69, + 72, + 289, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 289, + 127 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 289, + 127 + ], + "type": "text", + "content": "Brendan O'Connor, David Bamman, and Noah A Smith. 2011. Computational text analysis for social science: Model complexity and assumptions. In Proc. of the NIPS Workshop on Comptuational Social Science and the Wisdom of Crowds." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 137, + 289, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 137, + 289, + 170 + ], + "spans": [ + { + "bbox": [ + 69, + 137, + 289, + 170 + ], + "type": "text", + "content": "Jason Radford and Kenneth Joseph. 2020. Theory in, theory out: The uses of social theory in machine learning for social science. Frontiers in Big Data, 3." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 179, + 289, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 179, + 289, + 212 + ], + "spans": [ + { + "bbox": [ + 69, + 179, + 289, + 212 + ], + "type": "text", + "content": "William J Rapaport. 2005. Philosophy of computer science: An introductory course. Teaching philosophy, 28(4):319-341." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 222, + 289, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 222, + 289, + 256 + ], + "spans": [ + { + "bbox": [ + 69, + 222, + 289, + 256 + ], + "type": "text", + "content": "Stuart C. Shapiro. 2001. Computer science: The study of procedures. Technical report, Department of Computer Science and Engineering, University of Buffalo." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 265, + 289, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 265, + 289, + 353 + ], + "spans": [ + { + "bbox": [ + 69, + 265, + 289, + 353 + ], + "type": "text", + "content": "Han-Chin Shing, Suraj Nair, Ayah Zirikly, Meir Friedenberg, Hal Daumé III, and Philip Resnik. 2018. Expert, crowdsourced, and machine assessment of suicide risk via online postings. In Proceedings of the Fifth Workshop on Computational Linguistics and Clinical Psychology: From Keyboard to Clinic, pages 25-36, New Orleans, LA. Association for Computational Linguistics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 363, + 289, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 363, + 289, + 385 + ], + "spans": [ + { + "bbox": [ + 69, + 363, + 289, + 385 + ], + "type": "text", + "content": "David A. Siegel. 2018. Analyzing computational models. American Journal of Political Science, 62(3):745-759." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 395, + 289, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 395, + 289, + 428 + ], + "spans": [ + { + "bbox": [ + 69, + 395, + 289, + 428 + ], + "type": "text", + "content": "Lyn Spillman. 2015. Ghosts of straw men: A reply to Lee and Martin. American Journal of Cultural Sociology, 3(3):365-379." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 438, + 289, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 438, + 289, + 504 + ], + "spans": [ + { + "bbox": [ + 69, + 438, + 289, + 504 + ], + "type": "text", + "content": "Philip J. Stone and Earl B. Hunt. 1963. A computer approach to content analysis: Studies using the general inquirer system. In Proceedings of the May 21-23, 1963, Spring Joint Computer Conference, AFIPS '63 (Spring), page 241-256, New York, NY, USA. Association for Computing Machinery." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 513, + 289, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 513, + 289, + 546 + ], + "spans": [ + { + "bbox": [ + 69, + 513, + 289, + 546 + ], + "type": "text", + "content": "Sidney Tarrow. 1995. Bridging the quantitative-qualitative divide in political science. American Political Science Review, 89(2):471-474." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 555, + 289, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 555, + 289, + 577 + ], + "spans": [ + { + "bbox": [ + 69, + 555, + 289, + 577 + ], + "type": "text", + "content": "Paul Thagard. 1988. Computational Philosophy of Science. MIT Press." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 587, + 289, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 587, + 289, + 665 + ], + "spans": [ + { + "bbox": [ + 69, + 587, + 289, + 665 + ], + "type": "text", + "content": "Richard Thaler. 1980. Judgement And Decision Making Under Uncertainty: What Economists Can Learn From Psychology. Risk Analysis in Agriculture: Research and Educational Developments, January 16-18, 1980, Tucson, Arizona 271572, Regional Research Projects " + }, + { + "bbox": [ + 69, + 587, + 289, + 665 + ], + "type": "inline_equation", + "content": ">" + }, + { + "bbox": [ + 69, + 587, + 289, + 665 + ], + "type": "text", + "content": " W-149: An Economic Evaluation of Managing Market Risks in Agriculture." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 69, + 674, + 289, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 674, + 289, + 708 + ], + "spans": [ + { + "bbox": [ + 69, + 674, + 289, + 708 + ], + "type": "text", + "content": "Richard H. Thaler and Cass R. Sunstein. 2009. Nudge: Improving decisions about health, wealth, and happiness. Penguin." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 69, + 717, + 289, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 717, + 289, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 717, + 289, + 772 + ], + "type": "text", + "content": "Wouter van Atteveldt and Tai-Quan Peng. 2018. When communication meets computation: Opportunities, challenges, and pitfalls in computational communication science. Communication Methods and Measures, 12(2-3):81-92." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 524, + 316 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 304, + 72, + 524, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 524, + 116 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 524, + 116 + ], + "type": "text", + "content": "Sidney Verba. 1996. The citizen as respondent: Sample surveys and american democracy. American Political Science Review, 90(1):1-7. Presidential Address, American Political Science Association, 1995." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 125, + 524, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 125, + 524, + 169 + ], + "spans": [ + { + "bbox": [ + 304, + 125, + 524, + 169 + ], + "type": "text", + "content": "Sidney Verba, Kay Lehman Schlozman, Henry Brady, and Norman H. Nie. 1993. Citizen activity: Who participates? what do they say? The American Political Science Review, 87(2):303-318." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 178, + 524, + 211 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 178, + 524, + 211 + ], + "spans": [ + { + "bbox": [ + 304, + 178, + 524, + 211 + ], + "type": "text", + "content": "Sidney Verba, Kay Lehman Schlozman, and Henry E Brady. 1995. Voice and equality: Civic volunteerism in American politics. Harvard University Press." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 220, + 524, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 220, + 524, + 275 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 524, + 275 + ], + "type": "text", + "content": "Erin Walk, Elizabeth Parker-Magyar, Kiran Garimella, Ahmet Akbiyik, and Fotini Christia. 2022. Social media narratives across platforms in conflict: Evidence from Syria. MIT Political Science Department Research Paper No. 2022-2, available at SSRN." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 283, + 524, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 283, + 524, + 316 + ], + "spans": [ + { + "bbox": [ + 304, + 283, + 524, + 316 + ], + "type": "text", + "content": "Hanna Wallach. 2018. Computational social science " + }, + { + "bbox": [ + 304, + 283, + 524, + 316 + ], + "type": "inline_equation", + "content": "\\neq" + }, + { + "bbox": [ + 304, + 283, + 524, + 316 + ], + "type": "text", + "content": " computer science + social data. Commun. ACM, 61(3):42-44." + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1592" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "spans": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "type": "text", + "content": "A For every submission:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 106, + 317, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 106, + 317, + 121 + ], + "spans": [ + { + "bbox": [ + 77, + 106, + 317, + 121 + ], + "type": "text", + "content": "A1. Did you describe the limitations of your work?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 121, + 238, + 134 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 121, + 238, + 134 + ], + "spans": [ + { + "bbox": [ + 91, + 121, + 238, + 134 + ], + "type": "text", + "content": "Unnumbered; appears on page 5." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 143, + 329, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 143, + 329, + 157 + ], + "spans": [ + { + "bbox": [ + 77, + 143, + 329, + 157 + ], + "type": "text", + "content": "A2. Did you discuss any potential risks of your work?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 91, + 158, + 196, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 158, + 196, + 170 + ], + "spans": [ + { + "bbox": [ + 91, + 158, + 196, + 170 + ], + "type": "text", + "content": "This is a position paper." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 77, + 179, + 414, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 179, + 414, + 193 + ], + "spans": [ + { + "bbox": [ + 77, + 179, + 414, + 193 + ], + "type": "text", + "content": "A3. Do the abstract and introduction summarize the paper's main claims?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 91, + 194, + 138, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 194, + 138, + 206 + ], + "spans": [ + { + "bbox": [ + 91, + 194, + 138, + 206 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 77, + 215, + 398, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 215, + 398, + 229 + ], + "spans": [ + { + "bbox": [ + 77, + 215, + 398, + 229 + ], + "type": "text", + "content": "A4. Have you used AI writing assistants when working on this paper?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 91, + 230, + 138, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 230, + 138, + 242 + ], + "spans": [ + { + "bbox": [ + 91, + 230, + 138, + 242 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 68, + 253, + 290, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 253, + 290, + 266 + ], + "spans": [ + { + "bbox": [ + 68, + 253, + 290, + 266 + ], + "type": "text", + "content": "B Did you use or create scientific artifacts?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 79, + 270, + 127, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 270, + 127, + 283 + ], + "spans": [ + { + "bbox": [ + 79, + 270, + 127, + 283 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 77, + 292, + 315, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 292, + 315, + 306 + ], + "spans": [ + { + "bbox": [ + 77, + 292, + 315, + 306 + ], + "type": "text", + "content": "B1. Did you cite the creators of artifacts you used?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 91, + 306, + 208, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 306, + 208, + 319 + ], + "spans": [ + { + "bbox": [ + 91, + 306, + 208, + 319 + ], + "type": "text", + "content": "Not applicable. Left blank." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 77, + 328, + 463, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 328, + 463, + 342 + ], + "spans": [ + { + "bbox": [ + 77, + 328, + 463, + 342 + ], + "type": "text", + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 91, + 343, + 208, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 343, + 208, + 355 + ], + "spans": [ + { + "bbox": [ + 91, + 343, + 208, + 355 + ], + "type": "text", + "content": "Not applicable. Left blank." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 77, + 364, + 524, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 364, + 524, + 417 + ], + "spans": [ + { + "bbox": [ + 77, + 364, + 524, + 417 + ], + "type": "text", + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 91, + 418, + 208, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 418, + 208, + 432 + ], + "spans": [ + { + "bbox": [ + 91, + 418, + 208, + 432 + ], + "type": "text", + "content": "Not applicable. Left blank." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 77, + 441, + 524, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 441, + 524, + 481 + ], + "spans": [ + { + "bbox": [ + 77, + 441, + 524, + 481 + ], + "type": "text", + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 91, + 482, + 208, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 482, + 208, + 495 + ], + "spans": [ + { + "bbox": [ + 91, + 482, + 208, + 495 + ], + "type": "text", + "content": "Not applicable. Left blank." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 77, + 504, + 524, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 504, + 524, + 531 + ], + "spans": [ + { + "bbox": [ + 77, + 504, + 524, + 531 + ], + "type": "text", + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 91, + 532, + 208, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 532, + 208, + 544 + ], + "spans": [ + { + "bbox": [ + 91, + 532, + 208, + 544 + ], + "type": "text", + "content": "Not applicable. Left blank." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 77, + 554, + 524, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 554, + 524, + 622 + ], + "spans": [ + { + "bbox": [ + 77, + 554, + 524, + 622 + ], + "type": "text", + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 91, + 623, + 208, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 623, + 208, + 634 + ], + "spans": [ + { + "bbox": [ + 91, + 623, + 208, + 634 + ], + "type": "text", + "content": "Not applicable. Left blank." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 68, + 644, + 293, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 644, + 293, + 657 + ], + "spans": [ + { + "bbox": [ + 68, + 644, + 293, + 657 + ], + "type": "text", + "content": "C Did you run computational experiments?" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 79, + 661, + 127, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 661, + 127, + 674 + ], + "spans": [ + { + "bbox": [ + 79, + 661, + 127, + 674 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 77, + 684, + 524, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 684, + 524, + 711 + ], + "spans": [ + { + "bbox": [ + 77, + 684, + 524, + 711 + ], + "type": "text", + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 91, + 712, + 208, + 724 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 712, + 208, + 724 + ], + "spans": [ + { + "bbox": [ + 91, + 712, + 208, + 724 + ], + "type": "text", + "content": "Not applicable. Left blank." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "spans": [ + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "type": "text", + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "spans": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "text", + "content": "ACL 2023 Responsible NLP Checklist" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1593" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 76, + 71, + 523, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 71, + 523, + 97 + ], + "spans": [ + { + "bbox": [ + 76, + 71, + 523, + 97 + ], + "type": "text", + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 89, + 99, + 208, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 99, + 208, + 111 + ], + "spans": [ + { + "bbox": [ + 89, + 99, + 208, + 111 + ], + "type": "text", + "content": "Not applicable. Left blank." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 76, + 121, + 525, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 121, + 525, + 161 + ], + "spans": [ + { + "bbox": [ + 76, + 121, + 525, + 161 + ], + "type": "text", + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 162, + 208, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 162, + 208, + 174 + ], + "spans": [ + { + "bbox": [ + 89, + 162, + 208, + 174 + ], + "type": "text", + "content": "Not applicable. Left blank." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 76, + 184, + 525, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 184, + 525, + 223 + ], + "spans": [ + { + "bbox": [ + 76, + 184, + 525, + 223 + ], + "type": "text", + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 225, + 208, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 225, + 208, + 238 + ], + "spans": [ + { + "bbox": [ + 89, + 225, + 208, + 238 + ], + "type": "text", + "content": "Not applicable. Left blank." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 247, + 522, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 247, + 522, + 261 + ], + "spans": [ + { + "bbox": [ + 67, + 247, + 522, + 261 + ], + "type": "text", + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 79, + 264, + 127, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 264, + 127, + 277 + ], + "spans": [ + { + "bbox": [ + 79, + 264, + 127, + 277 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 76, + 286, + 525, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 286, + 525, + 313 + ], + "spans": [ + { + "bbox": [ + 76, + 286, + 525, + 313 + ], + "type": "text", + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 89, + 314, + 208, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 314, + 208, + 327 + ], + "spans": [ + { + "bbox": [ + 89, + 314, + 208, + 327 + ], + "type": "text", + "content": "Not applicable. Left blank." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 76, + 336, + 525, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 336, + 525, + 376 + ], + "spans": [ + { + "bbox": [ + 76, + 336, + 525, + 376 + ], + "type": "text", + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 89, + 377, + 208, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 377, + 208, + 391 + ], + "spans": [ + { + "bbox": [ + 89, + 377, + 208, + 391 + ], + "type": "text", + "content": "Not applicable. Left blank." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 76, + 400, + 525, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 400, + 525, + 439 + ], + "spans": [ + { + "bbox": [ + 76, + 400, + 525, + 439 + ], + "type": "text", + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 89, + 440, + 208, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 440, + 208, + 454 + ], + "spans": [ + { + "bbox": [ + 89, + 440, + 208, + 454 + ], + "type": "text", + "content": "Not applicable. Left blank." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 76, + 462, + 521, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 462, + 521, + 476 + ], + "spans": [ + { + "bbox": [ + 76, + 462, + 521, + 476 + ], + "type": "text", + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 89, + 476, + 208, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 476, + 208, + 490 + ], + "spans": [ + { + "bbox": [ + 89, + 476, + 208, + 490 + ], + "type": "text", + "content": "Not applicable. Left blank." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 76, + 498, + 525, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 498, + 525, + 524 + ], + "spans": [ + { + "bbox": [ + 76, + 498, + 525, + 524 + ], + "type": "text", + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 89, + 526, + 208, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 526, + 208, + 539 + ], + "spans": [ + { + "bbox": [ + 89, + 526, + 208, + 539 + ], + "type": "text", + "content": "Not applicable. Left blank." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1594" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Token-Level Self-Evolution Training for Sequence-to-Sequence Learning/8f4944a3-fb6e-4f22-bf90-4b7384f2525a_content_list.json b/2023/Token-Level Self-Evolution Training for Sequence-to-Sequence Learning/8f4944a3-fb6e-4f22-bf90-4b7384f2525a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b8c5724582ae075e75e828b57632f8b2014a7080 --- /dev/null +++ b/2023/Token-Level Self-Evolution Training for Sequence-to-Sequence Learning/8f4944a3-fb6e-4f22-bf90-4b7384f2525a_content_list.json @@ -0,0 +1,1616 @@ +[ + { + "type": "text", + "text": "Token-Level Self-Evolution Training for Sequence-to-Sequence Learning", + "text_level": 1, + "bbox": [ + 122, + 89, + 875, + 112 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keqin Peng $^{1*}$ , Liang Ding $^{2*}$ , Qihuang Zhong $^{3}$ \nYuanxin Ouyang $^{1†}$ , Wenge Rong $^{1}$ , Zhang Xiong $^{1}$ , Dacheng Tao $^{4}$", + "bbox": [ + 228, + 130, + 773, + 164 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Beihang University 2Zhejiang University 3Wuhan University 4The University of Sydney {keqin.peng, oyyx, w.rong, xiongz}@buaa.edu.cn zhongqihuang@whu.edu.cn, {liangding.liam, dacheng.tao}@gmail.com", + "bbox": [ + 122, + 165, + 875, + 216 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 252, + 342, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Adaptive training approaches, widely used in sequence-to-sequence models, commonly reweigh the losses of different target tokens based on priors, e.g. word frequency. However, most of them do not consider the variation of learning difficulty in different training steps, and overly emphasize the learning of difficult one-hot labels, making the learning deterministic and sub-optimal. In response, we present Token-Level Self-Evolution Training (SE), a simple and effective dynamic training method to fully and wisely exploit the knowledge from data. SE focuses on dynamically learning the under-explored tokens for each forward pass and adaptively regularizes the training by introducing a novel token-specific label smoothing approach. Empirically, SE yields consistent and significant improvements in three tasks, i.e. machine translation, summarization, and grammatical error correction. Encouragingly, we achieve averaging $+0.93$ BLEU improvement on three machine translation tasks. Analyses confirm that, besides improving lexical accuracy, SE enhances generation diversity and model generalization.", + "bbox": [ + 141, + 280, + 460, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 114, + 648, + 260, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sequence-to-sequence learning (Seq2Seq) with neural networks (Sutskever et al., 2014) has advanced the state-of-the-art in various NLP tasks, e.g. translation (Bahdanau et al., 2015; Vaswani et al., 2017), summarization (Cheng and Lapata, 2016), and grammatical error correction (Yuan and Briscoe, 2016). Generally, Seq2Seq models are trained with the cross-entropy loss, which equally weighs the training losses of different target tokens.", + "bbox": [ + 112, + 674, + 489, + 818 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, due to the token imbalance nature (Piantadosi, 2014) and the truth that different tokens contribute differently to the sentence meaning (Church and Hanks, 1990; Chen et al., 2020),", + "bbox": [ + 112, + 819, + 489, + 883 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/43bb068707a45de0c88ee7d411df1b2bc2378d00ca71a2a2df7381da1549eb90.jpg", + "image_caption": [ + "Figure 1: An example to illustrate the changing token difficulties in different training steps in WMT'14 En-De. The token \"abschreiben/ Sache\" is hard/ easy to learn at 50K while the trend is totally reversed at 100K." + ], + "image_footnote": [], + "bbox": [ + 512, + 249, + 880, + 367 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "several works are developed to reweigh the token-level training loss according to explicit (e.g. frequency) or implicit (uncertainty estimated by off-the-shelf language models) priors (Gu et al., 2020; Xu et al., 2021; Zhang et al., 2022a). For example, Gu et al. (2020) proposed two heuristic criteria based on word frequency to encourage the model to learn from larger-weight low-frequency tokens. Zhang et al. (2022a) introduce target-context-aware metric based on an additional target-side language model to adjust the weight of each target token.", + "bbox": [ + 507, + 464, + 884, + 640 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite some success, there are still limitations in these adaptive training approaches. First, most of them predetermine the difficult tokens and fix such prior to guiding the training. However, in our preliminary study, we find the hard-to-learn tokens are dynamically changing during training, rather than statically fixed. As shown in Figure 1, as the training progress goes, although the sentence-level loss is nicely converging, the difficult token is changing from \"abschreiben\" to \"Sache\" in terms of the token-level loss. Second, these adaptive training methods overly emphasize fitting the difficult tokens' one-hot labels by reweighing the loss, which empirically may cause overfitting and limit the generalization (Norouzi et al., 2016; Szegedy et al., 2016; Xiao et al., 2019; Miao et al., 2021). Also, a more recent study (Zhai et al., 2023) provides", + "bbox": [ + 507, + 645, + 884, + 919 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Keqin and Liang contributed equally.", + "bbox": [ + 139, + 891, + 376, + 904 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding Author.", + "bbox": [ + 139, + 904, + 294, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "841", + "bbox": [ + 484, + 927, + 515, + 940 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics", + "bbox": [ + 226, + 945, + 771, + 957 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Volume 2: Short Papers, pages 841-850", + "bbox": [ + 376, + 958, + 621, + 971 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "July 9-14, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 295, + 972, + 700, + 985 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "theoretical evidence to support that reweighting is not that effective to improve the generalization.", + "bbox": [ + 112, + 84, + 485, + 116 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Correspondingly, we design a simple and effective Token-Level Self-Evolution Training (SE) strategy to encourage Seq2Seq models to learn from difficult words that are dynamically selected by the model itself. Specifically, SE contains two stages: 1self-questioning and 2self-evolution training. In the first stage, the Seq2Seq models dynamically select the hard-to-learn tokens based on the token-level losses, then we encourage the Seq2Seq models to learn from them in the second stage, where, rather than adopting reweighing, we introduce a novel token-specific label smoothing approach to generate easily digestible soft label, which considers both the ground truth and model's prediction.", + "bbox": [ + 112, + 117, + 487, + 341 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experiments across tasks, language pairs, data scales, and model sizes show that SE consistently and significantly outperforms both the vanilla Seq2Seq model and the re-implemented advanced baselines. Analyses confirm that besides improved lexical accuracy, SE generates diverse and human-like generations with better model generalization.", + "bbox": [ + 112, + 343, + 489, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Methodology", + "text_level": 1, + "bbox": [ + 112, + 468, + 263, + 486 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Preliminary Sequence-to-sequence (Seq2Seq) learning aims to maximize the cross-entropy (CE) loss of the log-likelihood of each target word in $\\mathbf{y} = \\{y_1,\\dots ,y_N\\}$ , conditioned on source $\\mathbf{x}$ , where the optimization treats all tokens equally:", + "bbox": [ + 112, + 495, + 489, + 576 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C E}} (\\theta) = - \\sum_ {j = 1} ^ {N} \\log p \\left(y _ {j} \\mid \\mathbf {y} _ {< j}, \\mathbf {x}; \\theta\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 164, + 587, + 487, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, due to the different learning difficulties of each token, it is sub-optimal to treat all tokens equally (Gu et al., 2020). To address this limitation, a series of token-level adaptive training objectives were adopted to re-weight the losses of different target tokens (Xu et al., 2021; Zhang et al., 2022a). The common goal of these methods is to facilitate the model training by fully exploiting the informative but underexplored tokens.", + "bbox": [ + 112, + 645, + 487, + 788 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, our preliminary study shows that the hard tokens are dynamically changing (see Figure 1) in different training steps (or model structures), thus it is sub-optimal to employ static token priors (e.g. frequency) during training. Also, recent studies (Zhai et al., 2023) in the ML community theoretically show that reweighting is not that effective to improve the generalization. Based on the above", + "bbox": [ + 112, + 790, + 489, + 917 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "evidence, we present the self-evolution learning (SE) mechanism to encourage the model to adaptively and wisely learn from the informative yet under-explored tokens dynamically determined by the model itself (Stage① in §2.1), with an easy-to-learn label distribution (Stage② in §2.1). A similar work to ours is Hahn and Choi (2019). However, their method mainly considers the situation where the predicted answer is incorrect but close to the golden answer, while our method focuses on all dynamic hard tokens.", + "bbox": [ + 507, + 84, + 884, + 261 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 Token-Level Self-Evolution Learning", + "text_level": 1, + "bbox": [ + 507, + 274, + 845, + 288 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Self-questioning Stage. The goal is to select the hard-to-learn tokens that are questioned by the Seq2Seq model itself during training dynamics. Previously, these difficult tokens are predetermined by external models or specific statistical metrics. However, inspired by the finding of dynamic change of difficult tokens during the training stage as shown in Figure 1 and the finding that the trained model contains useful information (Li and Lu, 2021), e.g. synonym, we propose to straightforwardly leverage the behavior of the model to dynamically select target tokens. In practice, we first calculate the token-level CE loss, denoted as $\\{l_1, l_2, \\dots, l_n\\}$ , for each token for each forward pass. Then we set a loss threshold $\\Gamma$ and select the tokens whose losses exceed $\\Gamma$ as the target tokens, i.e., $D = \\{t_i | l_i > \\Gamma\\}$ where $i \\in N = \\{1, 2, \\dots, n\\}$ .", + "bbox": [ + 507, + 296, + 884, + 571 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "$\\Theta$ Self-evolution Training Stage. After selecting the difficult tokens, we encourage the model to carefully learn from them. Given the theoretical shortage (Zhai et al., 2023) and potentially caused overfitting or overconfidence problem (Miao et al., 2021) of reweighting and deliberately learning from difficult tokens, we propose to strengthen the learning from these tokens with a newly designed Token-specific Label Smoothing (TLS) approach. Specifically, motivated by the effect of label smoothing (LS) regularization (Szegedy et al., 2016), we combine the ground truth $p_i$ and the model's prediction $\\hat{p}_i$ to form a new soft label $\\widetilde{p}_i$ for the $i$ -th token. Then we use $\\widetilde{p}$ to guide the difficult tokens $D$ , while leaving label-smoothing CE loss for the other tokens. It is worth noting that we also apply the traditional label smoothing technique to $\\hat{p}_i$ to activate the information in the predicted distribution. Analogous to human learning, it is often easier for humans to grasp new things described by their familiar knowledge (Reder et al., 2016),", + "bbox": [ + 507, + 581, + 882, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "842", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/db7657c51f5d9b7bfc419ca03d3ac3fb73897200d49df934553d43ca66b80f82.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelWMT16 En→RoWMT14 En→DeWMT14 En→Fr
Transformer (Vaswani et al., 2017)35.1127.0840.65
+ Freq-Exponential (Gu et al., 2020)35.86 (+0.75)27.60 (+0.52)41.05 (+0.40)
+ Freq-Chi-Square (Gu et al., 2020)35.74 (+0.63)27.51 (+0.43)40.99 (+0.34)
+ D2GPo (Li et al., 2020)35.89 (+0.78)27.66 (+0.58)41.05 (+0.40)
+ BMI-adaptive (Xu et al., 2021)35.89 (+0.78)27.65 (+0.57)41.10 (+0.45)
+ MixCrossEntropy (Li and Lu, 2021)35.88 (+0.74)27.61 (+0.53)41.07 (+0.42)
+ CBMI-adaptive (Zhang et al., 2022a)35.90 (+0.79)27.69 (+0.61)41.13 (+0.48)
+ SPL (Wan et al., 2020)35.92 (+0.81)27.88 (+0.80)41.30 (+0.65)
+ Self-Evolution (ours)36.02 (+0.91)†28.02 (+0.94)†41.60 (+0.95)†
", + "bbox": [ + 119, + 82, + 880, + 250 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/37bb55b56009a665d91b1c75bed5dd5acdfe40f7f5fc607b9c7939cc24a20c2c.jpg", + "table_caption": [ + "Table 1: BLEU scores $(\\%)$ on three translation tasks spanning different data scales, i.e. $0.6\\mathrm{M}$ , $4.5\\mathrm{M}$ , $36\\mathrm{M}$ . “†” indicates a statistically significant difference from the powerful Transformer baseline $(p < 0.05)$ ." + ], + "table_footnote": [], + "table_body": "
Ro-EnXSUMGEC
BLEURG-1RG-2RG-LPrec.RecallF0.5
Baseline37.343.219.834.059.139.853.9
+ SE37.7†43.820.434.7†58.946.255.8†
", + "bbox": [ + 117, + 313, + 485, + 376 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 2: Performance on more tasks including translation, summarization, and grammar error correction, upon larger model BART (Lewis et al., 2020).", + "bbox": [ + 112, + 386, + 489, + 429 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "therefore the new soft label fused both accurate ground truth and model's self-distribution is easily digestible. Mathematically, for difficult tokens $t_i$ , $\\widetilde{p}_i$ is formulated as:", + "bbox": [ + 112, + 455, + 489, + 519 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {p _ {i}} = \\left(p _ {i} + \\hat {p _ {i}}\\right) / 2. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 231, + 532, + 485, + 550 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Then we calculate the losses of difficult tokens and the others, and combine the two losses:", + "bbox": [ + 112, + 562, + 487, + 594 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL = - \\left(\\sum_ {i} \\widetilde {p _ {i}} \\cdot \\log \\left(\\hat {p _ {i}}\\right) + \\sum_ {j} p _ {j} \\cdot \\log \\left(\\hat {p _ {j}}\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 608, + 487, + 644 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $i\\in D$ and $j\\in N\\setminus D$", + "bbox": [ + 112, + 655, + 332, + 671 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 Evaluation", + "text_level": 1, + "bbox": [ + 112, + 683, + 243, + 697 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Machine Translation on three widely-used benchmarks (Ding et al., 2020, 2021c, 2022): small-scale WMT16 English-Romanian (En-Ro; 0.6M), medium-scale WMT14 English-German (En-De; 4.5M), and large-scale WMT14 English-French (En-Fr; 36.0M). We implement the baselines and our approach under Transformer-base settings. We follow the previous adaptive training approach (Gu et al., 2020) to pretrain with the cross-entropy loss with $N$ steps, and further finetune the same steps with different adaptive training objectives, including Freq-Exponential (Gu et al., 2020), Freq-Chi-Square (Gu et al., 2020), D2GPo (Li et al., 2020),", + "bbox": [ + 112, + 709, + 490, + 919 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "BMI-adaptive (Xu et al., 2021), MixCrossEntropy (Li and Lu, 2021), CBMI-adaptive (Zhang et al., 2022a), and SPL (Wan et al., 2020). For $N$ , we adopt 100K and 30K for larger datasets, e.g. En-De and En-Fr, and small dataset, i.e. En-Ro, respectively. We empirically adopt 32K tokens per batch for large datasets, the learning rate warms up to 1e-7 for 10K steps, and then decays 90K, while for small dataset En-Ro, The learning rate warms up to 1e-7 for 4K steps, and then decays 26K steps. All the experiments are conducted on 4 NVIDIA Tesla A100 GPUs. The SacreBLEU (Post, 2018) was used for evaluation. Besides translation, we also follow previous works (Liu et al., 2021b; Zhong et al., 2022; Zhang et al., 2022b) to validate the universality of our method on more sequence-to-sequence learning tasks, e.g., summarization and grammatical error correction.", + "bbox": [ + 507, + 316, + 884, + 605 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text Summarization on XSUM corpus (0.2M). We follow fairseq (Ott et al., 2019) to preprocess the data and train the model, then finetune them for the same steps. We evaluated with the ROUGE (Lin, 2004), i.e. R-1, R-2, and R-L.", + "bbox": [ + 507, + 607, + 882, + 687 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Grammatical Error Correction on CoNLL14 (1.4M). We follow Chollampatt and Ng (2018) to preprocess the data and train the model, then finetune them for the same steps. The MaxMatch $(\\mathbf{M}^2)$ scores (Dahlmeier and Ng, 2012) were used for evaluation with precision, recall, and $\\mathrm{F_{0.5}}$ values.", + "bbox": [ + 507, + 688, + 882, + 785 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1 Main Results", + "text_level": 1, + "bbox": [ + 507, + 800, + 660, + 814 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SE brings gains across language pairs and scales. Results on machine translation across different data sizes ranging from 0.6M to 36M in Table 1 show that our SE-equipped Transformer “+ Self-Evolution (ours)” 1) considerably improves the performance by averaging +0.92 BLEU points; 2) out", + "bbox": [ + 507, + 822, + 884, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "843", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/006ecc2dac41eeeaf56e1bd722509822d02186c3e16235730351f19ca8ce05d4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Valid Loss Scale
0-11-22-3>3
Transformer + SE63.310.56.719.5
65.69.55.819.1
", + "bbox": [ + 137, + 80, + 467, + 162 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/0e148154138c0851214990282105ac370d251e6072edeb27c5e01a3da4c29c97.jpg", + "table_caption": [ + "Table 3: The token distribution $(\\%)$ on different loss scales. Shadowed areas mean accurate token prediction estimated with lower cross-entropy loss, i.e. \"0-1\"." + ], + "table_footnote": [], + "table_body": "
MethodWMT22 De⇒En
BLEUΔCOMETΔ
Transformer29.98-45.1
+SE30.38+0.446.3+1.2
", + "bbox": [ + 131, + 228, + 478, + 309 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 4: Performance on extremely large dataset WMT22 De-En (236M).", + "bbox": [ + 112, + 317, + 487, + 348 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "performs previous competitive method “+ CBMI-adaptive” by up to +0.47 BLEU points on large dataset WMT14 En-Fr. These results demonstrate the effectiveness and universality of our SE.", + "bbox": [ + 112, + 373, + 489, + 439 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SE brings gains across tasks and backbone sizes. Table 2 lists the performance on more tasks, including translation, summarization, and grammar error correction, upon large pretrained backbone - BART (Lewis et al., 2020), which has above 600M parameters. Compared to a stronger baseline, our SE significantly and incrementally improves the generation quality in all tasks, i.e. $+0.4$ BLEU, $+0.7$ RG-L, and $+1.9$ $\\mathrm{F_{0.5}}$ , respectively, showing our SE is robustly applicable to general scenarios.", + "bbox": [ + 112, + 447, + 489, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SE works well on extremely large dataset. To further verify the effectiveness of SE on extremely large dataset, we conducted an experiment on WMT22 De-En processed by Zan et al. (2022b), which contains 236M training examples. The results in Table 4 show that our method can achieve $+0.4$ and $+1.2$ improvement in BLEU and COMET respectively, which proves that our SE also works on extremely large datasets.", + "bbox": [ + 112, + 618, + 489, + 764 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 Analysis", + "text_level": 1, + "bbox": [ + 112, + 775, + 228, + 790 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We provide some insights to better understand the effectiveness of our approach. The ablation of important modules and parameters is in Appendix A.", + "bbox": [ + 112, + 796, + 489, + 846 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SE learns better token representation. To verify whether our method helps learn better tokens representation, we conduct analysis on WMT14 EnDe from learning loss and fine-grained generation", + "bbox": [ + 112, + 854, + 489, + 920 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4077ecf43a4abda12283cba9c9e154f7c27f83d31e008deae2104732133fd8d5.jpg", + "image_caption": [ + "Figure 2: Fine-grained translation quality across word frequencies and sentence lengths." + ], + "image_footnote": [], + "bbox": [ + 509, + 80, + 882, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "perspectives, respectively.", + "bbox": [ + 507, + 230, + 705, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, we count the token ratios distributed in different cross-entropy loss scales in Table 3 following Zan et al. (2022a). Cross-entropy is a good indicator to quantify the distance between the predicted distribution and the ground truth in the valid dataset, and a lower value means a more similar distribution. As shown, our method improves the low-loss token ratios by $+2.3\\%$ , indicating SE helps the model learn better token representations by reducing the token uncertainty. In addition, we follow Ding et al. (2021a); Liu et al. (2021a) to break the translation down into different granularities and measure their fined-grained performance. In particular, we calculate1 the F-measure of words by different frequency buckets and BLEU scores of buckets of different lengths in Figure 2. We see SE achieves better performance in all frequencies and sentence buckets, demonstrating our method can improve the performance of different granularities.", + "bbox": [ + 507, + 244, + 884, + 567 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SE encourages diverse generations. Lacking generation diversity is a notorious problem for Seq2Seq learning tasks (Sun et al., 2020; Lin et al., 2022). Benefiting from better exploring the model's prediction with corrected soft labels, SE is expected to improve generation diversity. We follow Wang et al. (2022) to examine this by analyzing the performance in an additional multiple-reference test of WMT'14 En-De (Ott et al., 2018). We choose additional references for each of the 500 test sentences taken from the original test. Table 5 shows SE consistently outperforms the baseline with the average improvement being 0.9/1.0 BLEU, which indicates that our SE can effectively generate diverse results.", + "bbox": [ + 507, + 577, + 882, + 819 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SE enhances model generalization. Benefiting from better hard token exploration, SE-equipped Transformers are expected to own better generalizations. We examine it by testing on domain shift", + "bbox": [ + 507, + 829, + 882, + 894 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "Using compare-mt (Neubig et al., 2019).", + "bbox": [ + 529, + 904, + 786, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "844", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/61fc329e7fdf7d3f06850dbdb4486eb540b9d0511a1847aa0f9d5960d68d19ed.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Ref.Avg.Top
Transformer+SETransformer+SE
#142.543.7 (+1.2)44.945.7 (+0.8)
#228.629.3 (+0.7)30.231.2 (+1.0)
#331.232.1 (+0.9)33.234.4 (+1.2)
#428.128.8 (+0.7)29.630.5 (+0.9)
Mean32.633.5 (+0.9)34.535.5 (+1.0)
", + "bbox": [ + 117, + 82, + 485, + 187 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f28cd44a350b3048880357db5a229d08709812f733b17bbf702c89bdf7b08d98.jpg", + "table_caption": [ + "Table 5: Multi-reference performance. 'Avg./ Top' means the averaging/ most-matching performance." + ], + "table_footnote": [], + "table_body": "
ModelLawMed.Kor.Sub.Avg.
Transformer41.230.97.414.523.5
+SE42.6†32.3†7.8†15.0†24.4
", + "bbox": [ + 119, + 240, + 485, + 299 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "scenarios following Ding et al. (2021b). In particular, we evaluate WMT14 En-De models over four out-of-domain test sets (Müller et al., 2020) in Table 6 and find that SE improves the translation by averaging $+0.9$ BLEU points, showing a better lexical generalization ability.", + "bbox": [ + 112, + 393, + 487, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SE encourages human-like generations. We design two types of evaluation on WMT14 En-Fr: 1) AUTOMATIC EVALUATION with COMET (Rei et al., 2020) and BLEURT (Sellam et al., 2020), which have a high-level correlation with human judgments. 2) HUMAN EVALUATION with three near-native French annotators who hold DALF C2 certificate2. Specifically, for human evaluation, we randomly sample 50 sentences from the test set to evaluate the translation adequacy and fluency, scoring $1 \\sim 5$ . For adequacy, 1 represents irrelevant to the source while 5 means semantically equal. For fluency, 1 means unintelligible while 5 means fluent and native. Table 7 shows the automatic and human evaluation results, where we find that our SE indeed achieves human-like translation.", + "bbox": [ + 112, + 500, + 487, + 757 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Conclusion", + "text_level": 1, + "bbox": [ + 112, + 771, + 245, + 785 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this paper, we propose a self-evolution learning mechanism to improve seq2seq learning, by exploiting the informative-yet-underexplored tokens dynamically. SE follows two stages, i.e. self-questioning and self-evolution training, and can be used to evolve any pretrained models with a sim", + "bbox": [ + 112, + 797, + 487, + 894 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/8f4ba288b0456c34d1b25ff851b1782cac8a3512a72bb365a00554ee58d9e557.jpg", + "table_caption": [ + "Table 6: Performance on domain shift setting. Models are trained on the news but evaluated on out-of-domain test sets, including law, medicine, koran, and subtitle. “†” indicates statistically significance $(p < 0.05)$ ." + ], + "table_footnote": [], + "table_body": "
AUTOMATIC EVAL.HUMAN EVAL.
COMETBLEURTAdequacyFluency
Transformer + SE61.668.64.324.58
63.769.54.504.68
", + "bbox": [ + 514, + 82, + 878, + 149 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 7: Human evaluation on WMT14 En-Fr.", + "bbox": [ + 531, + 159, + 857, + 173 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ple recipe: continue train with SE. We empirically demonstrated the effectiveness and universality of SE on a series of widely-used benchmarks, covering low, medium, high, and extremely-high data volumes.", + "bbox": [ + 507, + 200, + 882, + 278 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the future, besides generation tasks, we would like to verify the effectiveness of SE on language understanding tasks (Wu et al., 2020; Zhong et al., 2023). Also, it will be interesting to design SE-inspired instruction tuning or prompting strategy like Lu et al. (2023) to enhance the performance of large language models, e.g. ChatGPT3, which after all have already been fully validated on lots of conditional generation tasks (Hendy et al., 2023; Jiao et al., 2023; Peng et al., 2023; Wu et al., 2023).", + "bbox": [ + 507, + 281, + 884, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Limitations", + "text_level": 1, + "bbox": [ + 507, + 454, + 613, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our work has several potential limitations. First, we determine the threshold $\\Gamma$ by manual selection, which may limit the performance of Seq2Seq models, it will make our work more effective and elegant if we dynamically select the threshold. Second, besides the improvement on three widely used tasks, we believe that there are still other abilities, like code generation, of Seq2Seq models that can be improved by our method, which are not fully explored in this work.", + "bbox": [ + 507, + 480, + 882, + 640 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Ethics Statement", + "text_level": 1, + "bbox": [ + 507, + 653, + 660, + 668 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We take ethical considerations very seriously and strictly adhere to the ACL Ethics Policy. This paper focuses on effective training for sequence-to-sequence learning. The datasets used in this paper are publicly available and have been widely adopted by researchers. We ensure that the findings and conclusions of this paper are reported accurately and objectively.", + "bbox": [ + 507, + 678, + 882, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 507, + 820, + 672, + 837 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We are grateful to the anonymous reviewers and the area chair for their insightful comments and suggestions.", + "bbox": [ + 507, + 847, + 880, + 895 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "2http://www.delfdalf.fr/dalf-c2-en.html", + "bbox": [ + 134, + 903, + 376, + 917 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "3https://chat.openai.com/", + "bbox": [ + 529, + 903, + 757, + 917 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "845", + "bbox": [ + 485, + 927, + 515, + 940 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 115, + 84, + 213, + 98 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2015. Neural machine translation by jointly learning to align and translate. In ICLR.", + "Kehai Chen, Rui Wang, Masao Utiyama, and Eiichiro Sumita. 2020. Content word aware neural machine translation. In ACL.", + "Jianpeng Cheng and Mirella Lapata. 2016. Neural summarization by extracting sentences and words. In ACL.", + "Shamil Chollampatt and Hwee Tou Ng. 2018. A multilayer convolutional encoder-decoder neural network for grammatical error correction. In AAAI.", + "Kenneth Church and Patrick Hanks. 1990. Word association norms, mutual information, and lexicography. CL.", + "Daniel Dahlmeier and Hwee Tou Ng. 2012. Better evaluation for grammatical error correction. In NAACL.", + "Liang Ding, Longyue Wang, Xuebo Liu, Derek F Wong, Dacheng Tao, and Zhaopeng Tu. 2021a. Progressive multi-granularity training for non-autoregressive translation. In Findings of ACL.", + "Liang Ding, Longyue Wang, Xuebo Liu, Derek F Wong, Dacheng Tao, and Zhaopeng Tu. 2021b. Rejuvenating low-frequency words: Making the most of parallel data in non-autoregressive translation. In ACL.", + "Liang Ding, Longyue Wang, Xuebo Liu, Derek F Wong, Dacheng Tao, and Zhaopeng Tu. 2021c. Understanding and improving lexical choice in non-autoregressive translation. In ICLR.", + "Liang Ding, Longyue Wang, Shuming Shi, Dacheng Tao, and Zhaopeng Tu. 2022. Redistributing low-frequency words: Making the most of monolingual data in non-autoregressive translation. In ACL.", + "Liang Ding, Longyue Wang, and Dacheng Tao. 2020. Self-attention with cross-lingual position representation. In ACL.", + "Shuhao Gu, Jinchao Zhang, Fandong Meng, Yang Feng, Wanying Xie, Jie Zhou, and Dong Yu. 2020. Token-level adaptive training for neural machine translation. In EMNLP.", + "Sangchul Hahn and Heeyoul Choi. 2019. Self-knowledge distillation in natural language processing. In RANLP.", + "Amr Hendy, Mohamed Abdelrehim, et al. 2023. How good are gpt models at machine translation? a comprehensive evaluation. arXiv preprint.", + "Wenxiang Jiao, Wenxuan Wang, Jen-tse Huang, Xing Wang, and Zhaopeng Tu. 2023. Is chatgpt a good translator? a preliminary study. arXiv preprint." + ], + "bbox": [ + 115, + 107, + 485, + 917 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In ACL.", + "Haoran Li and Wei Lu. 2021. Mixed cross entropy loss for neural machine translation. In ICML.", + "Zuchao Li, Rui Wang, et al. 2020. Data-dependent gaussian prior objective for language generation. In ICLR.", + "Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out.", + "Huan Lin, Baosong Yang, Liang Yao, Dayiheng Liu, Haibo Zhang, Jun Xie, Min Zhang, and Jinsong Su. 2022. Bridging the gap between training and inference: Multi-candidate optimization for diverse neural machine translation. In Findings of NAACL.", + "Xuebo Liu, Longyue Wang, Derek F Wong, Liang Ding, Lidia S Chao, Shuming Shi, and Zhaopeng Tu. 2021a. On the copying behaviors of pre-training for neural machine translation. In Findings of ACL.", + "Xuebo Liu, Longyue Wang, Derek F Wong, Liang Ding, Lidia S Chao, and Zhaopeng Tu. 2021b. Understanding and improving encoder layer fusion in sequence-to-sequence learning. In ICLR.", + "Qingyu Lu, Baopu Qiu, Liang Ding, Liping Xie, and Dacheng Tao. 2023. Error analysis prompting enables human-like translation evaluation in large language models: A case study on chatgpt. arXiv preprint.", + "Mengqi Miao, Fandong Meng, Yijin Liu, Xiao-Hua Zhou, and Jie Zhou. 2021. Prevent the language model from being overconfident in neural machine translation. In ACL.", + "Mathias Müller, Annette Rios, and Rico Sennrich. 2020. Domain robustness in neural machine translation. In AMTA, Virtual.", + "Graham Neubig, Zi-Yi Dou, Junjie Hu, Paul Michel, Danish Pruthi, and Xinyi Wang. 2019. compare-mt: A tool for holistic comparison of language generation systems. In NAACL.", + "Mohammad Norouzi, Samy Bengio, Zhifeng Chen, et al. 2016. Reward augmented maximum likelihood for neural structured prediction. In NeurIPS.", + "Myle Ott, Michael Auli, David Grangier, and Marc'Aurelio Ranzato. 2018. Analyzing uncertainty in neural machine translation. In ICML.", + "Myle Ott, Sergey Edunov, Alexei Baevski, Angela Fan, Sam Gross, Nathan Ng, David Grangier, and Michael Auli. 2019. *fairoseq: A fast, extensible toolkit for sequence modeling*. In *NAACL Demonstration*." + ], + "bbox": [ + 510, + 85, + 880, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "846", + "bbox": [ + 485, + 928, + 515, + 939 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Keqin Peng, Liang Ding, Qihuang Zhong, Li Shen, Xuebo Liu, Min Zhang, Yuanxin Ouyang, and Dacheng Tao. 2023. Towards making the most of chatgpt for machine translation. arxiv preprint.", + "Steven T Piantadosi. 2014. Zipf's word frequency law in natural language: A critical review and future directions. Psychonomic bulletin & review.", + "Matt Post. 2018. A call for clarity in reporting BLEU scores. In WMT.", + "Lynne M Reder, Xiaonan L Liu, Alexander Keinath, and Vencislav Popov. 2016. Building knowledge requires bricks, not sand: The critical role of familiar constituents in learning. Psychonomic bulletin & review.", + "Ricardo Rei, Craig Stewart, Ana C. Farinha, and Alon Lavie. 2020. COMET: A neural framework for MT evaluation. In EMNLP.", + "Thibault Sellam, Dipanjan Das, and Ankur P. Parikh. 2020. BLEURT: learning robust metrics for text generation. In ACL.", + "Zewei Sun, Shujian Huang, Hao-Ran Wei, Xinyu Dai, and Jiajun Chen. 2020. Generating diverse translation by manipulating multi-head attention. In AAAI.", + "Ilya Sutskever, Oriol Vinyals, and Quoc V. Le. 2014. Sequence to sequence learning with neural networks. In NeurIPS.", + "Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. 2016. Rethinking the inception architecture for computer vision. In CVPR.", + "Ashish Vaswani, Noam Shazeer, et al. 2017. Attention is all you need. In NeurIPS.", + "Yu Wan, Baosong Yang, et al. 2020. Self-paced learning for neural machine translation. In EMNLP.", + "Wenxuan Wang, Wenxiang Jiao, Yongchang Hao, Xing Wang, Shuming Shi, Zhaopeng Tu, and Michael R. Lyu. 2022. Understanding and improving sequence-to-sequence pretraining for neural machine translation. In ACL.", + "Di Wu, Liang Ding, Fan Lu, and Jian Xie. 2020. Slotrefine: A fast non-autoregressive model for joint intent detection and slot filling. In EMNLP.", + "Haoran Wu, Wenxuan Wang, Yuxuan Wan, Wenxiang Jiao, and Michael Lyu. 2023. Chatgpt or grammarly? evaluating chatgpt on grammatical error correction benchmark. arXiv preprint.", + "Fengshun Xiao, Yingting Wu, Hai Zhao, Rui Wang, and Shu Jiang. 2019. Dual skew divergence loss for neural machine translation. CoRR.", + "Yangyifan Xu, Yijin Liu, Fandong Meng, Jiajun Zhang, Jinan Xu, and Jie Zhou. 2021. Bilingual mutual information based adaptive training for neural machine translation. In ACL." + ], + "bbox": [ + 115, + 85, + 485, + 916 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Zheng Yuan and Ted Briscoe. 2016. Grammatical error correction using neural machine translation. In NAACL.", + "Changtong Zan, Liang Ding, Li Shen, Yu Cao, Weifeng Liu, and Dacheng Tao. 2022a. On the complementarity between pre-training and random-initialization for resource-rich machine translation. In COLING.", + "Changtong Zan, Keqin Peng, Liang Ding, et al. 2022b. Vega-mt: The jd explore academy machine translation system for wmt22. In WMT.", + "Runtian Zhai, Chen Dan, J Zico Kolter, and Pradeep Kumar Ravikumar. 2023. Understanding why generalized reweighting does not improve over ERM. In ICLR.", + "Songming Zhang, Yijin Liu, Fandong Meng, Yufeng Chen, Jinan Xu, Jian Liu, and Jie Zhou. 2022a. Conditional bilingual mutual information based adaptive training for neural machine translation. In ACL.", + "Zheng Zhang, Liang Ding, Dazhao Cheng, Xuebo Liu, Min Zhang, and Dacheng Tao. 2022b. Bliss: Robust sequence-to-sequence learning via self-supervised input representation. arXiv preprint.", + "Qihuang Zhong, Liang Ding, Juhua Liu, Bo Du, and Dacheng Tao. 2022. E2s2: Encoding-enhanced sequence-to-sequence pretraining for language understanding and generation. arXiv preprint.", + "Qihuang Zhong, Liang Ding, Keqin Peng, Juhua Liu, Bo Du, Li Shen, Yibing Zhan, and Dacheng Tao. 2023. Bag of tricks for effective language model pretraining and downstream adaptation: A case study on glue. arXiv preprint." + ], + "bbox": [ + 510, + 85, + 880, + 563 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A Appendix", + "text_level": 1, + "bbox": [ + 510, + 577, + 631, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Parameter Analysis on $\\Gamma$ As stated in §2.1, we use the loss threshold $\\Gamma$ to dynamically select the hard-to-learn tokens. Here, we analyze the influence of different $\\Gamma$ in detail. In practice, we train the Transformer models with different $\\Gamma$ (in $\\{3,4,5,6\\}$ ) and evaluate the performance of the WMT14 En-De test set. Table 8 lists the performance of different $\\Gamma$ . The results of Table 8 show that SE is stable and insensitive to $\\Gamma$ within a certain range. Noting that we select $\\Gamma = 5$ for all experiment settings based on the results in Table 8.", + "bbox": [ + 509, + 602, + 882, + 778 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e282bc0bde55ed294ee627795f045b303dd5c6a93bb9f64a91d4611079e702d8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Γ=3Γ=4Γ=5Γ=6
BLEU27.727.828.027.8
", + "bbox": [ + 571, + 804, + 823, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 8: Parameter analysis of $\\Gamma$ on WMT14 En-De.", + "bbox": [ + 517, + 858, + 870, + 871 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "847", + "bbox": [ + 485, + 928, + 515, + 939 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation Study", + "text_level": 1, + "bbox": [ + 114, + 84, + 238, + 99 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Metric. In this work, we use the loss-based metric to dynamically select the hard-to-learn tokens. To validate the effectiveness of the metric, we use a simple adaptive training method (\"+ ADD\") that adds 1 to the weighting term of loss of the hard-to-learn tokens. The results on WMT16 EnRo are shown in Table 9, the simple Add method can achieve +0.3 BLEU improvement compared to the baseline model, which proves that our proposed self-questioning stage indeed mines informative difficult tokens. Also, we can observe that learning these dynamic difficult tokens with our SE framework (\"+ SE\") could outperform \"+\" ADD\" by +0.6 BLUE points, demonstrating the superiority of our token-specific label smoothing approach.", + "bbox": [ + 112, + 109, + 489, + 350 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/929cf37b53e86fa65783e09240e01496ba4e8c3f20c04c4725162757d3cf2107.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Baseline+ ADD+ SE
BLEU35.135.436.0
", + "bbox": [ + 176, + 360, + 431, + 405 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Learning objective. As stated in §2.1, our learning objective is the combination of the ground truth and the model's prediction. To validate the effectiveness of predicted distribution, we conduct ablation experiments on WMT16 En-Ro and WMT14 En-De. The results in Table 10 show that adding the predicted distribution will consistently improve the model's performance, which proves the effectiveness of the predicted distribution.", + "bbox": [ + 112, + 454, + 489, + 599 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a5062faacb13a04345d161149ed2f48791efd45af26686a5d85a458821cf830c.jpg", + "table_caption": [ + "Table 9: Ablation performance of our SE. on Metric." + ], + "table_footnote": [], + "table_body": "
MethodBLEU
EN⇒DEEN⇒Ro
Transformer27.0835.11
SE28.0236.02
-w/o predicted results27.8935.71
", + "bbox": [ + 154, + 609, + 450, + 697 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 10: Ablation performance of our SE. on learning objective.", + "bbox": [ + 112, + 706, + 485, + 733 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "848", + "bbox": [ + 485, + 928, + 515, + 939 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A For every submission:", + "bbox": [ + 114, + 107, + 322, + 122 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "A1. Did you describe the limitations of your work?", + "bbox": [ + 129, + 126, + 532, + 143 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The last section of the paper.", + "bbox": [ + 149, + 143, + 364, + 159 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "A2. Did you discuss any potential risks of your work?", + "bbox": [ + 129, + 170, + 552, + 186 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 149, + 187, + 231, + 200 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "A3. Do the abstract and introduction summarize the paper's main claims?", + "bbox": [ + 129, + 212, + 695, + 228 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The abstract and the introduction section.", + "bbox": [ + 149, + 230, + 458, + 244 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "A4. Have you used AI writing assistants when working on this paper?", + "bbox": [ + 129, + 255, + 668, + 272 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 149, + 273, + 231, + 287 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "B Did you use or create scientific artifacts?", + "bbox": [ + 114, + 300, + 489, + 316 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 132, + 321, + 213, + 336 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "B1. Did you cite the creators of artifacts you used?", + "bbox": [ + 127, + 347, + 529, + 363 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 363, + 248, + 379 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?", + "bbox": [ + 127, + 390, + 778, + 406 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 407, + 248, + 422 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?", + "bbox": [ + 127, + 432, + 880, + 495 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 498, + 248, + 513 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?", + "bbox": [ + 127, + 524, + 880, + 571 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 574, + 248, + 588 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?", + "bbox": [ + 127, + 599, + 880, + 631 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 633, + 248, + 646 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be.", + "bbox": [ + 127, + 658, + 880, + 739 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 740, + 248, + 753 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "C Did you run computational experiments?", + "bbox": [ + 114, + 765, + 494, + 781 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 132, + 787, + 213, + 801 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?", + "bbox": [ + 127, + 813, + 880, + 845 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 846, + 248, + 860 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.", + "bbox": [ + 112, + 868, + 877, + 892 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "ACL 2023 Responsible NLP Checklist", + "bbox": [ + 132, + 84, + 433, + 99 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "849", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? No response.", + "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? No response.", + "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? No response." + ], + "bbox": [ + 127, + 84, + 880, + 282 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?", + "text_level": 1, + "bbox": [ + 114, + 292, + 877, + 310 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Section 3.2", + "bbox": [ + 132, + 313, + 220, + 328 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Left blank.", + "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Left blank.", + "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Left blank.", + "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Left blank.", + "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Left blank." + ], + "bbox": [ + 129, + 340, + 880, + 640 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "850", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Token-Level Self-Evolution Training for Sequence-to-Sequence Learning/8f4944a3-fb6e-4f22-bf90-4b7384f2525a_model.json b/2023/Token-Level Self-Evolution Training for Sequence-to-Sequence Learning/8f4944a3-fb6e-4f22-bf90-4b7384f2525a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9290664db690c35426837125d173257f6baa0b24 --- /dev/null +++ b/2023/Token-Level Self-Evolution Training for Sequence-to-Sequence Learning/8f4944a3-fb6e-4f22-bf90-4b7384f2525a_model.json @@ -0,0 +1,2255 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.123, + 0.09, + 0.877, + 0.113 + ], + "angle": 0, + "content": "Token-Level Self-Evolution Training for Sequence-to-Sequence Learning" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.131, + 0.774, + 0.165 + ], + "angle": 0, + "content": "Keqin Peng\\(^{1*}\\), Liang Ding\\(^{2*}\\), Qihuang Zhong\\(^{3}\\) \nYuanxin Ouyang\\(^{1†}\\), Wenge Rong\\(^{1}\\), Zhang Xiong\\(^{1}\\), Dacheng Tao\\(^{4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.123, + 0.166, + 0.877, + 0.217 + ], + "angle": 0, + "content": "1Beihang University 2Zhejiang University 3Wuhan University 4The University of Sydney {keqin.peng, oyyx, w.rong, xiongz}@buaa.edu.cn zhongqihuang@whu.edu.cn, {liangding.liam, dacheng.tao}@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.253, + 0.343, + 0.269 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.281, + 0.462, + 0.637 + ], + "angle": 0, + "content": "Adaptive training approaches, widely used in sequence-to-sequence models, commonly reweigh the losses of different target tokens based on priors, e.g. word frequency. However, most of them do not consider the variation of learning difficulty in different training steps, and overly emphasize the learning of difficult one-hot labels, making the learning deterministic and sub-optimal. In response, we present Token-Level Self-Evolution Training (SE), a simple and effective dynamic training method to fully and wisely exploit the knowledge from data. SE focuses on dynamically learning the under-explored tokens for each forward pass and adaptively regularizes the training by introducing a novel token-specific label smoothing approach. Empirically, SE yields consistent and significant improvements in three tasks, i.e. machine translation, summarization, and grammatical error correction. Encouragingly, we achieve averaging \\(+0.93\\) BLEU improvement on three machine translation tasks. Analyses confirm that, besides improving lexical accuracy, SE enhances generation diversity and model generalization." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.649, + 0.262, + 0.665 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.675, + 0.49, + 0.819 + ], + "angle": 0, + "content": "Sequence-to-sequence learning (Seq2Seq) with neural networks (Sutskever et al., 2014) has advanced the state-of-the-art in various NLP tasks, e.g. translation (Bahdanau et al., 2015; Vaswani et al., 2017), summarization (Cheng and Lapata, 2016), and grammatical error correction (Yuan and Briscoe, 2016). Generally, Seq2Seq models are trained with the cross-entropy loss, which equally weighs the training losses of different target tokens." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.82, + 0.49, + 0.884 + ], + "angle": 0, + "content": "However, due to the token imbalance nature (Piantadosi, 2014) and the truth that different tokens contribute differently to the sentence meaning (Church and Hanks, 1990; Chen et al., 2020)," + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.25, + 0.882, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.373, + 0.885, + 0.43 + ], + "angle": 0, + "content": "Figure 1: An example to illustrate the changing token difficulties in different training steps in WMT'14 En-De. The token \"abschreiben/ Sache\" is hard/ easy to learn at 50K while the trend is totally reversed at 100K." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.465, + 0.885, + 0.642 + ], + "angle": 0, + "content": "several works are developed to reweigh the token-level training loss according to explicit (e.g. frequency) or implicit (uncertainty estimated by off-the-shelf language models) priors (Gu et al., 2020; Xu et al., 2021; Zhang et al., 2022a). For example, Gu et al. (2020) proposed two heuristic criteria based on word frequency to encourage the model to learn from larger-weight low-frequency tokens. Zhang et al. (2022a) introduce target-context-aware metric based on an additional target-side language model to adjust the weight of each target token." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.646, + 0.885, + 0.92 + ], + "angle": 0, + "content": "Despite some success, there are still limitations in these adaptive training approaches. First, most of them predetermine the difficult tokens and fix such prior to guiding the training. However, in our preliminary study, we find the hard-to-learn tokens are dynamically changing during training, rather than statically fixed. As shown in Figure 1, as the training progress goes, although the sentence-level loss is nicely converging, the difficult token is changing from \"abschreiben\" to \"Sache\" in terms of the token-level loss. Second, these adaptive training methods overly emphasize fitting the difficult tokens' one-hot labels by reweighing the loss, which empirically may cause overfitting and limit the generalization (Norouzi et al., 2016; Szegedy et al., 2016; Xiao et al., 2019; Miao et al., 2021). Also, a more recent study (Zhai et al., 2023) provides" + }, + { + "type": "page_footnote", + "bbox": [ + 0.141, + 0.892, + 0.378, + 0.906 + ], + "angle": 0, + "content": "* Keqin and Liang contributed equally." + }, + { + "type": "page_footnote", + "bbox": [ + 0.141, + 0.906, + 0.295, + 0.919 + ], + "angle": 0, + "content": "† Corresponding Author." + }, + { + "type": "list", + "bbox": [ + 0.141, + 0.892, + 0.378, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.928, + 0.516, + 0.941 + ], + "angle": 0, + "content": "841" + }, + { + "type": "footer", + "bbox": [ + 0.227, + 0.946, + 0.772, + 0.958 + ], + "angle": 0, + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + }, + { + "type": "footer", + "bbox": [ + 0.377, + 0.959, + 0.623, + 0.972 + ], + "angle": 0, + "content": "Volume 2: Short Papers, pages 841-850" + }, + { + "type": "footer", + "bbox": [ + 0.297, + 0.973, + 0.702, + 0.986 + ], + "angle": 0, + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.486, + 0.117 + ], + "angle": 0, + "content": "theoretical evidence to support that reweighting is not that effective to improve the generalization." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.118, + 0.489, + 0.342 + ], + "angle": 0, + "content": "Correspondingly, we design a simple and effective Token-Level Self-Evolution Training (SE) strategy to encourage Seq2Seq models to learn from difficult words that are dynamically selected by the model itself. Specifically, SE contains two stages: 1self-questioning and 2self-evolution training. In the first stage, the Seq2Seq models dynamically select the hard-to-learn tokens based on the token-level losses, then we encourage the Seq2Seq models to learn from them in the second stage, where, rather than adopting reweighing, we introduce a novel token-specific label smoothing approach to generate easily digestible soft label, which considers both the ground truth and model's prediction." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.344, + 0.49, + 0.456 + ], + "angle": 0, + "content": "Experiments across tasks, language pairs, data scales, and model sizes show that SE consistently and significantly outperforms both the vanilla Seq2Seq model and the re-implemented advanced baselines. Analyses confirm that besides improved lexical accuracy, SE generates diverse and human-like generations with better model generalization." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.469, + 0.265, + 0.487 + ], + "angle": 0, + "content": "2 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.496, + 0.49, + 0.577 + ], + "angle": 0, + "content": "Preliminary Sequence-to-sequence (Seq2Seq) learning aims to maximize the cross-entropy (CE) loss of the log-likelihood of each target word in \\(\\mathbf{y} = \\{y_1,\\dots ,y_N\\}\\), conditioned on source \\(\\mathbf{x}\\), where the optimization treats all tokens equally:" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.588, + 0.489, + 0.634 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C E}} (\\theta) = - \\sum_ {j = 1} ^ {N} \\log p \\left(y _ {j} \\mid \\mathbf {y} _ {< j}, \\mathbf {x}; \\theta\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.646, + 0.489, + 0.789 + ], + "angle": 0, + "content": "However, due to the different learning difficulties of each token, it is sub-optimal to treat all tokens equally (Gu et al., 2020). To address this limitation, a series of token-level adaptive training objectives were adopted to re-weight the losses of different target tokens (Xu et al., 2021; Zhang et al., 2022a). The common goal of these methods is to facilitate the model training by fully exploiting the informative but underexplored tokens." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.791, + 0.49, + 0.919 + ], + "angle": 0, + "content": "However, our preliminary study shows that the hard tokens are dynamically changing (see Figure 1) in different training steps (or model structures), thus it is sub-optimal to employ static token priors (e.g. frequency) during training. Also, recent studies (Zhai et al., 2023) in the ML community theoretically show that reweighting is not that effective to improve the generalization. Based on the above" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.262 + ], + "angle": 0, + "content": "evidence, we present the self-evolution learning (SE) mechanism to encourage the model to adaptively and wisely learn from the informative yet under-explored tokens dynamically determined by the model itself (Stage① in §2.1), with an easy-to-learn label distribution (Stage② in §2.1). A similar work to ours is Hahn and Choi (2019). However, their method mainly considers the situation where the predicted answer is incorrect but close to the golden answer, while our method focuses on all dynamic hard tokens." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.275, + 0.847, + 0.29 + ], + "angle": 0, + "content": "2.1 Token-Level Self-Evolution Learning" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.297, + 0.885, + 0.572 + ], + "angle": 0, + "content": "1 Self-questioning Stage. The goal is to select the hard-to-learn tokens that are questioned by the Seq2Seq model itself during training dynamics. Previously, these difficult tokens are predetermined by external models or specific statistical metrics. However, inspired by the finding of dynamic change of difficult tokens during the training stage as shown in Figure 1 and the finding that the trained model contains useful information (Li and Lu, 2021), e.g. synonym, we propose to straightforwardly leverage the behavior of the model to dynamically select target tokens. In practice, we first calculate the token-level CE loss, denoted as \\(\\{l_1, l_2, \\dots, l_n\\}\\), for each token for each forward pass. Then we set a loss threshold \\(\\Gamma\\) and select the tokens whose losses exceed \\(\\Gamma\\) as the target tokens, i.e., \\(D = \\{t_i | l_i > \\Gamma\\}\\) where \\(i \\in N = \\{1, 2, \\dots, n\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.582, + 0.884, + 0.919 + ], + "angle": 0, + "content": "\\(\\Theta\\) Self-evolution Training Stage. After selecting the difficult tokens, we encourage the model to carefully learn from them. Given the theoretical shortage (Zhai et al., 2023) and potentially caused overfitting or overconfidence problem (Miao et al., 2021) of reweighting and deliberately learning from difficult tokens, we propose to strengthen the learning from these tokens with a newly designed Token-specific Label Smoothing (TLS) approach. Specifically, motivated by the effect of label smoothing (LS) regularization (Szegedy et al., 2016), we combine the ground truth \\(p_i\\) and the model's prediction \\(\\hat{p}_i\\) to form a new soft label \\(\\widetilde{p}_i\\) for the \\(i\\)-th token. Then we use \\(\\widetilde{p}\\) to guide the difficult tokens \\(D\\), while leaving label-smoothing CE loss for the other tokens. It is worth noting that we also apply the traditional label smoothing technique to \\(\\hat{p}_i\\) to activate the information in the predicted distribution. Analogous to human learning, it is often easier for humans to grasp new things described by their familiar knowledge (Reder et al., 2016)," + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.517, + 0.941 + ], + "angle": 0, + "content": "842" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.12, + 0.083, + 0.881, + 0.251 + ], + "angle": 0, + "content": "
ModelWMT16 En→RoWMT14 En→DeWMT14 En→Fr
Transformer (Vaswani et al., 2017)35.1127.0840.65
+ Freq-Exponential (Gu et al., 2020)35.86 (+0.75)27.60 (+0.52)41.05 (+0.40)
+ Freq-Chi-Square (Gu et al., 2020)35.74 (+0.63)27.51 (+0.43)40.99 (+0.34)
+ D2GPo (Li et al., 2020)35.89 (+0.78)27.66 (+0.58)41.05 (+0.40)
+ BMI-adaptive (Xu et al., 2021)35.89 (+0.78)27.65 (+0.57)41.10 (+0.45)
+ MixCrossEntropy (Li and Lu, 2021)35.88 (+0.74)27.61 (+0.53)41.07 (+0.42)
+ CBMI-adaptive (Zhang et al., 2022a)35.90 (+0.79)27.69 (+0.61)41.13 (+0.48)
+ SPL (Wan et al., 2020)35.92 (+0.81)27.88 (+0.80)41.30 (+0.65)
+ Self-Evolution (ours)36.02 (+0.91)†28.02 (+0.94)†41.60 (+0.95)†
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.263, + 0.885, + 0.294 + ], + "angle": 0, + "content": "Table 1: BLEU scores \\((\\%)\\) on three translation tasks spanning different data scales, i.e. \\(0.6\\mathrm{M}\\), \\(4.5\\mathrm{M}\\), \\(36\\mathrm{M}\\). “†” indicates a statistically significant difference from the powerful Transformer baseline \\((p < 0.05)\\)." + }, + { + "type": "table", + "bbox": [ + 0.119, + 0.315, + 0.487, + 0.378 + ], + "angle": 0, + "content": "
Ro-EnXSUMGEC
BLEURG-1RG-2RG-LPrec.RecallF0.5
Baseline37.343.219.834.059.139.853.9
+ SE37.7†43.820.434.7†58.946.255.8†
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.387, + 0.49, + 0.43 + ], + "angle": 0, + "content": "Table 2: Performance on more tasks including translation, summarization, and grammar error correction, upon larger model BART (Lewis et al., 2020)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.456, + 0.49, + 0.52 + ], + "angle": 0, + "content": "therefore the new soft label fused both accurate ground truth and model's self-distribution is easily digestible. Mathematically, for difficult tokens \\( t_i \\), \\( \\widetilde{p}_i \\) is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.232, + 0.533, + 0.487, + 0.551 + ], + "angle": 0, + "content": "\\[\n\\widetilde {p _ {i}} = \\left(p _ {i} + \\hat {p _ {i}}\\right) / 2. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.563, + 0.488, + 0.595 + ], + "angle": 0, + "content": "Then we calculate the losses of difficult tokens and the others, and combine the two losses:" + }, + { + "type": "equation", + "bbox": [ + 0.129, + 0.609, + 0.489, + 0.645 + ], + "angle": 0, + "content": "\\[\nL = - \\left(\\sum_ {i} \\widetilde {p _ {i}} \\cdot \\log \\left(\\hat {p _ {i}}\\right) + \\sum_ {j} p _ {j} \\cdot \\log \\left(\\hat {p _ {j}}\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.656, + 0.334, + 0.673 + ], + "angle": 0, + "content": "where \\(i\\in D\\) and \\(j\\in N\\setminus D\\)" + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.684, + 0.245, + 0.699 + ], + "angle": 0, + "content": "3 Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.71, + 0.491, + 0.92 + ], + "angle": 0, + "content": "Machine Translation on three widely-used benchmarks (Ding et al., 2020, 2021c, 2022): small-scale WMT16 English-Romanian (En-Ro; 0.6M), medium-scale WMT14 English-German (En-De; 4.5M), and large-scale WMT14 English-French (En-Fr; 36.0M). We implement the baselines and our approach under Transformer-base settings. We follow the previous adaptive training approach (Gu et al., 2020) to pretrain with the cross-entropy loss with \\(N\\) steps, and further finetune the same steps with different adaptive training objectives, including Freq-Exponential (Gu et al., 2020), Freq-Chi-Square (Gu et al., 2020), D2GPo (Li et al., 2020)," + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.317, + 0.885, + 0.606 + ], + "angle": 0, + "content": "BMI-adaptive (Xu et al., 2021), MixCrossEntropy (Li and Lu, 2021), CBMI-adaptive (Zhang et al., 2022a), and SPL (Wan et al., 2020). For \\(N\\), we adopt 100K and 30K for larger datasets, e.g. En-De and En-Fr, and small dataset, i.e. En-Ro, respectively. We empirically adopt 32K tokens per batch for large datasets, the learning rate warms up to 1e-7 for 10K steps, and then decays 90K, while for small dataset En-Ro, The learning rate warms up to 1e-7 for 4K steps, and then decays 26K steps. All the experiments are conducted on 4 NVIDIA Tesla A100 GPUs. The SacreBLEU (Post, 2018) was used for evaluation. Besides translation, we also follow previous works (Liu et al., 2021b; Zhong et al., 2022; Zhang et al., 2022b) to validate the universality of our method on more sequence-to-sequence learning tasks, e.g., summarization and grammatical error correction." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.608, + 0.884, + 0.688 + ], + "angle": 0, + "content": "Text Summarization on XSUM corpus (0.2M). We follow fairseq (Ott et al., 2019) to preprocess the data and train the model, then finetune them for the same steps. We evaluated with the ROUGE (Lin, 2004), i.e. R-1, R-2, and R-L." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.689, + 0.884, + 0.787 + ], + "angle": 0, + "content": "Grammatical Error Correction on CoNLL14 (1.4M). We follow Chollampatt and Ng (2018) to preprocess the data and train the model, then finetune them for the same steps. The MaxMatch \\((\\mathbf{M}^2)\\) scores (Dahlmeier and Ng, 2012) were used for evaluation with precision, recall, and \\(\\mathrm{F_{0.5}}\\) values." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.801, + 0.661, + 0.815 + ], + "angle": 0, + "content": "3.1 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.823, + 0.885, + 0.919 + ], + "angle": 0, + "content": "SE brings gains across language pairs and scales. Results on machine translation across different data sizes ranging from 0.6M to 36M in Table 1 show that our SE-equipped Transformer “+ Self-Evolution (ours)” 1) considerably improves the performance by averaging +0.92 BLEU points; 2) out" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.516, + 0.941 + ], + "angle": 0, + "content": "843" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.139, + 0.082, + 0.468, + 0.163 + ], + "angle": 0, + "content": "
Valid Loss Scale
0-11-22-3>3
Transformer + SE63.310.56.719.5
65.69.55.819.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.172, + 0.49, + 0.219 + ], + "angle": 0, + "content": "Table 3: The token distribution \\((\\%)\\) on different loss scales. Shadowed areas mean accurate token prediction estimated with lower cross-entropy loss, i.e. \"0-1\"." + }, + { + "type": "table", + "bbox": [ + 0.132, + 0.229, + 0.479, + 0.31 + ], + "angle": 0, + "content": "
MethodWMT22 De⇒En
BLEUΔCOMETΔ
Transformer29.98-45.1
+SE30.38+0.446.3+1.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.318, + 0.489, + 0.349 + ], + "angle": 0, + "content": "Table 4: Performance on extremely large dataset WMT22 De-En (236M)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.374, + 0.49, + 0.44 + ], + "angle": 0, + "content": "performs previous competitive method “+ CBMI-adaptive” by up to +0.47 BLEU points on large dataset WMT14 En-Fr. These results demonstrate the effectiveness and universality of our SE." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.448, + 0.49, + 0.61 + ], + "angle": 0, + "content": "SE brings gains across tasks and backbone sizes. Table 2 lists the performance on more tasks, including translation, summarization, and grammar error correction, upon large pretrained backbone - BART (Lewis et al., 2020), which has above 600M parameters. Compared to a stronger baseline, our SE significantly and incrementally improves the generation quality in all tasks, i.e. \\(+0.4\\) BLEU, \\(+0.7\\) RG-L, and \\(+1.9\\) \\(\\mathrm{F_{0.5}}\\), respectively, showing our SE is robustly applicable to general scenarios." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.619, + 0.49, + 0.765 + ], + "angle": 0, + "content": "SE works well on extremely large dataset. To further verify the effectiveness of SE on extremely large dataset, we conducted an experiment on WMT22 De-En processed by Zan et al. (2022b), which contains 236M training examples. The results in Table 4 show that our method can achieve \\(+0.4\\) and \\(+1.2\\) improvement in BLEU and COMET respectively, which proves that our SE also works on extremely large datasets." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.776, + 0.229, + 0.791 + ], + "angle": 0, + "content": "3.2 Analysis" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.797, + 0.49, + 0.847 + ], + "angle": 0, + "content": "We provide some insights to better understand the effectiveness of our approach. The ablation of important modules and parameters is in Appendix A." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.855, + 0.49, + 0.921 + ], + "angle": 0, + "content": "SE learns better token representation. To verify whether our method helps learn better tokens representation, we conduct analysis on WMT14 EnDe from learning loss and fine-grained generation" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.081, + 0.883, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.174, + 0.883, + 0.204 + ], + "angle": 0, + "content": "Figure 2: Fine-grained translation quality across word frequencies and sentence lengths." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.231, + 0.707, + 0.245 + ], + "angle": 0, + "content": "perspectives, respectively." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.246, + 0.885, + 0.568 + ], + "angle": 0, + "content": "First, we count the token ratios distributed in different cross-entropy loss scales in Table 3 following Zan et al. (2022a). Cross-entropy is a good indicator to quantify the distance between the predicted distribution and the ground truth in the valid dataset, and a lower value means a more similar distribution. As shown, our method improves the low-loss token ratios by \\(+2.3\\%\\), indicating SE helps the model learn better token representations by reducing the token uncertainty. In addition, we follow Ding et al. (2021a); Liu et al. (2021a) to break the translation down into different granularities and measure their fined-grained performance. In particular, we calculate1 the F-measure of words by different frequency buckets and BLEU scores of buckets of different lengths in Figure 2. We see SE achieves better performance in all frequencies and sentence buckets, demonstrating our method can improve the performance of different granularities." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.578, + 0.884, + 0.82 + ], + "angle": 0, + "content": "SE encourages diverse generations. Lacking generation diversity is a notorious problem for Seq2Seq learning tasks (Sun et al., 2020; Lin et al., 2022). Benefiting from better exploring the model's prediction with corrected soft labels, SE is expected to improve generation diversity. We follow Wang et al. (2022) to examine this by analyzing the performance in an additional multiple-reference test of WMT'14 En-De (Ott et al., 2018). We choose additional references for each of the 500 test sentences taken from the original test. Table 5 shows SE consistently outperforms the baseline with the average improvement being 0.9/1.0 BLEU, which indicates that our SE can effectively generate diverse results." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.83, + 0.884, + 0.895 + ], + "angle": 0, + "content": "SE enhances model generalization. Benefiting from better hard token exploration, SE-equipped Transformers are expected to own better generalizations. We examine it by testing on domain shift" + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.905, + 0.787, + 0.919 + ], + "angle": 0, + "content": "Using compare-mt (Neubig et al., 2019)." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.516, + 0.941 + ], + "angle": 0, + "content": "844" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.119, + 0.083, + 0.486, + 0.188 + ], + "angle": 0, + "content": "
Ref.Avg.Top
Transformer+SETransformer+SE
#142.543.7 (+1.2)44.945.7 (+0.8)
#228.629.3 (+0.7)30.231.2 (+1.0)
#331.232.1 (+0.9)33.234.4 (+1.2)
#428.128.8 (+0.7)29.630.5 (+0.9)
Mean32.633.5 (+0.9)34.535.5 (+1.0)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.198, + 0.49, + 0.227 + ], + "angle": 0, + "content": "Table 5: Multi-reference performance. 'Avg./ Top' means the averaging/ most-matching performance." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.241, + 0.486, + 0.3 + ], + "angle": 0, + "content": "
ModelLawMed.Kor.Sub.Avg.
Transformer41.230.97.414.523.5
+SE42.6†32.3†7.8†15.0†24.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.31, + 0.489, + 0.368 + ], + "angle": 0, + "content": "Table 6: Performance on domain shift setting. Models are trained on the news but evaluated on out-of-domain test sets, including law, medicine, koran, and subtitle. “†” indicates statistically significance \\((p < 0.05)\\)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.394, + 0.489, + 0.492 + ], + "angle": 0, + "content": "scenarios following Ding et al. (2021b). In particular, we evaluate WMT14 En-De models over four out-of-domain test sets (Müller et al., 2020) in Table 6 and find that SE improves the translation by averaging \\(+0.9\\) BLEU points, showing a better lexical generalization ability." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.501, + 0.489, + 0.758 + ], + "angle": 0, + "content": "SE encourages human-like generations. We design two types of evaluation on WMT14 En-Fr: 1) AUTOMATIC EVALUATION with COMET (Rei et al., 2020) and BLEURT (Sellam et al., 2020), which have a high-level correlation with human judgments. 2) HUMAN EVALUATION with three near-native French annotators who hold DALF C2 certificate2. Specifically, for human evaluation, we randomly sample 50 sentences from the test set to evaluate the translation adequacy and fluency, scoring \\(1 \\sim 5\\). For adequacy, 1 represents irrelevant to the source while 5 means semantically equal. For fluency, 1 means unintelligible while 5 means fluent and native. Table 7 shows the automatic and human evaluation results, where we find that our SE indeed achieves human-like translation." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.772, + 0.247, + 0.787 + ], + "angle": 0, + "content": "4 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.798, + 0.489, + 0.895 + ], + "angle": 0, + "content": "In this paper, we propose a self-evolution learning mechanism to improve seq2seq learning, by exploiting the informative-yet-underexplored tokens dynamically. SE follows two stages, i.e. self-questioning and self-evolution training, and can be used to evolve any pretrained models with a sim" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.083, + 0.88, + 0.151 + ], + "angle": 0, + "content": "
AUTOMATIC EVAL.HUMAN EVAL.
COMETBLEURTAdequacyFluency
Transformer + SE61.668.64.324.58
63.769.54.504.68
" + }, + { + "type": "table_caption", + "bbox": [ + 0.532, + 0.16, + 0.858, + 0.174 + ], + "angle": 0, + "content": "Table 7: Human evaluation on WMT14 En-Fr." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.201, + 0.884, + 0.279 + ], + "angle": 0, + "content": "ple recipe: continue train with SE. We empirically demonstrated the effectiveness and universality of SE on a series of widely-used benchmarks, covering low, medium, high, and extremely-high data volumes." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.282, + 0.885, + 0.442 + ], + "angle": 0, + "content": "In the future, besides generation tasks, we would like to verify the effectiveness of SE on language understanding tasks (Wu et al., 2020; Zhong et al., 2023). Also, it will be interesting to design SE-inspired instruction tuning or prompting strategy like Lu et al. (2023) to enhance the performance of large language models, e.g. ChatGPT3, which after all have already been fully validated on lots of conditional generation tasks (Hendy et al., 2023; Jiao et al., 2023; Peng et al., 2023; Wu et al., 2023)." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.455, + 0.614, + 0.47 + ], + "angle": 0, + "content": "Limitations" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.481, + 0.884, + 0.641 + ], + "angle": 0, + "content": "Our work has several potential limitations. First, we determine the threshold \\(\\Gamma\\) by manual selection, which may limit the performance of Seq2Seq models, it will make our work more effective and elegant if we dynamically select the threshold. Second, besides the improvement on three widely used tasks, we believe that there are still other abilities, like code generation, of Seq2Seq models that can be improved by our method, which are not fully explored in this work." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.654, + 0.662, + 0.669 + ], + "angle": 0, + "content": "Ethics Statement" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.68, + 0.884, + 0.809 + ], + "angle": 0, + "content": "We take ethical considerations very seriously and strictly adhere to the ACL Ethics Policy. This paper focuses on effective training for sequence-to-sequence learning. The datasets used in this paper are publicly available and have been widely adopted by researchers. We ensure that the findings and conclusions of this paper are reported accurately and objectively." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.821, + 0.673, + 0.838 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.848, + 0.882, + 0.896 + ], + "angle": 0, + "content": "We are grateful to the anonymous reviewers and the area chair for their insightful comments and suggestions." + }, + { + "type": "page_footnote", + "bbox": [ + 0.136, + 0.904, + 0.378, + 0.919 + ], + "angle": 0, + "content": "2http://www.delfdalf.fr/dalf-c2-en.html" + }, + { + "type": "page_footnote", + "bbox": [ + 0.53, + 0.904, + 0.759, + 0.919 + ], + "angle": 0, + "content": "3https://chat.openai.com/" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.928, + 0.516, + 0.941 + ], + "angle": 0, + "content": "845" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.116, + 0.085, + 0.214, + 0.099 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.108, + 0.487, + 0.147 + ], + "angle": 0, + "content": "Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2015. Neural machine translation by jointly learning to align and translate. In ICLR." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.158, + 0.486, + 0.196 + ], + "angle": 0, + "content": "Kehai Chen, Rui Wang, Masao Utiyama, and Eiichiro Sumita. 2020. Content word aware neural machine translation. In ACL." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.209, + 0.487, + 0.247 + ], + "angle": 0, + "content": "Jianpeng Cheng and Mirella Lapata. 2016. Neural summarization by extracting sentences and words. In ACL." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.259, + 0.487, + 0.298 + ], + "angle": 0, + "content": "Shamil Chollampatt and Hwee Tou Ng. 2018. A multilayer convolutional encoder-decoder neural network for grammatical error correction. In AAAI." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.31, + 0.487, + 0.347 + ], + "angle": 0, + "content": "Kenneth Church and Patrick Hanks. 1990. Word association norms, mutual information, and lexicography. CL." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.36, + 0.487, + 0.386 + ], + "angle": 0, + "content": "Daniel Dahlmeier and Hwee Tou Ng. 2012. Better evaluation for grammatical error correction. In NAACL." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.398, + 0.487, + 0.45 + ], + "angle": 0, + "content": "Liang Ding, Longyue Wang, Xuebo Liu, Derek F Wong, Dacheng Tao, and Zhaopeng Tu. 2021a. Progressive multi-granularity training for non-autoregressive translation. In Findings of ACL." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.461, + 0.487, + 0.525 + ], + "angle": 0, + "content": "Liang Ding, Longyue Wang, Xuebo Liu, Derek F Wong, Dacheng Tao, and Zhaopeng Tu. 2021b. Rejuvenating low-frequency words: Making the most of parallel data in non-autoregressive translation. In ACL." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.538, + 0.487, + 0.589 + ], + "angle": 0, + "content": "Liang Ding, Longyue Wang, Xuebo Liu, Derek F Wong, Dacheng Tao, and Zhaopeng Tu. 2021c. Understanding and improving lexical choice in non-autoregressive translation. In ICLR." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.602, + 0.487, + 0.653 + ], + "angle": 0, + "content": "Liang Ding, Longyue Wang, Shuming Shi, Dacheng Tao, and Zhaopeng Tu. 2022. Redistributing low-frequency words: Making the most of monolingual data in non-autoregressive translation. In ACL." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.665, + 0.487, + 0.703 + ], + "angle": 0, + "content": "Liang Ding, Longyue Wang, and Dacheng Tao. 2020. Self-attention with cross-lingual position representation. In ACL." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.715, + 0.487, + 0.766 + ], + "angle": 0, + "content": "Shuhao Gu, Jinchao Zhang, Fandong Meng, Yang Feng, Wanying Xie, Jie Zhou, and Dong Yu. 2020. Token-level adaptive training for neural machine translation. In EMNLP." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.778, + 0.487, + 0.816 + ], + "angle": 0, + "content": "Sangchul Hahn and Heeyoul Choi. 2019. Self-knowledge distillation in natural language processing. In RANLP." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.829, + 0.487, + 0.868 + ], + "angle": 0, + "content": "Amr Hendy, Mohamed Abdelrehim, et al. 2023. How good are gpt models at machine translation? a comprehensive evaluation. arXiv preprint." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.879, + 0.486, + 0.919 + ], + "angle": 0, + "content": "Wenxiang Jiao, Wenxuan Wang, Jen-tse Huang, Xing Wang, and Zhaopeng Tu. 2023. Is chatgpt a good translator? a preliminary study. arXiv preprint." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.108, + 0.487, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.086, + 0.882, + 0.165 + ], + "angle": 0, + "content": "Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In ACL." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.175, + 0.881, + 0.201 + ], + "angle": 0, + "content": "Haoran Li and Wei Lu. 2021. Mixed cross entropy loss for neural machine translation. In ICML." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.213, + 0.881, + 0.25 + ], + "angle": 0, + "content": "Zuchao Li, Rui Wang, et al. 2020. Data-dependent gaussian prior objective for language generation. In ICLR." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.263, + 0.881, + 0.301 + ], + "angle": 0, + "content": "Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.313, + 0.882, + 0.377 + ], + "angle": 0, + "content": "Huan Lin, Baosong Yang, Liang Yao, Dayiheng Liu, Haibo Zhang, Jun Xie, Min Zhang, and Jinsong Su. 2022. Bridging the gap between training and inference: Multi-candidate optimization for diverse neural machine translation. In Findings of NAACL." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.388, + 0.882, + 0.441 + ], + "angle": 0, + "content": "Xuebo Liu, Longyue Wang, Derek F Wong, Liang Ding, Lidia S Chao, Shuming Shi, and Zhaopeng Tu. 2021a. On the copying behaviors of pre-training for neural machine translation. In Findings of ACL." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.452, + 0.882, + 0.504 + ], + "angle": 0, + "content": "Xuebo Liu, Longyue Wang, Derek F Wong, Liang Ding, Lidia S Chao, and Zhaopeng Tu. 2021b. Understanding and improving encoder layer fusion in sequence-to-sequence learning. In ICLR." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.515, + 0.882, + 0.58 + ], + "angle": 0, + "content": "Qingyu Lu, Baopu Qiu, Liang Ding, Liping Xie, and Dacheng Tao. 2023. Error analysis prompting enables human-like translation evaluation in large language models: A case study on chatgpt. arXiv preprint." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.591, + 0.881, + 0.641 + ], + "angle": 0, + "content": "Mengqi Miao, Fandong Meng, Yijin Liu, Xiao-Hua Zhou, and Jie Zhou. 2021. Prevent the language model from being overconfident in neural machine translation. In ACL." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.653, + 0.881, + 0.692 + ], + "angle": 0, + "content": "Mathias Müller, Annette Rios, and Rico Sennrich. 2020. Domain robustness in neural machine translation. In AMTA, Virtual." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.703, + 0.882, + 0.755 + ], + "angle": 0, + "content": "Graham Neubig, Zi-Yi Dou, Junjie Hu, Paul Michel, Danish Pruthi, and Xinyi Wang. 2019. compare-mt: A tool for holistic comparison of language generation systems. In NAACL." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.766, + 0.882, + 0.805 + ], + "angle": 0, + "content": "Mohammad Norouzi, Samy Bengio, Zhifeng Chen, et al. 2016. Reward augmented maximum likelihood for neural structured prediction. In NeurIPS." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.816, + 0.881, + 0.855 + ], + "angle": 0, + "content": "Myle Ott, Michael Auli, David Grangier, and Marc'Aurelio Ranzato. 2018. Analyzing uncertainty in neural machine translation. In ICML." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.866, + 0.882, + 0.918 + ], + "angle": 0, + "content": "Myle Ott, Sergey Edunov, Alexei Baevski, Angela Fan, Sam Gross, Nathan Ng, David Grangier, and Michael Auli. 2019. *fairoseq: A fast, extensible toolkit for sequence modeling*. In *NAACL Demonstration*." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.882, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.516, + 0.94 + ], + "angle": 0, + "content": "846" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.086, + 0.487, + 0.139 + ], + "angle": 0, + "content": "Keqin Peng, Liang Ding, Qihuang Zhong, Li Shen, Xuebo Liu, Min Zhang, Yuanxin Ouyang, and Dacheng Tao. 2023. Towards making the most of chatgpt for machine translation. arxiv preprint." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.148, + 0.487, + 0.186 + ], + "angle": 0, + "content": "Steven T Piantadosi. 2014. Zipf's word frequency law in natural language: A critical review and future directions. Psychonomic bulletin & review." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.197, + 0.485, + 0.222 + ], + "angle": 0, + "content": "Matt Post. 2018. A call for clarity in reporting BLEU scores. In WMT." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.232, + 0.487, + 0.295 + ], + "angle": 0, + "content": "Lynne M Reder, Xiaonan L Liu, Alexander Keinath, and Vencislav Popov. 2016. Building knowledge requires bricks, not sand: The critical role of familiar constituents in learning. Psychonomic bulletin & review." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.307, + 0.486, + 0.345 + ], + "angle": 0, + "content": "Ricardo Rei, Craig Stewart, Ana C. Farinha, and Alon Lavie. 2020. COMET: A neural framework for MT evaluation. In EMNLP." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.355, + 0.487, + 0.394 + ], + "angle": 0, + "content": "Thibault Sellam, Dipanjan Das, and Ankur P. Parikh. 2020. BLEURT: learning robust metrics for text generation. In ACL." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.404, + 0.487, + 0.442 + ], + "angle": 0, + "content": "Zewei Sun, Shujian Huang, Hao-Ran Wei, Xinyu Dai, and Jiajun Chen. 2020. Generating diverse translation by manipulating multi-head attention. In AAAI." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.452, + 0.487, + 0.489 + ], + "angle": 0, + "content": "Ilya Sutskever, Oriol Vinyals, and Quoc V. Le. 2014. Sequence to sequence learning with neural networks. In NeurIPS." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.501, + 0.487, + 0.551 + ], + "angle": 0, + "content": "Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. 2016. Rethinking the inception architecture for computer vision. In CVPR." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.562, + 0.485, + 0.587 + ], + "angle": 0, + "content": "Ashish Vaswani, Noam Shazeer, et al. 2017. Attention is all you need. In NeurIPS." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.597, + 0.485, + 0.623 + ], + "angle": 0, + "content": "Yu Wan, Baosong Yang, et al. 2020. Self-paced learning for neural machine translation. In EMNLP." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.633, + 0.487, + 0.697 + ], + "angle": 0, + "content": "Wenxuan Wang, Wenxiang Jiao, Yongchang Hao, Xing Wang, Shuming Shi, Zhaopeng Tu, and Michael R. Lyu. 2022. Understanding and improving sequence-to-sequence pretraining for neural machine translation. In ACL." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.707, + 0.487, + 0.746 + ], + "angle": 0, + "content": "Di Wu, Liang Ding, Fan Lu, and Jian Xie. 2020. Slotrefine: A fast non-autoregressive model for joint intent detection and slot filling. In EMNLP." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.756, + 0.486, + 0.809 + ], + "angle": 0, + "content": "Haoran Wu, Wenxuan Wang, Yuxuan Wan, Wenxiang Jiao, and Michael Lyu. 2023. Chatgpt or grammarly? evaluating chatgpt on grammatical error correction benchmark. arXiv preprint." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.817, + 0.487, + 0.856 + ], + "angle": 0, + "content": "Fengshun Xiao, Yingting Wu, Hai Zhao, Rui Wang, and Shu Jiang. 2019. Dual skew divergence loss for neural machine translation. CoRR." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.866, + 0.487, + 0.917 + ], + "angle": 0, + "content": "Yangyifan Xu, Yijin Liu, Fandong Meng, Jiajun Zhang, Jinan Xu, and Jie Zhou. 2021. Bilingual mutual information based adaptive training for neural machine translation. In ACL." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.487, + 0.917 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.086, + 0.882, + 0.124 + ], + "angle": 0, + "content": "Zheng Yuan and Ted Briscoe. 2016. Grammatical error correction using neural machine translation. In NAACL." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.136, + 0.882, + 0.187 + ], + "angle": 0, + "content": "Changtong Zan, Liang Ding, Li Shen, Yu Cao, Weifeng Liu, and Dacheng Tao. 2022a. On the complementarity between pre-training and random-initialization for resource-rich machine translation. In COLING." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.198, + 0.882, + 0.238 + ], + "angle": 0, + "content": "Changtong Zan, Keqin Peng, Liang Ding, et al. 2022b. Vega-mt: The jd explore academy machine translation system for wmt22. In WMT." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.248, + 0.882, + 0.299 + ], + "angle": 0, + "content": "Runtian Zhai, Chen Dan, J Zico Kolter, and Pradeep Kumar Ravikumar. 2023. Understanding why generalized reweighting does not improve over ERM. In ICLR." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.311, + 0.882, + 0.362 + ], + "angle": 0, + "content": "Songming Zhang, Yijin Liu, Fandong Meng, Yufeng Chen, Jinan Xu, Jian Liu, and Jie Zhou. 2022a. Conditional bilingual mutual information based adaptive training for neural machine translation. In ACL." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.374, + 0.882, + 0.426 + ], + "angle": 0, + "content": "Zheng Zhang, Liang Ding, Dazhao Cheng, Xuebo Liu, Min Zhang, and Dacheng Tao. 2022b. Bliss: Robust sequence-to-sequence learning via self-supervised input representation. arXiv preprint." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.437, + 0.882, + 0.489 + ], + "angle": 0, + "content": "Qihuang Zhong, Liang Ding, Juhua Liu, Bo Du, and Dacheng Tao. 2022. E2s2: Encoding-enhanced sequence-to-sequence pretraining for language understanding and generation. arXiv preprint." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.499, + 0.882, + 0.564 + ], + "angle": 0, + "content": "Qihuang Zhong, Liang Ding, Keqin Peng, Juhua Liu, Bo Du, Li Shen, Yibing Zhan, and Dacheng Tao. 2023. Bag of tricks for effective language model pretraining and downstream adaptation: A case study on glue. arXiv preprint." + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.086, + 0.882, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.511, + 0.579, + 0.633, + 0.595 + ], + "angle": 0, + "content": "A Appendix" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.604, + 0.884, + 0.779 + ], + "angle": 0, + "content": "Parameter Analysis on \\(\\Gamma\\) As stated in §2.1, we use the loss threshold \\(\\Gamma\\) to dynamically select the hard-to-learn tokens. Here, we analyze the influence of different \\(\\Gamma\\) in detail. In practice, we train the Transformer models with different \\(\\Gamma\\) (in \\(\\{3,4,5,6\\}\\)) and evaluate the performance of the WMT14 En-De test set. Table 8 lists the performance of different \\(\\Gamma\\). The results of Table 8 show that SE is stable and insensitive to \\(\\Gamma\\) within a certain range. Noting that we select \\(\\Gamma = 5\\) for all experiment settings based on the results in Table 8." + }, + { + "type": "table", + "bbox": [ + 0.572, + 0.805, + 0.824, + 0.849 + ], + "angle": 0, + "content": "
Γ=3Γ=4Γ=5Γ=6
BLEU27.727.828.027.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.518, + 0.859, + 0.872, + 0.872 + ], + "angle": 0, + "content": "Table 8: Parameter analysis of \\( \\Gamma \\) on WMT14 En-De." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.516, + 0.94 + ], + "angle": 0, + "content": "847" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.115, + 0.085, + 0.239, + 0.1 + ], + "angle": 0, + "content": "Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.11, + 0.49, + 0.351 + ], + "angle": 0, + "content": "Metric. In this work, we use the loss-based metric to dynamically select the hard-to-learn tokens. To validate the effectiveness of the metric, we use a simple adaptive training method (\"+ ADD\") that adds 1 to the weighting term of loss of the hard-to-learn tokens. The results on WMT16 EnRo are shown in Table 9, the simple Add method can achieve +0.3 BLEU improvement compared to the baseline model, which proves that our proposed self-questioning stage indeed mines informative difficult tokens. Also, we can observe that learning these dynamic difficult tokens with our SE framework (\"+ SE\") could outperform \"+\" ADD\" by +0.6 BLUE points, demonstrating the superiority of our token-specific label smoothing approach." + }, + { + "type": "table", + "bbox": [ + 0.177, + 0.361, + 0.432, + 0.406 + ], + "angle": 0, + "content": "
Baseline+ ADD+ SE
BLEU35.135.436.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.121, + 0.415, + 0.479, + 0.43 + ], + "angle": 0, + "content": "Table 9: Ablation performance of our SE. on Metric." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.455, + 0.49, + 0.6 + ], + "angle": 0, + "content": "Learning objective. As stated in §2.1, our learning objective is the combination of the ground truth and the model's prediction. To validate the effectiveness of predicted distribution, we conduct ablation experiments on WMT16 En-Ro and WMT14 En-De. The results in Table 10 show that adding the predicted distribution will consistently improve the model's performance, which proves the effectiveness of the predicted distribution." + }, + { + "type": "table", + "bbox": [ + 0.156, + 0.61, + 0.452, + 0.698 + ], + "angle": 0, + "content": "
MethodBLEU
EN⇒DEEN⇒Ro
Transformer27.0835.11
SE28.0236.02
-w/o predicted results27.8935.71
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.707, + 0.486, + 0.734 + ], + "angle": 0, + "content": "Table 10: Ablation performance of our SE. on learning objective." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.516, + 0.94 + ], + "angle": 0, + "content": "848" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.133, + 0.085, + 0.435, + 0.1 + ], + "angle": 0, + "content": "ACL 2023 Responsible NLP Checklist" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.108, + 0.323, + 0.123 + ], + "angle": 0, + "content": "A For every submission:" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.127, + 0.533, + 0.144 + ], + "angle": 0, + "content": "A1. Did you describe the limitations of your work?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.145, + 0.366, + 0.16 + ], + "angle": 0, + "content": "The last section of the paper." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.171, + 0.553, + 0.187 + ], + "angle": 0, + "content": "A2. Did you discuss any potential risks of your work?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.188, + 0.233, + 0.202 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.213, + 0.696, + 0.229 + ], + "angle": 0, + "content": "A3. Do the abstract and introduction summarize the paper's main claims?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.231, + 0.46, + 0.245 + ], + "angle": 0, + "content": "The abstract and the introduction section." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.256, + 0.669, + 0.273 + ], + "angle": 0, + "content": "A4. Have you used AI writing assistants when working on this paper?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.274, + 0.233, + 0.288 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.301, + 0.49, + 0.317 + ], + "angle": 0, + "content": "B Did you use or create scientific artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.322, + 0.215, + 0.337 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.348, + 0.53, + 0.364 + ], + "angle": 0, + "content": "B1. Did you cite the creators of artifacts you used?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.365, + 0.249, + 0.38 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.391, + 0.779, + 0.407 + ], + "angle": 0, + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.408, + 0.249, + 0.423 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.434, + 0.882, + 0.497 + ], + "angle": 0, + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.499, + 0.249, + 0.514 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.525, + 0.882, + 0.573 + ], + "angle": 0, + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.575, + 0.249, + 0.589 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.6, + 0.882, + 0.632 + ], + "angle": 0, + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.634, + 0.249, + 0.648 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.659, + 0.882, + 0.74 + ], + "angle": 0, + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be." + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.741, + 0.249, + 0.755 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.766, + 0.495, + 0.782 + ], + "angle": 0, + "content": "C Did you run computational experiments?" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.788, + 0.215, + 0.802 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.814, + 0.882, + 0.846 + ], + "angle": 0, + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.847, + 0.249, + 0.861 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.869, + 0.878, + 0.893 + ], + "angle": 0, + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.517, + 0.941 + ], + "angle": 0, + "content": "849" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.129, + 0.085, + 0.88, + 0.133 + ], + "angle": 0, + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.144, + 0.881, + 0.207 + ], + "angle": 0, + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.219, + 0.881, + 0.283 + ], + "angle": 0, + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? No response." + }, + { + "type": "list", + "bbox": [ + 0.129, + 0.085, + 0.881, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.293, + 0.878, + 0.311 + ], + "angle": 0, + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.315, + 0.221, + 0.329 + ], + "angle": 0, + "content": "Section 3.2" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.341, + 0.881, + 0.388 + ], + "angle": 0, + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.4, + 0.881, + 0.464 + ], + "angle": 0, + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.475, + 0.881, + 0.54 + ], + "angle": 0, + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.55, + 0.873, + 0.582 + ], + "angle": 0, + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.593, + 0.881, + 0.641 + ], + "angle": 0, + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Left blank." + }, + { + "type": "list", + "bbox": [ + 0.13, + 0.341, + 0.881, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.516, + 0.941 + ], + "angle": 0, + "content": "850" + } + ] +] \ No newline at end of file diff --git a/2023/Token-Level Self-Evolution Training for Sequence-to-Sequence Learning/8f4944a3-fb6e-4f22-bf90-4b7384f2525a_origin.pdf b/2023/Token-Level Self-Evolution Training for Sequence-to-Sequence Learning/8f4944a3-fb6e-4f22-bf90-4b7384f2525a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ccacd9ba6359a570527093244c323994bdea6542 --- /dev/null +++ b/2023/Token-Level Self-Evolution Training for Sequence-to-Sequence Learning/8f4944a3-fb6e-4f22-bf90-4b7384f2525a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6f6ba59816cfde7f8986fbc1b585e565c30c67a6da3db3dcb4b2879cf168f7d6 +size 2656167 diff --git a/2023/Token-Level Self-Evolution Training for Sequence-to-Sequence Learning/full.md b/2023/Token-Level Self-Evolution Training for Sequence-to-Sequence Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..db38c2935c379faae7ca8a933633e8ac861c7fb3 --- /dev/null +++ b/2023/Token-Level Self-Evolution Training for Sequence-to-Sequence Learning/full.md @@ -0,0 +1,303 @@ +# Token-Level Self-Evolution Training for Sequence-to-Sequence Learning + +Keqin Peng $^{1*}$ , Liang Ding $^{2*}$ , Qihuang Zhong $^{3}$ +Yuanxin Ouyang $^{1†}$ , Wenge Rong $^{1}$ , Zhang Xiong $^{1}$ , Dacheng Tao $^{4}$ + +1Beihang University 2Zhejiang University 3Wuhan University 4The University of Sydney {keqin.peng, oyyx, w.rong, xiongz}@buaa.edu.cn zhongqihuang@whu.edu.cn, {liangding.liam, dacheng.tao}@gmail.com + +# Abstract + +Adaptive training approaches, widely used in sequence-to-sequence models, commonly reweigh the losses of different target tokens based on priors, e.g. word frequency. However, most of them do not consider the variation of learning difficulty in different training steps, and overly emphasize the learning of difficult one-hot labels, making the learning deterministic and sub-optimal. In response, we present Token-Level Self-Evolution Training (SE), a simple and effective dynamic training method to fully and wisely exploit the knowledge from data. SE focuses on dynamically learning the under-explored tokens for each forward pass and adaptively regularizes the training by introducing a novel token-specific label smoothing approach. Empirically, SE yields consistent and significant improvements in three tasks, i.e. machine translation, summarization, and grammatical error correction. Encouragingly, we achieve averaging $+0.93$ BLEU improvement on three machine translation tasks. Analyses confirm that, besides improving lexical accuracy, SE enhances generation diversity and model generalization. + +# 1 Introduction + +Sequence-to-sequence learning (Seq2Seq) with neural networks (Sutskever et al., 2014) has advanced the state-of-the-art in various NLP tasks, e.g. translation (Bahdanau et al., 2015; Vaswani et al., 2017), summarization (Cheng and Lapata, 2016), and grammatical error correction (Yuan and Briscoe, 2016). Generally, Seq2Seq models are trained with the cross-entropy loss, which equally weighs the training losses of different target tokens. + +However, due to the token imbalance nature (Piantadosi, 2014) and the truth that different tokens contribute differently to the sentence meaning (Church and Hanks, 1990; Chen et al., 2020), + +![](images/43bb068707a45de0c88ee7d411df1b2bc2378d00ca71a2a2df7381da1549eb90.jpg) +Figure 1: An example to illustrate the changing token difficulties in different training steps in WMT'14 En-De. The token "abschreiben/ Sache" is hard/ easy to learn at 50K while the trend is totally reversed at 100K. + +several works are developed to reweigh the token-level training loss according to explicit (e.g. frequency) or implicit (uncertainty estimated by off-the-shelf language models) priors (Gu et al., 2020; Xu et al., 2021; Zhang et al., 2022a). For example, Gu et al. (2020) proposed two heuristic criteria based on word frequency to encourage the model to learn from larger-weight low-frequency tokens. Zhang et al. (2022a) introduce target-context-aware metric based on an additional target-side language model to adjust the weight of each target token. + +Despite some success, there are still limitations in these adaptive training approaches. First, most of them predetermine the difficult tokens and fix such prior to guiding the training. However, in our preliminary study, we find the hard-to-learn tokens are dynamically changing during training, rather than statically fixed. As shown in Figure 1, as the training progress goes, although the sentence-level loss is nicely converging, the difficult token is changing from "abschreiben" to "Sache" in terms of the token-level loss. Second, these adaptive training methods overly emphasize fitting the difficult tokens' one-hot labels by reweighing the loss, which empirically may cause overfitting and limit the generalization (Norouzi et al., 2016; Szegedy et al., 2016; Xiao et al., 2019; Miao et al., 2021). Also, a more recent study (Zhai et al., 2023) provides + +theoretical evidence to support that reweighting is not that effective to improve the generalization. + +Correspondingly, we design a simple and effective Token-Level Self-Evolution Training (SE) strategy to encourage Seq2Seq models to learn from difficult words that are dynamically selected by the model itself. Specifically, SE contains two stages: 1self-questioning and 2self-evolution training. In the first stage, the Seq2Seq models dynamically select the hard-to-learn tokens based on the token-level losses, then we encourage the Seq2Seq models to learn from them in the second stage, where, rather than adopting reweighing, we introduce a novel token-specific label smoothing approach to generate easily digestible soft label, which considers both the ground truth and model's prediction. + +Experiments across tasks, language pairs, data scales, and model sizes show that SE consistently and significantly outperforms both the vanilla Seq2Seq model and the re-implemented advanced baselines. Analyses confirm that besides improved lexical accuracy, SE generates diverse and human-like generations with better model generalization. + +# 2 Methodology + +Preliminary Sequence-to-sequence (Seq2Seq) learning aims to maximize the cross-entropy (CE) loss of the log-likelihood of each target word in $\mathbf{y} = \{y_1,\dots ,y_N\}$ , conditioned on source $\mathbf{x}$ , where the optimization treats all tokens equally: + +$$ +\mathcal {L} _ {\mathrm {C E}} (\theta) = - \sum_ {j = 1} ^ {N} \log p \left(y _ {j} \mid \mathbf {y} _ {< j}, \mathbf {x}; \theta\right) \tag {1} +$$ + +However, due to the different learning difficulties of each token, it is sub-optimal to treat all tokens equally (Gu et al., 2020). To address this limitation, a series of token-level adaptive training objectives were adopted to re-weight the losses of different target tokens (Xu et al., 2021; Zhang et al., 2022a). The common goal of these methods is to facilitate the model training by fully exploiting the informative but underexplored tokens. + +However, our preliminary study shows that the hard tokens are dynamically changing (see Figure 1) in different training steps (or model structures), thus it is sub-optimal to employ static token priors (e.g. frequency) during training. Also, recent studies (Zhai et al., 2023) in the ML community theoretically show that reweighting is not that effective to improve the generalization. Based on the above + +evidence, we present the self-evolution learning (SE) mechanism to encourage the model to adaptively and wisely learn from the informative yet under-explored tokens dynamically determined by the model itself (Stage① in §2.1), with an easy-to-learn label distribution (Stage② in §2.1). A similar work to ours is Hahn and Choi (2019). However, their method mainly considers the situation where the predicted answer is incorrect but close to the golden answer, while our method focuses on all dynamic hard tokens. + +# 2.1 Token-Level Self-Evolution Learning + +1 Self-questioning Stage. The goal is to select the hard-to-learn tokens that are questioned by the Seq2Seq model itself during training dynamics. Previously, these difficult tokens are predetermined by external models or specific statistical metrics. However, inspired by the finding of dynamic change of difficult tokens during the training stage as shown in Figure 1 and the finding that the trained model contains useful information (Li and Lu, 2021), e.g. synonym, we propose to straightforwardly leverage the behavior of the model to dynamically select target tokens. In practice, we first calculate the token-level CE loss, denoted as $\{l_1, l_2, \dots, l_n\}$ , for each token for each forward pass. Then we set a loss threshold $\Gamma$ and select the tokens whose losses exceed $\Gamma$ as the target tokens, i.e., $D = \{t_i | l_i > \Gamma\}$ where $i \in N = \{1, 2, \dots, n\}$ . + +$\Theta$ Self-evolution Training Stage. After selecting the difficult tokens, we encourage the model to carefully learn from them. Given the theoretical shortage (Zhai et al., 2023) and potentially caused overfitting or overconfidence problem (Miao et al., 2021) of reweighting and deliberately learning from difficult tokens, we propose to strengthen the learning from these tokens with a newly designed Token-specific Label Smoothing (TLS) approach. Specifically, motivated by the effect of label smoothing (LS) regularization (Szegedy et al., 2016), we combine the ground truth $p_i$ and the model's prediction $\hat{p}_i$ to form a new soft label $\widetilde{p}_i$ for the $i$ -th token. Then we use $\widetilde{p}$ to guide the difficult tokens $D$ , while leaving label-smoothing CE loss for the other tokens. It is worth noting that we also apply the traditional label smoothing technique to $\hat{p}_i$ to activate the information in the predicted distribution. Analogous to human learning, it is often easier for humans to grasp new things described by their familiar knowledge (Reder et al., 2016), + +
ModelWMT16 En→RoWMT14 En→DeWMT14 En→Fr
Transformer (Vaswani et al., 2017)35.1127.0840.65
+ Freq-Exponential (Gu et al., 2020)35.86 (+0.75)27.60 (+0.52)41.05 (+0.40)
+ Freq-Chi-Square (Gu et al., 2020)35.74 (+0.63)27.51 (+0.43)40.99 (+0.34)
+ D2GPo (Li et al., 2020)35.89 (+0.78)27.66 (+0.58)41.05 (+0.40)
+ BMI-adaptive (Xu et al., 2021)35.89 (+0.78)27.65 (+0.57)41.10 (+0.45)
+ MixCrossEntropy (Li and Lu, 2021)35.88 (+0.74)27.61 (+0.53)41.07 (+0.42)
+ CBMI-adaptive (Zhang et al., 2022a)35.90 (+0.79)27.69 (+0.61)41.13 (+0.48)
+ SPL (Wan et al., 2020)35.92 (+0.81)27.88 (+0.80)41.30 (+0.65)
+ Self-Evolution (ours)36.02 (+0.91)†28.02 (+0.94)†41.60 (+0.95)†
+ +Table 1: BLEU scores $(\%)$ on three translation tasks spanning different data scales, i.e. $0.6\mathrm{M}$ , $4.5\mathrm{M}$ , $36\mathrm{M}$ . “†” indicates a statistically significant difference from the powerful Transformer baseline $(p < 0.05)$ . + +
Ro-EnXSUMGEC
BLEURG-1RG-2RG-LPrec.RecallF0.5
Baseline37.343.219.834.059.139.853.9
+ SE37.7†43.820.434.7†58.946.255.8†
+ +Table 2: Performance on more tasks including translation, summarization, and grammar error correction, upon larger model BART (Lewis et al., 2020). + +therefore the new soft label fused both accurate ground truth and model's self-distribution is easily digestible. Mathematically, for difficult tokens $t_i$ , $\widetilde{p}_i$ is formulated as: + +$$ +\widetilde {p _ {i}} = \left(p _ {i} + \hat {p _ {i}}\right) / 2. \tag {2} +$$ + +Then we calculate the losses of difficult tokens and the others, and combine the two losses: + +$$ +L = - \left(\sum_ {i} \widetilde {p _ {i}} \cdot \log \left(\hat {p _ {i}}\right) + \sum_ {j} p _ {j} \cdot \log \left(\hat {p _ {j}}\right)\right), \tag {3} +$$ + +where $i\in D$ and $j\in N\setminus D$ + +# 3 Evaluation + +Machine Translation on three widely-used benchmarks (Ding et al., 2020, 2021c, 2022): small-scale WMT16 English-Romanian (En-Ro; 0.6M), medium-scale WMT14 English-German (En-De; 4.5M), and large-scale WMT14 English-French (En-Fr; 36.0M). We implement the baselines and our approach under Transformer-base settings. We follow the previous adaptive training approach (Gu et al., 2020) to pretrain with the cross-entropy loss with $N$ steps, and further finetune the same steps with different adaptive training objectives, including Freq-Exponential (Gu et al., 2020), Freq-Chi-Square (Gu et al., 2020), D2GPo (Li et al., 2020), + +BMI-adaptive (Xu et al., 2021), MixCrossEntropy (Li and Lu, 2021), CBMI-adaptive (Zhang et al., 2022a), and SPL (Wan et al., 2020). For $N$ , we adopt 100K and 30K for larger datasets, e.g. En-De and En-Fr, and small dataset, i.e. En-Ro, respectively. We empirically adopt 32K tokens per batch for large datasets, the learning rate warms up to 1e-7 for 10K steps, and then decays 90K, while for small dataset En-Ro, The learning rate warms up to 1e-7 for 4K steps, and then decays 26K steps. All the experiments are conducted on 4 NVIDIA Tesla A100 GPUs. The SacreBLEU (Post, 2018) was used for evaluation. Besides translation, we also follow previous works (Liu et al., 2021b; Zhong et al., 2022; Zhang et al., 2022b) to validate the universality of our method on more sequence-to-sequence learning tasks, e.g., summarization and grammatical error correction. + +Text Summarization on XSUM corpus (0.2M). We follow fairseq (Ott et al., 2019) to preprocess the data and train the model, then finetune them for the same steps. We evaluated with the ROUGE (Lin, 2004), i.e. R-1, R-2, and R-L. + +Grammatical Error Correction on CoNLL14 (1.4M). We follow Chollampatt and Ng (2018) to preprocess the data and train the model, then finetune them for the same steps. The MaxMatch $(\mathbf{M}^2)$ scores (Dahlmeier and Ng, 2012) were used for evaluation with precision, recall, and $\mathrm{F_{0.5}}$ values. + +# 3.1 Main Results + +SE brings gains across language pairs and scales. Results on machine translation across different data sizes ranging from 0.6M to 36M in Table 1 show that our SE-equipped Transformer “+ Self-Evolution (ours)” 1) considerably improves the performance by averaging +0.92 BLEU points; 2) out + +
Valid Loss Scale
0-11-22-3>3
Transformer + SE63.310.56.719.5
65.69.55.819.1
+ +Table 3: The token distribution $(\%)$ on different loss scales. Shadowed areas mean accurate token prediction estimated with lower cross-entropy loss, i.e. "0-1". + +
MethodWMT22 De⇒En
BLEUΔCOMETΔ
Transformer29.98-45.1
+SE30.38+0.446.3+1.2
+ +Table 4: Performance on extremely large dataset WMT22 De-En (236M). + +performs previous competitive method “+ CBMI-adaptive” by up to +0.47 BLEU points on large dataset WMT14 En-Fr. These results demonstrate the effectiveness and universality of our SE. + +SE brings gains across tasks and backbone sizes. Table 2 lists the performance on more tasks, including translation, summarization, and grammar error correction, upon large pretrained backbone - BART (Lewis et al., 2020), which has above 600M parameters. Compared to a stronger baseline, our SE significantly and incrementally improves the generation quality in all tasks, i.e. $+0.4$ BLEU, $+0.7$ RG-L, and $+1.9$ $\mathrm{F_{0.5}}$ , respectively, showing our SE is robustly applicable to general scenarios. + +SE works well on extremely large dataset. To further verify the effectiveness of SE on extremely large dataset, we conducted an experiment on WMT22 De-En processed by Zan et al. (2022b), which contains 236M training examples. The results in Table 4 show that our method can achieve $+0.4$ and $+1.2$ improvement in BLEU and COMET respectively, which proves that our SE also works on extremely large datasets. + +# 3.2 Analysis + +We provide some insights to better understand the effectiveness of our approach. The ablation of important modules and parameters is in Appendix A. + +SE learns better token representation. To verify whether our method helps learn better tokens representation, we conduct analysis on WMT14 EnDe from learning loss and fine-grained generation + +![](images/4077ecf43a4abda12283cba9c9e154f7c27f83d31e008deae2104732133fd8d5.jpg) +Figure 2: Fine-grained translation quality across word frequencies and sentence lengths. + +perspectives, respectively. + +First, we count the token ratios distributed in different cross-entropy loss scales in Table 3 following Zan et al. (2022a). Cross-entropy is a good indicator to quantify the distance between the predicted distribution and the ground truth in the valid dataset, and a lower value means a more similar distribution. As shown, our method improves the low-loss token ratios by $+2.3\%$ , indicating SE helps the model learn better token representations by reducing the token uncertainty. In addition, we follow Ding et al. (2021a); Liu et al. (2021a) to break the translation down into different granularities and measure their fined-grained performance. In particular, we calculate1 the F-measure of words by different frequency buckets and BLEU scores of buckets of different lengths in Figure 2. We see SE achieves better performance in all frequencies and sentence buckets, demonstrating our method can improve the performance of different granularities. + +SE encourages diverse generations. Lacking generation diversity is a notorious problem for Seq2Seq learning tasks (Sun et al., 2020; Lin et al., 2022). Benefiting from better exploring the model's prediction with corrected soft labels, SE is expected to improve generation diversity. We follow Wang et al. (2022) to examine this by analyzing the performance in an additional multiple-reference test of WMT'14 En-De (Ott et al., 2018). We choose additional references for each of the 500 test sentences taken from the original test. Table 5 shows SE consistently outperforms the baseline with the average improvement being 0.9/1.0 BLEU, which indicates that our SE can effectively generate diverse results. + +SE enhances model generalization. Benefiting from better hard token exploration, SE-equipped Transformers are expected to own better generalizations. We examine it by testing on domain shift + +
Ref.Avg.Top
Transformer+SETransformer+SE
#142.543.7 (+1.2)44.945.7 (+0.8)
#228.629.3 (+0.7)30.231.2 (+1.0)
#331.232.1 (+0.9)33.234.4 (+1.2)
#428.128.8 (+0.7)29.630.5 (+0.9)
Mean32.633.5 (+0.9)34.535.5 (+1.0)
+ +Table 5: Multi-reference performance. 'Avg./ Top' means the averaging/ most-matching performance. + +
ModelLawMed.Kor.Sub.Avg.
Transformer41.230.97.414.523.5
+SE42.6†32.3†7.8†15.0†24.4
+ +scenarios following Ding et al. (2021b). In particular, we evaluate WMT14 En-De models over four out-of-domain test sets (Müller et al., 2020) in Table 6 and find that SE improves the translation by averaging $+0.9$ BLEU points, showing a better lexical generalization ability. + +SE encourages human-like generations. We design two types of evaluation on WMT14 En-Fr: 1) AUTOMATIC EVALUATION with COMET (Rei et al., 2020) and BLEURT (Sellam et al., 2020), which have a high-level correlation with human judgments. 2) HUMAN EVALUATION with three near-native French annotators who hold DALF C2 certificate2. Specifically, for human evaluation, we randomly sample 50 sentences from the test set to evaluate the translation adequacy and fluency, scoring $1 \sim 5$ . For adequacy, 1 represents irrelevant to the source while 5 means semantically equal. For fluency, 1 means unintelligible while 5 means fluent and native. Table 7 shows the automatic and human evaluation results, where we find that our SE indeed achieves human-like translation. + +# 4 Conclusion + +In this paper, we propose a self-evolution learning mechanism to improve seq2seq learning, by exploiting the informative-yet-underexplored tokens dynamically. SE follows two stages, i.e. self-questioning and self-evolution training, and can be used to evolve any pretrained models with a sim + +Table 6: Performance on domain shift setting. Models are trained on the news but evaluated on out-of-domain test sets, including law, medicine, koran, and subtitle. “†” indicates statistically significance $(p < 0.05)$ . + +
AUTOMATIC EVAL.HUMAN EVAL.
COMETBLEURTAdequacyFluency
Transformer + SE61.668.64.324.58
63.769.54.504.68
+ +Table 7: Human evaluation on WMT14 En-Fr. + +ple recipe: continue train with SE. We empirically demonstrated the effectiveness and universality of SE on a series of widely-used benchmarks, covering low, medium, high, and extremely-high data volumes. + +In the future, besides generation tasks, we would like to verify the effectiveness of SE on language understanding tasks (Wu et al., 2020; Zhong et al., 2023). Also, it will be interesting to design SE-inspired instruction tuning or prompting strategy like Lu et al. (2023) to enhance the performance of large language models, e.g. ChatGPT3, which after all have already been fully validated on lots of conditional generation tasks (Hendy et al., 2023; Jiao et al., 2023; Peng et al., 2023; Wu et al., 2023). + +# Limitations + +Our work has several potential limitations. First, we determine the threshold $\Gamma$ by manual selection, which may limit the performance of Seq2Seq models, it will make our work more effective and elegant if we dynamically select the threshold. Second, besides the improvement on three widely used tasks, we believe that there are still other abilities, like code generation, of Seq2Seq models that can be improved by our method, which are not fully explored in this work. + +# Ethics Statement + +We take ethical considerations very seriously and strictly adhere to the ACL Ethics Policy. This paper focuses on effective training for sequence-to-sequence learning. The datasets used in this paper are publicly available and have been widely adopted by researchers. We ensure that the findings and conclusions of this paper are reported accurately and objectively. + +# Acknowledgement + +We are grateful to the anonymous reviewers and the area chair for their insightful comments and suggestions. + +# References + +Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2015. Neural machine translation by jointly learning to align and translate. In ICLR. +Kehai Chen, Rui Wang, Masao Utiyama, and Eiichiro Sumita. 2020. Content word aware neural machine translation. In ACL. +Jianpeng Cheng and Mirella Lapata. 2016. Neural summarization by extracting sentences and words. In ACL. +Shamil Chollampatt and Hwee Tou Ng. 2018. A multilayer convolutional encoder-decoder neural network for grammatical error correction. In AAAI. +Kenneth Church and Patrick Hanks. 1990. Word association norms, mutual information, and lexicography. CL. +Daniel Dahlmeier and Hwee Tou Ng. 2012. Better evaluation for grammatical error correction. In NAACL. +Liang Ding, Longyue Wang, Xuebo Liu, Derek F Wong, Dacheng Tao, and Zhaopeng Tu. 2021a. Progressive multi-granularity training for non-autoregressive translation. In Findings of ACL. +Liang Ding, Longyue Wang, Xuebo Liu, Derek F Wong, Dacheng Tao, and Zhaopeng Tu. 2021b. Rejuvenating low-frequency words: Making the most of parallel data in non-autoregressive translation. In ACL. +Liang Ding, Longyue Wang, Xuebo Liu, Derek F Wong, Dacheng Tao, and Zhaopeng Tu. 2021c. Understanding and improving lexical choice in non-autoregressive translation. In ICLR. +Liang Ding, Longyue Wang, Shuming Shi, Dacheng Tao, and Zhaopeng Tu. 2022. Redistributing low-frequency words: Making the most of monolingual data in non-autoregressive translation. In ACL. +Liang Ding, Longyue Wang, and Dacheng Tao. 2020. Self-attention with cross-lingual position representation. In ACL. +Shuhao Gu, Jinchao Zhang, Fandong Meng, Yang Feng, Wanying Xie, Jie Zhou, and Dong Yu. 2020. Token-level adaptive training for neural machine translation. In EMNLP. +Sangchul Hahn and Heeyoul Choi. 2019. Self-knowledge distillation in natural language processing. In RANLP. +Amr Hendy, Mohamed Abdelrehim, et al. 2023. How good are gpt models at machine translation? a comprehensive evaluation. arXiv preprint. +Wenxiang Jiao, Wenxuan Wang, Jen-tse Huang, Xing Wang, and Zhaopeng Tu. 2023. Is chatgpt a good translator? a preliminary study. arXiv preprint. + +Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In ACL. +Haoran Li and Wei Lu. 2021. Mixed cross entropy loss for neural machine translation. In ICML. +Zuchao Li, Rui Wang, et al. 2020. Data-dependent gaussian prior objective for language generation. In ICLR. +Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out. +Huan Lin, Baosong Yang, Liang Yao, Dayiheng Liu, Haibo Zhang, Jun Xie, Min Zhang, and Jinsong Su. 2022. Bridging the gap between training and inference: Multi-candidate optimization for diverse neural machine translation. In Findings of NAACL. +Xuebo Liu, Longyue Wang, Derek F Wong, Liang Ding, Lidia S Chao, Shuming Shi, and Zhaopeng Tu. 2021a. On the copying behaviors of pre-training for neural machine translation. In Findings of ACL. +Xuebo Liu, Longyue Wang, Derek F Wong, Liang Ding, Lidia S Chao, and Zhaopeng Tu. 2021b. Understanding and improving encoder layer fusion in sequence-to-sequence learning. In ICLR. +Qingyu Lu, Baopu Qiu, Liang Ding, Liping Xie, and Dacheng Tao. 2023. Error analysis prompting enables human-like translation evaluation in large language models: A case study on chatgpt. arXiv preprint. +Mengqi Miao, Fandong Meng, Yijin Liu, Xiao-Hua Zhou, and Jie Zhou. 2021. Prevent the language model from being overconfident in neural machine translation. In ACL. +Mathias Müller, Annette Rios, and Rico Sennrich. 2020. Domain robustness in neural machine translation. In AMTA, Virtual. +Graham Neubig, Zi-Yi Dou, Junjie Hu, Paul Michel, Danish Pruthi, and Xinyi Wang. 2019. compare-mt: A tool for holistic comparison of language generation systems. In NAACL. +Mohammad Norouzi, Samy Bengio, Zhifeng Chen, et al. 2016. Reward augmented maximum likelihood for neural structured prediction. In NeurIPS. +Myle Ott, Michael Auli, David Grangier, and Marc'Aurelio Ranzato. 2018. Analyzing uncertainty in neural machine translation. In ICML. +Myle Ott, Sergey Edunov, Alexei Baevski, Angela Fan, Sam Gross, Nathan Ng, David Grangier, and Michael Auli. 2019. *fairoseq: A fast, extensible toolkit for sequence modeling*. In *NAACL Demonstration*. + +Keqin Peng, Liang Ding, Qihuang Zhong, Li Shen, Xuebo Liu, Min Zhang, Yuanxin Ouyang, and Dacheng Tao. 2023. Towards making the most of chatgpt for machine translation. arxiv preprint. +Steven T Piantadosi. 2014. Zipf's word frequency law in natural language: A critical review and future directions. Psychonomic bulletin & review. +Matt Post. 2018. A call for clarity in reporting BLEU scores. In WMT. +Lynne M Reder, Xiaonan L Liu, Alexander Keinath, and Vencislav Popov. 2016. Building knowledge requires bricks, not sand: The critical role of familiar constituents in learning. Psychonomic bulletin & review. +Ricardo Rei, Craig Stewart, Ana C. Farinha, and Alon Lavie. 2020. COMET: A neural framework for MT evaluation. In EMNLP. +Thibault Sellam, Dipanjan Das, and Ankur P. Parikh. 2020. BLEURT: learning robust metrics for text generation. In ACL. +Zewei Sun, Shujian Huang, Hao-Ran Wei, Xinyu Dai, and Jiajun Chen. 2020. Generating diverse translation by manipulating multi-head attention. In AAAI. +Ilya Sutskever, Oriol Vinyals, and Quoc V. Le. 2014. Sequence to sequence learning with neural networks. In NeurIPS. +Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. 2016. Rethinking the inception architecture for computer vision. In CVPR. +Ashish Vaswani, Noam Shazeer, et al. 2017. Attention is all you need. In NeurIPS. +Yu Wan, Baosong Yang, et al. 2020. Self-paced learning for neural machine translation. In EMNLP. +Wenxuan Wang, Wenxiang Jiao, Yongchang Hao, Xing Wang, Shuming Shi, Zhaopeng Tu, and Michael R. Lyu. 2022. Understanding and improving sequence-to-sequence pretraining for neural machine translation. In ACL. +Di Wu, Liang Ding, Fan Lu, and Jian Xie. 2020. Slotrefine: A fast non-autoregressive model for joint intent detection and slot filling. In EMNLP. +Haoran Wu, Wenxuan Wang, Yuxuan Wan, Wenxiang Jiao, and Michael Lyu. 2023. Chatgpt or grammarly? evaluating chatgpt on grammatical error correction benchmark. arXiv preprint. +Fengshun Xiao, Yingting Wu, Hai Zhao, Rui Wang, and Shu Jiang. 2019. Dual skew divergence loss for neural machine translation. CoRR. +Yangyifan Xu, Yijin Liu, Fandong Meng, Jiajun Zhang, Jinan Xu, and Jie Zhou. 2021. Bilingual mutual information based adaptive training for neural machine translation. In ACL. + +Zheng Yuan and Ted Briscoe. 2016. Grammatical error correction using neural machine translation. In NAACL. +Changtong Zan, Liang Ding, Li Shen, Yu Cao, Weifeng Liu, and Dacheng Tao. 2022a. On the complementarity between pre-training and random-initialization for resource-rich machine translation. In COLING. +Changtong Zan, Keqin Peng, Liang Ding, et al. 2022b. Vega-mt: The jd explore academy machine translation system for wmt22. In WMT. +Runtian Zhai, Chen Dan, J Zico Kolter, and Pradeep Kumar Ravikumar. 2023. Understanding why generalized reweighting does not improve over ERM. In ICLR. +Songming Zhang, Yijin Liu, Fandong Meng, Yufeng Chen, Jinan Xu, Jian Liu, and Jie Zhou. 2022a. Conditional bilingual mutual information based adaptive training for neural machine translation. In ACL. +Zheng Zhang, Liang Ding, Dazhao Cheng, Xuebo Liu, Min Zhang, and Dacheng Tao. 2022b. Bliss: Robust sequence-to-sequence learning via self-supervised input representation. arXiv preprint. +Qihuang Zhong, Liang Ding, Juhua Liu, Bo Du, and Dacheng Tao. 2022. E2s2: Encoding-enhanced sequence-to-sequence pretraining for language understanding and generation. arXiv preprint. +Qihuang Zhong, Liang Ding, Keqin Peng, Juhua Liu, Bo Du, Li Shen, Yibing Zhan, and Dacheng Tao. 2023. Bag of tricks for effective language model pretraining and downstream adaptation: A case study on glue. arXiv preprint. + +# A Appendix + +Parameter Analysis on $\Gamma$ As stated in §2.1, we use the loss threshold $\Gamma$ to dynamically select the hard-to-learn tokens. Here, we analyze the influence of different $\Gamma$ in detail. In practice, we train the Transformer models with different $\Gamma$ (in $\{3,4,5,6\}$ ) and evaluate the performance of the WMT14 En-De test set. Table 8 lists the performance of different $\Gamma$ . The results of Table 8 show that SE is stable and insensitive to $\Gamma$ within a certain range. Noting that we select $\Gamma = 5$ for all experiment settings based on the results in Table 8. + +
Γ=3Γ=4Γ=5Γ=6
BLEU27.727.828.027.8
+ +Table 8: Parameter analysis of $\Gamma$ on WMT14 En-De. + +# Ablation Study + +Metric. In this work, we use the loss-based metric to dynamically select the hard-to-learn tokens. To validate the effectiveness of the metric, we use a simple adaptive training method ("+ ADD") that adds 1 to the weighting term of loss of the hard-to-learn tokens. The results on WMT16 EnRo are shown in Table 9, the simple Add method can achieve +0.3 BLEU improvement compared to the baseline model, which proves that our proposed self-questioning stage indeed mines informative difficult tokens. Also, we can observe that learning these dynamic difficult tokens with our SE framework ("+ SE") could outperform "+" ADD" by +0.6 BLUE points, demonstrating the superiority of our token-specific label smoothing approach. + +
Baseline+ ADD+ SE
BLEU35.135.436.0
+ +Learning objective. As stated in §2.1, our learning objective is the combination of the ground truth and the model's prediction. To validate the effectiveness of predicted distribution, we conduct ablation experiments on WMT16 En-Ro and WMT14 En-De. The results in Table 10 show that adding the predicted distribution will consistently improve the model's performance, which proves the effectiveness of the predicted distribution. + +Table 9: Ablation performance of our SE. on Metric. + +
MethodBLEU
EN⇒DEEN⇒Ro
Transformer27.0835.11
SE28.0236.02
-w/o predicted results27.8935.71
+ +Table 10: Ablation performance of our SE. on learning objective. + +A For every submission: + +A1. Did you describe the limitations of your work? + +The last section of the paper. + +A2. Did you discuss any potential risks of your work? + +Left blank. + +A3. Do the abstract and introduction summarize the paper's main claims? + +The abstract and the introduction section. + +A4. Have you used AI writing assistants when working on this paper? + +Left blank. + +B Did you use or create scientific artifacts? + +Left blank. + +B1. Did you cite the creators of artifacts you used? + +No response. + +B2. Did you discuss the license or terms for use and / or distribution of any artifacts? + +No response. + +B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? + +No response. + +B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? + +No response. + +B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? + +No response. + +B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. + +No response. + +C Did you run computational experiments? + +Left blank. + +C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? + +No response. + +The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance. + +C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? No response. +C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? No response. +C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? No response. + +# D Did you use human annotators (e.g., crowdworkers) or research with human participants? + +Section 3.2 + +D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Left blank. +D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Left blank. +D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Left blank. +D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Left blank. +D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Left blank. \ No newline at end of file diff --git a/2023/Token-Level Self-Evolution Training for Sequence-to-Sequence Learning/images.zip b/2023/Token-Level Self-Evolution Training for Sequence-to-Sequence Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..36dff0240889f4d4986ec22e3be861030f45cde9 --- /dev/null +++ b/2023/Token-Level Self-Evolution Training for Sequence-to-Sequence Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9f3af8e31a3d3aeaf0a7c6482231e1dfd76e7dc2ddbd2b6a6fc3c2824c12afb5 +size 259860 diff --git a/2023/Token-Level Self-Evolution Training for Sequence-to-Sequence Learning/layout.json b/2023/Token-Level Self-Evolution Training for Sequence-to-Sequence Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d6a1c6df7972644f777a58705c80b429d7f8688c --- /dev/null +++ b/2023/Token-Level Self-Evolution Training for Sequence-to-Sequence Learning/layout.json @@ -0,0 +1,7925 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 73, + 75, + 521, + 95 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 75, + 521, + 95 + ], + "spans": [ + { + "bbox": [ + 73, + 75, + 521, + 95 + ], + "type": "text", + "content": "Token-Level Self-Evolution Training for Sequence-to-Sequence Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 136, + 110, + 460, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 110, + 460, + 138 + ], + "spans": [ + { + "bbox": [ + 136, + 110, + 460, + 138 + ], + "type": "text", + "content": "Keqin Peng" + }, + { + "bbox": [ + 136, + 110, + 460, + 138 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 136, + 110, + 460, + 138 + ], + "type": "text", + "content": ", Liang Ding" + }, + { + "bbox": [ + 136, + 110, + 460, + 138 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 136, + 110, + 460, + 138 + ], + "type": "text", + "content": ", Qihuang Zhong" + }, + { + "bbox": [ + 136, + 110, + 460, + 138 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 136, + 110, + 460, + 138 + ], + "type": "text", + "content": " \nYuanxin Ouyang" + }, + { + "bbox": [ + 136, + 110, + 460, + 138 + ], + "type": "inline_equation", + "content": "^{1†}" + }, + { + "bbox": [ + 136, + 110, + 460, + 138 + ], + "type": "text", + "content": ", Wenge Rong" + }, + { + "bbox": [ + 136, + 110, + 460, + 138 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 136, + 110, + 460, + 138 + ], + "type": "text", + "content": ", Zhang Xiong" + }, + { + "bbox": [ + 136, + 110, + 460, + 138 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 136, + 110, + 460, + 138 + ], + "type": "text", + "content": ", Dacheng Tao" + }, + { + "bbox": [ + 136, + 110, + 460, + 138 + ], + "type": "inline_equation", + "content": "^{4}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 73, + 139, + 521, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 139, + 521, + 182 + ], + "spans": [ + { + "bbox": [ + 73, + 139, + 521, + 182 + ], + "type": "text", + "content": "1Beihang University 2Zhejiang University 3Wuhan University 4The University of Sydney {keqin.peng, oyyx, w.rong, xiongz}@buaa.edu.cn zhongqihuang@whu.edu.cn, {liangding.liam, dacheng.tao}@gmail.com" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 155, + 212, + 204, + 226 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 212, + 204, + 226 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 204, + 226 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 84, + 236, + 274, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 236, + 274, + 535 + ], + "spans": [ + { + "bbox": [ + 84, + 236, + 274, + 535 + ], + "type": "text", + "content": "Adaptive training approaches, widely used in sequence-to-sequence models, commonly reweigh the losses of different target tokens based on priors, e.g. word frequency. However, most of them do not consider the variation of learning difficulty in different training steps, and overly emphasize the learning of difficult one-hot labels, making the learning deterministic and sub-optimal. In response, we present Token-Level Self-Evolution Training (SE), a simple and effective dynamic training method to fully and wisely exploit the knowledge from data. SE focuses on dynamically learning the under-explored tokens for each forward pass and adaptively regularizes the training by introducing a novel token-specific label smoothing approach. Empirically, SE yields consistent and significant improvements in three tasks, i.e. machine translation, summarization, and grammatical error correction. Encouragingly, we achieve averaging " + }, + { + "bbox": [ + 84, + 236, + 274, + 535 + ], + "type": "inline_equation", + "content": "+0.93" + }, + { + "bbox": [ + 84, + 236, + 274, + 535 + ], + "type": "text", + "content": " BLEU improvement on three machine translation tasks. Analyses confirm that, besides improving lexical accuracy, SE enhances generation diversity and model generalization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 545, + 155, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 545, + 155, + 559 + ], + "spans": [ + { + "bbox": [ + 68, + 545, + 155, + 559 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 567, + 291, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 567, + 291, + 688 + ], + "spans": [ + { + "bbox": [ + 67, + 567, + 291, + 688 + ], + "type": "text", + "content": "Sequence-to-sequence learning (Seq2Seq) with neural networks (Sutskever et al., 2014) has advanced the state-of-the-art in various NLP tasks, e.g. translation (Bahdanau et al., 2015; Vaswani et al., 2017), summarization (Cheng and Lapata, 2016), and grammatical error correction (Yuan and Briscoe, 2016). Generally, Seq2Seq models are trained with the cross-entropy loss, which equally weighs the training losses of different target tokens." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 689, + 291, + 743 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 689, + 291, + 743 + ], + "spans": [ + { + "bbox": [ + 67, + 689, + 291, + 743 + ], + "type": "text", + "content": "However, due to the token imbalance nature (Piantadosi, 2014) and the truth that different tokens contribute differently to the sentence meaning (Church and Hanks, 1990; Chen et al., 2020)," + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 305, + 210, + 524, + 309 + ], + "blocks": [ + { + "bbox": [ + 305, + 210, + 524, + 309 + ], + "lines": [ + { + "bbox": [ + 305, + 210, + 524, + 309 + ], + "spans": [ + { + "bbox": [ + 305, + 210, + 524, + 309 + ], + "type": "image", + "image_path": "43bb068707a45de0c88ee7d411df1b2bc2378d00ca71a2a2df7381da1549eb90.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 313, + 526, + 361 + ], + "lines": [ + { + "bbox": [ + 302, + 313, + 526, + 361 + ], + "spans": [ + { + "bbox": [ + 302, + 313, + 526, + 361 + ], + "type": "text", + "content": "Figure 1: An example to illustrate the changing token difficulties in different training steps in WMT'14 En-De. The token \"abschreiben/ Sache\" is hard/ easy to learn at 50K while the trend is totally reversed at 100K." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 391, + 526, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 391, + 526, + 539 + ], + "spans": [ + { + "bbox": [ + 302, + 391, + 526, + 539 + ], + "type": "text", + "content": "several works are developed to reweigh the token-level training loss according to explicit (e.g. frequency) or implicit (uncertainty estimated by off-the-shelf language models) priors (Gu et al., 2020; Xu et al., 2021; Zhang et al., 2022a). For example, Gu et al. (2020) proposed two heuristic criteria based on word frequency to encourage the model to learn from larger-weight low-frequency tokens. Zhang et al. (2022a) introduce target-context-aware metric based on an additional target-side language model to adjust the weight of each target token." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 543, + 526, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 543, + 526, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 543, + 526, + 773 + ], + "type": "text", + "content": "Despite some success, there are still limitations in these adaptive training approaches. First, most of them predetermine the difficult tokens and fix such prior to guiding the training. However, in our preliminary study, we find the hard-to-learn tokens are dynamically changing during training, rather than statically fixed. As shown in Figure 1, as the training progress goes, although the sentence-level loss is nicely converging, the difficult token is changing from \"abschreiben\" to \"Sache\" in terms of the token-level loss. Second, these adaptive training methods overly emphasize fitting the difficult tokens' one-hot labels by reweighing the loss, which empirically may cause overfitting and limit the generalization (Norouzi et al., 2016; Szegedy et al., 2016; Xiao et al., 2019; Miao et al., 2021). Also, a more recent study (Zhai et al., 2023) provides" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 83, + 750, + 224, + 761 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 750, + 224, + 761 + ], + "spans": [ + { + "bbox": [ + 83, + 750, + 224, + 761 + ], + "type": "text", + "content": "* Keqin and Liang contributed equally." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 83, + 761, + 175, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 761, + 175, + 772 + ], + "spans": [ + { + "bbox": [ + 83, + 761, + 175, + 772 + ], + "type": "text", + "content": "† Corresponding Author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 288, + 780, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 288, + 780, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 288, + 780, + 307, + 791 + ], + "type": "text", + "content": "841" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 135, + 795, + 459, + 805 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 795, + 459, + 805 + ], + "spans": [ + { + "bbox": [ + 135, + 795, + 459, + 805 + ], + "type": "text", + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 224, + 806, + 370, + 817 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 806, + 370, + 817 + ], + "spans": [ + { + "bbox": [ + 224, + 806, + 370, + 817 + ], + "type": "text", + "content": "Volume 2: Short Papers, pages 841-850" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "spans": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "text", + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 289, + 98 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 289, + 98 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 289, + 98 + ], + "type": "text", + "content": "theoretical evidence to support that reweighting is not that effective to improve the generalization." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 99, + 290, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 99, + 290, + 287 + ], + "spans": [ + { + "bbox": [ + 67, + 99, + 290, + 287 + ], + "type": "text", + "content": "Correspondingly, we design a simple and effective Token-Level Self-Evolution Training (SE) strategy to encourage Seq2Seq models to learn from difficult words that are dynamically selected by the model itself. Specifically, SE contains two stages: 1self-questioning and 2self-evolution training. In the first stage, the Seq2Seq models dynamically select the hard-to-learn tokens based on the token-level losses, then we encourage the Seq2Seq models to learn from them in the second stage, where, rather than adopting reweighing, we introduce a novel token-specific label smoothing approach to generate easily digestible soft label, which considers both the ground truth and model's prediction." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 289, + 291, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 289, + 291, + 383 + ], + "spans": [ + { + "bbox": [ + 67, + 289, + 291, + 383 + ], + "type": "text", + "content": "Experiments across tasks, language pairs, data scales, and model sizes show that SE consistently and significantly outperforms both the vanilla Seq2Seq model and the re-implemented advanced baselines. Analyses confirm that besides improved lexical accuracy, SE generates diverse and human-like generations with better model generalization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 394, + 157, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 394, + 157, + 409 + ], + "spans": [ + { + "bbox": [ + 67, + 394, + 157, + 409 + ], + "type": "text", + "content": "2 Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 417, + 291, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 417, + 291, + 485 + ], + "spans": [ + { + "bbox": [ + 67, + 417, + 291, + 485 + ], + "type": "text", + "content": "Preliminary Sequence-to-sequence (Seq2Seq) learning aims to maximize the cross-entropy (CE) loss of the log-likelihood of each target word in " + }, + { + "bbox": [ + 67, + 417, + 291, + 485 + ], + "type": "inline_equation", + "content": "\\mathbf{y} = \\{y_1,\\dots ,y_N\\}" + }, + { + "bbox": [ + 67, + 417, + 291, + 485 + ], + "type": "text", + "content": ", conditioned on source " + }, + { + "bbox": [ + 67, + 417, + 291, + 485 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 67, + 417, + 291, + 485 + ], + "type": "text", + "content": ", where the optimization treats all tokens equally:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 98, + 494, + 290, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 494, + 290, + 533 + ], + "spans": [ + { + "bbox": [ + 98, + 494, + 290, + 533 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C E}} (\\theta) = - \\sum_ {j = 1} ^ {N} \\log p \\left(y _ {j} \\mid \\mathbf {y} _ {< j}, \\mathbf {x}; \\theta\\right) \\tag {1}", + "image_path": "eef9eee565614e28a67670ce41a10a38cb048c2670a0b8b4bc9bf6154a1e0622.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 543, + 290, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 543, + 290, + 663 + ], + "spans": [ + { + "bbox": [ + 67, + 543, + 290, + 663 + ], + "type": "text", + "content": "However, due to the different learning difficulties of each token, it is sub-optimal to treat all tokens equally (Gu et al., 2020). To address this limitation, a series of token-level adaptive training objectives were adopted to re-weight the losses of different target tokens (Xu et al., 2021; Zhang et al., 2022a). The common goal of these methods is to facilitate the model training by fully exploiting the informative but underexplored tokens." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 665, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 665, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 665, + 291, + 772 + ], + "type": "text", + "content": "However, our preliminary study shows that the hard tokens are dynamically changing (see Figure 1) in different training steps (or model structures), thus it is sub-optimal to employ static token priors (e.g. frequency) during training. Also, recent studies (Zhai et al., 2023) in the ML community theoretically show that reweighting is not that effective to improve the generalization. Based on the above" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 71, + 526, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 220 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 220 + ], + "type": "text", + "content": "evidence, we present the self-evolution learning (SE) mechanism to encourage the model to adaptively and wisely learn from the informative yet under-explored tokens dynamically determined by the model itself (Stage① in §2.1), with an easy-to-learn label distribution (Stage② in §2.1). A similar work to ours is Hahn and Choi (2019). However, their method mainly considers the situation where the predicted answer is incorrect but close to the golden answer, while our method focuses on all dynamic hard tokens." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 231, + 503, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 231, + 503, + 243 + ], + "spans": [ + { + "bbox": [ + 302, + 231, + 503, + 243 + ], + "type": "text", + "content": "2.1 Token-Level Self-Evolution Learning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 249, + 526, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 249, + 526, + 481 + ], + "spans": [ + { + "bbox": [ + 302, + 249, + 526, + 481 + ], + "type": "text", + "content": "1 Self-questioning Stage. The goal is to select the hard-to-learn tokens that are questioned by the Seq2Seq model itself during training dynamics. Previously, these difficult tokens are predetermined by external models or specific statistical metrics. However, inspired by the finding of dynamic change of difficult tokens during the training stage as shown in Figure 1 and the finding that the trained model contains useful information (Li and Lu, 2021), e.g. synonym, we propose to straightforwardly leverage the behavior of the model to dynamically select target tokens. In practice, we first calculate the token-level CE loss, denoted as " + }, + { + "bbox": [ + 302, + 249, + 526, + 481 + ], + "type": "inline_equation", + "content": "\\{l_1, l_2, \\dots, l_n\\}" + }, + { + "bbox": [ + 302, + 249, + 526, + 481 + ], + "type": "text", + "content": ", for each token for each forward pass. Then we set a loss threshold " + }, + { + "bbox": [ + 302, + 249, + 526, + 481 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 302, + 249, + 526, + 481 + ], + "type": "text", + "content": " and select the tokens whose losses exceed " + }, + { + "bbox": [ + 302, + 249, + 526, + 481 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 302, + 249, + 526, + 481 + ], + "type": "text", + "content": " as the target tokens, i.e., " + }, + { + "bbox": [ + 302, + 249, + 526, + 481 + ], + "type": "inline_equation", + "content": "D = \\{t_i | l_i > \\Gamma\\}" + }, + { + "bbox": [ + 302, + 249, + 526, + 481 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 302, + 249, + 526, + 481 + ], + "type": "inline_equation", + "content": "i \\in N = \\{1, 2, \\dots, n\\}" + }, + { + "bbox": [ + 302, + 249, + 526, + 481 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 489, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 489, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 489, + 525, + 772 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 302, + 489, + 525, + 772 + ], + "type": "text", + "content": " Self-evolution Training Stage. After selecting the difficult tokens, we encourage the model to carefully learn from them. Given the theoretical shortage (Zhai et al., 2023) and potentially caused overfitting or overconfidence problem (Miao et al., 2021) of reweighting and deliberately learning from difficult tokens, we propose to strengthen the learning from these tokens with a newly designed Token-specific Label Smoothing (TLS) approach. Specifically, motivated by the effect of label smoothing (LS) regularization (Szegedy et al., 2016), we combine the ground truth " + }, + { + "bbox": [ + 302, + 489, + 525, + 772 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 302, + 489, + 525, + 772 + ], + "type": "text", + "content": " and the model's prediction " + }, + { + "bbox": [ + 302, + 489, + 525, + 772 + ], + "type": "inline_equation", + "content": "\\hat{p}_i" + }, + { + "bbox": [ + 302, + 489, + 525, + 772 + ], + "type": "text", + "content": " to form a new soft label " + }, + { + "bbox": [ + 302, + 489, + 525, + 772 + ], + "type": "inline_equation", + "content": "\\widetilde{p}_i" + }, + { + "bbox": [ + 302, + 489, + 525, + 772 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 302, + 489, + 525, + 772 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 302, + 489, + 525, + 772 + ], + "type": "text", + "content": "-th token. Then we use " + }, + { + "bbox": [ + 302, + 489, + 525, + 772 + ], + "type": "inline_equation", + "content": "\\widetilde{p}" + }, + { + "bbox": [ + 302, + 489, + 525, + 772 + ], + "type": "text", + "content": " to guide the difficult tokens " + }, + { + "bbox": [ + 302, + 489, + 525, + 772 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 302, + 489, + 525, + 772 + ], + "type": "text", + "content": ", while leaving label-smoothing CE loss for the other tokens. It is worth noting that we also apply the traditional label smoothing technique to " + }, + { + "bbox": [ + 302, + 489, + 525, + 772 + ], + "type": "inline_equation", + "content": "\\hat{p}_i" + }, + { + "bbox": [ + 302, + 489, + 525, + 772 + ], + "type": "text", + "content": " to activate the information in the predicted distribution. Analogous to human learning, it is often easier for humans to grasp new things described by their familiar knowledge (Reder et al., 2016)," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "842" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 71, + 69, + 524, + 211 + ], + "blocks": [ + { + "bbox": [ + 71, + 69, + 524, + 211 + ], + "lines": [ + { + "bbox": [ + 71, + 69, + 524, + 211 + ], + "spans": [ + { + "bbox": [ + 71, + 69, + 524, + 211 + ], + "type": "table", + "html": "
ModelWMT16 En→RoWMT14 En→DeWMT14 En→Fr
Transformer (Vaswani et al., 2017)35.1127.0840.65
+ Freq-Exponential (Gu et al., 2020)35.86 (+0.75)27.60 (+0.52)41.05 (+0.40)
+ Freq-Chi-Square (Gu et al., 2020)35.74 (+0.63)27.51 (+0.43)40.99 (+0.34)
+ D2GPo (Li et al., 2020)35.89 (+0.78)27.66 (+0.58)41.05 (+0.40)
+ BMI-adaptive (Xu et al., 2021)35.89 (+0.78)27.65 (+0.57)41.10 (+0.45)
+ MixCrossEntropy (Li and Lu, 2021)35.88 (+0.74)27.61 (+0.53)41.07 (+0.42)
+ CBMI-adaptive (Zhang et al., 2022a)35.90 (+0.79)27.69 (+0.61)41.13 (+0.48)
+ SPL (Wan et al., 2020)35.92 (+0.81)27.88 (+0.80)41.30 (+0.65)
+ Self-Evolution (ours)36.02 (+0.91)†28.02 (+0.94)†41.60 (+0.95)†
", + "image_path": "db7657c51f5d9b7bfc419ca03d3ac3fb73897200d49df934553d43ca66b80f82.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 70, + 264, + 289, + 317 + ], + "blocks": [ + { + "bbox": [ + 67, + 221, + 526, + 247 + ], + "lines": [ + { + "bbox": [ + 67, + 221, + 526, + 247 + ], + "spans": [ + { + "bbox": [ + 67, + 221, + 526, + 247 + ], + "type": "text", + "content": "Table 1: BLEU scores " + }, + { + "bbox": [ + 67, + 221, + 526, + 247 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 67, + 221, + 526, + 247 + ], + "type": "text", + "content": " on three translation tasks spanning different data scales, i.e. " + }, + { + "bbox": [ + 67, + 221, + 526, + 247 + ], + "type": "inline_equation", + "content": "0.6\\mathrm{M}" + }, + { + "bbox": [ + 67, + 221, + 526, + 247 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 67, + 221, + 526, + 247 + ], + "type": "inline_equation", + "content": "4.5\\mathrm{M}" + }, + { + "bbox": [ + 67, + 221, + 526, + 247 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 67, + 221, + 526, + 247 + ], + "type": "inline_equation", + "content": "36\\mathrm{M}" + }, + { + "bbox": [ + 67, + 221, + 526, + 247 + ], + "type": "text", + "content": ". “†” indicates a statistically significant difference from the powerful Transformer baseline " + }, + { + "bbox": [ + 67, + 221, + 526, + 247 + ], + "type": "inline_equation", + "content": "(p < 0.05)" + }, + { + "bbox": [ + 67, + 221, + 526, + 247 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 264, + 289, + 317 + ], + "lines": [ + { + "bbox": [ + 70, + 264, + 289, + 317 + ], + "spans": [ + { + "bbox": [ + 70, + 264, + 289, + 317 + ], + "type": "table", + "html": "
Ro-EnXSUMGEC
BLEURG-1RG-2RG-LPrec.RecallF0.5
Baseline37.343.219.834.059.139.853.9
+ SE37.7†43.820.434.7†58.946.255.8†
", + "image_path": "37bb55b56009a665d91b1c75bed5dd5acdfe40f7f5fc607b9c7939cc24a20c2c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 325, + 291, + 361 + ], + "lines": [ + { + "bbox": [ + 67, + 325, + 291, + 361 + ], + "spans": [ + { + "bbox": [ + 67, + 325, + 291, + 361 + ], + "type": "text", + "content": "Table 2: Performance on more tasks including translation, summarization, and grammar error correction, upon larger model BART (Lewis et al., 2020)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 383, + 291, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 383, + 291, + 437 + ], + "spans": [ + { + "bbox": [ + 67, + 383, + 291, + 437 + ], + "type": "text", + "content": "therefore the new soft label fused both accurate ground truth and model's self-distribution is easily digestible. Mathematically, for difficult tokens " + }, + { + "bbox": [ + 67, + 383, + 291, + 437 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 67, + 383, + 291, + 437 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 67, + 383, + 291, + 437 + ], + "type": "inline_equation", + "content": "\\widetilde{p}_i" + }, + { + "bbox": [ + 67, + 383, + 291, + 437 + ], + "type": "text", + "content": " is formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 448, + 289, + 463 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 448, + 289, + 463 + ], + "spans": [ + { + "bbox": [ + 138, + 448, + 289, + 463 + ], + "type": "interline_equation", + "content": "\\widetilde {p _ {i}} = \\left(p _ {i} + \\hat {p _ {i}}\\right) / 2. \\tag {2}", + "image_path": "e2ce87ba3a2dd50c5cc40c9ed225d490ba66782caea367fc058e22afca9fb803.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 473, + 290, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 473, + 290, + 500 + ], + "spans": [ + { + "bbox": [ + 67, + 473, + 290, + 500 + ], + "type": "text", + "content": "Then we calculate the losses of difficult tokens and the others, and combine the two losses:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 76, + 512, + 290, + 542 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 512, + 290, + 542 + ], + "spans": [ + { + "bbox": [ + 76, + 512, + 290, + 542 + ], + "type": "interline_equation", + "content": "L = - \\left(\\sum_ {i} \\widetilde {p _ {i}} \\cdot \\log \\left(\\hat {p _ {i}}\\right) + \\sum_ {j} p _ {j} \\cdot \\log \\left(\\hat {p _ {j}}\\right)\\right), \\tag {3}", + "image_path": "847465342969b8afbdc86fd5ce7b5fdef44922ab524e6c6900fbcbdd385bfa8f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 551, + 198, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 551, + 198, + 565 + ], + "spans": [ + { + "bbox": [ + 67, + 551, + 198, + 565 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 67, + 551, + 198, + 565 + ], + "type": "inline_equation", + "content": "i\\in D" + }, + { + "bbox": [ + 67, + 551, + 198, + 565 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 551, + 198, + 565 + ], + "type": "inline_equation", + "content": "j\\in N\\setminus D" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 575, + 145, + 587 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 575, + 145, + 587 + ], + "spans": [ + { + "bbox": [ + 67, + 575, + 145, + 587 + ], + "type": "text", + "content": "3 Evaluation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 597, + 292, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 597, + 292, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 597, + 292, + 773 + ], + "type": "text", + "content": "Machine Translation on three widely-used benchmarks (Ding et al., 2020, 2021c, 2022): small-scale WMT16 English-Romanian (En-Ro; 0.6M), medium-scale WMT14 English-German (En-De; 4.5M), and large-scale WMT14 English-French (En-Fr; 36.0M). We implement the baselines and our approach under Transformer-base settings. We follow the previous adaptive training approach (Gu et al., 2020) to pretrain with the cross-entropy loss with " + }, + { + "bbox": [ + 67, + 597, + 292, + 773 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 67, + 597, + 292, + 773 + ], + "type": "text", + "content": " steps, and further finetune the same steps with different adaptive training objectives, including Freq-Exponential (Gu et al., 2020), Freq-Chi-Square (Gu et al., 2020), D2GPo (Li et al., 2020)," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 266, + 526, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 266, + 526, + 509 + ], + "spans": [ + { + "bbox": [ + 302, + 266, + 526, + 509 + ], + "type": "text", + "content": "BMI-adaptive (Xu et al., 2021), MixCrossEntropy (Li and Lu, 2021), CBMI-adaptive (Zhang et al., 2022a), and SPL (Wan et al., 2020). For " + }, + { + "bbox": [ + 302, + 266, + 526, + 509 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 302, + 266, + 526, + 509 + ], + "type": "text", + "content": ", we adopt 100K and 30K for larger datasets, e.g. En-De and En-Fr, and small dataset, i.e. En-Ro, respectively. We empirically adopt 32K tokens per batch for large datasets, the learning rate warms up to 1e-7 for 10K steps, and then decays 90K, while for small dataset En-Ro, The learning rate warms up to 1e-7 for 4K steps, and then decays 26K steps. All the experiments are conducted on 4 NVIDIA Tesla A100 GPUs. The SacreBLEU (Post, 2018) was used for evaluation. Besides translation, we also follow previous works (Liu et al., 2021b; Zhong et al., 2022; Zhang et al., 2022b) to validate the universality of our method on more sequence-to-sequence learning tasks, e.g., summarization and grammatical error correction." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 511, + 525, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 511, + 525, + 578 + ], + "spans": [ + { + "bbox": [ + 302, + 511, + 525, + 578 + ], + "type": "text", + "content": "Text Summarization on XSUM corpus (0.2M). We follow fairseq (Ott et al., 2019) to preprocess the data and train the model, then finetune them for the same steps. We evaluated with the ROUGE (Lin, 2004), i.e. R-1, R-2, and R-L." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 579, + 525, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 579, + 525, + 661 + ], + "spans": [ + { + "bbox": [ + 302, + 579, + 525, + 661 + ], + "type": "text", + "content": "Grammatical Error Correction on CoNLL14 (1.4M). We follow Chollampatt and Ng (2018) to preprocess the data and train the model, then finetune them for the same steps. The MaxMatch " + }, + { + "bbox": [ + 302, + 579, + 525, + 661 + ], + "type": "inline_equation", + "content": "(\\mathbf{M}^2)" + }, + { + "bbox": [ + 302, + 579, + 525, + 661 + ], + "type": "text", + "content": " scores (Dahlmeier and Ng, 2012) were used for evaluation with precision, recall, and " + }, + { + "bbox": [ + 302, + 579, + 525, + 661 + ], + "type": "inline_equation", + "content": "\\mathrm{F_{0.5}}" + }, + { + "bbox": [ + 302, + 579, + 525, + 661 + ], + "type": "text", + "content": " values." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 673, + 393, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 673, + 393, + 685 + ], + "spans": [ + { + "bbox": [ + 302, + 673, + 393, + 685 + ], + "type": "text", + "content": "3.1 Main Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 692, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 692, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 692, + 526, + 772 + ], + "type": "text", + "content": "SE brings gains across language pairs and scales. Results on machine translation across different data sizes ranging from 0.6M to 36M in Table 1 show that our SE-equipped Transformer “+ Self-Evolution (ours)” 1) considerably improves the performance by averaging +0.92 BLEU points; 2) out" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "843" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 68, + 278, + 137 + ], + "blocks": [ + { + "bbox": [ + 82, + 68, + 278, + 137 + ], + "lines": [ + { + "bbox": [ + 82, + 68, + 278, + 137 + ], + "spans": [ + { + "bbox": [ + 82, + 68, + 278, + 137 + ], + "type": "table", + "html": "
Valid Loss Scale
0-11-22-3>3
Transformer + SE63.310.56.719.5
65.69.55.819.1
", + "image_path": "006ecc2dac41eeeaf56e1bd722509822d02186c3e16235730351f19ca8ce05d4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 78, + 192, + 285, + 260 + ], + "blocks": [ + { + "bbox": [ + 67, + 144, + 291, + 184 + ], + "lines": [ + { + "bbox": [ + 67, + 144, + 291, + 184 + ], + "spans": [ + { + "bbox": [ + 67, + 144, + 291, + 184 + ], + "type": "text", + "content": "Table 3: The token distribution " + }, + { + "bbox": [ + 67, + 144, + 291, + 184 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 67, + 144, + 291, + 184 + ], + "type": "text", + "content": " on different loss scales. Shadowed areas mean accurate token prediction estimated with lower cross-entropy loss, i.e. \"0-1\"." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 78, + 192, + 285, + 260 + ], + "lines": [ + { + "bbox": [ + 78, + 192, + 285, + 260 + ], + "spans": [ + { + "bbox": [ + 78, + 192, + 285, + 260 + ], + "type": "table", + "html": "
MethodWMT22 De⇒En
BLEUΔCOMETΔ
Transformer29.98-45.1
+SE30.38+0.446.3+1.2
", + "image_path": "0e148154138c0851214990282105ac370d251e6072edeb27c5e01a3da4c29c97.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 267, + 290, + 293 + ], + "lines": [ + { + "bbox": [ + 67, + 267, + 290, + 293 + ], + "spans": [ + { + "bbox": [ + 67, + 267, + 290, + 293 + ], + "type": "text", + "content": "Table 4: Performance on extremely large dataset WMT22 De-En (236M)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 314, + 291, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 314, + 291, + 370 + ], + "spans": [ + { + "bbox": [ + 67, + 314, + 291, + 370 + ], + "type": "text", + "content": "performs previous competitive method “+ CBMI-adaptive” by up to +0.47 BLEU points on large dataset WMT14 En-Fr. These results demonstrate the effectiveness and universality of our SE." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 376, + 291, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 376, + 291, + 513 + ], + "spans": [ + { + "bbox": [ + 67, + 376, + 291, + 513 + ], + "type": "text", + "content": "SE brings gains across tasks and backbone sizes. Table 2 lists the performance on more tasks, including translation, summarization, and grammar error correction, upon large pretrained backbone - BART (Lewis et al., 2020), which has above 600M parameters. Compared to a stronger baseline, our SE significantly and incrementally improves the generation quality in all tasks, i.e. " + }, + { + "bbox": [ + 67, + 376, + 291, + 513 + ], + "type": "inline_equation", + "content": "+0.4" + }, + { + "bbox": [ + 67, + 376, + 291, + 513 + ], + "type": "text", + "content": " BLEU, " + }, + { + "bbox": [ + 67, + 376, + 291, + 513 + ], + "type": "inline_equation", + "content": "+0.7" + }, + { + "bbox": [ + 67, + 376, + 291, + 513 + ], + "type": "text", + "content": " RG-L, and " + }, + { + "bbox": [ + 67, + 376, + 291, + 513 + ], + "type": "inline_equation", + "content": "+1.9" + }, + { + "bbox": [ + 67, + 376, + 291, + 513 + ], + "type": "inline_equation", + "content": "\\mathrm{F_{0.5}}" + }, + { + "bbox": [ + 67, + 376, + 291, + 513 + ], + "type": "text", + "content": ", respectively, showing our SE is robustly applicable to general scenarios." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 520, + 291, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 520, + 291, + 643 + ], + "spans": [ + { + "bbox": [ + 67, + 520, + 291, + 643 + ], + "type": "text", + "content": "SE works well on extremely large dataset. To further verify the effectiveness of SE on extremely large dataset, we conducted an experiment on WMT22 De-En processed by Zan et al. (2022b), which contains 236M training examples. The results in Table 4 show that our method can achieve " + }, + { + "bbox": [ + 67, + 520, + 291, + 643 + ], + "type": "inline_equation", + "content": "+0.4" + }, + { + "bbox": [ + 67, + 520, + 291, + 643 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 520, + 291, + 643 + ], + "type": "inline_equation", + "content": "+1.2" + }, + { + "bbox": [ + 67, + 520, + 291, + 643 + ], + "type": "text", + "content": " improvement in BLEU and COMET respectively, which proves that our SE also works on extremely large datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 652, + 136, + 665 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 652, + 136, + 665 + ], + "spans": [ + { + "bbox": [ + 67, + 652, + 136, + 665 + ], + "type": "text", + "content": "3.2 Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 670, + 291, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 670, + 291, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 670, + 291, + 712 + ], + "type": "text", + "content": "We provide some insights to better understand the effectiveness of our approach. The ablation of important modules and parameters is in Appendix A." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 719, + 291, + 774 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 719, + 291, + 774 + ], + "spans": [ + { + "bbox": [ + 67, + 719, + 291, + 774 + ], + "type": "text", + "content": "SE learns better token representation. To verify whether our method helps learn better tokens representation, we conduct analysis on WMT14 EnDe from learning loss and fine-grained generation" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 303, + 68, + 525, + 139 + ], + "blocks": [ + { + "bbox": [ + 303, + 68, + 525, + 139 + ], + "lines": [ + { + "bbox": [ + 303, + 68, + 525, + 139 + ], + "spans": [ + { + "bbox": [ + 303, + 68, + 525, + 139 + ], + "type": "image", + "image_path": "4077ecf43a4abda12283cba9c9e154f7c27f83d31e008deae2104732133fd8d5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 146, + 525, + 171 + ], + "lines": [ + { + "bbox": [ + 302, + 146, + 525, + 171 + ], + "spans": [ + { + "bbox": [ + 302, + 146, + 525, + 171 + ], + "type": "text", + "content": "Figure 2: Fine-grained translation quality across word frequencies and sentence lengths." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 194, + 420, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 194, + 420, + 206 + ], + "spans": [ + { + "bbox": [ + 302, + 194, + 420, + 206 + ], + "type": "text", + "content": "perspectives, respectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 206, + 526, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 206, + 526, + 477 + ], + "spans": [ + { + "bbox": [ + 302, + 206, + 526, + 477 + ], + "type": "text", + "content": "First, we count the token ratios distributed in different cross-entropy loss scales in Table 3 following Zan et al. (2022a). Cross-entropy is a good indicator to quantify the distance between the predicted distribution and the ground truth in the valid dataset, and a lower value means a more similar distribution. As shown, our method improves the low-loss token ratios by " + }, + { + "bbox": [ + 302, + 206, + 526, + 477 + ], + "type": "inline_equation", + "content": "+2.3\\%" + }, + { + "bbox": [ + 302, + 206, + 526, + 477 + ], + "type": "text", + "content": ", indicating SE helps the model learn better token representations by reducing the token uncertainty. In addition, we follow Ding et al. (2021a); Liu et al. (2021a) to break the translation down into different granularities and measure their fined-grained performance. In particular, we calculate1 the F-measure of words by different frequency buckets and BLEU scores of buckets of different lengths in Figure 2. We see SE achieves better performance in all frequencies and sentence buckets, demonstrating our method can improve the performance of different granularities." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 486, + 525, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 486, + 525, + 689 + ], + "spans": [ + { + "bbox": [ + 302, + 486, + 525, + 689 + ], + "type": "text", + "content": "SE encourages diverse generations. Lacking generation diversity is a notorious problem for Seq2Seq learning tasks (Sun et al., 2020; Lin et al., 2022). Benefiting from better exploring the model's prediction with corrected soft labels, SE is expected to improve generation diversity. We follow Wang et al. (2022) to examine this by analyzing the performance in an additional multiple-reference test of WMT'14 En-De (Ott et al., 2018). We choose additional references for each of the 500 test sentences taken from the original test. Table 5 shows SE consistently outperforms the baseline with the average improvement being 0.9/1.0 BLEU, which indicates that our SE can effectively generate diverse results." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 698, + 525, + 752 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 698, + 525, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 698, + 525, + 752 + ], + "type": "text", + "content": "SE enhances model generalization. Benefiting from better hard token exploration, SE-equipped Transformers are expected to own better generalizations. We examine it by testing on domain shift" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 761, + 468, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 761, + 468, + 772 + ], + "spans": [ + { + "bbox": [ + 315, + 761, + 468, + 772 + ], + "type": "text", + "content": "Using compare-mt (Neubig et al., 2019)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "844" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 69, + 289, + 158 + ], + "blocks": [ + { + "bbox": [ + 70, + 69, + 289, + 158 + ], + "lines": [ + { + "bbox": [ + 70, + 69, + 289, + 158 + ], + "spans": [ + { + "bbox": [ + 70, + 69, + 289, + 158 + ], + "type": "table", + "html": "
Ref.Avg.Top
Transformer+SETransformer+SE
#142.543.7 (+1.2)44.945.7 (+0.8)
#228.629.3 (+0.7)30.231.2 (+1.0)
#331.232.1 (+0.9)33.234.4 (+1.2)
#428.128.8 (+0.7)29.630.5 (+0.9)
Mean32.633.5 (+0.9)34.535.5 (+1.0)
", + "image_path": "61fc329e7fdf7d3f06850dbdb4486eb540b9d0511a1847aa0f9d5960d68d19ed.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 71, + 202, + 289, + 252 + ], + "blocks": [ + { + "bbox": [ + 67, + 166, + 291, + 190 + ], + "lines": [ + { + "bbox": [ + 67, + 166, + 291, + 190 + ], + "spans": [ + { + "bbox": [ + 67, + 166, + 291, + 190 + ], + "type": "text", + "content": "Table 5: Multi-reference performance. 'Avg./ Top' means the averaging/ most-matching performance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 71, + 202, + 289, + 252 + ], + "lines": [ + { + "bbox": [ + 71, + 202, + 289, + 252 + ], + "spans": [ + { + "bbox": [ + 71, + 202, + 289, + 252 + ], + "type": "table", + "html": "
ModelLawMed.Kor.Sub.Avg.
Transformer41.230.97.414.523.5
+SE42.6†32.3†7.8†15.0†24.4
", + "image_path": "f28cd44a350b3048880357db5a229d08709812f733b17bbf702c89bdf7b08d98.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 331, + 290, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 331, + 290, + 413 + ], + "spans": [ + { + "bbox": [ + 67, + 331, + 290, + 413 + ], + "type": "text", + "content": "scenarios following Ding et al. (2021b). In particular, we evaluate WMT14 En-De models over four out-of-domain test sets (Müller et al., 2020) in Table 6 and find that SE improves the translation by averaging " + }, + { + "bbox": [ + 67, + 331, + 290, + 413 + ], + "type": "inline_equation", + "content": "+0.9" + }, + { + "bbox": [ + 67, + 331, + 290, + 413 + ], + "type": "text", + "content": " BLEU points, showing a better lexical generalization ability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 421, + 290, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 421, + 290, + 637 + ], + "spans": [ + { + "bbox": [ + 67, + 421, + 290, + 637 + ], + "type": "text", + "content": "SE encourages human-like generations. We design two types of evaluation on WMT14 En-Fr: 1) AUTOMATIC EVALUATION with COMET (Rei et al., 2020) and BLEURT (Sellam et al., 2020), which have a high-level correlation with human judgments. 2) HUMAN EVALUATION with three near-native French annotators who hold DALF C2 certificate2. Specifically, for human evaluation, we randomly sample 50 sentences from the test set to evaluate the translation adequacy and fluency, scoring " + }, + { + "bbox": [ + 67, + 421, + 290, + 637 + ], + "type": "inline_equation", + "content": "1 \\sim 5" + }, + { + "bbox": [ + 67, + 421, + 290, + 637 + ], + "type": "text", + "content": ". For adequacy, 1 represents irrelevant to the source while 5 means semantically equal. For fluency, 1 means unintelligible while 5 means fluent and native. Table 7 shows the automatic and human evaluation results, where we find that our SE indeed achieves human-like translation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 649, + 146, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 649, + 146, + 661 + ], + "spans": [ + { + "bbox": [ + 67, + 649, + 146, + 661 + ], + "type": "text", + "content": "4 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 671, + 290, + 752 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 671, + 290, + 752 + ], + "spans": [ + { + "bbox": [ + 67, + 671, + 290, + 752 + ], + "type": "text", + "content": "In this paper, we propose a self-evolution learning mechanism to improve seq2seq learning, by exploiting the informative-yet-underexplored tokens dynamically. SE follows two stages, i.e. self-questioning and self-evolution training, and can be used to evolve any pretrained models with a sim" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 306, + 69, + 523, + 126 + ], + "blocks": [ + { + "bbox": [ + 67, + 260, + 290, + 309 + ], + "lines": [ + { + "bbox": [ + 67, + 260, + 290, + 309 + ], + "spans": [ + { + "bbox": [ + 67, + 260, + 290, + 309 + ], + "type": "text", + "content": "Table 6: Performance on domain shift setting. Models are trained on the news but evaluated on out-of-domain test sets, including law, medicine, koran, and subtitle. “†” indicates statistically significance " + }, + { + "bbox": [ + 67, + 260, + 290, + 309 + ], + "type": "inline_equation", + "content": "(p < 0.05)" + }, + { + "bbox": [ + 67, + 260, + 290, + 309 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 69, + 523, + 126 + ], + "lines": [ + { + "bbox": [ + 306, + 69, + 523, + 126 + ], + "spans": [ + { + "bbox": [ + 306, + 69, + 523, + 126 + ], + "type": "table", + "html": "
AUTOMATIC EVAL.HUMAN EVAL.
COMETBLEURTAdequacyFluency
Transformer + SE61.668.64.324.58
63.769.54.504.68
", + "image_path": "8f4ba288b0456c34d1b25ff851b1782cac8a3512a72bb365a00554ee58d9e557.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 316, + 134, + 510, + 146 + ], + "lines": [ + { + "bbox": [ + 316, + 134, + 510, + 146 + ], + "spans": [ + { + "bbox": [ + 316, + 134, + 510, + 146 + ], + "type": "text", + "content": "Table 7: Human evaluation on WMT14 En-Fr." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 302, + 169, + 525, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 169, + 525, + 234 + ], + "spans": [ + { + "bbox": [ + 302, + 169, + 525, + 234 + ], + "type": "text", + "content": "ple recipe: continue train with SE. We empirically demonstrated the effectiveness and universality of SE on a series of widely-used benchmarks, covering low, medium, high, and extremely-high data volumes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 237, + 526, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 237, + 526, + 371 + ], + "spans": [ + { + "bbox": [ + 302, + 237, + 526, + 371 + ], + "type": "text", + "content": "In the future, besides generation tasks, we would like to verify the effectiveness of SE on language understanding tasks (Wu et al., 2020; Zhong et al., 2023). Also, it will be interesting to design SE-inspired instruction tuning or prompting strategy like Lu et al. (2023) to enhance the performance of large language models, e.g. ChatGPT3, which after all have already been fully validated on lots of conditional generation tasks (Hendy et al., 2023; Jiao et al., 2023; Peng et al., 2023; Wu et al., 2023)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 382, + 365, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 382, + 365, + 395 + ], + "spans": [ + { + "bbox": [ + 302, + 382, + 365, + 395 + ], + "type": "text", + "content": "Limitations" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 404, + 525, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 404, + 525, + 539 + ], + "spans": [ + { + "bbox": [ + 302, + 404, + 525, + 539 + ], + "type": "text", + "content": "Our work has several potential limitations. First, we determine the threshold " + }, + { + "bbox": [ + 302, + 404, + 525, + 539 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 302, + 404, + 525, + 539 + ], + "type": "text", + "content": " by manual selection, which may limit the performance of Seq2Seq models, it will make our work more effective and elegant if we dynamically select the threshold. Second, besides the improvement on three widely used tasks, we believe that there are still other abilities, like code generation, of Seq2Seq models that can be improved by our method, which are not fully explored in this work." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 550, + 393, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 550, + 393, + 562 + ], + "spans": [ + { + "bbox": [ + 302, + 550, + 393, + 562 + ], + "type": "text", + "content": "Ethics Statement" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 571, + 525, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 571, + 525, + 680 + ], + "spans": [ + { + "bbox": [ + 302, + 571, + 525, + 680 + ], + "type": "text", + "content": "We take ethical considerations very seriously and strictly adhere to the ACL Ethics Policy. This paper focuses on effective training for sequence-to-sequence learning. The datasets used in this paper are publicly available and have been widely adopted by researchers. We ensure that the findings and conclusions of this paper are reported accurately and objectively." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 690, + 400, + 704 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 690, + 400, + 704 + ], + "spans": [ + { + "bbox": [ + 302, + 690, + 400, + 704 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 302, + 713, + 524, + 753 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 713, + 524, + 753 + ], + "spans": [ + { + "bbox": [ + 302, + 713, + 524, + 753 + ], + "type": "text", + "content": "We are grateful to the anonymous reviewers and the area chair for their insightful comments and suggestions." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 80, + 760, + 224, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 760, + 224, + 772 + ], + "spans": [ + { + "bbox": [ + 80, + 760, + 224, + 772 + ], + "type": "text", + "content": "2http://www.delfdalf.fr/dalf-c2-en.html" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 760, + 451, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 760, + 451, + 772 + ], + "spans": [ + { + "bbox": [ + 315, + 760, + 451, + 772 + ], + "type": "text", + "content": "3https://chat.openai.com/" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "text", + "content": "845" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 71, + 127, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 71, + 127, + 83 + ], + "spans": [ + { + "bbox": [ + 69, + 71, + 127, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 90, + 289, + 772 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 69, + 90, + 289, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 90, + 289, + 123 + ], + "spans": [ + { + "bbox": [ + 69, + 90, + 289, + 123 + ], + "type": "text", + "content": "Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2015. Neural machine translation by jointly learning to align and translate. In ICLR." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 132, + 289, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 132, + 289, + 164 + ], + "spans": [ + { + "bbox": [ + 69, + 132, + 289, + 164 + ], + "type": "text", + "content": "Kehai Chen, Rui Wang, Masao Utiyama, and Eiichiro Sumita. 2020. Content word aware neural machine translation. In ACL." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 175, + 289, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 175, + 289, + 207 + ], + "spans": [ + { + "bbox": [ + 69, + 175, + 289, + 207 + ], + "type": "text", + "content": "Jianpeng Cheng and Mirella Lapata. 2016. Neural summarization by extracting sentences and words. In ACL." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 217, + 289, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 217, + 289, + 250 + ], + "spans": [ + { + "bbox": [ + 69, + 217, + 289, + 250 + ], + "type": "text", + "content": "Shamil Chollampatt and Hwee Tou Ng. 2018. A multilayer convolutional encoder-decoder neural network for grammatical error correction. In AAAI." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 260, + 289, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 260, + 289, + 291 + ], + "spans": [ + { + "bbox": [ + 69, + 260, + 289, + 291 + ], + "type": "text", + "content": "Kenneth Church and Patrick Hanks. 1990. Word association norms, mutual information, and lexicography. CL." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 302, + 289, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 302, + 289, + 324 + ], + "spans": [ + { + "bbox": [ + 69, + 302, + 289, + 324 + ], + "type": "text", + "content": "Daniel Dahlmeier and Hwee Tou Ng. 2012. Better evaluation for grammatical error correction. In NAACL." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 334, + 289, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 334, + 289, + 378 + ], + "spans": [ + { + "bbox": [ + 69, + 334, + 289, + 378 + ], + "type": "text", + "content": "Liang Ding, Longyue Wang, Xuebo Liu, Derek F Wong, Dacheng Tao, and Zhaopeng Tu. 2021a. Progressive multi-granularity training for non-autoregressive translation. In Findings of ACL." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 387, + 289, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 387, + 289, + 441 + ], + "spans": [ + { + "bbox": [ + 69, + 387, + 289, + 441 + ], + "type": "text", + "content": "Liang Ding, Longyue Wang, Xuebo Liu, Derek F Wong, Dacheng Tao, and Zhaopeng Tu. 2021b. Rejuvenating low-frequency words: Making the most of parallel data in non-autoregressive translation. In ACL." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 452, + 289, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 452, + 289, + 495 + ], + "spans": [ + { + "bbox": [ + 69, + 452, + 289, + 495 + ], + "type": "text", + "content": "Liang Ding, Longyue Wang, Xuebo Liu, Derek F Wong, Dacheng Tao, and Zhaopeng Tu. 2021c. Understanding and improving lexical choice in non-autoregressive translation. In ICLR." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 506, + 289, + 549 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 506, + 289, + 549 + ], + "spans": [ + { + "bbox": [ + 69, + 506, + 289, + 549 + ], + "type": "text", + "content": "Liang Ding, Longyue Wang, Shuming Shi, Dacheng Tao, and Zhaopeng Tu. 2022. Redistributing low-frequency words: Making the most of monolingual data in non-autoregressive translation. In ACL." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 69, + 559, + 289, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 559, + 289, + 591 + ], + "spans": [ + { + "bbox": [ + 69, + 559, + 289, + 591 + ], + "type": "text", + "content": "Liang Ding, Longyue Wang, and Dacheng Tao. 2020. Self-attention with cross-lingual position representation. In ACL." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 69, + 601, + 289, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 601, + 289, + 644 + ], + "spans": [ + { + "bbox": [ + 69, + 601, + 289, + 644 + ], + "type": "text", + "content": "Shuhao Gu, Jinchao Zhang, Fandong Meng, Yang Feng, Wanying Xie, Jie Zhou, and Dong Yu. 2020. Token-level adaptive training for neural machine translation. In EMNLP." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 69, + 654, + 289, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 654, + 289, + 686 + ], + "spans": [ + { + "bbox": [ + 69, + 654, + 289, + 686 + ], + "type": "text", + "content": "Sangchul Hahn and Heeyoul Choi. 2019. Self-knowledge distillation in natural language processing. In RANLP." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 69, + 697, + 289, + 729 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 697, + 289, + 729 + ], + "spans": [ + { + "bbox": [ + 69, + 697, + 289, + 729 + ], + "type": "text", + "content": "Amr Hendy, Mohamed Abdelrehim, et al. 2023. How good are gpt models at machine translation? a comprehensive evaluation. arXiv preprint." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 69, + 739, + 289, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 739, + 289, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 739, + 289, + 772 + ], + "type": "text", + "content": "Wenxiang Jiao, Wenxuan Wang, Jen-tse Huang, Xing Wang, and Zhaopeng Tu. 2023. Is chatgpt a good translator? a preliminary study. arXiv preprint." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 524, + 772 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 304, + 72, + 524, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 524, + 138 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 524, + 138 + ], + "type": "text", + "content": "Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In ACL." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 147, + 524, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 147, + 524, + 169 + ], + "spans": [ + { + "bbox": [ + 304, + 147, + 524, + 169 + ], + "type": "text", + "content": "Haoran Li and Wei Lu. 2021. Mixed cross entropy loss for neural machine translation. In ICML." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 179, + 524, + 210 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 179, + 524, + 210 + ], + "spans": [ + { + "bbox": [ + 304, + 179, + 524, + 210 + ], + "type": "text", + "content": "Zuchao Li, Rui Wang, et al. 2020. Data-dependent gaussian prior objective for language generation. In ICLR." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 221, + 524, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 221, + 524, + 253 + ], + "spans": [ + { + "bbox": [ + 304, + 221, + 524, + 253 + ], + "type": "text", + "content": "Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 263, + 524, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 263, + 524, + 317 + ], + "spans": [ + { + "bbox": [ + 304, + 263, + 524, + 317 + ], + "type": "text", + "content": "Huan Lin, Baosong Yang, Liang Yao, Dayiheng Liu, Haibo Zhang, Jun Xie, Min Zhang, and Jinsong Su. 2022. Bridging the gap between training and inference: Multi-candidate optimization for diverse neural machine translation. In Findings of NAACL." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 326, + 524, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 326, + 524, + 370 + ], + "spans": [ + { + "bbox": [ + 304, + 326, + 524, + 370 + ], + "type": "text", + "content": "Xuebo Liu, Longyue Wang, Derek F Wong, Liang Ding, Lidia S Chao, Shuming Shi, and Zhaopeng Tu. 2021a. On the copying behaviors of pre-training for neural machine translation. In Findings of ACL." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 380, + 524, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 380, + 524, + 423 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 524, + 423 + ], + "type": "text", + "content": "Xuebo Liu, Longyue Wang, Derek F Wong, Liang Ding, Lidia S Chao, and Zhaopeng Tu. 2021b. Understanding and improving encoder layer fusion in sequence-to-sequence learning. In ICLR." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 433, + 524, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 433, + 524, + 487 + ], + "spans": [ + { + "bbox": [ + 304, + 433, + 524, + 487 + ], + "type": "text", + "content": "Qingyu Lu, Baopu Qiu, Liang Ding, Liping Xie, and Dacheng Tao. 2023. Error analysis prompting enables human-like translation evaluation in large language models: A case study on chatgpt. arXiv preprint." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 497, + 524, + 539 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 524, + 539 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 524, + 539 + ], + "type": "text", + "content": "Mengqi Miao, Fandong Meng, Yijin Liu, Xiao-Hua Zhou, and Jie Zhou. 2021. Prevent the language model from being overconfident in neural machine translation. In ACL." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 549, + 524, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 549, + 524, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 549, + 524, + 581 + ], + "type": "text", + "content": "Mathias Müller, Annette Rios, and Rico Sennrich. 2020. Domain robustness in neural machine translation. In AMTA, Virtual." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 591, + 524, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 591, + 524, + 634 + ], + "spans": [ + { + "bbox": [ + 304, + 591, + 524, + 634 + ], + "type": "text", + "content": "Graham Neubig, Zi-Yi Dou, Junjie Hu, Paul Michel, Danish Pruthi, and Xinyi Wang. 2019. compare-mt: A tool for holistic comparison of language generation systems. In NAACL." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 644, + 524, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 644, + 524, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 644, + 524, + 677 + ], + "type": "text", + "content": "Mohammad Norouzi, Samy Bengio, Zhifeng Chen, et al. 2016. Reward augmented maximum likelihood for neural structured prediction. In NeurIPS." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 686, + 524, + 719 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 686, + 524, + 719 + ], + "spans": [ + { + "bbox": [ + 304, + 686, + 524, + 719 + ], + "type": "text", + "content": "Myle Ott, Michael Auli, David Grangier, and Marc'Aurelio Ranzato. 2018. Analyzing uncertainty in neural machine translation. In ICML." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 728, + 524, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 728, + 524, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 728, + 524, + 772 + ], + "type": "text", + "content": "Myle Ott, Sergey Edunov, Alexei Baevski, Angela Fan, Sam Gross, Nathan Ng, David Grangier, and Michael Auli. 2019. *fairoseq: A fast, extensible toolkit for sequence modeling*. In *NAACL Demonstration*." + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 790 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 790 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 790 + ], + "type": "text", + "content": "846" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 289, + 771 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 69, + 72, + 289, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 289, + 116 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 289, + 116 + ], + "type": "text", + "content": "Keqin Peng, Liang Ding, Qihuang Zhong, Li Shen, Xuebo Liu, Min Zhang, Yuanxin Ouyang, and Dacheng Tao. 2023. Towards making the most of chatgpt for machine translation. arxiv preprint." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 124, + 289, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 124, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 69, + 124, + 289, + 156 + ], + "type": "text", + "content": "Steven T Piantadosi. 2014. Zipf's word frequency law in natural language: A critical review and future directions. Psychonomic bulletin & review." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 165, + 288, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 165, + 288, + 186 + ], + "spans": [ + { + "bbox": [ + 69, + 165, + 288, + 186 + ], + "type": "text", + "content": "Matt Post. 2018. A call for clarity in reporting BLEU scores. In WMT." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 195, + 289, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 195, + 289, + 248 + ], + "spans": [ + { + "bbox": [ + 69, + 195, + 289, + 248 + ], + "type": "text", + "content": "Lynne M Reder, Xiaonan L Liu, Alexander Keinath, and Vencislav Popov. 2016. Building knowledge requires bricks, not sand: The critical role of familiar constituents in learning. Psychonomic bulletin & review." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 258, + 289, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 258, + 289, + 290 + ], + "spans": [ + { + "bbox": [ + 69, + 258, + 289, + 290 + ], + "type": "text", + "content": "Ricardo Rei, Craig Stewart, Ana C. Farinha, and Alon Lavie. 2020. COMET: A neural framework for MT evaluation. In EMNLP." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 298, + 289, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 298, + 289, + 331 + ], + "spans": [ + { + "bbox": [ + 69, + 298, + 289, + 331 + ], + "type": "text", + "content": "Thibault Sellam, Dipanjan Das, and Ankur P. Parikh. 2020. BLEURT: learning robust metrics for text generation. In ACL." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 339, + 289, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 339, + 289, + 371 + ], + "spans": [ + { + "bbox": [ + 69, + 339, + 289, + 371 + ], + "type": "text", + "content": "Zewei Sun, Shujian Huang, Hao-Ran Wei, Xinyu Dai, and Jiajun Chen. 2020. Generating diverse translation by manipulating multi-head attention. In AAAI." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 380, + 289, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 380, + 289, + 411 + ], + "spans": [ + { + "bbox": [ + 69, + 380, + 289, + 411 + ], + "type": "text", + "content": "Ilya Sutskever, Oriol Vinyals, and Quoc V. Le. 2014. Sequence to sequence learning with neural networks. In NeurIPS." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 421, + 289, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 421, + 289, + 463 + ], + "spans": [ + { + "bbox": [ + 69, + 421, + 289, + 463 + ], + "type": "text", + "content": "Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. 2016. Rethinking the inception architecture for computer vision. In CVPR." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 472, + 288, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 472, + 288, + 493 + ], + "spans": [ + { + "bbox": [ + 69, + 472, + 288, + 493 + ], + "type": "text", + "content": "Ashish Vaswani, Noam Shazeer, et al. 2017. Attention is all you need. In NeurIPS." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 502, + 288, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 502, + 288, + 523 + ], + "spans": [ + { + "bbox": [ + 69, + 502, + 288, + 523 + ], + "type": "text", + "content": "Yu Wan, Baosong Yang, et al. 2020. Self-paced learning for neural machine translation. In EMNLP." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 69, + 532, + 289, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 532, + 289, + 586 + ], + "spans": [ + { + "bbox": [ + 69, + 532, + 289, + 586 + ], + "type": "text", + "content": "Wenxuan Wang, Wenxiang Jiao, Yongchang Hao, Xing Wang, Shuming Shi, Zhaopeng Tu, and Michael R. Lyu. 2022. Understanding and improving sequence-to-sequence pretraining for neural machine translation. In ACL." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 69, + 594, + 289, + 627 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 594, + 289, + 627 + ], + "spans": [ + { + "bbox": [ + 69, + 594, + 289, + 627 + ], + "type": "text", + "content": "Di Wu, Liang Ding, Fan Lu, and Jian Xie. 2020. Slotrefine: A fast non-autoregressive model for joint intent detection and slot filling. In EMNLP." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 69, + 635, + 289, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 635, + 289, + 680 + ], + "spans": [ + { + "bbox": [ + 69, + 635, + 289, + 680 + ], + "type": "text", + "content": "Haoran Wu, Wenxuan Wang, Yuxuan Wan, Wenxiang Jiao, and Michael Lyu. 2023. Chatgpt or grammarly? evaluating chatgpt on grammatical error correction benchmark. arXiv preprint." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 69, + 687, + 289, + 719 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 687, + 289, + 719 + ], + "spans": [ + { + "bbox": [ + 69, + 687, + 289, + 719 + ], + "type": "text", + "content": "Fengshun Xiao, Yingting Wu, Hai Zhao, Rui Wang, and Shu Jiang. 2019. Dual skew divergence loss for neural machine translation. CoRR." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 69, + 728, + 289, + 771 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 728, + 289, + 771 + ], + "spans": [ + { + "bbox": [ + 69, + 728, + 289, + 771 + ], + "type": "text", + "content": "Yangyifan Xu, Yijin Liu, Fandong Meng, Jiajun Zhang, Jinan Xu, and Jie Zhou. 2021. Bilingual mutual information based adaptive training for neural machine translation. In ACL." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 524, + 474 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 304, + 72, + 524, + 104 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 524, + 104 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 524, + 104 + ], + "type": "text", + "content": "Zheng Yuan and Ted Briscoe. 2016. Grammatical error correction using neural machine translation. In NAACL." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 114, + 524, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 114, + 524, + 157 + ], + "spans": [ + { + "bbox": [ + 304, + 114, + 524, + 157 + ], + "type": "text", + "content": "Changtong Zan, Liang Ding, Li Shen, Yu Cao, Weifeng Liu, and Dacheng Tao. 2022a. On the complementarity between pre-training and random-initialization for resource-rich machine translation. In COLING." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 166, + 524, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 166, + 524, + 200 + ], + "spans": [ + { + "bbox": [ + 304, + 166, + 524, + 200 + ], + "type": "text", + "content": "Changtong Zan, Keqin Peng, Liang Ding, et al. 2022b. Vega-mt: The jd explore academy machine translation system for wmt22. In WMT." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 208, + 524, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 208, + 524, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 208, + 524, + 251 + ], + "type": "text", + "content": "Runtian Zhai, Chen Dan, J Zico Kolter, and Pradeep Kumar Ravikumar. 2023. Understanding why generalized reweighting does not improve over ERM. In ICLR." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 261, + 524, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 261, + 524, + 304 + ], + "spans": [ + { + "bbox": [ + 304, + 261, + 524, + 304 + ], + "type": "text", + "content": "Songming Zhang, Yijin Liu, Fandong Meng, Yufeng Chen, Jinan Xu, Jian Liu, and Jie Zhou. 2022a. Conditional bilingual mutual information based adaptive training for neural machine translation. In ACL." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 314, + 524, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 314, + 524, + 358 + ], + "spans": [ + { + "bbox": [ + 304, + 314, + 524, + 358 + ], + "type": "text", + "content": "Zheng Zhang, Liang Ding, Dazhao Cheng, Xuebo Liu, Min Zhang, and Dacheng Tao. 2022b. Bliss: Robust sequence-to-sequence learning via self-supervised input representation. arXiv preprint." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 367, + 524, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 367, + 524, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 367, + 524, + 411 + ], + "type": "text", + "content": "Qihuang Zhong, Liang Ding, Juhua Liu, Bo Du, and Dacheng Tao. 2022. E2s2: Encoding-enhanced sequence-to-sequence pretraining for language understanding and generation. arXiv preprint." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 419, + 524, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 419, + 524, + 474 + ], + "spans": [ + { + "bbox": [ + 304, + 419, + 524, + 474 + ], + "type": "text", + "content": "Qihuang Zhong, Liang Ding, Keqin Peng, Juhua Liu, Bo Du, Li Shen, Yibing Zhan, and Dacheng Tao. 2023. Bag of tricks for effective language model pretraining and downstream adaptation: A case study on glue. arXiv preprint." + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 486, + 376, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 376, + 500 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 376, + 500 + ], + "type": "text", + "content": "A Appendix" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 303, + 507, + 525, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 507, + 525, + 655 + ], + "spans": [ + { + "bbox": [ + 303, + 507, + 525, + 655 + ], + "type": "text", + "content": "Parameter Analysis on " + }, + { + "bbox": [ + 303, + 507, + 525, + 655 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 303, + 507, + 525, + 655 + ], + "type": "text", + "content": " As stated in §2.1, we use the loss threshold " + }, + { + "bbox": [ + 303, + 507, + 525, + 655 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 303, + 507, + 525, + 655 + ], + "type": "text", + "content": " to dynamically select the hard-to-learn tokens. Here, we analyze the influence of different " + }, + { + "bbox": [ + 303, + 507, + 525, + 655 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 303, + 507, + 525, + 655 + ], + "type": "text", + "content": " in detail. In practice, we train the Transformer models with different " + }, + { + "bbox": [ + 303, + 507, + 525, + 655 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 303, + 507, + 525, + 655 + ], + "type": "text", + "content": " (in " + }, + { + "bbox": [ + 303, + 507, + 525, + 655 + ], + "type": "inline_equation", + "content": "\\{3,4,5,6\\}" + }, + { + "bbox": [ + 303, + 507, + 525, + 655 + ], + "type": "text", + "content": ") and evaluate the performance of the WMT14 En-De test set. Table 8 lists the performance of different " + }, + { + "bbox": [ + 303, + 507, + 525, + 655 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 303, + 507, + 525, + 655 + ], + "type": "text", + "content": ". The results of Table 8 show that SE is stable and insensitive to " + }, + { + "bbox": [ + 303, + 507, + 525, + 655 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 303, + 507, + 525, + 655 + ], + "type": "text", + "content": " within a certain range. Noting that we select " + }, + { + "bbox": [ + 303, + 507, + 525, + 655 + ], + "type": "inline_equation", + "content": "\\Gamma = 5" + }, + { + "bbox": [ + 303, + 507, + 525, + 655 + ], + "type": "text", + "content": " for all experiment settings based on the results in Table 8." + } + ] + } + ], + "index": 27 + }, + { + "type": "table", + "bbox": [ + 340, + 677, + 490, + 714 + ], + "blocks": [ + { + "bbox": [ + 340, + 677, + 490, + 714 + ], + "lines": [ + { + "bbox": [ + 340, + 677, + 490, + 714 + ], + "spans": [ + { + "bbox": [ + 340, + 677, + 490, + 714 + ], + "type": "table", + "html": "
Γ=3Γ=4Γ=5Γ=6
BLEU27.727.828.027.8
", + "image_path": "e282bc0bde55ed294ee627795f045b303dd5c6a93bb9f64a91d4611079e702d8.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "table_body" + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 722, + 518, + 733 + ], + "lines": [ + { + "bbox": [ + 308, + 722, + 518, + 733 + ], + "spans": [ + { + "bbox": [ + 308, + 722, + 518, + 733 + ], + "type": "text", + "content": "Table 8: Parameter analysis of " + }, + { + "bbox": [ + 308, + 722, + 518, + 733 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 308, + 722, + 518, + 733 + ], + "type": "text", + "content": " on WMT14 En-De." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 790 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 790 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 790 + ], + "type": "text", + "content": "847" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 71, + 142, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 71, + 142, + 84 + ], + "spans": [ + { + "bbox": [ + 68, + 71, + 142, + 84 + ], + "type": "text", + "content": "Ablation Study" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 92, + 291, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 92, + 291, + 295 + ], + "spans": [ + { + "bbox": [ + 67, + 92, + 291, + 295 + ], + "type": "text", + "content": "Metric. In this work, we use the loss-based metric to dynamically select the hard-to-learn tokens. To validate the effectiveness of the metric, we use a simple adaptive training method (\"+ ADD\") that adds 1 to the weighting term of loss of the hard-to-learn tokens. The results on WMT16 EnRo are shown in Table 9, the simple Add method can achieve +0.3 BLEU improvement compared to the baseline model, which proves that our proposed self-questioning stage indeed mines informative difficult tokens. Also, we can observe that learning these dynamic difficult tokens with our SE framework (\"+ SE\") could outperform \"+\" ADD\" by +0.6 BLUE points, demonstrating the superiority of our token-specific label smoothing approach." + } + ] + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 105, + 303, + 257, + 341 + ], + "blocks": [ + { + "bbox": [ + 105, + 303, + 257, + 341 + ], + "lines": [ + { + "bbox": [ + 105, + 303, + 257, + 341 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 257, + 341 + ], + "type": "table", + "html": "
Baseline+ ADD+ SE
BLEU35.135.436.0
", + "image_path": "929cf37b53e86fa65783e09240e01496ba4e8c3f20c04c4725162757d3cf2107.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 382, + 291, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 382, + 291, + 504 + ], + "spans": [ + { + "bbox": [ + 67, + 382, + 291, + 504 + ], + "type": "text", + "content": "Learning objective. As stated in §2.1, our learning objective is the combination of the ground truth and the model's prediction. To validate the effectiveness of predicted distribution, we conduct ablation experiments on WMT16 En-Ro and WMT14 En-De. The results in Table 10 show that adding the predicted distribution will consistently improve the model's performance, which proves the effectiveness of the predicted distribution." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 92, + 513, + 268, + 587 + ], + "blocks": [ + { + "bbox": [ + 71, + 349, + 285, + 361 + ], + "lines": [ + { + "bbox": [ + 71, + 349, + 285, + 361 + ], + "spans": [ + { + "bbox": [ + 71, + 349, + 285, + 361 + ], + "type": "text", + "content": "Table 9: Ablation performance of our SE. on Metric." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 92, + 513, + 268, + 587 + ], + "lines": [ + { + "bbox": [ + 92, + 513, + 268, + 587 + ], + "spans": [ + { + "bbox": [ + 92, + 513, + 268, + 587 + ], + "type": "table", + "html": "
MethodBLEU
EN⇒DEEN⇒Ro
Transformer27.0835.11
SE28.0236.02
-w/o predicted results27.8935.71
", + "image_path": "a5062faacb13a04345d161149ed2f48791efd45af26686a5d85a458821cf830c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 594, + 289, + 617 + ], + "lines": [ + { + "bbox": [ + 67, + 594, + 289, + 617 + ], + "spans": [ + { + "bbox": [ + 67, + 594, + 289, + 617 + ], + "type": "text", + "content": "Table 10: Ablation performance of our SE. on learning objective." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 790 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 790 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 790 + ], + "type": "text", + "content": "848" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 90, + 192, + 103 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 90, + 192, + 103 + ], + "spans": [ + { + "bbox": [ + 68, + 90, + 192, + 103 + ], + "type": "text", + "content": "A For every submission:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 106, + 317, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 106, + 317, + 121 + ], + "spans": [ + { + "bbox": [ + 77, + 106, + 317, + 121 + ], + "type": "text", + "content": "A1. Did you describe the limitations of your work?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 121, + 217, + 134 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 121, + 217, + 134 + ], + "spans": [ + { + "bbox": [ + 89, + 121, + 217, + 134 + ], + "type": "text", + "content": "The last section of the paper." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 143, + 329, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 143, + 329, + 157 + ], + "spans": [ + { + "bbox": [ + 77, + 143, + 329, + 157 + ], + "type": "text", + "content": "A2. Did you discuss any potential risks of your work?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 158, + 138, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 158, + 138, + 169 + ], + "spans": [ + { + "bbox": [ + 89, + 158, + 138, + 169 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 77, + 179, + 414, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 179, + 414, + 192 + ], + "spans": [ + { + "bbox": [ + 77, + 179, + 414, + 192 + ], + "type": "text", + "content": "A3. Do the abstract and introduction summarize the paper's main claims?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 89, + 194, + 273, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 194, + 273, + 206 + ], + "spans": [ + { + "bbox": [ + 89, + 194, + 273, + 206 + ], + "type": "text", + "content": "The abstract and the introduction section." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 77, + 215, + 398, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 215, + 398, + 229 + ], + "spans": [ + { + "bbox": [ + 77, + 215, + 398, + 229 + ], + "type": "text", + "content": "A4. Have you used AI writing assistants when working on this paper?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 89, + 230, + 138, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 230, + 138, + 242 + ], + "spans": [ + { + "bbox": [ + 89, + 230, + 138, + 242 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 68, + 253, + 291, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 253, + 291, + 266 + ], + "spans": [ + { + "bbox": [ + 68, + 253, + 291, + 266 + ], + "type": "text", + "content": "B Did you use or create scientific artifacts?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 79, + 270, + 127, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 270, + 127, + 283 + ], + "spans": [ + { + "bbox": [ + 79, + 270, + 127, + 283 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 76, + 292, + 315, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 292, + 315, + 306 + ], + "spans": [ + { + "bbox": [ + 76, + 292, + 315, + 306 + ], + "type": "text", + "content": "B1. Did you cite the creators of artifacts you used?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 89, + 306, + 148, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 306, + 148, + 319 + ], + "spans": [ + { + "bbox": [ + 89, + 306, + 148, + 319 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 76, + 328, + 463, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 328, + 463, + 342 + ], + "spans": [ + { + "bbox": [ + 76, + 328, + 463, + 342 + ], + "type": "text", + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 89, + 343, + 148, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 343, + 148, + 355 + ], + "spans": [ + { + "bbox": [ + 89, + 343, + 148, + 355 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 76, + 364, + 524, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 364, + 524, + 417 + ], + "spans": [ + { + "bbox": [ + 76, + 364, + 524, + 417 + ], + "type": "text", + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 89, + 419, + 148, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 419, + 148, + 432 + ], + "spans": [ + { + "bbox": [ + 89, + 419, + 148, + 432 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 76, + 441, + 524, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 441, + 524, + 481 + ], + "spans": [ + { + "bbox": [ + 76, + 441, + 524, + 481 + ], + "type": "text", + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 89, + 483, + 148, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 483, + 148, + 495 + ], + "spans": [ + { + "bbox": [ + 89, + 483, + 148, + 495 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 76, + 504, + 524, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 504, + 524, + 531 + ], + "spans": [ + { + "bbox": [ + 76, + 504, + 524, + 531 + ], + "type": "text", + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 89, + 533, + 148, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 533, + 148, + 544 + ], + "spans": [ + { + "bbox": [ + 89, + 533, + 148, + 544 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 76, + 554, + 524, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 554, + 524, + 622 + ], + "spans": [ + { + "bbox": [ + 76, + 554, + 524, + 622 + ], + "type": "text", + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 89, + 623, + 148, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 623, + 148, + 634 + ], + "spans": [ + { + "bbox": [ + 89, + 623, + 148, + 634 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 68, + 644, + 294, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 644, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 68, + 644, + 294, + 657 + ], + "type": "text", + "content": "C Did you run computational experiments?" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 79, + 662, + 127, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 662, + 127, + 674 + ], + "spans": [ + { + "bbox": [ + 79, + 662, + 127, + 674 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 76, + 684, + 524, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 684, + 524, + 711 + ], + "spans": [ + { + "bbox": [ + 76, + 684, + 524, + 711 + ], + "type": "text", + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 89, + 712, + 148, + 724 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 712, + 148, + 724 + ], + "spans": [ + { + "bbox": [ + 89, + 712, + 148, + 724 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "spans": [ + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "type": "text", + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "spans": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "text", + "content": "ACL 2023 Responsible NLP Checklist" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "849" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 76, + 71, + 524, + 238 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 76, + 71, + 523, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 71, + 523, + 111 + ], + "spans": [ + { + "bbox": [ + 76, + 71, + 523, + 111 + ], + "type": "text", + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? No response." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 76, + 121, + 524, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 121, + 524, + 174 + ], + "spans": [ + { + "bbox": [ + 76, + 121, + 524, + 174 + ], + "type": "text", + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? No response." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 76, + 184, + 524, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 184, + 524, + 238 + ], + "spans": [ + { + "bbox": [ + 76, + 184, + 524, + 238 + ], + "type": "text", + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? No response." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 68, + 246, + 522, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 246, + 522, + 261 + ], + "spans": [ + { + "bbox": [ + 68, + 246, + 522, + 261 + ], + "type": "text", + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 79, + 264, + 131, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 264, + 131, + 276 + ], + "spans": [ + { + "bbox": [ + 79, + 264, + 131, + 276 + ], + "type": "text", + "content": "Section 3.2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 77, + 286, + 524, + 539 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 77, + 286, + 524, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 286, + 524, + 326 + ], + "spans": [ + { + "bbox": [ + 77, + 286, + 524, + 326 + ], + "type": "text", + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Left blank." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 77, + 336, + 524, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 336, + 524, + 390 + ], + "spans": [ + { + "bbox": [ + 77, + 336, + 524, + 390 + ], + "type": "text", + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Left blank." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 77, + 399, + 524, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 399, + 524, + 454 + ], + "spans": [ + { + "bbox": [ + 77, + 399, + 524, + 454 + ], + "type": "text", + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Left blank." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 77, + 462, + 519, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 462, + 519, + 489 + ], + "spans": [ + { + "bbox": [ + 77, + 462, + 519, + 489 + ], + "type": "text", + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Left blank." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 77, + 498, + 524, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 498, + 524, + 539 + ], + "spans": [ + { + "bbox": [ + 77, + 498, + 524, + 539 + ], + "type": "text", + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Left blank." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "850" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Toward Expanding the Scope of Radiology Report Summarization to Multiple Anatomies and Modalities/5f10cf4c-c72c-4642-9a79-12d04b80e5f5_content_list.json b/2023/Toward Expanding the Scope of Radiology Report Summarization to Multiple Anatomies and Modalities/5f10cf4c-c72c-4642-9a79-12d04b80e5f5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..98fb2523e69c3ed28d1aaf4769add3983a0d3d0d --- /dev/null +++ b/2023/Toward Expanding the Scope of Radiology Report Summarization to Multiple Anatomies and Modalities/5f10cf4c-c72c-4642-9a79-12d04b80e5f5_content_list.json @@ -0,0 +1,2340 @@ +[ + { + "type": "text", + "text": "Toward Expanding the Scope of Radiology Report Summarization to Multiple Anatomies and Modalities", + "text_level": 1, + "bbox": [ + 142, + 78, + 855, + 118 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhihong Chen $^{2,3*}$ , Maya Varma $^{1*}$ , Xiang Wan $^{2,3}$ , Curtis P. Langlotz $^{1}$ , Jean-Benoit Delbrouck $^{1*}$", + "bbox": [ + 268, + 124, + 742, + 158 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Stanford University", + "bbox": [ + 416, + 159, + 584, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2The Chinese University of Hong Kong, Shenzhen", + "bbox": [ + 295, + 175, + 705, + 191 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Shenzhen Research Institute of Big Data", + "bbox": [ + 332, + 192, + 670, + 208 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zhihongchen@link.cuhk.edu.cn wanxiang@sribd.cn", + "bbox": [ + 263, + 210, + 737, + 224 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{mvarma2,langlotz,jbdel}@stanford.edu", + "bbox": [ + 314, + 225, + 687, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 252, + 339, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Radiology report summarization (RRS) is a growing area of research. Given the Findings section of a radiology report, the goal is to generate a summary (called an Impression section) that highlights the key observations and conclusions of the radiology study. However, RRS currently faces essential limitations. First, many prior studies conduct experiments on private datasets, preventing the reproduction of results and fair comparisons across different systems and solutions. Second, most prior approaches are evaluated solely on chest X-rays. To address these limitations, we propose a dataset (MIMIC-RRS) involving three new modalities and seven new anatomies based on the MIMIC-III and MIMIC-CXR datasets. We then conduct extensive experiments to evaluate the performance of models both within and across modality-anatomy pairs in MIMIC-RRS. In addition, we evaluate their clinical efficacy via RadGraph, a factual correctness metric.", + "bbox": [ + 141, + 281, + 460, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 114, + 594, + 258, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A radiology report is a document that provides information about the results of a radiology study. It usually includes a Findings section with key observations from the study and an Impression section with the radiologist's overall conclusions. The latter is the most critical part of the report and is typically based on both the findings and the patient's condition. It can be helpful to automate the process of generating the impression section because it can be time-consuming and prone to errors when done manually (Bhargavan et al., 2009; Alexander et al., 2022). Recently, substantial progress has been made towards research on automated radiology report summarization (RRS) (Zhang et al., 2020; Ben Abacha et al., 2021; Hu et al., 2022). However, the field of RRS faces several key limitations. First, the experimental results of many", + "bbox": [ + 112, + 621, + 489, + 895 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "prior studies (Zhang et al., 2018, 2020) are reported on private datasets, making it difficult to replicate results or compare approaches. Second, existing studies are mainly limited to a single modality (i.e., X-ray) and a single anatomy (i.e., chest) (Zhang et al., 2020; Ben Abacha et al., 2021; Hu et al., 2021). In some cases, researchers omit to disclose the modality and anatomy of the radiology reports used for their experiments (Karn et al., 2022). Finally, recent models (Karn et al., 2022; Hu et al., 2022) present an increased complexity in architecture that offers only marginal improvements on the existing evaluation metrics for summarization. This further makes the replication of studies more difficult.", + "bbox": [ + 507, + 252, + 884, + 492 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address the aforementioned limitations, we construct a brand-new open-source dataset (named MIMIC-RRS) for radiology report summarization involving three modalities (X-ray, MRI, and CT) and seven anatomies (chest, head, neck, sinus, spine, abdomen, and pelvis). MIMIC-RRS is based on the MIMIC-CXR (Johnson et al., 2019) and MIMIC-III (Johnson et al., 2016) datasets and introduces data from 12 new modality-anatomy pairs. As a result, we introduce a new setting for evaluating the generalization capabilities of RRS models across different modalities and anatomies.", + "bbox": [ + 507, + 494, + 884, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In addition, we benchmark various pre-trained language models on MIMIC-RRS. Through extensive experiments within and across modality-anatomy pairs, we show that adopting an appropriate pretrained model can achieve promising results comparable to previous studies. We also introduce a metric to evaluate factual correctness of generated summaries for any modality-anatomy pair.", + "bbox": [ + 507, + 688, + 884, + 816 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Dataset Construction", + "text_level": 1, + "bbox": [ + 507, + 829, + 727, + 843 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this section, we present the new MIMIC-RRS dataset designed for radiology report summarization across multiple modalities and anatomies. Comparisons with existing datasets are shown in", + "bbox": [ + 507, + 854, + 884, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contribution.", + "bbox": [ + 136, + 904, + 265, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "469", + "bbox": [ + 485, + 927, + 515, + 940 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics", + "bbox": [ + 226, + 945, + 769, + 957 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Volume 2: Short Papers, pages 469-484", + "bbox": [ + 376, + 958, + 620, + 971 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "July 9-14, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 295, + 972, + 700, + 984 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/dead11df0c968ac5b18f4a4fb547a6419d07c1cb0ff012cbf6524e7faeb437e2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetAnatomyModalityLanguageNumber
Zhang et al. (2018)MultipleMultipleEnglish87,127
Zhang et al. (2020)MultipleMultipleEnglish130,850
RIH (Zhang et al., 2020)MultipleMultipleEnglish139,654
OpenI (Demner-Fushman et al., 2016)ChestX-rayEnglish3,268
MIMIC-CXR (Johnson et al., 2019)ChestX-rayEnglish128,003
PadChest (Bustos et al., 2020)ChestX-raySpanish206,222
MIMIC-RRS (ours)MultipleMultipleEnglish207,782
", + "bbox": [ + 181, + 80, + 816, + 234 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 Data Collection", + "text_level": 1, + "bbox": [ + 112, + 331, + 282, + 346 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "MIMIC-III One of our main contributions is to generate RRS data from MIMIC-II involving distinct combinations of modalities (i.e., medical imaging techniques) and anatomies (i.e., body parts). To this end, we first select five of the most frequently-occurring modality-anatomy pairs in the pool of MIMIC-III reports: \"CT Head\", \"CT Spine\", \"CT Chest\", \"CT Abdomen-Pelvis\" and \"MR Head\". Note that we discard chest X-rays as they are included in the MIMIC-CXR dataset. In addition, we pick six modality-anatomy pairs that occur infrequently in MIMIC-III to serve as out-of-domain (OOD) test sets: \"CT Neck\", \"CT Sinus\", \"MR Pelvis\", \"MR Neck\", \"MR Abdomen\", \"MR Spine\". This set of pairs represents two types of OOD cases: (1) the modality has not been seen during training (one could train on CT neck and test on MR Neck), and (2) the anatomy has not been seen during training (for example, CT Sinus is the only \"sinus\" dataset).", + "bbox": [ + 110, + 354, + 489, + 675 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For each report, we extract the findings and impression section. However, the findings section is not always clearly labeled as \"findings\". With the help of a board-certified radiologist, we identify alternate section headers that reference findings for each modality-anatomy pair. As an example, for CT head, findings may be referenced in reports with the section headings \"non-contrast head ct\", \"ct head\", \"ct head without contrast\", \"ct head without iv contrast\", \"head ct\", \"head ct without iv contrast\", or \"cta head\". We identify 537 candidate section headers that reference findings across our dataset. We also discarded reports where multiple studies are pooled in the same radiology report, leading to multiple intricate observations in the impression", + "bbox": [ + 110, + 678, + 489, + 919 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/6f4c2d9977b757eeec360c53e1cb287960a57f571619e3da871c15c7d9adb65a.jpg", + "table_caption": [ + "Table 1: Comparisons with existing datasets for radiology report summarization.", + "Table 1. We detail the collection process and the dataset statistics in the following subsections." + ], + "table_footnote": [], + "table_body": "
CT Abd-pelvCT ChestCT Head
15,98912,78631,402
CT SpineMR HeadCT Neck
5,5177,3131,140
CT SinusMR SpineMR Abdomen
1,2672,8211,061
MR NeckMR PelvisX-ray Chest
230253128,003
", + "bbox": [ + 551, + 281, + 847, + 445 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 2: Dataset statistics for MIMIC-RRS. We report the number of radiology reports from each modality-anatomy pair.", + "bbox": [ + 507, + 455, + 885, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "section1. Our resulting dataset consists of 79,779 selected reports across 11 modality-anatomy pairs.", + "bbox": [ + 507, + 521, + 882, + 556 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "MIMIC-CXR MIMIC-CXR studies are chest X-ray examinations. We follow preprocessing steps reported in previous work (Delbrouck et al., 2022b), and we only include reports with both a Findings and an Impression section. This yields 128,003 reports.", + "bbox": [ + 507, + 563, + 882, + 659 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2 Data statistics", + "text_level": 1, + "bbox": [ + 507, + 671, + 665, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In total, there are 207,782 samples in the MIMIC-RRS dataset. The number of examples for each modality and anatomy is provided in Table 2. To further analyze this dataset, we report in Figure 1 the text lengths and vocabulary sizes associated with reports from each modality-anatomy pair. We find that for all modality-anatomy pairs, the findings section is significantly longer than the impression section (up to $+315\\%$ for MR abdomen). Additionally, the findings sections of chest X-ray reports, which average only 49 words, are much shorter than reports from other modality-anatomy", + "bbox": [ + 505, + 690, + 884, + 885 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1We release our candidate section headers as well as code to recreate the dataset from scratch (Appendix B).", + "bbox": [ + 507, + 892, + 882, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "470", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3e0061ba8be51b6cb07e2785a6215aeb7af3386453565b0e4de0bf8819bd6c4a.jpg", + "image_caption": [ + "Figure 1: Section length and vocabulary size for reports from each modality-anatomy pair." + ], + "image_footnote": [], + "bbox": [ + 114, + 79, + 487, + 195 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "pairs. In contrast, MR Abdomen and MR Pelvis reports including findings sections that average 205 and 174 words, respectively. We see that CT Chest, CT Head, and CT Abdomen-Pelvis reports have a relatively large vocabulary size (given their sample size) with 20,909, 19,813, and 18,933 words. Surprisingly, the CT Abdomen-Pelvis impressions include a larger vocabulary than the findings. On the other hand, MR pelvis and MR abdomen impressions contain $36\\%$ and $37\\%$ fewer words than their corresponding findings, respectively.", + "bbox": [ + 110, + 261, + 490, + 438 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We assign reports from the following modality-anatomy pairs to training, validation, and test sets due to their large sample sizes: \"CT abdomen/pelvis\", \"CT Chest\", \"CT Neck\", \"CT Spine\", \"CT Head\", \"MR Head\", and \"X-ray Chest\". The remaining reports (i.e., \"MR Pelvis\", \"MR Spine\", \"MR Neck\", \"MR Abdomen\", and \"CT Sinus\") are used for OOD test sets $^2$ .", + "bbox": [ + 110, + 439, + 489, + 565 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 Algorithmic Analysis", + "text_level": 1, + "bbox": [ + 112, + 577, + 332, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we conduct experiments to analyze the performance of different models on MIMIC-RRS. We provide three categories of analyses: inmodality-anatomy, cross-modality-anatomy, and clinical efficacy.", + "bbox": [ + 112, + 602, + 489, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1 In-modality-anatomy", + "text_level": 1, + "bbox": [ + 112, + 695, + 329, + 711 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To benchmark the performance of different models on the proposed MIMIC-RRS dataset, we conduct experiments within each modality-anatomy pair (i.e., the training and test procedures are performed using only one modality-anatomy pair). We evaluate three types of pre-trained sequence-to-sequence models, namely T5 (Raffel et al., 2020), BART (Lewis et al., 2020), BioBART (Yuan et al., 2022), and their variants. Results are reported in", + "bbox": [ + 112, + 715, + 490, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 3.", + "text_level": 1, + "bbox": [ + 507, + 84, + 571, + 97 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Several observations can be drawn from these experiments. First, simply adopting pretrained sequence-to-sequence language models can achieve results comparable to previous state-of-the-art approaches designed for radiology summarization. Indeed, using BART-L as a backbone achieves the best performance, confirming the necessity of exploiting appropriate pre-trained language models. Secondly, the performances across different model types vary (i.e., BART-L/BART-B, BioBART-L/BioBART-B). Yet, we notice that the number of training parameters matters; large models report the best results. According to our evaluations, the BART models achieve better results across all modality-anatomy pairs. Surprisingly, it is worth noting that the BioBART models do not achieve better results than BART, although BioBART is pre-trained on a biomedical corpus. One explanation could be that BioBART models are pre-trained on abstracts from PubMed, which are not within the same domain as radiology reports.", + "bbox": [ + 507, + 99, + 884, + 437 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, we note several key findings for future studies: (i) \"Less is more\": starting from an appropriate backbone instead of designing complicated modules; (ii) the model size matters; (iii) the pretraining domain matters: knowledge from clinical notes or medical literature does not easily translate to radiology reports.", + "bbox": [ + 507, + 438, + 882, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2 Cross-modality-anatomy", + "text_level": 1, + "bbox": [ + 507, + 561, + 752, + 577 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we conduct experiments across modality-anatomy pairs (i.e., models are trained on reports from a subset of modality-anatomy pairs and then evaluated on all pairs, including the OOD test sets). We report the cross-modality-anatomy scores in Figure 2. A few interesting observations can be made. First, there are some associations between different anatomies and modalities. For example, the model trained on \"CT Head\" can also achieve promising results on the \"MR Head\" set. Secondly, training the model with all the modality-anatomy pairs (denoted as ALL) achieves the best generalization, obtaining the best results across all modalities and anatomies including the OOD test sets. We leave further exploration of cross-modality-anatomy associations and zero-shot OOD", + "bbox": [ + 507, + 582, + 882, + 840 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "et al., 2019), and Clinical-T5 (Lu et al., 2022)) that specialize in the clinical text since they were trained on the text from MIMIC-III, which overlaps with our dataset. The MIMIC-RRS test set is included in their pre-training data. Thus, we do not adopt them in our experiments to avoid potential data leakage and ensure a fair comparison.", + "bbox": [ + 507, + 845, + 882, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "2We release data splits publicly so that future work can fairly compare new results. 3We do not evaluate several pre-trained models (e.g., ClinicalBERT (Alsentzer et al., 2019), BioClinicalBERT (Alsentzer", + "bbox": [ + 112, + 866, + 487, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "471", + "bbox": [ + 485, + 928, + 512, + 940 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/09e4b9479b7f43318fa103b744bf6f33be27f15ad6e3cb2cec4d7db1b8cf9b39.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelsMR HeadCT SpineCT NeckCT HeadCT ChestCT Abd/PelX-ray Chest
R1R2RLR1R2RLR1R2RLR1R2RLR1R2RLR1R2RLR1R2RL
WGSum------------------48.433.346.7
AIG-CL------------------51.035.246.7
T5-S38.218.328.535.818.628.939.020.029.143.125.336.539.518.529.328.910.621.247.832.243.5
BioBART-B42.421.232.047.827.940.040.419.629.346.027.438.941.419.130.333.112.523.249.633.845.3
BioBART-L42.121.432.647.828.140.840.319.429.645.526.738.640.217.828.932.511.722.649.333.344.9
BART-B42.021.532.149.029.741.641.420.930.246.428.139.541.619.530.633.112.923.651.034.946.4
BART-L43.722.132.849.829.741.442.020.530.446.627.339.041.818.629.633.912.423.251.734.946.8
", + "bbox": [ + 121, + 80, + 884, + 200 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ed8891320f89fb9c958d4384a55d93432cad0a4283efacef08f26d99ac01d6f8.jpg", + "image_caption": [ + "Figure 2: Cross-modality-anatomy results from BART-L are visualized here using heatmaps. Colors from light to dark represent the values from low to high in each column. As discussed in Section 3.2, the model variant \"ALL\" reports the strongest performances." + ], + "image_footnote": [], + "bbox": [ + 117, + 268, + 359, + 495 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/88649b0013ced896b46bb9e32da3860d62d31f6ea58df1bcb8f5d0ea01738702.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 268, + 603, + 495 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/069e420a9ffef436dad2c0f5ad445bb82185a743120a22dc906b16229cdc250c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 608, + 268, + 880, + 495 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/2aecc3e1e0b16efb51de9eaebcd734880748117ece9fbd3ff59a3909b7d87798.jpg", + "table_caption": [ + "Table 3: The benchmarking comparisons of different approaches, including task-specific models (i.e., WGSum (Hu et al., 2021) and AIG-CL (Hu et al., 2022)) and pre-trained language models (i.e., T5-S, BioBART-B, BioBART-L, BART-B, and BART-L). R1, R2, and RL denote ROUGE-1, ROUGE-2, and ROUGE-L, respectively." + ], + "table_footnote": [], + "table_body": "
T5-SBioBART-BBioBART-LBART-BBART-L
MR Head21.524.825.325.026.1
CT Spine23.837.037.038.538.3
CT Neck21.223.623.624.024.9
CT Head31.834.234.035.234.7
CT Chest24.026.024.326.025.2
CT Abd/Pel12.615.915.316.115.9
X-ray Chest39.840.941.042.343.0
", + "bbox": [ + 117, + 576, + 485, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 4: F1-RadGraph scores on MIMIC-RRS test sets across different anatomies and modalities.", + "bbox": [ + 112, + 689, + 485, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "transfer for future work.", + "bbox": [ + 112, + 743, + 294, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3 Clinical-Efficacy", + "text_level": 1, + "bbox": [ + 112, + 769, + 290, + 785 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition to evaluating our systems using the ROUGE-1, ROUGE-2, and ROUGE-L metrics (Lin, 2004), we use a factual correctness metric to analyze clinical efficacy. Most prior works (Zhang et al., 2020; Smit et al., 2020; Hu et al., 2022) mainly use the $\\mathrm{F_1}$ CheXbert metric, an F1-score that evaluates the factual correctness of the generated impressions using 14 chest radio", + "bbox": [ + 112, + 790, + 490, + 919 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "graphic observations. Unfortunately, this metric is unsuitable for MIMIC-RRS, which contains reports from other modality-anatomy pairs beyond chest X-rays.", + "bbox": [ + 507, + 579, + 884, + 644 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For this reason, instead of using $\\mathrm{F_1}$ CheXbert, we propose to use RadGraph (Jain et al., 2021) to evaluate the clinical correctness of the generated impressions. RadGraph is a dataset containing board-certified radiologist annotations of radiology reports corresponding to 14,579 entities and 10,889 relations (Appendix A.1). We used the released pretrained model to annotate our reports and asked one board-certified radiologist to subjectively validate that the printed entities of the RadGraph model on our data are correct (examples are shown in Table 5). After confirming the effectiveness of the model, we follow Delbrouck et al. (2022a) to compute the F1-RadGraph scores. The score evaluates the correctness of the generated named entities in the hypothesis impression compared to the ground-truth impression. We report these results in Ta", + "bbox": [ + 507, + 645, + 885, + 919 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "472", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ble 4. It can be observed that the BART models can achieve the best performance with respect to clinical efficacy. The results are consistent with the ROUGE scores, further confirming the effectiveness of adopting BART as the backbone instead of designing complicated solutions.", + "bbox": [ + 112, + 84, + 489, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Related Work", + "text_level": 1, + "bbox": [ + 112, + 212, + 268, + 228 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we discuss prior research related to the radiology report summarization task. The first attempt at automatic summarization of radiology findings into natural language impression statements was proposed by Zhang et al. (2018). Their contribution was to propose a first baseline on the task, using a bidirectional-LSTM as encoder and decoder. Importantly, they found that about $30\\%$ of the summaries generated from neural models contained factual errors. Subsequently, Zhang et al. (2020) proposed the $\\mathrm{F_1}$ CheXbert score to evaluate the factual correctness of the generated impression. They also used reinforcement learning to optimize the $\\mathrm{F_1}$ CheXbert score directly. Finally, both Hu et al. (2021) and Hu et al. (2022) used the Biomedical and Clinical English Model Packages in the Stanza Python NLP Library (Zhang et al., 2021) to extract medical entities. The former study used the entities to construct a Graph Neural Network, which was used as input in their summarization pipeline. In contrast, the latter study used the entities to mask the findings during contrastive pre-training.", + "bbox": [ + 112, + 253, + 489, + 624 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We believe this paper is an original contribution to the aforementioned line of work. As instigated by Zhang et al. (2018), our goal is to release a new summarization corpus and baselines on new modalities and anatomies. We do so by releasing an RRS dataset with data from 11 new modality-anatomy pairs. In addition, we extend the work performed by Zhang et al. (2020) by proposing a new metric to evaluate the factual correctness and completeness of the generated impression, namely the RadGraph score. Finally, we improve on the work of Hu et al. (2021, 2022) in two ways: (1) we use semantic annotations from a pre-trained model trained using annotations from board-certified radiologists, as opposed to Stanza which leverages unsupervised biomedical and clinical text data; (2) we leverage relation annotations between entities, a feature that was not available in prior work.", + "bbox": [ + 112, + 629, + 489, + 917 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5 Conclusion and Discussion", + "text_level": 1, + "bbox": [ + 507, + 83, + 774, + 99 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this paper, we highlight and address several weaknesses associated with the radiology report summarization task. First, from a data perspective, we propose a publicly available dataset named MIMIC-RRS involving data samples from twelve modality-anatomy pairs, with 79,779 samples from MIMIC-III and 128,003 samples from MIMIC-CXR.", + "bbox": [ + 507, + 109, + 884, + 237 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Second, we conducted more than 40 experiments and over 400 cross-modality-anatomy evaluations to benchmark the performance of different models. We show that instead of designing complicated modules, we can start from an appropriate backbone model such as BART.", + "bbox": [ + 507, + 254, + 884, + 350 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we proposed an elegant and simple metric, F1-RadGraph, to evaluate the factual correctness of summaries generated for any modality and anatomy. In the future, we hope that our work broadens the scope of the radiology report summarization task and contributes to the development of reliable RRS models that generalize well to new anatomies and modalities.", + "bbox": [ + 507, + 367, + 884, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Limitations", + "text_level": 1, + "bbox": [ + 507, + 507, + 616, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We note two limitations of our paper. First, our work does not extensively evaluate all the available pre-trained models that could be suitable for this task, e.g., ELECTRA (Clark et al., 2020), BioLinkBERT (Yasunaga et al., 2022), GatorTron (Yang et al., 2022), RadBERT (Yan et al., 2022), and PubMedBERT (Gu et al., 2021). The aim of this work is not to report the strongest possible score but rather to address weaknesses of existing radiology report summarization studies (in terms of data and evaluation). Yet, we are confident our proposed solutions report a strong baseline for future work. Second, although F1-RadGraph seems like an appropriate metric to evaluate our new modalities and anatomies (and appears to be consistent with ROUGE scores), it has only been evaluated subjectively and not systematically.", + "bbox": [ + 507, + 532, + 884, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 507, + 818, + 672, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Maya Varma is supported by graduate fellowship awards from the Department of Defense (NDSEG) and the Knight-Hennessy Scholars program at Stanford University.", + "bbox": [ + 507, + 843, + 884, + 907 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "473", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 115, + 84, + 213, + 98 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Robert Alexander, Stephen Waite, Michael A Bruno, Elizabeth A Krupinski, Leonard Berlin, Stephen Macknik, and Susana Martinez-Conde. 2022. Mandating limits on workload, duty, and speed in radiology. *Radiology*, 304(2):274-282.", + "bbox": [ + 112, + 105, + 489, + 172 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Emily Alsentzer, John Murphy, William Boag, WeiHung Weng, Di Jindi, Tristan Naumann, and Matthew McDermott. 2019. Publicly available clinical bert embeddings. In Proceedings of the 2nd Clinical Natural Language Processing Workshop, pages 72-78.", + "bbox": [ + 112, + 181, + 489, + 249 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Asma Ben Abacha, Yassine Mrabet, Yuhao Zhang, Chaitanya Shivade, Curtis Langlotz, and Dina Demner-Fushman. 2021. Overview of the MEDIQA 2021 shared task on summarization in the medical domain. In Proceedings of the 20th Workshop on Biomedical Language Processing, pages 74-85, Online. Association for Computational Linguistics.", + "bbox": [ + 112, + 256, + 489, + 350 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Mythreyi Bhargavan, Adam H Kaye, Howard P Forman, and Jonathan H Sunshine. 2009. Workload of radiologists in united states in 2006-2007 and trends since 1991-1992. Radiology, 252(2):458-467.", + "bbox": [ + 112, + 359, + 489, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Aurelia Bustos, Antonio Pertusa, Jose-Maria Salinas, and Maria de la Iglesia-Vaya. 2020. Padchest: A large chest x-ray image dataset with multi-label annotated reports. Medical image analysis, 66:101797.", + "bbox": [ + 112, + 420, + 489, + 475 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Kevin Clark, Minh-Thang Luong, Quoc V. Le, and Christopher D. Manning. 2020. ELECTRA: Pretraining text encoders as discriminators rather than generators. In ICLR.", + "bbox": [ + 112, + 483, + 489, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Jean-Benoit Delbrouck, Pierre Chambon, Christian Bluethgen, Emily Tsai, Omar Almusa, and Curtis P Langlotz. 2022a. Improving the factual correctness of radiology report generation with semantic rewards. arXiv preprint arXiv:2210.12186.", + "bbox": [ + 112, + 546, + 489, + 613 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Jean-benoit Delbrouck, Khaled Saab, Maya Varma, Sabri Eyuboglu, Pierre Chambon, Jared Dunnmon, Juan Zambrano, Akshay Chaudhari, and Curtis Langlotz. 2022b. ViLMedic: a framework for research at the intersection of vision and language in medical AI. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 23-34, Dublin, Ireland. Association for Computational Linguistics.", + "bbox": [ + 112, + 621, + 489, + 741 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dina Demner-Fushman, Marc D Kohli, Marc B Rosenman, Sonya E Shooshan, Laritza Rodriguez, Sameer Antani, George R Thoma, and Clement J McDonald. 2016. Preparing a collection of radiology examinations for distribution and retrieval. Journal of the American Medical Informatics Association, 23(2):304-310.", + "bbox": [ + 112, + 750, + 489, + 829 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Yu Gu, Robert Tinn, Hao Cheng, Michael Lucas, Naoto Usuyama, Xiaodong Liu, Tristan Naumann, Jianfeng Gao, and Hoifung Poon. 2021. Domain-specific language model pretraining for biomedical natural language processing. ACM Transactions on Computing for Healthcare (HEALTH), 3(1):1-23.", + "bbox": [ + 112, + 839, + 489, + 917 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Jinpeng Hu, Jianling Li, Zhihong Chen, Yaling Shen, Yan Song, Xiang Wan, and Tsung-Hui Chang. 2021. Word graph guided summarization for radiology findings. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 4980-4990, Online. Association for Computational Linguistics.", + "bbox": [ + 507, + 84, + 884, + 165 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Jinpeng Hu, Zhuo Li, Zhihong Chen, Zhen Li, Xiang Wan, and Tsung-Hui Chang. 2022. Graph enhanced contrastive learning for radiology findings summarization. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4677-4688.", + "bbox": [ + 507, + 174, + 884, + 255 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Saahil Jain, Ashwin Agrawal, Adriel Saporta, Steven Truong, Du Nguyen Duong, Tan Bui, Pierre Chambon, Yuhao Zhang, Matthew Lungren, Andrew Ng, Curtis Langlotz, Pranav Rajpurkar, and Pranav Rajpurkar. 2021. Radgraph: Extracting clinical entities and relations from radiology reports. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, volume 1.", + "bbox": [ + 507, + 263, + 884, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Alistair EW Johnson, Tom J Pollard, Seth J Berkowitz, Nathaniel R Greenbaum, Matthew P Lungren, Chihying Deng, Roger G Mark, and Steven Horng. 2019. Mimic-cxr, a de-identified publicly available database of chest radiographs with free-text reports. Scientific data, 6(1):1-8.", + "bbox": [ + 507, + 379, + 884, + 458 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Alistair EW Johnson, Tom J Pollard, Lu Shen, Li-wei H Lehman, Mengling Feng, Mohammad Ghassemi, Benjamin Moody, Peter Szolovits, Leo Anthony Celi, and Roger G Mark. 2016. Mimic-iii, a freely accessible critical care database. Scientific data, 3(1):1-9.", + "bbox": [ + 507, + 469, + 884, + 535 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Sanjeev Kumar Karn, Ning Liu, Hinrich Schütze, and Oladimeji Farri. 2022. Differentiable multi-agent actor-critic for multi-step radiology report summarization. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1542-1553.", + "bbox": [ + 507, + 546, + 884, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871-7880.", + "bbox": [ + 507, + 634, + 884, + 728 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81.", + "bbox": [ + 507, + 738, + 884, + 778 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qiuhao Lu, Dejing Dou, and Thien Nguyen. 2022. Clinical5: A generative language model for clinical text. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 5436-5443.", + "bbox": [ + 507, + 788, + 884, + 841 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J Liu, et al. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(140):1-67.", + "bbox": [ + 507, + 851, + 884, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "474", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Akshay Smit, Saahil Jain, Pranav Rajpurkar, Anuj Parek, Andrew Y Ng, and Matthew Lungren. 2020. Combining automatic labelers and expert annotations for accurate radiology report labeling using bert. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1500-1519.", + "An Yan, Julian McAuley, Xing Lu, Jiang Du, Eric Y Chang, Amilcare Gentili, and Chun-Nan Hsu. 2022. Radbert: Adapting transformer-based language models to radiology. Radiology: Artificial Intelligence, 4(4):e210258.", + "Xi Yang, Nima PourNejatian, Hoo Chang Shin, Kaleb E Smith, Christopher Parisien, Colin Compas, Cheryl Martin, Mona G Flores, Ying Zhang, Tanja Magoc, et al. 2022. Gatortron: A large clinical language model to unlock patient information from unstructured electronic health records. arXiv preprint arXiv:2203.03540.", + "Michihiro Yasunaga, Jure Leskovec, and Percy Liang. 2022. Linkbert: Pretraining language models with document links. In Association for Computational Linguistics (ACL).", + "Hongyi Yuan, Zheng Yuan, Ruyi Gan, Jiaxing Zhang, Yutao Xie, and Sheng Yu. 2022. Biobart: Pretraining and evaluation of a biomedical generative language model. In Proceedings of the 21st Workshop on Biomedical Language Processing, pages 97-109.", + "Yuhao Zhang, Daisy Yi Ding, Tianpei Qian, Christopher D Manning, and Curtis P Langlotz. 2018. Learning to summarize radiology findings. In Proceedings of the Ninth International Workshop on Health Text Mining and Information Analysis, pages 204-213.", + "Yuhao Zhang, Derek Merck, Emily Tsai, Christopher D Manning, and Curtis Langlotz. 2020. Optimizing the factual correctness of a summary: A study of summarizing radiology reports. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5108-5120.", + "Yuhao Zhang, Yuhui Zhang, Peng Qi, Christopher D Manning, and Curtis P Langlotz. 2021. Biomedical and clinical english model packages for the stanza python nlp library. Journal of the American Medical Informatics Association, 28(9):1892-1899." + ], + "bbox": [ + 114, + 85, + 489, + 719 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "475", + "bbox": [ + 485, + 928, + 515, + 939 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A Details of RadGraph Scores", + "text_level": 1, + "bbox": [ + 114, + 84, + 394, + 99 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A.1 The Introduction of RadGraph", + "text_level": 1, + "bbox": [ + 114, + 112, + 408, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c2b5f95e63ae9279e70dfd8e06c1f438fa698ba2922a020d4ebdc1831e88daa4.jpg", + "image_caption": [ + "Figure 3: Example of the RadGraph annotations. Figure taken from (Jain et al., 2021)." + ], + "image_footnote": [], + "bbox": [ + 117, + 147, + 485, + 236 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To design our new evaluation metric, we leverage the RadGraph dataset (Jain et al., 2021) containing board-certified radiologist annotations of chest X-ray reports, which correspond to 14,579 entities and 10,889 relations. RadGraph has released a PubMedBERT model (Gu et al., 2021) pre-trained on these annotations to annotate new reports. An example of annotation can be seen in Figure 3. Before moving on to the next section, we quickly describe the concept of entities and relations:", + "bbox": [ + 112, + 293, + 489, + 454 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Entities An entity is defined as a continuous span of text that can include one or more adjacent words. Entities in RadGraph center around two concepts: Anatomy and Observation. Three uncertainty levels exist for Observation, leading to four different entities: Anatomy (ANAT-DP), Observation: Definitely Present (OBS-DP), Observation: Uncertain (OBS-U), and Observation: Definitely Absent (OBS-DA).", + "bbox": [ + 112, + 466, + 489, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Relations A relation is defined as a directed edge between two entities. Three levels exist: Suggestive $\\text{Of}(\\cdot, \\cdot)$ , Located At ( $. \\cdot$ ), and Modify ( $. \\cdot$ ).", + "bbox": [ + 112, + 607, + 487, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A.2 Metric Computation", + "text_level": 1, + "bbox": [ + 114, + 670, + 326, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Using the RadGraph annotation scheme and pretrained model, we designed an F-score style reward that measures the factual consistency and completeness of the generated impression (also called hypothesis impression) compared to the reference impression.", + "bbox": [ + 112, + 692, + 487, + 788 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To do so, we treat the RadGraph annotations of an impression as a graph $\\mathcal{G}(V,E)$ with the set of nodes $V = \\{v_{1},v_{2},\\ldots ,v_{|V|}\\}$ containing the entities and the set of edges $E = \\{e_1,e_2,\\dots ,e_{|E|}\\}$ the relations between pairs of entities. The graph is directed, meaning that the edge $e = (v_{1},v_{2})\\neq$ $(v_{2},v_{1})$ . An example is depicted in Figure 4. Each node or edge of the graph also has a label, which", + "bbox": [ + 112, + 790, + 489, + 917 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "we denote as $v_{i_L}$ for an entity $i$ (for example \"OBS-DP\" or \"ANAT\") and $e_{ij_L}$ for a relation $e = (v_i, v_j)$ (such as \"modified\" or \"located at\").", + "bbox": [ + 507, + 84, + 882, + 131 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To design our RadGraph score, we focus on the nodes $V$ and whether or not a node has a relation in $E$ . For a hypothesis impression $y$ , we create a new set of triplets $T_{y} = \\{(v_{i}, v_{i_{L}}, \\mathcal{R})\\}_{i=1:|V|}$ . The value $\\mathcal{R}$ is 1 if $(v_{i}, v_{j})_{j=1:|E|, i \\neq j} \\in E$ , 0 otherwise. In other words, a triplet contains an entity, the entity label, and whether or not this entity has a relation. We proceed to construct the same set for the reference report $\\hat{y}$ and denote this set $T_{\\hat{y}}$ .", + "bbox": [ + 507, + 133, + 882, + 277 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, our score is defined as the harmonic mean of precision and recall between the hypothesis set $T_{y}$ and the reference set $T_{\\hat{y}}$ , giving a value between 0 and 100. As an illustration, the set $V$ , $E$ and $T$ of the graph $\\mathcal{G}$ in Figure 4 are shown as follows:", + "bbox": [ + 507, + 278, + 882, + 357 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} V = \\{\\text {m i l d , f l u i d , o v e r l o a d , o v e r t , p u l m o n a r y}, \\\\ \\text {e d e m a} \\} \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 507, + 359, + 882, + 391 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} E = \\left\\{\\text {(m i l d , o v e r l o a d)}, \\text {(o v e r l o a d , f l u i d)}, \\text {(e d e m a ,} \\right. \\\\ \\text {p u l m o n a r y)} \\} \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 507, + 392, + 882, + 423 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nT = \\left\\{\\left(\\text {m i l d , o b s - d p , 1}\\right), \\left(\\text {f l u i d , o b s - d p , 0}\\right), \\left(\\text {o v e r - l o a d , o b s - d p , 1}\\right), \\left(\\text {o v e r t , o b s - d a , 0}\\right), \\left(\\text {p u l m o n a r y , a n a t - d p , 0}\\right), \\left(\\text {e d e m a , o b s - d a , 1}\\right) \\right\\}\n$$\n", + "text_format": "latex", + "bbox": [ + 507, + 424, + 882, + 472 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/97416453b9d6654c726b9901018e3bd8c8599e607683b9ecf6ff0d7dac451c8d.jpg", + "image_caption": [ + "Figure 4: Graph view of the RadGraph annotations for the report in Figure 3." + ], + "image_footnote": [], + "bbox": [ + 514, + 485, + 880, + 680 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "B Code and Data Release", + "text_level": 1, + "bbox": [ + 507, + 747, + 746, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our research has been carried out using the ViLMedic library (Delbrouck et al., 2022b). Our code is available at https://github.com/jbdel/vilmedic. This link is anonymized and complies with the double-blind review process. More specifically, we release the code of the RadGraph score as well as the training of our baseline. We also release the script to download, pre-process, and split the radiology reports of the MIMIC-III database", + "bbox": [ + 507, + 774, + 882, + 917 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "476", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/34e2f0ac6cd5bf87c3b5498c91b5cafc67dbf29861bc3bb08274e0e3d60f88a0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CT SpineCT SinusMR NeckMR Head
low resolution study reveals degenerative OBS-DP change OBS-OP and foraminal ANAT-OP narrowing OBS-OP without gross OBS-DA acute OBS-DA pathology OBS-DA1. sinusitis OBS-DP affecting the left ANAT-DS pheoid anat-OP and ethmoid anat-OP sinus ANAT-OP .2 . opacification OBS-OP of bilateral ANAT-OP mastoid ANAT-OP air cells and fluid OBS-OP seen in the middle ANAT-OP ear ANAT-OP cavities ANAT-OP which may indicate infection OBS-OP .slightly OBS-OP prominent OBS-OP lymph OBS-OP node OBS-OP in the posterior ANAT-OP chain ANAT-OP on the left side ANAT-OP side unchanged OBS-OP from previous examination . no definite evidence of infiltrating OBS-DA mass OBS-DA or definite pathologic adenopathy OBS-DA .1. no acute OBS-DA ischemia OBS-DA .2 . age -appropriate OBS-OP -appropriate atrophy OBS-OP , and chronic OBS-OP small OBS-OP vessel ANAT-OP ischemic OBS-OP changes OBS-OP .3 . there is no occlusion OBS-DA or flow -limiting OBS-DA - limiting stenosis OBS-DA of the arterial ANAT-OP system ANAT-OP of the head and neck
", + "bbox": [ + 119, + 80, + 877, + 195 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 5: Examples of entities detected by RadGraph (used in the $\\mathrm{RG_{ER}}$ metric) on out-of-domain anatomy/modality radiology reports. Relations are omitted for clarity.", + "bbox": [ + 112, + 205, + 882, + 233 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "as per our experiments. To download the MIMIC-III database, researchers are required to formally request access via a process documented on the MIMIC website. There are two key steps that must be completed before access is granted: (i) the researcher must complete a recognized course in protecting human research participants, including Health Insurance Portability and Accountability Act (HIPAA) requirements. (ii) the researcher must sign a data use agreement, which outlines appropriate data usage and security standards, and forbids efforts to identify individual patients.", + "bbox": [ + 112, + 259, + 489, + 451 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "C More Results", + "text_level": 1, + "bbox": [ + 112, + 464, + 268, + 479 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We present the results (including four metrics, i.e., ROUGE-1, ROUGE-2, ROUGE-L, and RadGraph scores) of all the experiments on Figure 5-9 for further research in this field. We also show the output of RadGraph (for entities) on a few samples of our new dataset in Table 5.", + "bbox": [ + 112, + 489, + 487, + 583 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "D Ethics Statement", + "text_level": 1, + "bbox": [ + 112, + 596, + 302, + 612 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The MIMIC-CXR and MIMIC-III datasets are de-identified to satisfy the US Health Insurance Portability and Accountability Act of 1996 (HIPAA) Safe Harbor requirements. Protected health information (PHI) has been removed. Therefore, the ethical approval statement and the need for informed consent were waived for the studies on this database, which was approved by the Massachusetts Institute of Technology (Cambridge, MA) and Beth Israel Deaconess Medical Center (Boston, MA). This research was conducted in accordance with the Declaration of Helsinki, describing the ethical principles of medical research involving human subjects.", + "bbox": [ + 112, + 621, + 489, + 848 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "477", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/ee33f4d0af3955c15a0fd299c144cb6820be34977e226e921b4cabcc16a93d7f.jpg", + "image_caption": [ + "ROUGE-1" + ], + "image_footnote": [], + "bbox": [ + 115, + 191, + 455, + 448 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/8a3b25fab03d88a532ab372bc72d1847fa90c4d1d60a277843c25981bf34ca3c.jpg", + "image_caption": [ + "ROUGE-2" + ], + "image_footnote": [], + "bbox": [ + 514, + 191, + 855, + 448 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/7dc05ddbe4482c7299afe712e7f22401b67815d2b9d8544b501949b0adefcf19.jpg", + "image_caption": [ + "ROUGE-L" + ], + "image_footnote": [], + "bbox": [ + 117, + 481, + 455, + 736 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/064769bd195bbfdfbc4a5af3200ac07641ef25a517951c0330bbb57cc6599299.jpg", + "image_caption": [ + "RadGraph Score", + "Figure 5: Cross-modality-anatomy results from T5-S are visualized here using heatmpas. Colors from light to dark represent the values from low to high in each column." + ], + "image_footnote": [], + "bbox": [ + 514, + 481, + 855, + 737 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "478", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/7c5bc00fd7b4a190f8a1537f7e1fbd1efb2df1611316e2587fd11f12751c4438.jpg", + "image_caption": [ + "ROUGE-1" + ], + "image_footnote": [], + "bbox": [ + 115, + 191, + 455, + 448 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/38a5b92cf5ca425c6d7f47c2742119166fdea4b13e43deee0673fe95a707fb12.jpg", + "image_caption": [ + "ROUGE-2" + ], + "image_footnote": [], + "bbox": [ + 514, + 191, + 855, + 448 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/570950fb54e59fb1aaa2dd7efcdb9a61be3e109e0a36e840befe716be5f1124e.jpg", + "image_caption": [ + "ROUGE-L" + ], + "image_footnote": [], + "bbox": [ + 117, + 481, + 455, + 736 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/d22930f2b39cc1358c376d653ad55150a3d60b85e9dd3217201ec04711ada013.jpg", + "image_caption": [ + "RadGraph Score", + "Figure 6: Cross-modality-anatomy results from BART-B are visualized here using heatmaps. Colors from light to dark represent the values from low to high in each column." + ], + "image_footnote": [], + "bbox": [ + 514, + 481, + 855, + 737 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "479", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/fa7dd4a837009612dd7b747453bc1f6037ecbe4aa426564fbf098aca89742521.jpg", + "image_caption": [ + "ROUGE-1" + ], + "image_footnote": [], + "bbox": [ + 117, + 191, + 453, + 448 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d428477be2e0fa4e49f3ce17689dfe14c2f95bdb265c7e0684d79bf4a0641d93.jpg", + "image_caption": [ + "ROUGE-2" + ], + "image_footnote": [], + "bbox": [ + 512, + 191, + 855, + 448 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/f2076082bc66504433f5ecf5dd5d1cc86629e8339887f9c2da1660ccd44be3e0.jpg", + "image_caption": [ + "ROUGE-L" + ], + "image_footnote": [], + "bbox": [ + 117, + 481, + 453, + 736 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/c6bb4b722501db9c7d686922487bf329fb3a05e8ebdc2ffa0e397584b238d8e6.jpg", + "image_caption": [ + "RadGraph Score", + "Figure 7: Cross-modality-anatomy results from BART-L are visualized here using heatmaps. Colors from light to dark represent the values from low to high in each column." + ], + "image_footnote": [], + "bbox": [ + 514, + 481, + 855, + 737 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "480", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/4c520bb5abbe22fcabc27d0bd6870d48ba950708bae7b5ea54dbf79586afc079.jpg", + "image_caption": [ + "ROUGE-1" + ], + "image_footnote": [], + "bbox": [ + 117, + 191, + 453, + 448 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/05a17ab6cdbe137561b71e15c1fa2835b1748e4f96d5e7c4d8f489d55f60bdad.jpg", + "image_caption": [ + "ROUGE-2" + ], + "image_footnote": [], + "bbox": [ + 514, + 191, + 855, + 448 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/045ea34b4c69f0539d8800597e79569ca0c1f6dfd64f841defd447710dc7015d.jpg", + "image_caption": [ + "ROUGE-L" + ], + "image_footnote": [], + "bbox": [ + 117, + 481, + 453, + 736 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a28e93aedb83d89a195a25518651ffc2bb5fdd3d196e274255a6c3bb30f7f9bd.jpg", + "image_caption": [ + "RadGraph Score", + "Figure 8: Cross-modality-anatomy results from BioBART-B are visualized here using heatmaps. Colors from light to dark represent the values from low to high in each column." + ], + "image_footnote": [], + "bbox": [ + 514, + 481, + 855, + 737 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "481", + "bbox": [ + 485, + 928, + 512, + 940 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e3e86c1bb2e64c7703a64064869fce07d9fd81f177a8d6aeec13ebd30406634f.jpg", + "image_caption": [ + "ROUGE-1" + ], + "image_footnote": [], + "bbox": [ + 117, + 191, + 453, + 448 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/08186fd00caf82df86aa3a9469e966fcfe201d0246ac986a51811a4d7c681764.jpg", + "image_caption": [ + "ROUGE-2" + ], + "image_footnote": [], + "bbox": [ + 514, + 191, + 855, + 448 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/3fbc8aab9bfa93c77c1f164ea14aa387df22ae9d8124b7a18a56f875f014bf02.jpg", + "image_caption": [ + "ROUGE-L" + ], + "image_footnote": [], + "bbox": [ + 117, + 481, + 455, + 736 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/f0a5090e77c9e5fd49ca1a9227b19307f6e367765d2763718c8d172171a135cf.jpg", + "image_caption": [ + "RadGraph Score", + "Figure 9: Cross-modality-anatomy results from BioBART-L are visualized here using heatmaps. Colors from light to dark represent the values from low to high in each column." + ], + "image_footnote": [], + "bbox": [ + 514, + 481, + 855, + 737 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "482", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "A For every submission:", + "bbox": [ + 114, + 107, + 322, + 122 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "A1. Did you describe the limitations of your work?", + "bbox": [ + 129, + 126, + 532, + 143 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "On Page 5.", + "bbox": [ + 151, + 143, + 233, + 158 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "A2. Did you discuss any potential risks of your work?", + "bbox": [ + 129, + 168, + 552, + 185 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "On Page 5.", + "bbox": [ + 151, + 186, + 233, + 200 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "A3. Do the abstract and introduction summarize the paper's main claims?", + "bbox": [ + 129, + 212, + 695, + 229 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "On Pages 1 and 4.", + "bbox": [ + 151, + 230, + 287, + 244 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "A4. Have you used AI writing assistants when working on this paper?", + "bbox": [ + 129, + 255, + 668, + 272 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 151, + 273, + 231, + 287 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "B Did you use or create scientific artifacts?", + "bbox": [ + 114, + 299, + 487, + 316 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Section 2.", + "bbox": [ + 132, + 321, + 208, + 335 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "B1. Did you cite the creators of artifacts you used?", + "bbox": [ + 129, + 346, + 529, + 363 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Section 2.", + "bbox": [ + 151, + 363, + 226, + 376 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?", + "bbox": [ + 129, + 388, + 778, + 406 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "On Page 5.", + "bbox": [ + 151, + 407, + 233, + 422 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?", + "bbox": [ + 129, + 432, + 880, + 495 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Section 2.", + "bbox": [ + 151, + 498, + 226, + 511 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?", + "bbox": [ + 129, + 523, + 880, + 571 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Section 2.", + "bbox": [ + 151, + 573, + 226, + 586 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?", + "bbox": [ + 129, + 598, + 880, + 631 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Section 2.", + "bbox": [ + 151, + 632, + 226, + 646 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be.", + "bbox": [ + 129, + 657, + 880, + 739 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Section 2.", + "bbox": [ + 151, + 740, + 226, + 753 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "C Did you run computational experiments?", + "bbox": [ + 114, + 764, + 492, + 781 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Section 3.", + "bbox": [ + 132, + 785, + 208, + 800 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?", + "bbox": [ + 127, + 813, + 880, + 845 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Not applicable. We use the common pre-trained models in our experiments.", + "bbox": [ + 149, + 846, + 705, + 860 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.", + "bbox": [ + 112, + 868, + 877, + 892 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "ACL 2023 Responsible NLP Checklist", + "bbox": [ + 132, + 84, + 433, + 99 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "483", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?", + "bbox": [ + 129, + 83, + 878, + 115 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Section 3.", + "bbox": [ + 149, + 117, + 226, + 131 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?", + "bbox": [ + 129, + 142, + 882, + 190 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Section 3.", + "bbox": [ + 149, + 192, + 226, + 205 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?", + "bbox": [ + 129, + 217, + 882, + 265 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Section 3.", + "bbox": [ + 149, + 267, + 226, + 281 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?", + "bbox": [ + 114, + 292, + 877, + 310 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Sections 2 and 3.", + "bbox": [ + 132, + 313, + 260, + 328 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?", + "bbox": [ + 129, + 338, + 882, + 372 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Section 2.", + "bbox": [ + 149, + 374, + 226, + 387 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?", + "bbox": [ + 129, + 398, + 882, + 447 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Section 2.", + "bbox": [ + 149, + 449, + 226, + 462 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?", + "bbox": [ + 129, + 473, + 882, + 521 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Section 2.", + "bbox": [ + 149, + 524, + 226, + 537 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?", + "bbox": [ + 129, + 548, + 875, + 565 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Section 2 and Page 5.", + "bbox": [ + 149, + 565, + 312, + 581 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?", + "bbox": [ + 129, + 590, + 882, + 623 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Section 2.", + "bbox": [ + 149, + 626, + 226, + 639 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "484", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 15 + } +] \ No newline at end of file diff --git a/2023/Toward Expanding the Scope of Radiology Report Summarization to Multiple Anatomies and Modalities/5f10cf4c-c72c-4642-9a79-12d04b80e5f5_model.json b/2023/Toward Expanding the Scope of Radiology Report Summarization to Multiple Anatomies and Modalities/5f10cf4c-c72c-4642-9a79-12d04b80e5f5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3371bae8edc73e542ec3e627f23f133c5e2494f5 --- /dev/null +++ b/2023/Toward Expanding the Scope of Radiology Report Summarization to Multiple Anatomies and Modalities/5f10cf4c-c72c-4642-9a79-12d04b80e5f5_model.json @@ -0,0 +1,2652 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.143, + 0.079, + 0.856, + 0.119 + ], + "angle": 0, + "content": "Toward Expanding the Scope of Radiology Report Summarization to Multiple Anatomies and Modalities" + }, + { + "type": "text", + "bbox": [ + 0.269, + 0.125, + 0.744, + 0.159 + ], + "angle": 0, + "content": "Zhihong Chen\\(^{2,3*}\\), Maya Varma\\(^{1*}\\), Xiang Wan\\(^{2,3}\\), Curtis P. Langlotz\\(^{1}\\), Jean-Benoit Delbrouck\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.418, + 0.16, + 0.586, + 0.176 + ], + "angle": 0, + "content": "\\(^{1}\\)Stanford University" + }, + { + "type": "text", + "bbox": [ + 0.297, + 0.177, + 0.706, + 0.192 + ], + "angle": 0, + "content": "2The Chinese University of Hong Kong, Shenzhen" + }, + { + "type": "text", + "bbox": [ + 0.333, + 0.193, + 0.671, + 0.209 + ], + "angle": 0, + "content": "3Shenzhen Research Institute of Big Data" + }, + { + "type": "text", + "bbox": [ + 0.265, + 0.211, + 0.739, + 0.225 + ], + "angle": 0, + "content": "zhihongchen@link.cuhk.edu.cn wanxiang@sribd.cn" + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.227, + 0.689, + 0.242 + ], + "angle": 0, + "content": "{mvarma2,langlotz,jbdel}@stanford.edu" + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.253, + 0.341, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.282, + 0.461, + 0.581 + ], + "angle": 0, + "content": "Radiology report summarization (RRS) is a growing area of research. Given the Findings section of a radiology report, the goal is to generate a summary (called an Impression section) that highlights the key observations and conclusions of the radiology study. However, RRS currently faces essential limitations. First, many prior studies conduct experiments on private datasets, preventing the reproduction of results and fair comparisons across different systems and solutions. Second, most prior approaches are evaluated solely on chest X-rays. To address these limitations, we propose a dataset (MIMIC-RRS) involving three new modalities and seven new anatomies based on the MIMIC-III and MIMIC-CXR datasets. We then conduct extensive experiments to evaluate the performance of models both within and across modality-anatomy pairs in MIMIC-RRS. In addition, we evaluate their clinical efficacy via RadGraph, a factual correctness metric." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.595, + 0.26, + 0.61 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.622, + 0.49, + 0.896 + ], + "angle": 0, + "content": "A radiology report is a document that provides information about the results of a radiology study. It usually includes a Findings section with key observations from the study and an Impression section with the radiologist's overall conclusions. The latter is the most critical part of the report and is typically based on both the findings and the patient's condition. It can be helpful to automate the process of generating the impression section because it can be time-consuming and prone to errors when done manually (Bhargavan et al., 2009; Alexander et al., 2022). Recently, substantial progress has been made towards research on automated radiology report summarization (RRS) (Zhang et al., 2020; Ben Abacha et al., 2021; Hu et al., 2022). However, the field of RRS faces several key limitations. First, the experimental results of many" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.253, + 0.885, + 0.493 + ], + "angle": 0, + "content": "prior studies (Zhang et al., 2018, 2020) are reported on private datasets, making it difficult to replicate results or compare approaches. Second, existing studies are mainly limited to a single modality (i.e., X-ray) and a single anatomy (i.e., chest) (Zhang et al., 2020; Ben Abacha et al., 2021; Hu et al., 2021). In some cases, researchers omit to disclose the modality and anatomy of the radiology reports used for their experiments (Karn et al., 2022). Finally, recent models (Karn et al., 2022; Hu et al., 2022) present an increased complexity in architecture that offers only marginal improvements on the existing evaluation metrics for summarization. This further makes the replication of studies more difficult." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.495, + 0.885, + 0.687 + ], + "angle": 0, + "content": "To address the aforementioned limitations, we construct a brand-new open-source dataset (named MIMIC-RRS) for radiology report summarization involving three modalities (X-ray, MRI, and CT) and seven anatomies (chest, head, neck, sinus, spine, abdomen, and pelvis). MIMIC-RRS is based on the MIMIC-CXR (Johnson et al., 2019) and MIMIC-III (Johnson et al., 2016) datasets and introduces data from 12 new modality-anatomy pairs. As a result, we introduce a new setting for evaluating the generalization capabilities of RRS models across different modalities and anatomies." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.689, + 0.885, + 0.817 + ], + "angle": 0, + "content": "In addition, we benchmark various pre-trained language models on MIMIC-RRS. Through extensive experiments within and across modality-anatomy pairs, we show that adopting an appropriate pretrained model can achieve promising results comparable to previous studies. We also introduce a metric to evaluate factual correctness of generated summaries for any modality-anatomy pair." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.83, + 0.729, + 0.844 + ], + "angle": 0, + "content": "2 Dataset Construction" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.855, + 0.885, + 0.919 + ], + "angle": 0, + "content": "In this section, we present the new MIMIC-RRS dataset designed for radiology report summarization across multiple modalities and anatomies. Comparisons with existing datasets are shown in" + }, + { + "type": "page_footnote", + "bbox": [ + 0.137, + 0.905, + 0.266, + 0.919 + ], + "angle": 0, + "content": "*Equal Contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.928, + 0.516, + 0.941 + ], + "angle": 0, + "content": "469" + }, + { + "type": "footer", + "bbox": [ + 0.227, + 0.946, + 0.77, + 0.958 + ], + "angle": 0, + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + }, + { + "type": "footer", + "bbox": [ + 0.377, + 0.959, + 0.621, + 0.972 + ], + "angle": 0, + "content": "Volume 2: Short Papers, pages 469-484" + }, + { + "type": "footer", + "bbox": [ + 0.297, + 0.973, + 0.702, + 0.985 + ], + "angle": 0, + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.182, + 0.082, + 0.818, + 0.235 + ], + "angle": 0, + "content": "
DatasetAnatomyModalityLanguageNumber
Zhang et al. (2018)MultipleMultipleEnglish87,127
Zhang et al. (2020)MultipleMultipleEnglish130,850
RIH (Zhang et al., 2020)MultipleMultipleEnglish139,654
OpenI (Demner-Fushman et al., 2016)ChestX-rayEnglish3,268
MIMIC-CXR (Johnson et al., 2019)ChestX-rayEnglish128,003
PadChest (Bustos et al., 2020)ChestX-raySpanish206,222
MIMIC-RRS (ours)MultipleMultipleEnglish207,782
" + }, + { + "type": "table_caption", + "bbox": [ + 0.225, + 0.245, + 0.772, + 0.26 + ], + "angle": 0, + "content": "Table 1: Comparisons with existing datasets for radiology report summarization." + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.285, + 0.489, + 0.318 + ], + "angle": 0, + "content": "Table 1. We detail the collection process and the dataset statistics in the following subsections." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.332, + 0.284, + 0.347 + ], + "angle": 0, + "content": "2.1 Data Collection" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.355, + 0.49, + 0.676 + ], + "angle": 0, + "content": "MIMIC-III One of our main contributions is to generate RRS data from MIMIC-II involving distinct combinations of modalities (i.e., medical imaging techniques) and anatomies (i.e., body parts). To this end, we first select five of the most frequently-occurring modality-anatomy pairs in the pool of MIMIC-III reports: \"CT Head\", \"CT Spine\", \"CT Chest\", \"CT Abdomen-Pelvis\" and \"MR Head\". Note that we discard chest X-rays as they are included in the MIMIC-CXR dataset. In addition, we pick six modality-anatomy pairs that occur infrequently in MIMIC-III to serve as out-of-domain (OOD) test sets: \"CT Neck\", \"CT Sinus\", \"MR Pelvis\", \"MR Neck\", \"MR Abdomen\", \"MR Spine\". This set of pairs represents two types of OOD cases: (1) the modality has not been seen during training (one could train on CT neck and test on MR Neck), and (2) the anatomy has not been seen during training (for example, CT Sinus is the only \"sinus\" dataset)." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.679, + 0.49, + 0.92 + ], + "angle": 0, + "content": "For each report, we extract the findings and impression section. However, the findings section is not always clearly labeled as \"findings\". With the help of a board-certified radiologist, we identify alternate section headers that reference findings for each modality-anatomy pair. As an example, for CT head, findings may be referenced in reports with the section headings \"non-contrast head ct\", \"ct head\", \"ct head without contrast\", \"ct head without iv contrast\", \"head ct\", \"head ct without iv contrast\", or \"cta head\". We identify 537 candidate section headers that reference findings across our dataset. We also discarded reports where multiple studies are pooled in the same radiology report, leading to multiple intricate observations in the impression" + }, + { + "type": "table", + "bbox": [ + 0.552, + 0.282, + 0.848, + 0.447 + ], + "angle": 0, + "content": "
CT Abd-pelvCT ChestCT Head
15,98912,78631,402
CT SpineMR HeadCT Neck
5,5177,3131,140
CT SinusMR SpineMR Abdomen
1,2672,8211,061
MR NeckMR PelvisX-ray Chest
230253128,003
" + }, + { + "type": "table_caption", + "bbox": [ + 0.508, + 0.456, + 0.886, + 0.501 + ], + "angle": 0, + "content": "Table 2: Dataset statistics for MIMIC-RRS. We report the number of radiology reports from each modality-anatomy pair." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.523, + 0.884, + 0.557 + ], + "angle": 0, + "content": "section1. Our resulting dataset consists of 79,779 selected reports across 11 modality-anatomy pairs." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.564, + 0.884, + 0.66 + ], + "angle": 0, + "content": "MIMIC-CXR MIMIC-CXR studies are chest X-ray examinations. We follow preprocessing steps reported in previous work (Delbrouck et al., 2022b), and we only include reports with both a Findings and an Impression section. This yields 128,003 reports." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.672, + 0.667, + 0.685 + ], + "angle": 0, + "content": "2.2 Data statistics" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.692, + 0.885, + 0.886 + ], + "angle": 0, + "content": "In total, there are 207,782 samples in the MIMIC-RRS dataset. The number of examples for each modality and anatomy is provided in Table 2. To further analyze this dataset, we report in Figure 1 the text lengths and vocabulary sizes associated with reports from each modality-anatomy pair. We find that for all modality-anatomy pairs, the findings section is significantly longer than the impression section (up to \\(+315\\%\\) for MR abdomen). Additionally, the findings sections of chest X-ray reports, which average only 49 words, are much shorter than reports from other modality-anatomy" + }, + { + "type": "page_footnote", + "bbox": [ + 0.509, + 0.893, + 0.883, + 0.919 + ], + "angle": 0, + "content": "1We release our candidate section headers as well as code to recreate the dataset from scratch (Appendix B)." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.517, + 0.941 + ], + "angle": 0, + "content": "470" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.08, + 0.488, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.209, + 0.487, + 0.239 + ], + "angle": 0, + "content": "Figure 1: Section length and vocabulary size for reports from each modality-anatomy pair." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.262, + 0.491, + 0.439 + ], + "angle": 0, + "content": "pairs. In contrast, MR Abdomen and MR Pelvis reports including findings sections that average 205 and 174 words, respectively. We see that CT Chest, CT Head, and CT Abdomen-Pelvis reports have a relatively large vocabulary size (given their sample size) with 20,909, 19,813, and 18,933 words. Surprisingly, the CT Abdomen-Pelvis impressions include a larger vocabulary than the findings. On the other hand, MR pelvis and MR abdomen impressions contain \\(36\\%\\) and \\(37\\%\\) fewer words than their corresponding findings, respectively." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.44, + 0.49, + 0.567 + ], + "angle": 0, + "content": "We assign reports from the following modality-anatomy pairs to training, validation, and test sets due to their large sample sizes: \"CT abdomen/pelvis\", \"CT Chest\", \"CT Neck\", \"CT Spine\", \"CT Head\", \"MR Head\", and \"X-ray Chest\". The remaining reports (i.e., \"MR Pelvis\", \"MR Spine\", \"MR Neck\", \"MR Abdomen\", and \"CT Sinus\") are used for OOD test sets\\(^2\\)." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.579, + 0.334, + 0.596 + ], + "angle": 0, + "content": "3 Algorithmic Analysis" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.604, + 0.49, + 0.685 + ], + "angle": 0, + "content": "In this section, we conduct experiments to analyze the performance of different models on MIMIC-RRS. We provide three categories of analyses: inmodality-anatomy, cross-modality-anatomy, and clinical efficacy." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.696, + 0.33, + 0.712 + ], + "angle": 0, + "content": "3.1 In-modality-anatomy" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.716, + 0.491, + 0.86 + ], + "angle": 0, + "content": "To benchmark the performance of different models on the proposed MIMIC-RRS dataset, we conduct experiments within each modality-anatomy pair (i.e., the training and test procedures are performed using only one modality-anatomy pair). We evaluate three types of pre-trained sequence-to-sequence models, namely T5 (Raffel et al., 2020), BART (Lewis et al., 2020), BioBART (Yuan et al., 2022), and their variants. Results are reported in" + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.085, + 0.572, + 0.098 + ], + "angle": 0, + "content": "Table 3." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.101, + 0.885, + 0.438 + ], + "angle": 0, + "content": "Several observations can be drawn from these experiments. First, simply adopting pretrained sequence-to-sequence language models can achieve results comparable to previous state-of-the-art approaches designed for radiology summarization. Indeed, using BART-L as a backbone achieves the best performance, confirming the necessity of exploiting appropriate pre-trained language models. Secondly, the performances across different model types vary (i.e., BART-L/BART-B, BioBART-L/BioBART-B). Yet, we notice that the number of training parameters matters; large models report the best results. According to our evaluations, the BART models achieve better results across all modality-anatomy pairs. Surprisingly, it is worth noting that the BioBART models do not achieve better results than BART, although BioBART is pre-trained on a biomedical corpus. One explanation could be that BioBART models are pre-trained on abstracts from PubMed, which are not within the same domain as radiology reports." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.439, + 0.884, + 0.552 + ], + "angle": 0, + "content": "In summary, we note several key findings for future studies: (i) \"Less is more\": starting from an appropriate backbone instead of designing complicated modules; (ii) the model size matters; (iii) the pretraining domain matters: knowledge from clinical notes or medical literature does not easily translate to radiology reports." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.562, + 0.753, + 0.578 + ], + "angle": 0, + "content": "3.2 Cross-modality-anatomy" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.583, + 0.884, + 0.841 + ], + "angle": 0, + "content": "In this section, we conduct experiments across modality-anatomy pairs (i.e., models are trained on reports from a subset of modality-anatomy pairs and then evaluated on all pairs, including the OOD test sets). We report the cross-modality-anatomy scores in Figure 2. A few interesting observations can be made. First, there are some associations between different anatomies and modalities. For example, the model trained on \"CT Head\" can also achieve promising results on the \"MR Head\" set. Secondly, training the model with all the modality-anatomy pairs (denoted as ALL) achieves the best generalization, obtaining the best results across all modalities and anatomies including the OOD test sets. We leave further exploration of cross-modality-anatomy associations and zero-shot OOD" + }, + { + "type": "page_footnote", + "bbox": [ + 0.508, + 0.846, + 0.883, + 0.919 + ], + "angle": 0, + "content": "et al., 2019), and Clinical-T5 (Lu et al., 2022)) that specialize in the clinical text since they were trained on the text from MIMIC-III, which overlaps with our dataset. The MIMIC-RRS test set is included in their pre-training data. Thus, we do not adopt them in our experiments to avoid potential data leakage and ensure a fair comparison." + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.868, + 0.488, + 0.918 + ], + "angle": 0, + "content": "2We release data splits publicly so that future work can fairly compare new results. 3We do not evaluate several pre-trained models (e.g., ClinicalBERT (Alsentzer et al., 2019), BioClinicalBERT (Alsentzer" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.514, + 0.941 + ], + "angle": 0, + "content": "471" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.122, + 0.082, + 0.885, + 0.201 + ], + "angle": 0, + "content": "
ModelsMR HeadCT SpineCT NeckCT HeadCT ChestCT Abd/PelX-ray Chest
R1R2RLR1R2RLR1R2RLR1R2RLR1R2RLR1R2RLR1R2RL
WGSum------------------48.433.346.7
AIG-CL------------------51.035.246.7
T5-S38.218.328.535.818.628.939.020.029.143.125.336.539.518.529.328.910.621.247.832.243.5
BioBART-B42.421.232.047.827.940.040.419.629.346.027.438.941.419.130.333.112.523.249.633.845.3
BioBART-L42.121.432.647.828.140.840.319.429.645.526.738.640.217.828.932.511.722.649.333.344.9
BART-B42.021.532.149.029.741.641.420.930.246.428.139.541.619.530.633.112.923.651.034.946.4
BART-L43.722.132.849.829.741.442.020.530.446.627.339.041.818.629.633.912.423.251.734.946.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.21, + 0.884, + 0.255 + ], + "angle": 0, + "content": "Table 3: The benchmarking comparisons of different approaches, including task-specific models (i.e., WGSum (Hu et al., 2021) and AIG-CL (Hu et al., 2022)) and pre-trained language models (i.e., T5-S, BioBART-B, BioBART-L, BART-B, and BART-L). R1, R2, and RL denote ROUGE-1, ROUGE-2, and ROUGE-L, respectively." + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.269, + 0.361, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.269, + 0.605, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.269, + 0.882, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.512, + 0.884, + 0.556 + ], + "angle": 0, + "content": "Figure 2: Cross-modality-anatomy results from BART-L are visualized here using heatmaps. Colors from light to dark represent the values from low to high in each column. As discussed in Section 3.2, the model variant \"ALL\" reports the strongest performances." + }, + { + "type": "table", + "bbox": [ + 0.119, + 0.577, + 0.487, + 0.681 + ], + "angle": 0, + "content": "
T5-SBioBART-BBioBART-LBART-BBART-L
MR Head21.524.825.325.026.1
CT Spine23.837.037.038.538.3
CT Neck21.223.623.624.024.9
CT Head31.834.234.035.234.7
CT Chest24.026.024.326.025.2
CT Abd/Pel12.615.915.316.115.9
X-ray Chest39.840.941.042.343.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.69, + 0.487, + 0.719 + ], + "angle": 0, + "content": "Table 4: F1-RadGraph scores on MIMIC-RRS test sets across different anatomies and modalities." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.744, + 0.295, + 0.759 + ], + "angle": 0, + "content": "transfer for future work." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.77, + 0.292, + 0.787 + ], + "angle": 0, + "content": "3.3 Clinical-Efficacy" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.791, + 0.491, + 0.92 + ], + "angle": 0, + "content": "In addition to evaluating our systems using the ROUGE-1, ROUGE-2, and ROUGE-L metrics (Lin, 2004), we use a factual correctness metric to analyze clinical efficacy. Most prior works (Zhang et al., 2020; Smit et al., 2020; Hu et al., 2022) mainly use the \\(\\mathrm{F_1}\\) CheXbert metric, an F1-score that evaluates the factual correctness of the generated impressions using 14 chest radio" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.58, + 0.885, + 0.645 + ], + "angle": 0, + "content": "graphic observations. Unfortunately, this metric is unsuitable for MIMIC-RRS, which contains reports from other modality-anatomy pairs beyond chest X-rays." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.646, + 0.886, + 0.92 + ], + "angle": 0, + "content": "For this reason, instead of using \\(\\mathrm{F_1}\\) CheXbert, we propose to use RadGraph (Jain et al., 2021) to evaluate the clinical correctness of the generated impressions. RadGraph is a dataset containing board-certified radiologist annotations of radiology reports corresponding to 14,579 entities and 10,889 relations (Appendix A.1). We used the released pretrained model to annotate our reports and asked one board-certified radiologist to subjectively validate that the printed entities of the RadGraph model on our data are correct (examples are shown in Table 5). After confirming the effectiveness of the model, we follow Delbrouck et al. (2022a) to compute the F1-RadGraph scores. The score evaluates the correctness of the generated named entities in the hypothesis impression compared to the ground-truth impression. We report these results in Ta" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.517, + 0.941 + ], + "angle": 0, + "content": "472" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.49, + 0.182 + ], + "angle": 0, + "content": "ble 4. It can be observed that the BART models can achieve the best performance with respect to clinical efficacy. The results are consistent with the ROUGE scores, further confirming the effectiveness of adopting BART as the backbone instead of designing complicated solutions." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.214, + 0.27, + 0.229 + ], + "angle": 0, + "content": "4 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.254, + 0.49, + 0.625 + ], + "angle": 0, + "content": "In this section, we discuss prior research related to the radiology report summarization task. The first attempt at automatic summarization of radiology findings into natural language impression statements was proposed by Zhang et al. (2018). Their contribution was to propose a first baseline on the task, using a bidirectional-LSTM as encoder and decoder. Importantly, they found that about \\(30\\%\\) of the summaries generated from neural models contained factual errors. Subsequently, Zhang et al. (2020) proposed the \\(\\mathrm{F_1}\\) CheXbert score to evaluate the factual correctness of the generated impression. They also used reinforcement learning to optimize the \\(\\mathrm{F_1}\\) CheXbert score directly. Finally, both Hu et al. (2021) and Hu et al. (2022) used the Biomedical and Clinical English Model Packages in the Stanza Python NLP Library (Zhang et al., 2021) to extract medical entities. The former study used the entities to construct a Graph Neural Network, which was used as input in their summarization pipeline. In contrast, the latter study used the entities to mask the findings during contrastive pre-training." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.63, + 0.49, + 0.919 + ], + "angle": 0, + "content": "We believe this paper is an original contribution to the aforementioned line of work. As instigated by Zhang et al. (2018), our goal is to release a new summarization corpus and baselines on new modalities and anatomies. We do so by releasing an RRS dataset with data from 11 new modality-anatomy pairs. In addition, we extend the work performed by Zhang et al. (2020) by proposing a new metric to evaluate the factual correctness and completeness of the generated impression, namely the RadGraph score. Finally, we improve on the work of Hu et al. (2021, 2022) in two ways: (1) we use semantic annotations from a pre-trained model trained using annotations from board-certified radiologists, as opposed to Stanza which leverages unsupervised biomedical and clinical text data; (2) we leverage relation annotations between entities, a feature that was not available in prior work." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.084, + 0.776, + 0.1 + ], + "angle": 0, + "content": "5 Conclusion and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.11, + 0.885, + 0.238 + ], + "angle": 0, + "content": "In this paper, we highlight and address several weaknesses associated with the radiology report summarization task. First, from a data perspective, we propose a publicly available dataset named MIMIC-RRS involving data samples from twelve modality-anatomy pairs, with 79,779 samples from MIMIC-III and 128,003 samples from MIMIC-CXR." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.255, + 0.885, + 0.351 + ], + "angle": 0, + "content": "Second, we conducted more than 40 experiments and over 400 cross-modality-anatomy evaluations to benchmark the performance of different models. We show that instead of designing complicated modules, we can start from an appropriate backbone model such as BART." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.368, + 0.885, + 0.495 + ], + "angle": 0, + "content": "Finally, we proposed an elegant and simple metric, F1-RadGraph, to evaluate the factual correctness of summaries generated for any modality and anatomy. In the future, we hope that our work broadens the scope of the radiology report summarization task and contributes to the development of reliable RRS models that generalize well to new anatomies and modalities." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.508, + 0.617, + 0.523 + ], + "angle": 0, + "content": "Limitations" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.533, + 0.885, + 0.807 + ], + "angle": 0, + "content": "We note two limitations of our paper. First, our work does not extensively evaluate all the available pre-trained models that could be suitable for this task, e.g., ELECTRA (Clark et al., 2020), BioLinkBERT (Yasunaga et al., 2022), GatorTron (Yang et al., 2022), RadBERT (Yan et al., 2022), and PubMedBERT (Gu et al., 2021). The aim of this work is not to report the strongest possible score but rather to address weaknesses of existing radiology report summarization studies (in terms of data and evaluation). Yet, we are confident our proposed solutions report a strong baseline for future work. Second, although F1-RadGraph seems like an appropriate metric to evaluate our new modalities and anatomies (and appears to be consistent with ROUGE scores), it has only been evaluated subjectively and not systematically." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.819, + 0.673, + 0.835 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.844, + 0.885, + 0.908 + ], + "angle": 0, + "content": "Maya Varma is supported by graduate fellowship awards from the Department of Defense (NDSEG) and the Knight-Hennessy Scholars program at Stanford University." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.517, + 0.941 + ], + "angle": 0, + "content": "473" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.116, + 0.085, + 0.214, + 0.099 + ], + "angle": 0, + "content": "References" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.107, + 0.49, + 0.173 + ], + "angle": 0, + "content": "Robert Alexander, Stephen Waite, Michael A Bruno, Elizabeth A Krupinski, Leonard Berlin, Stephen Macknik, and Susana Martinez-Conde. 2022. Mandating limits on workload, duty, and speed in radiology. *Radiology*, 304(2):274-282." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.182, + 0.49, + 0.25 + ], + "angle": 0, + "content": "Emily Alsentzer, John Murphy, William Boag, WeiHung Weng, Di Jindi, Tristan Naumann, and Matthew McDermott. 2019. Publicly available clinical bert embeddings. In Proceedings of the 2nd Clinical Natural Language Processing Workshop, pages 72-78." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.258, + 0.49, + 0.351 + ], + "angle": 0, + "content": "Asma Ben Abacha, Yassine Mrabet, Yuhao Zhang, Chaitanya Shivade, Curtis Langlotz, and Dina Demner-Fushman. 2021. Overview of the MEDIQA 2021 shared task on summarization in the medical domain. In Proceedings of the 20th Workshop on Biomedical Language Processing, pages 74-85, Online. Association for Computational Linguistics." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.36, + 0.49, + 0.413 + ], + "angle": 0, + "content": "Mythreyi Bhargavan, Adam H Kaye, Howard P Forman, and Jonathan H Sunshine. 2009. Workload of radiologists in united states in 2006-2007 and trends since 1991-1992. Radiology, 252(2):458-467." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.422, + 0.49, + 0.476 + ], + "angle": 0, + "content": "Aurelia Bustos, Antonio Pertusa, Jose-Maria Salinas, and Maria de la Iglesia-Vaya. 2020. Padchest: A large chest x-ray image dataset with multi-label annotated reports. Medical image analysis, 66:101797." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.485, + 0.49, + 0.538 + ], + "angle": 0, + "content": "Kevin Clark, Minh-Thang Luong, Quoc V. Le, and Christopher D. Manning. 2020. ELECTRA: Pretraining text encoders as discriminators rather than generators. In ICLR." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.548, + 0.49, + 0.614 + ], + "angle": 0, + "content": "Jean-Benoit Delbrouck, Pierre Chambon, Christian Bluethgen, Emily Tsai, Omar Almusa, and Curtis P Langlotz. 2022a. Improving the factual correctness of radiology report generation with semantic rewards. arXiv preprint arXiv:2210.12186." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.623, + 0.49, + 0.743 + ], + "angle": 0, + "content": "Jean-benoit Delbrouck, Khaled Saab, Maya Varma, Sabri Eyuboglu, Pierre Chambon, Jared Dunnmon, Juan Zambrano, Akshay Chaudhari, and Curtis Langlotz. 2022b. ViLMedic: a framework for research at the intersection of vision and language in medical AI. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 23-34, Dublin, Ireland. Association for Computational Linguistics." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.751, + 0.49, + 0.83 + ], + "angle": 0, + "content": "Dina Demner-Fushman, Marc D Kohli, Marc B Rosenman, Sonya E Shooshan, Laritza Rodriguez, Sameer Antani, George R Thoma, and Clement J McDonald. 2016. Preparing a collection of radiology examinations for distribution and retrieval. Journal of the American Medical Informatics Association, 23(2):304-310." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.84, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Yu Gu, Robert Tinn, Hao Cheng, Michael Lucas, Naoto Usuyama, Xiaodong Liu, Tristan Naumann, Jianfeng Gao, and Hoifung Poon. 2021. Domain-specific language model pretraining for biomedical natural language processing. ACM Transactions on Computing for Healthcare (HEALTH), 3(1):1-23." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.166 + ], + "angle": 0, + "content": "Jinpeng Hu, Jianling Li, Zhihong Chen, Yaling Shen, Yan Song, Xiang Wan, and Tsung-Hui Chang. 2021. Word graph guided summarization for radiology findings. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 4980-4990, Online. Association for Computational Linguistics." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.175, + 0.885, + 0.256 + ], + "angle": 0, + "content": "Jinpeng Hu, Zhuo Li, Zhihong Chen, Zhen Li, Xiang Wan, and Tsung-Hui Chang. 2022. Graph enhanced contrastive learning for radiology findings summarization. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4677-4688." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.265, + 0.885, + 0.37 + ], + "angle": 0, + "content": "Saahil Jain, Ashwin Agrawal, Adriel Saporta, Steven Truong, Du Nguyen Duong, Tan Bui, Pierre Chambon, Yuhao Zhang, Matthew Lungren, Andrew Ng, Curtis Langlotz, Pranav Rajpurkar, and Pranav Rajpurkar. 2021. Radgraph: Extracting clinical entities and relations from radiology reports. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, volume 1." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.38, + 0.885, + 0.46 + ], + "angle": 0, + "content": "Alistair EW Johnson, Tom J Pollard, Seth J Berkowitz, Nathaniel R Greenbaum, Matthew P Lungren, Chihying Deng, Roger G Mark, and Steven Horng. 2019. Mimic-cxr, a de-identified publicly available database of chest radiographs with free-text reports. Scientific data, 6(1):1-8." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.47, + 0.885, + 0.536 + ], + "angle": 0, + "content": "Alistair EW Johnson, Tom J Pollard, Lu Shen, Li-wei H Lehman, Mengling Feng, Mohammad Ghassemi, Benjamin Moody, Peter Szolovits, Leo Anthony Celi, and Roger G Mark. 2016. Mimic-iii, a freely accessible critical care database. Scientific data, 3(1):1-9." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.547, + 0.885, + 0.626 + ], + "angle": 0, + "content": "Sanjeev Kumar Karn, Ning Liu, Hinrich Schütze, and Oladimeji Farri. 2022. Differentiable multi-agent actor-critic for multi-step radiology report summarization. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1542-1553." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.636, + 0.885, + 0.73 + ], + "angle": 0, + "content": "Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871-7880." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.739, + 0.885, + 0.78 + ], + "angle": 0, + "content": "Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.789, + 0.885, + 0.843 + ], + "angle": 0, + "content": "Qiuhao Lu, Dejing Dou, and Thien Nguyen. 2022. Clinical5: A generative language model for clinical text. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 5436-5443." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.852, + 0.885, + 0.919 + ], + "angle": 0, + "content": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J Liu, et al. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(140):1-67." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.517, + 0.941 + ], + "angle": 0, + "content": "474" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.086, + 0.49, + 0.177 + ], + "angle": 0, + "content": "Akshay Smit, Saahil Jain, Pranav Rajpurkar, Anuj Parek, Andrew Y Ng, and Matthew Lungren. 2020. Combining automatic labelers and expert annotations for accurate radiology report labeling using bert. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1500-1519." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.188, + 0.49, + 0.254 + ], + "angle": 0, + "content": "An Yan, Julian McAuley, Xing Lu, Jiang Du, Eric Y Chang, Amilcare Gentili, and Chun-Nan Hsu. 2022. Radbert: Adapting transformer-based language models to radiology. Radiology: Artificial Intelligence, 4(4):e210258." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.263, + 0.488, + 0.343 + ], + "angle": 0, + "content": "Xi Yang, Nima PourNejatian, Hoo Chang Shin, Kaleb E Smith, Christopher Parisien, Colin Compas, Cheryl Martin, Mona G Flores, Ying Zhang, Tanja Magoc, et al. 2022. Gatortron: A large clinical language model to unlock patient information from unstructured electronic health records. arXiv preprint arXiv:2203.03540." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.352, + 0.488, + 0.404 + ], + "angle": 0, + "content": "Michihiro Yasunaga, Jure Leskovec, and Percy Liang. 2022. Linkbert: Pretraining language models with document links. In Association for Computational Linguistics (ACL)." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.415, + 0.488, + 0.481 + ], + "angle": 0, + "content": "Hongyi Yuan, Zheng Yuan, Ruyi Gan, Jiaxing Zhang, Yutao Xie, and Sheng Yu. 2022. Biobart: Pretraining and evaluation of a biomedical generative language model. In Proceedings of the 21st Workshop on Biomedical Language Processing, pages 97-109." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.49, + 0.488, + 0.557 + ], + "angle": 0, + "content": "Yuhao Zhang, Daisy Yi Ding, Tianpei Qian, Christopher D Manning, and Curtis P Langlotz. 2018. Learning to summarize radiology findings. In Proceedings of the Ninth International Workshop on Health Text Mining and Information Analysis, pages 204-213." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.566, + 0.488, + 0.645 + ], + "angle": 0, + "content": "Yuhao Zhang, Derek Merck, Emily Tsai, Christopher D Manning, and Curtis Langlotz. 2020. Optimizing the factual correctness of a summary: A study of summarizing radiology reports. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5108-5120." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.655, + 0.488, + 0.72 + ], + "angle": 0, + "content": "Yuhao Zhang, Yuhui Zhang, Peng Qi, Christopher D Manning, and Curtis P Langlotz. 2021. Biomedical and clinical english model packages for the stanza python nlp library. Journal of the American Medical Informatics Association, 28(9):1892-1899." + }, + { + "type": "list", + "bbox": [ + 0.115, + 0.086, + 0.49, + 0.72 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.516, + 0.94 + ], + "angle": 0, + "content": "475" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.115, + 0.085, + 0.395, + 0.101 + ], + "angle": 0, + "content": "A Details of RadGraph Scores" + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.113, + 0.409, + 0.129 + ], + "angle": 0, + "content": "A.1 The Introduction of RadGraph" + }, + { + "type": "image", + "bbox": [ + 0.118, + 0.148, + 0.486, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.246, + 0.487, + 0.276 + ], + "angle": 0, + "content": "Figure 3: Example of the RadGraph annotations. Figure taken from (Jain et al., 2021)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.294, + 0.49, + 0.455 + ], + "angle": 0, + "content": "To design our new evaluation metric, we leverage the RadGraph dataset (Jain et al., 2021) containing board-certified radiologist annotations of chest X-ray reports, which correspond to 14,579 entities and 10,889 relations. RadGraph has released a PubMedBERT model (Gu et al., 2021) pre-trained on these annotations to annotate new reports. An example of annotation can be seen in Figure 3. Before moving on to the next section, we quickly describe the concept of entities and relations:" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.467, + 0.49, + 0.596 + ], + "angle": 0, + "content": "Entities An entity is defined as a continuous span of text that can include one or more adjacent words. Entities in RadGraph center around two concepts: Anatomy and Observation. Three uncertainty levels exist for Observation, leading to four different entities: Anatomy (ANAT-DP), Observation: Definitely Present (OBS-DP), Observation: Uncertain (OBS-U), and Observation: Definitely Absent (OBS-DA)." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.608, + 0.489, + 0.658 + ], + "angle": 0, + "content": "Relations A relation is defined as a directed edge between two entities. Three levels exist: Suggestive \\( \\text{Of}(\\cdot, \\cdot) \\), Located At (\\(. \\cdot\\)), and Modify (\\(. \\cdot\\))." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.671, + 0.327, + 0.687 + ], + "angle": 0, + "content": "A.2 Metric Computation" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.693, + 0.489, + 0.789 + ], + "angle": 0, + "content": "Using the RadGraph annotation scheme and pretrained model, we designed an F-score style reward that measures the factual consistency and completeness of the generated impression (also called hypothesis impression) compared to the reference impression." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.791, + 0.49, + 0.919 + ], + "angle": 0, + "content": "To do so, we treat the RadGraph annotations of an impression as a graph \\(\\mathcal{G}(V,E)\\) with the set of nodes \\(V = \\{v_{1},v_{2},\\ldots ,v_{|V|}\\}\\) containing the entities and the set of edges \\(E = \\{e_1,e_2,\\dots ,e_{|E|}\\}\\) the relations between pairs of entities. The graph is directed, meaning that the edge \\(e = (v_{1},v_{2})\\neq\\) \\((v_{2},v_{1})\\) . An example is depicted in Figure 4. Each node or edge of the graph also has a label, which" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.883, + 0.133 + ], + "angle": 0, + "content": "we denote as \\( v_{i_L} \\) for an entity \\( i \\) (for example \"OBS-DP\" or \"ANAT\") and \\( e_{ij_L} \\) for a relation \\( e = (v_i, v_j) \\) (such as \"modified\" or \"located at\")." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.134, + 0.884, + 0.278 + ], + "angle": 0, + "content": "To design our RadGraph score, we focus on the nodes \\(V\\) and whether or not a node has a relation in \\(E\\). For a hypothesis impression \\(y\\), we create a new set of triplets \\(T_{y} = \\{(v_{i}, v_{i_{L}}, \\mathcal{R})\\}_{i=1:|V|}\\). The value \\(\\mathcal{R}\\) is 1 if \\((v_{i}, v_{j})_{j=1:|E|, i \\neq j} \\in E\\), 0 otherwise. In other words, a triplet contains an entity, the entity label, and whether or not this entity has a relation. We proceed to construct the same set for the reference report \\(\\hat{y}\\) and denote this set \\(T_{\\hat{y}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.279, + 0.884, + 0.359 + ], + "angle": 0, + "content": "Finally, our score is defined as the harmonic mean of precision and recall between the hypothesis set \\( T_{y} \\) and the reference set \\( T_{\\hat{y}} \\), giving a value between 0 and 100. As an illustration, the set \\( V \\), \\( E \\) and \\( T \\) of the graph \\( \\mathcal{G} \\) in Figure 4 are shown as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.36, + 0.883, + 0.392 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} V = \\{\\text {m i l d , f l u i d , o v e r l o a d , o v e r t , p u l m o n a r y}, \\\\ \\text {e d e m a} \\} \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.393, + 0.884, + 0.424 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} E = \\left\\{\\text {(m i l d , o v e r l o a d)}, \\text {(o v e r l o a d , f l u i d)}, \\text {(e d e m a ,} \\right. \\\\ \\text {p u l m o n a r y)} \\} \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.425, + 0.884, + 0.473 + ], + "angle": 0, + "content": "\\[\nT = \\left\\{\\left(\\text {m i l d , o b s - d p , 1}\\right), \\left(\\text {f l u i d , o b s - d p , 0}\\right), \\left(\\text {o v e r - l o a d , o b s - d p , 1}\\right), \\left(\\text {o v e r t , o b s - d a , 0}\\right), \\left(\\text {p u l m o n a r y , a n a t - d p , 0}\\right), \\left(\\text {e d e m a , o b s - d a , 1}\\right) \\right\\}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.486, + 0.882, + 0.681 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.69, + 0.883, + 0.72 + ], + "angle": 0, + "content": "Figure 4: Graph view of the RadGraph annotations for the report in Figure 3." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.749, + 0.747, + 0.764 + ], + "angle": 0, + "content": "B Code and Data Release" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.775, + 0.884, + 0.919 + ], + "angle": 0, + "content": "Our research has been carried out using the ViLMedic library (Delbrouck et al., 2022b). Our code is available at https://github.com/jbdel/vilmedic. This link is anonymized and complies with the double-blind review process. More specifically, we release the code of the RadGraph score as well as the training of our baseline. We also release the script to download, pre-process, and split the radiology reports of the MIMIC-III database" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.517, + 0.941 + ], + "angle": 0, + "content": "476" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.12, + 0.082, + 0.878, + 0.196 + ], + "angle": 0, + "content": "
CT SpineCT SinusMR NeckMR Head
low resolution study reveals degenerative OBS-DP change OBS-OP and foraminal ANAT-OP narrowing OBS-OP without gross OBS-DA acute OBS-DA pathology OBS-DA1. sinusitis OBS-DP affecting the left ANAT-DS pheoid anat-OP and ethmoid anat-OP sinus ANAT-OP .2 . opacification OBS-OP of bilateral ANAT-OP mastoid ANAT-OP air cells and fluid OBS-OP seen in the middle ANAT-OP ear ANAT-OP cavities ANAT-OP which may indicate infection OBS-OP .slightly OBS-OP prominent OBS-OP lymph OBS-OP node OBS-OP in the posterior ANAT-OP chain ANAT-OP on the left side ANAT-OP side unchanged OBS-OP from previous examination . no definite evidence of infiltrating OBS-DA mass OBS-DA or definite pathologic adenopathy OBS-DA .1. no acute OBS-DA ischemia OBS-DA .2 . age -appropriate OBS-OP -appropriate atrophy OBS-OP , and chronic OBS-OP small OBS-OP vessel ANAT-OP ischemic OBS-OP changes OBS-OP .3 . there is no occlusion OBS-DA or flow -limiting OBS-DA - limiting stenosis OBS-DA of the arterial ANAT-OP system ANAT-OP of the head and neck
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.206, + 0.883, + 0.234 + ], + "angle": 0, + "content": "Table 5: Examples of entities detected by RadGraph (used in the \\(\\mathrm{RG_{ER}}\\) metric) on out-of-domain anatomy/modality radiology reports. Relations are omitted for clarity." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.26, + 0.49, + 0.452 + ], + "angle": 0, + "content": "as per our experiments. To download the MIMIC-III database, researchers are required to formally request access via a process documented on the MIMIC website. There are two key steps that must be completed before access is granted: (i) the researcher must complete a recognized course in protecting human research participants, including Health Insurance Portability and Accountability Act (HIPAA) requirements. (ii) the researcher must sign a data use agreement, which outlines appropriate data usage and security standards, and forbids efforts to identify individual patients." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.465, + 0.269, + 0.48 + ], + "angle": 0, + "content": "C More Results" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.491, + 0.489, + 0.584 + ], + "angle": 0, + "content": "We present the results (including four metrics, i.e., ROUGE-1, ROUGE-2, ROUGE-L, and RadGraph scores) of all the experiments on Figure 5-9 for further research in this field. We also show the output of RadGraph (for entities) on a few samples of our new dataset in Table 5." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.598, + 0.303, + 0.613 + ], + "angle": 0, + "content": "D Ethics Statement" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.623, + 0.49, + 0.849 + ], + "angle": 0, + "content": "The MIMIC-CXR and MIMIC-III datasets are de-identified to satisfy the US Health Insurance Portability and Accountability Act of 1996 (HIPAA) Safe Harbor requirements. Protected health information (PHI) has been removed. Therefore, the ethical approval statement and the need for informed consent were waived for the studies on this database, which was approved by the Massachusetts Institute of Technology (Cambridge, MA) and Beth Israel Deaconess Medical Center (Boston, MA). This research was conducted in accordance with the Declaration of Helsinki, describing the ethical principles of medical research involving human subjects." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.516, + 0.941 + ], + "angle": 0, + "content": "477" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.192, + 0.456, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.249, + 0.459, + 0.315, + 0.471 + ], + "angle": 0, + "content": "ROUGE-1" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.192, + 0.856, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.459, + 0.717, + 0.471 + ], + "angle": 0, + "content": "ROUGE-2" + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.482, + 0.456, + 0.737 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.749, + 0.316, + 0.76 + ], + "angle": 0, + "content": "ROUGE-L" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.482, + 0.856, + 0.738 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.634, + 0.749, + 0.742, + 0.761 + ], + "angle": 0, + "content": "RadGraph Score" + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.777, + 0.885, + 0.809 + ], + "angle": 0, + "content": "Figure 5: Cross-modality-anatomy results from T5-S are visualized here using heatmpas. Colors from light to dark represent the values from low to high in each column." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.516, + 0.941 + ], + "angle": 0, + "content": "478" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.192, + 0.456, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.249, + 0.459, + 0.315, + 0.471 + ], + "angle": 0, + "content": "ROUGE-1" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.192, + 0.856, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.459, + 0.717, + 0.471 + ], + "angle": 0, + "content": "ROUGE-2" + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.482, + 0.456, + 0.737 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.749, + 0.316, + 0.76 + ], + "angle": 0, + "content": "ROUGE-L" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.482, + 0.856, + 0.738 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.634, + 0.749, + 0.742, + 0.761 + ], + "angle": 0, + "content": "RadGraph Score" + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.777, + 0.885, + 0.809 + ], + "angle": 0, + "content": "Figure 6: Cross-modality-anatomy results from BART-B are visualized here using heatmaps. Colors from light to dark represent the values from low to high in each column." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.516, + 0.941 + ], + "angle": 0, + "content": "479" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.192, + 0.455, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.249, + 0.459, + 0.315, + 0.471 + ], + "angle": 0, + "content": "ROUGE-1" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.192, + 0.856, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.459, + 0.717, + 0.471 + ], + "angle": 0, + "content": "ROUGE-2" + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.482, + 0.455, + 0.737 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.749, + 0.316, + 0.76 + ], + "angle": 0, + "content": "ROUGE-L" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.482, + 0.856, + 0.738 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.634, + 0.749, + 0.742, + 0.761 + ], + "angle": 0, + "content": "RadGraph Score" + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.777, + 0.885, + 0.808 + ], + "angle": 0, + "content": "Figure 7: Cross-modality-anatomy results from BART-L are visualized here using heatmaps. Colors from light to dark represent the values from low to high in each column." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.517, + 0.941 + ], + "angle": 0, + "content": "480" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.192, + 0.455, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.249, + 0.459, + 0.315, + 0.471 + ], + "angle": 0, + "content": "ROUGE-1" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.192, + 0.856, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.459, + 0.717, + 0.471 + ], + "angle": 0, + "content": "ROUGE-2" + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.482, + 0.455, + 0.737 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.749, + 0.316, + 0.76 + ], + "angle": 0, + "content": "ROUGE-L" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.482, + 0.856, + 0.738 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.634, + 0.749, + 0.742, + 0.761 + ], + "angle": 0, + "content": "RadGraph Score" + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.777, + 0.885, + 0.809 + ], + "angle": 0, + "content": "Figure 8: Cross-modality-anatomy results from BioBART-B are visualized here using heatmaps. Colors from light to dark represent the values from low to high in each column." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.514, + 0.941 + ], + "angle": 0, + "content": "481" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.192, + 0.455, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.249, + 0.459, + 0.315, + 0.471 + ], + "angle": 0, + "content": "ROUGE-1" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.192, + 0.856, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.459, + 0.717, + 0.471 + ], + "angle": 0, + "content": "ROUGE-2" + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.482, + 0.456, + 0.737 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.749, + 0.316, + 0.76 + ], + "angle": 0, + "content": "ROUGE-L" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.482, + 0.856, + 0.738 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.634, + 0.749, + 0.742, + 0.761 + ], + "angle": 0, + "content": "RadGraph Score" + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.777, + 0.885, + 0.808 + ], + "angle": 0, + "content": "Figure 9: Cross-modality-anatomy results from BioBART-L are visualized here using heatmaps. Colors from light to dark represent the values from low to high in each column." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.516, + 0.941 + ], + "angle": 0, + "content": "482" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.133, + 0.085, + 0.435, + 0.1 + ], + "angle": 0, + "content": "ACL 2023 Responsible NLP Checklist" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.108, + 0.323, + 0.123 + ], + "angle": 0, + "content": "A For every submission:" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.127, + 0.533, + 0.144 + ], + "angle": 0, + "content": "A1. Did you describe the limitations of your work?" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.145, + 0.235, + 0.159 + ], + "angle": 0, + "content": "On Page 5." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.169, + 0.553, + 0.186 + ], + "angle": 0, + "content": "A2. Did you discuss any potential risks of your work?" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.187, + 0.235, + 0.202 + ], + "angle": 0, + "content": "On Page 5." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.213, + 0.696, + 0.23 + ], + "angle": 0, + "content": "A3. Do the abstract and introduction summarize the paper's main claims?" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.231, + 0.288, + 0.245 + ], + "angle": 0, + "content": "On Pages 1 and 4." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.256, + 0.669, + 0.273 + ], + "angle": 0, + "content": "A4. Have you used AI writing assistants when working on this paper?" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.274, + 0.232, + 0.288 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.3, + 0.489, + 0.317 + ], + "angle": 0, + "content": "B Did you use or create scientific artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.322, + 0.209, + 0.336 + ], + "angle": 0, + "content": "Section 2." + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.347, + 0.53, + 0.364 + ], + "angle": 0, + "content": "B1. Did you cite the creators of artifacts you used?" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.365, + 0.227, + 0.378 + ], + "angle": 0, + "content": "Section 2." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.39, + 0.779, + 0.407 + ], + "angle": 0, + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.408, + 0.235, + 0.423 + ], + "angle": 0, + "content": "On Page 5." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.433, + 0.882, + 0.497 + ], + "angle": 0, + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.499, + 0.227, + 0.512 + ], + "angle": 0, + "content": "Section 2." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.524, + 0.882, + 0.573 + ], + "angle": 0, + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.574, + 0.227, + 0.587 + ], + "angle": 0, + "content": "Section 2." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.599, + 0.882, + 0.632 + ], + "angle": 0, + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.633, + 0.227, + 0.647 + ], + "angle": 0, + "content": "Section 2." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.658, + 0.882, + 0.74 + ], + "angle": 0, + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be." + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.741, + 0.227, + 0.754 + ], + "angle": 0, + "content": "Section 2." + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.765, + 0.494, + 0.782 + ], + "angle": 0, + "content": "C Did you run computational experiments?" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.787, + 0.209, + 0.801 + ], + "angle": 0, + "content": "Section 3." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.814, + 0.882, + 0.846 + ], + "angle": 0, + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.847, + 0.706, + 0.862 + ], + "angle": 0, + "content": "Not applicable. We use the common pre-trained models in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.869, + 0.878, + 0.893 + ], + "angle": 0, + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.516, + 0.941 + ], + "angle": 0, + "content": "483" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.13, + 0.084, + 0.88, + 0.116 + ], + "angle": 0, + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.118, + 0.227, + 0.132 + ], + "angle": 0, + "content": "Section 3." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.143, + 0.884, + 0.191 + ], + "angle": 0, + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.193, + 0.227, + 0.206 + ], + "angle": 0, + "content": "Section 3." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.218, + 0.884, + 0.266 + ], + "angle": 0, + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.268, + 0.227, + 0.282 + ], + "angle": 0, + "content": "Section 3." + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.293, + 0.878, + 0.311 + ], + "angle": 0, + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.315, + 0.262, + 0.329 + ], + "angle": 0, + "content": "Sections 2 and 3." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.34, + 0.884, + 0.373 + ], + "angle": 0, + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.375, + 0.227, + 0.388 + ], + "angle": 0, + "content": "Section 2." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.399, + 0.884, + 0.448 + ], + "angle": 0, + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.45, + 0.227, + 0.463 + ], + "angle": 0, + "content": "Section 2." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.474, + 0.884, + 0.523 + ], + "angle": 0, + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.525, + 0.227, + 0.538 + ], + "angle": 0, + "content": "Section 2." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.549, + 0.876, + 0.566 + ], + "angle": 0, + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.567, + 0.314, + 0.582 + ], + "angle": 0, + "content": "Section 2 and Page 5." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.592, + 0.884, + 0.624 + ], + "angle": 0, + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.627, + 0.227, + 0.64 + ], + "angle": 0, + "content": "Section 2." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.516, + 0.941 + ], + "angle": 0, + "content": "484" + } + ] +] \ No newline at end of file diff --git a/2023/Toward Expanding the Scope of Radiology Report Summarization to Multiple Anatomies and Modalities/5f10cf4c-c72c-4642-9a79-12d04b80e5f5_origin.pdf b/2023/Toward Expanding the Scope of Radiology Report Summarization to Multiple Anatomies and Modalities/5f10cf4c-c72c-4642-9a79-12d04b80e5f5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..995dd3143a4a2158c33b800cb5bce5b4b57f24e3 --- /dev/null +++ b/2023/Toward Expanding the Scope of Radiology Report Summarization to Multiple Anatomies and Modalities/5f10cf4c-c72c-4642-9a79-12d04b80e5f5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1c5f974b9a1f4aeaa96c90904af5d28f3ddc51312e5b95c8f231b73069782bf8 +size 6206213 diff --git a/2023/Toward Expanding the Scope of Radiology Report Summarization to Multiple Anatomies and Modalities/full.md b/2023/Toward Expanding the Scope of Radiology Report Summarization to Multiple Anatomies and Modalities/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0b25c081a1f8ee2fad9dc04ace4cc63db05edcea --- /dev/null +++ b/2023/Toward Expanding the Scope of Radiology Report Summarization to Multiple Anatomies and Modalities/full.md @@ -0,0 +1,394 @@ +# Toward Expanding the Scope of Radiology Report Summarization to Multiple Anatomies and Modalities + +Zhihong Chen $^{2,3*}$ , Maya Varma $^{1*}$ , Xiang Wan $^{2,3}$ , Curtis P. Langlotz $^{1}$ , Jean-Benoit Delbrouck $^{1*}$ + +$^{1}$ Stanford University + +2The Chinese University of Hong Kong, Shenzhen + +3Shenzhen Research Institute of Big Data + +zhihongchen@link.cuhk.edu.cn wanxiang@sribd.cn + +{mvarma2,langlotz,jbdel}@stanford.edu + +# Abstract + +Radiology report summarization (RRS) is a growing area of research. Given the Findings section of a radiology report, the goal is to generate a summary (called an Impression section) that highlights the key observations and conclusions of the radiology study. However, RRS currently faces essential limitations. First, many prior studies conduct experiments on private datasets, preventing the reproduction of results and fair comparisons across different systems and solutions. Second, most prior approaches are evaluated solely on chest X-rays. To address these limitations, we propose a dataset (MIMIC-RRS) involving three new modalities and seven new anatomies based on the MIMIC-III and MIMIC-CXR datasets. We then conduct extensive experiments to evaluate the performance of models both within and across modality-anatomy pairs in MIMIC-RRS. In addition, we evaluate their clinical efficacy via RadGraph, a factual correctness metric. + +# 1 Introduction + +A radiology report is a document that provides information about the results of a radiology study. It usually includes a Findings section with key observations from the study and an Impression section with the radiologist's overall conclusions. The latter is the most critical part of the report and is typically based on both the findings and the patient's condition. It can be helpful to automate the process of generating the impression section because it can be time-consuming and prone to errors when done manually (Bhargavan et al., 2009; Alexander et al., 2022). Recently, substantial progress has been made towards research on automated radiology report summarization (RRS) (Zhang et al., 2020; Ben Abacha et al., 2021; Hu et al., 2022). However, the field of RRS faces several key limitations. First, the experimental results of many + +prior studies (Zhang et al., 2018, 2020) are reported on private datasets, making it difficult to replicate results or compare approaches. Second, existing studies are mainly limited to a single modality (i.e., X-ray) and a single anatomy (i.e., chest) (Zhang et al., 2020; Ben Abacha et al., 2021; Hu et al., 2021). In some cases, researchers omit to disclose the modality and anatomy of the radiology reports used for their experiments (Karn et al., 2022). Finally, recent models (Karn et al., 2022; Hu et al., 2022) present an increased complexity in architecture that offers only marginal improvements on the existing evaluation metrics for summarization. This further makes the replication of studies more difficult. + +To address the aforementioned limitations, we construct a brand-new open-source dataset (named MIMIC-RRS) for radiology report summarization involving three modalities (X-ray, MRI, and CT) and seven anatomies (chest, head, neck, sinus, spine, abdomen, and pelvis). MIMIC-RRS is based on the MIMIC-CXR (Johnson et al., 2019) and MIMIC-III (Johnson et al., 2016) datasets and introduces data from 12 new modality-anatomy pairs. As a result, we introduce a new setting for evaluating the generalization capabilities of RRS models across different modalities and anatomies. + +In addition, we benchmark various pre-trained language models on MIMIC-RRS. Through extensive experiments within and across modality-anatomy pairs, we show that adopting an appropriate pretrained model can achieve promising results comparable to previous studies. We also introduce a metric to evaluate factual correctness of generated summaries for any modality-anatomy pair. + +# 2 Dataset Construction + +In this section, we present the new MIMIC-RRS dataset designed for radiology report summarization across multiple modalities and anatomies. Comparisons with existing datasets are shown in + +
DatasetAnatomyModalityLanguageNumber
Zhang et al. (2018)MultipleMultipleEnglish87,127
Zhang et al. (2020)MultipleMultipleEnglish130,850
RIH (Zhang et al., 2020)MultipleMultipleEnglish139,654
OpenI (Demner-Fushman et al., 2016)ChestX-rayEnglish3,268
MIMIC-CXR (Johnson et al., 2019)ChestX-rayEnglish128,003
PadChest (Bustos et al., 2020)ChestX-raySpanish206,222
MIMIC-RRS (ours)MultipleMultipleEnglish207,782
+ +# 2.1 Data Collection + +MIMIC-III One of our main contributions is to generate RRS data from MIMIC-II involving distinct combinations of modalities (i.e., medical imaging techniques) and anatomies (i.e., body parts). To this end, we first select five of the most frequently-occurring modality-anatomy pairs in the pool of MIMIC-III reports: "CT Head", "CT Spine", "CT Chest", "CT Abdomen-Pelvis" and "MR Head". Note that we discard chest X-rays as they are included in the MIMIC-CXR dataset. In addition, we pick six modality-anatomy pairs that occur infrequently in MIMIC-III to serve as out-of-domain (OOD) test sets: "CT Neck", "CT Sinus", "MR Pelvis", "MR Neck", "MR Abdomen", "MR Spine". This set of pairs represents two types of OOD cases: (1) the modality has not been seen during training (one could train on CT neck and test on MR Neck), and (2) the anatomy has not been seen during training (for example, CT Sinus is the only "sinus" dataset). + +For each report, we extract the findings and impression section. However, the findings section is not always clearly labeled as "findings". With the help of a board-certified radiologist, we identify alternate section headers that reference findings for each modality-anatomy pair. As an example, for CT head, findings may be referenced in reports with the section headings "non-contrast head ct", "ct head", "ct head without contrast", "ct head without iv contrast", "head ct", "head ct without iv contrast", or "cta head". We identify 537 candidate section headers that reference findings across our dataset. We also discarded reports where multiple studies are pooled in the same radiology report, leading to multiple intricate observations in the impression + +Table 1: Comparisons with existing datasets for radiology report summarization. +Table 1. We detail the collection process and the dataset statistics in the following subsections. + +
CT Abd-pelvCT ChestCT Head
15,98912,78631,402
CT SpineMR HeadCT Neck
5,5177,3131,140
CT SinusMR SpineMR Abdomen
1,2672,8211,061
MR NeckMR PelvisX-ray Chest
230253128,003
+ +Table 2: Dataset statistics for MIMIC-RRS. We report the number of radiology reports from each modality-anatomy pair. + +section1. Our resulting dataset consists of 79,779 selected reports across 11 modality-anatomy pairs. + +MIMIC-CXR MIMIC-CXR studies are chest X-ray examinations. We follow preprocessing steps reported in previous work (Delbrouck et al., 2022b), and we only include reports with both a Findings and an Impression section. This yields 128,003 reports. + +# 2.2 Data statistics + +In total, there are 207,782 samples in the MIMIC-RRS dataset. The number of examples for each modality and anatomy is provided in Table 2. To further analyze this dataset, we report in Figure 1 the text lengths and vocabulary sizes associated with reports from each modality-anatomy pair. We find that for all modality-anatomy pairs, the findings section is significantly longer than the impression section (up to $+315\%$ for MR abdomen). Additionally, the findings sections of chest X-ray reports, which average only 49 words, are much shorter than reports from other modality-anatomy + +![](images/3e0061ba8be51b6cb07e2785a6215aeb7af3386453565b0e4de0bf8819bd6c4a.jpg) +Figure 1: Section length and vocabulary size for reports from each modality-anatomy pair. + +pairs. In contrast, MR Abdomen and MR Pelvis reports including findings sections that average 205 and 174 words, respectively. We see that CT Chest, CT Head, and CT Abdomen-Pelvis reports have a relatively large vocabulary size (given their sample size) with 20,909, 19,813, and 18,933 words. Surprisingly, the CT Abdomen-Pelvis impressions include a larger vocabulary than the findings. On the other hand, MR pelvis and MR abdomen impressions contain $36\%$ and $37\%$ fewer words than their corresponding findings, respectively. + +We assign reports from the following modality-anatomy pairs to training, validation, and test sets due to their large sample sizes: "CT abdomen/pelvis", "CT Chest", "CT Neck", "CT Spine", "CT Head", "MR Head", and "X-ray Chest". The remaining reports (i.e., "MR Pelvis", "MR Spine", "MR Neck", "MR Abdomen", and "CT Sinus") are used for OOD test sets $^2$ . + +# 3 Algorithmic Analysis + +In this section, we conduct experiments to analyze the performance of different models on MIMIC-RRS. We provide three categories of analyses: inmodality-anatomy, cross-modality-anatomy, and clinical efficacy. + +# 3.1 In-modality-anatomy + +To benchmark the performance of different models on the proposed MIMIC-RRS dataset, we conduct experiments within each modality-anatomy pair (i.e., the training and test procedures are performed using only one modality-anatomy pair). We evaluate three types of pre-trained sequence-to-sequence models, namely T5 (Raffel et al., 2020), BART (Lewis et al., 2020), BioBART (Yuan et al., 2022), and their variants. Results are reported in + +# Table 3. + +Several observations can be drawn from these experiments. First, simply adopting pretrained sequence-to-sequence language models can achieve results comparable to previous state-of-the-art approaches designed for radiology summarization. Indeed, using BART-L as a backbone achieves the best performance, confirming the necessity of exploiting appropriate pre-trained language models. Secondly, the performances across different model types vary (i.e., BART-L/BART-B, BioBART-L/BioBART-B). Yet, we notice that the number of training parameters matters; large models report the best results. According to our evaluations, the BART models achieve better results across all modality-anatomy pairs. Surprisingly, it is worth noting that the BioBART models do not achieve better results than BART, although BioBART is pre-trained on a biomedical corpus. One explanation could be that BioBART models are pre-trained on abstracts from PubMed, which are not within the same domain as radiology reports. + +In summary, we note several key findings for future studies: (i) "Less is more": starting from an appropriate backbone instead of designing complicated modules; (ii) the model size matters; (iii) the pretraining domain matters: knowledge from clinical notes or medical literature does not easily translate to radiology reports. + +# 3.2 Cross-modality-anatomy + +In this section, we conduct experiments across modality-anatomy pairs (i.e., models are trained on reports from a subset of modality-anatomy pairs and then evaluated on all pairs, including the OOD test sets). We report the cross-modality-anatomy scores in Figure 2. A few interesting observations can be made. First, there are some associations between different anatomies and modalities. For example, the model trained on "CT Head" can also achieve promising results on the "MR Head" set. Secondly, training the model with all the modality-anatomy pairs (denoted as ALL) achieves the best generalization, obtaining the best results across all modalities and anatomies including the OOD test sets. We leave further exploration of cross-modality-anatomy associations and zero-shot OOD + +
ModelsMR HeadCT SpineCT NeckCT HeadCT ChestCT Abd/PelX-ray Chest
R1R2RLR1R2RLR1R2RLR1R2RLR1R2RLR1R2RLR1R2RL
WGSum------------------48.433.346.7
AIG-CL------------------51.035.246.7
T5-S38.218.328.535.818.628.939.020.029.143.125.336.539.518.529.328.910.621.247.832.243.5
BioBART-B42.421.232.047.827.940.040.419.629.346.027.438.941.419.130.333.112.523.249.633.845.3
BioBART-L42.121.432.647.828.140.840.319.429.645.526.738.640.217.828.932.511.722.649.333.344.9
BART-B42.021.532.149.029.741.641.420.930.246.428.139.541.619.530.633.112.923.651.034.946.4
BART-L43.722.132.849.829.741.442.020.530.446.627.339.041.818.629.633.912.423.251.734.946.8
+ +![](images/ed8891320f89fb9c958d4384a55d93432cad0a4283efacef08f26d99ac01d6f8.jpg) +Figure 2: Cross-modality-anatomy results from BART-L are visualized here using heatmaps. Colors from light to dark represent the values from low to high in each column. As discussed in Section 3.2, the model variant "ALL" reports the strongest performances. + +![](images/88649b0013ced896b46bb9e32da3860d62d31f6ea58df1bcb8f5d0ea01738702.jpg) + +![](images/069e420a9ffef436dad2c0f5ad445bb82185a743120a22dc906b16229cdc250c.jpg) + +Table 3: The benchmarking comparisons of different approaches, including task-specific models (i.e., WGSum (Hu et al., 2021) and AIG-CL (Hu et al., 2022)) and pre-trained language models (i.e., T5-S, BioBART-B, BioBART-L, BART-B, and BART-L). R1, R2, and RL denote ROUGE-1, ROUGE-2, and ROUGE-L, respectively. + +
T5-SBioBART-BBioBART-LBART-BBART-L
MR Head21.524.825.325.026.1
CT Spine23.837.037.038.538.3
CT Neck21.223.623.624.024.9
CT Head31.834.234.035.234.7
CT Chest24.026.024.326.025.2
CT Abd/Pel12.615.915.316.115.9
X-ray Chest39.840.941.042.343.0
+ +Table 4: F1-RadGraph scores on MIMIC-RRS test sets across different anatomies and modalities. + +transfer for future work. + +# 3.3 Clinical-Efficacy + +In addition to evaluating our systems using the ROUGE-1, ROUGE-2, and ROUGE-L metrics (Lin, 2004), we use a factual correctness metric to analyze clinical efficacy. Most prior works (Zhang et al., 2020; Smit et al., 2020; Hu et al., 2022) mainly use the $\mathrm{F_1}$ CheXbert metric, an F1-score that evaluates the factual correctness of the generated impressions using 14 chest radio + +graphic observations. Unfortunately, this metric is unsuitable for MIMIC-RRS, which contains reports from other modality-anatomy pairs beyond chest X-rays. + +For this reason, instead of using $\mathrm{F_1}$ CheXbert, we propose to use RadGraph (Jain et al., 2021) to evaluate the clinical correctness of the generated impressions. RadGraph is a dataset containing board-certified radiologist annotations of radiology reports corresponding to 14,579 entities and 10,889 relations (Appendix A.1). We used the released pretrained model to annotate our reports and asked one board-certified radiologist to subjectively validate that the printed entities of the RadGraph model on our data are correct (examples are shown in Table 5). After confirming the effectiveness of the model, we follow Delbrouck et al. (2022a) to compute the F1-RadGraph scores. The score evaluates the correctness of the generated named entities in the hypothesis impression compared to the ground-truth impression. We report these results in Ta + +ble 4. It can be observed that the BART models can achieve the best performance with respect to clinical efficacy. The results are consistent with the ROUGE scores, further confirming the effectiveness of adopting BART as the backbone instead of designing complicated solutions. + +# 4 Related Work + +In this section, we discuss prior research related to the radiology report summarization task. The first attempt at automatic summarization of radiology findings into natural language impression statements was proposed by Zhang et al. (2018). Their contribution was to propose a first baseline on the task, using a bidirectional-LSTM as encoder and decoder. Importantly, they found that about $30\%$ of the summaries generated from neural models contained factual errors. Subsequently, Zhang et al. (2020) proposed the $\mathrm{F_1}$ CheXbert score to evaluate the factual correctness of the generated impression. They also used reinforcement learning to optimize the $\mathrm{F_1}$ CheXbert score directly. Finally, both Hu et al. (2021) and Hu et al. (2022) used the Biomedical and Clinical English Model Packages in the Stanza Python NLP Library (Zhang et al., 2021) to extract medical entities. The former study used the entities to construct a Graph Neural Network, which was used as input in their summarization pipeline. In contrast, the latter study used the entities to mask the findings during contrastive pre-training. + +We believe this paper is an original contribution to the aforementioned line of work. As instigated by Zhang et al. (2018), our goal is to release a new summarization corpus and baselines on new modalities and anatomies. We do so by releasing an RRS dataset with data from 11 new modality-anatomy pairs. In addition, we extend the work performed by Zhang et al. (2020) by proposing a new metric to evaluate the factual correctness and completeness of the generated impression, namely the RadGraph score. Finally, we improve on the work of Hu et al. (2021, 2022) in two ways: (1) we use semantic annotations from a pre-trained model trained using annotations from board-certified radiologists, as opposed to Stanza which leverages unsupervised biomedical and clinical text data; (2) we leverage relation annotations between entities, a feature that was not available in prior work. + +# 5 Conclusion and Discussion + +In this paper, we highlight and address several weaknesses associated with the radiology report summarization task. First, from a data perspective, we propose a publicly available dataset named MIMIC-RRS involving data samples from twelve modality-anatomy pairs, with 79,779 samples from MIMIC-III and 128,003 samples from MIMIC-CXR. + +Second, we conducted more than 40 experiments and over 400 cross-modality-anatomy evaluations to benchmark the performance of different models. We show that instead of designing complicated modules, we can start from an appropriate backbone model such as BART. + +Finally, we proposed an elegant and simple metric, F1-RadGraph, to evaluate the factual correctness of summaries generated for any modality and anatomy. In the future, we hope that our work broadens the scope of the radiology report summarization task and contributes to the development of reliable RRS models that generalize well to new anatomies and modalities. + +# Limitations + +We note two limitations of our paper. First, our work does not extensively evaluate all the available pre-trained models that could be suitable for this task, e.g., ELECTRA (Clark et al., 2020), BioLinkBERT (Yasunaga et al., 2022), GatorTron (Yang et al., 2022), RadBERT (Yan et al., 2022), and PubMedBERT (Gu et al., 2021). The aim of this work is not to report the strongest possible score but rather to address weaknesses of existing radiology report summarization studies (in terms of data and evaluation). Yet, we are confident our proposed solutions report a strong baseline for future work. Second, although F1-RadGraph seems like an appropriate metric to evaluate our new modalities and anatomies (and appears to be consistent with ROUGE scores), it has only been evaluated subjectively and not systematically. + +# Acknowledgments + +Maya Varma is supported by graduate fellowship awards from the Department of Defense (NDSEG) and the Knight-Hennessy Scholars program at Stanford University. + +# References + +Robert Alexander, Stephen Waite, Michael A Bruno, Elizabeth A Krupinski, Leonard Berlin, Stephen Macknik, and Susana Martinez-Conde. 2022. Mandating limits on workload, duty, and speed in radiology. *Radiology*, 304(2):274-282. + +Emily Alsentzer, John Murphy, William Boag, WeiHung Weng, Di Jindi, Tristan Naumann, and Matthew McDermott. 2019. Publicly available clinical bert embeddings. In Proceedings of the 2nd Clinical Natural Language Processing Workshop, pages 72-78. + +Asma Ben Abacha, Yassine Mrabet, Yuhao Zhang, Chaitanya Shivade, Curtis Langlotz, and Dina Demner-Fushman. 2021. Overview of the MEDIQA 2021 shared task on summarization in the medical domain. In Proceedings of the 20th Workshop on Biomedical Language Processing, pages 74-85, Online. Association for Computational Linguistics. + +Mythreyi Bhargavan, Adam H Kaye, Howard P Forman, and Jonathan H Sunshine. 2009. Workload of radiologists in united states in 2006-2007 and trends since 1991-1992. Radiology, 252(2):458-467. + +Aurelia Bustos, Antonio Pertusa, Jose-Maria Salinas, and Maria de la Iglesia-Vaya. 2020. Padchest: A large chest x-ray image dataset with multi-label annotated reports. Medical image analysis, 66:101797. + +Kevin Clark, Minh-Thang Luong, Quoc V. Le, and Christopher D. Manning. 2020. ELECTRA: Pretraining text encoders as discriminators rather than generators. In ICLR. + +Jean-Benoit Delbrouck, Pierre Chambon, Christian Bluethgen, Emily Tsai, Omar Almusa, and Curtis P Langlotz. 2022a. Improving the factual correctness of radiology report generation with semantic rewards. arXiv preprint arXiv:2210.12186. + +Jean-benoit Delbrouck, Khaled Saab, Maya Varma, Sabri Eyuboglu, Pierre Chambon, Jared Dunnmon, Juan Zambrano, Akshay Chaudhari, and Curtis Langlotz. 2022b. ViLMedic: a framework for research at the intersection of vision and language in medical AI. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 23-34, Dublin, Ireland. Association for Computational Linguistics. + +Dina Demner-Fushman, Marc D Kohli, Marc B Rosenman, Sonya E Shooshan, Laritza Rodriguez, Sameer Antani, George R Thoma, and Clement J McDonald. 2016. Preparing a collection of radiology examinations for distribution and retrieval. Journal of the American Medical Informatics Association, 23(2):304-310. + +Yu Gu, Robert Tinn, Hao Cheng, Michael Lucas, Naoto Usuyama, Xiaodong Liu, Tristan Naumann, Jianfeng Gao, and Hoifung Poon. 2021. Domain-specific language model pretraining for biomedical natural language processing. ACM Transactions on Computing for Healthcare (HEALTH), 3(1):1-23. + +Jinpeng Hu, Jianling Li, Zhihong Chen, Yaling Shen, Yan Song, Xiang Wan, and Tsung-Hui Chang. 2021. Word graph guided summarization for radiology findings. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 4980-4990, Online. Association for Computational Linguistics. + +Jinpeng Hu, Zhuo Li, Zhihong Chen, Zhen Li, Xiang Wan, and Tsung-Hui Chang. 2022. Graph enhanced contrastive learning for radiology findings summarization. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4677-4688. + +Saahil Jain, Ashwin Agrawal, Adriel Saporta, Steven Truong, Du Nguyen Duong, Tan Bui, Pierre Chambon, Yuhao Zhang, Matthew Lungren, Andrew Ng, Curtis Langlotz, Pranav Rajpurkar, and Pranav Rajpurkar. 2021. Radgraph: Extracting clinical entities and relations from radiology reports. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, volume 1. + +Alistair EW Johnson, Tom J Pollard, Seth J Berkowitz, Nathaniel R Greenbaum, Matthew P Lungren, Chihying Deng, Roger G Mark, and Steven Horng. 2019. Mimic-cxr, a de-identified publicly available database of chest radiographs with free-text reports. Scientific data, 6(1):1-8. + +Alistair EW Johnson, Tom J Pollard, Lu Shen, Li-wei H Lehman, Mengling Feng, Mohammad Ghassemi, Benjamin Moody, Peter Szolovits, Leo Anthony Celi, and Roger G Mark. 2016. Mimic-iii, a freely accessible critical care database. Scientific data, 3(1):1-9. + +Sanjeev Kumar Karn, Ning Liu, Hinrich Schütze, and Oladimeji Farri. 2022. Differentiable multi-agent actor-critic for multi-step radiology report summarization. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1542-1553. + +Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871-7880. + +Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81. + +Qiuhao Lu, Dejing Dou, and Thien Nguyen. 2022. Clinical5: A generative language model for clinical text. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 5436-5443. + +Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J Liu, et al. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(140):1-67. + +Akshay Smit, Saahil Jain, Pranav Rajpurkar, Anuj Parek, Andrew Y Ng, and Matthew Lungren. 2020. Combining automatic labelers and expert annotations for accurate radiology report labeling using bert. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1500-1519. +An Yan, Julian McAuley, Xing Lu, Jiang Du, Eric Y Chang, Amilcare Gentili, and Chun-Nan Hsu. 2022. Radbert: Adapting transformer-based language models to radiology. Radiology: Artificial Intelligence, 4(4):e210258. +Xi Yang, Nima PourNejatian, Hoo Chang Shin, Kaleb E Smith, Christopher Parisien, Colin Compas, Cheryl Martin, Mona G Flores, Ying Zhang, Tanja Magoc, et al. 2022. Gatortron: A large clinical language model to unlock patient information from unstructured electronic health records. arXiv preprint arXiv:2203.03540. +Michihiro Yasunaga, Jure Leskovec, and Percy Liang. 2022. Linkbert: Pretraining language models with document links. In Association for Computational Linguistics (ACL). +Hongyi Yuan, Zheng Yuan, Ruyi Gan, Jiaxing Zhang, Yutao Xie, and Sheng Yu. 2022. Biobart: Pretraining and evaluation of a biomedical generative language model. In Proceedings of the 21st Workshop on Biomedical Language Processing, pages 97-109. +Yuhao Zhang, Daisy Yi Ding, Tianpei Qian, Christopher D Manning, and Curtis P Langlotz. 2018. Learning to summarize radiology findings. In Proceedings of the Ninth International Workshop on Health Text Mining and Information Analysis, pages 204-213. +Yuhao Zhang, Derek Merck, Emily Tsai, Christopher D Manning, and Curtis Langlotz. 2020. Optimizing the factual correctness of a summary: A study of summarizing radiology reports. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5108-5120. +Yuhao Zhang, Yuhui Zhang, Peng Qi, Christopher D Manning, and Curtis P Langlotz. 2021. Biomedical and clinical english model packages for the stanza python nlp library. Journal of the American Medical Informatics Association, 28(9):1892-1899. + +# A Details of RadGraph Scores + +# A.1 The Introduction of RadGraph + +![](images/c2b5f95e63ae9279e70dfd8e06c1f438fa698ba2922a020d4ebdc1831e88daa4.jpg) +Figure 3: Example of the RadGraph annotations. Figure taken from (Jain et al., 2021). + +To design our new evaluation metric, we leverage the RadGraph dataset (Jain et al., 2021) containing board-certified radiologist annotations of chest X-ray reports, which correspond to 14,579 entities and 10,889 relations. RadGraph has released a PubMedBERT model (Gu et al., 2021) pre-trained on these annotations to annotate new reports. An example of annotation can be seen in Figure 3. Before moving on to the next section, we quickly describe the concept of entities and relations: + +Entities An entity is defined as a continuous span of text that can include one or more adjacent words. Entities in RadGraph center around two concepts: Anatomy and Observation. Three uncertainty levels exist for Observation, leading to four different entities: Anatomy (ANAT-DP), Observation: Definitely Present (OBS-DP), Observation: Uncertain (OBS-U), and Observation: Definitely Absent (OBS-DA). + +Relations A relation is defined as a directed edge between two entities. Three levels exist: Suggestive $\text{Of}(\cdot, \cdot)$ , Located At ( $. \cdot$ ), and Modify ( $. \cdot$ ). + +# A.2 Metric Computation + +Using the RadGraph annotation scheme and pretrained model, we designed an F-score style reward that measures the factual consistency and completeness of the generated impression (also called hypothesis impression) compared to the reference impression. + +To do so, we treat the RadGraph annotations of an impression as a graph $\mathcal{G}(V,E)$ with the set of nodes $V = \{v_{1},v_{2},\ldots ,v_{|V|}\}$ containing the entities and the set of edges $E = \{e_1,e_2,\dots ,e_{|E|}\}$ the relations between pairs of entities. The graph is directed, meaning that the edge $e = (v_{1},v_{2})\neq$ $(v_{2},v_{1})$ . An example is depicted in Figure 4. Each node or edge of the graph also has a label, which + +we denote as $v_{i_L}$ for an entity $i$ (for example "OBS-DP" or "ANAT") and $e_{ij_L}$ for a relation $e = (v_i, v_j)$ (such as "modified" or "located at"). + +To design our RadGraph score, we focus on the nodes $V$ and whether or not a node has a relation in $E$ . For a hypothesis impression $y$ , we create a new set of triplets $T_{y} = \{(v_{i}, v_{i_{L}}, \mathcal{R})\}_{i=1:|V|}$ . The value $\mathcal{R}$ is 1 if $(v_{i}, v_{j})_{j=1:|E|, i \neq j} \in E$ , 0 otherwise. In other words, a triplet contains an entity, the entity label, and whether or not this entity has a relation. We proceed to construct the same set for the reference report $\hat{y}$ and denote this set $T_{\hat{y}}$ . + +Finally, our score is defined as the harmonic mean of precision and recall between the hypothesis set $T_{y}$ and the reference set $T_{\hat{y}}$ , giving a value between 0 and 100. As an illustration, the set $V$ , $E$ and $T$ of the graph $\mathcal{G}$ in Figure 4 are shown as follows: + +$$ +\begin{array}{l} V = \{\text {m i l d , f l u i d , o v e r l o a d , o v e r t , p u l m o n a r y}, \\ \text {e d e m a} \} \end{array} +$$ + +$$ +\begin{array}{l} E = \left\{\text {(m i l d , o v e r l o a d)}, \text {(o v e r l o a d , f l u i d)}, \text {(e d e m a ,} \right. \\ \text {p u l m o n a r y)} \} \end{array} +$$ + +$$ +T = \left\{\left(\text {m i l d , o b s - d p , 1}\right), \left(\text {f l u i d , o b s - d p , 0}\right), \left(\text {o v e r - l o a d , o b s - d p , 1}\right), \left(\text {o v e r t , o b s - d a , 0}\right), \left(\text {p u l m o n a r y , a n a t - d p , 0}\right), \left(\text {e d e m a , o b s - d a , 1}\right) \right\} +$$ + +![](images/97416453b9d6654c726b9901018e3bd8c8599e607683b9ecf6ff0d7dac451c8d.jpg) +Figure 4: Graph view of the RadGraph annotations for the report in Figure 3. + +# B Code and Data Release + +Our research has been carried out using the ViLMedic library (Delbrouck et al., 2022b). Our code is available at https://github.com/jbdel/vilmedic. This link is anonymized and complies with the double-blind review process. More specifically, we release the code of the RadGraph score as well as the training of our baseline. We also release the script to download, pre-process, and split the radiology reports of the MIMIC-III database + +
CT SpineCT SinusMR NeckMR Head
low resolution study reveals degenerative OBS-DP change OBS-OP and foraminal ANAT-OP narrowing OBS-OP without gross OBS-DA acute OBS-DA pathology OBS-DA1. sinusitis OBS-DP affecting the left ANAT-DS pheoid anat-OP and ethmoid anat-OP sinus ANAT-OP .2 . opacification OBS-OP of bilateral ANAT-OP mastoid ANAT-OP air cells and fluid OBS-OP seen in the middle ANAT-OP ear ANAT-OP cavities ANAT-OP which may indicate infection OBS-OP .slightly OBS-OP prominent OBS-OP lymph OBS-OP node OBS-OP in the posterior ANAT-OP chain ANAT-OP on the left side ANAT-OP side unchanged OBS-OP from previous examination . no definite evidence of infiltrating OBS-DA mass OBS-DA or definite pathologic adenopathy OBS-DA .1. no acute OBS-DA ischemia OBS-DA .2 . age -appropriate OBS-OP -appropriate atrophy OBS-OP , and chronic OBS-OP small OBS-OP vessel ANAT-OP ischemic OBS-OP changes OBS-OP .3 . there is no occlusion OBS-DA or flow -limiting OBS-DA - limiting stenosis OBS-DA of the arterial ANAT-OP system ANAT-OP of the head and neck
+ +Table 5: Examples of entities detected by RadGraph (used in the $\mathrm{RG_{ER}}$ metric) on out-of-domain anatomy/modality radiology reports. Relations are omitted for clarity. + +as per our experiments. To download the MIMIC-III database, researchers are required to formally request access via a process documented on the MIMIC website. There are two key steps that must be completed before access is granted: (i) the researcher must complete a recognized course in protecting human research participants, including Health Insurance Portability and Accountability Act (HIPAA) requirements. (ii) the researcher must sign a data use agreement, which outlines appropriate data usage and security standards, and forbids efforts to identify individual patients. + +# C More Results + +We present the results (including four metrics, i.e., ROUGE-1, ROUGE-2, ROUGE-L, and RadGraph scores) of all the experiments on Figure 5-9 for further research in this field. We also show the output of RadGraph (for entities) on a few samples of our new dataset in Table 5. + +# D Ethics Statement + +The MIMIC-CXR and MIMIC-III datasets are de-identified to satisfy the US Health Insurance Portability and Accountability Act of 1996 (HIPAA) Safe Harbor requirements. Protected health information (PHI) has been removed. Therefore, the ethical approval statement and the need for informed consent were waived for the studies on this database, which was approved by the Massachusetts Institute of Technology (Cambridge, MA) and Beth Israel Deaconess Medical Center (Boston, MA). This research was conducted in accordance with the Declaration of Helsinki, describing the ethical principles of medical research involving human subjects. + +![](images/ee33f4d0af3955c15a0fd299c144cb6820be34977e226e921b4cabcc16a93d7f.jpg) +ROUGE-1 + +![](images/8a3b25fab03d88a532ab372bc72d1847fa90c4d1d60a277843c25981bf34ca3c.jpg) +ROUGE-2 + +![](images/7dc05ddbe4482c7299afe712e7f22401b67815d2b9d8544b501949b0adefcf19.jpg) +ROUGE-L + +![](images/064769bd195bbfdfbc4a5af3200ac07641ef25a517951c0330bbb57cc6599299.jpg) +RadGraph Score +Figure 5: Cross-modality-anatomy results from T5-S are visualized here using heatmpas. Colors from light to dark represent the values from low to high in each column. + +![](images/7c5bc00fd7b4a190f8a1537f7e1fbd1efb2df1611316e2587fd11f12751c4438.jpg) +ROUGE-1 + +![](images/38a5b92cf5ca425c6d7f47c2742119166fdea4b13e43deee0673fe95a707fb12.jpg) +ROUGE-2 + +![](images/570950fb54e59fb1aaa2dd7efcdb9a61be3e109e0a36e840befe716be5f1124e.jpg) +ROUGE-L + +![](images/d22930f2b39cc1358c376d653ad55150a3d60b85e9dd3217201ec04711ada013.jpg) +RadGraph Score +Figure 6: Cross-modality-anatomy results from BART-B are visualized here using heatmaps. Colors from light to dark represent the values from low to high in each column. + +![](images/fa7dd4a837009612dd7b747453bc1f6037ecbe4aa426564fbf098aca89742521.jpg) +ROUGE-1 + +![](images/d428477be2e0fa4e49f3ce17689dfe14c2f95bdb265c7e0684d79bf4a0641d93.jpg) +ROUGE-2 + +![](images/f2076082bc66504433f5ecf5dd5d1cc86629e8339887f9c2da1660ccd44be3e0.jpg) +ROUGE-L + +![](images/c6bb4b722501db9c7d686922487bf329fb3a05e8ebdc2ffa0e397584b238d8e6.jpg) +RadGraph Score +Figure 7: Cross-modality-anatomy results from BART-L are visualized here using heatmaps. Colors from light to dark represent the values from low to high in each column. + +![](images/4c520bb5abbe22fcabc27d0bd6870d48ba950708bae7b5ea54dbf79586afc079.jpg) +ROUGE-1 + +![](images/05a17ab6cdbe137561b71e15c1fa2835b1748e4f96d5e7c4d8f489d55f60bdad.jpg) +ROUGE-2 + +![](images/045ea34b4c69f0539d8800597e79569ca0c1f6dfd64f841defd447710dc7015d.jpg) +ROUGE-L + +![](images/a28e93aedb83d89a195a25518651ffc2bb5fdd3d196e274255a6c3bb30f7f9bd.jpg) +RadGraph Score +Figure 8: Cross-modality-anatomy results from BioBART-B are visualized here using heatmaps. Colors from light to dark represent the values from low to high in each column. + +![](images/e3e86c1bb2e64c7703a64064869fce07d9fd81f177a8d6aeec13ebd30406634f.jpg) +ROUGE-1 + +![](images/08186fd00caf82df86aa3a9469e966fcfe201d0246ac986a51811a4d7c681764.jpg) +ROUGE-2 + +![](images/3fbc8aab9bfa93c77c1f164ea14aa387df22ae9d8124b7a18a56f875f014bf02.jpg) +ROUGE-L + +![](images/f0a5090e77c9e5fd49ca1a9227b19307f6e367765d2763718c8d172171a135cf.jpg) +RadGraph Score +Figure 9: Cross-modality-anatomy results from BioBART-L are visualized here using heatmaps. Colors from light to dark represent the values from low to high in each column. + +A For every submission: + +A1. Did you describe the limitations of your work? + +On Page 5. + +A2. Did you discuss any potential risks of your work? + +On Page 5. + +A3. Do the abstract and introduction summarize the paper's main claims? + +On Pages 1 and 4. + +A4. Have you used AI writing assistants when working on this paper? + +Left blank. + +B Did you use or create scientific artifacts? + +Section 2. + +B1. Did you cite the creators of artifacts you used? + +Section 2. + +B2. Did you discuss the license or terms for use and / or distribution of any artifacts? + +On Page 5. + +B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? + +Section 2. + +B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? + +Section 2. + +B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? + +Section 2. + +B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. + +Section 2. + +C Did you run computational experiments? + +Section 3. + +C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? + +Not applicable. We use the common pre-trained models in our experiments. + +The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance. + +C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? + +Section 3. + +C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? + +Section 3. + +C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? + +Section 3. + +D Did you use human annotators (e.g., crowdworkers) or research with human participants? + +Sections 2 and 3. + +D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? + +Section 2. + +D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? + +Section 2. + +D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? + +Section 2. + +D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? + +Section 2 and Page 5. + +D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? + +Section 2. \ No newline at end of file diff --git a/2023/Toward Expanding the Scope of Radiology Report Summarization to Multiple Anatomies and Modalities/images.zip b/2023/Toward Expanding the Scope of Radiology Report Summarization to Multiple Anatomies and Modalities/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..65b7cacbe48a4966dbaa1a7974ffc184d331ed66 --- /dev/null +++ b/2023/Toward Expanding the Scope of Radiology Report Summarization to Multiple Anatomies and Modalities/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:67c8ebc1d6526c4995f2bbb6725164eca658b0551798a4ac1450ce2b03f5d9e2 +size 1134203 diff --git a/2023/Toward Expanding the Scope of Radiology Report Summarization to Multiple Anatomies and Modalities/layout.json b/2023/Toward Expanding the Scope of Radiology Report Summarization to Multiple Anatomies and Modalities/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e8a85e261753cbae2313dca82aaa320cb2241d58 --- /dev/null +++ b/2023/Toward Expanding the Scope of Radiology Report Summarization to Multiple Anatomies and Modalities/layout.json @@ -0,0 +1,9207 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 85, + 66, + 509, + 100 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 66, + 509, + 100 + ], + "spans": [ + { + "bbox": [ + 85, + 66, + 509, + 100 + ], + "type": "text", + "content": "Toward Expanding the Scope of Radiology Report Summarization to Multiple Anatomies and Modalities" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 160, + 105, + 442, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 105, + 442, + 133 + ], + "spans": [ + { + "bbox": [ + 160, + 105, + 442, + 133 + ], + "type": "text", + "content": "Zhihong Chen" + }, + { + "bbox": [ + 160, + 105, + 442, + 133 + ], + "type": "inline_equation", + "content": "^{2,3*}" + }, + { + "bbox": [ + 160, + 105, + 442, + 133 + ], + "type": "text", + "content": ", Maya Varma" + }, + { + "bbox": [ + 160, + 105, + 442, + 133 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 160, + 105, + 442, + 133 + ], + "type": "text", + "content": ", Xiang Wan" + }, + { + "bbox": [ + 160, + 105, + 442, + 133 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 160, + 105, + 442, + 133 + ], + "type": "text", + "content": ", Curtis P. Langlotz" + }, + { + "bbox": [ + 160, + 105, + 442, + 133 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 160, + 105, + 442, + 133 + ], + "type": "text", + "content": ", Jean-Benoit Delbrouck" + }, + { + "bbox": [ + 160, + 105, + 442, + 133 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 248, + 134, + 348, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 134, + 348, + 148 + ], + "spans": [ + { + "bbox": [ + 248, + 134, + 348, + 148 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 248, + 134, + 348, + 148 + ], + "type": "text", + "content": "Stanford University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 176, + 148, + 420, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 148, + 420, + 161 + ], + "spans": [ + { + "bbox": [ + 176, + 148, + 420, + 161 + ], + "type": "text", + "content": "2The Chinese University of Hong Kong, Shenzhen" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 198, + 162, + 399, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 162, + 399, + 175 + ], + "spans": [ + { + "bbox": [ + 198, + 162, + 399, + 175 + ], + "type": "text", + "content": "3Shenzhen Research Institute of Big Data" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 157, + 177, + 439, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 177, + 439, + 189 + ], + "spans": [ + { + "bbox": [ + 157, + 177, + 439, + 189 + ], + "type": "text", + "content": "zhihongchen@link.cuhk.edu.cn wanxiang@sribd.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 187, + 190, + 409, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 190, + 409, + 203 + ], + "spans": [ + { + "bbox": [ + 187, + 190, + 409, + 203 + ], + "type": "text", + "content": "{mvarma2,langlotz,jbdel}@stanford.edu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 84, + 237, + 274, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 237, + 274, + 488 + ], + "spans": [ + { + "bbox": [ + 84, + 237, + 274, + 488 + ], + "type": "text", + "content": "Radiology report summarization (RRS) is a growing area of research. Given the Findings section of a radiology report, the goal is to generate a summary (called an Impression section) that highlights the key observations and conclusions of the radiology study. However, RRS currently faces essential limitations. First, many prior studies conduct experiments on private datasets, preventing the reproduction of results and fair comparisons across different systems and solutions. Second, most prior approaches are evaluated solely on chest X-rays. To address these limitations, we propose a dataset (MIMIC-RRS) involving three new modalities and seven new anatomies based on the MIMIC-III and MIMIC-CXR datasets. We then conduct extensive experiments to evaluate the performance of models both within and across modality-anatomy pairs in MIMIC-RRS. In addition, we evaluate their clinical efficacy via RadGraph, a factual correctness metric." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 68, + 500, + 154, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 500, + 154, + 513 + ], + "spans": [ + { + "bbox": [ + 68, + 500, + 154, + 513 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 523, + 291, + 753 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 523, + 291, + 753 + ], + "spans": [ + { + "bbox": [ + 67, + 523, + 291, + 753 + ], + "type": "text", + "content": "A radiology report is a document that provides information about the results of a radiology study. It usually includes a Findings section with key observations from the study and an Impression section with the radiologist's overall conclusions. The latter is the most critical part of the report and is typically based on both the findings and the patient's condition. It can be helpful to automate the process of generating the impression section because it can be time-consuming and prone to errors when done manually (Bhargavan et al., 2009; Alexander et al., 2022). Recently, substantial progress has been made towards research on automated radiology report summarization (RRS) (Zhang et al., 2020; Ben Abacha et al., 2021; Hu et al., 2022). However, the field of RRS faces several key limitations. First, the experimental results of many" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 212, + 526, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 212, + 526, + 414 + ], + "spans": [ + { + "bbox": [ + 302, + 212, + 526, + 414 + ], + "type": "text", + "content": "prior studies (Zhang et al., 2018, 2020) are reported on private datasets, making it difficult to replicate results or compare approaches. Second, existing studies are mainly limited to a single modality (i.e., X-ray) and a single anatomy (i.e., chest) (Zhang et al., 2020; Ben Abacha et al., 2021; Hu et al., 2021). In some cases, researchers omit to disclose the modality and anatomy of the radiology reports used for their experiments (Karn et al., 2022). Finally, recent models (Karn et al., 2022; Hu et al., 2022) present an increased complexity in architecture that offers only marginal improvements on the existing evaluation metrics for summarization. This further makes the replication of studies more difficult." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 416, + 526, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 416, + 526, + 577 + ], + "spans": [ + { + "bbox": [ + 302, + 416, + 526, + 577 + ], + "type": "text", + "content": "To address the aforementioned limitations, we construct a brand-new open-source dataset (named MIMIC-RRS) for radiology report summarization involving three modalities (X-ray, MRI, and CT) and seven anatomies (chest, head, neck, sinus, spine, abdomen, and pelvis). MIMIC-RRS is based on the MIMIC-CXR (Johnson et al., 2019) and MIMIC-III (Johnson et al., 2016) datasets and introduces data from 12 new modality-anatomy pairs. As a result, we introduce a new setting for evaluating the generalization capabilities of RRS models across different modalities and anatomies." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 579, + 526, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 579, + 526, + 687 + ], + "spans": [ + { + "bbox": [ + 302, + 579, + 526, + 687 + ], + "type": "text", + "content": "In addition, we benchmark various pre-trained language models on MIMIC-RRS. Through extensive experiments within and across modality-anatomy pairs, we show that adopting an appropriate pretrained model can achieve promising results comparable to previous studies. We also introduce a metric to evaluate factual correctness of generated summaries for any modality-anatomy pair." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 698, + 433, + 709 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 698, + 433, + 709 + ], + "spans": [ + { + "bbox": [ + 302, + 698, + 433, + 709 + ], + "type": "text", + "content": "2 Dataset Construction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 719, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 719, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 719, + 526, + 772 + ], + "type": "text", + "content": "In this section, we present the new MIMIC-RRS dataset designed for radiology report summarization across multiple modalities and anatomies. Comparisons with existing datasets are shown in" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 81, + 761, + 158, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 761, + 158, + 772 + ], + "spans": [ + { + "bbox": [ + 81, + 761, + 158, + 772 + ], + "type": "text", + "content": "*Equal Contribution." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "text", + "content": "469" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "spans": [ + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "type": "text", + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 224, + 806, + 369, + 817 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 806, + 369, + 817 + ], + "spans": [ + { + "bbox": [ + 224, + 806, + 369, + 817 + ], + "type": "text", + "content": "Volume 2: Short Papers, pages 469-484" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 176, + 818, + 417, + 828 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 818, + 417, + 828 + ], + "spans": [ + { + "bbox": [ + 176, + 818, + 417, + 828 + ], + "type": "text", + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 108, + 68, + 486, + 197 + ], + "blocks": [ + { + "bbox": [ + 108, + 68, + 486, + 197 + ], + "lines": [ + { + "bbox": [ + 108, + 68, + 486, + 197 + ], + "spans": [ + { + "bbox": [ + 108, + 68, + 486, + 197 + ], + "type": "table", + "html": "
DatasetAnatomyModalityLanguageNumber
Zhang et al. (2018)MultipleMultipleEnglish87,127
Zhang et al. (2020)MultipleMultipleEnglish130,850
RIH (Zhang et al., 2020)MultipleMultipleEnglish139,654
OpenI (Demner-Fushman et al., 2016)ChestX-rayEnglish3,268
MIMIC-CXR (Johnson et al., 2019)ChestX-rayEnglish128,003
PadChest (Bustos et al., 2020)ChestX-raySpanish206,222
MIMIC-RRS (ours)MultipleMultipleEnglish207,782
", + "image_path": "dead11df0c968ac5b18f4a4fb547a6419d07c1cb0ff012cbf6524e7faeb437e2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 279, + 168, + 291 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 279, + 168, + 291 + ], + "spans": [ + { + "bbox": [ + 67, + 279, + 168, + 291 + ], + "type": "text", + "content": "2.1 Data Collection" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 298, + 291, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 298, + 291, + 568 + ], + "spans": [ + { + "bbox": [ + 66, + 298, + 291, + 568 + ], + "type": "text", + "content": "MIMIC-III One of our main contributions is to generate RRS data from MIMIC-II involving distinct combinations of modalities (i.e., medical imaging techniques) and anatomies (i.e., body parts). To this end, we first select five of the most frequently-occurring modality-anatomy pairs in the pool of MIMIC-III reports: \"CT Head\", \"CT Spine\", \"CT Chest\", \"CT Abdomen-Pelvis\" and \"MR Head\". Note that we discard chest X-rays as they are included in the MIMIC-CXR dataset. In addition, we pick six modality-anatomy pairs that occur infrequently in MIMIC-III to serve as out-of-domain (OOD) test sets: \"CT Neck\", \"CT Sinus\", \"MR Pelvis\", \"MR Neck\", \"MR Abdomen\", \"MR Spine\". This set of pairs represents two types of OOD cases: (1) the modality has not been seen during training (one could train on CT neck and test on MR Neck), and (2) the anatomy has not been seen during training (for example, CT Sinus is the only \"sinus\" dataset)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 571, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 571, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 66, + 571, + 291, + 773 + ], + "type": "text", + "content": "For each report, we extract the findings and impression section. However, the findings section is not always clearly labeled as \"findings\". With the help of a board-certified radiologist, we identify alternate section headers that reference findings for each modality-anatomy pair. As an example, for CT head, findings may be referenced in reports with the section headings \"non-contrast head ct\", \"ct head\", \"ct head without contrast\", \"ct head without iv contrast\", \"head ct\", \"head ct without iv contrast\", or \"cta head\". We identify 537 candidate section headers that reference findings across our dataset. We also discarded reports where multiple studies are pooled in the same radiology report, leading to multiple intricate observations in the impression" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 328, + 237, + 504, + 375 + ], + "blocks": [ + { + "bbox": [ + 133, + 206, + 459, + 218 + ], + "lines": [ + { + "bbox": [ + 133, + 206, + 459, + 218 + ], + "spans": [ + { + "bbox": [ + 133, + 206, + 459, + 218 + ], + "type": "text", + "content": "Table 1: Comparisons with existing datasets for radiology report summarization." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 239, + 290, + 267 + ], + "lines": [ + { + "bbox": [ + 67, + 239, + 290, + 267 + ], + "spans": [ + { + "bbox": [ + 67, + 239, + 290, + 267 + ], + "type": "text", + "content": "Table 1. We detail the collection process and the dataset statistics in the following subsections." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 237, + 504, + 375 + ], + "lines": [ + { + "bbox": [ + 328, + 237, + 504, + 375 + ], + "spans": [ + { + "bbox": [ + 328, + 237, + 504, + 375 + ], + "type": "table", + "html": "
CT Abd-pelvCT ChestCT Head
15,98912,78631,402
CT SpineMR HeadCT Neck
5,5177,3131,140
CT SinusMR SpineMR Abdomen
1,2672,8211,061
MR NeckMR PelvisX-ray Chest
230253128,003
", + "image_path": "6f4c2d9977b757eeec360c53e1cb287960a57f571619e3da871c15c7d9adb65a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 383, + 527, + 421 + ], + "lines": [ + { + "bbox": [ + 302, + 383, + 527, + 421 + ], + "spans": [ + { + "bbox": [ + 302, + 383, + 527, + 421 + ], + "type": "text", + "content": "Table 2: Dataset statistics for MIMIC-RRS. We report the number of radiology reports from each modality-anatomy pair." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 302, + 439, + 525, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 439, + 525, + 468 + ], + "spans": [ + { + "bbox": [ + 302, + 439, + 525, + 468 + ], + "type": "text", + "content": "section1. Our resulting dataset consists of 79,779 selected reports across 11 modality-anatomy pairs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 474, + 525, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 474, + 525, + 555 + ], + "spans": [ + { + "bbox": [ + 302, + 474, + 525, + 555 + ], + "type": "text", + "content": "MIMIC-CXR MIMIC-CXR studies are chest X-ray examinations. We follow preprocessing steps reported in previous work (Delbrouck et al., 2022b), and we only include reports with both a Findings and an Impression section. This yields 128,003 reports." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 565, + 396, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 565, + 396, + 576 + ], + "spans": [ + { + "bbox": [ + 302, + 565, + 396, + 576 + ], + "type": "text", + "content": "2.2 Data statistics" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 301, + 581, + 526, + 745 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 581, + 526, + 745 + ], + "spans": [ + { + "bbox": [ + 301, + 581, + 526, + 745 + ], + "type": "text", + "content": "In total, there are 207,782 samples in the MIMIC-RRS dataset. The number of examples for each modality and anatomy is provided in Table 2. To further analyze this dataset, we report in Figure 1 the text lengths and vocabulary sizes associated with reports from each modality-anatomy pair. We find that for all modality-anatomy pairs, the findings section is significantly longer than the impression section (up to " + }, + { + "bbox": [ + 301, + 581, + 526, + 745 + ], + "type": "inline_equation", + "content": "+315\\%" + }, + { + "bbox": [ + 301, + 581, + 526, + 745 + ], + "type": "text", + "content": " for MR abdomen). Additionally, the findings sections of chest X-ray reports, which average only 49 words, are much shorter than reports from other modality-anatomy" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 751, + 525, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 751, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 525, + 772 + ], + "type": "text", + "content": "1We release our candidate section headers as well as code to recreate the dataset from scratch (Appendix B)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "470" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 68, + 67, + 290, + 164 + ], + "blocks": [ + { + "bbox": [ + 68, + 67, + 290, + 164 + ], + "lines": [ + { + "bbox": [ + 68, + 67, + 290, + 164 + ], + "spans": [ + { + "bbox": [ + 68, + 67, + 290, + 164 + ], + "type": "image", + "image_path": "3e0061ba8be51b6cb07e2785a6215aeb7af3386453565b0e4de0bf8819bd6c4a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 175, + 289, + 200 + ], + "lines": [ + { + "bbox": [ + 67, + 175, + 289, + 200 + ], + "spans": [ + { + "bbox": [ + 67, + 175, + 289, + 200 + ], + "type": "text", + "content": "Figure 1: Section length and vocabulary size for reports from each modality-anatomy pair." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 220, + 292, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 220, + 292, + 369 + ], + "spans": [ + { + "bbox": [ + 66, + 220, + 292, + 369 + ], + "type": "text", + "content": "pairs. In contrast, MR Abdomen and MR Pelvis reports including findings sections that average 205 and 174 words, respectively. We see that CT Chest, CT Head, and CT Abdomen-Pelvis reports have a relatively large vocabulary size (given their sample size) with 20,909, 19,813, and 18,933 words. Surprisingly, the CT Abdomen-Pelvis impressions include a larger vocabulary than the findings. On the other hand, MR pelvis and MR abdomen impressions contain " + }, + { + "bbox": [ + 66, + 220, + 292, + 369 + ], + "type": "inline_equation", + "content": "36\\%" + }, + { + "bbox": [ + 66, + 220, + 292, + 369 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 66, + 220, + 292, + 369 + ], + "type": "inline_equation", + "content": "37\\%" + }, + { + "bbox": [ + 66, + 220, + 292, + 369 + ], + "type": "text", + "content": " fewer words than their corresponding findings, respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 370, + 291, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 370, + 291, + 476 + ], + "spans": [ + { + "bbox": [ + 66, + 370, + 291, + 476 + ], + "type": "text", + "content": "We assign reports from the following modality-anatomy pairs to training, validation, and test sets due to their large sample sizes: \"CT abdomen/pelvis\", \"CT Chest\", \"CT Neck\", \"CT Spine\", \"CT Head\", \"MR Head\", and \"X-ray Chest\". The remaining reports (i.e., \"MR Pelvis\", \"MR Spine\", \"MR Neck\", \"MR Abdomen\", and \"CT Sinus\") are used for OOD test sets" + }, + { + "bbox": [ + 66, + 370, + 291, + 476 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 66, + 370, + 291, + 476 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 486, + 198, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 486, + 198, + 501 + ], + "spans": [ + { + "bbox": [ + 67, + 486, + 198, + 501 + ], + "type": "text", + "content": "3 Algorithmic Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 507, + 291, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 507, + 291, + 576 + ], + "spans": [ + { + "bbox": [ + 67, + 507, + 291, + 576 + ], + "type": "text", + "content": "In this section, we conduct experiments to analyze the performance of different models on MIMIC-RRS. We provide three categories of analyses: inmodality-anatomy, cross-modality-anatomy, and clinical efficacy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 585, + 196, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 585, + 196, + 598 + ], + "spans": [ + { + "bbox": [ + 67, + 585, + 196, + 598 + ], + "type": "text", + "content": "3.1 In-modality-anatomy" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 602, + 292, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 602, + 292, + 723 + ], + "spans": [ + { + "bbox": [ + 67, + 602, + 292, + 723 + ], + "type": "text", + "content": "To benchmark the performance of different models on the proposed MIMIC-RRS dataset, we conduct experiments within each modality-anatomy pair (i.e., the training and test procedures are performed using only one modality-anatomy pair). We evaluate three types of pre-trained sequence-to-sequence models, namely T5 (Raffel et al., 2020), BART (Lewis et al., 2020), BioBART (Yuan et al., 2022), and their variants. Results are reported in" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 71, + 340, + 82 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 340, + 82 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 340, + 82 + ], + "type": "text", + "content": "Table 3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 84, + 526, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 84, + 526, + 368 + ], + "spans": [ + { + "bbox": [ + 302, + 84, + 526, + 368 + ], + "type": "text", + "content": "Several observations can be drawn from these experiments. First, simply adopting pretrained sequence-to-sequence language models can achieve results comparable to previous state-of-the-art approaches designed for radiology summarization. Indeed, using BART-L as a backbone achieves the best performance, confirming the necessity of exploiting appropriate pre-trained language models. Secondly, the performances across different model types vary (i.e., BART-L/BART-B, BioBART-L/BioBART-B). Yet, we notice that the number of training parameters matters; large models report the best results. According to our evaluations, the BART models achieve better results across all modality-anatomy pairs. Surprisingly, it is worth noting that the BioBART models do not achieve better results than BART, although BioBART is pre-trained on a biomedical corpus. One explanation could be that BioBART models are pre-trained on abstracts from PubMed, which are not within the same domain as radiology reports." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 369, + 525, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 369, + 525, + 464 + ], + "spans": [ + { + "bbox": [ + 302, + 369, + 525, + 464 + ], + "type": "text", + "content": "In summary, we note several key findings for future studies: (i) \"Less is more\": starting from an appropriate backbone instead of designing complicated modules; (ii) the model size matters; (iii) the pretraining domain matters: knowledge from clinical notes or medical literature does not easily translate to radiology reports." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 472, + 448, + 486 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 472, + 448, + 486 + ], + "spans": [ + { + "bbox": [ + 302, + 472, + 448, + 486 + ], + "type": "text", + "content": "3.2 Cross-modality-anatomy" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 490, + 525, + 707 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 490, + 525, + 707 + ], + "spans": [ + { + "bbox": [ + 302, + 490, + 525, + 707 + ], + "type": "text", + "content": "In this section, we conduct experiments across modality-anatomy pairs (i.e., models are trained on reports from a subset of modality-anatomy pairs and then evaluated on all pairs, including the OOD test sets). We report the cross-modality-anatomy scores in Figure 2. A few interesting observations can be made. First, there are some associations between different anatomies and modalities. For example, the model trained on \"CT Head\" can also achieve promising results on the \"MR Head\" set. Secondly, training the model with all the modality-anatomy pairs (denoted as ALL) achieves the best generalization, obtaining the best results across all modalities and anatomies including the OOD test sets. We leave further exploration of cross-modality-anatomy associations and zero-shot OOD" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 711, + 525, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 711, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 711, + 525, + 772 + ], + "type": "text", + "content": "et al., 2019), and Clinical-T5 (Lu et al., 2022)) that specialize in the clinical text since they were trained on the text from MIMIC-III, which overlaps with our dataset. The MIMIC-RRS test set is included in their pre-training data. Thus, we do not adopt them in our experiments to avoid potential data leakage and ensure a fair comparison." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 67, + 729, + 290, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 729, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 729, + 290, + 772 + ], + "type": "text", + "content": "2We release data splits publicly so that future work can fairly compare new results. 3We do not evaluate several pre-trained models (e.g., ClinicalBERT (Alsentzer et al., 2019), BioClinicalBERT (Alsentzer" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 289, + 781, + 305, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 305, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 305, + 791 + ], + "type": "text", + "content": "471" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 72, + 68, + 526, + 169 + ], + "blocks": [ + { + "bbox": [ + 72, + 68, + 526, + 169 + ], + "lines": [ + { + "bbox": [ + 72, + 68, + 526, + 169 + ], + "spans": [ + { + "bbox": [ + 72, + 68, + 526, + 169 + ], + "type": "table", + "html": "
ModelsMR HeadCT SpineCT NeckCT HeadCT ChestCT Abd/PelX-ray Chest
R1R2RLR1R2RLR1R2RLR1R2RLR1R2RLR1R2RLR1R2RL
WGSum------------------48.433.346.7
AIG-CL------------------51.035.246.7
T5-S38.218.328.535.818.628.939.020.029.143.125.336.539.518.529.328.910.621.247.832.243.5
BioBART-B42.421.232.047.827.940.040.419.629.346.027.438.941.419.130.333.112.523.249.633.845.3
BioBART-L42.121.432.647.828.140.840.319.429.645.526.738.640.217.828.932.511.722.649.333.344.9
BART-B42.021.532.149.029.741.641.420.930.246.428.139.541.619.530.633.112.923.651.034.946.4
BART-L43.722.132.849.829.741.442.020.530.446.627.339.041.818.629.633.912.423.251.734.946.8
", + "image_path": "09e4b9479b7f43318fa103b744bf6f33be27f15ad6e3cb2cec4d7db1b8cf9b39.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 70, + 226, + 214, + 417 + ], + "blocks": [ + { + "bbox": [ + 70, + 226, + 214, + 417 + ], + "lines": [ + { + "bbox": [ + 70, + 226, + 214, + 417 + ], + "spans": [ + { + "bbox": [ + 70, + 226, + 214, + 417 + ], + "type": "image", + "image_path": "ed8891320f89fb9c958d4384a55d93432cad0a4283efacef08f26d99ac01d6f8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 430, + 525, + 467 + ], + "lines": [ + { + "bbox": [ + 67, + 430, + 525, + 467 + ], + "spans": [ + { + "bbox": [ + 67, + 430, + 525, + 467 + ], + "type": "text", + "content": "Figure 2: Cross-modality-anatomy results from BART-L are visualized here using heatmaps. Colors from light to dark represent the values from low to high in each column. As discussed in Section 3.2, the model variant \"ALL\" reports the strongest performances." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 217, + 226, + 359, + 417 + ], + "blocks": [ + { + "bbox": [ + 217, + 226, + 359, + 417 + ], + "lines": [ + { + "bbox": [ + 217, + 226, + 359, + 417 + ], + "spans": [ + { + "bbox": [ + 217, + 226, + 359, + 417 + ], + "type": "image", + "image_path": "88649b0013ced896b46bb9e32da3860d62d31f6ea58df1bcb8f5d0ea01738702.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 362, + 226, + 524, + 417 + ], + "blocks": [ + { + "bbox": [ + 362, + 226, + 524, + 417 + ], + "lines": [ + { + "bbox": [ + 362, + 226, + 524, + 417 + ], + "spans": [ + { + "bbox": [ + 362, + 226, + 524, + 417 + ], + "type": "image", + "image_path": "069e420a9ffef436dad2c0f5ad445bb82185a743120a22dc906b16229cdc250c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 70, + 485, + 289, + 572 + ], + "blocks": [ + { + "bbox": [ + 67, + 176, + 525, + 214 + ], + "lines": [ + { + "bbox": [ + 67, + 176, + 525, + 214 + ], + "spans": [ + { + "bbox": [ + 67, + 176, + 525, + 214 + ], + "type": "text", + "content": "Table 3: The benchmarking comparisons of different approaches, including task-specific models (i.e., WGSum (Hu et al., 2021) and AIG-CL (Hu et al., 2022)) and pre-trained language models (i.e., T5-S, BioBART-B, BioBART-L, BART-B, and BART-L). R1, R2, and RL denote ROUGE-1, ROUGE-2, and ROUGE-L, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 485, + 289, + 572 + ], + "lines": [ + { + "bbox": [ + 70, + 485, + 289, + 572 + ], + "spans": [ + { + "bbox": [ + 70, + 485, + 289, + 572 + ], + "type": "table", + "html": "
T5-SBioBART-BBioBART-LBART-BBART-L
MR Head21.524.825.325.026.1
CT Spine23.837.037.038.538.3
CT Neck21.223.623.624.024.9
CT Head31.834.234.035.234.7
CT Chest24.026.024.326.025.2
CT Abd/Pel12.615.915.316.115.9
X-ray Chest39.840.941.042.343.0
", + "image_path": "2aecc3e1e0b16efb51de9eaebcd734880748117ece9fbd3ff59a3909b7d87798.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 580, + 289, + 604 + ], + "lines": [ + { + "bbox": [ + 67, + 580, + 289, + 604 + ], + "spans": [ + { + "bbox": [ + 67, + 580, + 289, + 604 + ], + "type": "text", + "content": "Table 4: F1-RadGraph scores on MIMIC-RRS test sets across different anatomies and modalities." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 625, + 175, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 625, + 175, + 638 + ], + "spans": [ + { + "bbox": [ + 67, + 625, + 175, + 638 + ], + "type": "text", + "content": "transfer for future work." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 647, + 173, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 647, + 173, + 661 + ], + "spans": [ + { + "bbox": [ + 67, + 647, + 173, + 661 + ], + "type": "text", + "content": "3.3 Clinical-Efficacy" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 665, + 292, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 665, + 292, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 665, + 292, + 773 + ], + "type": "text", + "content": "In addition to evaluating our systems using the ROUGE-1, ROUGE-2, and ROUGE-L metrics (Lin, 2004), we use a factual correctness metric to analyze clinical efficacy. Most prior works (Zhang et al., 2020; Smit et al., 2020; Hu et al., 2022) mainly use the " + }, + { + "bbox": [ + 67, + 665, + 292, + 773 + ], + "type": "inline_equation", + "content": "\\mathrm{F_1}" + }, + { + "bbox": [ + 67, + 665, + 292, + 773 + ], + "type": "text", + "content": " CheXbert metric, an F1-score that evaluates the factual correctness of the generated impressions using 14 chest radio" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 487, + 526, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 487, + 526, + 542 + ], + "spans": [ + { + "bbox": [ + 302, + 487, + 526, + 542 + ], + "type": "text", + "content": "graphic observations. Unfortunately, this metric is unsuitable for MIMIC-RRS, which contains reports from other modality-anatomy pairs beyond chest X-rays." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 543, + 527, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 543, + 527, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 543, + 527, + 773 + ], + "type": "text", + "content": "For this reason, instead of using " + }, + { + "bbox": [ + 302, + 543, + 527, + 773 + ], + "type": "inline_equation", + "content": "\\mathrm{F_1}" + }, + { + "bbox": [ + 302, + 543, + 527, + 773 + ], + "type": "text", + "content": " CheXbert, we propose to use RadGraph (Jain et al., 2021) to evaluate the clinical correctness of the generated impressions. RadGraph is a dataset containing board-certified radiologist annotations of radiology reports corresponding to 14,579 entities and 10,889 relations (Appendix A.1). We used the released pretrained model to annotate our reports and asked one board-certified radiologist to subjectively validate that the printed entities of the RadGraph model on our data are correct (examples are shown in Table 5). After confirming the effectiveness of the model, we follow Delbrouck et al. (2022a) to compute the F1-RadGraph scores. The score evaluates the correctness of the generated named entities in the hypothesis impression compared to the ground-truth impression. We report these results in Ta" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "472" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 291, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 291, + 153 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 291, + 153 + ], + "type": "text", + "content": "ble 4. It can be observed that the BART models can achieve the best performance with respect to clinical efficacy. The results are consistent with the ROUGE scores, further confirming the effectiveness of adopting BART as the backbone instead of designing complicated solutions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 179, + 160, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 179, + 160, + 192 + ], + "spans": [ + { + "bbox": [ + 67, + 179, + 160, + 192 + ], + "type": "text", + "content": "4 Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 213, + 291, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 213, + 291, + 525 + ], + "spans": [ + { + "bbox": [ + 67, + 213, + 291, + 525 + ], + "type": "text", + "content": "In this section, we discuss prior research related to the radiology report summarization task. The first attempt at automatic summarization of radiology findings into natural language impression statements was proposed by Zhang et al. (2018). Their contribution was to propose a first baseline on the task, using a bidirectional-LSTM as encoder and decoder. Importantly, they found that about " + }, + { + "bbox": [ + 67, + 213, + 291, + 525 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 67, + 213, + 291, + 525 + ], + "type": "text", + "content": " of the summaries generated from neural models contained factual errors. Subsequently, Zhang et al. (2020) proposed the " + }, + { + "bbox": [ + 67, + 213, + 291, + 525 + ], + "type": "inline_equation", + "content": "\\mathrm{F_1}" + }, + { + "bbox": [ + 67, + 213, + 291, + 525 + ], + "type": "text", + "content": " CheXbert score to evaluate the factual correctness of the generated impression. They also used reinforcement learning to optimize the " + }, + { + "bbox": [ + 67, + 213, + 291, + 525 + ], + "type": "inline_equation", + "content": "\\mathrm{F_1}" + }, + { + "bbox": [ + 67, + 213, + 291, + 525 + ], + "type": "text", + "content": " CheXbert score directly. Finally, both Hu et al. (2021) and Hu et al. (2022) used the Biomedical and Clinical English Model Packages in the Stanza Python NLP Library (Zhang et al., 2021) to extract medical entities. The former study used the entities to construct a Graph Neural Network, which was used as input in their summarization pipeline. In contrast, the latter study used the entities to mask the findings during contrastive pre-training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 529, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 529, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 529, + 291, + 772 + ], + "type": "text", + "content": "We believe this paper is an original contribution to the aforementioned line of work. As instigated by Zhang et al. (2018), our goal is to release a new summarization corpus and baselines on new modalities and anatomies. We do so by releasing an RRS dataset with data from 11 new modality-anatomy pairs. In addition, we extend the work performed by Zhang et al. (2020) by proposing a new metric to evaluate the factual correctness and completeness of the generated impression, namely the RadGraph score. Finally, we improve on the work of Hu et al. (2021, 2022) in two ways: (1) we use semantic annotations from a pre-trained model trained using annotations from board-certified radiologists, as opposed to Stanza which leverages unsupervised biomedical and clinical text data; (2) we leverage relation annotations between entities, a feature that was not available in prior work." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 302, + 70, + 461, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 70, + 461, + 84 + ], + "spans": [ + { + "bbox": [ + 302, + 70, + 461, + 84 + ], + "type": "text", + "content": "5 Conclusion and Discussion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 302, + 92, + 526, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 92, + 526, + 200 + ], + "spans": [ + { + "bbox": [ + 302, + 92, + 526, + 200 + ], + "type": "text", + "content": "In this paper, we highlight and address several weaknesses associated with the radiology report summarization task. First, from a data perspective, we propose a publicly available dataset named MIMIC-RRS involving data samples from twelve modality-anatomy pairs, with 79,779 samples from MIMIC-III and 128,003 samples from MIMIC-CXR." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 214, + 526, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 214, + 526, + 295 + ], + "spans": [ + { + "bbox": [ + 302, + 214, + 526, + 295 + ], + "type": "text", + "content": "Second, we conducted more than 40 experiments and over 400 cross-modality-anatomy evaluations to benchmark the performance of different models. We show that instead of designing complicated modules, we can start from an appropriate backbone model such as BART." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 309, + 526, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 309, + 526, + 416 + ], + "spans": [ + { + "bbox": [ + 302, + 309, + 526, + 416 + ], + "type": "text", + "content": "Finally, we proposed an elegant and simple metric, F1-RadGraph, to evaluate the factual correctness of summaries generated for any modality and anatomy. In the future, we hope that our work broadens the scope of the radiology report summarization task and contributes to the development of reliable RRS models that generalize well to new anatomies and modalities." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 427, + 367, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 427, + 367, + 439 + ], + "spans": [ + { + "bbox": [ + 302, + 427, + 367, + 439 + ], + "type": "text", + "content": "Limitations" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 448, + 526, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 448, + 526, + 678 + ], + "spans": [ + { + "bbox": [ + 302, + 448, + 526, + 678 + ], + "type": "text", + "content": "We note two limitations of our paper. First, our work does not extensively evaluate all the available pre-trained models that could be suitable for this task, e.g., ELECTRA (Clark et al., 2020), BioLinkBERT (Yasunaga et al., 2022), GatorTron (Yang et al., 2022), RadBERT (Yan et al., 2022), and PubMedBERT (Gu et al., 2021). The aim of this work is not to report the strongest possible score but rather to address weaknesses of existing radiology report summarization studies (in terms of data and evaluation). Yet, we are confident our proposed solutions report a strong baseline for future work. Second, although F1-RadGraph seems like an appropriate metric to evaluate our new modalities and anatomies (and appears to be consistent with ROUGE scores), it has only been evaluated subjectively and not systematically." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 688, + 400, + 702 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 688, + 400, + 702 + ], + "spans": [ + { + "bbox": [ + 302, + 688, + 400, + 702 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 709, + 526, + 763 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 709, + 526, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 709, + 526, + 763 + ], + "type": "text", + "content": "Maya Varma is supported by graduate fellowship awards from the Department of Defense (NDSEG) and the Knight-Hennessy Scholars program at Stanford University." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "473" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 71, + 127, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 71, + 127, + 83 + ], + "spans": [ + { + "bbox": [ + 69, + 71, + 127, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 89, + 291, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 89, + 291, + 145 + ], + "spans": [ + { + "bbox": [ + 67, + 89, + 291, + 145 + ], + "type": "text", + "content": "Robert Alexander, Stephen Waite, Michael A Bruno, Elizabeth A Krupinski, Leonard Berlin, Stephen Macknik, and Susana Martinez-Conde. 2022. Mandating limits on workload, duty, and speed in radiology. *Radiology*, 304(2):274-282." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 153, + 291, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 153, + 291, + 210 + ], + "spans": [ + { + "bbox": [ + 67, + 153, + 291, + 210 + ], + "type": "text", + "content": "Emily Alsentzer, John Murphy, William Boag, WeiHung Weng, Di Jindi, Tristan Naumann, and Matthew McDermott. 2019. Publicly available clinical bert embeddings. In Proceedings of the 2nd Clinical Natural Language Processing Workshop, pages 72-78." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 216, + 291, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 216, + 291, + 295 + ], + "spans": [ + { + "bbox": [ + 67, + 216, + 291, + 295 + ], + "type": "text", + "content": "Asma Ben Abacha, Yassine Mrabet, Yuhao Zhang, Chaitanya Shivade, Curtis Langlotz, and Dina Demner-Fushman. 2021. Overview of the MEDIQA 2021 shared task on summarization in the medical domain. In Proceedings of the 20th Workshop on Biomedical Language Processing, pages 74-85, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 302, + 291, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 302, + 291, + 347 + ], + "spans": [ + { + "bbox": [ + 67, + 302, + 291, + 347 + ], + "type": "text", + "content": "Mythreyi Bhargavan, Adam H Kaye, Howard P Forman, and Jonathan H Sunshine. 2009. Workload of radiologists in united states in 2006-2007 and trends since 1991-1992. Radiology, 252(2):458-467." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 354, + 291, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 354, + 291, + 400 + ], + "spans": [ + { + "bbox": [ + 67, + 354, + 291, + 400 + ], + "type": "text", + "content": "Aurelia Bustos, Antonio Pertusa, Jose-Maria Salinas, and Maria de la Iglesia-Vaya. 2020. Padchest: A large chest x-ray image dataset with multi-label annotated reports. Medical image analysis, 66:101797." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 407, + 291, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 407, + 291, + 452 + ], + "spans": [ + { + "bbox": [ + 67, + 407, + 291, + 452 + ], + "type": "text", + "content": "Kevin Clark, Minh-Thang Luong, Quoc V. Le, and Christopher D. Manning. 2020. ELECTRA: Pretraining text encoders as discriminators rather than generators. In ICLR." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 460, + 291, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 460, + 291, + 516 + ], + "spans": [ + { + "bbox": [ + 67, + 460, + 291, + 516 + ], + "type": "text", + "content": "Jean-Benoit Delbrouck, Pierre Chambon, Christian Bluethgen, Emily Tsai, Omar Almusa, and Curtis P Langlotz. 2022a. Improving the factual correctness of radiology report generation with semantic rewards. arXiv preprint arXiv:2210.12186." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 523, + 291, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 523, + 291, + 624 + ], + "spans": [ + { + "bbox": [ + 67, + 523, + 291, + 624 + ], + "type": "text", + "content": "Jean-benoit Delbrouck, Khaled Saab, Maya Varma, Sabri Eyuboglu, Pierre Chambon, Jared Dunnmon, Juan Zambrano, Akshay Chaudhari, and Curtis Langlotz. 2022b. ViLMedic: a framework for research at the intersection of vision and language in medical AI. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 23-34, Dublin, Ireland. Association for Computational Linguistics." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 631, + 291, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 631, + 291, + 698 + ], + "spans": [ + { + "bbox": [ + 67, + 631, + 291, + 698 + ], + "type": "text", + "content": "Dina Demner-Fushman, Marc D Kohli, Marc B Rosenman, Sonya E Shooshan, Laritza Rodriguez, Sameer Antani, George R Thoma, and Clement J McDonald. 2016. Preparing a collection of radiology examinations for distribution and retrieval. Journal of the American Medical Informatics Association, 23(2):304-310." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 706, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 706, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 706, + 291, + 772 + ], + "type": "text", + "content": "Yu Gu, Robert Tinn, Hao Cheng, Michael Lucas, Naoto Usuyama, Xiaodong Liu, Tristan Naumann, Jianfeng Gao, and Hoifung Poon. 2021. Domain-specific language model pretraining for biomedical natural language processing. ACM Transactions on Computing for Healthcare (HEALTH), 3(1):1-23." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 71, + 526, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 139 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 139 + ], + "type": "text", + "content": "Jinpeng Hu, Jianling Li, Zhihong Chen, Yaling Shen, Yan Song, Xiang Wan, and Tsung-Hui Chang. 2021. Word graph guided summarization for radiology findings. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 4980-4990, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 147, + 526, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 147, + 526, + 215 + ], + "spans": [ + { + "bbox": [ + 302, + 147, + 526, + 215 + ], + "type": "text", + "content": "Jinpeng Hu, Zhuo Li, Zhihong Chen, Zhen Li, Xiang Wan, and Tsung-Hui Chang. 2022. Graph enhanced contrastive learning for radiology findings summarization. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4677-4688." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 222, + 526, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 222, + 526, + 311 + ], + "spans": [ + { + "bbox": [ + 302, + 222, + 526, + 311 + ], + "type": "text", + "content": "Saahil Jain, Ashwin Agrawal, Adriel Saporta, Steven Truong, Du Nguyen Duong, Tan Bui, Pierre Chambon, Yuhao Zhang, Matthew Lungren, Andrew Ng, Curtis Langlotz, Pranav Rajpurkar, and Pranav Rajpurkar. 2021. Radgraph: Extracting clinical entities and relations from radiology reports. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, volume 1." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 319, + 526, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 319, + 526, + 386 + ], + "spans": [ + { + "bbox": [ + 302, + 319, + 526, + 386 + ], + "type": "text", + "content": "Alistair EW Johnson, Tom J Pollard, Seth J Berkowitz, Nathaniel R Greenbaum, Matthew P Lungren, Chihying Deng, Roger G Mark, and Steven Horng. 2019. Mimic-cxr, a de-identified publicly available database of chest radiographs with free-text reports. Scientific data, 6(1):1-8." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 395, + 526, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 395, + 526, + 450 + ], + "spans": [ + { + "bbox": [ + 302, + 395, + 526, + 450 + ], + "type": "text", + "content": "Alistair EW Johnson, Tom J Pollard, Lu Shen, Li-wei H Lehman, Mengling Feng, Mohammad Ghassemi, Benjamin Moody, Peter Szolovits, Leo Anthony Celi, and Roger G Mark. 2016. Mimic-iii, a freely accessible critical care database. Scientific data, 3(1):1-9." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 460, + 526, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 460, + 526, + 526 + ], + "spans": [ + { + "bbox": [ + 302, + 460, + 526, + 526 + ], + "type": "text", + "content": "Sanjeev Kumar Karn, Ning Liu, Hinrich Schütze, and Oladimeji Farri. 2022. Differentiable multi-agent actor-critic for multi-step radiology report summarization. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1542-1553." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 302, + 534, + 526, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 534, + 526, + 613 + ], + "spans": [ + { + "bbox": [ + 302, + 534, + 526, + 613 + ], + "type": "text", + "content": "Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871-7880." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 302, + 621, + 526, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 621, + 526, + 655 + ], + "spans": [ + { + "bbox": [ + 302, + 621, + 526, + 655 + ], + "type": "text", + "content": "Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 302, + 663, + 526, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 663, + 526, + 708 + ], + "spans": [ + { + "bbox": [ + 302, + 663, + 526, + 708 + ], + "type": "text", + "content": "Qiuhao Lu, Dejing Dou, and Thien Nguyen. 2022. Clinical5: A generative language model for clinical text. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 5436-5443." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 302, + 716, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 716, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 716, + 526, + 772 + ], + "type": "text", + "content": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J Liu, et al. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(140):1-67." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "474" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 72, + 291, + 605 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 68, + 72, + 291, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 72, + 291, + 148 + ], + "spans": [ + { + "bbox": [ + 68, + 72, + 291, + 148 + ], + "type": "text", + "content": "Akshay Smit, Saahil Jain, Pranav Rajpurkar, Anuj Parek, Andrew Y Ng, and Matthew Lungren. 2020. Combining automatic labelers and expert annotations for accurate radiology report labeling using bert. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1500-1519." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 158, + 291, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 158, + 291, + 213 + ], + "spans": [ + { + "bbox": [ + 69, + 158, + 291, + 213 + ], + "type": "text", + "content": "An Yan, Julian McAuley, Xing Lu, Jiang Du, Eric Y Chang, Amilcare Gentili, and Chun-Nan Hsu. 2022. Radbert: Adapting transformer-based language models to radiology. Radiology: Artificial Intelligence, 4(4):e210258." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 221, + 290, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 221, + 290, + 288 + ], + "spans": [ + { + "bbox": [ + 69, + 221, + 290, + 288 + ], + "type": "text", + "content": "Xi Yang, Nima PourNejatian, Hoo Chang Shin, Kaleb E Smith, Christopher Parisien, Colin Compas, Cheryl Martin, Mona G Flores, Ying Zhang, Tanja Magoc, et al. 2022. Gatortron: A large clinical language model to unlock patient information from unstructured electronic health records. arXiv preprint arXiv:2203.03540." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 296, + 290, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 296, + 290, + 339 + ], + "spans": [ + { + "bbox": [ + 69, + 296, + 290, + 339 + ], + "type": "text", + "content": "Michihiro Yasunaga, Jure Leskovec, and Percy Liang. 2022. Linkbert: Pretraining language models with document links. In Association for Computational Linguistics (ACL)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 349, + 290, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 349, + 290, + 404 + ], + "spans": [ + { + "bbox": [ + 69, + 349, + 290, + 404 + ], + "type": "text", + "content": "Hongyi Yuan, Zheng Yuan, Ruyi Gan, Jiaxing Zhang, Yutao Xie, and Sheng Yu. 2022. Biobart: Pretraining and evaluation of a biomedical generative language model. In Proceedings of the 21st Workshop on Biomedical Language Processing, pages 97-109." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 412, + 290, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 412, + 290, + 468 + ], + "spans": [ + { + "bbox": [ + 69, + 412, + 290, + 468 + ], + "type": "text", + "content": "Yuhao Zhang, Daisy Yi Ding, Tianpei Qian, Christopher D Manning, and Curtis P Langlotz. 2018. Learning to summarize radiology findings. In Proceedings of the Ninth International Workshop on Health Text Mining and Information Analysis, pages 204-213." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 476, + 290, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 476, + 290, + 542 + ], + "spans": [ + { + "bbox": [ + 69, + 476, + 290, + 542 + ], + "type": "text", + "content": "Yuhao Zhang, Derek Merck, Emily Tsai, Christopher D Manning, and Curtis Langlotz. 2020. Optimizing the factual correctness of a summary: A study of summarizing radiology reports. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5108-5120." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 550, + 290, + 605 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 550, + 290, + 605 + ], + "spans": [ + { + "bbox": [ + 69, + 550, + 290, + 605 + ], + "type": "text", + "content": "Yuhao Zhang, Yuhui Zhang, Peng Qi, Christopher D Manning, and Curtis P Langlotz. 2021. Biomedical and clinical english model packages for the stanza python nlp library. Journal of the American Medical Informatics Association, 28(9):1892-1899." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 790 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 790 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 790 + ], + "type": "text", + "content": "475" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 71, + 235, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 71, + 235, + 84 + ], + "spans": [ + { + "bbox": [ + 68, + 71, + 235, + 84 + ], + "type": "text", + "content": "A Details of RadGraph Scores" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 95, + 243, + 108 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 95, + 243, + 108 + ], + "spans": [ + { + "bbox": [ + 68, + 95, + 243, + 108 + ], + "type": "text", + "content": "A.1 The Introduction of RadGraph" + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 70, + 124, + 289, + 199 + ], + "blocks": [ + { + "bbox": [ + 70, + 124, + 289, + 199 + ], + "lines": [ + { + "bbox": [ + 70, + 124, + 289, + 199 + ], + "spans": [ + { + "bbox": [ + 70, + 124, + 289, + 199 + ], + "type": "image", + "image_path": "c2b5f95e63ae9279e70dfd8e06c1f438fa698ba2922a020d4ebdc1831e88daa4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 206, + 289, + 232 + ], + "lines": [ + { + "bbox": [ + 67, + 206, + 289, + 232 + ], + "spans": [ + { + "bbox": [ + 67, + 206, + 289, + 232 + ], + "type": "text", + "content": "Figure 3: Example of the RadGraph annotations. Figure taken from (Jain et al., 2021)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 247, + 291, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 247, + 291, + 382 + ], + "spans": [ + { + "bbox": [ + 67, + 247, + 291, + 382 + ], + "type": "text", + "content": "To design our new evaluation metric, we leverage the RadGraph dataset (Jain et al., 2021) containing board-certified radiologist annotations of chest X-ray reports, which correspond to 14,579 entities and 10,889 relations. RadGraph has released a PubMedBERT model (Gu et al., 2021) pre-trained on these annotations to annotate new reports. An example of annotation can be seen in Figure 3. Before moving on to the next section, we quickly describe the concept of entities and relations:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 392, + 291, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 392, + 291, + 501 + ], + "spans": [ + { + "bbox": [ + 67, + 392, + 291, + 501 + ], + "type": "text", + "content": "Entities An entity is defined as a continuous span of text that can include one or more adjacent words. Entities in RadGraph center around two concepts: Anatomy and Observation. Three uncertainty levels exist for Observation, leading to four different entities: Anatomy (ANAT-DP), Observation: Definitely Present (OBS-DP), Observation: Uncertain (OBS-U), and Observation: Definitely Absent (OBS-DA)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 511, + 290, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 511, + 290, + 553 + ], + "spans": [ + { + "bbox": [ + 67, + 511, + 290, + 553 + ], + "type": "text", + "content": "Relations A relation is defined as a directed edge between two entities. Three levels exist: Suggestive " + }, + { + "bbox": [ + 67, + 511, + 290, + 553 + ], + "type": "inline_equation", + "content": "\\text{Of}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 67, + 511, + 290, + 553 + ], + "type": "text", + "content": ", Located At (" + }, + { + "bbox": [ + 67, + 511, + 290, + 553 + ], + "type": "inline_equation", + "content": ". \\cdot" + }, + { + "bbox": [ + 67, + 511, + 290, + 553 + ], + "type": "text", + "content": "), and Modify (" + }, + { + "bbox": [ + 67, + 511, + 290, + 553 + ], + "type": "inline_equation", + "content": ". \\cdot" + }, + { + "bbox": [ + 67, + 511, + 290, + 553 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 68, + 564, + 194, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 564, + 194, + 577 + ], + "spans": [ + { + "bbox": [ + 68, + 564, + 194, + 577 + ], + "type": "text", + "content": "A.2 Metric Computation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 582, + 290, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 582, + 290, + 663 + ], + "spans": [ + { + "bbox": [ + 67, + 582, + 290, + 663 + ], + "type": "text", + "content": "Using the RadGraph annotation scheme and pretrained model, we designed an F-score style reward that measures the factual consistency and completeness of the generated impression (also called hypothesis impression) compared to the reference impression." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 665, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 665, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 665, + 291, + 772 + ], + "type": "text", + "content": "To do so, we treat the RadGraph annotations of an impression as a graph " + }, + { + "bbox": [ + 67, + 665, + 291, + 772 + ], + "type": "inline_equation", + "content": "\\mathcal{G}(V,E)" + }, + { + "bbox": [ + 67, + 665, + 291, + 772 + ], + "type": "text", + "content": " with the set of nodes " + }, + { + "bbox": [ + 67, + 665, + 291, + 772 + ], + "type": "inline_equation", + "content": "V = \\{v_{1},v_{2},\\ldots ,v_{|V|}\\}" + }, + { + "bbox": [ + 67, + 665, + 291, + 772 + ], + "type": "text", + "content": " containing the entities and the set of edges " + }, + { + "bbox": [ + 67, + 665, + 291, + 772 + ], + "type": "inline_equation", + "content": "E = \\{e_1,e_2,\\dots ,e_{|E|}\\}" + }, + { + "bbox": [ + 67, + 665, + 291, + 772 + ], + "type": "text", + "content": " the relations between pairs of entities. The graph is directed, meaning that the edge " + }, + { + "bbox": [ + 67, + 665, + 291, + 772 + ], + "type": "inline_equation", + "content": "e = (v_{1},v_{2})\\neq" + }, + { + "bbox": [ + 67, + 665, + 291, + 772 + ], + "type": "inline_equation", + "content": "(v_{2},v_{1})" + }, + { + "bbox": [ + 67, + 665, + 291, + 772 + ], + "type": "text", + "content": " . An example is depicted in Figure 4. Each node or edge of the graph also has a label, which" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 71, + 525, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 525, + 111 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 525, + 111 + ], + "type": "text", + "content": "we denote as " + }, + { + "bbox": [ + 302, + 71, + 525, + 111 + ], + "type": "inline_equation", + "content": "v_{i_L}" + }, + { + "bbox": [ + 302, + 71, + 525, + 111 + ], + "type": "text", + "content": " for an entity " + }, + { + "bbox": [ + 302, + 71, + 525, + 111 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 302, + 71, + 525, + 111 + ], + "type": "text", + "content": " (for example \"OBS-DP\" or \"ANAT\") and " + }, + { + "bbox": [ + 302, + 71, + 525, + 111 + ], + "type": "inline_equation", + "content": "e_{ij_L}" + }, + { + "bbox": [ + 302, + 71, + 525, + 111 + ], + "type": "text", + "content": " for a relation " + }, + { + "bbox": [ + 302, + 71, + 525, + 111 + ], + "type": "inline_equation", + "content": "e = (v_i, v_j)" + }, + { + "bbox": [ + 302, + 71, + 525, + 111 + ], + "type": "text", + "content": " (such as \"modified\" or \"located at\")." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 112, + 525, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 112, + 525, + 233 + ], + "spans": [ + { + "bbox": [ + 302, + 112, + 525, + 233 + ], + "type": "text", + "content": "To design our RadGraph score, we focus on the nodes " + }, + { + "bbox": [ + 302, + 112, + 525, + 233 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 302, + 112, + 525, + 233 + ], + "type": "text", + "content": " and whether or not a node has a relation in " + }, + { + "bbox": [ + 302, + 112, + 525, + 233 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 302, + 112, + 525, + 233 + ], + "type": "text", + "content": ". For a hypothesis impression " + }, + { + "bbox": [ + 302, + 112, + 525, + 233 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 302, + 112, + 525, + 233 + ], + "type": "text", + "content": ", we create a new set of triplets " + }, + { + "bbox": [ + 302, + 112, + 525, + 233 + ], + "type": "inline_equation", + "content": "T_{y} = \\{(v_{i}, v_{i_{L}}, \\mathcal{R})\\}_{i=1:|V|}" + }, + { + "bbox": [ + 302, + 112, + 525, + 233 + ], + "type": "text", + "content": ". The value " + }, + { + "bbox": [ + 302, + 112, + 525, + 233 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 302, + 112, + 525, + 233 + ], + "type": "text", + "content": " is 1 if " + }, + { + "bbox": [ + 302, + 112, + 525, + 233 + ], + "type": "inline_equation", + "content": "(v_{i}, v_{j})_{j=1:|E|, i \\neq j} \\in E" + }, + { + "bbox": [ + 302, + 112, + 525, + 233 + ], + "type": "text", + "content": ", 0 otherwise. In other words, a triplet contains an entity, the entity label, and whether or not this entity has a relation. We proceed to construct the same set for the reference report " + }, + { + "bbox": [ + 302, + 112, + 525, + 233 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 302, + 112, + 525, + 233 + ], + "type": "text", + "content": " and denote this set " + }, + { + "bbox": [ + 302, + 112, + 525, + 233 + ], + "type": "inline_equation", + "content": "T_{\\hat{y}}" + }, + { + "bbox": [ + 302, + 112, + 525, + 233 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 234, + 525, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 234, + 525, + 301 + ], + "spans": [ + { + "bbox": [ + 302, + 234, + 525, + 301 + ], + "type": "text", + "content": "Finally, our score is defined as the harmonic mean of precision and recall between the hypothesis set " + }, + { + "bbox": [ + 302, + 234, + 525, + 301 + ], + "type": "inline_equation", + "content": "T_{y}" + }, + { + "bbox": [ + 302, + 234, + 525, + 301 + ], + "type": "text", + "content": " and the reference set " + }, + { + "bbox": [ + 302, + 234, + 525, + 301 + ], + "type": "inline_equation", + "content": "T_{\\hat{y}}" + }, + { + "bbox": [ + 302, + 234, + 525, + 301 + ], + "type": "text", + "content": ", giving a value between 0 and 100. As an illustration, the set " + }, + { + "bbox": [ + 302, + 234, + 525, + 301 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 302, + 234, + 525, + 301 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 302, + 234, + 525, + 301 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 302, + 234, + 525, + 301 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 302, + 234, + 525, + 301 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 302, + 234, + 525, + 301 + ], + "type": "text", + "content": " of the graph " + }, + { + "bbox": [ + 302, + 234, + 525, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 302, + 234, + 525, + 301 + ], + "type": "text", + "content": " in Figure 4 are shown as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 302, + 525, + 329 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 302, + 525, + 329 + ], + "spans": [ + { + "bbox": [ + 302, + 302, + 525, + 329 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} V = \\{\\text {m i l d , f l u i d , o v e r l o a d , o v e r t , p u l m o n a r y}, \\\\ \\text {e d e m a} \\} \\end{array}", + "image_path": "8b083ca90ed743409b17176b585d0aefe4c13b93a687f23025023ecb7a8c7c13.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 330, + 525, + 356 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 330, + 525, + 356 + ], + "spans": [ + { + "bbox": [ + 302, + 330, + 525, + 356 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} E = \\left\\{\\text {(m i l d , o v e r l o a d)}, \\text {(o v e r l o a d , f l u i d)}, \\text {(e d e m a ,} \\right. \\\\ \\text {p u l m o n a r y)} \\} \\end{array}", + "image_path": "dd0757c21def506de51363b6ca60306636751b3a8df6018d50d60b013fe8f9aa.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 357, + 525, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 357, + 525, + 397 + ], + "spans": [ + { + "bbox": [ + 302, + 357, + 525, + 397 + ], + "type": "interline_equation", + "content": "T = \\left\\{\\left(\\text {m i l d , o b s - d p , 1}\\right), \\left(\\text {f l u i d , o b s - d p , 0}\\right), \\left(\\text {o v e r - l o a d , o b s - d p , 1}\\right), \\left(\\text {o v e r t , o b s - d a , 0}\\right), \\left(\\text {p u l m o n a r y , a n a t - d p , 0}\\right), \\left(\\text {e d e m a , o b s - d a , 1}\\right) \\right\\}", + "image_path": "65d2c5ba401452539d596874b142f760dd1a975a862cc00cdf5fc292e5dd1c43.jpg" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 306, + 408, + 524, + 572 + ], + "blocks": [ + { + "bbox": [ + 306, + 408, + 524, + 572 + ], + "lines": [ + { + "bbox": [ + 306, + 408, + 524, + 572 + ], + "spans": [ + { + "bbox": [ + 306, + 408, + 524, + 572 + ], + "type": "image", + "image_path": "97416453b9d6654c726b9901018e3bd8c8599e607683b9ecf6ff0d7dac451c8d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 580, + 525, + 605 + ], + "lines": [ + { + "bbox": [ + 302, + 580, + 525, + 605 + ], + "spans": [ + { + "bbox": [ + 302, + 580, + 525, + 605 + ], + "type": "text", + "content": "Figure 4: Graph view of the RadGraph annotations for the report in Figure 3." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 302, + 629, + 444, + 642 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 629, + 444, + 642 + ], + "spans": [ + { + "bbox": [ + 302, + 629, + 444, + 642 + ], + "type": "text", + "content": "B Code and Data Release" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 302, + 651, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 651, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 651, + 525, + 772 + ], + "type": "text", + "content": "Our research has been carried out using the ViLMedic library (Delbrouck et al., 2022b). Our code is available at https://github.com/jbdel/vilmedic. This link is anonymized and complies with the double-blind review process. More specifically, we release the code of the RadGraph score as well as the training of our baseline. We also release the script to download, pre-process, and split the radiology reports of the MIMIC-III database" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "476" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 71, + 68, + 522, + 164 + ], + "blocks": [ + { + "bbox": [ + 71, + 68, + 522, + 164 + ], + "lines": [ + { + "bbox": [ + 71, + 68, + 522, + 164 + ], + "spans": [ + { + "bbox": [ + 71, + 68, + 522, + 164 + ], + "type": "table", + "html": "
CT SpineCT SinusMR NeckMR Head
low resolution study reveals degenerative OBS-DP change OBS-OP and foraminal ANAT-OP narrowing OBS-OP without gross OBS-DA acute OBS-DA pathology OBS-DA1. sinusitis OBS-DP affecting the left ANAT-DS pheoid anat-OP and ethmoid anat-OP sinus ANAT-OP .2 . opacification OBS-OP of bilateral ANAT-OP mastoid ANAT-OP air cells and fluid OBS-OP seen in the middle ANAT-OP ear ANAT-OP cavities ANAT-OP which may indicate infection OBS-OP .slightly OBS-OP prominent OBS-OP lymph OBS-OP node OBS-OP in the posterior ANAT-OP chain ANAT-OP on the left side ANAT-OP side unchanged OBS-OP from previous examination . no definite evidence of infiltrating OBS-DA mass OBS-DA or definite pathologic adenopathy OBS-DA .1. no acute OBS-DA ischemia OBS-DA .2 . age -appropriate OBS-OP -appropriate atrophy OBS-OP , and chronic OBS-OP small OBS-OP vessel ANAT-OP ischemic OBS-OP changes OBS-OP .3 . there is no occlusion OBS-DA or flow -limiting OBS-DA - limiting stenosis OBS-DA of the arterial ANAT-OP system ANAT-OP of the head and neck
", + "image_path": "34e2f0ac6cd5bf87c3b5498c91b5cafc67dbf29861bc3bb08274e0e3d60f88a0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 173, + 525, + 196 + ], + "lines": [ + { + "bbox": [ + 67, + 173, + 525, + 196 + ], + "spans": [ + { + "bbox": [ + 67, + 173, + 525, + 196 + ], + "type": "text", + "content": "Table 5: Examples of entities detected by RadGraph (used in the " + }, + { + "bbox": [ + 67, + 173, + 525, + 196 + ], + "type": "inline_equation", + "content": "\\mathrm{RG_{ER}}" + }, + { + "bbox": [ + 67, + 173, + 525, + 196 + ], + "type": "text", + "content": " metric) on out-of-domain anatomy/modality radiology reports. Relations are omitted for clarity." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 218, + 291, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 218, + 291, + 380 + ], + "spans": [ + { + "bbox": [ + 67, + 218, + 291, + 380 + ], + "type": "text", + "content": "as per our experiments. To download the MIMIC-III database, researchers are required to formally request access via a process documented on the MIMIC website. There are two key steps that must be completed before access is granted: (i) the researcher must complete a recognized course in protecting human research participants, including Health Insurance Portability and Accountability Act (HIPAA) requirements. (ii) the researcher must sign a data use agreement, which outlines appropriate data usage and security standards, and forbids efforts to identify individual patients." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 391, + 160, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 391, + 160, + 403 + ], + "spans": [ + { + "bbox": [ + 67, + 391, + 160, + 403 + ], + "type": "text", + "content": "C More Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 412, + 290, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 412, + 290, + 491 + ], + "spans": [ + { + "bbox": [ + 67, + 412, + 290, + 491 + ], + "type": "text", + "content": "We present the results (including four metrics, i.e., ROUGE-1, ROUGE-2, ROUGE-L, and RadGraph scores) of all the experiments on Figure 5-9 for further research in this field. We also show the output of RadGraph (for entities) on a few samples of our new dataset in Table 5." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 502, + 180, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 502, + 180, + 515 + ], + "spans": [ + { + "bbox": [ + 67, + 502, + 180, + 515 + ], + "type": "text", + "content": "D Ethics Statement" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 523, + 291, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 523, + 291, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 523, + 291, + 714 + ], + "type": "text", + "content": "The MIMIC-CXR and MIMIC-III datasets are de-identified to satisfy the US Health Insurance Portability and Accountability Act of 1996 (HIPAA) Safe Harbor requirements. Protected health information (PHI) has been removed. Therefore, the ethical approval statement and the need for informed consent were waived for the studies on this database, which was approved by the Massachusetts Institute of Technology (Cambridge, MA) and Beth Israel Deaconess Medical Center (Boston, MA). This research was conducted in accordance with the Declaration of Helsinki, describing the ethical principles of medical research involving human subjects." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "477" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 161, + 271, + 377 + ], + "blocks": [ + { + "bbox": [ + 69, + 161, + 271, + 377 + ], + "lines": [ + { + "bbox": [ + 69, + 161, + 271, + 377 + ], + "spans": [ + { + "bbox": [ + 69, + 161, + 271, + 377 + ], + "type": "image", + "image_path": "ee33f4d0af3955c15a0fd299c144cb6820be34977e226e921b4cabcc16a93d7f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 148, + 386, + 187, + 396 + ], + "lines": [ + { + "bbox": [ + 148, + 386, + 187, + 396 + ], + "spans": [ + { + "bbox": [ + 148, + 386, + 187, + 396 + ], + "type": "text", + "content": "ROUGE-1" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 306, + 161, + 509, + 377 + ], + "blocks": [ + { + "bbox": [ + 306, + 161, + 509, + 377 + ], + "lines": [ + { + "bbox": [ + 306, + 161, + 509, + 377 + ], + "spans": [ + { + "bbox": [ + 306, + 161, + 509, + 377 + ], + "type": "image", + "image_path": "8a3b25fab03d88a532ab372bc72d1847fa90c4d1d60a277843c25981bf34ca3c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 386, + 426, + 396 + ], + "lines": [ + { + "bbox": [ + 388, + 386, + 426, + 396 + ], + "spans": [ + { + "bbox": [ + 388, + 386, + 426, + 396 + ], + "type": "text", + "content": "ROUGE-2" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 70, + 405, + 271, + 619 + ], + "blocks": [ + { + "bbox": [ + 70, + 405, + 271, + 619 + ], + "lines": [ + { + "bbox": [ + 70, + 405, + 271, + 619 + ], + "spans": [ + { + "bbox": [ + 70, + 405, + 271, + 619 + ], + "type": "image", + "image_path": "7dc05ddbe4482c7299afe712e7f22401b67815d2b9d8544b501949b0adefcf19.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 629, + 188, + 639 + ], + "lines": [ + { + "bbox": [ + 149, + 629, + 188, + 639 + ], + "spans": [ + { + "bbox": [ + 149, + 629, + 188, + 639 + ], + "type": "text", + "content": "ROUGE-L" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 306, + 405, + 509, + 620 + ], + "blocks": [ + { + "bbox": [ + 306, + 405, + 509, + 620 + ], + "lines": [ + { + "bbox": [ + 306, + 405, + 509, + 620 + ], + "spans": [ + { + "bbox": [ + 306, + 405, + 509, + 620 + ], + "type": "image", + "image_path": "064769bd195bbfdfbc4a5af3200ac07641ef25a517951c0330bbb57cc6599299.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 377, + 629, + 441, + 640 + ], + "lines": [ + { + "bbox": [ + 377, + 629, + 441, + 640 + ], + "spans": [ + { + "bbox": [ + 377, + 629, + 441, + 640 + ], + "type": "text", + "content": "RadGraph Score" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 653, + 526, + 680 + ], + "lines": [ + { + "bbox": [ + 67, + 653, + 526, + 680 + ], + "spans": [ + { + "bbox": [ + 67, + 653, + 526, + 680 + ], + "type": "text", + "content": "Figure 5: Cross-modality-anatomy results from T5-S are visualized here using heatmpas. Colors from light to dark represent the values from low to high in each column." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "478" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 161, + 271, + 377 + ], + "blocks": [ + { + "bbox": [ + 69, + 161, + 271, + 377 + ], + "lines": [ + { + "bbox": [ + 69, + 161, + 271, + 377 + ], + "spans": [ + { + "bbox": [ + 69, + 161, + 271, + 377 + ], + "type": "image", + "image_path": "7c5bc00fd7b4a190f8a1537f7e1fbd1efb2df1611316e2587fd11f12751c4438.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 148, + 386, + 187, + 396 + ], + "lines": [ + { + "bbox": [ + 148, + 386, + 187, + 396 + ], + "spans": [ + { + "bbox": [ + 148, + 386, + 187, + 396 + ], + "type": "text", + "content": "ROUGE-1" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 306, + 161, + 509, + 377 + ], + "blocks": [ + { + "bbox": [ + 306, + 161, + 509, + 377 + ], + "lines": [ + { + "bbox": [ + 306, + 161, + 509, + 377 + ], + "spans": [ + { + "bbox": [ + 306, + 161, + 509, + 377 + ], + "type": "image", + "image_path": "38a5b92cf5ca425c6d7f47c2742119166fdea4b13e43deee0673fe95a707fb12.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 386, + 426, + 396 + ], + "lines": [ + { + "bbox": [ + 388, + 386, + 426, + 396 + ], + "spans": [ + { + "bbox": [ + 388, + 386, + 426, + 396 + ], + "type": "text", + "content": "ROUGE-2" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 70, + 405, + 271, + 619 + ], + "blocks": [ + { + "bbox": [ + 70, + 405, + 271, + 619 + ], + "lines": [ + { + "bbox": [ + 70, + 405, + 271, + 619 + ], + "spans": [ + { + "bbox": [ + 70, + 405, + 271, + 619 + ], + "type": "image", + "image_path": "570950fb54e59fb1aaa2dd7efcdb9a61be3e109e0a36e840befe716be5f1124e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 629, + 188, + 639 + ], + "lines": [ + { + "bbox": [ + 149, + 629, + 188, + 639 + ], + "spans": [ + { + "bbox": [ + 149, + 629, + 188, + 639 + ], + "type": "text", + "content": "ROUGE-L" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 306, + 405, + 509, + 620 + ], + "blocks": [ + { + "bbox": [ + 306, + 405, + 509, + 620 + ], + "lines": [ + { + "bbox": [ + 306, + 405, + 509, + 620 + ], + "spans": [ + { + "bbox": [ + 306, + 405, + 509, + 620 + ], + "type": "image", + "image_path": "d22930f2b39cc1358c376d653ad55150a3d60b85e9dd3217201ec04711ada013.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 377, + 629, + 441, + 640 + ], + "lines": [ + { + "bbox": [ + 377, + 629, + 441, + 640 + ], + "spans": [ + { + "bbox": [ + 377, + 629, + 441, + 640 + ], + "type": "text", + "content": "RadGraph Score" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 653, + 526, + 680 + ], + "lines": [ + { + "bbox": [ + 67, + 653, + 526, + 680 + ], + "spans": [ + { + "bbox": [ + 67, + 653, + 526, + 680 + ], + "type": "text", + "content": "Figure 6: Cross-modality-anatomy results from BART-B are visualized here using heatmaps. Colors from light to dark represent the values from low to high in each column." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "479" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 161, + 270, + 377 + ], + "blocks": [ + { + "bbox": [ + 70, + 161, + 270, + 377 + ], + "lines": [ + { + "bbox": [ + 70, + 161, + 270, + 377 + ], + "spans": [ + { + "bbox": [ + 70, + 161, + 270, + 377 + ], + "type": "image", + "image_path": "fa7dd4a837009612dd7b747453bc1f6037ecbe4aa426564fbf098aca89742521.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 148, + 386, + 187, + 396 + ], + "lines": [ + { + "bbox": [ + 148, + 386, + 187, + 396 + ], + "spans": [ + { + "bbox": [ + 148, + 386, + 187, + 396 + ], + "type": "text", + "content": "ROUGE-1" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 305, + 161, + 509, + 377 + ], + "blocks": [ + { + "bbox": [ + 305, + 161, + 509, + 377 + ], + "lines": [ + { + "bbox": [ + 305, + 161, + 509, + 377 + ], + "spans": [ + { + "bbox": [ + 305, + 161, + 509, + 377 + ], + "type": "image", + "image_path": "d428477be2e0fa4e49f3ce17689dfe14c2f95bdb265c7e0684d79bf4a0641d93.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 386, + 426, + 396 + ], + "lines": [ + { + "bbox": [ + 388, + 386, + 426, + 396 + ], + "spans": [ + { + "bbox": [ + 388, + 386, + 426, + 396 + ], + "type": "text", + "content": "ROUGE-2" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 70, + 405, + 270, + 619 + ], + "blocks": [ + { + "bbox": [ + 70, + 405, + 270, + 619 + ], + "lines": [ + { + "bbox": [ + 70, + 405, + 270, + 619 + ], + "spans": [ + { + "bbox": [ + 70, + 405, + 270, + 619 + ], + "type": "image", + "image_path": "f2076082bc66504433f5ecf5dd5d1cc86629e8339887f9c2da1660ccd44be3e0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 629, + 188, + 639 + ], + "lines": [ + { + "bbox": [ + 149, + 629, + 188, + 639 + ], + "spans": [ + { + "bbox": [ + 149, + 629, + 188, + 639 + ], + "type": "text", + "content": "ROUGE-L" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 306, + 405, + 509, + 620 + ], + "blocks": [ + { + "bbox": [ + 306, + 405, + 509, + 620 + ], + "lines": [ + { + "bbox": [ + 306, + 405, + 509, + 620 + ], + "spans": [ + { + "bbox": [ + 306, + 405, + 509, + 620 + ], + "type": "image", + "image_path": "c6bb4b722501db9c7d686922487bf329fb3a05e8ebdc2ffa0e397584b238d8e6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 377, + 629, + 441, + 640 + ], + "lines": [ + { + "bbox": [ + 377, + 629, + 441, + 640 + ], + "spans": [ + { + "bbox": [ + 377, + 629, + 441, + 640 + ], + "type": "text", + "content": "RadGraph Score" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 653, + 526, + 679 + ], + "lines": [ + { + "bbox": [ + 67, + 653, + 526, + 679 + ], + "spans": [ + { + "bbox": [ + 67, + 653, + 526, + 679 + ], + "type": "text", + "content": "Figure 7: Cross-modality-anatomy results from BART-L are visualized here using heatmaps. Colors from light to dark represent the values from low to high in each column." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "480" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 161, + 270, + 377 + ], + "blocks": [ + { + "bbox": [ + 70, + 161, + 270, + 377 + ], + "lines": [ + { + "bbox": [ + 70, + 161, + 270, + 377 + ], + "spans": [ + { + "bbox": [ + 70, + 161, + 270, + 377 + ], + "type": "image", + "image_path": "4c520bb5abbe22fcabc27d0bd6870d48ba950708bae7b5ea54dbf79586afc079.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 148, + 386, + 187, + 396 + ], + "lines": [ + { + "bbox": [ + 148, + 386, + 187, + 396 + ], + "spans": [ + { + "bbox": [ + 148, + 386, + 187, + 396 + ], + "type": "text", + "content": "ROUGE-1" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 306, + 161, + 509, + 377 + ], + "blocks": [ + { + "bbox": [ + 306, + 161, + 509, + 377 + ], + "lines": [ + { + "bbox": [ + 306, + 161, + 509, + 377 + ], + "spans": [ + { + "bbox": [ + 306, + 161, + 509, + 377 + ], + "type": "image", + "image_path": "05a17ab6cdbe137561b71e15c1fa2835b1748e4f96d5e7c4d8f489d55f60bdad.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 386, + 426, + 396 + ], + "lines": [ + { + "bbox": [ + 388, + 386, + 426, + 396 + ], + "spans": [ + { + "bbox": [ + 388, + 386, + 426, + 396 + ], + "type": "text", + "content": "ROUGE-2" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 70, + 405, + 270, + 619 + ], + "blocks": [ + { + "bbox": [ + 70, + 405, + 270, + 619 + ], + "lines": [ + { + "bbox": [ + 70, + 405, + 270, + 619 + ], + "spans": [ + { + "bbox": [ + 70, + 405, + 270, + 619 + ], + "type": "image", + "image_path": "045ea34b4c69f0539d8800597e79569ca0c1f6dfd64f841defd447710dc7015d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 629, + 188, + 639 + ], + "lines": [ + { + "bbox": [ + 149, + 629, + 188, + 639 + ], + "spans": [ + { + "bbox": [ + 149, + 629, + 188, + 639 + ], + "type": "text", + "content": "ROUGE-L" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 306, + 405, + 509, + 620 + ], + "blocks": [ + { + "bbox": [ + 306, + 405, + 509, + 620 + ], + "lines": [ + { + "bbox": [ + 306, + 405, + 509, + 620 + ], + "spans": [ + { + "bbox": [ + 306, + 405, + 509, + 620 + ], + "type": "image", + "image_path": "a28e93aedb83d89a195a25518651ffc2bb5fdd3d196e274255a6c3bb30f7f9bd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 377, + 629, + 441, + 640 + ], + "lines": [ + { + "bbox": [ + 377, + 629, + 441, + 640 + ], + "spans": [ + { + "bbox": [ + 377, + 629, + 441, + 640 + ], + "type": "text", + "content": "RadGraph Score" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 653, + 526, + 680 + ], + "lines": [ + { + "bbox": [ + 67, + 653, + 526, + 680 + ], + "spans": [ + { + "bbox": [ + 67, + 653, + 526, + 680 + ], + "type": "text", + "content": "Figure 8: Cross-modality-anatomy results from BioBART-B are visualized here using heatmaps. Colors from light to dark represent the values from low to high in each column." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 305, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 305, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 305, + 791 + ], + "type": "text", + "content": "481" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 161, + 270, + 377 + ], + "blocks": [ + { + "bbox": [ + 70, + 161, + 270, + 377 + ], + "lines": [ + { + "bbox": [ + 70, + 161, + 270, + 377 + ], + "spans": [ + { + "bbox": [ + 70, + 161, + 270, + 377 + ], + "type": "image", + "image_path": "e3e86c1bb2e64c7703a64064869fce07d9fd81f177a8d6aeec13ebd30406634f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 148, + 386, + 187, + 396 + ], + "lines": [ + { + "bbox": [ + 148, + 386, + 187, + 396 + ], + "spans": [ + { + "bbox": [ + 148, + 386, + 187, + 396 + ], + "type": "text", + "content": "ROUGE-1" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 306, + 161, + 509, + 377 + ], + "blocks": [ + { + "bbox": [ + 306, + 161, + 509, + 377 + ], + "lines": [ + { + "bbox": [ + 306, + 161, + 509, + 377 + ], + "spans": [ + { + "bbox": [ + 306, + 161, + 509, + 377 + ], + "type": "image", + "image_path": "08186fd00caf82df86aa3a9469e966fcfe201d0246ac986a51811a4d7c681764.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 386, + 426, + 396 + ], + "lines": [ + { + "bbox": [ + 388, + 386, + 426, + 396 + ], + "spans": [ + { + "bbox": [ + 388, + 386, + 426, + 396 + ], + "type": "text", + "content": "ROUGE-2" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 70, + 405, + 271, + 619 + ], + "blocks": [ + { + "bbox": [ + 70, + 405, + 271, + 619 + ], + "lines": [ + { + "bbox": [ + 70, + 405, + 271, + 619 + ], + "spans": [ + { + "bbox": [ + 70, + 405, + 271, + 619 + ], + "type": "image", + "image_path": "3fbc8aab9bfa93c77c1f164ea14aa387df22ae9d8124b7a18a56f875f014bf02.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 629, + 188, + 639 + ], + "lines": [ + { + "bbox": [ + 149, + 629, + 188, + 639 + ], + "spans": [ + { + "bbox": [ + 149, + 629, + 188, + 639 + ], + "type": "text", + "content": "ROUGE-L" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 306, + 405, + 509, + 620 + ], + "blocks": [ + { + "bbox": [ + 306, + 405, + 509, + 620 + ], + "lines": [ + { + "bbox": [ + 306, + 405, + 509, + 620 + ], + "spans": [ + { + "bbox": [ + 306, + 405, + 509, + 620 + ], + "type": "image", + "image_path": "f0a5090e77c9e5fd49ca1a9227b19307f6e367765d2763718c8d172171a135cf.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 377, + 629, + 441, + 640 + ], + "lines": [ + { + "bbox": [ + 377, + 629, + 441, + 640 + ], + "spans": [ + { + "bbox": [ + 377, + 629, + 441, + 640 + ], + "type": "text", + "content": "RadGraph Score" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 653, + 526, + 679 + ], + "lines": [ + { + "bbox": [ + 67, + 653, + 526, + 679 + ], + "spans": [ + { + "bbox": [ + 67, + 653, + 526, + 679 + ], + "type": "text", + "content": "Figure 9: Cross-modality-anatomy results from BioBART-L are visualized here using heatmaps. Colors from light to dark represent the values from low to high in each column." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "482" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 90, + 192, + 103 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 90, + 192, + 103 + ], + "spans": [ + { + "bbox": [ + 68, + 90, + 192, + 103 + ], + "type": "text", + "content": "A For every submission:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 106, + 317, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 106, + 317, + 121 + ], + "spans": [ + { + "bbox": [ + 77, + 106, + 317, + 121 + ], + "type": "text", + "content": "A1. Did you describe the limitations of your work?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 90, + 121, + 139, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 121, + 139, + 133 + ], + "spans": [ + { + "bbox": [ + 90, + 121, + 139, + 133 + ], + "type": "text", + "content": "On Page 5." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 142, + 329, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 142, + 329, + 156 + ], + "spans": [ + { + "bbox": [ + 77, + 142, + 329, + 156 + ], + "type": "text", + "content": "A2. Did you discuss any potential risks of your work?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 90, + 157, + 139, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 157, + 139, + 169 + ], + "spans": [ + { + "bbox": [ + 90, + 157, + 139, + 169 + ], + "type": "text", + "content": "On Page 5." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 77, + 179, + 414, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 179, + 414, + 193 + ], + "spans": [ + { + "bbox": [ + 77, + 179, + 414, + 193 + ], + "type": "text", + "content": "A3. Do the abstract and introduction summarize the paper's main claims?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 90, + 194, + 171, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 194, + 171, + 206 + ], + "spans": [ + { + "bbox": [ + 90, + 194, + 171, + 206 + ], + "type": "text", + "content": "On Pages 1 and 4." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 77, + 215, + 398, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 215, + 398, + 229 + ], + "spans": [ + { + "bbox": [ + 77, + 215, + 398, + 229 + ], + "type": "text", + "content": "A4. Have you used AI writing assistants when working on this paper?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 90, + 230, + 138, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 230, + 138, + 242 + ], + "spans": [ + { + "bbox": [ + 90, + 230, + 138, + 242 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 68, + 252, + 290, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 252, + 290, + 266 + ], + "spans": [ + { + "bbox": [ + 68, + 252, + 290, + 266 + ], + "type": "text", + "content": "B Did you use or create scientific artifacts?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 79, + 270, + 124, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 270, + 124, + 282 + ], + "spans": [ + { + "bbox": [ + 79, + 270, + 124, + 282 + ], + "type": "text", + "content": "Section 2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 77, + 291, + 315, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 291, + 315, + 306 + ], + "spans": [ + { + "bbox": [ + 77, + 291, + 315, + 306 + ], + "type": "text", + "content": "B1. Did you cite the creators of artifacts you used?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 90, + 306, + 135, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 306, + 135, + 317 + ], + "spans": [ + { + "bbox": [ + 90, + 306, + 135, + 317 + ], + "type": "text", + "content": "Section 2." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 77, + 327, + 463, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 327, + 463, + 342 + ], + "spans": [ + { + "bbox": [ + 77, + 327, + 463, + 342 + ], + "type": "text", + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 90, + 343, + 139, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 343, + 139, + 355 + ], + "spans": [ + { + "bbox": [ + 90, + 343, + 139, + 355 + ], + "type": "text", + "content": "On Page 5." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 77, + 364, + 524, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 364, + 524, + 417 + ], + "spans": [ + { + "bbox": [ + 77, + 364, + 524, + 417 + ], + "type": "text", + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 90, + 419, + 135, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 419, + 135, + 430 + ], + "spans": [ + { + "bbox": [ + 90, + 419, + 135, + 430 + ], + "type": "text", + "content": "Section 2." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 77, + 440, + 524, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 440, + 524, + 481 + ], + "spans": [ + { + "bbox": [ + 77, + 440, + 524, + 481 + ], + "type": "text", + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 90, + 482, + 135, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 482, + 135, + 493 + ], + "spans": [ + { + "bbox": [ + 90, + 482, + 135, + 493 + ], + "type": "text", + "content": "Section 2." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 77, + 503, + 524, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 503, + 524, + 531 + ], + "spans": [ + { + "bbox": [ + 77, + 503, + 524, + 531 + ], + "type": "text", + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 90, + 532, + 135, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 532, + 135, + 544 + ], + "spans": [ + { + "bbox": [ + 90, + 532, + 135, + 544 + ], + "type": "text", + "content": "Section 2." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 77, + 553, + 524, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 553, + 524, + 622 + ], + "spans": [ + { + "bbox": [ + 77, + 553, + 524, + 622 + ], + "type": "text", + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 90, + 623, + 135, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 623, + 135, + 634 + ], + "spans": [ + { + "bbox": [ + 90, + 623, + 135, + 634 + ], + "type": "text", + "content": "Section 2." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "spans": [ + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "type": "text", + "content": "C Did you run computational experiments?" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 79, + 661, + 124, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 661, + 124, + 673 + ], + "spans": [ + { + "bbox": [ + 79, + 661, + 124, + 673 + ], + "type": "text", + "content": "Section 3." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 76, + 684, + 524, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 684, + 524, + 711 + ], + "spans": [ + { + "bbox": [ + 76, + 684, + 524, + 711 + ], + "type": "text", + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 89, + 712, + 420, + 724 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 712, + 420, + 724 + ], + "spans": [ + { + "bbox": [ + 89, + 712, + 420, + 724 + ], + "type": "text", + "content": "Not applicable. We use the common pre-trained models in our experiments." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "spans": [ + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "type": "text", + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "spans": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "text", + "content": "ACL 2023 Responsible NLP Checklist" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "483" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 77, + 70, + 523, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 70, + 523, + 97 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 523, + 97 + ], + "type": "text", + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 89, + 99, + 135, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 99, + 135, + 111 + ], + "spans": [ + { + "bbox": [ + 89, + 99, + 135, + 111 + ], + "type": "text", + "content": "Section 3." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 120, + 525, + 160 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 120, + 525, + 160 + ], + "spans": [ + { + "bbox": [ + 77, + 120, + 525, + 160 + ], + "type": "text", + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 162, + 135, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 162, + 135, + 173 + ], + "spans": [ + { + "bbox": [ + 89, + 162, + 135, + 173 + ], + "type": "text", + "content": "Section 3." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 183, + 525, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 183, + 525, + 223 + ], + "spans": [ + { + "bbox": [ + 77, + 183, + 525, + 223 + ], + "type": "text", + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 225, + 135, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 225, + 135, + 237 + ], + "spans": [ + { + "bbox": [ + 89, + 225, + 135, + 237 + ], + "type": "text", + "content": "Section 3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 246, + 522, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 246, + 522, + 261 + ], + "spans": [ + { + "bbox": [ + 68, + 246, + 522, + 261 + ], + "type": "text", + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 79, + 264, + 155, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 264, + 155, + 276 + ], + "spans": [ + { + "bbox": [ + 79, + 264, + 155, + 276 + ], + "type": "text", + "content": "Sections 2 and 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 77, + 285, + 525, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 285, + 525, + 313 + ], + "spans": [ + { + "bbox": [ + 77, + 285, + 525, + 313 + ], + "type": "text", + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 89, + 315, + 135, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 315, + 135, + 326 + ], + "spans": [ + { + "bbox": [ + 89, + 315, + 135, + 326 + ], + "type": "text", + "content": "Section 2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 77, + 335, + 525, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 335, + 525, + 376 + ], + "spans": [ + { + "bbox": [ + 77, + 335, + 525, + 376 + ], + "type": "text", + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 89, + 378, + 135, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 378, + 135, + 389 + ], + "spans": [ + { + "bbox": [ + 89, + 378, + 135, + 389 + ], + "type": "text", + "content": "Section 2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 77, + 398, + 525, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 398, + 525, + 439 + ], + "spans": [ + { + "bbox": [ + 77, + 398, + 525, + 439 + ], + "type": "text", + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 89, + 441, + 135, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 441, + 135, + 452 + ], + "spans": [ + { + "bbox": [ + 89, + 441, + 135, + 452 + ], + "type": "text", + "content": "Section 2." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 77, + 461, + 521, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 461, + 521, + 476 + ], + "spans": [ + { + "bbox": [ + 77, + 461, + 521, + 476 + ], + "type": "text", + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 89, + 476, + 186, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 476, + 186, + 489 + ], + "spans": [ + { + "bbox": [ + 89, + 476, + 186, + 489 + ], + "type": "text", + "content": "Section 2 and Page 5." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 77, + 497, + 525, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 497, + 525, + 524 + ], + "spans": [ + { + "bbox": [ + 77, + 497, + 525, + 524 + ], + "type": "text", + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 89, + 527, + 135, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 527, + 135, + 538 + ], + "spans": [ + { + "bbox": [ + 89, + 527, + 135, + 538 + ], + "type": "text", + "content": "Section 2." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "484" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 15 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Towards Adaptive Prefix Tuning for Parameter-Efficient Language Model Fine-tuning/4802d16c-8245-489f-8ead-82b990dc5bd2_content_list.json b/2023/Towards Adaptive Prefix Tuning for Parameter-Efficient Language Model Fine-tuning/4802d16c-8245-489f-8ead-82b990dc5bd2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e954b2f9c90115287b59e6a4504aca231ac396d9 --- /dev/null +++ b/2023/Towards Adaptive Prefix Tuning for Parameter-Efficient Language Model Fine-tuning/4802d16c-8245-489f-8ead-82b990dc5bd2_content_list.json @@ -0,0 +1,1348 @@ +[ + { + "type": "text", + "text": "Towards Adaptive Prefix Tuning for Parameter-Efficient Language Model Fine-tuning", + "text_level": 1, + "bbox": [ + 221, + 89, + 774, + 129 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhen-Ru Zhang, Chuanqi Tan, Haiyang Xu, Chengyu Wang, Jun Huang, Songfang Huang", + "bbox": [ + 238, + 142, + 766, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Alibaba Group", + "bbox": [ + 438, + 177, + 563, + 193 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhangzhenru.zzr, chuanqi.tcq, shuofeng.xhy}@alibaba-inc.com", + "bbox": [ + 206, + 193, + 794, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{chengyu.wcy,huangjun.hj,songfang.hsf}@alibaba-inc.com", + "bbox": [ + 226, + 210, + 774, + 225 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 252, + 339, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fine-tuning large pre-trained language models on various downstream tasks with whole parameters is prohibitively expensive. Hence, Parameter-efficient fine-tuning has attracted attention that only optimizes a few task-specific parameters with the frozen pre-trained model. In this work, we focus on prefix tuning, which only optimizes continuous prefix vectors (i.e. pseudo tokens) inserted into Transformer layers. Based on the observation that the learned syntax and semantics representation varies a lot at different layers, we argue that the adaptive prefix will be further tailored to each layer than the fixed one, enabling the fine-tuning more effective and efficient. Thus, we propose Adaptive Prefix Tuning (APT) to adjust the prefix in terms of both fine-grained token level and coarse-grained layer level with a gate mechanism. Experiments on the SuperGLUE and NER datasets show the effectiveness of APT. In addition, taking the gate as a probing, we validate the efficiency and effectiveness of the variable prefix.", + "bbox": [ + 141, + 280, + 460, + 607 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 114, + 619, + 260, + 634 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vanilla fine-tuning strategy usually adjusts all the parameters to adapt the pre-trained language model to downstream tasks. Parameter-efficient learning (He et al., 2022; Houlsby et al., 2019; Lester et al., 2021; Guo et al., 2021; Ben Zaken et al., 2022) is an emerging framework that freezes the pre-trained model and only tunes a few number of task-specific parameters for downstream tasks. For instance, Prefix tuning (Li and Liang, 2021; Liu et al., 2022) prepends length-equivalent pseudo prefix tokens, i.e. continuous task-specific vectors to each layer of the pre-trained model, achieving comparable even superior performance with only $0.1 - 3\\%$ parameters.", + "bbox": [ + 112, + 645, + 489, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In previous works, the length of prefix tokens (or the number of trainable parameters) is usually the same at each layer. However, a potential observation lies in that the structure information and", + "bbox": [ + 112, + 854, + 489, + 917 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fe15aa65e38d6f4206afbd88889ca9617d6365cb5ba39e97b808c870e3a4cff1.jpg", + "image_caption": [ + "Figure 1: An illustration of the proposed approach APT where the left is the internal structure of Transformer with inserted prefixes, and the right is the schematic of prefix gate mechanism." + ], + "image_footnote": [], + "bbox": [ + 512, + 253, + 884, + 370 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "representational capacity embedded in each layer are prone to be inconsistent (Jawahar et al., 2019). It is generally considered that the bottom layers of the language model tend to capture concrete and shallow phrase-level features, while the top layers concerns more with abstract semantic information (Tenney et al., 2019). Based on the perspective, we assume adaptive prefix can grab the emphasis more flexibly to adapt to various downstream tasks.", + "bbox": [ + 507, + 466, + 884, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In light of above motivation, we investigate the adaptive prefix in this work. We propose Adaptive Prefix Tuning (APT) with an adaptive gate mechanism at both fine-grained token level and coarse-grained layer level. Specifically, as shown in Figure 1, for fine granularity, APT scores each individual prefix token via gated weight assignment. Then, the scaled weight is utilized to balance the inserted task-specific prefix tokens and original input tokens for current layer at coarse-grained level.", + "bbox": [ + 507, + 612, + 885, + 772 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Extensive experiments against prefix tuning on the sentence and token classification tasks in full data and low resources setting validate the effectiveness of APT. In addition, the gate learned from APT could be served as a probing for the number of necessary parameters in different layers, guiding us to directly apply variable prefix to the original prefix tuning. The probing experiment further demonstrates the effectiveness of adaptive prefix.", + "bbox": [ + 507, + 774, + 885, + 919 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1239", + "bbox": [ + 480, + 927, + 521, + 940 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics", + "bbox": [ + 226, + 945, + 771, + 957 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Volume 2: Short Papers, pages 1239-1248", + "bbox": [ + 368, + 958, + 630, + 971 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "July 9-14, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 295, + 972, + 700, + 985 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 114, + 83, + 278, + 98 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Since fine-tuning the whole model is prohibitively expensive, parameter-efficient language model fine-tuning becomes a lightweight alternative that only optimizes a small number of parameters while keeping most pre-trained parameters frozen (He et al., 2022). Adapter tuning (Houlsby et al., 2019) inserts two tunable task-specific modules after multi-head attention and feed-forward network, achieving comparable performance with only $2 - 4\\%$ of the parameters. Prompt tuning (Lester et al., 2021) and Prefix-Tuning (Li and Liang, 2021) only train soft prompts by adding prefix tokens to the input or hidden states. Recently, Liu et al. (2022) extend the prefix tuning to the natural language understanding tasks, which matches the performance of fine-tuning with only $0.1\\% -3\\%$ tuned parameters.", + "bbox": [ + 112, + 109, + 489, + 366 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, with an overlap of our motivations that each layer of the pre-trained language model focuses on different aspects of feature for various tasks (Jawahar et al., 2019; Clark et al., 2019b) and extra parameters are probably not necessary for certain tasks (Houlsby et al., 2019; Fan et al., 2020; Rücklé et al., 2021), Adaptable Adapters (Moosavi et al., 2022) selects beneficial adapter layers and learns task-specific activation function for downstream tasks to make adaptor dynamic for each task and layer. In addition to different frameworks (adapter versa prefix tuning), our key difference from their work lies in that we aim to dynamically filter required information at each layer in a soft way, while they choose whether to add trainable modules at the layer level in a hard manner.", + "bbox": [ + 115, + 367, + 489, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 112, + 636, + 263, + 653 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1 Prefix Tuning", + "text_level": 1, + "bbox": [ + 112, + 662, + 267, + 678 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As prefix tuning is an extension on Transformer (Vaswani et al., 2017), we first recap the structure of Transformer. Transformer is the block consisting of multi-head attention concatenated by multiple single self-attention functions and a fully connected feed-forward network. Formally speaking, the Transformer block is calculated as follows:", + "bbox": [ + 112, + 682, + 489, + 794 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t n} (\\boldsymbol {Q}, \\boldsymbol {K}, \\boldsymbol {V}) = \\operatorname {s o f t m a x} \\left(\\frac {\\boldsymbol {Q} \\boldsymbol {K} ^ {T}}{\\sqrt {d}} \\boldsymbol {V}\\right) \\quad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 803, + 487, + 839 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {F F N} (\\boldsymbol {x}) = \\operatorname {R e L U} (\\boldsymbol {x} \\boldsymbol {W} _ {1} + \\boldsymbol {b} _ {1}) \\boldsymbol {W} _ {2} + \\boldsymbol {b} _ {2} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 139, + 841, + 487, + 858 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prefix tuning prepends pseudo prefix tokens of length $l$ to each layer of the language model, which is implemented by concatenating inserted", + "bbox": [ + 112, + 871, + 489, + 919 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "keys and values matrix with original corresponding items in each multi-head attention. Specifically, let $P_{k}, P_{v} \\in \\mathbb{R}^{l \\times d}$ be the keys and values of the engaged prefix separately, where $l$ denotes the length of prefix and $d$ corresponds to the dimension, thus self-attention function can be reformatted as:", + "bbox": [ + 507, + 84, + 884, + 180 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t n} \\left(\\boldsymbol {Q}, \\boldsymbol {K} ^ {\\prime}, \\boldsymbol {V} ^ {\\prime}\\right) = \\operatorname {s o f t m a x} \\left(\\frac {\\boldsymbol {Q} \\left(\\boldsymbol {K} ^ {\\prime}\\right) ^ {T}}{\\sqrt {d}} \\boldsymbol {V} ^ {\\prime}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 190, + 882, + 225 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\text {w h e r e} \\boldsymbol {K} ^ {\\prime} = \\left[ \\boldsymbol {P} _ {k}; \\boldsymbol {K} \\right], \\boldsymbol {V} ^ {\\prime} = \\left[ \\boldsymbol {P} _ {v}; \\boldsymbol {V} \\right]\n$$\n", + "text_format": "latex", + "bbox": [ + 546, + 227, + 823, + 244 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Here, $[;]$ donates concatenation function.", + "bbox": [ + 507, + 259, + 815, + 274 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.2 Adaptive Prefix Tuning", + "text_level": 1, + "bbox": [ + 507, + 288, + 741, + 304 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The length of prefix is usually a manually set hyperparameter for each task and fixed in distinct layers of the model. However, existing work demonstrates each layer of the language model pays attention to different aspects of the input feature. We assume the prefix in fixed length is insufficient to tailor different layers and tasks. To dynamically customize the prefix at each layer, APT performs a gate mechanism via fine-grained gated weight assignment and coarse-grained scaled weight specification.", + "bbox": [ + 505, + 309, + 884, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, to capture the diversity of information utilization at different layers, we go deep into the token level at the fine-grained granularity. The token-level gate can inspire us on how many trainable parameters (i.e. pseudo tokens in prefix tuning) are required for this layer, which will be discussed in Section 4.4. Thus, APT yields the gated weights of $l$ pseudo tokens at each layer. We use the hidden states to represent the information encoded in the layer and calculate the gated weights $\\alpha_{i} = [\\alpha_{i1},\\alpha_{i2},\\dots,\\alpha_{il}]$ for $i$ -th layer as:", + "bbox": [ + 507, + 472, + 885, + 650 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\alpha} _ {i} = \\operatorname {s i g m o i d} \\left(\\boldsymbol {h} _ {i - 1} \\boldsymbol {W} _ {i}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 663, + 882, + 678 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Here, $\\pmb{h}_{i-1}$ is the $d$ -dimensional hidden states from the previous layer, and $\\pmb{W}_i \\in \\mathbb{R}^{d \\times l}$ corresponds to the parameters to be learned.", + "bbox": [ + 507, + 694, + 882, + 741 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Besides, we also design a coarse-level gate to balance the information brought from task-specific prefix tokens and original input tokens by learning a layer-level weight. A learnable scaled weight $\\lambda_{i}$ is added to the representation of pseudo prefix tokens at the $i$ -th layer.", + "bbox": [ + 507, + 743, + 880, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the above strategy, the keys-values pair $P_{i} = [P_{ik}, P_{iv}]$ derived from pseudo prefix tokens in $i$ -th layer is updated to $\\hat{P}_{i}$ as:", + "bbox": [ + 507, + 840, + 880, + 888 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {P}} _ {i} = \\lambda_ {i} \\boldsymbol {\\alpha} _ {i} \\odot [ \\boldsymbol {P} _ {i k}, \\boldsymbol {P} _ {i v} ] \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 901, + 882, + 920 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1240", + "bbox": [ + 480, + 927, + 521, + 940 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/b1ad81093b446b5a1b22ef26c897bc77c34543cce612f60a3c3492bffd22c670.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelSuperGLUENER
BoolQCOPARTEWiCWSCAvg.CoNLL03CoNLL04OntoNotesAvg.
BERT-base (110M)FT72.967.068.471.163.568.6----
PT-272.567.471.369.565.469.289.382.687.186.3
APT72.670.072.771.266.970.789.784.187.287.0
BERT-large (335M)FT77.769.070.474.968.372.192.885.689.289.2
PT-275.873.078.375.168.374.190.284.586.487.0
APT76.079.079.475.170.275.990.785.888.688.4
RoBRETa-large (355M)FT86.994.086.675.663.581.392.688.889.890.4
PT-284.893.089.573.463.580.892.888.489.890.3
APT84.894.089.974.668.382.392.789.089.890.5
DeBERTa-xlarge (750M)FT------93.189.190.490.9
PT-2------93.186.590.490.0
APT------93.089.190.590.8
", + "bbox": [ + 117, + 80, + 878, + 286 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/eb2e5d9db59f3eeb55d734dc7c9f437ea7e56af43ffdbed2ddf9d9d8f4e5a108.jpg", + "table_caption": [ + "Table 1: The results on SuperGLUE development set and NER test set in full data setting. The metric of SuperGLUE is accuracy and other is micro-f1 score. Results for FT and PT-2 on BERT-large, RoBRETa-large and DeBERTa-large are token from (Liu et al., 2022). Results for FT on BERT-base are from (Liu et al., 2021). (FT: vanilla fine-tuning; PT-2: P-Tuning v2; APT: Adaptive Prefix Tuning; bold: the best score; underline: the second best)" + ], + "table_footnote": [], + "table_body": "
SettingMethodBoolQCOPARTEWiCWSCAvg.
BERT-base (16-shot)FT47.27.554.06.549.42.750.32.346.26.849.4
PT-252.47.254.23.350.83.148.23.348.54.350.8
APT55.76.557.42.753.14.453.72.255.23.855.0
BERT-large (16-shot)FT57.39.752.02.449.52.750.00.038.72.249.5
PT-250.35.758.25.349.93.449.32.248.14.251.2
APT51.73.560.06.353.94.651.84.855.42.354.6
BERT-base (32-shot)FT48.19.452.26.449.52.749.40.960.43.851.9
PT-250.15.555.03.253.83.452.04.151.54.652.5
APT53.55.357.62.256.51.654.83.954.66.555.4
BERT-large (32-shot)FT47.611.945.03.648.42.250.00.047.313.247.6
PT-245.55.157.46.951.32.353.32.146.07.150.7
APT49.95.962.05.055.53.654.92.849.04.454.3
", + "bbox": [ + 226, + 366, + 771, + 558 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 2: The mean ${}_{std}$ experimental results within 5 random seeds on SuperGLUE development set in 16-shot and 32-shot setting where all metrics are accuracy. bold: the best score.", + "bbox": [ + 112, + 567, + 882, + 596 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\odot$ is the element-wise multiplication. Accordingly, the calculation of the self-attention function in APT is similar to Eq.(3) without further elaboration.", + "bbox": [ + 112, + 621, + 487, + 671 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 112, + 681, + 260, + 697 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 112, + 705, + 317, + 721 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We conduct 5 NLU tasks on SuperGLUE (Wang et al., 2019) benchmark including BoolQ (Clark et al., 2019a), COPA (Roemmle et al., 2011), RTE (Wang et al., 2018), WiC (Pilehvar and Camacho-Collados, 2019) and WSC (Levesque et al., 2012) as well as 3 Named Entity Recognition (NER) tasks including CoNLL03 (Tjong Kim Sang and De Meulder, 2003), CoNLL04 (Carreras and Marquez, 2004), and OntoNotes 5.0 (Weischedel et al., 2013). With BERT-base / large (Devlin et al., 2019) and RoBERTa-large (Liu et al., 2019) instantiated by HuggingFace Transformers (Wolf et al.,", + "bbox": [ + 112, + 726, + 489, + 917 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2020), we compare APT with vanilla fine-tuning and P-Tuning v2 (Liu et al., 2022) which is an implementation of the prefix tuning, configured with hyper-parameters public in the released code1. We also verify our method with DeBERTa-xlarge (He et al., 2020) on NER tasks following P-Tuning v2.", + "bbox": [ + 507, + 621, + 884, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.2 Results", + "text_level": 1, + "bbox": [ + 507, + 730, + 613, + 743 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We report the main results in Table 1. For BERT-base, we can observe that APT achieves $1.5\\%$ and $0.7\\%$ improvements over P-Tuning v2 on SuperGLUE and NER tasks, respectively. For BERT-large, APT outperforms P-Tuning v2 by $1.8\\%$ on SuperGLUE tasks and $1.4\\%$ on NER tasks. For RoBERTa-large, APT surpasses P-Tuning v2 by $1.5\\%$ on SuperGLUE tasks and $0.2\\%$ on NER tasks. On NER tasks with DeBERTa-xlarge, APT is supe", + "bbox": [ + 507, + 750, + 884, + 896 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "", + "bbox": [ + 529, + 904, + 815, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1241", + "bbox": [ + 482, + 927, + 517, + 940 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/bbe35785b8e676de1bdba268ae802c861f00eaab2011d16b85d7b2665c040990.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SettingSuperGLUENER
BoolQCOPARTEWiCWSCAvg.CoNLL03CoNLL04OntoNotesAvg.
APT72.670.072.771.266.970.789.784.187.287.0
w/o token-level α72.669.069.970.865.869.689.583.787.286.8
w/o layer-level λ72.167.471.369.665.469.189.082.686.986.2
w/o hidden states h72.068.868.770.264.668.989.183.687.186.6
", + "bbox": [ + 127, + 80, + 868, + 183 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/abf16e47a310187c3b9123d75cc78ffa492095ff72242b6f5790af9177b45f22.jpg", + "table_caption": [ + "Table 3: Ablation study on BERT-base for two different level gate mechanisms and the hidden states from the previous layer. **bold:** the best score." + ], + "table_footnote": [], + "table_body": "
ModelSuperGLUENER
BoolQCOPARTEWiCWSCAvg.CoNLL03CoNLL04OntoNotesAvg.
PT-272.567.471.369.565.469.289.382.687.186.3
PT-2*72.668.871.970.065.869.889.383.087.286.5
PT-2+72.865.469.171.165.868.889.483.287.186.6
APT72.670.072.771.266.970.789.784.187.287.0
", + "bbox": [ + 166, + 234, + 828, + 337 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 4: Comparison between PT-2 and PT-2*, PT-2+ and APT on BERT-base. (PT-2: P-Tuning v2; PT-2*: PT-2 with variable prefix; PT-2+: PT-2 with enlarged prefix)", + "bbox": [ + 112, + 346, + 882, + 376 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "rior to P-Tuning v2 by an average of $0.8\\%$ . Compared with vanilla fine-tuning, APT is comparable or even better on part of tasks. In addition, we explore the experimental performance under low resource settings on SuperGLUE benchmark. As shown in Table 2, APT is a better few-shot learner than P-Tuning v2, which exceeds $4.2\\%$ , $3.4\\%$ in 16-shot setting, and $2.9\\%$ , $3.6\\%$ in 32-shot setting for BERT-base and BERT-large, respectively.", + "bbox": [ + 112, + 401, + 487, + 546 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 112, + 558, + 280, + 573 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We conduct an ablation study in order to explore the separate effect of token-level gated weight $\\alpha$ , layer-level scaled weight $\\lambda$ and the hidden states $h$ from the previous layer which is used to calculate token-level gated weight $\\alpha$ in Eq.(4). As shown in Table 3, it can be found that removing any strategy hurts the performance to varying degrees, demonstrating that they are all advantageous. Specifically, the beneficial effect of $\\lambda$ for APT is slightly greater than $\\alpha$ overall. Besides, it is effective and meaningful to introduce the context (i.e. the hidden states $h$ from the previous layer) when obtaining the gated weight, especially for SuperGLUE tasks.", + "bbox": [ + 112, + 580, + 489, + 789 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.4 Discussion", + "text_level": 1, + "bbox": [ + 112, + 801, + 243, + 815 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "What is prefix weight distribution learned by APT? The gate mechanism for prefix serves as the key strategy of the proposed APT, where the learned prefix weight distribution turns out to be a critical point. Figure 2 illustrates the gate weights of the pseudo prefix token for COPA and CoNLL04,", + "bbox": [ + 112, + 822, + 489, + 917 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/07649a75104e661b3dbf87701602fa9852bc7bcb688c3443c4cb5b878095dfe4.jpg", + "image_caption": [ + "(a) COPA" + ], + "image_footnote": [], + "bbox": [ + 517, + 403, + 694, + 504 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0416bb02e236d8865949902aacebeab38efa4e6c4fd673a2716986be20f5be9f.jpg", + "image_caption": [ + "(b) CoNLL04", + "Figure 2: Visualization of the learned weights of the prefix token for SuperGLUE task COPA on BERT-large and NER task CoNLL04 on BERT-base, with darker colors indicating higher weights." + ], + "image_footnote": [], + "bbox": [ + 699, + 403, + 875, + 504 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "respectively. It can be found that CoNLL04 is concerned with bottom layers in the language model which are regarded as phrase-level features, while COPA pays more attention to the higher layers, indicating semantic information. The observation is consistent with the characteristics of corresponding tasks. NER is a token-level task while COPA is a causal reasoning task sensitive to the semantics of sentences, which reminds us that it is worth placing various prefix tokens on specific layers according to the task properties.", + "bbox": [ + 505, + 620, + 884, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Does variable prefix work better than fixed one? To verify the effectiveness of adaptive prefix under the proposed architecture, we wonder if the learned ratio at each layer can be directly transferred to P-Tuning v2. Taking the gate as a probing indicator, we reset the prefix length of P-Tuning v2 from fixed to variable in different layers based on the ob", + "bbox": [ + 507, + 806, + 882, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1242", + "bbox": [ + 482, + 927, + 519, + 940 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "servation of the learned ratio (e.g. the distribution shown in Figure 2). From the comparison between PT-2 and $\\mathrm{PT - }2^{*}$ in Table 4, we demonstrate that the variable prefix with less trainable parameters surprisingly outperforms the original implementation in fixed prefix. Nonetheless, it is also worth noting that there is still a gap between P-Tuning v2 with variable prefix and APT, where the latter continuously adjusts the weight of prefix during the training phase while the former only initializes with a one-time mask probing.", + "bbox": [ + 112, + 84, + 489, + 261 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Whether the adaptive structure benefits the fine-tuning? Compared to P-Tuning v2, APT learns extra gated and scaled weights. To figure it out whether the improvement of APT is brought from more trainable parameters or the adaptive model structure, we adjust the hyper-parameter, i.e., enlarge the prefix length of P-Tuning v2 by 1.5 times to align the number of parameters with our APT. As shown in the comparison between PT- $2^{+}$ and APT of Table 4, we observe that APT still outperforms enlarged P-Tuning v2 with $1.9\\%$ , $0.4\\%$ on average for SuperGLUE and NER tasks respectively, validating the superiority of the gate mechanism.", + "bbox": [ + 112, + 275, + 490, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 112, + 502, + 247, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this paper, we investigate prefix tuning and assume that adaptive prefix is probably more efficient and effective than fixed prefix. Firstly, we propose APT that leverages the token-level and the layer-level gate mechanism which achieves an improvement of performance over original prefix tuning. Then, we illustrate the weight distribution learned by APT and take it as a probe, which validates the variable prefix can work better than the fixed one. The above experiments and analysis demonstrate that the adaptive prefix can be served as a promising strategy for parameter-efficient fine-tuning.", + "bbox": [ + 112, + 533, + 489, + 727 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Limitations", + "text_level": 1, + "bbox": [ + 112, + 744, + 220, + 759 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The proposed approach in this paper also suffers from certain limitations, i.e. we adapt APT on the encoder model and lack design for the other architectures such as decoder-only and encoder-decoder. In addition, it is better to generalize the key idea to other parameter-efficient learning approaches. A unified solution for existing work may be worth exploring in the future.", + "bbox": [ + 112, + 774, + 489, + 903 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 510, + 83, + 608, + 98 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Elad Ben Zaken, Yoav Goldberg, and Shauli Ravfogel. 2022. BitFit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 1-9, Dublin, Ireland. Association for Computational Linguistics.", + "Xavier Carreras and Lluis Márquez. 2004. Introduction to the CoNLL-2004 shared task: Semantic role labeling. In Proceedings of the Eighth Conference on Computational Natural Language Learning (CoNLL-2004) at HLT-NAACL 2004, pages 89-97, Boston, Massachusetts, USA. Association for Computational Linguistics.", + "Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. 2019a. BoolQ: Exploring the surprising difficulty of natural yes/no questions. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2924–2936, Minneapolis, Minnesota. Association for Computational Linguistics.", + "Kevin Clark, Urvashi Khandelwal, Omer Levy, and Christopher D. Manning. 2019b. What does BERT look at? an analysis of BERT's attention. In Proceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 276-286, Florence, Italy. Association for Computational Linguistics.", + "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics.", + "Angela Fan, Edouard Grave, and Armand Joulin. 2020. Reducing transformer depth on demand with structured dropout. In International Conference on Learning Representations.", + "Demi Guo, Alexander Rush, and Yoon Kim. 2021. Parameter-efficient transfer learning with diff pruning. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4884-4896, Online. Association for Computational Linguistics.", + "Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. 2022. Towards a unified view of parameter-efficient transfer learning. In International Conference on Learning Representations." + ], + "bbox": [ + 510, + 107, + 884, + 917 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1243", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. 2020. Deberta: Decoding-enhanced bert with disentangled attention. arXiv preprint arXiv:2006.03654.", + "Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. 2019. Parameter-efficient transfer learning for NLP. In Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pages 2790-2799. PMLR.", + "Ganesh Jawahar, Benoit Sagot, and Djamé Seddah. 2019. What does BERT learn about the structure of language? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3651-3657, Florence, Italy. Association for Computational Linguistics.", + "Brian Lester, Rami Al-Rfou, and Noah Constant. 2021. The power of scale for parameter-efficient prompt tuning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3045-3059, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.", + "Hector Levesque, Ernest Davis, and Leora Morgenstern. 2012. The winograd schema challenge. In Thirteenth international conference on the principles of knowledge representation and reasoning.", + "Xiang Lisa Li and Percy Liang. 2021. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4582-4597, Online. Association for Computational Linguistics.", + "Xiao Liu, Kaixuan Ji, Yicheng Fu, Weng Tam, Zhengxiao Du, Zhilin Yang, and Jie Tang. 2022. P-tuning: Prompt tuning can be comparable to fine-tuning across scales and tasks. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 61-68, Dublin, Ireland. Association for Computational Linguistics.", + "Xiao Liu, Yanan Zheng, Zhengxiao Du, Ming Ding, Yujie Qian, Zhilin Yang, and Jie Tang. 2021. Gpt understands, too. arXiv:2103.10385.", + "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692.", + "Nafise Sadat Moosavi, Quentin Delfosse, Kristian Kersting, and Iryna Gurevych. 2022. Adaptable Adapters. In Proceedings of the 2022 Annual Conference of" + ], + "bbox": [ + 115, + 85, + 485, + 917 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the North American Chapter of the Association for Computational Linguistics, Seattle, WA, USA. Association for Computational Linguistics.", + "Mohammad Taher Pilehvar and Jose Camacho-Collados. 2019. WiC: the word-in-context dataset for evaluating context-sensitive meaning representations. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 1267-1273, Minneapolis, Minnesota. Association for Computational Linguistics.", + "Melissa Roemmele, Cosmin Adrian Bejan, and Andrew S Gordon. 2011. Choice of plausible alternatives: An evaluation of commonsense causal reasoning. In AAAI spring symposium: logical formalizations of commonsense reasoning, pages 90-95.", + "Andreas Rückle, Gregor Geigle, Max Glockner, Tilman Beck, Jonas Pfeiffer, Nils Reimers, and Iryna Gurevych. 2021. AdapterDrop: On the efficiency of adapters in transformers. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7930-7946, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.", + "Ian Tenney, Dipanjan Das, and Ellie Pavlick. 2019. BERT rediscovers the classical NLP pipeline. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4593-4601, Florence, Italy. Association for Computational Linguistics.", + "Erik F. Tjong Kim Sang and Fien De Meulder. 2003. Introduction to the CoNLL-2003 shared task: Language-independent named entity recognition. In Proceedings of the Seventh Conference on Natural Language Learning at HLT-NAACL 2003, pages 142-147.", + "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc.", + "Alex Wang, Yada Pruksachatkun, Nikita Nangia, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. 2019. SuperGLUE: A stickier benchmark for general-purpose language understanding systems. In Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc.", + "Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. 2018. GLUE: A multi-task benchmark and analysis platform for natural language understanding. In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 353-355, Brussels, Belgium. Association for Computational Linguistics." + ], + "bbox": [ + 510, + 85, + 880, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1244", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ralph Weischedel, Martha Palmer, Mitchell Marcus, Eduard Hovy, Sameer Pradhan, Lance Ramshaw, Nianwen Xue, Ann Taylor, Jeff Kaufman, Michelle Franchini, Mohammed El-Bachouti, Robert Belvin, and Ann Houston. 2013. OntoNotes Release 5.0. Abacus Data Network.", + "bbox": [ + 115, + 85, + 489, + 164 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online. Association for Computational Linguistics.", + "bbox": [ + 115, + 181, + 489, + 338 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A Experimental Details", + "text_level": 1, + "bbox": [ + 114, + 359, + 337, + 376 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets In the full data setting, all train-dev-test splits follow P-Tuning v2 (Liu et al., 2022). For low resources setting, to generate k-shot ( $k = 16, 32$ ) datasets on SuperGLUE, the fixed set of random seed [11,21,42,87,100] is utilized to sample instances in training and development set, while the entire development set is treated as test set, where the average performance is reported in Table 2.", + "bbox": [ + 112, + 388, + 487, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Experimental Setting We grid search the learning rate over [5e-3, 7e-3, 1e-2, 1e-4], training epoch over [20, 40, 60, 80, 100, 120], batch size over [8, 16, 32], and random seeds over [11, 21, 42, 87, 100]. For a fair comparison, the prefix length utilized by APT is consistent with P-Tuning v2. In low resources setting, the batch size we used is 2. In Eq.(4), we take the hidden states of the first input token as representation in previous layer.", + "bbox": [ + 112, + 533, + 489, + 678 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Experimental Computation We use the pretrained model BERT-base with 110M parameters, BERT-large with 335M parameters, RoBERTa-large with 355M parameters and DeBERTa-xlarge with 750M parameters. We conduct experiments on NVIDIA V100 or A100 GPUs for each task.", + "bbox": [ + 112, + 694, + 489, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "B Further Ablation Results", + "text_level": 1, + "bbox": [ + 112, + 808, + 369, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We demonstrate further ablation results on BERT-large and RoBERTa-large as shown in Table 5. It can be found that the beneficial impact of the three strategies and the observation is consistent with BERT-base in Section 4.3 in general.", + "bbox": [ + 112, + 838, + 489, + 917 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e6bc50894731e36b1bee2fca83254350a9cafe9c3c5973e104559060a17484c6.jpg", + "image_caption": [ + "(a) COPA" + ], + "image_footnote": [], + "bbox": [ + 519, + 86, + 694, + 178 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/904200878bce5b4c5915b7b1a9a571384d8ddbb69cc1d9df19b652004f0281fb.jpg", + "image_caption": [ + "(b) WSC", + "Figure 3: The performance of APT and PT-2 on COPA and WSC in a range of prefix length on BERT-large." + ], + "image_footnote": [], + "bbox": [ + 700, + 86, + 875, + 177 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "C Prefix Length", + "text_level": 1, + "bbox": [ + 507, + 265, + 668, + 282 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The prefix length is an important hyper-parameter for prefix tuning and APT. Figure 3 illustrates the performance of APT and P-Tuning v2 with different prefix lengths over a range. It can be observed that APT is superior to P-Tuning v2 in most prefix length settings, indicating that APT has a relatively wider range of prefix length to achieve better performance.", + "bbox": [ + 507, + 290, + 884, + 418 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "D Scientific Artifacts", + "text_level": 1, + "bbox": [ + 507, + 431, + 709, + 445 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We use datasets involving SuperGLUE (Wang et al., 2019) benchmark including BoolQ (Clark et al., 2019a), COPA (Roemmle et al., 2011), RTE (Wang et al., 2018), WiC (Pilehvar and Camacho-Collados, 2019) and WSC (Levesque et al., 2012) as well as 3 Named Entity Recognition (NER) tasks including CoNLL03 (Tjong Kim Sang and De Meulder, 2003), CoNLL04 (Carreras and Marquez, 2004), and OntoNotes 5.0 (Weischedel et al., 2013). The pre-trained model we used are BERT-base / large (Devlin et al., 2019), RoBERTa-large (Liu et al., 2019) and DeBERTa-xlarge (He et al., 2020). We use HuggingFace Transformers (Wolf et al., 2020) and P-Tuning v2 (Liu et al., 2022) as the codebase implemented by PyTorch. They are all open-source and we only use for academic research which is consistent with their intended use.", + "bbox": [ + 507, + 456, + 884, + 745 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "$^{2}$ https://pytorch.org/", + "bbox": [ + 531, + 903, + 658, + 919 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1245", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7eb934468cc6e92a90aad0be5fb3ea83bf16a423f5b869a2a4501151ef10badc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelSettingSuperGLUENER
BoolQCOPARTEWiCWSCAvg.CoNLL03CoNLL04OntoNotesAvg.
BERT-largeAPT76.079.079.475.170.275.990.785.888.688.4
w/o token-level α75.877.077.374.868.374.691.184.488.588.0
w/o layer-level λ75.474.076.974.668.373.890.783.788.487.6
w/o hidden states h74.776.075.874.668.373.991.284.088.687.9
RoBERTa-largeAPT84.894.089.974.668.382.392.789.089.890.5
w/o token-level α84.388.088.173.065.479.892.288.789.590.1
w/o layer-level λ84.788.086.372.164.479.192.088.789.890.2
w/o hidden states h83.991.087.072.964.479.892.288.789.490.1
", + "bbox": [ + 117, + 399, + 884, + 558 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5: Ablation experiments on BERT-large and RoBERTa-large for two different level gate mechanisms and the hidden states from the previous layer. **bold:** the best score.", + "bbox": [ + 112, + 567, + 882, + 596 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1246", + "bbox": [ + 482, + 928, + 521, + 940 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A For every submission:", + "bbox": [ + 114, + 107, + 322, + 122 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "A1. Did you describe the limitations of your work? section limitations", + "A2. Did you discuss any potential risks of your work? Not applicable. Left blank.", + "A3. Do the abstract and introduction summarize the paper's main claims? section abstract and section 1 introduction", + "A4. Have you used AI writing assistants when working on this paper? Left blank." + ], + "bbox": [ + 127, + 126, + 695, + 287 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "B Did you use or create scientific artifacts?", + "bbox": [ + 114, + 299, + 489, + 316 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "section 4.1", + "bbox": [ + 132, + 322, + 216, + 334 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "B1. Did you cite the creators of artifacts you used? section 4.1", + "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? section D Scientific Artifacts", + "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? section D Scientific Artifacts", + "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? We use open-source datasets and do not change datasets for a fair comparison.", + "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Not applicable. Left blank.", + "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. It can be found in the cited paper." + ], + "bbox": [ + 127, + 346, + 880, + 753 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "C Did you run computational experiments?", + "bbox": [ + 114, + 764, + 492, + 781 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "section 4 Experiments", + "bbox": [ + 132, + 787, + 299, + 801 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Table 1 and section appendix A Experimental Computation", + "bbox": [ + 129, + 812, + 880, + 860 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "ACL 2023 Responsible NLP Checklist", + "bbox": [ + 132, + 84, + 433, + 99 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.", + "bbox": [ + 112, + 868, + 877, + 892 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1247", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? section appendix A Experimental Details", + "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Table 2 report the mean and std results.", + "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? We follow the existing work and keep consistent with them." + ], + "bbox": [ + 129, + 83, + 878, + 282 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank.", + "bbox": [ + 112, + 293, + 877, + 330 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response.", + "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response.", + "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response.", + "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response.", + "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response." + ], + "bbox": [ + 127, + 340, + 878, + 640 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1248", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Towards Adaptive Prefix Tuning for Parameter-Efficient Language Model Fine-tuning/4802d16c-8245-489f-8ead-82b990dc5bd2_model.json b/2023/Towards Adaptive Prefix Tuning for Parameter-Efficient Language Model Fine-tuning/4802d16c-8245-489f-8ead-82b990dc5bd2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b9f36a2d49219992d802c0ef47fce97e87a06231 --- /dev/null +++ b/2023/Towards Adaptive Prefix Tuning for Parameter-Efficient Language Model Fine-tuning/4802d16c-8245-489f-8ead-82b990dc5bd2_model.json @@ -0,0 +1,1837 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.223, + 0.09, + 0.776, + 0.13 + ], + "angle": 0, + "content": "Towards Adaptive Prefix Tuning for Parameter-Efficient Language Model Fine-tuning" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.143, + 0.768, + 0.177 + ], + "angle": 0, + "content": "Zhen-Ru Zhang, Chuanqi Tan, Haiyang Xu, Chengyu Wang, Jun Huang, Songfang Huang" + }, + { + "type": "text", + "bbox": [ + 0.44, + 0.178, + 0.564, + 0.194 + ], + "angle": 0, + "content": "Alibaba Group" + }, + { + "type": "text", + "bbox": [ + 0.208, + 0.194, + 0.796, + 0.21 + ], + "angle": 0, + "content": "{zhangzhenru.zzr, chuanqi.tcq, shuofeng.xhy}@alibaba-inc.com" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.211, + 0.776, + 0.227 + ], + "angle": 0, + "content": "{chengyu.wcy,huangjun.hj,songfang.hsf}@alibaba-inc.com" + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.253, + 0.341, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.281, + 0.461, + 0.608 + ], + "angle": 0, + "content": "Fine-tuning large pre-trained language models on various downstream tasks with whole parameters is prohibitively expensive. Hence, Parameter-efficient fine-tuning has attracted attention that only optimizes a few task-specific parameters with the frozen pre-trained model. In this work, we focus on prefix tuning, which only optimizes continuous prefix vectors (i.e. pseudo tokens) inserted into Transformer layers. Based on the observation that the learned syntax and semantics representation varies a lot at different layers, we argue that the adaptive prefix will be further tailored to each layer than the fixed one, enabling the fine-tuning more effective and efficient. Thus, we propose Adaptive Prefix Tuning (APT) to adjust the prefix in terms of both fine-grained token level and coarse-grained layer level with a gate mechanism. Experiments on the SuperGLUE and NER datasets show the effectiveness of APT. In addition, taking the gate as a probing, we validate the efficiency and effectiveness of the variable prefix." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.62, + 0.262, + 0.636 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.646, + 0.49, + 0.854 + ], + "angle": 0, + "content": "Vanilla fine-tuning strategy usually adjusts all the parameters to adapt the pre-trained language model to downstream tasks. Parameter-efficient learning (He et al., 2022; Houlsby et al., 2019; Lester et al., 2021; Guo et al., 2021; Ben Zaken et al., 2022) is an emerging framework that freezes the pre-trained model and only tunes a few number of task-specific parameters for downstream tasks. For instance, Prefix tuning (Li and Liang, 2021; Liu et al., 2022) prepends length-equivalent pseudo prefix tokens, i.e. continuous task-specific vectors to each layer of the pre-trained model, achieving comparable even superior performance with only \\(0.1 - 3\\%\\) parameters." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.856, + 0.49, + 0.919 + ], + "angle": 0, + "content": "In previous works, the length of prefix tokens (or the number of trainable parameters) is usually the same at each layer. However, a potential observation lies in that the structure information and" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.254, + 0.885, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.382, + 0.885, + 0.44 + ], + "angle": 0, + "content": "Figure 1: An illustration of the proposed approach APT where the left is the internal structure of Transformer with inserted prefixes, and the right is the schematic of prefix gate mechanism." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.467, + 0.885, + 0.611 + ], + "angle": 0, + "content": "representational capacity embedded in each layer are prone to be inconsistent (Jawahar et al., 2019). It is generally considered that the bottom layers of the language model tend to capture concrete and shallow phrase-level features, while the top layers concerns more with abstract semantic information (Tenney et al., 2019). Based on the perspective, we assume adaptive prefix can grab the emphasis more flexibly to adapt to various downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.613, + 0.886, + 0.773 + ], + "angle": 0, + "content": "In light of above motivation, we investigate the adaptive prefix in this work. We propose Adaptive Prefix Tuning (APT) with an adaptive gate mechanism at both fine-grained token level and coarse-grained layer level. Specifically, as shown in Figure 1, for fine granularity, APT scores each individual prefix token via gated weight assignment. Then, the scaled weight is utilized to balance the inserted task-specific prefix tokens and original input tokens for current layer at coarse-grained level." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.775, + 0.886, + 0.92 + ], + "angle": 0, + "content": "Extensive experiments against prefix tuning on the sentence and token classification tasks in full data and low resources setting validate the effectiveness of APT. In addition, the gate learned from APT could be served as a probing for the number of necessary parameters in different layers, guiding us to directly apply variable prefix to the original prefix tuning. The probing experiment further demonstrates the effectiveness of adaptive prefix." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.928, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1239" + }, + { + "type": "footer", + "bbox": [ + 0.227, + 0.946, + 0.772, + 0.958 + ], + "angle": 0, + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + }, + { + "type": "footer", + "bbox": [ + 0.369, + 0.959, + 0.631, + 0.972 + ], + "angle": 0, + "content": "Volume 2: Short Papers, pages 1239-1248" + }, + { + "type": "footer", + "bbox": [ + 0.297, + 0.973, + 0.702, + 0.986 + ], + "angle": 0, + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.115, + 0.084, + 0.279, + 0.099 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.11, + 0.49, + 0.367 + ], + "angle": 0, + "content": "Since fine-tuning the whole model is prohibitively expensive, parameter-efficient language model fine-tuning becomes a lightweight alternative that only optimizes a small number of parameters while keeping most pre-trained parameters frozen (He et al., 2022). Adapter tuning (Houlsby et al., 2019) inserts two tunable task-specific modules after multi-head attention and feed-forward network, achieving comparable performance with only \\(2 - 4\\%\\) of the parameters. Prompt tuning (Lester et al., 2021) and Prefix-Tuning (Li and Liang, 2021) only train soft prompts by adding prefix tokens to the input or hidden states. Recently, Liu et al. (2022) extend the prefix tuning to the natural language understanding tasks, which matches the performance of fine-tuning with only \\(0.1\\% -3\\%\\) tuned parameters." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.368, + 0.49, + 0.626 + ], + "angle": 0, + "content": "Furthermore, with an overlap of our motivations that each layer of the pre-trained language model focuses on different aspects of feature for various tasks (Jawahar et al., 2019; Clark et al., 2019b) and extra parameters are probably not necessary for certain tasks (Houlsby et al., 2019; Fan et al., 2020; Rücklé et al., 2021), Adaptable Adapters (Moosavi et al., 2022) selects beneficial adapter layers and learns task-specific activation function for downstream tasks to make adaptor dynamic for each task and layer. In addition to different frameworks (adapter versa prefix tuning), our key difference from their work lies in that we aim to dynamically filter required information at each layer in a soft way, while they choose whether to add trainable modules at the layer level in a hard manner." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.637, + 0.264, + 0.654 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.663, + 0.268, + 0.679 + ], + "angle": 0, + "content": "3.1 Prefix Tuning" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.683, + 0.49, + 0.795 + ], + "angle": 0, + "content": "As prefix tuning is an extension on Transformer (Vaswani et al., 2017), we first recap the structure of Transformer. Transformer is the block consisting of multi-head attention concatenated by multiple single self-attention functions and a fully connected feed-forward network. Formally speaking, the Transformer block is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.804, + 0.488, + 0.84 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t n} (\\boldsymbol {Q}, \\boldsymbol {K}, \\boldsymbol {V}) = \\operatorname {s o f t m a x} \\left(\\frac {\\boldsymbol {Q} \\boldsymbol {K} ^ {T}}{\\sqrt {d}} \\boldsymbol {V}\\right) \\quad (1)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.842, + 0.488, + 0.859 + ], + "angle": 0, + "content": "\\[\n\\operatorname {F F N} (\\boldsymbol {x}) = \\operatorname {R e L U} (\\boldsymbol {x} \\boldsymbol {W} _ {1} + \\boldsymbol {b} _ {1}) \\boldsymbol {W} _ {2} + \\boldsymbol {b} _ {2} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.872, + 0.49, + 0.92 + ], + "angle": 0, + "content": "Prefix tuning prepends pseudo prefix tokens of length \\( l \\) to each layer of the language model, which is implemented by concatenating inserted" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.181 + ], + "angle": 0, + "content": "keys and values matrix with original corresponding items in each multi-head attention. Specifically, let \\( P_{k}, P_{v} \\in \\mathbb{R}^{l \\times d} \\) be the keys and values of the engaged prefix separately, where \\( l \\) denotes the length of prefix and \\( d \\) corresponds to the dimension, thus self-attention function can be reformatted as:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.191, + 0.883, + 0.226 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t n} \\left(\\boldsymbol {Q}, \\boldsymbol {K} ^ {\\prime}, \\boldsymbol {V} ^ {\\prime}\\right) = \\operatorname {s o f t m a x} \\left(\\frac {\\boldsymbol {Q} \\left(\\boldsymbol {K} ^ {\\prime}\\right) ^ {T}}{\\sqrt {d}} \\boldsymbol {V} ^ {\\prime}\\right) \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.228, + 0.825, + 0.246 + ], + "angle": 0, + "content": "\\[\n\\text {w h e r e} \\boldsymbol {K} ^ {\\prime} = \\left[ \\boldsymbol {P} _ {k}; \\boldsymbol {K} \\right], \\boldsymbol {V} ^ {\\prime} = \\left[ \\boldsymbol {P} _ {v}; \\boldsymbol {V} \\right]\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.26, + 0.816, + 0.275 + ], + "angle": 0, + "content": "Here, \\([;]\\) donates concatenation function." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.289, + 0.742, + 0.305 + ], + "angle": 0, + "content": "3.2 Adaptive Prefix Tuning" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.31, + 0.885, + 0.472 + ], + "angle": 0, + "content": "The length of prefix is usually a manually set hyperparameter for each task and fixed in distinct layers of the model. However, existing work demonstrates each layer of the language model pays attention to different aspects of the input feature. We assume the prefix in fixed length is insufficient to tailor different layers and tasks. To dynamically customize the prefix at each layer, APT performs a gate mechanism via fine-grained gated weight assignment and coarse-grained scaled weight specification." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.473, + 0.886, + 0.651 + ], + "angle": 0, + "content": "Specifically, to capture the diversity of information utilization at different layers, we go deep into the token level at the fine-grained granularity. The token-level gate can inspire us on how many trainable parameters (i.e. pseudo tokens in prefix tuning) are required for this layer, which will be discussed in Section 4.4. Thus, APT yields the gated weights of \\( l \\) pseudo tokens at each layer. We use the hidden states to represent the information encoded in the layer and calculate the gated weights \\( \\alpha_{i} = [\\alpha_{i1},\\alpha_{i2},\\dots,\\alpha_{il}] \\) for \\( i \\)-th layer as:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.664, + 0.883, + 0.68 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\alpha} _ {i} = \\operatorname {s i g m o i d} \\left(\\boldsymbol {h} _ {i - 1} \\boldsymbol {W} _ {i}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.695, + 0.883, + 0.743 + ], + "angle": 0, + "content": "Here, \\( \\pmb{h}_{i-1} \\) is the \\( d \\)-dimensional hidden states from the previous layer, and \\( \\pmb{W}_i \\in \\mathbb{R}^{d \\times l} \\) corresponds to the parameters to be learned." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.744, + 0.882, + 0.84 + ], + "angle": 0, + "content": "Besides, we also design a coarse-level gate to balance the information brought from task-specific prefix tokens and original input tokens by learning a layer-level weight. A learnable scaled weight \\(\\lambda_{i}\\) is added to the representation of pseudo prefix tokens at the \\(i\\)-th layer." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.841, + 0.882, + 0.889 + ], + "angle": 0, + "content": "With the above strategy, the keys-values pair \\( P_{i} = [P_{ik}, P_{iv}] \\) derived from pseudo prefix tokens in \\( i \\)-th layer is updated to \\( \\hat{P}_{i} \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.902, + 0.883, + 0.921 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {P}} _ {i} = \\lambda_ {i} \\boldsymbol {\\alpha} _ {i} \\odot [ \\boldsymbol {P} _ {i k}, \\boldsymbol {P} _ {i v} ] \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.928, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1240" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.118, + 0.082, + 0.88, + 0.287 + ], + "angle": 0, + "content": "
ModelSuperGLUENER
BoolQCOPARTEWiCWSCAvg.CoNLL03CoNLL04OntoNotesAvg.
BERT-base (110M)FT72.967.068.471.163.568.6----
PT-272.567.471.369.565.469.289.382.687.186.3
APT72.670.072.771.266.970.789.784.187.287.0
BERT-large (335M)FT77.769.070.474.968.372.192.885.689.289.2
PT-275.873.078.375.168.374.190.284.586.487.0
APT76.079.079.475.170.275.990.785.888.688.4
RoBRETa-large (355M)FT86.994.086.675.663.581.392.688.889.890.4
PT-284.893.089.573.463.580.892.888.489.890.3
APT84.894.089.974.668.382.392.789.089.890.5
DeBERTa-xlarge (750M)FT------93.189.190.490.9
PT-2------93.186.590.490.0
APT------93.089.190.590.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.296, + 0.885, + 0.355 + ], + "angle": 0, + "content": "Table 1: The results on SuperGLUE development set and NER test set in full data setting. The metric of SuperGLUE is accuracy and other is micro-f1 score. Results for FT and PT-2 on BERT-large, RoBRETa-large and DeBERTa-large are token from (Liu et al., 2022). Results for FT on BERT-base are from (Liu et al., 2021). (FT: vanilla fine-tuning; PT-2: P-Tuning v2; APT: Adaptive Prefix Tuning; bold: the best score; underline: the second best)" + }, + { + "type": "table", + "bbox": [ + 0.227, + 0.367, + 0.773, + 0.56 + ], + "angle": 0, + "content": "
SettingMethodBoolQCOPARTEWiCWSCAvg.
BERT-base (16-shot)FT47.27.554.06.549.42.750.32.346.26.849.4
PT-252.47.254.23.350.83.148.23.348.54.350.8
APT55.76.557.42.753.14.453.72.255.23.855.0
BERT-large (16-shot)FT57.39.752.02.449.52.750.00.038.72.249.5
PT-250.35.758.25.349.93.449.32.248.14.251.2
APT51.73.560.06.353.94.651.84.855.42.354.6
BERT-base (32-shot)FT48.19.452.26.449.52.749.40.960.43.851.9
PT-250.15.555.03.253.83.452.04.151.54.652.5
APT53.55.357.62.256.51.654.83.954.66.555.4
BERT-large (32-shot)FT47.611.945.03.648.42.250.00.047.313.247.6
PT-245.55.157.46.951.32.353.32.146.07.150.7
APT49.95.962.05.055.53.654.92.849.04.454.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.568, + 0.884, + 0.598 + ], + "angle": 0, + "content": "Table 2: The mean \\( {}_{std} \\) experimental results within 5 random seeds on SuperGLUE development set in 16-shot and 32-shot setting where all metrics are accuracy. bold: the best score." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.623, + 0.489, + 0.673 + ], + "angle": 0, + "content": "\\(\\odot\\) is the element-wise multiplication. Accordingly, the calculation of the self-attention function in APT is similar to Eq.(3) without further elaboration." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.682, + 0.262, + 0.699 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.706, + 0.318, + 0.722 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.727, + 0.49, + 0.919 + ], + "angle": 0, + "content": "We conduct 5 NLU tasks on SuperGLUE (Wang et al., 2019) benchmark including BoolQ (Clark et al., 2019a), COPA (Roemmle et al., 2011), RTE (Wang et al., 2018), WiC (Pilehvar and Camacho-Collados, 2019) and WSC (Levesque et al., 2012) as well as 3 Named Entity Recognition (NER) tasks including CoNLL03 (Tjong Kim Sang and De Meulder, 2003), CoNLL04 (Carreras and Marquez, 2004), and OntoNotes 5.0 (Weischedel et al., 2013). With BERT-base / large (Devlin et al., 2019) and RoBERTa-large (Liu et al., 2019) instantiated by HuggingFace Transformers (Wolf et al.," + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.623, + 0.885, + 0.721 + ], + "angle": 0, + "content": "2020), we compare APT with vanilla fine-tuning and P-Tuning v2 (Liu et al., 2022) which is an implementation of the prefix tuning, configured with hyper-parameters public in the released code1. We also verify our method with DeBERTa-xlarge (He et al., 2020) on NER tasks following P-Tuning v2." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.731, + 0.615, + 0.744 + ], + "angle": 0, + "content": "4.2 Results" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.751, + 0.885, + 0.897 + ], + "angle": 0, + "content": "We report the main results in Table 1. For BERT-base, we can observe that APT achieves \\(1.5\\%\\) and \\(0.7\\%\\) improvements over P-Tuning v2 on SuperGLUE and NER tasks, respectively. For BERT-large, APT outperforms P-Tuning v2 by \\(1.8\\%\\) on SuperGLUE tasks and \\(1.4\\%\\) on NER tasks. For RoBERTa-large, APT surpasses P-Tuning v2 by \\(1.5\\%\\) on SuperGLUE tasks and \\(0.2\\%\\) on NER tasks. On NER tasks with DeBERTa-xlarge, APT is supe" + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.905, + 0.816, + 0.919 + ], + "angle": 0, + "content": "" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.928, + 0.519, + 0.941 + ], + "angle": 0, + "content": "1241" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.129, + 0.082, + 0.87, + 0.184 + ], + "angle": 0, + "content": "
SettingSuperGLUENER
BoolQCOPARTEWiCWSCAvg.CoNLL03CoNLL04OntoNotesAvg.
APT72.670.072.771.266.970.789.784.187.287.0
w/o token-level α72.669.069.970.865.869.689.583.787.286.8
w/o layer-level λ72.167.471.369.665.469.189.082.686.986.2
w/o hidden states h72.068.868.770.264.668.989.183.687.186.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.194, + 0.883, + 0.223 + ], + "angle": 0, + "content": "Table 3: Ablation study on BERT-base for two different level gate mechanisms and the hidden states from the previous layer. **bold:** the best score." + }, + { + "type": "table", + "bbox": [ + 0.167, + 0.235, + 0.83, + 0.338 + ], + "angle": 0, + "content": "
ModelSuperGLUENER
BoolQCOPARTEWiCWSCAvg.CoNLL03CoNLL04OntoNotesAvg.
PT-272.567.471.369.565.469.289.382.687.186.3
PT-2*72.668.871.970.065.869.889.383.087.286.5
PT-2+72.865.469.171.165.868.889.483.287.186.6
APT72.670.072.771.266.970.789.784.187.287.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.347, + 0.884, + 0.378 + ], + "angle": 0, + "content": "Table 4: Comparison between PT-2 and PT-2*, PT-2+ and APT on BERT-base. (PT-2: P-Tuning v2; PT-2*: PT-2 with variable prefix; PT-2+: PT-2 with enlarged prefix)" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.402, + 0.489, + 0.548 + ], + "angle": 0, + "content": "rior to P-Tuning v2 by an average of \\(0.8\\%\\). Compared with vanilla fine-tuning, APT is comparable or even better on part of tasks. In addition, we explore the experimental performance under low resource settings on SuperGLUE benchmark. As shown in Table 2, APT is a better few-shot learner than P-Tuning v2, which exceeds \\(4.2\\%\\), \\(3.4\\%\\) in 16-shot setting, and \\(2.9\\%\\), \\(3.6\\%\\) in 32-shot setting for BERT-base and BERT-large, respectively." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.559, + 0.281, + 0.574 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.581, + 0.49, + 0.79 + ], + "angle": 0, + "content": "We conduct an ablation study in order to explore the separate effect of token-level gated weight \\(\\alpha\\), layer-level scaled weight \\(\\lambda\\) and the hidden states \\(h\\) from the previous layer which is used to calculate token-level gated weight \\(\\alpha\\) in Eq.(4). As shown in Table 3, it can be found that removing any strategy hurts the performance to varying degrees, demonstrating that they are all advantageous. Specifically, the beneficial effect of \\(\\lambda\\) for APT is slightly greater than \\(\\alpha\\) overall. Besides, it is effective and meaningful to introduce the context (i.e. the hidden states \\(h\\) from the previous layer) when obtaining the gated weight, especially for SuperGLUE tasks." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.802, + 0.245, + 0.816 + ], + "angle": 0, + "content": "4.4 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.823, + 0.49, + 0.919 + ], + "angle": 0, + "content": "What is prefix weight distribution learned by APT? The gate mechanism for prefix serves as the key strategy of the proposed APT, where the learned prefix weight distribution turns out to be a critical point. Figure 2 illustrates the gate weights of the pseudo prefix token for COPA and CoNLL04," + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.404, + 0.695, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.574, + 0.51, + 0.639, + 0.522 + ], + "angle": 0, + "content": "(a) COPA" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.404, + 0.876, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.743, + 0.509, + 0.832, + 0.522 + ], + "angle": 0, + "content": "(b) CoNLL04" + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.538, + 0.883, + 0.596 + ], + "angle": 0, + "content": "Figure 2: Visualization of the learned weights of the prefix token for SuperGLUE task COPA on BERT-large and NER task CoNLL04 on BERT-base, with darker colors indicating higher weights." + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.621, + 0.885, + 0.798 + ], + "angle": 0, + "content": "respectively. It can be found that CoNLL04 is concerned with bottom layers in the language model which are regarded as phrase-level features, while COPA pays more attention to the higher layers, indicating semantic information. The observation is consistent with the characteristics of corresponding tasks. NER is a token-level task while COPA is a causal reasoning task sensitive to the semantics of sentences, which reminds us that it is worth placing various prefix tokens on specific layers according to the task properties." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.807, + 0.884, + 0.919 + ], + "angle": 0, + "content": "Does variable prefix work better than fixed one? To verify the effectiveness of adaptive prefix under the proposed architecture, we wonder if the learned ratio at each layer can be directly transferred to P-Tuning v2. Taking the gate as a probing indicator, we reset the prefix length of P-Tuning v2 from fixed to variable in different layers based on the ob" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1242" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.49, + 0.262 + ], + "angle": 0, + "content": "servation of the learned ratio (e.g. the distribution shown in Figure 2). From the comparison between PT-2 and \\(\\mathrm{PT - }2^{*}\\) in Table 4, we demonstrate that the variable prefix with less trainable parameters surprisingly outperforms the original implementation in fixed prefix. Nonetheless, it is also worth noting that there is still a gap between P-Tuning v2 with variable prefix and APT, where the latter continuously adjusts the weight of prefix during the training phase while the former only initializes with a one-time mask probing." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.277, + 0.492, + 0.487 + ], + "angle": 0, + "content": "Whether the adaptive structure benefits the fine-tuning? Compared to P-Tuning v2, APT learns extra gated and scaled weights. To figure it out whether the improvement of APT is brought from more trainable parameters or the adaptive model structure, we adjust the hyper-parameter, i.e., enlarge the prefix length of P-Tuning v2 by 1.5 times to align the number of parameters with our APT. As shown in the comparison between PT- \\(2^{+}\\) and APT of Table 4, we observe that APT still outperforms enlarged P-Tuning v2 with \\(1.9\\%\\), \\(0.4\\%\\) on average for SuperGLUE and NER tasks respectively, validating the superiority of the gate mechanism." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.504, + 0.248, + 0.519 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.535, + 0.49, + 0.728 + ], + "angle": 0, + "content": "In this paper, we investigate prefix tuning and assume that adaptive prefix is probably more efficient and effective than fixed prefix. Firstly, we propose APT that leverages the token-level and the layer-level gate mechanism which achieves an improvement of performance over original prefix tuning. Then, we illustrate the weight distribution learned by APT and take it as a probe, which validates the variable prefix can work better than the fixed one. The above experiments and analysis demonstrate that the adaptive prefix can be served as a promising strategy for parameter-efficient fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.745, + 0.221, + 0.76 + ], + "angle": 0, + "content": "Limitations" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.775, + 0.49, + 0.904 + ], + "angle": 0, + "content": "The proposed approach in this paper also suffers from certain limitations, i.e. we adapt APT on the encoder model and lack design for the other architectures such as decoder-only and encoder-decoder. In addition, it is better to generalize the key idea to other parameter-efficient learning approaches. A unified solution for existing work may be worth exploring in the future." + }, + { + "type": "title", + "bbox": [ + 0.511, + 0.084, + 0.61, + 0.099 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.108, + 0.885, + 0.201 + ], + "angle": 0, + "content": "Elad Ben Zaken, Yoav Goldberg, and Shauli Ravfogel. 2022. BitFit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 1-9, Dublin, Ireland. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.21, + 0.885, + 0.303 + ], + "angle": 0, + "content": "Xavier Carreras and Lluis Márquez. 2004. Introduction to the CoNLL-2004 shared task: Semantic role labeling. In Proceedings of the Eighth Conference on Computational Natural Language Learning (CoNLL-2004) at HLT-NAACL 2004, pages 89-97, Boston, Massachusetts, USA. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.313, + 0.885, + 0.432 + ], + "angle": 0, + "content": "Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. 2019a. BoolQ: Exploring the surprising difficulty of natural yes/no questions. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2924–2936, Minneapolis, Minnesota. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.442, + 0.885, + 0.535 + ], + "angle": 0, + "content": "Kevin Clark, Urvashi Khandelwal, Omer Levy, and Christopher D. Manning. 2019b. What does BERT look at? an analysis of BERT's attention. In Proceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 276-286, Florence, Italy. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.545, + 0.885, + 0.663 + ], + "angle": 0, + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.673, + 0.885, + 0.727 + ], + "angle": 0, + "content": "Angela Fan, Edouard Grave, and Armand Joulin. 2020. Reducing transformer depth on demand with structured dropout. In International Conference on Learning Representations." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.737, + 0.885, + 0.843 + ], + "angle": 0, + "content": "Demi Guo, Alexander Rush, and Yoon Kim. 2021. Parameter-efficient transfer learning with diff pruning. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4884-4896, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.853, + 0.885, + 0.918 + ], + "angle": 0, + "content": "Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. 2022. Towards a unified view of parameter-efficient transfer learning. In International Conference on Learning Representations." + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.108, + 0.885, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1243" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.086, + 0.487, + 0.138 + ], + "angle": 0, + "content": "Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. 2020. Deberta: Decoding-enhanced bert with disentangled attention. arXiv preprint arXiv:2006.03654." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.15, + 0.487, + 0.253 + ], + "angle": 0, + "content": "Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. 2019. Parameter-efficient transfer learning for NLP. In Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pages 2790-2799. PMLR." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.266, + 0.487, + 0.343 + ], + "angle": 0, + "content": "Ganesh Jawahar, Benoit Sagot, and Djamé Seddah. 2019. What does BERT learn about the structure of language? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3651-3657, Florence, Italy. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.355, + 0.487, + 0.446 + ], + "angle": 0, + "content": "Brian Lester, Rami Al-Rfou, and Noah Constant. 2021. The power of scale for parameter-efficient prompt tuning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3045-3059, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.458, + 0.487, + 0.51 + ], + "angle": 0, + "content": "Hector Levesque, Ernest Davis, and Leora Morgenstern. 2012. The winograd schema challenge. In Thirteenth international conference on the principles of knowledge representation and reasoning." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.521, + 0.487, + 0.625 + ], + "angle": 0, + "content": "Xiang Lisa Li and Percy Liang. 2021. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4582-4597, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.637, + 0.487, + 0.741 + ], + "angle": 0, + "content": "Xiao Liu, Kaixuan Ji, Yicheng Fu, Weng Tam, Zhengxiao Du, Zhilin Yang, and Jie Tang. 2022. P-tuning: Prompt tuning can be comparable to fine-tuning across scales and tasks. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 61-68, Dublin, Ireland. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.752, + 0.487, + 0.791 + ], + "angle": 0, + "content": "Xiao Liu, Yanan Zheng, Zhengxiao Du, Ming Ding, Yujie Qian, Zhilin Yang, and Jie Tang. 2021. Gpt understands, too. arXiv:2103.10385." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.803, + 0.487, + 0.868 + ], + "angle": 0, + "content": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.879, + 0.487, + 0.918 + ], + "angle": 0, + "content": "Nafise Sadat Moosavi, Quentin Delfosse, Kristian Kersting, and Iryna Gurevych. 2022. Adaptable Adapters. In Proceedings of the 2022 Annual Conference of" + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.487, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.529, + 0.086, + 0.882, + 0.126 + ], + "angle": 0, + "content": "the North American Chapter of the Association for Computational Linguistics, Seattle, WA, USA. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.137, + 0.882, + 0.254 + ], + "angle": 0, + "content": "Mohammad Taher Pilehvar and Jose Camacho-Collados. 2019. WiC: the word-in-context dataset for evaluating context-sensitive meaning representations. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 1267-1273, Minneapolis, Minnesota. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.266, + 0.882, + 0.331 + ], + "angle": 0, + "content": "Melissa Roemmele, Cosmin Adrian Bejan, and Andrew S Gordon. 2011. Choice of plausible alternatives: An evaluation of commonsense causal reasoning. In AAAI spring symposium: logical formalizations of commonsense reasoning, pages 90-95." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.342, + 0.882, + 0.446 + ], + "angle": 0, + "content": "Andreas Rückle, Gregor Geigle, Max Glockner, Tilman Beck, Jonas Pfeiffer, Nils Reimers, and Iryna Gurevych. 2021. AdapterDrop: On the efficiency of adapters in transformers. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7930-7946, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.457, + 0.882, + 0.535 + ], + "angle": 0, + "content": "Ian Tenney, Dipanjan Das, and Ellie Pavlick. 2019. BERT rediscovers the classical NLP pipeline. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4593-4601, Florence, Italy. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.546, + 0.882, + 0.623 + ], + "angle": 0, + "content": "Erik F. Tjong Kim Sang and Fien De Meulder. 2003. Introduction to the CoNLL-2003 shared task: Language-independent named entity recognition. In Proceedings of the Seventh Conference on Natural Language Learning at HLT-NAACL 2003, pages 142-147." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.636, + 0.882, + 0.701 + ], + "angle": 0, + "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.712, + 0.882, + 0.802 + ], + "angle": 0, + "content": "Alex Wang, Yada Pruksachatkun, Nikita Nangia, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. 2019. SuperGLUE: A stickier benchmark for general-purpose language understanding systems. In Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.814, + 0.882, + 0.918 + ], + "angle": 0, + "content": "Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. 2018. GLUE: A multi-task benchmark and analysis platform for natural language understanding. In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 353-355, Brussels, Belgium. Association for Computational Linguistics." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.882, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1244" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.116, + 0.086, + 0.49, + 0.165 + ], + "angle": 0, + "content": "Ralph Weischedel, Martha Palmer, Mitchell Marcus, Eduard Hovy, Sameer Pradhan, Lance Ramshaw, Nianwen Xue, Ann Taylor, Jeff Kaufman, Michelle Franchini, Mohammed El-Bachouti, Robert Belvin, and Ann Houston. 2013. OntoNotes Release 5.0. Abacus Data Network." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.182, + 0.49, + 0.34 + ], + "angle": 0, + "content": "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online. Association for Computational Linguistics." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.36, + 0.338, + 0.377 + ], + "angle": 0, + "content": "A Experimental Details" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.39, + 0.489, + 0.519 + ], + "angle": 0, + "content": "Datasets In the full data setting, all train-dev-test splits follow P-Tuning v2 (Liu et al., 2022). For low resources setting, to generate k-shot (\\(k = 16, 32\\)) datasets on SuperGLUE, the fixed set of random seed [11,21,42,87,100] is utilized to sample instances in training and development set, while the entire development set is treated as test set, where the average performance is reported in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.535, + 0.49, + 0.68 + ], + "angle": 0, + "content": "Experimental Setting We grid search the learning rate over [5e-3, 7e-3, 1e-2, 1e-4], training epoch over [20, 40, 60, 80, 100, 120], batch size over [8, 16, 32], and random seeds over [11, 21, 42, 87, 100]. For a fair comparison, the prefix length utilized by APT is consistent with P-Tuning v2. In low resources setting, the batch size we used is 2. In Eq.(4), we take the hidden states of the first input token as representation in previous layer." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.695, + 0.49, + 0.79 + ], + "angle": 0, + "content": "Experimental Computation We use the pretrained model BERT-base with 110M parameters, BERT-large with 335M parameters, RoBERTa-large with 355M parameters and DeBERTa-xlarge with 750M parameters. We conduct experiments on NVIDIA V100 or A100 GPUs for each task." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.809, + 0.37, + 0.825 + ], + "angle": 0, + "content": "B Further Ablation Results" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.839, + 0.49, + 0.919 + ], + "angle": 0, + "content": "We demonstrate further ablation results on BERT-large and RoBERTa-large as shown in Table 5. It can be found that the beneficial impact of the three strategies and the observation is consistent with BERT-base in Section 4.3 in general." + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.087, + 0.695, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.574, + 0.184, + 0.639, + 0.197 + ], + "angle": 0, + "content": "(a) COPA" + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.087, + 0.877, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.758, + 0.184, + 0.818, + 0.197 + ], + "angle": 0, + "content": "(b) WSC" + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.212, + 0.882, + 0.241 + ], + "angle": 0, + "content": "Figure 3: The performance of APT and PT-2 on COPA and WSC in a range of prefix length on BERT-large." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.266, + 0.669, + 0.283 + ], + "angle": 0, + "content": "C Prefix Length" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.291, + 0.885, + 0.419 + ], + "angle": 0, + "content": "The prefix length is an important hyper-parameter for prefix tuning and APT. Figure 3 illustrates the performance of APT and P-Tuning v2 with different prefix lengths over a range. It can be observed that APT is superior to P-Tuning v2 in most prefix length settings, indicating that APT has a relatively wider range of prefix length to achieve better performance." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.432, + 0.71, + 0.447 + ], + "angle": 0, + "content": "D Scientific Artifacts" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.457, + 0.885, + 0.746 + ], + "angle": 0, + "content": "We use datasets involving SuperGLUE (Wang et al., 2019) benchmark including BoolQ (Clark et al., 2019a), COPA (Roemmle et al., 2011), RTE (Wang et al., 2018), WiC (Pilehvar and Camacho-Collados, 2019) and WSC (Levesque et al., 2012) as well as 3 Named Entity Recognition (NER) tasks including CoNLL03 (Tjong Kim Sang and De Meulder, 2003), CoNLL04 (Carreras and Marquez, 2004), and OntoNotes 5.0 (Weischedel et al., 2013). The pre-trained model we used are BERT-base / large (Devlin et al., 2019), RoBERTa-large (Liu et al., 2019) and DeBERTa-xlarge (He et al., 2020). We use HuggingFace Transformers (Wolf et al., 2020) and P-Tuning v2 (Liu et al., 2022) as the codebase implemented by PyTorch. They are all open-source and we only use for academic research which is consistent with their intended use." + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.904, + 0.659, + 0.92 + ], + "angle": 0, + "content": "\\(^{2}\\)https://pytorch.org/" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1245" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.118, + 0.4, + 0.885, + 0.559 + ], + "angle": 0, + "content": "
ModelSettingSuperGLUENER
BoolQCOPARTEWiCWSCAvg.CoNLL03CoNLL04OntoNotesAvg.
BERT-largeAPT76.079.079.475.170.275.990.785.888.688.4
w/o token-level α75.877.077.374.868.374.691.184.488.588.0
w/o layer-level λ75.474.076.974.668.373.890.783.788.487.6
w/o hidden states h74.776.075.874.668.373.991.284.088.687.9
RoBERTa-largeAPT84.894.089.974.668.382.392.789.089.890.5
w/o token-level α84.388.088.173.065.479.892.288.789.590.1
w/o layer-level λ84.788.086.372.164.479.192.088.789.890.2
w/o hidden states h83.991.087.072.964.479.892.288.789.490.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.568, + 0.883, + 0.597 + ], + "angle": 0, + "content": "Table 5: Ablation experiments on BERT-large and RoBERTa-large for two different level gate mechanisms and the hidden states from the previous layer. **bold:** the best score." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1246" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.133, + 0.085, + 0.435, + 0.1 + ], + "angle": 0, + "content": "ACL 2023 Responsible NLP Checklist" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.108, + 0.323, + 0.123 + ], + "angle": 0, + "content": "A For every submission:" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.127, + 0.533, + 0.158 + ], + "angle": 0, + "content": "A1. Did you describe the limitations of your work? section limitations" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.171, + 0.553, + 0.202 + ], + "angle": 0, + "content": "A2. Did you discuss any potential risks of your work? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.213, + 0.696, + 0.244 + ], + "angle": 0, + "content": "A3. Do the abstract and introduction summarize the paper's main claims? section abstract and section 1 introduction" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.256, + 0.669, + 0.288 + ], + "angle": 0, + "content": "A4. Have you used AI writing assistants when working on this paper? Left blank." + }, + { + "type": "list", + "bbox": [ + 0.129, + 0.127, + 0.696, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.3, + 0.49, + 0.317 + ], + "angle": 0, + "content": "B Did you use or create scientific artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.323, + 0.217, + 0.335 + ], + "angle": 0, + "content": "section 4.1" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.347, + 0.53, + 0.378 + ], + "angle": 0, + "content": "B1. Did you cite the creators of artifacts you used? section 4.1" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.39, + 0.779, + 0.423 + ], + "angle": 0, + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? section D Scientific Artifacts" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.433, + 0.882, + 0.513 + ], + "angle": 0, + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? section D Scientific Artifacts" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.525, + 0.882, + 0.59 + ], + "angle": 0, + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? We use open-source datasets and do not change datasets for a fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.6, + 0.881, + 0.648 + ], + "angle": 0, + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.659, + 0.882, + 0.755 + ], + "angle": 0, + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. It can be found in the cited paper." + }, + { + "type": "list", + "bbox": [ + 0.129, + 0.347, + 0.882, + 0.755 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.765, + 0.494, + 0.782 + ], + "angle": 0, + "content": "C Did you run computational experiments?" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.788, + 0.3, + 0.802 + ], + "angle": 0, + "content": "section 4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.813, + 0.882, + 0.861 + ], + "angle": 0, + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Table 1 and section appendix A Experimental Computation" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.869, + 0.878, + 0.893 + ], + "angle": 0, + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1247" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.13, + 0.084, + 0.88, + 0.133 + ], + "angle": 0, + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? section appendix A Experimental Details" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.143, + 0.88, + 0.207 + ], + "angle": 0, + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Table 2 report the mean and std results." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.219, + 0.88, + 0.283 + ], + "angle": 0, + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? We follow the existing work and keep consistent with them." + }, + { + "type": "list", + "bbox": [ + 0.13, + 0.084, + 0.88, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.294, + 0.878, + 0.331 + ], + "angle": 0, + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.341, + 0.88, + 0.388 + ], + "angle": 0, + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.4, + 0.88, + 0.464 + ], + "angle": 0, + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.476, + 0.88, + 0.539 + ], + "angle": 0, + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.55, + 0.873, + 0.582 + ], + "angle": 0, + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.593, + 0.88, + 0.641 + ], + "angle": 0, + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response." + }, + { + "type": "list", + "bbox": [ + 0.128, + 0.341, + 0.88, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1248" + } + ] +] \ No newline at end of file diff --git a/2023/Towards Adaptive Prefix Tuning for Parameter-Efficient Language Model Fine-tuning/4802d16c-8245-489f-8ead-82b990dc5bd2_origin.pdf b/2023/Towards Adaptive Prefix Tuning for Parameter-Efficient Language Model Fine-tuning/4802d16c-8245-489f-8ead-82b990dc5bd2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b4edad48f4f8818b57a902d9fe6c7da7965242e8 --- /dev/null +++ b/2023/Towards Adaptive Prefix Tuning for Parameter-Efficient Language Model Fine-tuning/4802d16c-8245-489f-8ead-82b990dc5bd2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1f97d94e024d729ea725bb8d3a4bccd9f9a6475c6e0f7e6e368b7df7e9d267cd +size 420873 diff --git a/2023/Towards Adaptive Prefix Tuning for Parameter-Efficient Language Model Fine-tuning/full.md b/2023/Towards Adaptive Prefix Tuning for Parameter-Efficient Language Model Fine-tuning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7566f77677b875f7ecfef0495d98be98d4a5503f --- /dev/null +++ b/2023/Towards Adaptive Prefix Tuning for Parameter-Efficient Language Model Fine-tuning/full.md @@ -0,0 +1,248 @@ +# Towards Adaptive Prefix Tuning for Parameter-Efficient Language Model Fine-tuning + +Zhen-Ru Zhang, Chuanqi Tan, Haiyang Xu, Chengyu Wang, Jun Huang, Songfang Huang + +Alibaba Group + +{zhangzhenru.zzr, chuanqi.tcq, shuofeng.xhy}@alibaba-inc.com + +{chengyu.wcy,huangjun.hj,songfang.hsf}@alibaba-inc.com + +# Abstract + +Fine-tuning large pre-trained language models on various downstream tasks with whole parameters is prohibitively expensive. Hence, Parameter-efficient fine-tuning has attracted attention that only optimizes a few task-specific parameters with the frozen pre-trained model. In this work, we focus on prefix tuning, which only optimizes continuous prefix vectors (i.e. pseudo tokens) inserted into Transformer layers. Based on the observation that the learned syntax and semantics representation varies a lot at different layers, we argue that the adaptive prefix will be further tailored to each layer than the fixed one, enabling the fine-tuning more effective and efficient. Thus, we propose Adaptive Prefix Tuning (APT) to adjust the prefix in terms of both fine-grained token level and coarse-grained layer level with a gate mechanism. Experiments on the SuperGLUE and NER datasets show the effectiveness of APT. In addition, taking the gate as a probing, we validate the efficiency and effectiveness of the variable prefix. + +# 1 Introduction + +Vanilla fine-tuning strategy usually adjusts all the parameters to adapt the pre-trained language model to downstream tasks. Parameter-efficient learning (He et al., 2022; Houlsby et al., 2019; Lester et al., 2021; Guo et al., 2021; Ben Zaken et al., 2022) is an emerging framework that freezes the pre-trained model and only tunes a few number of task-specific parameters for downstream tasks. For instance, Prefix tuning (Li and Liang, 2021; Liu et al., 2022) prepends length-equivalent pseudo prefix tokens, i.e. continuous task-specific vectors to each layer of the pre-trained model, achieving comparable even superior performance with only $0.1 - 3\%$ parameters. + +In previous works, the length of prefix tokens (or the number of trainable parameters) is usually the same at each layer. However, a potential observation lies in that the structure information and + +![](images/fe15aa65e38d6f4206afbd88889ca9617d6365cb5ba39e97b808c870e3a4cff1.jpg) +Figure 1: An illustration of the proposed approach APT where the left is the internal structure of Transformer with inserted prefixes, and the right is the schematic of prefix gate mechanism. + +representational capacity embedded in each layer are prone to be inconsistent (Jawahar et al., 2019). It is generally considered that the bottom layers of the language model tend to capture concrete and shallow phrase-level features, while the top layers concerns more with abstract semantic information (Tenney et al., 2019). Based on the perspective, we assume adaptive prefix can grab the emphasis more flexibly to adapt to various downstream tasks. + +In light of above motivation, we investigate the adaptive prefix in this work. We propose Adaptive Prefix Tuning (APT) with an adaptive gate mechanism at both fine-grained token level and coarse-grained layer level. Specifically, as shown in Figure 1, for fine granularity, APT scores each individual prefix token via gated weight assignment. Then, the scaled weight is utilized to balance the inserted task-specific prefix tokens and original input tokens for current layer at coarse-grained level. + +Extensive experiments against prefix tuning on the sentence and token classification tasks in full data and low resources setting validate the effectiveness of APT. In addition, the gate learned from APT could be served as a probing for the number of necessary parameters in different layers, guiding us to directly apply variable prefix to the original prefix tuning. The probing experiment further demonstrates the effectiveness of adaptive prefix. + +# 2 Related Works + +Since fine-tuning the whole model is prohibitively expensive, parameter-efficient language model fine-tuning becomes a lightweight alternative that only optimizes a small number of parameters while keeping most pre-trained parameters frozen (He et al., 2022). Adapter tuning (Houlsby et al., 2019) inserts two tunable task-specific modules after multi-head attention and feed-forward network, achieving comparable performance with only $2 - 4\%$ of the parameters. Prompt tuning (Lester et al., 2021) and Prefix-Tuning (Li and Liang, 2021) only train soft prompts by adding prefix tokens to the input or hidden states. Recently, Liu et al. (2022) extend the prefix tuning to the natural language understanding tasks, which matches the performance of fine-tuning with only $0.1\% -3\%$ tuned parameters. + +Furthermore, with an overlap of our motivations that each layer of the pre-trained language model focuses on different aspects of feature for various tasks (Jawahar et al., 2019; Clark et al., 2019b) and extra parameters are probably not necessary for certain tasks (Houlsby et al., 2019; Fan et al., 2020; Rücklé et al., 2021), Adaptable Adapters (Moosavi et al., 2022) selects beneficial adapter layers and learns task-specific activation function for downstream tasks to make adaptor dynamic for each task and layer. In addition to different frameworks (adapter versa prefix tuning), our key difference from their work lies in that we aim to dynamically filter required information at each layer in a soft way, while they choose whether to add trainable modules at the layer level in a hard manner. + +# 3 Methodology + +# 3.1 Prefix Tuning + +As prefix tuning is an extension on Transformer (Vaswani et al., 2017), we first recap the structure of Transformer. Transformer is the block consisting of multi-head attention concatenated by multiple single self-attention functions and a fully connected feed-forward network. Formally speaking, the Transformer block is calculated as follows: + +$$ +\operatorname {A t t n} (\boldsymbol {Q}, \boldsymbol {K}, \boldsymbol {V}) = \operatorname {s o f t m a x} \left(\frac {\boldsymbol {Q} \boldsymbol {K} ^ {T}}{\sqrt {d}} \boldsymbol {V}\right) \quad (1) +$$ + +$$ +\operatorname {F F N} (\boldsymbol {x}) = \operatorname {R e L U} (\boldsymbol {x} \boldsymbol {W} _ {1} + \boldsymbol {b} _ {1}) \boldsymbol {W} _ {2} + \boldsymbol {b} _ {2} \tag {2} +$$ + +Prefix tuning prepends pseudo prefix tokens of length $l$ to each layer of the language model, which is implemented by concatenating inserted + +keys and values matrix with original corresponding items in each multi-head attention. Specifically, let $P_{k}, P_{v} \in \mathbb{R}^{l \times d}$ be the keys and values of the engaged prefix separately, where $l$ denotes the length of prefix and $d$ corresponds to the dimension, thus self-attention function can be reformatted as: + +$$ +\operatorname {A t t n} \left(\boldsymbol {Q}, \boldsymbol {K} ^ {\prime}, \boldsymbol {V} ^ {\prime}\right) = \operatorname {s o f t m a x} \left(\frac {\boldsymbol {Q} \left(\boldsymbol {K} ^ {\prime}\right) ^ {T}}{\sqrt {d}} \boldsymbol {V} ^ {\prime}\right) \tag {3} +$$ + +$$ +\text {w h e r e} \boldsymbol {K} ^ {\prime} = \left[ \boldsymbol {P} _ {k}; \boldsymbol {K} \right], \boldsymbol {V} ^ {\prime} = \left[ \boldsymbol {P} _ {v}; \boldsymbol {V} \right] +$$ + +Here, $[;]$ donates concatenation function. + +# 3.2 Adaptive Prefix Tuning + +The length of prefix is usually a manually set hyperparameter for each task and fixed in distinct layers of the model. However, existing work demonstrates each layer of the language model pays attention to different aspects of the input feature. We assume the prefix in fixed length is insufficient to tailor different layers and tasks. To dynamically customize the prefix at each layer, APT performs a gate mechanism via fine-grained gated weight assignment and coarse-grained scaled weight specification. + +Specifically, to capture the diversity of information utilization at different layers, we go deep into the token level at the fine-grained granularity. The token-level gate can inspire us on how many trainable parameters (i.e. pseudo tokens in prefix tuning) are required for this layer, which will be discussed in Section 4.4. Thus, APT yields the gated weights of $l$ pseudo tokens at each layer. We use the hidden states to represent the information encoded in the layer and calculate the gated weights $\alpha_{i} = [\alpha_{i1},\alpha_{i2},\dots,\alpha_{il}]$ for $i$ -th layer as: + +$$ +\boldsymbol {\alpha} _ {i} = \operatorname {s i g m o i d} \left(\boldsymbol {h} _ {i - 1} \boldsymbol {W} _ {i}\right) \tag {4} +$$ + +Here, $\pmb{h}_{i-1}$ is the $d$ -dimensional hidden states from the previous layer, and $\pmb{W}_i \in \mathbb{R}^{d \times l}$ corresponds to the parameters to be learned. + +Besides, we also design a coarse-level gate to balance the information brought from task-specific prefix tokens and original input tokens by learning a layer-level weight. A learnable scaled weight $\lambda_{i}$ is added to the representation of pseudo prefix tokens at the $i$ -th layer. + +With the above strategy, the keys-values pair $P_{i} = [P_{ik}, P_{iv}]$ derived from pseudo prefix tokens in $i$ -th layer is updated to $\hat{P}_{i}$ as: + +$$ +\hat {\boldsymbol {P}} _ {i} = \lambda_ {i} \boldsymbol {\alpha} _ {i} \odot [ \boldsymbol {P} _ {i k}, \boldsymbol {P} _ {i v} ] \tag {5} +$$ + +
ModelSuperGLUENER
BoolQCOPARTEWiCWSCAvg.CoNLL03CoNLL04OntoNotesAvg.
BERT-base (110M)FT72.967.068.471.163.568.6----
PT-272.567.471.369.565.469.289.382.687.186.3
APT72.670.072.771.266.970.789.784.187.287.0
BERT-large (335M)FT77.769.070.474.968.372.192.885.689.289.2
PT-275.873.078.375.168.374.190.284.586.487.0
APT76.079.079.475.170.275.990.785.888.688.4
RoBRETa-large (355M)FT86.994.086.675.663.581.392.688.889.890.4
PT-284.893.089.573.463.580.892.888.489.890.3
APT84.894.089.974.668.382.392.789.089.890.5
DeBERTa-xlarge (750M)FT------93.189.190.490.9
PT-2------93.186.590.490.0
APT------93.089.190.590.8
+ +Table 1: The results on SuperGLUE development set and NER test set in full data setting. The metric of SuperGLUE is accuracy and other is micro-f1 score. Results for FT and PT-2 on BERT-large, RoBRETa-large and DeBERTa-large are token from (Liu et al., 2022). Results for FT on BERT-base are from (Liu et al., 2021). (FT: vanilla fine-tuning; PT-2: P-Tuning v2; APT: Adaptive Prefix Tuning; bold: the best score; underline: the second best) + +
SettingMethodBoolQCOPARTEWiCWSCAvg.
BERT-base (16-shot)FT47.27.554.06.549.42.750.32.346.26.849.4
PT-252.47.254.23.350.83.148.23.348.54.350.8
APT55.76.557.42.753.14.453.72.255.23.855.0
BERT-large (16-shot)FT57.39.752.02.449.52.750.00.038.72.249.5
PT-250.35.758.25.349.93.449.32.248.14.251.2
APT51.73.560.06.353.94.651.84.855.42.354.6
BERT-base (32-shot)FT48.19.452.26.449.52.749.40.960.43.851.9
PT-250.15.555.03.253.83.452.04.151.54.652.5
APT53.55.357.62.256.51.654.83.954.66.555.4
BERT-large (32-shot)FT47.611.945.03.648.42.250.00.047.313.247.6
PT-245.55.157.46.951.32.353.32.146.07.150.7
APT49.95.962.05.055.53.654.92.849.04.454.3
+ +Table 2: The mean ${}_{std}$ experimental results within 5 random seeds on SuperGLUE development set in 16-shot and 32-shot setting where all metrics are accuracy. bold: the best score. + +$\odot$ is the element-wise multiplication. Accordingly, the calculation of the self-attention function in APT is similar to Eq.(3) without further elaboration. + +# 4 Experiments + +# 4.1 Experimental Setup + +We conduct 5 NLU tasks on SuperGLUE (Wang et al., 2019) benchmark including BoolQ (Clark et al., 2019a), COPA (Roemmle et al., 2011), RTE (Wang et al., 2018), WiC (Pilehvar and Camacho-Collados, 2019) and WSC (Levesque et al., 2012) as well as 3 Named Entity Recognition (NER) tasks including CoNLL03 (Tjong Kim Sang and De Meulder, 2003), CoNLL04 (Carreras and Marquez, 2004), and OntoNotes 5.0 (Weischedel et al., 2013). With BERT-base / large (Devlin et al., 2019) and RoBERTa-large (Liu et al., 2019) instantiated by HuggingFace Transformers (Wolf et al., + +2020), we compare APT with vanilla fine-tuning and P-Tuning v2 (Liu et al., 2022) which is an implementation of the prefix tuning, configured with hyper-parameters public in the released code1. We also verify our method with DeBERTa-xlarge (He et al., 2020) on NER tasks following P-Tuning v2. + +# 4.2 Results + +We report the main results in Table 1. For BERT-base, we can observe that APT achieves $1.5\%$ and $0.7\%$ improvements over P-Tuning v2 on SuperGLUE and NER tasks, respectively. For BERT-large, APT outperforms P-Tuning v2 by $1.8\%$ on SuperGLUE tasks and $1.4\%$ on NER tasks. For RoBERTa-large, APT surpasses P-Tuning v2 by $1.5\%$ on SuperGLUE tasks and $0.2\%$ on NER tasks. On NER tasks with DeBERTa-xlarge, APT is supe + +
SettingSuperGLUENER
BoolQCOPARTEWiCWSCAvg.CoNLL03CoNLL04OntoNotesAvg.
APT72.670.072.771.266.970.789.784.187.287.0
w/o token-level α72.669.069.970.865.869.689.583.787.286.8
w/o layer-level λ72.167.471.369.665.469.189.082.686.986.2
w/o hidden states h72.068.868.770.264.668.989.183.687.186.6
+ +Table 3: Ablation study on BERT-base for two different level gate mechanisms and the hidden states from the previous layer. **bold:** the best score. + +
ModelSuperGLUENER
BoolQCOPARTEWiCWSCAvg.CoNLL03CoNLL04OntoNotesAvg.
PT-272.567.471.369.565.469.289.382.687.186.3
PT-2*72.668.871.970.065.869.889.383.087.286.5
PT-2+72.865.469.171.165.868.889.483.287.186.6
APT72.670.072.771.266.970.789.784.187.287.0
+ +Table 4: Comparison between PT-2 and PT-2*, PT-2+ and APT on BERT-base. (PT-2: P-Tuning v2; PT-2*: PT-2 with variable prefix; PT-2+: PT-2 with enlarged prefix) + +rior to P-Tuning v2 by an average of $0.8\%$ . Compared with vanilla fine-tuning, APT is comparable or even better on part of tasks. In addition, we explore the experimental performance under low resource settings on SuperGLUE benchmark. As shown in Table 2, APT is a better few-shot learner than P-Tuning v2, which exceeds $4.2\%$ , $3.4\%$ in 16-shot setting, and $2.9\%$ , $3.6\%$ in 32-shot setting for BERT-base and BERT-large, respectively. + +# 4.3 Ablation Study + +We conduct an ablation study in order to explore the separate effect of token-level gated weight $\alpha$ , layer-level scaled weight $\lambda$ and the hidden states $h$ from the previous layer which is used to calculate token-level gated weight $\alpha$ in Eq.(4). As shown in Table 3, it can be found that removing any strategy hurts the performance to varying degrees, demonstrating that they are all advantageous. Specifically, the beneficial effect of $\lambda$ for APT is slightly greater than $\alpha$ overall. Besides, it is effective and meaningful to introduce the context (i.e. the hidden states $h$ from the previous layer) when obtaining the gated weight, especially for SuperGLUE tasks. + +# 4.4 Discussion + +What is prefix weight distribution learned by APT? The gate mechanism for prefix serves as the key strategy of the proposed APT, where the learned prefix weight distribution turns out to be a critical point. Figure 2 illustrates the gate weights of the pseudo prefix token for COPA and CoNLL04, + +![](images/07649a75104e661b3dbf87701602fa9852bc7bcb688c3443c4cb5b878095dfe4.jpg) +(a) COPA + +![](images/0416bb02e236d8865949902aacebeab38efa4e6c4fd673a2716986be20f5be9f.jpg) +(b) CoNLL04 +Figure 2: Visualization of the learned weights of the prefix token for SuperGLUE task COPA on BERT-large and NER task CoNLL04 on BERT-base, with darker colors indicating higher weights. + +respectively. It can be found that CoNLL04 is concerned with bottom layers in the language model which are regarded as phrase-level features, while COPA pays more attention to the higher layers, indicating semantic information. The observation is consistent with the characteristics of corresponding tasks. NER is a token-level task while COPA is a causal reasoning task sensitive to the semantics of sentences, which reminds us that it is worth placing various prefix tokens on specific layers according to the task properties. + +Does variable prefix work better than fixed one? To verify the effectiveness of adaptive prefix under the proposed architecture, we wonder if the learned ratio at each layer can be directly transferred to P-Tuning v2. Taking the gate as a probing indicator, we reset the prefix length of P-Tuning v2 from fixed to variable in different layers based on the ob + +servation of the learned ratio (e.g. the distribution shown in Figure 2). From the comparison between PT-2 and $\mathrm{PT - }2^{*}$ in Table 4, we demonstrate that the variable prefix with less trainable parameters surprisingly outperforms the original implementation in fixed prefix. Nonetheless, it is also worth noting that there is still a gap between P-Tuning v2 with variable prefix and APT, where the latter continuously adjusts the weight of prefix during the training phase while the former only initializes with a one-time mask probing. + +Whether the adaptive structure benefits the fine-tuning? Compared to P-Tuning v2, APT learns extra gated and scaled weights. To figure it out whether the improvement of APT is brought from more trainable parameters or the adaptive model structure, we adjust the hyper-parameter, i.e., enlarge the prefix length of P-Tuning v2 by 1.5 times to align the number of parameters with our APT. As shown in the comparison between PT- $2^{+}$ and APT of Table 4, we observe that APT still outperforms enlarged P-Tuning v2 with $1.9\%$ , $0.4\%$ on average for SuperGLUE and NER tasks respectively, validating the superiority of the gate mechanism. + +# 5 Conclusion + +In this paper, we investigate prefix tuning and assume that adaptive prefix is probably more efficient and effective than fixed prefix. Firstly, we propose APT that leverages the token-level and the layer-level gate mechanism which achieves an improvement of performance over original prefix tuning. Then, we illustrate the weight distribution learned by APT and take it as a probe, which validates the variable prefix can work better than the fixed one. The above experiments and analysis demonstrate that the adaptive prefix can be served as a promising strategy for parameter-efficient fine-tuning. + +# Limitations + +The proposed approach in this paper also suffers from certain limitations, i.e. we adapt APT on the encoder model and lack design for the other architectures such as decoder-only and encoder-decoder. In addition, it is better to generalize the key idea to other parameter-efficient learning approaches. A unified solution for existing work may be worth exploring in the future. + +# References + +Elad Ben Zaken, Yoav Goldberg, and Shauli Ravfogel. 2022. BitFit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 1-9, Dublin, Ireland. Association for Computational Linguistics. +Xavier Carreras and Lluis Márquez. 2004. Introduction to the CoNLL-2004 shared task: Semantic role labeling. In Proceedings of the Eighth Conference on Computational Natural Language Learning (CoNLL-2004) at HLT-NAACL 2004, pages 89-97, Boston, Massachusetts, USA. Association for Computational Linguistics. +Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. 2019a. BoolQ: Exploring the surprising difficulty of natural yes/no questions. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2924–2936, Minneapolis, Minnesota. Association for Computational Linguistics. +Kevin Clark, Urvashi Khandelwal, Omer Levy, and Christopher D. Manning. 2019b. What does BERT look at? an analysis of BERT's attention. In Proceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 276-286, Florence, Italy. Association for Computational Linguistics. +Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics. +Angela Fan, Edouard Grave, and Armand Joulin. 2020. Reducing transformer depth on demand with structured dropout. In International Conference on Learning Representations. +Demi Guo, Alexander Rush, and Yoon Kim. 2021. Parameter-efficient transfer learning with diff pruning. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4884-4896, Online. Association for Computational Linguistics. +Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. 2022. Towards a unified view of parameter-efficient transfer learning. In International Conference on Learning Representations. + +Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. 2020. Deberta: Decoding-enhanced bert with disentangled attention. arXiv preprint arXiv:2006.03654. +Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. 2019. Parameter-efficient transfer learning for NLP. In Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pages 2790-2799. PMLR. +Ganesh Jawahar, Benoit Sagot, and Djamé Seddah. 2019. What does BERT learn about the structure of language? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3651-3657, Florence, Italy. Association for Computational Linguistics. +Brian Lester, Rami Al-Rfou, and Noah Constant. 2021. The power of scale for parameter-efficient prompt tuning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3045-3059, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics. +Hector Levesque, Ernest Davis, and Leora Morgenstern. 2012. The winograd schema challenge. In Thirteenth international conference on the principles of knowledge representation and reasoning. +Xiang Lisa Li and Percy Liang. 2021. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4582-4597, Online. Association for Computational Linguistics. +Xiao Liu, Kaixuan Ji, Yicheng Fu, Weng Tam, Zhengxiao Du, Zhilin Yang, and Jie Tang. 2022. P-tuning: Prompt tuning can be comparable to fine-tuning across scales and tasks. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 61-68, Dublin, Ireland. Association for Computational Linguistics. +Xiao Liu, Yanan Zheng, Zhengxiao Du, Ming Ding, Yujie Qian, Zhilin Yang, and Jie Tang. 2021. Gpt understands, too. arXiv:2103.10385. +Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692. +Nafise Sadat Moosavi, Quentin Delfosse, Kristian Kersting, and Iryna Gurevych. 2022. Adaptable Adapters. In Proceedings of the 2022 Annual Conference of + +the North American Chapter of the Association for Computational Linguistics, Seattle, WA, USA. Association for Computational Linguistics. +Mohammad Taher Pilehvar and Jose Camacho-Collados. 2019. WiC: the word-in-context dataset for evaluating context-sensitive meaning representations. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 1267-1273, Minneapolis, Minnesota. Association for Computational Linguistics. +Melissa Roemmele, Cosmin Adrian Bejan, and Andrew S Gordon. 2011. Choice of plausible alternatives: An evaluation of commonsense causal reasoning. In AAAI spring symposium: logical formalizations of commonsense reasoning, pages 90-95. +Andreas Rückle, Gregor Geigle, Max Glockner, Tilman Beck, Jonas Pfeiffer, Nils Reimers, and Iryna Gurevych. 2021. AdapterDrop: On the efficiency of adapters in transformers. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7930-7946, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics. +Ian Tenney, Dipanjan Das, and Ellie Pavlick. 2019. BERT rediscovers the classical NLP pipeline. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4593-4601, Florence, Italy. Association for Computational Linguistics. +Erik F. Tjong Kim Sang and Fien De Meulder. 2003. Introduction to the CoNLL-2003 shared task: Language-independent named entity recognition. In Proceedings of the Seventh Conference on Natural Language Learning at HLT-NAACL 2003, pages 142-147. +Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc. +Alex Wang, Yada Pruksachatkun, Nikita Nangia, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. 2019. SuperGLUE: A stickier benchmark for general-purpose language understanding systems. In Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc. +Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. 2018. GLUE: A multi-task benchmark and analysis platform for natural language understanding. In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 353-355, Brussels, Belgium. Association for Computational Linguistics. + +Ralph Weischedel, Martha Palmer, Mitchell Marcus, Eduard Hovy, Sameer Pradhan, Lance Ramshaw, Nianwen Xue, Ann Taylor, Jeff Kaufman, Michelle Franchini, Mohammed El-Bachouti, Robert Belvin, and Ann Houston. 2013. OntoNotes Release 5.0. Abacus Data Network. + +Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online. Association for Computational Linguistics. + +# A Experimental Details + +Datasets In the full data setting, all train-dev-test splits follow P-Tuning v2 (Liu et al., 2022). For low resources setting, to generate k-shot ( $k = 16, 32$ ) datasets on SuperGLUE, the fixed set of random seed [11,21,42,87,100] is utilized to sample instances in training and development set, while the entire development set is treated as test set, where the average performance is reported in Table 2. + +Experimental Setting We grid search the learning rate over [5e-3, 7e-3, 1e-2, 1e-4], training epoch over [20, 40, 60, 80, 100, 120], batch size over [8, 16, 32], and random seeds over [11, 21, 42, 87, 100]. For a fair comparison, the prefix length utilized by APT is consistent with P-Tuning v2. In low resources setting, the batch size we used is 2. In Eq.(4), we take the hidden states of the first input token as representation in previous layer. + +Experimental Computation We use the pretrained model BERT-base with 110M parameters, BERT-large with 335M parameters, RoBERTa-large with 355M parameters and DeBERTa-xlarge with 750M parameters. We conduct experiments on NVIDIA V100 or A100 GPUs for each task. + +# B Further Ablation Results + +We demonstrate further ablation results on BERT-large and RoBERTa-large as shown in Table 5. It can be found that the beneficial impact of the three strategies and the observation is consistent with BERT-base in Section 4.3 in general. + +![](images/e6bc50894731e36b1bee2fca83254350a9cafe9c3c5973e104559060a17484c6.jpg) +(a) COPA + +![](images/904200878bce5b4c5915b7b1a9a571384d8ddbb69cc1d9df19b652004f0281fb.jpg) +(b) WSC +Figure 3: The performance of APT and PT-2 on COPA and WSC in a range of prefix length on BERT-large. + +# C Prefix Length + +The prefix length is an important hyper-parameter for prefix tuning and APT. Figure 3 illustrates the performance of APT and P-Tuning v2 with different prefix lengths over a range. It can be observed that APT is superior to P-Tuning v2 in most prefix length settings, indicating that APT has a relatively wider range of prefix length to achieve better performance. + +# D Scientific Artifacts + +We use datasets involving SuperGLUE (Wang et al., 2019) benchmark including BoolQ (Clark et al., 2019a), COPA (Roemmle et al., 2011), RTE (Wang et al., 2018), WiC (Pilehvar and Camacho-Collados, 2019) and WSC (Levesque et al., 2012) as well as 3 Named Entity Recognition (NER) tasks including CoNLL03 (Tjong Kim Sang and De Meulder, 2003), CoNLL04 (Carreras and Marquez, 2004), and OntoNotes 5.0 (Weischedel et al., 2013). The pre-trained model we used are BERT-base / large (Devlin et al., 2019), RoBERTa-large (Liu et al., 2019) and DeBERTa-xlarge (He et al., 2020). We use HuggingFace Transformers (Wolf et al., 2020) and P-Tuning v2 (Liu et al., 2022) as the codebase implemented by PyTorch. They are all open-source and we only use for academic research which is consistent with their intended use. + +
ModelSettingSuperGLUENER
BoolQCOPARTEWiCWSCAvg.CoNLL03CoNLL04OntoNotesAvg.
BERT-largeAPT76.079.079.475.170.275.990.785.888.688.4
w/o token-level α75.877.077.374.868.374.691.184.488.588.0
w/o layer-level λ75.474.076.974.668.373.890.783.788.487.6
w/o hidden states h74.776.075.874.668.373.991.284.088.687.9
RoBERTa-largeAPT84.894.089.974.668.382.392.789.089.890.5
w/o token-level α84.388.088.173.065.479.892.288.789.590.1
w/o layer-level λ84.788.086.372.164.479.192.088.789.890.2
w/o hidden states h83.991.087.072.964.479.892.288.789.490.1
+ +Table 5: Ablation experiments on BERT-large and RoBERTa-large for two different level gate mechanisms and the hidden states from the previous layer. **bold:** the best score. + +A For every submission: + +A1. Did you describe the limitations of your work? section limitations +A2. Did you discuss any potential risks of your work? Not applicable. Left blank. +A3. Do the abstract and introduction summarize the paper's main claims? section abstract and section 1 introduction +A4. Have you used AI writing assistants when working on this paper? Left blank. + +B Did you use or create scientific artifacts? + +section 4.1 + +B1. Did you cite the creators of artifacts you used? section 4.1 +B2. Did you discuss the license or terms for use and / or distribution of any artifacts? section D Scientific Artifacts +B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? section D Scientific Artifacts +B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? We use open-source datasets and do not change datasets for a fair comparison. +B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Not applicable. Left blank. +B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. It can be found in the cited paper. + +C Did you run computational experiments? + +section 4 Experiments + +C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Table 1 and section appendix A Experimental Computation + +C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? section appendix A Experimental Details +C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Table 2 report the mean and std results. +C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? We follow the existing work and keep consistent with them. + +D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank. + +D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response. +D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response. +D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response. +D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response. +D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response. \ No newline at end of file diff --git a/2023/Towards Adaptive Prefix Tuning for Parameter-Efficient Language Model Fine-tuning/images.zip b/2023/Towards Adaptive Prefix Tuning for Parameter-Efficient Language Model Fine-tuning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f29df51577b98b133551bd872727432a5fb90771 --- /dev/null +++ b/2023/Towards Adaptive Prefix Tuning for Parameter-Efficient Language Model Fine-tuning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3a02a636cf0ab0757e6eae6e0fa60e1f9e5f1df0ec6c8667b898c8a64d3c7690 +size 397894 diff --git a/2023/Towards Adaptive Prefix Tuning for Parameter-Efficient Language Model Fine-tuning/layout.json b/2023/Towards Adaptive Prefix Tuning for Parameter-Efficient Language Model Fine-tuning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4f3c8b8cf5f9564ec1325563cacaac6519ccfcc6 --- /dev/null +++ b/2023/Towards Adaptive Prefix Tuning for Parameter-Efficient Language Model Fine-tuning/layout.json @@ -0,0 +1,6439 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 132, + 75, + 461, + 109 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 75, + 461, + 109 + ], + "spans": [ + { + "bbox": [ + 132, + 75, + 461, + 109 + ], + "type": "text", + "content": "Towards Adaptive Prefix Tuning for Parameter-Efficient Language Model Fine-tuning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 142, + 120, + 456, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 120, + 456, + 148 + ], + "spans": [ + { + "bbox": [ + 142, + 120, + 456, + 148 + ], + "type": "text", + "content": "Zhen-Ru Zhang, Chuanqi Tan, Haiyang Xu, Chengyu Wang, Jun Huang, Songfang Huang" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 261, + 149, + 335, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 149, + 335, + 163 + ], + "spans": [ + { + "bbox": [ + 261, + 149, + 335, + 163 + ], + "type": "text", + "content": "Alibaba Group" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 123, + 163, + 473, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 163, + 473, + 176 + ], + "spans": [ + { + "bbox": [ + 123, + 163, + 473, + 176 + ], + "type": "text", + "content": "{zhangzhenru.zzr, chuanqi.tcq, shuofeng.xhy}@alibaba-inc.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 135, + 177, + 461, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 177, + 461, + 190 + ], + "spans": [ + { + "bbox": [ + 135, + 177, + 461, + 190 + ], + "type": "text", + "content": "{chengyu.wcy,huangjun.hj,songfang.hsf}@alibaba-inc.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 84, + 236, + 274, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 236, + 274, + 511 + ], + "spans": [ + { + "bbox": [ + 84, + 236, + 274, + 511 + ], + "type": "text", + "content": "Fine-tuning large pre-trained language models on various downstream tasks with whole parameters is prohibitively expensive. Hence, Parameter-efficient fine-tuning has attracted attention that only optimizes a few task-specific parameters with the frozen pre-trained model. In this work, we focus on prefix tuning, which only optimizes continuous prefix vectors (i.e. pseudo tokens) inserted into Transformer layers. Based on the observation that the learned syntax and semantics representation varies a lot at different layers, we argue that the adaptive prefix will be further tailored to each layer than the fixed one, enabling the fine-tuning more effective and efficient. Thus, we propose Adaptive Prefix Tuning (APT) to adjust the prefix in terms of both fine-grained token level and coarse-grained layer level with a gate mechanism. Experiments on the SuperGLUE and NER datasets show the effectiveness of APT. In addition, taking the gate as a probing, we validate the efficiency and effectiveness of the variable prefix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 68, + 521, + 155, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 521, + 155, + 534 + ], + "spans": [ + { + "bbox": [ + 68, + 521, + 155, + 534 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 543, + 291, + 718 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 543, + 291, + 718 + ], + "spans": [ + { + "bbox": [ + 67, + 543, + 291, + 718 + ], + "type": "text", + "content": "Vanilla fine-tuning strategy usually adjusts all the parameters to adapt the pre-trained language model to downstream tasks. Parameter-efficient learning (He et al., 2022; Houlsby et al., 2019; Lester et al., 2021; Guo et al., 2021; Ben Zaken et al., 2022) is an emerging framework that freezes the pre-trained model and only tunes a few number of task-specific parameters for downstream tasks. For instance, Prefix tuning (Li and Liang, 2021; Liu et al., 2022) prepends length-equivalent pseudo prefix tokens, i.e. continuous task-specific vectors to each layer of the pre-trained model, achieving comparable even superior performance with only " + }, + { + "bbox": [ + 67, + 543, + 291, + 718 + ], + "type": "inline_equation", + "content": "0.1 - 3\\%" + }, + { + "bbox": [ + 67, + 543, + 291, + 718 + ], + "type": "text", + "content": " parameters." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 719, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 719, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 719, + 291, + 772 + ], + "type": "text", + "content": "In previous works, the length of prefix tokens (or the number of trainable parameters) is usually the same at each layer. However, a potential observation lies in that the structure information and" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 305, + 213, + 526, + 312 + ], + "blocks": [ + { + "bbox": [ + 305, + 213, + 526, + 312 + ], + "lines": [ + { + "bbox": [ + 305, + 213, + 526, + 312 + ], + "spans": [ + { + "bbox": [ + 305, + 213, + 526, + 312 + ], + "type": "image", + "image_path": "fe15aa65e38d6f4206afbd88889ca9617d6365cb5ba39e97b808c870e3a4cff1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 321, + 526, + 370 + ], + "lines": [ + { + "bbox": [ + 302, + 321, + 526, + 370 + ], + "spans": [ + { + "bbox": [ + 302, + 321, + 526, + 370 + ], + "type": "text", + "content": "Figure 1: An illustration of the proposed approach APT where the left is the internal structure of Transformer with inserted prefixes, and the right is the schematic of prefix gate mechanism." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 392, + 526, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 392, + 526, + 513 + ], + "spans": [ + { + "bbox": [ + 302, + 392, + 526, + 513 + ], + "type": "text", + "content": "representational capacity embedded in each layer are prone to be inconsistent (Jawahar et al., 2019). It is generally considered that the bottom layers of the language model tend to capture concrete and shallow phrase-level features, while the top layers concerns more with abstract semantic information (Tenney et al., 2019). Based on the perspective, we assume adaptive prefix can grab the emphasis more flexibly to adapt to various downstream tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 515, + 527, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 515, + 527, + 650 + ], + "spans": [ + { + "bbox": [ + 302, + 515, + 527, + 650 + ], + "type": "text", + "content": "In light of above motivation, we investigate the adaptive prefix in this work. We propose Adaptive Prefix Tuning (APT) with an adaptive gate mechanism at both fine-grained token level and coarse-grained layer level. Specifically, as shown in Figure 1, for fine granularity, APT scores each individual prefix token via gated weight assignment. Then, the scaled weight is utilized to balance the inserted task-specific prefix tokens and original input tokens for current layer at coarse-grained level." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 651, + 527, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 651, + 527, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 651, + 527, + 773 + ], + "type": "text", + "content": "Extensive experiments against prefix tuning on the sentence and token classification tasks in full data and low resources setting validate the effectiveness of APT. In addition, the gate learned from APT could be served as a probing for the number of necessary parameters in different layers, guiding us to directly apply variable prefix to the original prefix tuning. The probing experiment further demonstrates the effectiveness of adaptive prefix." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 780, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 310, + 791 + ], + "type": "text", + "content": "1239" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 135, + 795, + 459, + 805 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 795, + 459, + 805 + ], + "spans": [ + { + "bbox": [ + 135, + 795, + 459, + 805 + ], + "type": "text", + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 219, + 806, + 375, + 817 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 806, + 375, + 817 + ], + "spans": [ + { + "bbox": [ + 219, + 806, + 375, + 817 + ], + "type": "text", + "content": "Volume 2: Short Papers, pages 1239-1248" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "spans": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "text", + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 70, + 166, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 70, + 166, + 83 + ], + "spans": [ + { + "bbox": [ + 68, + 70, + 166, + 83 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 92, + 291, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 92, + 291, + 308 + ], + "spans": [ + { + "bbox": [ + 67, + 92, + 291, + 308 + ], + "type": "text", + "content": "Since fine-tuning the whole model is prohibitively expensive, parameter-efficient language model fine-tuning becomes a lightweight alternative that only optimizes a small number of parameters while keeping most pre-trained parameters frozen (He et al., 2022). Adapter tuning (Houlsby et al., 2019) inserts two tunable task-specific modules after multi-head attention and feed-forward network, achieving comparable performance with only " + }, + { + "bbox": [ + 67, + 92, + 291, + 308 + ], + "type": "inline_equation", + "content": "2 - 4\\%" + }, + { + "bbox": [ + 67, + 92, + 291, + 308 + ], + "type": "text", + "content": " of the parameters. Prompt tuning (Lester et al., 2021) and Prefix-Tuning (Li and Liang, 2021) only train soft prompts by adding prefix tokens to the input or hidden states. Recently, Liu et al. (2022) extend the prefix tuning to the natural language understanding tasks, which matches the performance of fine-tuning with only " + }, + { + "bbox": [ + 67, + 92, + 291, + 308 + ], + "type": "inline_equation", + "content": "0.1\\% -3\\%" + }, + { + "bbox": [ + 67, + 92, + 291, + 308 + ], + "type": "text", + "content": " tuned parameters." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 309, + 291, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 309, + 291, + 526 + ], + "spans": [ + { + "bbox": [ + 69, + 309, + 291, + 526 + ], + "type": "text", + "content": "Furthermore, with an overlap of our motivations that each layer of the pre-trained language model focuses on different aspects of feature for various tasks (Jawahar et al., 2019; Clark et al., 2019b) and extra parameters are probably not necessary for certain tasks (Houlsby et al., 2019; Fan et al., 2020; Rücklé et al., 2021), Adaptable Adapters (Moosavi et al., 2022) selects beneficial adapter layers and learns task-specific activation function for downstream tasks to make adaptor dynamic for each task and layer. In addition to different frameworks (adapter versa prefix tuning), our key difference from their work lies in that we aim to dynamically filter required information at each layer in a soft way, while they choose whether to add trainable modules at the layer level in a hard manner." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 535, + 157, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 535, + 157, + 550 + ], + "spans": [ + { + "bbox": [ + 67, + 535, + 157, + 550 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 557, + 159, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 557, + 159, + 571 + ], + "spans": [ + { + "bbox": [ + 67, + 557, + 159, + 571 + ], + "type": "text", + "content": "3.1 Prefix Tuning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 574, + 291, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 574, + 291, + 668 + ], + "spans": [ + { + "bbox": [ + 67, + 574, + 291, + 668 + ], + "type": "text", + "content": "As prefix tuning is an extension on Transformer (Vaswani et al., 2017), we first recap the structure of Transformer. Transformer is the block consisting of multi-head attention concatenated by multiple single self-attention functions and a fully connected feed-forward network. Formally speaking, the Transformer block is calculated as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 88, + 676, + 290, + 706 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 676, + 290, + 706 + ], + "spans": [ + { + "bbox": [ + 88, + 676, + 290, + 706 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t n} (\\boldsymbol {Q}, \\boldsymbol {K}, \\boldsymbol {V}) = \\operatorname {s o f t m a x} \\left(\\frac {\\boldsymbol {Q} \\boldsymbol {K} ^ {T}}{\\sqrt {d}} \\boldsymbol {V}\\right) \\quad (1)", + "image_path": "fb1377f436180bde3f97d638a96c8b834273575235231672cbde80f98a910588.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 83, + 708, + 290, + 722 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 708, + 290, + 722 + ], + "spans": [ + { + "bbox": [ + 83, + 708, + 290, + 722 + ], + "type": "interline_equation", + "content": "\\operatorname {F F N} (\\boldsymbol {x}) = \\operatorname {R e L U} (\\boldsymbol {x} \\boldsymbol {W} _ {1} + \\boldsymbol {b} _ {1}) \\boldsymbol {W} _ {2} + \\boldsymbol {b} _ {2} \\tag {2}", + "image_path": "d38f2aaf58717958b086c0c699003c89fc83915bb4e3ef991e94f3684bbb26b0.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 733, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 733, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 733, + 291, + 773 + ], + "type": "text", + "content": "Prefix tuning prepends pseudo prefix tokens of length " + }, + { + "bbox": [ + 67, + 733, + 291, + 773 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 67, + 733, + 291, + 773 + ], + "type": "text", + "content": " to each layer of the language model, which is implemented by concatenating inserted" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 71, + 526, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 152 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 152 + ], + "type": "text", + "content": "keys and values matrix with original corresponding items in each multi-head attention. Specifically, let " + }, + { + "bbox": [ + 302, + 71, + 526, + 152 + ], + "type": "inline_equation", + "content": "P_{k}, P_{v} \\in \\mathbb{R}^{l \\times d}" + }, + { + "bbox": [ + 302, + 71, + 526, + 152 + ], + "type": "text", + "content": " be the keys and values of the engaged prefix separately, where " + }, + { + "bbox": [ + 302, + 71, + 526, + 152 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 302, + 71, + 526, + 152 + ], + "type": "text", + "content": " denotes the length of prefix and " + }, + { + "bbox": [ + 302, + 71, + 526, + 152 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 302, + 71, + 526, + 152 + ], + "type": "text", + "content": " corresponds to the dimension, thus self-attention function can be reformatted as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 160, + 525, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 160, + 525, + 190 + ], + "spans": [ + { + "bbox": [ + 306, + 160, + 525, + 190 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t n} \\left(\\boldsymbol {Q}, \\boldsymbol {K} ^ {\\prime}, \\boldsymbol {V} ^ {\\prime}\\right) = \\operatorname {s o f t m a x} \\left(\\frac {\\boldsymbol {Q} \\left(\\boldsymbol {K} ^ {\\prime}\\right) ^ {T}}{\\sqrt {d}} \\boldsymbol {V} ^ {\\prime}\\right) \\tag {3}", + "image_path": "96ec2576da6b265e0f96ac0d3c541c2d1ca1d8a57c8accffb94f93fda8064dd2.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 325, + 191, + 490, + 206 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 191, + 490, + 206 + ], + "spans": [ + { + "bbox": [ + 325, + 191, + 490, + 206 + ], + "type": "interline_equation", + "content": "\\text {w h e r e} \\boldsymbol {K} ^ {\\prime} = \\left[ \\boldsymbol {P} _ {k}; \\boldsymbol {K} \\right], \\boldsymbol {V} ^ {\\prime} = \\left[ \\boldsymbol {P} _ {v}; \\boldsymbol {V} \\right]", + "image_path": "382913230644488eae6b51782429babd93dacb30009a8a6ebce4ba2ab567d4ea.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 218, + 485, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 218, + 485, + 231 + ], + "spans": [ + { + "bbox": [ + 302, + 218, + 485, + 231 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 302, + 218, + 485, + 231 + ], + "type": "inline_equation", + "content": "[;]" + }, + { + "bbox": [ + 302, + 218, + 485, + 231 + ], + "type": "text", + "content": " donates concatenation function." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 243, + 441, + 256 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 243, + 441, + 256 + ], + "spans": [ + { + "bbox": [ + 302, + 243, + 441, + 256 + ], + "type": "text", + "content": "3.2 Adaptive Prefix Tuning" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 301, + 260, + 526, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 260, + 526, + 396 + ], + "spans": [ + { + "bbox": [ + 301, + 260, + 526, + 396 + ], + "type": "text", + "content": "The length of prefix is usually a manually set hyperparameter for each task and fixed in distinct layers of the model. However, existing work demonstrates each layer of the language model pays attention to different aspects of the input feature. We assume the prefix in fixed length is insufficient to tailor different layers and tasks. To dynamically customize the prefix at each layer, APT performs a gate mechanism via fine-grained gated weight assignment and coarse-grained scaled weight specification." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 397, + 527, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 397, + 527, + 547 + ], + "spans": [ + { + "bbox": [ + 302, + 397, + 527, + 547 + ], + "type": "text", + "content": "Specifically, to capture the diversity of information utilization at different layers, we go deep into the token level at the fine-grained granularity. The token-level gate can inspire us on how many trainable parameters (i.e. pseudo tokens in prefix tuning) are required for this layer, which will be discussed in Section 4.4. Thus, APT yields the gated weights of " + }, + { + "bbox": [ + 302, + 397, + 527, + 547 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 302, + 397, + 527, + 547 + ], + "type": "text", + "content": " pseudo tokens at each layer. We use the hidden states to represent the information encoded in the layer and calculate the gated weights " + }, + { + "bbox": [ + 302, + 397, + 527, + 547 + ], + "type": "inline_equation", + "content": "\\alpha_{i} = [\\alpha_{i1},\\alpha_{i2},\\dots,\\alpha_{il}]" + }, + { + "bbox": [ + 302, + 397, + 527, + 547 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 302, + 397, + 527, + 547 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 302, + 397, + 527, + 547 + ], + "type": "text", + "content": "-th layer as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 348, + 558, + 525, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 558, + 525, + 571 + ], + "spans": [ + { + "bbox": [ + 348, + 558, + 525, + 571 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\alpha} _ {i} = \\operatorname {s i g m o i d} \\left(\\boldsymbol {h} _ {i - 1} \\boldsymbol {W} _ {i}\\right) \\tag {4}", + "image_path": "69d177e7de911a94630f02a17c718bbf8e96b59974dafdc83162501c372a704e.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 302, + 584, + 525, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 584, + 525, + 624 + ], + "spans": [ + { + "bbox": [ + 302, + 584, + 525, + 624 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 302, + 584, + 525, + 624 + ], + "type": "inline_equation", + "content": "\\pmb{h}_{i-1}" + }, + { + "bbox": [ + 302, + 584, + 525, + 624 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 302, + 584, + 525, + 624 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 302, + 584, + 525, + 624 + ], + "type": "text", + "content": "-dimensional hidden states from the previous layer, and " + }, + { + "bbox": [ + 302, + 584, + 525, + 624 + ], + "type": "inline_equation", + "content": "\\pmb{W}_i \\in \\mathbb{R}^{d \\times l}" + }, + { + "bbox": [ + 302, + 584, + 525, + 624 + ], + "type": "text", + "content": " corresponds to the parameters to be learned." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 302, + 625, + 524, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 625, + 524, + 706 + ], + "spans": [ + { + "bbox": [ + 302, + 625, + 524, + 706 + ], + "type": "text", + "content": "Besides, we also design a coarse-level gate to balance the information brought from task-specific prefix tokens and original input tokens by learning a layer-level weight. A learnable scaled weight " + }, + { + "bbox": [ + 302, + 625, + 524, + 706 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 302, + 625, + 524, + 706 + ], + "type": "text", + "content": " is added to the representation of pseudo prefix tokens at the " + }, + { + "bbox": [ + 302, + 625, + 524, + 706 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 302, + 625, + 524, + 706 + ], + "type": "text", + "content": "-th layer." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 302, + 707, + 524, + 747 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 707, + 524, + 747 + ], + "spans": [ + { + "bbox": [ + 302, + 707, + 524, + 747 + ], + "type": "text", + "content": "With the above strategy, the keys-values pair " + }, + { + "bbox": [ + 302, + 707, + 524, + 747 + ], + "type": "inline_equation", + "content": "P_{i} = [P_{ik}, P_{iv}]" + }, + { + "bbox": [ + 302, + 707, + 524, + 747 + ], + "type": "text", + "content": " derived from pseudo prefix tokens in " + }, + { + "bbox": [ + 302, + 707, + 524, + 747 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 302, + 707, + 524, + 747 + ], + "type": "text", + "content": "-th layer is updated to " + }, + { + "bbox": [ + 302, + 707, + 524, + 747 + ], + "type": "inline_equation", + "content": "\\hat{P}_{i}" + }, + { + "bbox": [ + 302, + 707, + 524, + 747 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 351, + 758, + 525, + 774 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 758, + 525, + 774 + ], + "spans": [ + { + "bbox": [ + 351, + 758, + 525, + 774 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {P}} _ {i} = \\lambda_ {i} \\boldsymbol {\\alpha} _ {i} \\odot [ \\boldsymbol {P} _ {i k}, \\boldsymbol {P} _ {i v} ] \\tag {5}", + "image_path": "8b15d77f72ef823d40c3b8a921c0dd772b64ff31dfc89613be408a69a3da09d7.jpg" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 780, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 310, + 791 + ], + "type": "text", + "content": "1240" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 68, + 523, + 241 + ], + "blocks": [ + { + "bbox": [ + 70, + 68, + 523, + 241 + ], + "lines": [ + { + "bbox": [ + 70, + 68, + 523, + 241 + ], + "spans": [ + { + "bbox": [ + 70, + 68, + 523, + 241 + ], + "type": "table", + "html": "
ModelSuperGLUENER
BoolQCOPARTEWiCWSCAvg.CoNLL03CoNLL04OntoNotesAvg.
BERT-base (110M)FT72.967.068.471.163.568.6----
PT-272.567.471.369.565.469.289.382.687.186.3
APT72.670.072.771.266.970.789.784.187.287.0
BERT-large (335M)FT77.769.070.474.968.372.192.885.689.289.2
PT-275.873.078.375.168.374.190.284.586.487.0
APT76.079.079.475.170.275.990.785.888.688.4
RoBRETa-large (355M)FT86.994.086.675.663.581.392.688.889.890.4
PT-284.893.089.573.463.580.892.888.489.890.3
APT84.894.089.974.668.382.392.789.089.890.5
DeBERTa-xlarge (750M)FT------93.189.190.490.9
PT-2------93.186.590.490.0
APT------93.089.190.590.8
", + "image_path": "b1ad81093b446b5a1b22ef26c897bc77c34543cce612f60a3c3492bffd22c670.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 135, + 308, + 459, + 470 + ], + "blocks": [ + { + "bbox": [ + 67, + 248, + 526, + 298 + ], + "lines": [ + { + "bbox": [ + 67, + 248, + 526, + 298 + ], + "spans": [ + { + "bbox": [ + 67, + 248, + 526, + 298 + ], + "type": "text", + "content": "Table 1: The results on SuperGLUE development set and NER test set in full data setting. The metric of SuperGLUE is accuracy and other is micro-f1 score. Results for FT and PT-2 on BERT-large, RoBRETa-large and DeBERTa-large are token from (Liu et al., 2022). Results for FT on BERT-base are from (Liu et al., 2021). (FT: vanilla fine-tuning; PT-2: P-Tuning v2; APT: Adaptive Prefix Tuning; bold: the best score; underline: the second best)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 308, + 459, + 470 + ], + "lines": [ + { + "bbox": [ + 135, + 308, + 459, + 470 + ], + "spans": [ + { + "bbox": [ + 135, + 308, + 459, + 470 + ], + "type": "table", + "html": "
SettingMethodBoolQCOPARTEWiCWSCAvg.
BERT-base (16-shot)FT47.27.554.06.549.42.750.32.346.26.849.4
PT-252.47.254.23.350.83.148.23.348.54.350.8
APT55.76.557.42.753.14.453.72.255.23.855.0
BERT-large (16-shot)FT57.39.752.02.449.52.750.00.038.72.249.5
PT-250.35.758.25.349.93.449.32.248.14.251.2
APT51.73.560.06.353.94.651.84.855.42.354.6
BERT-base (32-shot)FT48.19.452.26.449.52.749.40.960.43.851.9
PT-250.15.555.03.253.83.452.04.151.54.652.5
APT53.55.357.62.256.51.654.83.954.66.555.4
BERT-large (32-shot)FT47.611.945.03.648.42.250.00.047.313.247.6
PT-245.55.157.46.951.32.353.32.146.07.150.7
APT49.95.962.05.055.53.654.92.849.04.454.3
", + "image_path": "eb2e5d9db59f3eeb55d734dc7c9f437ea7e56af43ffdbed2ddf9d9d8f4e5a108.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 477, + 525, + 502 + ], + "lines": [ + { + "bbox": [ + 67, + 477, + 525, + 502 + ], + "spans": [ + { + "bbox": [ + 67, + 477, + 525, + 502 + ], + "type": "text", + "content": "Table 2: The mean " + }, + { + "bbox": [ + 67, + 477, + 525, + 502 + ], + "type": "inline_equation", + "content": "{}_{std}" + }, + { + "bbox": [ + 67, + 477, + 525, + 502 + ], + "type": "text", + "content": " experimental results within 5 random seeds on SuperGLUE development set in 16-shot and 32-shot setting where all metrics are accuracy. bold: the best score." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 523, + 290, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 523, + 290, + 565 + ], + "spans": [ + { + "bbox": [ + 67, + 523, + 290, + 565 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 67, + 523, + 290, + 565 + ], + "type": "text", + "content": " is the element-wise multiplication. Accordingly, the calculation of the self-attention function in APT is similar to Eq.(3) without further elaboration." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 573, + 155, + 587 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 573, + 155, + 587 + ], + "spans": [ + { + "bbox": [ + 67, + 573, + 155, + 587 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 593, + 189, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 593, + 189, + 607 + ], + "spans": [ + { + "bbox": [ + 67, + 593, + 189, + 607 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 611, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 611, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 611, + 291, + 772 + ], + "type": "text", + "content": "We conduct 5 NLU tasks on SuperGLUE (Wang et al., 2019) benchmark including BoolQ (Clark et al., 2019a), COPA (Roemmle et al., 2011), RTE (Wang et al., 2018), WiC (Pilehvar and Camacho-Collados, 2019) and WSC (Levesque et al., 2012) as well as 3 Named Entity Recognition (NER) tasks including CoNLL03 (Tjong Kim Sang and De Meulder, 2003), CoNLL04 (Carreras and Marquez, 2004), and OntoNotes 5.0 (Weischedel et al., 2013). With BERT-base / large (Devlin et al., 2019) and RoBERTa-large (Liu et al., 2019) instantiated by HuggingFace Transformers (Wolf et al.," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 523, + 526, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 523, + 526, + 606 + ], + "spans": [ + { + "bbox": [ + 302, + 523, + 526, + 606 + ], + "type": "text", + "content": "2020), we compare APT with vanilla fine-tuning and P-Tuning v2 (Liu et al., 2022) which is an implementation of the prefix tuning, configured with hyper-parameters public in the released code1. We also verify our method with DeBERTa-xlarge (He et al., 2020) on NER tasks following P-Tuning v2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 614, + 365, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 614, + 365, + 625 + ], + "spans": [ + { + "bbox": [ + 302, + 614, + 365, + 625 + ], + "type": "text", + "content": "4.2 Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 631, + 526, + 754 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 631, + 526, + 754 + ], + "spans": [ + { + "bbox": [ + 302, + 631, + 526, + 754 + ], + "type": "text", + "content": "We report the main results in Table 1. For BERT-base, we can observe that APT achieves " + }, + { + "bbox": [ + 302, + 631, + 526, + 754 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 302, + 631, + 526, + 754 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 302, + 631, + 526, + 754 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 302, + 631, + 526, + 754 + ], + "type": "text", + "content": " improvements over P-Tuning v2 on SuperGLUE and NER tasks, respectively. For BERT-large, APT outperforms P-Tuning v2 by " + }, + { + "bbox": [ + 302, + 631, + 526, + 754 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 302, + 631, + 526, + 754 + ], + "type": "text", + "content": " on SuperGLUE tasks and " + }, + { + "bbox": [ + 302, + 631, + 526, + 754 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 302, + 631, + 526, + 754 + ], + "type": "text", + "content": " on NER tasks. For RoBERTa-large, APT surpasses P-Tuning v2 by " + }, + { + "bbox": [ + 302, + 631, + 526, + 754 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 302, + 631, + 526, + 754 + ], + "type": "text", + "content": " on SuperGLUE tasks and " + }, + { + "bbox": [ + 302, + 631, + 526, + 754 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 302, + 631, + 526, + 754 + ], + "type": "text", + "content": " on NER tasks. On NER tasks with DeBERTa-xlarge, APT is supe" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 761, + 485, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 761, + 485, + 772 + ], + "spans": [ + { + "bbox": [ + 315, + 761, + 485, + 772 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 287, + 780, + 308, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 780, + 308, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 780, + 308, + 791 + ], + "type": "text", + "content": "1241" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 76, + 68, + 517, + 154 + ], + "blocks": [ + { + "bbox": [ + 76, + 68, + 517, + 154 + ], + "lines": [ + { + "bbox": [ + 76, + 68, + 517, + 154 + ], + "spans": [ + { + "bbox": [ + 76, + 68, + 517, + 154 + ], + "type": "table", + "html": "
SettingSuperGLUENER
BoolQCOPARTEWiCWSCAvg.CoNLL03CoNLL04OntoNotesAvg.
APT72.670.072.771.266.970.789.784.187.287.0
w/o token-level α72.669.069.970.865.869.689.583.787.286.8
w/o layer-level λ72.167.471.369.665.469.189.082.686.986.2
w/o hidden states h72.068.868.770.264.668.989.183.687.186.6
", + "image_path": "bbe35785b8e676de1bdba268ae802c861f00eaab2011d16b85d7b2665c040990.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 99, + 197, + 493, + 284 + ], + "blocks": [ + { + "bbox": [ + 67, + 163, + 525, + 187 + ], + "lines": [ + { + "bbox": [ + 67, + 163, + 525, + 187 + ], + "spans": [ + { + "bbox": [ + 67, + 163, + 525, + 187 + ], + "type": "text", + "content": "Table 3: Ablation study on BERT-base for two different level gate mechanisms and the hidden states from the previous layer. **bold:** the best score." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 99, + 197, + 493, + 284 + ], + "lines": [ + { + "bbox": [ + 99, + 197, + 493, + 284 + ], + "spans": [ + { + "bbox": [ + 99, + 197, + 493, + 284 + ], + "type": "table", + "html": "
ModelSuperGLUENER
BoolQCOPARTEWiCWSCAvg.CoNLL03CoNLL04OntoNotesAvg.
PT-272.567.471.369.565.469.289.382.687.186.3
PT-2*72.668.871.970.065.869.889.383.087.286.5
PT-2+72.865.469.171.165.868.889.483.287.186.6
APT72.670.072.771.266.970.789.784.187.287.0
", + "image_path": "abf16e47a310187c3b9123d75cc78ffa492095ff72242b6f5790af9177b45f22.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 291, + 525, + 317 + ], + "lines": [ + { + "bbox": [ + 67, + 291, + 525, + 317 + ], + "spans": [ + { + "bbox": [ + 67, + 291, + 525, + 317 + ], + "type": "text", + "content": "Table 4: Comparison between PT-2 and PT-2*, PT-2+ and APT on BERT-base. (PT-2: P-Tuning v2; PT-2*: PT-2 with variable prefix; PT-2+: PT-2 with enlarged prefix)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 338, + 290, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 338, + 290, + 460 + ], + "spans": [ + { + "bbox": [ + 67, + 338, + 290, + 460 + ], + "type": "text", + "content": "rior to P-Tuning v2 by an average of " + }, + { + "bbox": [ + 67, + 338, + 290, + 460 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 67, + 338, + 290, + 460 + ], + "type": "text", + "content": ". Compared with vanilla fine-tuning, APT is comparable or even better on part of tasks. In addition, we explore the experimental performance under low resource settings on SuperGLUE benchmark. As shown in Table 2, APT is a better few-shot learner than P-Tuning v2, which exceeds " + }, + { + "bbox": [ + 67, + 338, + 290, + 460 + ], + "type": "inline_equation", + "content": "4.2\\%" + }, + { + "bbox": [ + 67, + 338, + 290, + 460 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 67, + 338, + 290, + 460 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 67, + 338, + 290, + 460 + ], + "type": "text", + "content": " in 16-shot setting, and " + }, + { + "bbox": [ + 67, + 338, + 290, + 460 + ], + "type": "inline_equation", + "content": "2.9\\%" + }, + { + "bbox": [ + 67, + 338, + 290, + 460 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 67, + 338, + 290, + 460 + ], + "type": "inline_equation", + "content": "3.6\\%" + }, + { + "bbox": [ + 67, + 338, + 290, + 460 + ], + "type": "text", + "content": " in 32-shot setting for BERT-base and BERT-large, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 470, + 167, + 482 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 470, + 167, + 482 + ], + "spans": [ + { + "bbox": [ + 67, + 470, + 167, + 482 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 488, + 291, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 488, + 291, + 664 + ], + "spans": [ + { + "bbox": [ + 67, + 488, + 291, + 664 + ], + "type": "text", + "content": "We conduct an ablation study in order to explore the separate effect of token-level gated weight " + }, + { + "bbox": [ + 67, + 488, + 291, + 664 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 67, + 488, + 291, + 664 + ], + "type": "text", + "content": ", layer-level scaled weight " + }, + { + "bbox": [ + 67, + 488, + 291, + 664 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 67, + 488, + 291, + 664 + ], + "type": "text", + "content": " and the hidden states " + }, + { + "bbox": [ + 67, + 488, + 291, + 664 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 67, + 488, + 291, + 664 + ], + "type": "text", + "content": " from the previous layer which is used to calculate token-level gated weight " + }, + { + "bbox": [ + 67, + 488, + 291, + 664 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 67, + 488, + 291, + 664 + ], + "type": "text", + "content": " in Eq.(4). As shown in Table 3, it can be found that removing any strategy hurts the performance to varying degrees, demonstrating that they are all advantageous. Specifically, the beneficial effect of " + }, + { + "bbox": [ + 67, + 488, + 291, + 664 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 67, + 488, + 291, + 664 + ], + "type": "text", + "content": " for APT is slightly greater than " + }, + { + "bbox": [ + 67, + 488, + 291, + 664 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 67, + 488, + 291, + 664 + ], + "type": "text", + "content": " overall. Besides, it is effective and meaningful to introduce the context (i.e. the hidden states " + }, + { + "bbox": [ + 67, + 488, + 291, + 664 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 67, + 488, + 291, + 664 + ], + "type": "text", + "content": " from the previous layer) when obtaining the gated weight, especially for SuperGLUE tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 674, + 145, + 686 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 674, + 145, + 686 + ], + "spans": [ + { + "bbox": [ + 67, + 674, + 145, + 686 + ], + "type": "text", + "content": "4.4 Discussion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 692, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 692, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 692, + 291, + 772 + ], + "type": "text", + "content": "What is prefix weight distribution learned by APT? The gate mechanism for prefix serves as the key strategy of the proposed APT, where the learned prefix weight distribution turns out to be a critical point. Figure 2 illustrates the gate weights of the pseudo prefix token for COPA and CoNLL04," + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 308, + 339, + 413, + 424 + ], + "blocks": [ + { + "bbox": [ + 308, + 339, + 413, + 424 + ], + "lines": [ + { + "bbox": [ + 308, + 339, + 413, + 424 + ], + "spans": [ + { + "bbox": [ + 308, + 339, + 413, + 424 + ], + "type": "image", + "image_path": "07649a75104e661b3dbf87701602fa9852bc7bcb688c3443c4cb5b878095dfe4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 341, + 428, + 380, + 439 + ], + "lines": [ + { + "bbox": [ + 341, + 428, + 380, + 439 + ], + "spans": [ + { + "bbox": [ + 341, + 428, + 380, + 439 + ], + "type": "text", + "content": "(a) COPA" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 416, + 339, + 521, + 424 + ], + "blocks": [ + { + "bbox": [ + 416, + 339, + 521, + 424 + ], + "lines": [ + { + "bbox": [ + 416, + 339, + 521, + 424 + ], + "spans": [ + { + "bbox": [ + 416, + 339, + 521, + 424 + ], + "type": "image", + "image_path": "0416bb02e236d8865949902aacebeab38efa4e6c4fd673a2716986be20f5be9f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 442, + 428, + 495, + 439 + ], + "lines": [ + { + "bbox": [ + 442, + 428, + 495, + 439 + ], + "spans": [ + { + "bbox": [ + 442, + 428, + 495, + 439 + ], + "type": "text", + "content": "(b) CoNLL04" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 302, + 452, + 525, + 501 + ], + "lines": [ + { + "bbox": [ + 302, + 452, + 525, + 501 + ], + "spans": [ + { + "bbox": [ + 302, + 452, + 525, + 501 + ], + "type": "text", + "content": "Figure 2: Visualization of the learned weights of the prefix token for SuperGLUE task COPA on BERT-large and NER task CoNLL04 on BERT-base, with darker colors indicating higher weights." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 301, + 522, + 526, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 522, + 526, + 671 + ], + "spans": [ + { + "bbox": [ + 301, + 522, + 526, + 671 + ], + "type": "text", + "content": "respectively. It can be found that CoNLL04 is concerned with bottom layers in the language model which are regarded as phrase-level features, while COPA pays more attention to the higher layers, indicating semantic information. The observation is consistent with the characteristics of corresponding tasks. NER is a token-level task while COPA is a causal reasoning task sensitive to the semantics of sentences, which reminds us that it is worth placing various prefix tokens on specific layers according to the task properties." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 678, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 678, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 678, + 525, + 772 + ], + "type": "text", + "content": "Does variable prefix work better than fixed one? To verify the effectiveness of adaptive prefix under the proposed architecture, we wonder if the learned ratio at each layer can be directly transferred to P-Tuning v2. Taking the gate as a probing indicator, we reset the prefix length of P-Tuning v2 from fixed to variable in different layers based on the ob" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 780, + 309, + 791 + ], + "type": "text", + "content": "1242" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 291, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 291, + 220 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 291, + 220 + ], + "type": "text", + "content": "servation of the learned ratio (e.g. the distribution shown in Figure 2). From the comparison between PT-2 and " + }, + { + "bbox": [ + 67, + 71, + 291, + 220 + ], + "type": "inline_equation", + "content": "\\mathrm{PT - }2^{*}" + }, + { + "bbox": [ + 67, + 71, + 291, + 220 + ], + "type": "text", + "content": " in Table 4, we demonstrate that the variable prefix with less trainable parameters surprisingly outperforms the original implementation in fixed prefix. Nonetheless, it is also worth noting that there is still a gap between P-Tuning v2 with variable prefix and APT, where the latter continuously adjusts the weight of prefix during the training phase while the former only initializes with a one-time mask probing." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 232, + 292, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 232, + 292, + 409 + ], + "spans": [ + { + "bbox": [ + 67, + 232, + 292, + 409 + ], + "type": "text", + "content": "Whether the adaptive structure benefits the fine-tuning? Compared to P-Tuning v2, APT learns extra gated and scaled weights. To figure it out whether the improvement of APT is brought from more trainable parameters or the adaptive model structure, we adjust the hyper-parameter, i.e., enlarge the prefix length of P-Tuning v2 by 1.5 times to align the number of parameters with our APT. As shown in the comparison between PT- " + }, + { + "bbox": [ + 67, + 232, + 292, + 409 + ], + "type": "inline_equation", + "content": "2^{+}" + }, + { + "bbox": [ + 67, + 232, + 292, + 409 + ], + "type": "text", + "content": " and APT of Table 4, we observe that APT still outperforms enlarged P-Tuning v2 with " + }, + { + "bbox": [ + 67, + 232, + 292, + 409 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 67, + 232, + 292, + 409 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 67, + 232, + 292, + 409 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 67, + 232, + 292, + 409 + ], + "type": "text", + "content": " on average for SuperGLUE and NER tasks respectively, validating the superiority of the gate mechanism." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 423, + 147, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 423, + 147, + 436 + ], + "spans": [ + { + "bbox": [ + 67, + 423, + 147, + 436 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 449, + 291, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 449, + 291, + 612 + ], + "spans": [ + { + "bbox": [ + 67, + 449, + 291, + 612 + ], + "type": "text", + "content": "In this paper, we investigate prefix tuning and assume that adaptive prefix is probably more efficient and effective than fixed prefix. Firstly, we propose APT that leverages the token-level and the layer-level gate mechanism which achieves an improvement of performance over original prefix tuning. Then, we illustrate the weight distribution learned by APT and take it as a probe, which validates the variable prefix can work better than the fixed one. The above experiments and analysis demonstrate that the adaptive prefix can be served as a promising strategy for parameter-efficient fine-tuning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 626, + 131, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 626, + 131, + 639 + ], + "spans": [ + { + "bbox": [ + 67, + 626, + 131, + 639 + ], + "type": "text", + "content": "Limitations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 651, + 291, + 760 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 651, + 291, + 760 + ], + "spans": [ + { + "bbox": [ + 67, + 651, + 291, + 760 + ], + "type": "text", + "content": "The proposed approach in this paper also suffers from certain limitations, i.e. we adapt APT on the encoder model and lack design for the other architectures such as decoder-only and encoder-decoder. In addition, it is better to generalize the key idea to other parameter-efficient learning approaches. A unified solution for existing work may be worth exploring in the future." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 70, + 362, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 70, + 362, + 83 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 362, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 90, + 526, + 772 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 304, + 90, + 526, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 526, + 169 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 526, + 169 + ], + "type": "text", + "content": "Elad Ben Zaken, Yoav Goldberg, and Shauli Ravfogel. 2022. BitFit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 1-9, Dublin, Ireland. Association for Computational Linguistics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 176, + 526, + 254 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 176, + 526, + 254 + ], + "spans": [ + { + "bbox": [ + 304, + 176, + 526, + 254 + ], + "type": "text", + "content": "Xavier Carreras and Lluis Márquez. 2004. Introduction to the CoNLL-2004 shared task: Semantic role labeling. In Proceedings of the Eighth Conference on Computational Natural Language Learning (CoNLL-2004) at HLT-NAACL 2004, pages 89-97, Boston, Massachusetts, USA. Association for Computational Linguistics." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 263, + 526, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 263, + 526, + 363 + ], + "spans": [ + { + "bbox": [ + 304, + 263, + 526, + 363 + ], + "type": "text", + "content": "Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. 2019a. BoolQ: Exploring the surprising difficulty of natural yes/no questions. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2924–2936, Minneapolis, Minnesota. Association for Computational Linguistics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 371, + 526, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 371, + 526, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 371, + 526, + 449 + ], + "type": "text", + "content": "Kevin Clark, Urvashi Khandelwal, Omer Levy, and Christopher D. Manning. 2019b. What does BERT look at? an analysis of BERT's attention. In Proceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 276-286, Florence, Italy. Association for Computational Linguistics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 458, + 526, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 458, + 526, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 526, + 557 + ], + "type": "text", + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 565, + 526, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 565, + 526, + 611 + ], + "spans": [ + { + "bbox": [ + 304, + 565, + 526, + 611 + ], + "type": "text", + "content": "Angela Fan, Edouard Grave, and Armand Joulin. 2020. Reducing transformer depth on demand with structured dropout. In International Conference on Learning Representations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 619, + 526, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 619, + 526, + 708 + ], + "spans": [ + { + "bbox": [ + 304, + 619, + 526, + 708 + ], + "type": "text", + "content": "Demi Guo, Alexander Rush, and Yoon Kim. 2021. Parameter-efficient transfer learning with diff pruning. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4884-4896, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 717, + 526, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 717, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 717, + 526, + 772 + ], + "type": "text", + "content": "Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. 2022. Towards a unified view of parameter-efficient transfer learning. In International Conference on Learning Representations." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1243" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 289, + 772 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 69, + 72, + 289, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 289, + 116 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 289, + 116 + ], + "type": "text", + "content": "Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. 2020. Deberta: Decoding-enhanced bert with disentangled attention. arXiv preprint arXiv:2006.03654." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 126, + 289, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 126, + 289, + 212 + ], + "spans": [ + { + "bbox": [ + 69, + 126, + 289, + 212 + ], + "type": "text", + "content": "Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. 2019. Parameter-efficient transfer learning for NLP. In Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pages 2790-2799. PMLR." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 223, + 289, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 223, + 289, + 288 + ], + "spans": [ + { + "bbox": [ + 69, + 223, + 289, + 288 + ], + "type": "text", + "content": "Ganesh Jawahar, Benoit Sagot, and Djamé Seddah. 2019. What does BERT learn about the structure of language? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3651-3657, Florence, Italy. Association for Computational Linguistics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 298, + 289, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 298, + 289, + 375 + ], + "spans": [ + { + "bbox": [ + 69, + 298, + 289, + 375 + ], + "type": "text", + "content": "Brian Lester, Rami Al-Rfou, and Noah Constant. 2021. The power of scale for parameter-efficient prompt tuning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3045-3059, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 385, + 289, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 385, + 289, + 428 + ], + "spans": [ + { + "bbox": [ + 69, + 385, + 289, + 428 + ], + "type": "text", + "content": "Hector Levesque, Ernest Davis, and Leora Morgenstern. 2012. The winograd schema challenge. In Thirteenth international conference on the principles of knowledge representation and reasoning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 438, + 289, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 438, + 289, + 525 + ], + "spans": [ + { + "bbox": [ + 69, + 438, + 289, + 525 + ], + "type": "text", + "content": "Xiang Lisa Li and Percy Liang. 2021. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4582-4597, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 535, + 289, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 535, + 289, + 623 + ], + "spans": [ + { + "bbox": [ + 69, + 535, + 289, + 623 + ], + "type": "text", + "content": "Xiao Liu, Kaixuan Ji, Yicheng Fu, Weng Tam, Zhengxiao Du, Zhilin Yang, and Jie Tang. 2022. P-tuning: Prompt tuning can be comparable to fine-tuning across scales and tasks. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 61-68, Dublin, Ireland. Association for Computational Linguistics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 632, + 289, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 632, + 289, + 665 + ], + "spans": [ + { + "bbox": [ + 69, + 632, + 289, + 665 + ], + "type": "text", + "content": "Xiao Liu, Yanan Zheng, Zhengxiao Du, Ming Ding, Yujie Qian, Zhilin Yang, and Jie Tang. 2021. Gpt understands, too. arXiv:2103.10385." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 675, + 289, + 729 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 675, + 289, + 729 + ], + "spans": [ + { + "bbox": [ + 69, + 675, + 289, + 729 + ], + "type": "text", + "content": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 739, + 289, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 739, + 289, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 739, + 289, + 772 + ], + "type": "text", + "content": "Nafise Sadat Moosavi, Quentin Delfosse, Kristian Kersting, and Iryna Gurevych. 2022. Adaptable Adapters. In Proceedings of the 2022 Annual Conference of" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 524, + 772 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 314, + 72, + 524, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 524, + 105 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 524, + 105 + ], + "type": "text", + "content": "the North American Chapter of the Association for Computational Linguistics, Seattle, WA, USA. Association for Computational Linguistics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 115, + 524, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 115, + 524, + 213 + ], + "spans": [ + { + "bbox": [ + 304, + 115, + 524, + 213 + ], + "type": "text", + "content": "Mohammad Taher Pilehvar and Jose Camacho-Collados. 2019. WiC: the word-in-context dataset for evaluating context-sensitive meaning representations. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 1267-1273, Minneapolis, Minnesota. Association for Computational Linguistics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 223, + 524, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 223, + 524, + 278 + ], + "spans": [ + { + "bbox": [ + 304, + 223, + 524, + 278 + ], + "type": "text", + "content": "Melissa Roemmele, Cosmin Adrian Bejan, and Andrew S Gordon. 2011. Choice of plausible alternatives: An evaluation of commonsense causal reasoning. In AAAI spring symposium: logical formalizations of commonsense reasoning, pages 90-95." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 287, + 524, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 287, + 524, + 375 + ], + "spans": [ + { + "bbox": [ + 304, + 287, + 524, + 375 + ], + "type": "text", + "content": "Andreas Rückle, Gregor Geigle, Max Glockner, Tilman Beck, Jonas Pfeiffer, Nils Reimers, and Iryna Gurevych. 2021. AdapterDrop: On the efficiency of adapters in transformers. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7930-7946, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 384, + 524, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 524, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 524, + 449 + ], + "type": "text", + "content": "Ian Tenney, Dipanjan Das, and Ellie Pavlick. 2019. BERT rediscovers the classical NLP pipeline. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4593-4601, Florence, Italy. Association for Computational Linguistics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 459, + 524, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 524, + 523 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 524, + 523 + ], + "type": "text", + "content": "Erik F. Tjong Kim Sang and Fien De Meulder. 2003. Introduction to the CoNLL-2003 shared task: Language-independent named entity recognition. In Proceedings of the Seventh Conference on Natural Language Learning at HLT-NAACL 2003, pages 142-147." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 534, + 524, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 524, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 524, + 589 + ], + "type": "text", + "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 598, + 524, + 674 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 598, + 524, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 598, + 524, + 674 + ], + "type": "text", + "content": "Alex Wang, Yada Pruksachatkun, Nikita Nangia, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. 2019. SuperGLUE: A stickier benchmark for general-purpose language understanding systems. In Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 684, + 524, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 684, + 524, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 684, + 524, + 772 + ], + "type": "text", + "content": "Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. 2018. GLUE: A multi-task benchmark and analysis platform for natural language understanding. In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 353-355, Brussels, Belgium. Association for Computational Linguistics." + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1244" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 291, + 138 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 291, + 138 + ], + "type": "text", + "content": "Ralph Weischedel, Martha Palmer, Mitchell Marcus, Eduard Hovy, Sameer Pradhan, Lance Ramshaw, Nianwen Xue, Ann Taylor, Jeff Kaufman, Michelle Franchini, Mohammed El-Bachouti, Robert Belvin, and Ann Houston. 2013. OntoNotes Release 5.0. Abacus Data Network." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 153, + 291, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 153, + 291, + 285 + ], + "spans": [ + { + "bbox": [ + 69, + 153, + 291, + 285 + ], + "type": "text", + "content": "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 68, + 302, + 201, + 317 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 302, + 201, + 317 + ], + "spans": [ + { + "bbox": [ + 68, + 302, + 201, + 317 + ], + "type": "text", + "content": "A Experimental Details" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 327, + 290, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 327, + 290, + 436 + ], + "spans": [ + { + "bbox": [ + 67, + 327, + 290, + 436 + ], + "type": "text", + "content": "Datasets In the full data setting, all train-dev-test splits follow P-Tuning v2 (Liu et al., 2022). For low resources setting, to generate k-shot (" + }, + { + "bbox": [ + 67, + 327, + 290, + 436 + ], + "type": "inline_equation", + "content": "k = 16, 32" + }, + { + "bbox": [ + 67, + 327, + 290, + 436 + ], + "type": "text", + "content": ") datasets on SuperGLUE, the fixed set of random seed [11,21,42,87,100] is utilized to sample instances in training and development set, while the entire development set is treated as test set, where the average performance is reported in Table 2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 449, + 291, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 449, + 291, + 571 + ], + "spans": [ + { + "bbox": [ + 67, + 449, + 291, + 571 + ], + "type": "text", + "content": "Experimental Setting We grid search the learning rate over [5e-3, 7e-3, 1e-2, 1e-4], training epoch over [20, 40, 60, 80, 100, 120], batch size over [8, 16, 32], and random seeds over [11, 21, 42, 87, 100]. For a fair comparison, the prefix length utilized by APT is consistent with P-Tuning v2. In low resources setting, the batch size we used is 2. In Eq.(4), we take the hidden states of the first input token as representation in previous layer." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 584, + 291, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 584, + 291, + 664 + ], + "spans": [ + { + "bbox": [ + 67, + 584, + 291, + 664 + ], + "type": "text", + "content": "Experimental Computation We use the pretrained model BERT-base with 110M parameters, BERT-large with 335M parameters, RoBERTa-large with 355M parameters and DeBERTa-xlarge with 750M parameters. We conduct experiments on NVIDIA V100 or A100 GPUs for each task." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 680, + 220, + 693 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 680, + 220, + 693 + ], + "spans": [ + { + "bbox": [ + 67, + 680, + 220, + 693 + ], + "type": "text", + "content": "B Further Ablation Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 705, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 705, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 705, + 291, + 772 + ], + "type": "text", + "content": "We demonstrate further ablation results on BERT-large and RoBERTa-large as shown in Table 5. It can be found that the beneficial impact of the three strategies and the observation is consistent with BERT-base in Section 4.3 in general." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 73, + 413, + 150 + ], + "blocks": [ + { + "bbox": [ + 309, + 73, + 413, + 150 + ], + "lines": [ + { + "bbox": [ + 309, + 73, + 413, + 150 + ], + "spans": [ + { + "bbox": [ + 309, + 73, + 413, + 150 + ], + "type": "image", + "image_path": "e6bc50894731e36b1bee2fca83254350a9cafe9c3c5973e104559060a17484c6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 341, + 154, + 380, + 165 + ], + "lines": [ + { + "bbox": [ + 341, + 154, + 380, + 165 + ], + "spans": [ + { + "bbox": [ + 341, + 154, + 380, + 165 + ], + "type": "text", + "content": "(a) COPA" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 417, + 73, + 521, + 149 + ], + "blocks": [ + { + "bbox": [ + 417, + 73, + 521, + 149 + ], + "lines": [ + { + "bbox": [ + 417, + 73, + 521, + 149 + ], + "spans": [ + { + "bbox": [ + 417, + 73, + 521, + 149 + ], + "type": "image", + "image_path": "904200878bce5b4c5915b7b1a9a571384d8ddbb69cc1d9df19b652004f0281fb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 451, + 154, + 486, + 165 + ], + "lines": [ + { + "bbox": [ + 451, + 154, + 486, + 165 + ], + "spans": [ + { + "bbox": [ + 451, + 154, + 486, + 165 + ], + "type": "text", + "content": "(b) WSC" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 302, + 178, + 524, + 202 + ], + "lines": [ + { + "bbox": [ + 302, + 178, + 524, + 202 + ], + "spans": [ + { + "bbox": [ + 302, + 178, + 524, + 202 + ], + "type": "text", + "content": "Figure 3: The performance of APT and PT-2 on COPA and WSC in a range of prefix length on BERT-large." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 223, + 398, + 238 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 223, + 398, + 238 + ], + "spans": [ + { + "bbox": [ + 302, + 223, + 398, + 238 + ], + "type": "text", + "content": "C Prefix Length" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 244, + 526, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 244, + 526, + 352 + ], + "spans": [ + { + "bbox": [ + 302, + 244, + 526, + 352 + ], + "type": "text", + "content": "The prefix length is an important hyper-parameter for prefix tuning and APT. Figure 3 illustrates the performance of APT and P-Tuning v2 with different prefix lengths over a range. It can be observed that APT is superior to P-Tuning v2 in most prefix length settings, indicating that APT has a relatively wider range of prefix length to achieve better performance." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 363, + 422, + 375 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 363, + 422, + 375 + ], + "spans": [ + { + "bbox": [ + 302, + 363, + 422, + 375 + ], + "type": "text", + "content": "D Scientific Artifacts" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 384, + 526, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 384, + 526, + 627 + ], + "spans": [ + { + "bbox": [ + 302, + 384, + 526, + 627 + ], + "type": "text", + "content": "We use datasets involving SuperGLUE (Wang et al., 2019) benchmark including BoolQ (Clark et al., 2019a), COPA (Roemmle et al., 2011), RTE (Wang et al., 2018), WiC (Pilehvar and Camacho-Collados, 2019) and WSC (Levesque et al., 2012) as well as 3 Named Entity Recognition (NER) tasks including CoNLL03 (Tjong Kim Sang and De Meulder, 2003), CoNLL04 (Carreras and Marquez, 2004), and OntoNotes 5.0 (Weischedel et al., 2013). The pre-trained model we used are BERT-base / large (Devlin et al., 2019), RoBERTa-large (Liu et al., 2019) and DeBERTa-xlarge (He et al., 2020). We use HuggingFace Transformers (Wolf et al., 2020) and P-Tuning v2 (Liu et al., 2022) as the codebase implemented by PyTorch. They are all open-source and we only use for academic research which is consistent with their intended use." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 316, + 760, + 392, + 773 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 760, + 392, + 773 + ], + "spans": [ + { + "bbox": [ + 316, + 760, + 392, + 773 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 316, + 760, + 392, + 773 + ], + "type": "text", + "content": "https://pytorch.org/" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "1245" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 336, + 526, + 470 + ], + "blocks": [ + { + "bbox": [ + 70, + 336, + 526, + 470 + ], + "lines": [ + { + "bbox": [ + 70, + 336, + 526, + 470 + ], + "spans": [ + { + "bbox": [ + 70, + 336, + 526, + 470 + ], + "type": "table", + "html": "
ModelSettingSuperGLUENER
BoolQCOPARTEWiCWSCAvg.CoNLL03CoNLL04OntoNotesAvg.
BERT-largeAPT76.079.079.475.170.275.990.785.888.688.4
w/o token-level α75.877.077.374.868.374.691.184.488.588.0
w/o layer-level λ75.474.076.974.668.373.890.783.788.487.6
w/o hidden states h74.776.075.874.668.373.991.284.088.687.9
RoBERTa-largeAPT84.894.089.974.668.382.392.789.089.890.5
w/o token-level α84.388.088.173.065.479.892.288.789.590.1
w/o layer-level λ84.788.086.372.164.479.192.088.789.890.2
w/o hidden states h83.991.087.072.964.479.892.288.789.490.1
", + "image_path": "7eb934468cc6e92a90aad0be5fb3ea83bf16a423f5b869a2a4501151ef10badc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 477, + 525, + 502 + ], + "lines": [ + { + "bbox": [ + 67, + 477, + 525, + 502 + ], + "spans": [ + { + "bbox": [ + 67, + 477, + 525, + 502 + ], + "type": "text", + "content": "Table 5: Ablation experiments on BERT-large and RoBERTa-large for two different level gate mechanisms and the hidden states from the previous layer. **bold:** the best score." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "text", + "content": "1246" + } + ] + } + ], + "index": 2 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 90, + 192, + 103 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 90, + 192, + 103 + ], + "spans": [ + { + "bbox": [ + 68, + 90, + 192, + 103 + ], + "type": "text", + "content": "A For every submission:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 76, + 106, + 414, + 242 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "spans": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "type": "text", + "content": "A1. Did you describe the limitations of your work? section limitations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 76, + 143, + 329, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 143, + 329, + 169 + ], + "spans": [ + { + "bbox": [ + 76, + 143, + 329, + 169 + ], + "type": "text", + "content": "A2. Did you discuss any potential risks of your work? Not applicable. Left blank." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 179, + 414, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 179, + 414, + 205 + ], + "spans": [ + { + "bbox": [ + 77, + 179, + 414, + 205 + ], + "type": "text", + "content": "A3. Do the abstract and introduction summarize the paper's main claims? section abstract and section 1 introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 76, + 215, + 398, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 215, + 398, + 242 + ], + "spans": [ + { + "bbox": [ + 76, + 215, + 398, + 242 + ], + "type": "text", + "content": "A4. Have you used AI writing assistants when working on this paper? Left blank." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 68, + 252, + 291, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 252, + 291, + 266 + ], + "spans": [ + { + "bbox": [ + 68, + 252, + 291, + 266 + ], + "type": "text", + "content": "B Did you use or create scientific artifacts?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 79, + 271, + 129, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 271, + 129, + 281 + ], + "spans": [ + { + "bbox": [ + 79, + 271, + 129, + 281 + ], + "type": "text", + "content": "section 4.1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 76, + 291, + 524, + 634 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 77, + 291, + 315, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 291, + 315, + 317 + ], + "spans": [ + { + "bbox": [ + 77, + 291, + 315, + 317 + ], + "type": "text", + "content": "B1. Did you cite the creators of artifacts you used? section 4.1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 77, + 327, + 463, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 327, + 463, + 355 + ], + "spans": [ + { + "bbox": [ + 77, + 327, + 463, + 355 + ], + "type": "text", + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? section D Scientific Artifacts" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 77, + 364, + 524, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 364, + 524, + 431 + ], + "spans": [ + { + "bbox": [ + 77, + 364, + 524, + 431 + ], + "type": "text", + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? section D Scientific Artifacts" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 77, + 441, + 524, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 441, + 524, + 496 + ], + "spans": [ + { + "bbox": [ + 77, + 441, + 524, + 496 + ], + "type": "text", + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? We use open-source datasets and do not change datasets for a fair comparison." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 76, + 504, + 524, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 504, + 524, + 544 + ], + "spans": [ + { + "bbox": [ + 76, + 504, + 524, + 544 + ], + "type": "text", + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Not applicable. Left blank." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 77, + 554, + 524, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 554, + 524, + 634 + ], + "spans": [ + { + "bbox": [ + 77, + 554, + 524, + 634 + ], + "type": "text", + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. It can be found in the cited paper." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "spans": [ + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "type": "text", + "content": "C Did you run computational experiments?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 79, + 662, + 178, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 662, + 178, + 674 + ], + "spans": [ + { + "bbox": [ + 79, + 662, + 178, + 674 + ], + "type": "text", + "content": "section 4 Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 77, + 683, + 524, + 724 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 683, + 524, + 724 + ], + "spans": [ + { + "bbox": [ + 77, + 683, + 524, + 724 + ], + "type": "text", + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Table 1 and section appendix A Experimental Computation" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "spans": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "text", + "content": "ACL 2023 Responsible NLP Checklist" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "spans": [ + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "type": "text", + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1247" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 77, + 70, + 523, + 238 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 77, + 70, + 523, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 70, + 523, + 111 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 523, + 111 + ], + "type": "text", + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? section appendix A Experimental Details" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 77, + 120, + 523, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 120, + 523, + 174 + ], + "spans": [ + { + "bbox": [ + 77, + 120, + 523, + 174 + ], + "type": "text", + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Table 2 report the mean and std results." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 184, + 523, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 184, + 523, + 238 + ], + "spans": [ + { + "bbox": [ + 77, + 184, + 523, + 238 + ], + "type": "text", + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? We follow the existing work and keep consistent with them." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 67, + 247, + 522, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 247, + 522, + 278 + ], + "spans": [ + { + "bbox": [ + 67, + 247, + 522, + 278 + ], + "type": "text", + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 76, + 286, + 523, + 539 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 76, + 286, + 523, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 286, + 523, + 326 + ], + "spans": [ + { + "bbox": [ + 76, + 286, + 523, + 326 + ], + "type": "text", + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 76, + 336, + 523, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 336, + 523, + 390 + ], + "spans": [ + { + "bbox": [ + 76, + 336, + 523, + 390 + ], + "type": "text", + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 76, + 400, + 523, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 400, + 523, + 453 + ], + "spans": [ + { + "bbox": [ + 76, + 400, + 523, + 453 + ], + "type": "text", + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 76, + 462, + 519, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 462, + 519, + 489 + ], + "spans": [ + { + "bbox": [ + 76, + 462, + 519, + 489 + ], + "type": "text", + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 76, + 498, + 523, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 498, + 523, + 539 + ], + "spans": [ + { + "bbox": [ + 76, + 498, + 523, + 539 + ], + "type": "text", + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1248" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Towards Fewer Hallucinations in Knowledge-Grounded Dialogue Generation via Augmentative and Contrastive Knowledge-Dialogue/989c66cd-1d3d-4fda-ba4f-e2869fc255a6_content_list.json b/2023/Towards Fewer Hallucinations in Knowledge-Grounded Dialogue Generation via Augmentative and Contrastive Knowledge-Dialogue/989c66cd-1d3d-4fda-ba4f-e2869fc255a6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5b299bea51beeb8be25b41ca3a7c498239c56c82 --- /dev/null +++ b/2023/Towards Fewer Hallucinations in Knowledge-Grounded Dialogue Generation via Augmentative and Contrastive Knowledge-Dialogue/989c66cd-1d3d-4fda-ba4f-e2869fc255a6_content_list.json @@ -0,0 +1,1591 @@ +[ + { + "type": "text", + "text": "Towards Fewer Hallucinations in Knowledge-Grounded Dialogue Generation via Augmentative and Contrastive Knowledge-Dialogue", + "text_level": 1, + "bbox": [ + 149, + 87, + 848, + 129 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bin Sun $^{1}$ , Yitong Li $^{2,3}$ , Fei Mi $^{2}$ , FanHu Bie $^{3}$ , Yiwei Li $^{1}$ , Kan Li $^{1*}$", + "bbox": [ + 221, + 141, + 776, + 158 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science & Technology, Beijing Institute of Technology", + "bbox": [ + 191, + 159, + 813, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Huawei Noah's Ark Lab $^{3}$ Huawei Technologies Ltd.", + "bbox": [ + 272, + 175, + 731, + 192 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{binsun,liyiwei,likan}@bit.edu.cn", + "bbox": [ + 374, + 195, + 626, + 208 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{liyitong3,mifei2,biefanhu}@huawei.com", + "bbox": [ + 356, + 211, + 645, + 225 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 252, + 339, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing knowledge-grounded open-domain dialogue generation models often face the hallucination problem, i.e. the dialogue generative model will persist in an inappropriate knowledge and generate responses that inconsistent with the facts. We argue that this problem mainly stems from the polarized optimization objectives and weak knowledge generation ability. To mitigate the hallucination, we take inspiration from human communicating that people will replay euphemistic responses for the unclear or unrecognizable knowledge, and propose an Augmentative and Contrastive Knowledge Dialogue Expansion Framework (ACK-DEF). ACK-DEF constructs the augmentative and contrastive knowledge dialogue samples, which consist of the knowledge of different degrees of errors and the response of manual design, to expand the original training set and smooth the polarized optimization objective that enables models to generate ground-truth with or without gold knowledge. Not only the knowledge, ACK-DEF also provides the tactful responses of manual design corresponding to the incomplete correct knowledge. Experimental results on the Wikipedia of Wizard dataset show that employing the ACK-DEF is effective to alleviate the hallucination problem.", + "bbox": [ + 144, + 277, + 460, + 675 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 114, + 684, + 258, + 699 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, Knowledge-Grounded Dialogue Generation draws dramatic attentions in artificial intelligence community. Many efforts incorporate knowledge information to improve the performance of dialogue generation models (Zhou et al., 2018; Dinan et al., 2019; Gopalakrishnan et al., 2019; Kim et al., 2020; Zhao et al., 2020a; Zheng et al., 2021; Zhao et al., 2022a; Bao et al., 2022). However, these methods always face the hallucination problem, that is, the dialogue generation model may insist on an inappropriate knowledge and generate responses that inconsistent with the facts (Rashkin et al., 2021; Zhao et al., 2022a; Dziri et al., 2022).", + "bbox": [ + 112, + 709, + 489, + 917 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We argue that the hallucination problem primarily caused by two aspects: (1) The optimization objective is usually polarized by the gold knowledge dialogue samples and general dialogue samples without knowledge in current knowledge-grounded dialogue datasets (Zhou et al., 2018; Gopalakrishnan et al., 2019; Dinan et al., 2019; Wu et al., 2019; Komeili et al., 2022). Few datasets consider teaching models how to respond when dealing with incomplete correct knowledge, which makes the models tend to believe in the given knowledge, regardless of whether the knowledge is appropriate or not, resulting in hallucination problems. In addition, the knowledge retrieval system tends to extract irrelevant knowledge rather than relevant knowledge when the database is large, aggravating the hallucinations (Reimers and Gurevych, 2021; Liu et al., 2022). (2) The generation of knowledge may also face the hallucination problem and obtain the inappropriate knowledge, leading the generation of hallucination responses (Kim et al., 2020; Zhao et al., 2020a; Liu et al., 2022; Adolphs et al., 2021; Bao et al., 2022).", + "bbox": [ + 507, + 253, + 884, + 621 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To mitigate the hallucination problem, we propose an Augmentative and Contrastive Knowledge Dialogue Expansion Framework (ACK-DEF), which is inspired by human communicating that people will replay euphemistic response for the unrecognizable knowledge. ACK-DEF is proposed to smooth the polarized optimization objective by augmenting training set with augmentative and contrastive knowledge-dialogue samples. Not only the knowledge, we also designed the reply patterns for the knowledge with different level of errors. For this, we propose the augmentative knowledge dialogue expansion (AK), and contrastive knowledge dialogue expansion (CK). AK is proposed to boost the generalization ability of models on knowledge with minor noise. On the contrary, inspired from the contrastive learning paradigm (Cai et al., 2020; Chen et al., 2020a,b; Sun et al., 2021, 2022), CK", + "bbox": [ + 507, + 629, + 884, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1741", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics", + "bbox": [ + 226, + 945, + 771, + 957 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Volume 2: Short Papers, pages 1741-1750", + "bbox": [ + 368, + 958, + 630, + 971 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "July 9-14, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 295, + 972, + 700, + 985 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b3d345dea917129cfa22a2ad09d6cca3aea481adcf1f8a95852176977677e589.jpg", + "image_caption": [ + "Figure 1: A diagram of our Augmentative Knowledge dialogue expansion method. We replace different proportion of words in the original knowledge with synonyms to construct incomplete correct knowledge, and design response for different knowledge. We also use prompts to guide the dialogue generation process." + ], + "image_footnote": [], + "bbox": [ + 129, + 80, + 473, + 262 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "reconstructs incorrect knowledge and designs euphemistic responses, which aims to push the model learn the reply pattern of incorrect knowledge and a better boundary between correct and incorrect knowledge.", + "bbox": [ + 112, + 384, + 487, + 463 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions: We propose an ACK-DEF to construct new knowledge-dialogue samples that consist of knowledge with different level of errors and manual responses, to soften the training optimization objectives of models, which will mitigate the hallucination. Finally, we conduct extension experiments to show the superior performance of ACK-DEF on alleviating the hallucination.", + "bbox": [ + 112, + 464, + 487, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Methodology", + "text_level": 1, + "bbox": [ + 112, + 605, + 263, + 621 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To mitigate the hallucination problem that caused by the polarized optimization objectives in knowledge grounded dialogue generation, we take inspiration from human communicating, and propose the Augmentative and Contrastive Knowledge Dialogue Expansion Framework (ACK-DEF). Our ACK-DEF aims to soften the polarized training optimization objectives of current knowledge-grounded dialogue generation methods, and guide the dialogue system reply patterns for the knowledge with different level of errors. To achieve this end, we design two effective expansion method, which will be detailed in below.", + "bbox": [ + 112, + 630, + 489, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 Augmentative Knowledge Dialogue", + "text_level": 1, + "bbox": [ + 112, + 850, + 436, + 866 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose the Augmentative Knowledge (AK) dialogue expansion to boost the generalization ability of the dialogue model on the knowledge with simi", + "bbox": [ + 112, + 871, + 489, + 917 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/19f02e3a689096203540d697ce684f8bf997410565aa7090e0aeacf6395192ce.jpg", + "image_caption": [ + "Figure 2: A diagram of our Contrastive Knowledge dialogue expansion method. We use the antonym to reconstruct the knowledge information and design multiple responses for such knowledge. Since antonyms transform the semantics of the original knowledge, the noise knowledge often contains wrong facts. By this, the model can learn a better boundary between correct and incorrect knowledge, and a safety reply pattern for incorrect knowledge." + ], + "image_footnote": [], + "bbox": [ + 524, + 80, + 870, + 200 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "lar semantics but different expressions, which can prevent the model from being interfered by partialrelevant knowledge retrieved by the retrieval systems (Lian et al., 2019; Zhao et al., 2020b; Hedayatnia et al., 2020; Zheng et al., 2021; Shuster et al., 2021; Komeili et al., 2022). As shown in Figure 1, we employ the synonym data augmentation tool, which replaces words in the original knowledge with synonyms, to reconstruct the knowledge information (Miller, 1995). Considering that the synonym may disrupt the original semantics of new constructed knowledge, we constrain the replace possibility within [0.1,0.2]. Hence, we can obtain the approximate knowledge. Combining this knowledge and the original dialogue, we obtain the \"ak-less sample\". In addition, we also replace $30\\%$ to $50\\%$ words with their synonyms to construct the less similar knowledge. Inspired from prompt learning paradigm (Yao et al., 2022; Valvoda et al., 2022; Zhao et al., 2022b), we manually produce some Prefix-prompts and Post-prompts (see Appendix) to (1) make the new response more tactful for the less similar knowledge; (2) regulate and guide the dialogue generation process of the model. We call the sample consist of less-similar knowledge and designed response as \"ak-more sample\".", + "bbox": [ + 505, + 367, + 884, + 785 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2 Contrastive Knowledge Dialogue", + "text_level": 1, + "bbox": [ + 507, + 800, + 815, + 815 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose the Contrastive Knowledge (CK) dialogue expansion, inspired from the contrastive learning paradigm (Chen et al., 2020b; Cai et al., 2020), not only construct the incorrect knowledge as negative samples for original knowledge, but also build the euphemistic responses as positive", + "bbox": [ + 507, + 822, + 884, + 919 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1742", + "bbox": [ + 480, + 927, + 521, + 940 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "samples for the original response with incorrect knowledge. To help the model learn a boundary between correct and incorrect knowledge, we employ the antonym to make up new incorrect knowledge. For example, given the knowledge \"nintendo was founded on 23 september 1889 ...\", the \"founded\" will be replaced with \"abolish\", which greatly changes the semantics but little changes the expression. After that, we random choose an euphemistic response to replace the original response of the dialogue. Finally, The incorrect knowledge and the replaced euphemistic response are combined as the \"ck-sample\".", + "bbox": [ + 110, + 84, + 492, + 294 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 Experiment and Results", + "text_level": 1, + "bbox": [ + 112, + 312, + 357, + 328 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1 Experiment Settings", + "text_level": 1, + "bbox": [ + 112, + 342, + 319, + 357 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1 Dataset", + "text_level": 1, + "bbox": [ + 112, + 367, + 233, + 381 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We use the Wikipedia of Wizard (WoW) data, a well-established knowledge-grounded open-domain dialogue dataset, for our experiment. We pre-precess the WoW dataset and extract the single-turn knowledge dialogue samples. To evaluate the performance of our method in detail, we perform four test sets: normal, ak-less, ak-more and ck. The normal set is the original test set. And the ak-less, ak-more and ck are the sets consist of ak-less, ak-more and ck samples, respectively. We also follow the settings of WoW data and divide the test set into two groups (seen test and unseen test): the topic of the knowledge in the unseen test set is missing in the training set.", + "bbox": [ + 110, + 390, + 489, + 615 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.2 Baseline", + "text_level": 1, + "bbox": [ + 112, + 631, + 240, + 645 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We employ the released PLATO-v1 (Bao et al., 2020) model, a pre-trained dialogue generation model based on UniLM, for our experiment.", + "bbox": [ + 112, + 653, + 487, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fine-tuning We directly finetune a model on the original WoW training set. By this, the model can only see gold knowledge dialogue samples and general dialogue samples without knowledge. Hence, we call the fine-tuned model PLATO+GOLD.", + "bbox": [ + 112, + 719, + 489, + 797 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fine-tuning with ACK-DEF We finetune the model with the original set and the expansion samples that obtained through ACK-DEF. Thence, we call it PLATO+ACK-DEF.", + "bbox": [ + 112, + 815, + 489, + 878 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.3 AutoEvaluation Metrics", + "text_level": 1, + "bbox": [ + 507, + 84, + 757, + 98 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dialogue Metrics Our primary metrics of interest are Distinct-n (Li et al., 2016), Response Length (Len.) (Csaky et al., 2019), BLEU (Papineni et al., 2002), Embedding-based (Greedy (GRE), Average (AVG), Extrema (EXT)) (Liu et al., 2016), and Coherence (COH) (Xu et al., 2018). Distinct-n evaluates the diversity of generated responses, which is calculated through the ratio of distinct $n$ -grams and all generated $n$ -grams. Len. is the average number of words of all generated responses. BLEU validates the degree of the word-overlap between the generated response and the ground-truth, which denotes the consistence between generated response and ground-truth. Embedding-based metrics (GRE, AVG and EXT) are introduced to evaluate the semantic relationship of generated responses and ground-truth responses, illustrating the consistence in semantic level. COH. mainly assesses the relevance between contexts and generated responses.", + "bbox": [ + 507, + 103, + 884, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Knowledge Metrics We follow the PLATO(Bao et al., 2020) and use the knowledge precision, recall and f1 scores. These metrics are used to calculate the ratio of tokens that exist in common in ground-truth knowledge and generated responses to tokens in generated responses. \"Recall\" is the average ratio of the number of overlapping tokens in response and knowledge to the number of tokens in knowledge. And \"Precision\" is the average ratio of the number of overlapping tokens to the number of tokens in response. In other words, \"Recall\" indicates how much knowledge information is contained in the response, while \"Precision\" indicates the proportion of knowledge information in the response. Even we involve the negative and incorrect knowledge in response generation, we still use the ground-truth knowledge to calculate the metrics in Table 3,4.", + "bbox": [ + 507, + 419, + 884, + 707 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2 Dialogue Performance Analysis", + "text_level": 1, + "bbox": [ + 507, + 721, + 801, + 736 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1 and Table 2 report the automatic results on four test sets and four unseen test sets, respectively. In these Tables, it can be observed that (1) the PLATO+ACK-DEF has a competitive performance with PLATO+GOLD on the normal set, which means that the PLATO+ACK-DEF can recognize the golden knowledge and produce appropriate responses. (2) the PLATO+GOLD perform worse than PLATO+ACK-DEF on ak-less, which means that the robustness of the dialogue model only trained with golden knowledge is very weak.", + "bbox": [ + 507, + 741, + 884, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1We manually construct some responses, please see Appendix for the detail.", + "bbox": [ + 112, + 891, + 489, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1743", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/a67f259174ebbf0e1f9043752f8d5f4f805c1123dc506c842abda3e436d84a4e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
test setDistinct-1/2/3Len.BLEU-1/2/3/4GREAVGEXTCOH
normal0.10680.453313.690.42800.29650.21100.15290.73920.86890.63610.7808
0.09020.398416.200.44280.30170.21090.14990.73660.86830.63300.7878
ak-less0.11940.502413.500.38610.25740.17450.11920.71600.86070.61480.7755
0.08230.353218.780.45020.29820.20150.13800.73070.86960.62930.7948
ak-more0.12340.517412.810.16750.10620.06800.04350.69080.85510.59940.7706
0.06750.294621.830.43580.30010.21230.15420.77450.91510.70930.8098
ck0.11090.477913.230.29650.17790.10800.06570.58380.76220.53730.7712
0.06520.202913.360.42300.27050.18090.12660.65720.83060.61620.8049
", + "bbox": [ + 126, + 82, + 870, + 225 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/c881791a65bea10b76e98faf787df69b22ddb480eee1e130354fa0ebb2ebc361.jpg", + "table_caption": [ + "Table 1: The automatic results of PLATO+GOLD (up) and PLATO+ACK-DEF (down) on four test seen sets." + ], + "table_footnote": [], + "table_body": "
test setDistinct-1/2Len.BLUE-1/2/3/4GREAVGEXTCOH
normal0.05030.242212.430.35160.23310.15820.10900.69880.85680.63060.8094
0.04670.231113.140.34630.22810.15360.10490.69680.85410.63380.8105
ak-less0.09660.391713.390.38710.25650.17240.11640.71430.86000.61220.7836
0.06230.266419.180.44430.29070.19360.13010.72320.86630.61940.8026
ak-more0.10640.444012.710.16520.10460.06680.04260.68880.85380.59800.7797
0.05610.240021.820.43310.29680.20910.15110.76970.91140.70370.8197
ck0.08130.332413.240.30110.18090.11000.06690.58540.76760.54790.7794
0.04650.149013.520.43290.27750.18610.13070.66120.83340.62150.8145
", + "bbox": [ + 126, + 263, + 870, + 407 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/a87608cbd02b6c4e4d49c9a8af9a6556762e9af95f42b39c1c2d58fb13c5a0fc.jpg", + "table_caption": [ + "Table 2: The automatic results of PLATO+GOLD (up) and PLATO+ACK-DEF (down) on four test sets with unseen knowledge." + ], + "table_footnote": [], + "table_body": "
test setRecallPrecisionF1avg. Dec.
normal0.36070.70090.4546-
ak-less0.28830.55850.3618∇ 0.1026
ak-more0.17520.36320.2228∇ 0.2517
ck0.31930.61330.4003∇ 0.0611
normal0.36950.65380.4520-
ak-less0.32510.56360.3927∇ 0.0647
ak-more0.23350.39830.2775∇ 0.1887
ck0.10650.20410.1337∇ 0.3437
", + "bbox": [ + 122, + 468, + 478, + 600 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/a03261e80b612cb54fd6234ecea69baca569b4fe3c58a8a91e06df4c1a9e1370.jpg", + "table_caption": [ + "Table 3: The knowledge correlation results of PLATO+GOLD (up) and PLATO+ACK-DEF (down) on four test sets with seen knowledge." + ], + "table_footnote": [], + "table_body": "
test setRecallPrecisionF1avg. Dec.
normal0.37320.74420.4736-
ak-less0.27280.54750.3452∇ 0.1418
ak-more0.16650.36270.2152∇ 0.2822
ck0.30280.60680.3830∇ 0.0995
normal0.36550.68820.4535-
ak-less0.29380.53480.3579∇ 0.1069
ak-more0.20460.37140.2481∇ 0.2277
ck0.08700.18470.1116∇ 0.3747
", + "bbox": [ + 124, + 671, + 478, + 802 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 4: The knowledge correlation results of PLATO+GOLD (up) and PLATO+ACK-DEF (down) on four test sets with unseen knowledge.", + "bbox": [ + 112, + 812, + 487, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Even if the knowledge information only changes by $10\\%$ to $20\\%$ , the performance of the model will", + "bbox": [ + 112, + 887, + 487, + 917 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "significantly decline, especially consistency metrics (i.e. BLEU, GRE, AVG and EXT). (3) the PLATO+GOLD achieve better Distinct scores but weaker BLEU and embedding-based scores, which means that the PLATO+GOLD is easy to generate responses that are very different from ground-truth responses, that is, the hallucinations.", + "bbox": [ + 507, + 472, + 882, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3 Knowledge Correlation Analysis", + "text_level": 1, + "bbox": [ + 507, + 604, + 810, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 3 and Table 4 report the knowledge correlation result of PLATO+GOLD and PLATO+ACK-DEF on four test sets and four test unseen sets, respectively. From these table, we can observe that the performance of PLATO+GOLD is reduced when the given knowledge changed, which illustrates the danger that the model generate responses based on incorrect knowledge. In addition to the above findings, we also observed that the recall, precision and f1 scores of PLATO+ACK-DEF are better than PLATO+GOLD on ak-less and ak-more sets, which demonstrates that using ACK-DEF effectively enhance the model's capability for the similar knowledge information. Moreover, the result of PLATO+ACK-DEF on the ck set is significantly reduced, which shows that the model distinguishes the wrong knowledge constructed with antonyms and gives an appropriate response with", + "bbox": [ + 507, + 629, + 884, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1744", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/aa53b36205c9ac306af3b3709673d3aa885824f3643542f0bb83eecc70b81481.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
test setw. GOLD (%)w. ACK-DEF (%)kappa
normal13.0014.000.481
ak-less23.6717.330.513
ak-more33.6724.330.479
ck21.675.670.597
total23.0015.330.552
", + "bbox": [ + 115, + 80, + 485, + 177 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "out knowledge (see Table 1 and Table 2 for the effect). These results are inline with our exception that incorporating noised knowledge dialogue samples in training stages can smooth the polarized optimization objective, and mitigate the hallucination problem.", + "bbox": [ + 112, + 225, + 487, + 322 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "According to the results of test seen sets and unseen sets), we can notice that the PLATO+ACK-DEF achieves a good performance on ground-truth seen knowledge and a weak performance on ground-truth unseen knowledge. This illustrates that the PLATO+ACK-DEF may doubt the authenticity of unseen given knowledge (even if the knowledge is the ground-truth), and will not fully use it to generate responses. This may alleviate the hallucination, and we believe it is caused by (1) the Augmentative knowledge dialogue introduces similar knowledge to improve the generalization of the model; (2) the Contrastive knowledge dialogue introduces knowledge independent responses, which tell the model to generate responses without knowledge; (3) the ACK-DEF smooths the polarized optimization, which ensures the model not to directly use the given knowledge.", + "bbox": [ + 115, + 324, + 489, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4 Human Evaluation", + "text_level": 1, + "bbox": [ + 112, + 625, + 310, + 639 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To further evaluation the ability of our ACK-DEF on reducing the hallucination problem, we randomly select 400 samples form four test sets, and hire three annotators to do human evaluations by assessing whether the responses generated by PLATO +GOLD and +ACK-DEL have hallucinations. Table 5 reports the results of human evaluation, from which we can notice that the PLATO+ACK-DEF generate less hallucinations than PLATO+GOLD. This shows the effectiveness of our ACK-DEF.", + "bbox": [ + 112, + 645, + 489, + 804 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5 Case Study", + "text_level": 1, + "bbox": [ + 112, + 816, + 248, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 6 shows a generated case of PLATO+GOLD (GOLD) and PLATO+ACK-DEF (ACK-DEF) based on different knowledge information. We can observed that the GOLD is convinced of the given knowledge, regardless of whether the knowledge", + "bbox": [ + 112, + 838, + 487, + 917 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/111b9f5f83d9f38c8bc93b98e00d9c5f0b4f838d443cf77e47a60d152e7e7fd8.jpg", + "table_caption": [ + "Table 5: The human evaluation results." + ], + "table_footnote": [], + "table_body": "
QueryThey really are super cool looking dogs\nthey are so amazing
Normal\nKnowledgegolden retriever are well suited to resi-\ndency in suburban or country environ-\nments.
GOLDI love golden retrievers! They are well\nsuited to residency in suburban or country.
ACK-DEFThey are! They are well suited to suburban\nor country environments.
AK-Less\nKnowledgegolden retriever are good suited to resi-\ndency in suburban operating elbow room\nland environments.
GOLDGolden retriever are great suited to resi-\ndency in suburban operating elbow
ACK-DEFGolden retriever are great suited to subur-\nban environments.
AK-More\nKnowledgegild retriever make up considerably suit\nto residency in suburban surgery common-\nwealth environments.
GOLDgild retrievers make up considerably suit\nto residency in suburban areas.
ACK-DEFNot very clear, maybe they are! They can\nbe found in suburban areas. If it is wrong,\nplease tell me.
CK Knowl-\nedgegolden retriever unequal disadvantage-\ngeously suited to residency in suburban or\ncountry environments.
GOLDThey sure are! I have a golden retriever\nwho is unequally suited to residency in sub-\nurban environments.
ACK-DEFThat sounds pretty good.
", + "bbox": [ + 512, + 80, + 882, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 6: A case of PLATO +GOLD and +ACK-DEF.", + "bbox": [ + 514, + 495, + 875, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "is appropriate or not, and more easily to copy the knowledge information into responses. Even the GOLD has seen the knowledge topic, it could not remember the knowledge in their parameters. On the contrary, the ACK-DEF has good resistance to incomplete correct knowledge.", + "bbox": [ + 507, + 538, + 882, + 634 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Conclusion", + "text_level": 1, + "bbox": [ + 507, + 650, + 640, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This paper focuses on the hallucinations caused by polarized optimization objective in knowledge-grounded dialogue generation (KGDG), and proposes an augmentative and contrastive knowledge dialogue expansion framework (ACK-DEF) to mitigate it. The optimization objective of KGDG is to train the model could generate proper response with or without knowledge, which inevitably weaken the model's ability on unrecognized knowledge and lead hallucinations. Therefore, ACK-DEF constructs multiple level knowledge-dialogue samples to soften the optimization objective of KGDG. Extension experimental results show the superior performance of using our methods on dialogue metrics and knowledge correlations.", + "bbox": [ + 505, + 677, + 884, + 917 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1745", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Limitations", + "text_level": 1, + "bbox": [ + 114, + 84, + 220, + 99 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our limitations are as follow:", + "bbox": [ + 114, + 108, + 334, + 123 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Data Scale: This paper only employs the Wikipedia of Wizard dataset, a small scale and well-established knowledge conversation dataset, and lack of the validation on large-scale dataset.", + "- **Backbones:** This paper lacks the evaluating of other knowledge dialogue model on the proposed method. Actually, we have two reasons to employ the PLATO. First, the PLATO can better handle the one-to-many phenomenon, which is suitable for learning our expansion samples. Second, the PLATO is a pre-trained dialogue model, and its performance on knowledge dialogue generation task has been proved. We will evaluating the performance of other knowledge dialogue model on our method for our future work.", + "- Knowledge Expansion Methods: This paper only uses the synonym and antonym to construct the noised knowledge, which lacks of the comparison of using other data augmentation method. Indeed, we use two token-level data augmentation methods (synonym and antonym augmentation) to prove our statements on hallucination problem in knowledge dialogue generation task. Based on this study, we believe that incorporating other data augmentation methods will also mitigate the hallucinations.", + "- Manual Prompts and Responses: This paper designed five prefix prompts, four post-prompts and nineteen euphemistic responses. For $AK$ -More method, we simply randomly choose one prefix-prompt and one post-prompt and concatenate them with the ground-truth response. This leads to some irregular responses. As for $CK$ method, we randomly select one euphemistic response for the incorrect knowledge. However, we found that the response may not coherent with the query. We will design more smooth expansion ways to construct more human-like training samples for our future work." + ], + "bbox": [ + 136, + 131, + 485, + 850 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ethics Statement", + "text_level": 1, + "bbox": [ + 114, + 862, + 265, + 876 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We acknowledge and ensure that our study is compatible with the provided Code of Ethics.", + "bbox": [ + 112, + 887, + 487, + 917 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Knowledge-grounded open-domain dialogue generation is crucial for building a knowledgeable dialogue system, which is beyond the wildest dreams in natural language process field. All our experiments are conducted on public available datasets to avoid ethical concerns. All terms for using these datasets are strictly followed in our study. There are no direct ethical concerns in our research.", + "bbox": [ + 507, + 84, + 882, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 509, + 224, + 672, + 240 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We would like to thank the anonymous reviewers for their constructive comments. This research is supported by Beijing Natural Science Foundation (No.4222037 and L181010) and BIT Research and Innovation Promoting Project (Grant No.2022YCXY021). Kan Li is the corresponding author.", + "bbox": [ + 507, + 248, + 882, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 510, + 388, + 608, + 401 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Leonard Adolphs, Kurt Shuster, Jack Urbanek, Arthur Szlam, and Jason Weston. 2021. Reason first, then respond: Modular generation for knowledge-infused dialogue. CoRR, abs/2111.05204.", + "Siqi Bao, Huang He, Fan Wang, Hua Wu, and Haifeng Wang. 2020. PLATO: pre-trained dialogue generation model with discrete latent variable. In ACL, pages 85-96. ACL.", + "Siqi Bao, Huang He, Jun Xu, Hua Lu, Fan Wang, Hua Wu, Han Zhou, Wenquan Wu, Zheng-Yu Niu, and Haifeng Wang. 2022. PLATO-K: internal and external knowledge enhanced dialogue generation. CoRR, abs/2211.00910.", + "Hengyi Cai, Hongshen Chen, Yonghao Song, Zhuoye Ding, Yongjun Bao, Weipeng Yan, and Xiaofang Zhao. 2020. Group-wise contrastive learning for neural dialogue generation. In EMNLP, pages 793-802. Association for Computational Linguistics.", + "Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. 2020a. A simple framework for contrastive learning of visual representations. In ICML, volume 119 of Proceedings of Machine Learning Research, pages 1597-1607. PMLR.", + "Xinlei Chen, Haoqi Fan, Ross B. Girshick, and Kaiming He. 2020b. Improved baselines with momentum contrastive learning. CoRR, abs/2003.04297.", + "Richard Csaky, Patrik Purgai, and Gábor Recski. 2019. Improving neural conversational models with entropy-based data filtering. In ACL (1), pages 5650-5669.", + "Emily Dinan, Stephen Roller, Kurt Shuster, Angela Fan, Michael Auli, and Jason Weston. 2019. Wizard of wikipedia: Knowledge-powered conversational agents. In ICLR. OpenReview.net." + ], + "bbox": [ + 510, + 409, + 884, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1746", + "bbox": [ + 482, + 928, + 521, + 940 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Nouha Dziri, Sivan Milton, Mo Yu, Osmar R. Zaïane, and Siva Reddy. 2022. On the origin of hallucinations in conversational models: Is it the datasets or the models? In *NAACL*, pages 5271-5285. Association for Computational Linguistics.", + "Karthik Gopalakrishnan, Behnam Hedayatnia, Qinglang Chen, Anna Gottardi, Sanjeev Kwatra, Anu Venkatesh, Raefer Gabriel, and Dilek Hakkani-Tur. 2019. Topical-chat: Towards knowledge-grounded open-domain conversations. In Interspeech, pages 1891-1895. ISCA.", + "Behnam Hedayatnia, Karthik Gopalakrishnan, Seokhwan Kim, Yang Liu, Mihail Eric, and Dilek Hakkani-Tur. 2020. Policy-driven neural response generation for knowledge-grounded dialog systems. In INLG, pages 412-421. Association for Computational Linguistics.", + "Byeongchang Kim, Jaewoo Ahn, and Gunhee Kim. 2020. Sequential latent knowledge selection for knowledge-grounded dialogue. In ICLR. OpenReview.net.", + "Mojtaba Komeili, Kurt Shuster, and Jason Weston. 2022. Internet-augmented dialogue generation. In ACL, pages 8460-8478. Association for Computational Linguistics.", + "Jiwei Li, Michel Galley, Chris Brockett, Jianfeng Gao, and Bill Dolan. 2016. A diversity-promoting objective function for neural conversation models. In HLT-NAACL, pages 110-119.", + "Rongzhong Lian, Min Xie, Fan Wang, Jinhua Peng, and Hua Wu. 2019. Learning to select knowledge for response generation in dialog systems. In *IJCAI*, pages 5081-5087. ijcai.org.", + "Chia-Wei Liu, Ryan Lowe, Iulian Serban, Michael Noseworthy, Laurent Charlin, and Joelle Pineau. 2016. How NOT to evaluate your dialogue system: An empirical study of unsupervised evaluation metrics for dialogue response generation. In EMNLP, pages 2122-2132.", + "Zihan Liu, Mostofa Patwary, Ryan Prenger, Shrimai Prabhumoye, Wei Ping, Mohammad Shoeybi, and Bryan Catanzaro. 2022. Multi-stage prompting for knowledgeable dialogue generation. In *Findings of ACL*, pages 1317-1337. Association for Computational Linguistics.", + "George A. Miller. 1995. Wordnet: A lexical database for english. Commun. ACM, 38(11):39-41.", + "Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In ACL, pages 311-318.", + "Hannah Rashkin, David Reitter, Gaurav Singh Tomar, and Dipanjan Das. 2021. Increasing faithfulness in knowledge-grounded dialogue with controllable features. In ACL/IJCNLP, pages 704-718. Association for Computational Linguistics." + ], + "bbox": [ + 115, + 85, + 485, + 917 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Nils Reimers and Iryna Gurevych. 2021. The curse of dense low-dimensional information retrieval for large index sizes. In ACL/IJCNLP, pages 605-611. Association for Computational Linguistics.", + "Kurt Shuster, Spencer Poff, Moya Chen, Douwe Kiela, and Jason Weston. 2021. Retrieval augmentation reduces hallucination in conversation. In *Findings of EMNLP*, pages 3784–3803. Association for Computational Linguistics.", + "Bin Sun, Shaoxiong Feng, Yiwei Li, Jiamou Liu, and Kan Li. 2021. Generating relevant and coherent dialogue responses using self-separated conditional variational autoencoders. In ACL/IJCNLP, pages 5624-5637. Association for Computational Linguistics.", + "Weiwei Sun, Zhengliang Shi, Shen Gao, Pengjie Ren, Maarten de Rijke, and Zhaochun Ren. 2022. Contrastive learning reduces hallucination in conversations. CoRR, abs/2212.10400.", + "Josef Valvoda, Yimai Fang, and David Vandyke. 2022. Prompting for a conversation: How to control a dialog model? In Proceedings of the Second Workshop on When Creative AI Meets Conversational AI, pages 1-8, Gyeongju, Republic of Korea. Association for Computational Linguistics.", + "Wenquan Wu, Zhen Guo, Xiangyang Zhou, Hua Wu, Xiyuan Zhang, Rongzhong Lian, and Haifeng Wang. 2019. Proactive human-machine conversation with explicit conversation goal. In ACL, pages 3794-3804. Association for Computational Linguistics.", + "Xinnuo Xu, Ondrej Dusek, Ioannis Konstas, and Verena Rieser. 2018. Better conversations by modeling, filtering, and optimizing for coherence and diversity. In EMNLP, pages 3981-3991.", + "Yuan Yao, Bowen Dong, Ao Zhang, Zhengyan Zhang, Ruobing Xie, Zhiyuan Liu, Leyu Lin, Maosong Sun, and Jianyong Wang. 2022. Prompt tuning for discriminative pre-trained language models. In *Findings of the Association for Computational Linguistics: ACL* 2022, pages 3468-3473, Dublin, Ireland. Association for Computational Linguistics.", + "Xueliang Zhao, Tingchen Fu, Chongyang Tao, and Rui Yan. 2022a. There is no standard answer: Knowledge-grounded dialogue generation with adversarial activated multi-reference learning. CoRR, abs/2210.12459.", + "Xueliang Zhao, Wei Wu, Chongyang Tao, Can Xu, Dongyan Zhao, and Rui Yan. 2020a. Low-resource knowledge-grounded dialogue generation. In ICLR. OpenReview.net.", + "Xueliang Zhao, Wei Wu, Can Xu, Chongyang Tao, Dongyan Zhao, and Rui Yan. 2020b. Knowledge-grounded dialogue generation with pre-trained language models. In EMNLP, pages 3377-3390. Association for Computational Linguistics." + ], + "bbox": [ + 510, + 85, + 882, + 917 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1747", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Yingxiu Zhao, Yinhe Zheng, Zhiliang Tian, Chang Gao, Bowen Yu, Haiyang Yu, Yongbin Li, Jian Sun, and Nevin L. Zhang. 2022b. Prompt conditioned VAE: enhancing generative replay for lifelong learning in task-oriented dialogue. CoRR, abs/2210.07783.", + "bbox": [ + 115, + 85, + 487, + 151 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Wen Zheng, Natasa Milic-Frayling, and Ke Zhou. 2021. Knowledge-grounded dialogue generation with term-level de-noising. In Findings of ACL/IJCNLP, volume ACL/IJCNLP 2021 of Findings of ACL, pages 2972-2983. Association for Computational Linguistics.", + "bbox": [ + 115, + 162, + 489, + 240 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Kangyan Zhou, Shrimai Prabhumoye, and Alan W. Black. 2018. A dataset for document grounded conversations. In EMNLP, pages 708-713. Association for Computational Linguistics.", + "bbox": [ + 115, + 253, + 489, + 307 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b08d270a0120c4730f3e02269fa6eb542e02689cca965b90e7cb474a6f182ffa.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Prefix PromptsPost Prompts
I was thinking that perhaps I am not sure, maybe that Not very clear, maybe Not very clear, perhaps I was thinking that maybeMaybe i am wrong. If I am wrong, please correct me. If I am wrong, please for-give me. If it is wrong, please tell me.
", + "bbox": [ + 115, + 332, + 485, + 432 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e84dbfc0526ffd29c63be0d7023df44e5b0fb52d5743ea9dd3c01d1e8c378a7d.jpg", + "table_caption": [ + "Table 7: The designed prefix and post prompts." + ], + "table_footnote": [], + "table_body": "
Euphemistic Responses
Interesting, do you know that? \nThat sounds pretty good. Are there any way to visit? \nOh, I had not heard. \nHmm, I have never heard of that. What is that one about? \nI have never heard. Can you tell me more about it? \nOh, wow, that is remarkable. \nI have never played those, are they fun? \nCan I ask you about it? \nPlease tell me more about that. \nCan you tell me more about that? \nI have never had that. Anything else you can tell me? \nThat's really interesting! But I have never heard of that. \nI literally know nothing about that! \nI have no idea about that. \nI have not heard that one. I will have to check it out. \nHuh, maybe I will need to check that out then. \nOh, I misunderstood then. \nOh, i do not know about that. \nWow, that's a lot! I haven't heard of those.
", + "bbox": [ + 115, + 487, + 485, + 740 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8: The designed euphemistic responses.", + "bbox": [ + 142, + 750, + 455, + 765 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A Prefix and Post Prompts", + "text_level": 1, + "bbox": [ + 114, + 795, + 363, + 812 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We manually design five prefix prompts and four post prompts, which are shown in Table 7. We discuss below about the prefixes and posts.", + "bbox": [ + 112, + 822, + 487, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We designed the prefixes and posts based on the WoW dataset and our daily conversation habits. In WoW dataset, one role is \"0_Wizard\", and the other", + "bbox": [ + 112, + 871, + 487, + 917 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "is \"1_Aprentice\". We noticed that the 1_Aprentice will give the sentences such as \"correct my if I am wrong ...\", which is also easy to appear in our daily conversation. Taking inspiration of this, we manually designed the prefixes and posts. Moreover, since the PLATO is pre-trained on conversation datasets, these prefixes may introduce the pre-knowledge that the model learned during the pre-training process.", + "bbox": [ + 507, + 84, + 884, + 228 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In fact, we declare the weakness of our manual prefixes and posts, i.e. direct connections of prefixes, responses, and posts do not fit all contexts. Therefore, we are exploring a new way of constructing replies, such as passing the design prefix, response, post, and context into the large-language-model to rewrite the appropriate response. We believe that better prefixes and posts will lead to more benefits in solving the hallucination problem.", + "bbox": [ + 507, + 229, + 885, + 373 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "B Euphemistic Responses", + "text_level": 1, + "bbox": [ + 507, + 385, + 749, + 401 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We manually design nineteen euphemistic responses, which are shown in Table 8.", + "bbox": [ + 507, + 411, + 882, + 442 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "C Dissuasion about the boundary between ak-less and ak-more", + "text_level": 1, + "bbox": [ + 507, + 454, + 815, + 486 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Below we provide an example in our dataset:", + "bbox": [ + 509, + 497, + 843, + 512 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Ground-truth Knowledge: laziness | thesis (\"thesis\") is a 1996 spanish thriller film.", + "- AK-Less Knowledge: acedia | thesis (\"thesis\") is a 1996 spanish thriller film.", + "- AK_More Knowledge: laziness | thesis (\"thesis\") personate a 1996 spanish thriller picture show." + ], + "bbox": [ + 531, + 523, + 882, + 655 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "It can be noted that the more synonyms are introduced into a sentence, the semantics of the sentence will become more and more different from the original semantics. Therefore, we suppose that replacing at least $30\\%$ of words at once will make a big difference in sentence semantics. Then, we decided the boundary between ak-less and ak-more.", + "bbox": [ + 507, + 668, + 884, + 780 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1748", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A For every submission:", + "bbox": [ + 115, + 107, + 322, + 122 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "A1. Did you describe the limitations of your work?", + "bbox": [ + 129, + 126, + 532, + 143 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We provide a section of Limitations after the Conclusion and before the Ethics Statement", + "bbox": [ + 149, + 143, + 805, + 159 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "A2. Did you discuss any potential risks of your work?", + "bbox": [ + 129, + 170, + 552, + 186 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Not applicable. Left blank.", + "bbox": [ + 149, + 187, + 349, + 200 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "A3. Do the abstract and introduction summarize the paper's main claims?", + "bbox": [ + 129, + 212, + 695, + 228 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "1", + "bbox": [ + 152, + 230, + 163, + 242 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "A4. Have you used AI writing assistants when working on this paper?", + "bbox": [ + 129, + 255, + 668, + 272 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 152, + 273, + 231, + 287 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "B Did you use or create scientific artifacts?", + "bbox": [ + 115, + 300, + 489, + 316 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 132, + 321, + 213, + 336 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "B1. Did you cite the creators of artifacts you used?", + "bbox": [ + 129, + 347, + 529, + 363 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Not applicable. Left blank.", + "bbox": [ + 149, + 363, + 349, + 379 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?", + "bbox": [ + 129, + 390, + 778, + 406 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Not applicable. Left blank.", + "bbox": [ + 151, + 407, + 349, + 422 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?", + "bbox": [ + 129, + 432, + 880, + 495 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Not applicable. Left blank.", + "bbox": [ + 151, + 498, + 349, + 513 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?", + "bbox": [ + 129, + 524, + 880, + 571 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We use a publicly well-established dataset.", + "bbox": [ + 151, + 573, + 467, + 588 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?", + "bbox": [ + 129, + 599, + 880, + 631 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Not applicable. Left blank.", + "bbox": [ + 151, + 632, + 349, + 646 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be.", + "bbox": [ + 129, + 658, + 880, + 739 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 152, + 740, + 231, + 753 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "C Did you run computational experiments?", + "bbox": [ + 115, + 764, + 492, + 781 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3", + "bbox": [ + 134, + 787, + 146, + 799 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?", + "bbox": [ + 129, + 813, + 880, + 845 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We use the released code and checkpoints. We cite the source of our model.", + "bbox": [ + 149, + 846, + 704, + 860 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.", + "bbox": [ + 112, + 868, + 877, + 892 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "ACL 2023 Responsible NLP Checklist", + "bbox": [ + 132, + 84, + 433, + 99 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1749", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Not applicable. Left blank.", + "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Not applicable. Left blank.", + "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? 3" + ], + "bbox": [ + 127, + 84, + 880, + 280 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? 3", + "text_level": 1, + "bbox": [ + 114, + 292, + 877, + 326 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Not applicable. Left blank.", + "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Not applicable. Left blank.", + "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? 2", + "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Not applicable. Left blank.", + "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Not applicable. Left blank." + ], + "bbox": [ + 127, + 341, + 880, + 640 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1750", + "bbox": [ + 482, + 928, + 521, + 940 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Towards Fewer Hallucinations in Knowledge-Grounded Dialogue Generation via Augmentative and Contrastive Knowledge-Dialogue/989c66cd-1d3d-4fda-ba4f-e2869fc255a6_model.json b/2023/Towards Fewer Hallucinations in Knowledge-Grounded Dialogue Generation via Augmentative and Contrastive Knowledge-Dialogue/989c66cd-1d3d-4fda-ba4f-e2869fc255a6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0df534c11e04f266a131234fce02c4d47e28ec75 --- /dev/null +++ b/2023/Towards Fewer Hallucinations in Knowledge-Grounded Dialogue Generation via Augmentative and Contrastive Knowledge-Dialogue/989c66cd-1d3d-4fda-ba4f-e2869fc255a6_model.json @@ -0,0 +1,2068 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.15, + 0.089, + 0.849, + 0.13 + ], + "angle": 0, + "content": "Towards Fewer Hallucinations in Knowledge-Grounded Dialogue Generation via Augmentative and Contrastive Knowledge-Dialogue" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.142, + 0.777, + 0.159 + ], + "angle": 0, + "content": "Bin Sun\\(^{1}\\), Yitong Li\\(^{2,3}\\), Fei Mi\\(^{2}\\), FanHu Bie\\(^{3}\\), Yiwei Li\\(^{1}\\), Kan Li\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.192, + 0.16, + 0.815, + 0.177 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Computer Science & Technology, Beijing Institute of Technology" + }, + { + "type": "text", + "bbox": [ + 0.273, + 0.177, + 0.732, + 0.193 + ], + "angle": 0, + "content": "\\(^{2}\\) Huawei Noah's Ark Lab \\(^{3}\\)Huawei Technologies Ltd." + }, + { + "type": "text", + "bbox": [ + 0.375, + 0.196, + 0.628, + 0.209 + ], + "angle": 0, + "content": "{binsun,liyiwei,likan}@bit.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.357, + 0.212, + 0.647, + 0.226 + ], + "angle": 0, + "content": "{liyitong3,mifei2,biefanhu}@huawei.com" + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.253, + 0.341, + 0.268 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.278, + 0.461, + 0.676 + ], + "angle": 0, + "content": "Existing knowledge-grounded open-domain dialogue generation models often face the hallucination problem, i.e. the dialogue generative model will persist in an inappropriate knowledge and generate responses that inconsistent with the facts. We argue that this problem mainly stems from the polarized optimization objectives and weak knowledge generation ability. To mitigate the hallucination, we take inspiration from human communicating that people will replay euphemistic responses for the unclear or unrecognizable knowledge, and propose an Augmentative and Contrastive Knowledge Dialogue Expansion Framework (ACK-DEF). ACK-DEF constructs the augmentative and contrastive knowledge dialogue samples, which consist of the knowledge of different degrees of errors and the response of manual design, to expand the original training set and smooth the polarized optimization objective that enables models to generate ground-truth with or without gold knowledge. Not only the knowledge, ACK-DEF also provides the tactful responses of manual design corresponding to the incomplete correct knowledge. Experimental results on the Wikipedia of Wizard dataset show that employing the ACK-DEF is effective to alleviate the hallucination problem." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.685, + 0.26, + 0.7 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.71, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Recently, Knowledge-Grounded Dialogue Generation draws dramatic attentions in artificial intelligence community. Many efforts incorporate knowledge information to improve the performance of dialogue generation models (Zhou et al., 2018; Dinan et al., 2019; Gopalakrishnan et al., 2019; Kim et al., 2020; Zhao et al., 2020a; Zheng et al., 2021; Zhao et al., 2022a; Bao et al., 2022). However, these methods always face the hallucination problem, that is, the dialogue generation model may insist on an inappropriate knowledge and generate responses that inconsistent with the facts (Rashkin et al., 2021; Zhao et al., 2022a; Dziri et al., 2022)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.254, + 0.885, + 0.623 + ], + "angle": 0, + "content": "We argue that the hallucination problem primarily caused by two aspects: (1) The optimization objective is usually polarized by the gold knowledge dialogue samples and general dialogue samples without knowledge in current knowledge-grounded dialogue datasets (Zhou et al., 2018; Gopalakrishnan et al., 2019; Dinan et al., 2019; Wu et al., 2019; Komeili et al., 2022). Few datasets consider teaching models how to respond when dealing with incomplete correct knowledge, which makes the models tend to believe in the given knowledge, regardless of whether the knowledge is appropriate or not, resulting in hallucination problems. In addition, the knowledge retrieval system tends to extract irrelevant knowledge rather than relevant knowledge when the database is large, aggravating the hallucinations (Reimers and Gurevych, 2021; Liu et al., 2022). (2) The generation of knowledge may also face the hallucination problem and obtain the inappropriate knowledge, leading the generation of hallucination responses (Kim et al., 2020; Zhao et al., 2020a; Liu et al., 2022; Adolphs et al., 2021; Bao et al., 2022)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.63, + 0.885, + 0.919 + ], + "angle": 0, + "content": "To mitigate the hallucination problem, we propose an Augmentative and Contrastive Knowledge Dialogue Expansion Framework (ACK-DEF), which is inspired by human communicating that people will replay euphemistic response for the unrecognizable knowledge. ACK-DEF is proposed to smooth the polarized optimization objective by augmenting training set with augmentative and contrastive knowledge-dialogue samples. Not only the knowledge, we also designed the reply patterns for the knowledge with different level of errors. For this, we propose the augmentative knowledge dialogue expansion (AK), and contrastive knowledge dialogue expansion (CK). AK is proposed to boost the generalization ability of models on knowledge with minor noise. On the contrary, inspired from the contrastive learning paradigm (Cai et al., 2020; Chen et al., 2020a,b; Sun et al., 2021, 2022), CK" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.928, + 0.52, + 0.941 + ], + "angle": 0, + "content": "1741" + }, + { + "type": "footer", + "bbox": [ + 0.227, + 0.946, + 0.772, + 0.958 + ], + "angle": 0, + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + }, + { + "type": "footer", + "bbox": [ + 0.369, + 0.959, + 0.631, + 0.972 + ], + "angle": 0, + "content": "Volume 2: Short Papers, pages 1741-1750" + }, + { + "type": "footer", + "bbox": [ + 0.297, + 0.973, + 0.702, + 0.986 + ], + "angle": 0, + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.131, + 0.082, + 0.475, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.274, + 0.49, + 0.362 + ], + "angle": 0, + "content": "Figure 1: A diagram of our Augmentative Knowledge dialogue expansion method. We replace different proportion of words in the original knowledge with synonyms to construct incomplete correct knowledge, and design response for different knowledge. We also use prompts to guide the dialogue generation process." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.385, + 0.489, + 0.464 + ], + "angle": 0, + "content": "reconstructs incorrect knowledge and designs euphemistic responses, which aims to push the model learn the reply pattern of incorrect knowledge and a better boundary between correct and incorrect knowledge." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.466, + 0.489, + 0.595 + ], + "angle": 0, + "content": "Contributions: We propose an ACK-DEF to construct new knowledge-dialogue samples that consist of knowledge with different level of errors and manual responses, to soften the training optimization objectives of models, which will mitigate the hallucination. Finally, we conduct extension experiments to show the superior performance of ACK-DEF on alleviating the hallucination." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.606, + 0.265, + 0.623 + ], + "angle": 0, + "content": "2 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.631, + 0.49, + 0.839 + ], + "angle": 0, + "content": "To mitigate the hallucination problem that caused by the polarized optimization objectives in knowledge grounded dialogue generation, we take inspiration from human communicating, and propose the Augmentative and Contrastive Knowledge Dialogue Expansion Framework (ACK-DEF). Our ACK-DEF aims to soften the polarized training optimization objectives of current knowledge-grounded dialogue generation methods, and guide the dialogue system reply patterns for the knowledge with different level of errors. To achieve this end, we design two effective expansion method, which will be detailed in below." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.851, + 0.438, + 0.867 + ], + "angle": 0, + "content": "2.1 Augmentative Knowledge Dialogue" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.872, + 0.49, + 0.919 + ], + "angle": 0, + "content": "We propose the Augmentative Knowledge (AK) dialogue expansion to boost the generalization ability of the dialogue model on the knowledge with simi" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.082, + 0.872, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.211, + 0.885, + 0.34 + ], + "angle": 0, + "content": "Figure 2: A diagram of our Contrastive Knowledge dialogue expansion method. We use the antonym to reconstruct the knowledge information and design multiple responses for such knowledge. Since antonyms transform the semantics of the original knowledge, the noise knowledge often contains wrong facts. By this, the model can learn a better boundary between correct and incorrect knowledge, and a safety reply pattern for incorrect knowledge." + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.368, + 0.885, + 0.787 + ], + "angle": 0, + "content": "lar semantics but different expressions, which can prevent the model from being interfered by partialrelevant knowledge retrieved by the retrieval systems (Lian et al., 2019; Zhao et al., 2020b; Hedayatnia et al., 2020; Zheng et al., 2021; Shuster et al., 2021; Komeili et al., 2022). As shown in Figure 1, we employ the synonym data augmentation tool, which replaces words in the original knowledge with synonyms, to reconstruct the knowledge information (Miller, 1995). Considering that the synonym may disrupt the original semantics of new constructed knowledge, we constrain the replace possibility within [0.1,0.2]. Hence, we can obtain the approximate knowledge. Combining this knowledge and the original dialogue, we obtain the \"ak-less sample\". In addition, we also replace \\(30\\%\\) to \\(50\\%\\) words with their synonyms to construct the less similar knowledge. Inspired from prompt learning paradigm (Yao et al., 2022; Valvoda et al., 2022; Zhao et al., 2022b), we manually produce some Prefix-prompts and Post-prompts (see Appendix) to (1) make the new response more tactful for the less similar knowledge; (2) regulate and guide the dialogue generation process of the model. We call the sample consist of less-similar knowledge and designed response as \"ak-more sample\"." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.801, + 0.816, + 0.816 + ], + "angle": 0, + "content": "2.2 Contrastive Knowledge Dialogue" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.823, + 0.885, + 0.92 + ], + "angle": 0, + "content": "We propose the Contrastive Knowledge (CK) dialogue expansion, inspired from the contrastive learning paradigm (Chen et al., 2020b; Cai et al., 2020), not only construct the incorrect knowledge as negative samples for original knowledge, but also build the euphemistic responses as positive" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.928, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1742" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.112, + 0.085, + 0.493, + 0.295 + ], + "angle": 0, + "content": "samples for the original response with incorrect knowledge. To help the model learn a boundary between correct and incorrect knowledge, we employ the antonym to make up new incorrect knowledge. For example, given the knowledge \"nintendo was founded on 23 september 1889 ...\", the \"founded\" will be replaced with \"abolish\", which greatly changes the semantics but little changes the expression. After that, we random choose an euphemistic response to replace the original response of the dialogue. Finally, The incorrect knowledge and the replaced euphemistic response are combined as the \"ck-sample\"." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.313, + 0.358, + 0.329 + ], + "angle": 0, + "content": "3 Experiment and Results" + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.343, + 0.321, + 0.358 + ], + "angle": 0, + "content": "3.1 Experiment Settings" + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.368, + 0.235, + 0.382 + ], + "angle": 0, + "content": "3.1.1 Dataset" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.391, + 0.49, + 0.617 + ], + "angle": 0, + "content": "We use the Wikipedia of Wizard (WoW) data, a well-established knowledge-grounded open-domain dialogue dataset, for our experiment. We pre-precess the WoW dataset and extract the single-turn knowledge dialogue samples. To evaluate the performance of our method in detail, we perform four test sets: normal, ak-less, ak-more and ck. The normal set is the original test set. And the ak-less, ak-more and ck are the sets consist of ak-less, ak-more and ck samples, respectively. We also follow the settings of WoW data and divide the test set into two groups (seen test and unseen test): the topic of the knowledge in the unseen test set is missing in the training set." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.632, + 0.242, + 0.646 + ], + "angle": 0, + "content": "3.1.2 Baseline" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.655, + 0.489, + 0.703 + ], + "angle": 0, + "content": "We employ the released PLATO-v1 (Bao et al., 2020) model, a pre-trained dialogue generation model based on UniLM, for our experiment." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.72, + 0.49, + 0.799 + ], + "angle": 0, + "content": "Fine-tuning We directly finetune a model on the original WoW training set. By this, the model can only see gold knowledge dialogue samples and general dialogue samples without knowledge. Hence, we call the fine-tuned model PLATO+GOLD." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.816, + 0.49, + 0.879 + ], + "angle": 0, + "content": "Fine-tuning with ACK-DEF We finetune the model with the original set and the expansion samples that obtained through ACK-DEF. Thence, we call it PLATO+ACK-DEF." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.085, + 0.759, + 0.099 + ], + "angle": 0, + "content": "3.1.3 AutoEvaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.104, + 0.885, + 0.41 + ], + "angle": 0, + "content": "Dialogue Metrics Our primary metrics of interest are Distinct-n (Li et al., 2016), Response Length (Len.) (Csaky et al., 2019), BLEU (Papineni et al., 2002), Embedding-based (Greedy (GRE), Average (AVG), Extrema (EXT)) (Liu et al., 2016), and Coherence (COH) (Xu et al., 2018). Distinct-n evaluates the diversity of generated responses, which is calculated through the ratio of distinct \\( n \\)-grams and all generated \\( n \\)-grams. Len. is the average number of words of all generated responses. BLEU validates the degree of the word-overlap between the generated response and the ground-truth, which denotes the consistence between generated response and ground-truth. Embedding-based metrics (GRE, AVG and EXT) are introduced to evaluate the semantic relationship of generated responses and ground-truth responses, illustrating the consistence in semantic level. COH. mainly assesses the relevance between contexts and generated responses." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.42, + 0.885, + 0.708 + ], + "angle": 0, + "content": "Knowledge Metrics We follow the PLATO(Bao et al., 2020) and use the knowledge precision, recall and f1 scores. These metrics are used to calculate the ratio of tokens that exist in common in ground-truth knowledge and generated responses to tokens in generated responses. \"Recall\" is the average ratio of the number of overlapping tokens in response and knowledge to the number of tokens in knowledge. And \"Precision\" is the average ratio of the number of overlapping tokens to the number of tokens in response. In other words, \"Recall\" indicates how much knowledge information is contained in the response, while \"Precision\" indicates the proportion of knowledge information in the response. Even we involve the negative and incorrect knowledge in response generation, we still use the ground-truth knowledge to calculate the metrics in Table 3,4." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.722, + 0.803, + 0.737 + ], + "angle": 0, + "content": "3.2 Dialogue Performance Analysis" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.743, + 0.885, + 0.919 + ], + "angle": 0, + "content": "Table 1 and Table 2 report the automatic results on four test sets and four unseen test sets, respectively. In these Tables, it can be observed that (1) the PLATO+ACK-DEF has a competitive performance with PLATO+GOLD on the normal set, which means that the PLATO+ACK-DEF can recognize the golden knowledge and produce appropriate responses. (2) the PLATO+GOLD perform worse than PLATO+ACK-DEF on ak-less, which means that the robustness of the dialogue model only trained with golden knowledge is very weak." + }, + { + "type": "page_footnote", + "bbox": [ + 0.113, + 0.892, + 0.49, + 0.919 + ], + "angle": 0, + "content": "1We manually construct some responses, please see Appendix for the detail." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1743" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.127, + 0.083, + 0.871, + 0.226 + ], + "angle": 0, + "content": "
test setDistinct-1/2/3Len.BLEU-1/2/3/4GREAVGEXTCOH
normal0.10680.453313.690.42800.29650.21100.15290.73920.86890.63610.7808
0.09020.398416.200.44280.30170.21090.14990.73660.86830.63300.7878
ak-less0.11940.502413.500.38610.25740.17450.11920.71600.86070.61480.7755
0.08230.353218.780.45020.29820.20150.13800.73070.86960.62930.7948
ak-more0.12340.517412.810.16750.10620.06800.04350.69080.85510.59940.7706
0.06750.294621.830.43580.30010.21230.15420.77450.91510.70930.8098
ck0.11090.477913.230.29650.17790.10800.06570.58380.76220.53730.7712
0.06520.202913.360.42300.27050.18090.12660.65720.83060.61620.8049
" + }, + { + "type": "table_caption", + "bbox": [ + 0.13, + 0.236, + 0.865, + 0.251 + ], + "angle": 0, + "content": "Table 1: The automatic results of PLATO+GOLD (up) and PLATO+ACK-DEF (down) on four test seen sets." + }, + { + "type": "table", + "bbox": [ + 0.127, + 0.265, + 0.871, + 0.408 + ], + "angle": 0, + "content": "
test setDistinct-1/2Len.BLUE-1/2/3/4GREAVGEXTCOH
normal0.05030.242212.430.35160.23310.15820.10900.69880.85680.63060.8094
0.04670.231113.140.34630.22810.15360.10490.69680.85410.63380.8105
ak-less0.09660.391713.390.38710.25650.17240.11640.71430.86000.61220.7836
0.06230.266419.180.44430.29070.19360.13010.72320.86630.61940.8026
ak-more0.10640.444012.710.16520.10460.06680.04260.68880.85380.59800.7797
0.05610.240021.820.43310.29680.20910.15110.76970.91140.70370.8197
ck0.08130.332413.240.30110.18090.11000.06690.58540.76760.54790.7794
0.04650.149013.520.43290.27750.18610.13070.66120.83340.62150.8145
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.419, + 0.882, + 0.447 + ], + "angle": 0, + "content": "Table 2: The automatic results of PLATO+GOLD (up) and PLATO+ACK-DEF (down) on four test sets with unseen knowledge." + }, + { + "type": "table", + "bbox": [ + 0.124, + 0.469, + 0.479, + 0.601 + ], + "angle": 0, + "content": "
test setRecallPrecisionF1avg. Dec.
normal0.36070.70090.4546-
ak-less0.28830.55850.3618∇ 0.1026
ak-more0.17520.36320.2228∇ 0.2517
ck0.31930.61330.4003∇ 0.0611
normal0.36950.65380.4520-
ak-less0.32510.56360.3927∇ 0.0647
ak-more0.23350.39830.2775∇ 0.1887
ck0.10650.20410.1337∇ 0.3437
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.611, + 0.488, + 0.655 + ], + "angle": 0, + "content": "Table 3: The knowledge correlation results of PLATO+GOLD (up) and PLATO+ACK-DEF (down) on four test sets with seen knowledge." + }, + { + "type": "table", + "bbox": [ + 0.125, + 0.672, + 0.479, + 0.803 + ], + "angle": 0, + "content": "
test setRecallPrecisionF1avg. Dec.
normal0.37320.74420.4736-
ak-less0.27280.54750.3452∇ 0.1418
ak-more0.16650.36270.2152∇ 0.2822
ck0.30280.60680.3830∇ 0.0995
normal0.36550.68820.4535-
ak-less0.29380.53480.3579∇ 0.1069
ak-more0.20460.37140.2481∇ 0.2277
ck0.08700.18470.1116∇ 0.3747
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.813, + 0.489, + 0.856 + ], + "angle": 0, + "content": "Table 4: The knowledge correlation results of PLATO+GOLD (up) and PLATO+ACK-DEF (down) on four test sets with unseen knowledge." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.888, + 0.489, + 0.919 + ], + "angle": 0, + "content": "Even if the knowledge information only changes by \\(10\\%\\) to \\(20\\%\\), the performance of the model will" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.473, + 0.884, + 0.584 + ], + "angle": 0, + "content": "significantly decline, especially consistency metrics (i.e. BLEU, GRE, AVG and EXT). (3) the PLATO+GOLD achieve better Distinct scores but weaker BLEU and embedding-based scores, which means that the PLATO+GOLD is easy to generate responses that are very different from ground-truth responses, that is, the hallucinations." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.605, + 0.811, + 0.62 + ], + "angle": 0, + "content": "3.3 Knowledge Correlation Analysis" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.63, + 0.885, + 0.919 + ], + "angle": 0, + "content": "Table 3 and Table 4 report the knowledge correlation result of PLATO+GOLD and PLATO+ACK-DEF on four test sets and four test unseen sets, respectively. From these table, we can observe that the performance of PLATO+GOLD is reduced when the given knowledge changed, which illustrates the danger that the model generate responses based on incorrect knowledge. In addition to the above findings, we also observed that the recall, precision and f1 scores of PLATO+ACK-DEF are better than PLATO+GOLD on ak-less and ak-more sets, which demonstrates that using ACK-DEF effectively enhance the model's capability for the similar knowledge information. Moreover, the result of PLATO+ACK-DEF on the ck set is significantly reduced, which shows that the model distinguishes the wrong knowledge constructed with antonyms and gives an appropriate response with" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1744" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.117, + 0.082, + 0.486, + 0.178 + ], + "angle": 0, + "content": "
test setw. GOLD (%)w. ACK-DEF (%)kappa
normal13.0014.000.481
ak-less23.6717.330.513
ak-more33.6724.330.479
ck21.675.670.597
total23.0015.330.552
" + }, + { + "type": "table_caption", + "bbox": [ + 0.168, + 0.188, + 0.432, + 0.201 + ], + "angle": 0, + "content": "Table 5: The human evaluation results." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.227, + 0.489, + 0.323 + ], + "angle": 0, + "content": "out knowledge (see Table 1 and Table 2 for the effect). These results are inline with our exception that incorporating noised knowledge dialogue samples in training stages can smooth the polarized optimization objective, and mitigate the hallucination problem." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.325, + 0.49, + 0.614 + ], + "angle": 0, + "content": "According to the results of test seen sets and unseen sets), we can notice that the PLATO+ACK-DEF achieves a good performance on ground-truth seen knowledge and a weak performance on ground-truth unseen knowledge. This illustrates that the PLATO+ACK-DEF may doubt the authenticity of unseen given knowledge (even if the knowledge is the ground-truth), and will not fully use it to generate responses. This may alleviate the hallucination, and we believe it is caused by (1) the Augmentative knowledge dialogue introduces similar knowledge to improve the generalization of the model; (2) the Contrastive knowledge dialogue introduces knowledge independent responses, which tell the model to generate responses without knowledge; (3) the ACK-DEF smooths the polarized optimization, which ensures the model not to directly use the given knowledge." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.626, + 0.312, + 0.64 + ], + "angle": 0, + "content": "3.4 Human Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.646, + 0.49, + 0.806 + ], + "angle": 0, + "content": "To further evaluation the ability of our ACK-DEF on reducing the hallucination problem, we randomly select 400 samples form four test sets, and hire three annotators to do human evaluations by assessing whether the responses generated by PLATO +GOLD and +ACK-DEL have hallucinations. Table 5 reports the results of human evaluation, from which we can notice that the PLATO+ACK-DEF generate less hallucinations than PLATO+GOLD. This shows the effectiveness of our ACK-DEF." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.818, + 0.249, + 0.833 + ], + "angle": 0, + "content": "3.5 Case Study" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.839, + 0.489, + 0.919 + ], + "angle": 0, + "content": "Table 6 shows a generated case of PLATO+GOLD (GOLD) and PLATO+ACK-DEF (ACK-DEF) based on different knowledge information. We can observed that the GOLD is convinced of the given knowledge, regardless of whether the knowledge" + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.082, + 0.883, + 0.487 + ], + "angle": 0, + "content": "
QueryThey really are super cool looking dogs\nthey are so amazing
Normal\nKnowledgegolden retriever are well suited to resi-\ndency in suburban or country environ-\nments.
GOLDI love golden retrievers! They are well\nsuited to residency in suburban or country.
ACK-DEFThey are! They are well suited to suburban\nor country environments.
AK-Less\nKnowledgegolden retriever are good suited to resi-\ndency in suburban operating elbow room\nland environments.
GOLDGolden retriever are great suited to resi-\ndency in suburban operating elbow
ACK-DEFGolden retriever are great suited to subur-\nban environments.
AK-More\nKnowledgegild retriever make up considerably suit\nto residency in suburban surgery common-\nwealth environments.
GOLDgild retrievers make up considerably suit\nto residency in suburban areas.
ACK-DEFNot very clear, maybe they are! They can\nbe found in suburban areas. If it is wrong,\nplease tell me.
CK Knowl-\nedgegolden retriever unequal disadvantage-\ngeously suited to residency in suburban or\ncountry environments.
GOLDThey sure are! I have a golden retriever\nwho is unequally suited to residency in sub-\nurban environments.
ACK-DEFThat sounds pretty good.
" + }, + { + "type": "table_caption", + "bbox": [ + 0.515, + 0.496, + 0.877, + 0.51 + ], + "angle": 0, + "content": "Table 6: A case of PLATO +GOLD and +ACK-DEF." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.539, + 0.884, + 0.636 + ], + "angle": 0, + "content": "is appropriate or not, and more easily to copy the knowledge information into responses. Even the GOLD has seen the knowledge topic, it could not remember the knowledge in their parameters. On the contrary, the ACK-DEF has good resistance to incomplete correct knowledge." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.651, + 0.642, + 0.665 + ], + "angle": 0, + "content": "4 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.678, + 0.885, + 0.919 + ], + "angle": 0, + "content": "This paper focuses on the hallucinations caused by polarized optimization objective in knowledge-grounded dialogue generation (KGDG), and proposes an augmentative and contrastive knowledge dialogue expansion framework (ACK-DEF) to mitigate it. The optimization objective of KGDG is to train the model could generate proper response with or without knowledge, which inevitably weaken the model's ability on unrecognized knowledge and lead hallucinations. Therefore, ACK-DEF constructs multiple level knowledge-dialogue samples to soften the optimization objective of KGDG. Extension experimental results show the superior performance of using our methods on dialogue metrics and knowledge correlations." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1745" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.115, + 0.085, + 0.221, + 0.1 + ], + "angle": 0, + "content": "Limitations" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.109, + 0.336, + 0.124 + ], + "angle": 0, + "content": "Our limitations are as follow:" + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.133, + 0.486, + 0.211 + ], + "angle": 0, + "content": "- Data Scale: This paper only employs the Wikipedia of Wizard dataset, a small scale and well-established knowledge conversation dataset, and lack of the validation on large-scale dataset." + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.222, + 0.486, + 0.414 + ], + "angle": 0, + "content": "- **Backbones:** This paper lacks the evaluating of other knowledge dialogue model on the proposed method. Actually, we have two reasons to employ the PLATO. First, the PLATO can better handle the one-to-many phenomenon, which is suitable for learning our expansion samples. Second, the PLATO is a pre-trained dialogue model, and its performance on knowledge dialogue generation task has been proved. We will evaluating the performance of other knowledge dialogue model on our method for our future work." + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.424, + 0.486, + 0.616 + ], + "angle": 0, + "content": "- Knowledge Expansion Methods: This paper only uses the synonym and antonym to construct the noised knowledge, which lacks of the comparison of using other data augmentation method. Indeed, we use two token-level data augmentation methods (synonym and antonym augmentation) to prove our statements on hallucination problem in knowledge dialogue generation task. Based on this study, we believe that incorporating other data augmentation methods will also mitigate the hallucinations." + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.627, + 0.486, + 0.851 + ], + "angle": 0, + "content": "- Manual Prompts and Responses: This paper designed five prefix prompts, four post-prompts and nineteen euphemistic responses. For \\(AK\\)-More method, we simply randomly choose one prefix-prompt and one post-prompt and concatenate them with the ground-truth response. This leads to some irregular responses. As for \\(CK\\) method, we randomly select one euphemistic response for the incorrect knowledge. However, we found that the response may not coherent with the query. We will design more smooth expansion ways to construct more human-like training samples for our future work." + }, + { + "type": "list", + "bbox": [ + 0.137, + 0.133, + 0.486, + 0.851 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.863, + 0.266, + 0.877 + ], + "angle": 0, + "content": "Ethics Statement" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.888, + 0.489, + 0.919 + ], + "angle": 0, + "content": "We acknowledge and ensure that our study is compatible with the provided Code of Ethics." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.883, + 0.213 + ], + "angle": 0, + "content": "Knowledge-grounded open-domain dialogue generation is crucial for building a knowledgeable dialogue system, which is beyond the wildest dreams in natural language process field. All our experiments are conducted on public available datasets to avoid ethical concerns. All terms for using these datasets are strictly followed in our study. There are no direct ethical concerns in our research." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.225, + 0.673, + 0.241 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.249, + 0.883, + 0.36 + ], + "angle": 0, + "content": "We would like to thank the anonymous reviewers for their constructive comments. This research is supported by Beijing Natural Science Foundation (No.4222037 and L181010) and BIT Research and Innovation Promoting Project (Grant No.2022YCXY021). Kan Li is the corresponding author." + }, + { + "type": "title", + "bbox": [ + 0.511, + 0.389, + 0.61, + 0.403 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.41, + 0.882, + 0.464 + ], + "angle": 0, + "content": "Leonard Adolphs, Kurt Shuster, Jack Urbanek, Arthur Szlam, and Jason Weston. 2021. Reason first, then respond: Modular generation for knowledge-infused dialogue. CoRR, abs/2111.05204." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.471, + 0.884, + 0.525 + ], + "angle": 0, + "content": "Siqi Bao, Huang He, Fan Wang, Hua Wu, and Haifeng Wang. 2020. PLATO: pre-trained dialogue generation model with discrete latent variable. In ACL, pages 85-96. ACL." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.533, + 0.885, + 0.598 + ], + "angle": 0, + "content": "Siqi Bao, Huang He, Jun Xu, Hua Lu, Fan Wang, Hua Wu, Han Zhou, Wenquan Wu, Zheng-Yu Niu, and Haifeng Wang. 2022. PLATO-K: internal and external knowledge enhanced dialogue generation. CoRR, abs/2211.00910." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.607, + 0.884, + 0.674 + ], + "angle": 0, + "content": "Hengyi Cai, Hongshen Chen, Yonghao Song, Zhuoye Ding, Yongjun Bao, Weipeng Yan, and Xiaofang Zhao. 2020. Group-wise contrastive learning for neural dialogue generation. In EMNLP, pages 793-802. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.682, + 0.884, + 0.748 + ], + "angle": 0, + "content": "Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. 2020a. A simple framework for contrastive learning of visual representations. In ICML, volume 119 of Proceedings of Machine Learning Research, pages 1597-1607. PMLR." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.756, + 0.883, + 0.797 + ], + "angle": 0, + "content": "Xinlei Chen, Haoqi Fan, Ross B. Girshick, and Kaiming He. 2020b. Improved baselines with momentum contrastive learning. CoRR, abs/2003.04297." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.804, + 0.884, + 0.857 + ], + "angle": 0, + "content": "Richard Csaky, Patrik Purgai, and Gábor Recski. 2019. Improving neural conversational models with entropy-based data filtering. In ACL (1), pages 5650-5669." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.866, + 0.883, + 0.919 + ], + "angle": 0, + "content": "Emily Dinan, Stephen Roller, Kurt Shuster, Angela Fan, Michael Auli, and Jason Weston. 2019. Wizard of wikipedia: Knowledge-powered conversational agents. In ICLR. OpenReview.net." + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.41, + 0.885, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1746" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.086, + 0.487, + 0.152 + ], + "angle": 0, + "content": "Nouha Dziri, Sivan Milton, Mo Yu, Osmar R. Zaïane, and Siva Reddy. 2022. On the origin of hallucinations in conversational models: Is it the datasets or the models? In *NAACL*, pages 5271-5285. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.162, + 0.487, + 0.239 + ], + "angle": 0, + "content": "Karthik Gopalakrishnan, Behnam Hedayatnia, Qinglang Chen, Anna Gottardi, Sanjeev Kwatra, Anu Venkatesh, Raefer Gabriel, and Dilek Hakkani-Tur. 2019. Topical-chat: Towards knowledge-grounded open-domain conversations. In Interspeech, pages 1891-1895. ISCA." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.25, + 0.487, + 0.33 + ], + "angle": 0, + "content": "Behnam Hedayatnia, Karthik Gopalakrishnan, Seokhwan Kim, Yang Liu, Mihail Eric, and Dilek Hakkani-Tur. 2020. Policy-driven neural response generation for knowledge-grounded dialog systems. In INLG, pages 412-421. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.339, + 0.487, + 0.391 + ], + "angle": 0, + "content": "Byeongchang Kim, Jaewoo Ahn, and Gunhee Kim. 2020. Sequential latent knowledge selection for knowledge-grounded dialogue. In ICLR. OpenReview.net." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.402, + 0.487, + 0.454 + ], + "angle": 0, + "content": "Mojtaba Komeili, Kurt Shuster, and Jason Weston. 2022. Internet-augmented dialogue generation. In ACL, pages 8460-8478. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.465, + 0.487, + 0.517 + ], + "angle": 0, + "content": "Jiwei Li, Michel Galley, Chris Brockett, Jianfeng Gao, and Bill Dolan. 2016. A diversity-promoting objective function for neural conversation models. In HLT-NAACL, pages 110-119." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.527, + 0.487, + 0.58 + ], + "angle": 0, + "content": "Rongzhong Lian, Min Xie, Fan Wang, Jinhua Peng, and Hua Wu. 2019. Learning to select knowledge for response generation in dialog systems. In *IJCAI*, pages 5081-5087. ijcai.org." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.589, + 0.487, + 0.667 + ], + "angle": 0, + "content": "Chia-Wei Liu, Ryan Lowe, Iulian Serban, Michael Noseworthy, Laurent Charlin, and Joelle Pineau. 2016. How NOT to evaluate your dialogue system: An empirical study of unsupervised evaluation metrics for dialogue response generation. In EMNLP, pages 2122-2132." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.678, + 0.487, + 0.757 + ], + "angle": 0, + "content": "Zihan Liu, Mostofa Patwary, Ryan Prenger, Shrimai Prabhumoye, Wei Ping, Mohammad Shoeybi, and Bryan Catanzaro. 2022. Multi-stage prompting for knowledgeable dialogue generation. In *Findings of ACL*, pages 1317-1337. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.767, + 0.486, + 0.793 + ], + "angle": 0, + "content": "George A. Miller. 1995. Wordnet: A lexical database for english. Commun. ACM, 38(11):39-41." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.803, + 0.487, + 0.843 + ], + "angle": 0, + "content": "Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In ACL, pages 311-318." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.853, + 0.487, + 0.918 + ], + "angle": 0, + "content": "Hannah Rashkin, David Reitter, Gaurav Singh Tomar, and Dipanjan Das. 2021. Increasing faithfulness in knowledge-grounded dialogue with controllable features. In ACL/IJCNLP, pages 704-718. Association for Computational Linguistics." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.487, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.086, + 0.882, + 0.139 + ], + "angle": 0, + "content": "Nils Reimers and Iryna Gurevych. 2021. The curse of dense low-dimensional information retrieval for large index sizes. In ACL/IJCNLP, pages 605-611. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.151, + 0.883, + 0.217 + ], + "angle": 0, + "content": "Kurt Shuster, Spencer Poff, Moya Chen, Douwe Kiela, and Jason Weston. 2021. Retrieval augmentation reduces hallucination in conversation. In *Findings of EMNLP*, pages 3784–3803. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.229, + 0.883, + 0.295 + ], + "angle": 0, + "content": "Bin Sun, Shaoxiong Feng, Yiwei Li, Jiamou Liu, and Kan Li. 2021. Generating relevant and coherent dialogue responses using self-separated conditional variational autoencoders. In ACL/IJCNLP, pages 5624-5637. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.307, + 0.883, + 0.358 + ], + "angle": 0, + "content": "Weiwei Sun, Zhengliang Shi, Shen Gao, Pengjie Ren, Maarten de Rijke, and Zhaochun Ren. 2022. Contrastive learning reduces hallucination in conversations. CoRR, abs/2212.10400." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.372, + 0.883, + 0.451 + ], + "angle": 0, + "content": "Josef Valvoda, Yimai Fang, and David Vandyke. 2022. Prompting for a conversation: How to control a dialog model? In Proceedings of the Second Workshop on When Creative AI Meets Conversational AI, pages 1-8, Gyeongju, Republic of Korea. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.463, + 0.883, + 0.528 + ], + "angle": 0, + "content": "Wenquan Wu, Zhen Guo, Xiangyang Zhou, Hua Wu, Xiyuan Zhang, Rongzhong Lian, and Haifeng Wang. 2019. Proactive human-machine conversation with explicit conversation goal. In ACL, pages 3794-3804. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.541, + 0.883, + 0.594 + ], + "angle": 0, + "content": "Xinnuo Xu, Ondrej Dusek, Ioannis Konstas, and Verena Rieser. 2018. Better conversations by modeling, filtering, and optimizing for coherence and diversity. In EMNLP, pages 3981-3991." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.606, + 0.883, + 0.698 + ], + "angle": 0, + "content": "Yuan Yao, Bowen Dong, Ao Zhang, Zhengyan Zhang, Ruobing Xie, Zhiyuan Liu, Leyu Lin, Maosong Sun, and Jianyong Wang. 2022. Prompt tuning for discriminative pre-trained language models. In *Findings of the Association for Computational Linguistics: ACL* 2022, pages 3468-3473, Dublin, Ireland. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.71, + 0.883, + 0.774 + ], + "angle": 0, + "content": "Xueliang Zhao, Tingchen Fu, Chongyang Tao, and Rui Yan. 2022a. There is no standard answer: Knowledge-grounded dialogue generation with adversarial activated multi-reference learning. CoRR, abs/2210.12459." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.788, + 0.883, + 0.84 + ], + "angle": 0, + "content": "Xueliang Zhao, Wei Wu, Chongyang Tao, Can Xu, Dongyan Zhao, and Rui Yan. 2020a. Low-resource knowledge-grounded dialogue generation. In ICLR. OpenReview.net." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.853, + 0.883, + 0.918 + ], + "angle": 0, + "content": "Xueliang Zhao, Wei Wu, Can Xu, Chongyang Tao, Dongyan Zhao, and Rui Yan. 2020b. Knowledge-grounded dialogue generation with pre-trained language models. In EMNLP, pages 3377-3390. Association for Computational Linguistics." + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.086, + 0.883, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1747" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.116, + 0.086, + 0.489, + 0.152 + ], + "angle": 0, + "content": "Yingxiu Zhao, Yinhe Zheng, Zhiliang Tian, Chang Gao, Bowen Yu, Haiyang Yu, Yongbin Li, Jian Sun, and Nevin L. Zhang. 2022b. Prompt conditioned VAE: enhancing generative replay for lifelong learning in task-oriented dialogue. CoRR, abs/2210.07783." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.163, + 0.49, + 0.241 + ], + "angle": 0, + "content": "Wen Zheng, Natasa Milic-Frayling, and Ke Zhou. 2021. Knowledge-grounded dialogue generation with term-level de-noising. In Findings of ACL/IJCNLP, volume ACL/IJCNLP 2021 of Findings of ACL, pages 2972-2983. Association for Computational Linguistics." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.254, + 0.49, + 0.308 + ], + "angle": 0, + "content": "Kangyan Zhou, Shrimai Prabhumoye, and Alan W. Black. 2018. A dataset for document grounded conversations. In EMNLP, pages 708-713. Association for Computational Linguistics." + }, + { + "type": "table", + "bbox": [ + 0.117, + 0.333, + 0.486, + 0.434 + ], + "angle": 0, + "content": "
Prefix PromptsPost Prompts
I was thinking that perhaps I am not sure, maybe that Not very clear, maybe Not very clear, perhaps I was thinking that maybeMaybe i am wrong. If I am wrong, please correct me. If I am wrong, please for-give me. If it is wrong, please tell me.
" + }, + { + "type": "table_caption", + "bbox": [ + 0.14, + 0.444, + 0.46, + 0.459 + ], + "angle": 0, + "content": "Table 7: The designed prefix and post prompts." + }, + { + "type": "table", + "bbox": [ + 0.117, + 0.488, + 0.486, + 0.741 + ], + "angle": 0, + "content": "
Euphemistic Responses
Interesting, do you know that? \nThat sounds pretty good. Are there any way to visit? \nOh, I had not heard. \nHmm, I have never heard of that. What is that one about? \nI have never heard. Can you tell me more about it? \nOh, wow, that is remarkable. \nI have never played those, are they fun? \nCan I ask you about it? \nPlease tell me more about that. \nCan you tell me more about that? \nI have never had that. Anything else you can tell me? \nThat's really interesting! But I have never heard of that. \nI literally know nothing about that! \nI have no idea about that. \nI have not heard that one. I will have to check it out. \nHuh, maybe I will need to check that out then. \nOh, I misunderstood then. \nOh, i do not know about that. \nWow, that's a lot! I haven't heard of those.
" + }, + { + "type": "table_caption", + "bbox": [ + 0.144, + 0.751, + 0.456, + 0.766 + ], + "angle": 0, + "content": "Table 8: The designed euphemistic responses." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.796, + 0.364, + 0.813 + ], + "angle": 0, + "content": "A Prefix and Post Prompts" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.823, + 0.489, + 0.87 + ], + "angle": 0, + "content": "We manually design five prefix prompts and four post prompts, which are shown in Table 7. We discuss below about the prefixes and posts." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.872, + 0.489, + 0.919 + ], + "angle": 0, + "content": "We designed the prefixes and posts based on the WoW dataset and our daily conversation habits. In WoW dataset, one role is \"0_Wizard\", and the other" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.229 + ], + "angle": 0, + "content": "is \"1_Aprentice\". We noticed that the 1_Aprentice will give the sentences such as \"correct my if I am wrong ...\", which is also easy to appear in our daily conversation. Taking inspiration of this, we manually designed the prefixes and posts. Moreover, since the PLATO is pre-trained on conversation datasets, these prefixes may introduce the pre-knowledge that the model learned during the pre-training process." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.23, + 0.886, + 0.374 + ], + "angle": 0, + "content": "In fact, we declare the weakness of our manual prefixes and posts, i.e. direct connections of prefixes, responses, and posts do not fit all contexts. Therefore, we are exploring a new way of constructing replies, such as passing the design prefix, response, post, and context into the large-language-model to rewrite the appropriate response. We believe that better prefixes and posts will lead to more benefits in solving the hallucination problem." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.386, + 0.75, + 0.403 + ], + "angle": 0, + "content": "B Euphemistic Responses" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.412, + 0.884, + 0.443 + ], + "angle": 0, + "content": "We manually design nineteen euphemistic responses, which are shown in Table 8." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.455, + 0.816, + 0.487 + ], + "angle": 0, + "content": "C Dissuasion about the boundary between ak-less and ak-more" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.498, + 0.845, + 0.513 + ], + "angle": 0, + "content": "Below we provide an example in our dataset:" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.524, + 0.882, + 0.556 + ], + "angle": 0, + "content": "- Ground-truth Knowledge: laziness | thesis (\"thesis\") is a 1996 spanish thriller film." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.567, + 0.882, + 0.599 + ], + "angle": 0, + "content": "- AK-Less Knowledge: acedia | thesis (\"thesis\") is a 1996 spanish thriller film." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.61, + 0.884, + 0.656 + ], + "angle": 0, + "content": "- AK_More Knowledge: laziness | thesis (\"thesis\") personate a 1996 spanish thriller picture show." + }, + { + "type": "list", + "bbox": [ + 0.532, + 0.524, + 0.884, + 0.656 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.669, + 0.885, + 0.781 + ], + "angle": 0, + "content": "It can be noted that the more synonyms are introduced into a sentence, the semantics of the sentence will become more and more different from the original semantics. Therefore, we suppose that replacing at least \\(30\\%\\) of words at once will make a big difference in sentence semantics. Then, we decided the boundary between ak-less and ak-more." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1748" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.134, + 0.085, + 0.435, + 0.1 + ], + "angle": 0, + "content": "ACL 2023 Responsible NLP Checklist" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.108, + 0.323, + 0.123 + ], + "angle": 0, + "content": "A For every submission:" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.127, + 0.534, + 0.144 + ], + "angle": 0, + "content": "A1. Did you describe the limitations of your work?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.145, + 0.806, + 0.16 + ], + "angle": 0, + "content": "We provide a section of Limitations after the Conclusion and before the Ethics Statement" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.171, + 0.553, + 0.187 + ], + "angle": 0, + "content": "A2. Did you discuss any potential risks of your work?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.188, + 0.351, + 0.202 + ], + "angle": 0, + "content": "Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.213, + 0.697, + 0.229 + ], + "angle": 0, + "content": "A3. Do the abstract and introduction summarize the paper's main claims?" + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.231, + 0.164, + 0.243 + ], + "angle": 0, + "content": "1" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.256, + 0.67, + 0.273 + ], + "angle": 0, + "content": "A4. Have you used AI writing assistants when working on this paper?" + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.274, + 0.232, + 0.288 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.301, + 0.49, + 0.317 + ], + "angle": 0, + "content": "B Did you use or create scientific artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.322, + 0.215, + 0.337 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.348, + 0.53, + 0.364 + ], + "angle": 0, + "content": "B1. Did you cite the creators of artifacts you used?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.365, + 0.351, + 0.38 + ], + "angle": 0, + "content": "Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.391, + 0.779, + 0.407 + ], + "angle": 0, + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.408, + 0.351, + 0.423 + ], + "angle": 0, + "content": "Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.434, + 0.882, + 0.497 + ], + "angle": 0, + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.499, + 0.351, + 0.514 + ], + "angle": 0, + "content": "Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.525, + 0.882, + 0.573 + ], + "angle": 0, + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.574, + 0.468, + 0.589 + ], + "angle": 0, + "content": "We use a publicly well-established dataset." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.6, + 0.882, + 0.632 + ], + "angle": 0, + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.633, + 0.351, + 0.648 + ], + "angle": 0, + "content": "Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.659, + 0.882, + 0.74 + ], + "angle": 0, + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be." + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.741, + 0.232, + 0.755 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.765, + 0.494, + 0.782 + ], + "angle": 0, + "content": "C Did you run computational experiments?" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.788, + 0.147, + 0.8 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.814, + 0.882, + 0.846 + ], + "angle": 0, + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.847, + 0.705, + 0.861 + ], + "angle": 0, + "content": "We use the released code and checkpoints. We cite the source of our model." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.869, + 0.878, + 0.893 + ], + "angle": 0, + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1749" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.129, + 0.085, + 0.88, + 0.133 + ], + "angle": 0, + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.144, + 0.881, + 0.208 + ], + "angle": 0, + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.218, + 0.881, + 0.281 + ], + "angle": 0, + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? 3" + }, + { + "type": "list", + "bbox": [ + 0.129, + 0.085, + 0.881, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.293, + 0.878, + 0.328 + ], + "angle": 0, + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? 3" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.342, + 0.881, + 0.389 + ], + "angle": 0, + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.401, + 0.881, + 0.464 + ], + "angle": 0, + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.475, + 0.881, + 0.538 + ], + "angle": 0, + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? 2" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.551, + 0.873, + 0.582 + ], + "angle": 0, + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.593, + 0.881, + 0.641 + ], + "angle": 0, + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Not applicable. Left blank." + }, + { + "type": "list", + "bbox": [ + 0.129, + 0.342, + 0.881, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1750" + } + ] +] \ No newline at end of file diff --git a/2023/Towards Fewer Hallucinations in Knowledge-Grounded Dialogue Generation via Augmentative and Contrastive Knowledge-Dialogue/989c66cd-1d3d-4fda-ba4f-e2869fc255a6_origin.pdf b/2023/Towards Fewer Hallucinations in Knowledge-Grounded Dialogue Generation via Augmentative and Contrastive Knowledge-Dialogue/989c66cd-1d3d-4fda-ba4f-e2869fc255a6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b773ea100aa8e1125fe348a7efa6d9ed25c1f53a --- /dev/null +++ b/2023/Towards Fewer Hallucinations in Knowledge-Grounded Dialogue Generation via Augmentative and Contrastive Knowledge-Dialogue/989c66cd-1d3d-4fda-ba4f-e2869fc255a6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9fc5f2c093bb38ef9171bfe5ea04b8ba202f701797c7bdee9e93dfb5fa77fbab +size 364686 diff --git a/2023/Towards Fewer Hallucinations in Knowledge-Grounded Dialogue Generation via Augmentative and Contrastive Knowledge-Dialogue/full.md b/2023/Towards Fewer Hallucinations in Knowledge-Grounded Dialogue Generation via Augmentative and Contrastive Knowledge-Dialogue/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c9e7d4b410bcb5331a5bd2f9cde36e8500700f25 --- /dev/null +++ b/2023/Towards Fewer Hallucinations in Knowledge-Grounded Dialogue Generation via Augmentative and Contrastive Knowledge-Dialogue/full.md @@ -0,0 +1,326 @@ +# Towards Fewer Hallucinations in Knowledge-Grounded Dialogue Generation via Augmentative and Contrastive Knowledge-Dialogue + +Bin Sun $^{1}$ , Yitong Li $^{2,3}$ , Fei Mi $^{2}$ , FanHu Bie $^{3}$ , Yiwei Li $^{1}$ , Kan Li $^{1*}$ + +$^{1}$ School of Computer Science & Technology, Beijing Institute of Technology + +$^{2}$ Huawei Noah's Ark Lab $^{3}$ Huawei Technologies Ltd. + +{binsun,liyiwei,likan}@bit.edu.cn + +{liyitong3,mifei2,biefanhu}@huawei.com + +# Abstract + +Existing knowledge-grounded open-domain dialogue generation models often face the hallucination problem, i.e. the dialogue generative model will persist in an inappropriate knowledge and generate responses that inconsistent with the facts. We argue that this problem mainly stems from the polarized optimization objectives and weak knowledge generation ability. To mitigate the hallucination, we take inspiration from human communicating that people will replay euphemistic responses for the unclear or unrecognizable knowledge, and propose an Augmentative and Contrastive Knowledge Dialogue Expansion Framework (ACK-DEF). ACK-DEF constructs the augmentative and contrastive knowledge dialogue samples, which consist of the knowledge of different degrees of errors and the response of manual design, to expand the original training set and smooth the polarized optimization objective that enables models to generate ground-truth with or without gold knowledge. Not only the knowledge, ACK-DEF also provides the tactful responses of manual design corresponding to the incomplete correct knowledge. Experimental results on the Wikipedia of Wizard dataset show that employing the ACK-DEF is effective to alleviate the hallucination problem. + +# 1 Introduction + +Recently, Knowledge-Grounded Dialogue Generation draws dramatic attentions in artificial intelligence community. Many efforts incorporate knowledge information to improve the performance of dialogue generation models (Zhou et al., 2018; Dinan et al., 2019; Gopalakrishnan et al., 2019; Kim et al., 2020; Zhao et al., 2020a; Zheng et al., 2021; Zhao et al., 2022a; Bao et al., 2022). However, these methods always face the hallucination problem, that is, the dialogue generation model may insist on an inappropriate knowledge and generate responses that inconsistent with the facts (Rashkin et al., 2021; Zhao et al., 2022a; Dziri et al., 2022). + +We argue that the hallucination problem primarily caused by two aspects: (1) The optimization objective is usually polarized by the gold knowledge dialogue samples and general dialogue samples without knowledge in current knowledge-grounded dialogue datasets (Zhou et al., 2018; Gopalakrishnan et al., 2019; Dinan et al., 2019; Wu et al., 2019; Komeili et al., 2022). Few datasets consider teaching models how to respond when dealing with incomplete correct knowledge, which makes the models tend to believe in the given knowledge, regardless of whether the knowledge is appropriate or not, resulting in hallucination problems. In addition, the knowledge retrieval system tends to extract irrelevant knowledge rather than relevant knowledge when the database is large, aggravating the hallucinations (Reimers and Gurevych, 2021; Liu et al., 2022). (2) The generation of knowledge may also face the hallucination problem and obtain the inappropriate knowledge, leading the generation of hallucination responses (Kim et al., 2020; Zhao et al., 2020a; Liu et al., 2022; Adolphs et al., 2021; Bao et al., 2022). + +To mitigate the hallucination problem, we propose an Augmentative and Contrastive Knowledge Dialogue Expansion Framework (ACK-DEF), which is inspired by human communicating that people will replay euphemistic response for the unrecognizable knowledge. ACK-DEF is proposed to smooth the polarized optimization objective by augmenting training set with augmentative and contrastive knowledge-dialogue samples. Not only the knowledge, we also designed the reply patterns for the knowledge with different level of errors. For this, we propose the augmentative knowledge dialogue expansion (AK), and contrastive knowledge dialogue expansion (CK). AK is proposed to boost the generalization ability of models on knowledge with minor noise. On the contrary, inspired from the contrastive learning paradigm (Cai et al., 2020; Chen et al., 2020a,b; Sun et al., 2021, 2022), CK + +![](images/b3d345dea917129cfa22a2ad09d6cca3aea481adcf1f8a95852176977677e589.jpg) +Figure 1: A diagram of our Augmentative Knowledge dialogue expansion method. We replace different proportion of words in the original knowledge with synonyms to construct incomplete correct knowledge, and design response for different knowledge. We also use prompts to guide the dialogue generation process. + +reconstructs incorrect knowledge and designs euphemistic responses, which aims to push the model learn the reply pattern of incorrect knowledge and a better boundary between correct and incorrect knowledge. + +Contributions: We propose an ACK-DEF to construct new knowledge-dialogue samples that consist of knowledge with different level of errors and manual responses, to soften the training optimization objectives of models, which will mitigate the hallucination. Finally, we conduct extension experiments to show the superior performance of ACK-DEF on alleviating the hallucination. + +# 2 Methodology + +To mitigate the hallucination problem that caused by the polarized optimization objectives in knowledge grounded dialogue generation, we take inspiration from human communicating, and propose the Augmentative and Contrastive Knowledge Dialogue Expansion Framework (ACK-DEF). Our ACK-DEF aims to soften the polarized training optimization objectives of current knowledge-grounded dialogue generation methods, and guide the dialogue system reply patterns for the knowledge with different level of errors. To achieve this end, we design two effective expansion method, which will be detailed in below. + +# 2.1 Augmentative Knowledge Dialogue + +We propose the Augmentative Knowledge (AK) dialogue expansion to boost the generalization ability of the dialogue model on the knowledge with simi + +![](images/19f02e3a689096203540d697ce684f8bf997410565aa7090e0aeacf6395192ce.jpg) +Figure 2: A diagram of our Contrastive Knowledge dialogue expansion method. We use the antonym to reconstruct the knowledge information and design multiple responses for such knowledge. Since antonyms transform the semantics of the original knowledge, the noise knowledge often contains wrong facts. By this, the model can learn a better boundary between correct and incorrect knowledge, and a safety reply pattern for incorrect knowledge. + +lar semantics but different expressions, which can prevent the model from being interfered by partialrelevant knowledge retrieved by the retrieval systems (Lian et al., 2019; Zhao et al., 2020b; Hedayatnia et al., 2020; Zheng et al., 2021; Shuster et al., 2021; Komeili et al., 2022). As shown in Figure 1, we employ the synonym data augmentation tool, which replaces words in the original knowledge with synonyms, to reconstruct the knowledge information (Miller, 1995). Considering that the synonym may disrupt the original semantics of new constructed knowledge, we constrain the replace possibility within [0.1,0.2]. Hence, we can obtain the approximate knowledge. Combining this knowledge and the original dialogue, we obtain the "ak-less sample". In addition, we also replace $30\%$ to $50\%$ words with their synonyms to construct the less similar knowledge. Inspired from prompt learning paradigm (Yao et al., 2022; Valvoda et al., 2022; Zhao et al., 2022b), we manually produce some Prefix-prompts and Post-prompts (see Appendix) to (1) make the new response more tactful for the less similar knowledge; (2) regulate and guide the dialogue generation process of the model. We call the sample consist of less-similar knowledge and designed response as "ak-more sample". + +# 2.2 Contrastive Knowledge Dialogue + +We propose the Contrastive Knowledge (CK) dialogue expansion, inspired from the contrastive learning paradigm (Chen et al., 2020b; Cai et al., 2020), not only construct the incorrect knowledge as negative samples for original knowledge, but also build the euphemistic responses as positive + +samples for the original response with incorrect knowledge. To help the model learn a boundary between correct and incorrect knowledge, we employ the antonym to make up new incorrect knowledge. For example, given the knowledge "nintendo was founded on 23 september 1889 ...", the "founded" will be replaced with "abolish", which greatly changes the semantics but little changes the expression. After that, we random choose an euphemistic response to replace the original response of the dialogue. Finally, The incorrect knowledge and the replaced euphemistic response are combined as the "ck-sample". + +# 3 Experiment and Results + +# 3.1 Experiment Settings + +# 3.1.1 Dataset + +We use the Wikipedia of Wizard (WoW) data, a well-established knowledge-grounded open-domain dialogue dataset, for our experiment. We pre-precess the WoW dataset and extract the single-turn knowledge dialogue samples. To evaluate the performance of our method in detail, we perform four test sets: normal, ak-less, ak-more and ck. The normal set is the original test set. And the ak-less, ak-more and ck are the sets consist of ak-less, ak-more and ck samples, respectively. We also follow the settings of WoW data and divide the test set into two groups (seen test and unseen test): the topic of the knowledge in the unseen test set is missing in the training set. + +# 3.1.2 Baseline + +We employ the released PLATO-v1 (Bao et al., 2020) model, a pre-trained dialogue generation model based on UniLM, for our experiment. + +Fine-tuning We directly finetune a model on the original WoW training set. By this, the model can only see gold knowledge dialogue samples and general dialogue samples without knowledge. Hence, we call the fine-tuned model PLATO+GOLD. + +Fine-tuning with ACK-DEF We finetune the model with the original set and the expansion samples that obtained through ACK-DEF. Thence, we call it PLATO+ACK-DEF. + +# 3.1.3 AutoEvaluation Metrics + +Dialogue Metrics Our primary metrics of interest are Distinct-n (Li et al., 2016), Response Length (Len.) (Csaky et al., 2019), BLEU (Papineni et al., 2002), Embedding-based (Greedy (GRE), Average (AVG), Extrema (EXT)) (Liu et al., 2016), and Coherence (COH) (Xu et al., 2018). Distinct-n evaluates the diversity of generated responses, which is calculated through the ratio of distinct $n$ -grams and all generated $n$ -grams. Len. is the average number of words of all generated responses. BLEU validates the degree of the word-overlap between the generated response and the ground-truth, which denotes the consistence between generated response and ground-truth. Embedding-based metrics (GRE, AVG and EXT) are introduced to evaluate the semantic relationship of generated responses and ground-truth responses, illustrating the consistence in semantic level. COH. mainly assesses the relevance between contexts and generated responses. + +Knowledge Metrics We follow the PLATO(Bao et al., 2020) and use the knowledge precision, recall and f1 scores. These metrics are used to calculate the ratio of tokens that exist in common in ground-truth knowledge and generated responses to tokens in generated responses. "Recall" is the average ratio of the number of overlapping tokens in response and knowledge to the number of tokens in knowledge. And "Precision" is the average ratio of the number of overlapping tokens to the number of tokens in response. In other words, "Recall" indicates how much knowledge information is contained in the response, while "Precision" indicates the proportion of knowledge information in the response. Even we involve the negative and incorrect knowledge in response generation, we still use the ground-truth knowledge to calculate the metrics in Table 3,4. + +# 3.2 Dialogue Performance Analysis + +Table 1 and Table 2 report the automatic results on four test sets and four unseen test sets, respectively. In these Tables, it can be observed that (1) the PLATO+ACK-DEF has a competitive performance with PLATO+GOLD on the normal set, which means that the PLATO+ACK-DEF can recognize the golden knowledge and produce appropriate responses. (2) the PLATO+GOLD perform worse than PLATO+ACK-DEF on ak-less, which means that the robustness of the dialogue model only trained with golden knowledge is very weak. + +
test setDistinct-1/2/3Len.BLEU-1/2/3/4GREAVGEXTCOH
normal0.10680.453313.690.42800.29650.21100.15290.73920.86890.63610.7808
0.09020.398416.200.44280.30170.21090.14990.73660.86830.63300.7878
ak-less0.11940.502413.500.38610.25740.17450.11920.71600.86070.61480.7755
0.08230.353218.780.45020.29820.20150.13800.73070.86960.62930.7948
ak-more0.12340.517412.810.16750.10620.06800.04350.69080.85510.59940.7706
0.06750.294621.830.43580.30010.21230.15420.77450.91510.70930.8098
ck0.11090.477913.230.29650.17790.10800.06570.58380.76220.53730.7712
0.06520.202913.360.42300.27050.18090.12660.65720.83060.61620.8049
+ +Table 1: The automatic results of PLATO+GOLD (up) and PLATO+ACK-DEF (down) on four test seen sets. + +
test setDistinct-1/2Len.BLUE-1/2/3/4GREAVGEXTCOH
normal0.05030.242212.430.35160.23310.15820.10900.69880.85680.63060.8094
0.04670.231113.140.34630.22810.15360.10490.69680.85410.63380.8105
ak-less0.09660.391713.390.38710.25650.17240.11640.71430.86000.61220.7836
0.06230.266419.180.44430.29070.19360.13010.72320.86630.61940.8026
ak-more0.10640.444012.710.16520.10460.06680.04260.68880.85380.59800.7797
0.05610.240021.820.43310.29680.20910.15110.76970.91140.70370.8197
ck0.08130.332413.240.30110.18090.11000.06690.58540.76760.54790.7794
0.04650.149013.520.43290.27750.18610.13070.66120.83340.62150.8145
+ +Table 2: The automatic results of PLATO+GOLD (up) and PLATO+ACK-DEF (down) on four test sets with unseen knowledge. + +
test setRecallPrecisionF1avg. Dec.
normal0.36070.70090.4546-
ak-less0.28830.55850.3618∇ 0.1026
ak-more0.17520.36320.2228∇ 0.2517
ck0.31930.61330.4003∇ 0.0611
normal0.36950.65380.4520-
ak-less0.32510.56360.3927∇ 0.0647
ak-more0.23350.39830.2775∇ 0.1887
ck0.10650.20410.1337∇ 0.3437
+ +Table 3: The knowledge correlation results of PLATO+GOLD (up) and PLATO+ACK-DEF (down) on four test sets with seen knowledge. + +
test setRecallPrecisionF1avg. Dec.
normal0.37320.74420.4736-
ak-less0.27280.54750.3452∇ 0.1418
ak-more0.16650.36270.2152∇ 0.2822
ck0.30280.60680.3830∇ 0.0995
normal0.36550.68820.4535-
ak-less0.29380.53480.3579∇ 0.1069
ak-more0.20460.37140.2481∇ 0.2277
ck0.08700.18470.1116∇ 0.3747
+ +Table 4: The knowledge correlation results of PLATO+GOLD (up) and PLATO+ACK-DEF (down) on four test sets with unseen knowledge. + +Even if the knowledge information only changes by $10\%$ to $20\%$ , the performance of the model will + +significantly decline, especially consistency metrics (i.e. BLEU, GRE, AVG and EXT). (3) the PLATO+GOLD achieve better Distinct scores but weaker BLEU and embedding-based scores, which means that the PLATO+GOLD is easy to generate responses that are very different from ground-truth responses, that is, the hallucinations. + +# 3.3 Knowledge Correlation Analysis + +Table 3 and Table 4 report the knowledge correlation result of PLATO+GOLD and PLATO+ACK-DEF on four test sets and four test unseen sets, respectively. From these table, we can observe that the performance of PLATO+GOLD is reduced when the given knowledge changed, which illustrates the danger that the model generate responses based on incorrect knowledge. In addition to the above findings, we also observed that the recall, precision and f1 scores of PLATO+ACK-DEF are better than PLATO+GOLD on ak-less and ak-more sets, which demonstrates that using ACK-DEF effectively enhance the model's capability for the similar knowledge information. Moreover, the result of PLATO+ACK-DEF on the ck set is significantly reduced, which shows that the model distinguishes the wrong knowledge constructed with antonyms and gives an appropriate response with + +
test setw. GOLD (%)w. ACK-DEF (%)kappa
normal13.0014.000.481
ak-less23.6717.330.513
ak-more33.6724.330.479
ck21.675.670.597
total23.0015.330.552
+ +out knowledge (see Table 1 and Table 2 for the effect). These results are inline with our exception that incorporating noised knowledge dialogue samples in training stages can smooth the polarized optimization objective, and mitigate the hallucination problem. + +According to the results of test seen sets and unseen sets), we can notice that the PLATO+ACK-DEF achieves a good performance on ground-truth seen knowledge and a weak performance on ground-truth unseen knowledge. This illustrates that the PLATO+ACK-DEF may doubt the authenticity of unseen given knowledge (even if the knowledge is the ground-truth), and will not fully use it to generate responses. This may alleviate the hallucination, and we believe it is caused by (1) the Augmentative knowledge dialogue introduces similar knowledge to improve the generalization of the model; (2) the Contrastive knowledge dialogue introduces knowledge independent responses, which tell the model to generate responses without knowledge; (3) the ACK-DEF smooths the polarized optimization, which ensures the model not to directly use the given knowledge. + +# 3.4 Human Evaluation + +To further evaluation the ability of our ACK-DEF on reducing the hallucination problem, we randomly select 400 samples form four test sets, and hire three annotators to do human evaluations by assessing whether the responses generated by PLATO +GOLD and +ACK-DEL have hallucinations. Table 5 reports the results of human evaluation, from which we can notice that the PLATO+ACK-DEF generate less hallucinations than PLATO+GOLD. This shows the effectiveness of our ACK-DEF. + +# 3.5 Case Study + +Table 6 shows a generated case of PLATO+GOLD (GOLD) and PLATO+ACK-DEF (ACK-DEF) based on different knowledge information. We can observed that the GOLD is convinced of the given knowledge, regardless of whether the knowledge + +Table 5: The human evaluation results. + +
QueryThey really are super cool looking dogs +they are so amazing
Normal +Knowledgegolden retriever are well suited to resi- +dency in suburban or country environ- +ments.
GOLDI love golden retrievers! They are well +suited to residency in suburban or country.
ACK-DEFThey are! They are well suited to suburban +or country environments.
AK-Less +Knowledgegolden retriever are good suited to resi- +dency in suburban operating elbow room +land environments.
GOLDGolden retriever are great suited to resi- +dency in suburban operating elbow
ACK-DEFGolden retriever are great suited to subur- +ban environments.
AK-More +Knowledgegild retriever make up considerably suit +to residency in suburban surgery common- +wealth environments.
GOLDgild retrievers make up considerably suit +to residency in suburban areas.
ACK-DEFNot very clear, maybe they are! They can +be found in suburban areas. If it is wrong, +please tell me.
CK Knowl- +edgegolden retriever unequal disadvantage- +geously suited to residency in suburban or +country environments.
GOLDThey sure are! I have a golden retriever +who is unequally suited to residency in sub- +urban environments.
ACK-DEFThat sounds pretty good.
+ +Table 6: A case of PLATO +GOLD and +ACK-DEF. + +is appropriate or not, and more easily to copy the knowledge information into responses. Even the GOLD has seen the knowledge topic, it could not remember the knowledge in their parameters. On the contrary, the ACK-DEF has good resistance to incomplete correct knowledge. + +# 4 Conclusion + +This paper focuses on the hallucinations caused by polarized optimization objective in knowledge-grounded dialogue generation (KGDG), and proposes an augmentative and contrastive knowledge dialogue expansion framework (ACK-DEF) to mitigate it. The optimization objective of KGDG is to train the model could generate proper response with or without knowledge, which inevitably weaken the model's ability on unrecognized knowledge and lead hallucinations. Therefore, ACK-DEF constructs multiple level knowledge-dialogue samples to soften the optimization objective of KGDG. Extension experimental results show the superior performance of using our methods on dialogue metrics and knowledge correlations. + +# Limitations + +Our limitations are as follow: + +- Data Scale: This paper only employs the Wikipedia of Wizard dataset, a small scale and well-established knowledge conversation dataset, and lack of the validation on large-scale dataset. +- **Backbones:** This paper lacks the evaluating of other knowledge dialogue model on the proposed method. Actually, we have two reasons to employ the PLATO. First, the PLATO can better handle the one-to-many phenomenon, which is suitable for learning our expansion samples. Second, the PLATO is a pre-trained dialogue model, and its performance on knowledge dialogue generation task has been proved. We will evaluating the performance of other knowledge dialogue model on our method for our future work. +- Knowledge Expansion Methods: This paper only uses the synonym and antonym to construct the noised knowledge, which lacks of the comparison of using other data augmentation method. Indeed, we use two token-level data augmentation methods (synonym and antonym augmentation) to prove our statements on hallucination problem in knowledge dialogue generation task. Based on this study, we believe that incorporating other data augmentation methods will also mitigate the hallucinations. +- Manual Prompts and Responses: This paper designed five prefix prompts, four post-prompts and nineteen euphemistic responses. For $AK$ -More method, we simply randomly choose one prefix-prompt and one post-prompt and concatenate them with the ground-truth response. This leads to some irregular responses. As for $CK$ method, we randomly select one euphemistic response for the incorrect knowledge. However, we found that the response may not coherent with the query. We will design more smooth expansion ways to construct more human-like training samples for our future work. + +# Ethics Statement + +We acknowledge and ensure that our study is compatible with the provided Code of Ethics. + +Knowledge-grounded open-domain dialogue generation is crucial for building a knowledgeable dialogue system, which is beyond the wildest dreams in natural language process field. All our experiments are conducted on public available datasets to avoid ethical concerns. All terms for using these datasets are strictly followed in our study. There are no direct ethical concerns in our research. + +# Acknowledgments + +We would like to thank the anonymous reviewers for their constructive comments. This research is supported by Beijing Natural Science Foundation (No.4222037 and L181010) and BIT Research and Innovation Promoting Project (Grant No.2022YCXY021). Kan Li is the corresponding author. + +# References + +Leonard Adolphs, Kurt Shuster, Jack Urbanek, Arthur Szlam, and Jason Weston. 2021. Reason first, then respond: Modular generation for knowledge-infused dialogue. CoRR, abs/2111.05204. +Siqi Bao, Huang He, Fan Wang, Hua Wu, and Haifeng Wang. 2020. PLATO: pre-trained dialogue generation model with discrete latent variable. In ACL, pages 85-96. ACL. +Siqi Bao, Huang He, Jun Xu, Hua Lu, Fan Wang, Hua Wu, Han Zhou, Wenquan Wu, Zheng-Yu Niu, and Haifeng Wang. 2022. PLATO-K: internal and external knowledge enhanced dialogue generation. CoRR, abs/2211.00910. +Hengyi Cai, Hongshen Chen, Yonghao Song, Zhuoye Ding, Yongjun Bao, Weipeng Yan, and Xiaofang Zhao. 2020. Group-wise contrastive learning for neural dialogue generation. In EMNLP, pages 793-802. Association for Computational Linguistics. +Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. 2020a. A simple framework for contrastive learning of visual representations. In ICML, volume 119 of Proceedings of Machine Learning Research, pages 1597-1607. PMLR. +Xinlei Chen, Haoqi Fan, Ross B. Girshick, and Kaiming He. 2020b. Improved baselines with momentum contrastive learning. CoRR, abs/2003.04297. +Richard Csaky, Patrik Purgai, and Gábor Recski. 2019. Improving neural conversational models with entropy-based data filtering. In ACL (1), pages 5650-5669. +Emily Dinan, Stephen Roller, Kurt Shuster, Angela Fan, Michael Auli, and Jason Weston. 2019. Wizard of wikipedia: Knowledge-powered conversational agents. In ICLR. OpenReview.net. + +Nouha Dziri, Sivan Milton, Mo Yu, Osmar R. Zaïane, and Siva Reddy. 2022. On the origin of hallucinations in conversational models: Is it the datasets or the models? In *NAACL*, pages 5271-5285. Association for Computational Linguistics. +Karthik Gopalakrishnan, Behnam Hedayatnia, Qinglang Chen, Anna Gottardi, Sanjeev Kwatra, Anu Venkatesh, Raefer Gabriel, and Dilek Hakkani-Tur. 2019. Topical-chat: Towards knowledge-grounded open-domain conversations. In Interspeech, pages 1891-1895. ISCA. +Behnam Hedayatnia, Karthik Gopalakrishnan, Seokhwan Kim, Yang Liu, Mihail Eric, and Dilek Hakkani-Tur. 2020. Policy-driven neural response generation for knowledge-grounded dialog systems. In INLG, pages 412-421. Association for Computational Linguistics. +Byeongchang Kim, Jaewoo Ahn, and Gunhee Kim. 2020. Sequential latent knowledge selection for knowledge-grounded dialogue. In ICLR. OpenReview.net. +Mojtaba Komeili, Kurt Shuster, and Jason Weston. 2022. Internet-augmented dialogue generation. In ACL, pages 8460-8478. Association for Computational Linguistics. +Jiwei Li, Michel Galley, Chris Brockett, Jianfeng Gao, and Bill Dolan. 2016. A diversity-promoting objective function for neural conversation models. In HLT-NAACL, pages 110-119. +Rongzhong Lian, Min Xie, Fan Wang, Jinhua Peng, and Hua Wu. 2019. Learning to select knowledge for response generation in dialog systems. In *IJCAI*, pages 5081-5087. ijcai.org. +Chia-Wei Liu, Ryan Lowe, Iulian Serban, Michael Noseworthy, Laurent Charlin, and Joelle Pineau. 2016. How NOT to evaluate your dialogue system: An empirical study of unsupervised evaluation metrics for dialogue response generation. In EMNLP, pages 2122-2132. +Zihan Liu, Mostofa Patwary, Ryan Prenger, Shrimai Prabhumoye, Wei Ping, Mohammad Shoeybi, and Bryan Catanzaro. 2022. Multi-stage prompting for knowledgeable dialogue generation. In *Findings of ACL*, pages 1317-1337. Association for Computational Linguistics. +George A. Miller. 1995. Wordnet: A lexical database for english. Commun. ACM, 38(11):39-41. +Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In ACL, pages 311-318. +Hannah Rashkin, David Reitter, Gaurav Singh Tomar, and Dipanjan Das. 2021. Increasing faithfulness in knowledge-grounded dialogue with controllable features. In ACL/IJCNLP, pages 704-718. Association for Computational Linguistics. + +Nils Reimers and Iryna Gurevych. 2021. The curse of dense low-dimensional information retrieval for large index sizes. In ACL/IJCNLP, pages 605-611. Association for Computational Linguistics. +Kurt Shuster, Spencer Poff, Moya Chen, Douwe Kiela, and Jason Weston. 2021. Retrieval augmentation reduces hallucination in conversation. In *Findings of EMNLP*, pages 3784–3803. Association for Computational Linguistics. +Bin Sun, Shaoxiong Feng, Yiwei Li, Jiamou Liu, and Kan Li. 2021. Generating relevant and coherent dialogue responses using self-separated conditional variational autoencoders. In ACL/IJCNLP, pages 5624-5637. Association for Computational Linguistics. +Weiwei Sun, Zhengliang Shi, Shen Gao, Pengjie Ren, Maarten de Rijke, and Zhaochun Ren. 2022. Contrastive learning reduces hallucination in conversations. CoRR, abs/2212.10400. +Josef Valvoda, Yimai Fang, and David Vandyke. 2022. Prompting for a conversation: How to control a dialog model? In Proceedings of the Second Workshop on When Creative AI Meets Conversational AI, pages 1-8, Gyeongju, Republic of Korea. Association for Computational Linguistics. +Wenquan Wu, Zhen Guo, Xiangyang Zhou, Hua Wu, Xiyuan Zhang, Rongzhong Lian, and Haifeng Wang. 2019. Proactive human-machine conversation with explicit conversation goal. In ACL, pages 3794-3804. Association for Computational Linguistics. +Xinnuo Xu, Ondrej Dusek, Ioannis Konstas, and Verena Rieser. 2018. Better conversations by modeling, filtering, and optimizing for coherence and diversity. In EMNLP, pages 3981-3991. +Yuan Yao, Bowen Dong, Ao Zhang, Zhengyan Zhang, Ruobing Xie, Zhiyuan Liu, Leyu Lin, Maosong Sun, and Jianyong Wang. 2022. Prompt tuning for discriminative pre-trained language models. In *Findings of the Association for Computational Linguistics: ACL* 2022, pages 3468-3473, Dublin, Ireland. Association for Computational Linguistics. +Xueliang Zhao, Tingchen Fu, Chongyang Tao, and Rui Yan. 2022a. There is no standard answer: Knowledge-grounded dialogue generation with adversarial activated multi-reference learning. CoRR, abs/2210.12459. +Xueliang Zhao, Wei Wu, Chongyang Tao, Can Xu, Dongyan Zhao, and Rui Yan. 2020a. Low-resource knowledge-grounded dialogue generation. In ICLR. OpenReview.net. +Xueliang Zhao, Wei Wu, Can Xu, Chongyang Tao, Dongyan Zhao, and Rui Yan. 2020b. Knowledge-grounded dialogue generation with pre-trained language models. In EMNLP, pages 3377-3390. Association for Computational Linguistics. + +Yingxiu Zhao, Yinhe Zheng, Zhiliang Tian, Chang Gao, Bowen Yu, Haiyang Yu, Yongbin Li, Jian Sun, and Nevin L. Zhang. 2022b. Prompt conditioned VAE: enhancing generative replay for lifelong learning in task-oriented dialogue. CoRR, abs/2210.07783. + +Wen Zheng, Natasa Milic-Frayling, and Ke Zhou. 2021. Knowledge-grounded dialogue generation with term-level de-noising. In Findings of ACL/IJCNLP, volume ACL/IJCNLP 2021 of Findings of ACL, pages 2972-2983. Association for Computational Linguistics. + +Kangyan Zhou, Shrimai Prabhumoye, and Alan W. Black. 2018. A dataset for document grounded conversations. In EMNLP, pages 708-713. Association for Computational Linguistics. + +
Prefix PromptsPost Prompts
I was thinking that perhaps I am not sure, maybe that Not very clear, maybe Not very clear, perhaps I was thinking that maybeMaybe i am wrong. If I am wrong, please correct me. If I am wrong, please for-give me. If it is wrong, please tell me.
+ +Table 7: The designed prefix and post prompts. + +
Euphemistic Responses
Interesting, do you know that? +That sounds pretty good. Are there any way to visit? +Oh, I had not heard. +Hmm, I have never heard of that. What is that one about? +I have never heard. Can you tell me more about it? +Oh, wow, that is remarkable. +I have never played those, are they fun? +Can I ask you about it? +Please tell me more about that. +Can you tell me more about that? +I have never had that. Anything else you can tell me? +That's really interesting! But I have never heard of that. +I literally know nothing about that! +I have no idea about that. +I have not heard that one. I will have to check it out. +Huh, maybe I will need to check that out then. +Oh, I misunderstood then. +Oh, i do not know about that. +Wow, that's a lot! I haven't heard of those.
+ +Table 8: The designed euphemistic responses. + +# A Prefix and Post Prompts + +We manually design five prefix prompts and four post prompts, which are shown in Table 7. We discuss below about the prefixes and posts. + +We designed the prefixes and posts based on the WoW dataset and our daily conversation habits. In WoW dataset, one role is "0_Wizard", and the other + +is "1_Aprentice". We noticed that the 1_Aprentice will give the sentences such as "correct my if I am wrong ...", which is also easy to appear in our daily conversation. Taking inspiration of this, we manually designed the prefixes and posts. Moreover, since the PLATO is pre-trained on conversation datasets, these prefixes may introduce the pre-knowledge that the model learned during the pre-training process. + +In fact, we declare the weakness of our manual prefixes and posts, i.e. direct connections of prefixes, responses, and posts do not fit all contexts. Therefore, we are exploring a new way of constructing replies, such as passing the design prefix, response, post, and context into the large-language-model to rewrite the appropriate response. We believe that better prefixes and posts will lead to more benefits in solving the hallucination problem. + +# B Euphemistic Responses + +We manually design nineteen euphemistic responses, which are shown in Table 8. + +# C Dissuasion about the boundary between ak-less and ak-more + +Below we provide an example in our dataset: + +- Ground-truth Knowledge: laziness | thesis ("thesis") is a 1996 spanish thriller film. +- AK-Less Knowledge: acedia | thesis ("thesis") is a 1996 spanish thriller film. +- AK_More Knowledge: laziness | thesis ("thesis") personate a 1996 spanish thriller picture show. + +It can be noted that the more synonyms are introduced into a sentence, the semantics of the sentence will become more and more different from the original semantics. Therefore, we suppose that replacing at least $30\%$ of words at once will make a big difference in sentence semantics. Then, we decided the boundary between ak-less and ak-more. + +A For every submission: + +A1. Did you describe the limitations of your work? + +We provide a section of Limitations after the Conclusion and before the Ethics Statement + +A2. Did you discuss any potential risks of your work? + +Not applicable. Left blank. + +A3. Do the abstract and introduction summarize the paper's main claims? + +1 + +A4. Have you used AI writing assistants when working on this paper? + +Left blank. + +B Did you use or create scientific artifacts? + +Left blank. + +B1. Did you cite the creators of artifacts you used? + +Not applicable. Left blank. + +B2. Did you discuss the license or terms for use and / or distribution of any artifacts? + +Not applicable. Left blank. + +B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? + +Not applicable. Left blank. + +B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? + +We use a publicly well-established dataset. + +B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? + +Not applicable. Left blank. + +B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. + +Left blank. + +C Did you run computational experiments? + +3 + +C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? + +We use the released code and checkpoints. We cite the source of our model. + +The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance. + +C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Not applicable. Left blank. +C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Not applicable. Left blank. +C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? 3 + +# D Did you use human annotators (e.g., crowdworkers) or research with human participants? 3 + +D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Not applicable. Left blank. +D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Not applicable. Left blank. +D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? 2 +D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Not applicable. Left blank. +D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Not applicable. Left blank. \ No newline at end of file diff --git a/2023/Towards Fewer Hallucinations in Knowledge-Grounded Dialogue Generation via Augmentative and Contrastive Knowledge-Dialogue/images.zip b/2023/Towards Fewer Hallucinations in Knowledge-Grounded Dialogue Generation via Augmentative and Contrastive Knowledge-Dialogue/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8218f13d27bbb8186ee1052e368f87b302bf03f7 --- /dev/null +++ b/2023/Towards Fewer Hallucinations in Knowledge-Grounded Dialogue Generation via Augmentative and Contrastive Knowledge-Dialogue/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:31c51e4850efc618bf69c2001cc790482d8eeef614200087357f7596647020e7 +size 541112 diff --git a/2023/Towards Fewer Hallucinations in Knowledge-Grounded Dialogue Generation via Augmentative and Contrastive Knowledge-Dialogue/layout.json b/2023/Towards Fewer Hallucinations in Knowledge-Grounded Dialogue Generation via Augmentative and Contrastive Knowledge-Dialogue/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5c5b7340e6798efcaad6e65f40dbdb61c13496c1 --- /dev/null +++ b/2023/Towards Fewer Hallucinations in Knowledge-Grounded Dialogue Generation via Augmentative and Contrastive Knowledge-Dialogue/layout.json @@ -0,0 +1,6579 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 89, + 74, + 505, + 109 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 74, + 505, + 109 + ], + "spans": [ + { + "bbox": [ + 89, + 74, + 505, + 109 + ], + "type": "text", + "content": "Towards Fewer Hallucinations in Knowledge-Grounded Dialogue Generation via Augmentative and Contrastive Knowledge-Dialogue" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 119, + 462, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 119, + 462, + 133 + ], + "spans": [ + { + "bbox": [ + 132, + 119, + 462, + 133 + ], + "type": "text", + "content": "Bin Sun" + }, + { + "bbox": [ + 132, + 119, + 462, + 133 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 119, + 462, + 133 + ], + "type": "text", + "content": ", Yitong Li" + }, + { + "bbox": [ + 132, + 119, + 462, + 133 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 132, + 119, + 462, + 133 + ], + "type": "text", + "content": ", Fei Mi" + }, + { + "bbox": [ + 132, + 119, + 462, + 133 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 119, + 462, + 133 + ], + "type": "text", + "content": ", FanHu Bie" + }, + { + "bbox": [ + 132, + 119, + 462, + 133 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 132, + 119, + 462, + 133 + ], + "type": "text", + "content": ", Yiwei Li" + }, + { + "bbox": [ + 132, + 119, + 462, + 133 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 119, + 462, + 133 + ], + "type": "text", + "content": ", Kan Li" + }, + { + "bbox": [ + 132, + 119, + 462, + 133 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 114, + 134, + 484, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 134, + 484, + 148 + ], + "spans": [ + { + "bbox": [ + 114, + 134, + 484, + 148 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 114, + 134, + 484, + 148 + ], + "type": "text", + "content": "School of Computer Science & Technology, Beijing Institute of Technology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 162, + 148, + 435, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 148, + 435, + 162 + ], + "spans": [ + { + "bbox": [ + 162, + 148, + 435, + 162 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 162, + 148, + 435, + 162 + ], + "type": "text", + "content": " Huawei Noah's Ark Lab " + }, + { + "bbox": [ + 162, + 148, + 435, + 162 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 162, + 148, + 435, + 162 + ], + "type": "text", + "content": "Huawei Technologies Ltd." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 223, + 164, + 373, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 164, + 373, + 175 + ], + "spans": [ + { + "bbox": [ + 223, + 164, + 373, + 175 + ], + "type": "text", + "content": "{binsun,liyiwei,likan}@bit.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 212, + 178, + 384, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 178, + 384, + 190 + ], + "spans": [ + { + "bbox": [ + 212, + 178, + 384, + 190 + ], + "type": "text", + "content": "{liyitong3,mifei2,biefanhu}@huawei.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 155, + 212, + 202, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 212, + 202, + 225 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 202, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 86, + 233, + 274, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 233, + 274, + 568 + ], + "spans": [ + { + "bbox": [ + 86, + 233, + 274, + 568 + ], + "type": "text", + "content": "Existing knowledge-grounded open-domain dialogue generation models often face the hallucination problem, i.e. the dialogue generative model will persist in an inappropriate knowledge and generate responses that inconsistent with the facts. We argue that this problem mainly stems from the polarized optimization objectives and weak knowledge generation ability. To mitigate the hallucination, we take inspiration from human communicating that people will replay euphemistic responses for the unclear or unrecognizable knowledge, and propose an Augmentative and Contrastive Knowledge Dialogue Expansion Framework (ACK-DEF). ACK-DEF constructs the augmentative and contrastive knowledge dialogue samples, which consist of the knowledge of different degrees of errors and the response of manual design, to expand the original training set and smooth the polarized optimization objective that enables models to generate ground-truth with or without gold knowledge. Not only the knowledge, ACK-DEF also provides the tactful responses of manual design corresponding to the incomplete correct knowledge. Experimental results on the Wikipedia of Wizard dataset show that employing the ACK-DEF is effective to alleviate the hallucination problem." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 68, + 576, + 154, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 576, + 154, + 588 + ], + "spans": [ + { + "bbox": [ + 68, + 576, + 154, + 588 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 597, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 597, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 597, + 291, + 772 + ], + "type": "text", + "content": "Recently, Knowledge-Grounded Dialogue Generation draws dramatic attentions in artificial intelligence community. Many efforts incorporate knowledge information to improve the performance of dialogue generation models (Zhou et al., 2018; Dinan et al., 2019; Gopalakrishnan et al., 2019; Kim et al., 2020; Zhao et al., 2020a; Zheng et al., 2021; Zhao et al., 2022a; Bao et al., 2022). However, these methods always face the hallucination problem, that is, the dialogue generation model may insist on an inappropriate knowledge and generate responses that inconsistent with the facts (Rashkin et al., 2021; Zhao et al., 2022a; Dziri et al., 2022)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 213, + 526, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 213, + 526, + 523 + ], + "spans": [ + { + "bbox": [ + 302, + 213, + 526, + 523 + ], + "type": "text", + "content": "We argue that the hallucination problem primarily caused by two aspects: (1) The optimization objective is usually polarized by the gold knowledge dialogue samples and general dialogue samples without knowledge in current knowledge-grounded dialogue datasets (Zhou et al., 2018; Gopalakrishnan et al., 2019; Dinan et al., 2019; Wu et al., 2019; Komeili et al., 2022). Few datasets consider teaching models how to respond when dealing with incomplete correct knowledge, which makes the models tend to believe in the given knowledge, regardless of whether the knowledge is appropriate or not, resulting in hallucination problems. In addition, the knowledge retrieval system tends to extract irrelevant knowledge rather than relevant knowledge when the database is large, aggravating the hallucinations (Reimers and Gurevych, 2021; Liu et al., 2022). (2) The generation of knowledge may also face the hallucination problem and obtain the inappropriate knowledge, leading the generation of hallucination responses (Kim et al., 2020; Zhao et al., 2020a; Liu et al., 2022; Adolphs et al., 2021; Bao et al., 2022)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 529, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 529, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 529, + 526, + 772 + ], + "type": "text", + "content": "To mitigate the hallucination problem, we propose an Augmentative and Contrastive Knowledge Dialogue Expansion Framework (ACK-DEF), which is inspired by human communicating that people will replay euphemistic response for the unrecognizable knowledge. ACK-DEF is proposed to smooth the polarized optimization objective by augmenting training set with augmentative and contrastive knowledge-dialogue samples. Not only the knowledge, we also designed the reply patterns for the knowledge with different level of errors. For this, we propose the augmentative knowledge dialogue expansion (AK), and contrastive knowledge dialogue expansion (CK). AK is proposed to boost the generalization ability of models on knowledge with minor noise. On the contrary, inspired from the contrastive learning paradigm (Cai et al., 2020; Chen et al., 2020a,b; Sun et al., 2021, 2022), CK" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "1741" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 135, + 795, + 459, + 805 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 795, + 459, + 805 + ], + "spans": [ + { + "bbox": [ + 135, + 795, + 459, + 805 + ], + "type": "text", + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 219, + 806, + 375, + 817 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 806, + 375, + 817 + ], + "spans": [ + { + "bbox": [ + 219, + 806, + 375, + 817 + ], + "type": "text", + "content": "Volume 2: Short Papers, pages 1741-1750" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "spans": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "text", + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 68, + 282, + 221 + ], + "blocks": [ + { + "bbox": [ + 77, + 68, + 282, + 221 + ], + "lines": [ + { + "bbox": [ + 77, + 68, + 282, + 221 + ], + "spans": [ + { + "bbox": [ + 77, + 68, + 282, + 221 + ], + "type": "image", + "image_path": "b3d345dea917129cfa22a2ad09d6cca3aea481adcf1f8a95852176977677e589.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 230, + 291, + 304 + ], + "lines": [ + { + "bbox": [ + 67, + 230, + 291, + 304 + ], + "spans": [ + { + "bbox": [ + 67, + 230, + 291, + 304 + ], + "type": "text", + "content": "Figure 1: A diagram of our Augmentative Knowledge dialogue expansion method. We replace different proportion of words in the original knowledge with synonyms to construct incomplete correct knowledge, and design response for different knowledge. We also use prompts to guide the dialogue generation process." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 323, + 290, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 323, + 290, + 390 + ], + "spans": [ + { + "bbox": [ + 67, + 323, + 290, + 390 + ], + "type": "text", + "content": "reconstructs incorrect knowledge and designs euphemistic responses, which aims to push the model learn the reply pattern of incorrect knowledge and a better boundary between correct and incorrect knowledge." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 391, + 290, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 391, + 290, + 500 + ], + "spans": [ + { + "bbox": [ + 67, + 391, + 290, + 500 + ], + "type": "text", + "content": "Contributions: We propose an ACK-DEF to construct new knowledge-dialogue samples that consist of knowledge with different level of errors and manual responses, to soften the training optimization objectives of models, which will mitigate the hallucination. Finally, we conduct extension experiments to show the superior performance of ACK-DEF on alleviating the hallucination." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 509, + 157, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 509, + 157, + 523 + ], + "spans": [ + { + "bbox": [ + 67, + 509, + 157, + 523 + ], + "type": "text", + "content": "2 Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 530, + 291, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 530, + 291, + 705 + ], + "spans": [ + { + "bbox": [ + 67, + 530, + 291, + 705 + ], + "type": "text", + "content": "To mitigate the hallucination problem that caused by the polarized optimization objectives in knowledge grounded dialogue generation, we take inspiration from human communicating, and propose the Augmentative and Contrastive Knowledge Dialogue Expansion Framework (ACK-DEF). Our ACK-DEF aims to soften the polarized training optimization objectives of current knowledge-grounded dialogue generation methods, and guide the dialogue system reply patterns for the knowledge with different level of errors. To achieve this end, we design two effective expansion method, which will be detailed in below." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 715, + 260, + 729 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 715, + 260, + 729 + ], + "spans": [ + { + "bbox": [ + 67, + 715, + 260, + 729 + ], + "type": "text", + "content": "2.1 Augmentative Knowledge Dialogue" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 733, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 733, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 733, + 291, + 772 + ], + "type": "text", + "content": "We propose the Augmentative Knowledge (AK) dialogue expansion to boost the generalization ability of the dialogue model on the knowledge with simi" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 312, + 68, + 518, + 169 + ], + "blocks": [ + { + "bbox": [ + 312, + 68, + 518, + 169 + ], + "lines": [ + { + "bbox": [ + 312, + 68, + 518, + 169 + ], + "spans": [ + { + "bbox": [ + 312, + 68, + 518, + 169 + ], + "type": "image", + "image_path": "19f02e3a689096203540d697ce684f8bf997410565aa7090e0aeacf6395192ce.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 177, + 526, + 285 + ], + "lines": [ + { + "bbox": [ + 302, + 177, + 526, + 285 + ], + "spans": [ + { + "bbox": [ + 302, + 177, + 526, + 285 + ], + "type": "text", + "content": "Figure 2: A diagram of our Contrastive Knowledge dialogue expansion method. We use the antonym to reconstruct the knowledge information and design multiple responses for such knowledge. Since antonyms transform the semantics of the original knowledge, the noise knowledge often contains wrong facts. By this, the model can learn a better boundary between correct and incorrect knowledge, and a safety reply pattern for incorrect knowledge." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 301, + 309, + 526, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 309, + 526, + 661 + ], + "spans": [ + { + "bbox": [ + 301, + 309, + 526, + 661 + ], + "type": "text", + "content": "lar semantics but different expressions, which can prevent the model from being interfered by partialrelevant knowledge retrieved by the retrieval systems (Lian et al., 2019; Zhao et al., 2020b; Hedayatnia et al., 2020; Zheng et al., 2021; Shuster et al., 2021; Komeili et al., 2022). As shown in Figure 1, we employ the synonym data augmentation tool, which replaces words in the original knowledge with synonyms, to reconstruct the knowledge information (Miller, 1995). Considering that the synonym may disrupt the original semantics of new constructed knowledge, we constrain the replace possibility within [0.1,0.2]. Hence, we can obtain the approximate knowledge. Combining this knowledge and the original dialogue, we obtain the \"ak-less sample\". In addition, we also replace " + }, + { + "bbox": [ + 301, + 309, + 526, + 661 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 301, + 309, + 526, + 661 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 301, + 309, + 526, + 661 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 301, + 309, + 526, + 661 + ], + "type": "text", + "content": " words with their synonyms to construct the less similar knowledge. Inspired from prompt learning paradigm (Yao et al., 2022; Valvoda et al., 2022; Zhao et al., 2022b), we manually produce some Prefix-prompts and Post-prompts (see Appendix) to (1) make the new response more tactful for the less similar knowledge; (2) regulate and guide the dialogue generation process of the model. We call the sample consist of less-similar knowledge and designed response as \"ak-more sample\"." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 673, + 485, + 686 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 673, + 485, + 686 + ], + "spans": [ + { + "bbox": [ + 302, + 673, + 485, + 686 + ], + "type": "text", + "content": "2.2 Contrastive Knowledge Dialogue" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 692, + 526, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 692, + 526, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 692, + 526, + 773 + ], + "type": "text", + "content": "We propose the Contrastive Knowledge (CK) dialogue expansion, inspired from the contrastive learning paradigm (Chen et al., 2020b; Cai et al., 2020), not only construct the incorrect knowledge as negative samples for original knowledge, but also build the euphemistic responses as positive" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 780, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 310, + 791 + ], + "type": "text", + "content": "1742" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 66, + 71, + 293, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 71, + 293, + 248 + ], + "spans": [ + { + "bbox": [ + 66, + 71, + 293, + 248 + ], + "type": "text", + "content": "samples for the original response with incorrect knowledge. To help the model learn a boundary between correct and incorrect knowledge, we employ the antonym to make up new incorrect knowledge. For example, given the knowledge \"nintendo was founded on 23 september 1889 ...\", the \"founded\" will be replaced with \"abolish\", which greatly changes the semantics but little changes the expression. After that, we random choose an euphemistic response to replace the original response of the dialogue. Finally, The incorrect knowledge and the replaced euphemistic response are combined as the \"ck-sample\"." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 263, + 213, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 263, + 213, + 276 + ], + "spans": [ + { + "bbox": [ + 67, + 263, + 213, + 276 + ], + "type": "text", + "content": "3 Experiment and Results" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 288, + 190, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 288, + 190, + 301 + ], + "spans": [ + { + "bbox": [ + 67, + 288, + 190, + 301 + ], + "type": "text", + "content": "3.1 Experiment Settings" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 309, + 139, + 321 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 309, + 139, + 321 + ], + "spans": [ + { + "bbox": [ + 67, + 309, + 139, + 321 + ], + "type": "text", + "content": "3.1.1 Dataset" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 328, + 291, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 328, + 291, + 518 + ], + "spans": [ + { + "bbox": [ + 66, + 328, + 291, + 518 + ], + "type": "text", + "content": "We use the Wikipedia of Wizard (WoW) data, a well-established knowledge-grounded open-domain dialogue dataset, for our experiment. We pre-precess the WoW dataset and extract the single-turn knowledge dialogue samples. To evaluate the performance of our method in detail, we perform four test sets: normal, ak-less, ak-more and ck. The normal set is the original test set. And the ak-less, ak-more and ck are the sets consist of ak-less, ak-more and ck samples, respectively. We also follow the settings of WoW data and divide the test set into two groups (seen test and unseen test): the topic of the knowledge in the unseen test set is missing in the training set." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 531, + 143, + 543 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 531, + 143, + 543 + ], + "spans": [ + { + "bbox": [ + 67, + 531, + 143, + 543 + ], + "type": "text", + "content": "3.1.2 Baseline" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 550, + 290, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 550, + 290, + 591 + ], + "spans": [ + { + "bbox": [ + 67, + 550, + 290, + 591 + ], + "type": "text", + "content": "We employ the released PLATO-v1 (Bao et al., 2020) model, a pre-trained dialogue generation model based on UniLM, for our experiment." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 605, + 291, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 605, + 291, + 671 + ], + "spans": [ + { + "bbox": [ + 67, + 605, + 291, + 671 + ], + "type": "text", + "content": "Fine-tuning We directly finetune a model on the original WoW training set. By this, the model can only see gold knowledge dialogue samples and general dialogue samples without knowledge. Hence, we call the fine-tuned model PLATO+GOLD." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 686, + 291, + 739 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 686, + 291, + 739 + ], + "spans": [ + { + "bbox": [ + 67, + 686, + 291, + 739 + ], + "type": "text", + "content": "Fine-tuning with ACK-DEF We finetune the model with the original set and the expansion samples that obtained through ACK-DEF. Thence, we call it PLATO+ACK-DEF." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 71, + 451, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 451, + 83 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 451, + 83 + ], + "type": "text", + "content": "3.1.3 AutoEvaluation Metrics" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 87, + 526, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 87, + 526, + 344 + ], + "spans": [ + { + "bbox": [ + 302, + 87, + 526, + 344 + ], + "type": "text", + "content": "Dialogue Metrics Our primary metrics of interest are Distinct-n (Li et al., 2016), Response Length (Len.) (Csaky et al., 2019), BLEU (Papineni et al., 2002), Embedding-based (Greedy (GRE), Average (AVG), Extrema (EXT)) (Liu et al., 2016), and Coherence (COH) (Xu et al., 2018). Distinct-n evaluates the diversity of generated responses, which is calculated through the ratio of distinct " + }, + { + "bbox": [ + 302, + 87, + 526, + 344 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 302, + 87, + 526, + 344 + ], + "type": "text", + "content": "-grams and all generated " + }, + { + "bbox": [ + 302, + 87, + 526, + 344 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 302, + 87, + 526, + 344 + ], + "type": "text", + "content": "-grams. Len. is the average number of words of all generated responses. BLEU validates the degree of the word-overlap between the generated response and the ground-truth, which denotes the consistence between generated response and ground-truth. Embedding-based metrics (GRE, AVG and EXT) are introduced to evaluate the semantic relationship of generated responses and ground-truth responses, illustrating the consistence in semantic level. COH. mainly assesses the relevance between contexts and generated responses." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 353, + 526, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 353, + 526, + 595 + ], + "spans": [ + { + "bbox": [ + 302, + 353, + 526, + 595 + ], + "type": "text", + "content": "Knowledge Metrics We follow the PLATO(Bao et al., 2020) and use the knowledge precision, recall and f1 scores. These metrics are used to calculate the ratio of tokens that exist in common in ground-truth knowledge and generated responses to tokens in generated responses. \"Recall\" is the average ratio of the number of overlapping tokens in response and knowledge to the number of tokens in knowledge. And \"Precision\" is the average ratio of the number of overlapping tokens to the number of tokens in response. In other words, \"Recall\" indicates how much knowledge information is contained in the response, while \"Precision\" indicates the proportion of knowledge information in the response. Even we involve the negative and incorrect knowledge in response generation, we still use the ground-truth knowledge to calculate the metrics in Table 3,4." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 607, + 477, + 619 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 607, + 477, + 619 + ], + "spans": [ + { + "bbox": [ + 302, + 607, + 477, + 619 + ], + "type": "text", + "content": "3.2 Dialogue Performance Analysis" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 624, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 624, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 624, + 526, + 772 + ], + "type": "text", + "content": "Table 1 and Table 2 report the automatic results on four test sets and four unseen test sets, respectively. In these Tables, it can be observed that (1) the PLATO+ACK-DEF has a competitive performance with PLATO+GOLD on the normal set, which means that the PLATO+ACK-DEF can recognize the golden knowledge and produce appropriate responses. (2) the PLATO+GOLD perform worse than PLATO+ACK-DEF on ak-less, which means that the robustness of the dialogue model only trained with golden knowledge is very weak." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 750, + 291, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 750, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 750, + 291, + 772 + ], + "type": "text", + "content": "1We manually construct some responses, please see Appendix for the detail." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "1743" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 69, + 518, + 190 + ], + "blocks": [ + { + "bbox": [ + 75, + 69, + 518, + 190 + ], + "lines": [ + { + "bbox": [ + 75, + 69, + 518, + 190 + ], + "spans": [ + { + "bbox": [ + 75, + 69, + 518, + 190 + ], + "type": "table", + "html": "
test setDistinct-1/2/3Len.BLEU-1/2/3/4GREAVGEXTCOH
normal0.10680.453313.690.42800.29650.21100.15290.73920.86890.63610.7808
0.09020.398416.200.44280.30170.21090.14990.73660.86830.63300.7878
ak-less0.11940.502413.500.38610.25740.17450.11920.71600.86070.61480.7755
0.08230.353218.780.45020.29820.20150.13800.73070.86960.62930.7948
ak-more0.12340.517412.810.16750.10620.06800.04350.69080.85510.59940.7706
0.06750.294621.830.43580.30010.21230.15420.77450.91510.70930.8098
ck0.11090.477913.230.29650.17790.10800.06570.58380.76220.53730.7712
0.06520.202913.360.42300.27050.18090.12660.65720.83060.61620.8049
", + "image_path": "a67f259174ebbf0e1f9043752f8d5f4f805c1123dc506c842abda3e436d84a4e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 75, + 222, + 518, + 343 + ], + "blocks": [ + { + "bbox": [ + 77, + 198, + 514, + 211 + ], + "lines": [ + { + "bbox": [ + 77, + 198, + 514, + 211 + ], + "spans": [ + { + "bbox": [ + 77, + 198, + 514, + 211 + ], + "type": "text", + "content": "Table 1: The automatic results of PLATO+GOLD (up) and PLATO+ACK-DEF (down) on four test seen sets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 75, + 222, + 518, + 343 + ], + "lines": [ + { + "bbox": [ + 75, + 222, + 518, + 343 + ], + "spans": [ + { + "bbox": [ + 75, + 222, + 518, + 343 + ], + "type": "table", + "html": "
test setDistinct-1/2Len.BLUE-1/2/3/4GREAVGEXTCOH
normal0.05030.242212.430.35160.23310.15820.10900.69880.85680.63060.8094
0.04670.231113.140.34630.22810.15360.10490.69680.85410.63380.8105
ak-less0.09660.391713.390.38710.25650.17240.11640.71430.86000.61220.7836
0.06230.266419.180.44430.29070.19360.13010.72320.86630.61940.8026
ak-more0.10640.444012.710.16520.10460.06680.04260.68880.85380.59800.7797
0.05610.240021.820.43310.29680.20910.15110.76970.91140.70370.8197
ck0.08130.332413.240.30110.18090.11000.06690.58540.76760.54790.7794
0.04650.149013.520.43290.27750.18610.13070.66120.83340.62150.8145
", + "image_path": "c881791a65bea10b76e98faf787df69b22ddb480eee1e130354fa0ebb2ebc361.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 73, + 394, + 285, + 505 + ], + "blocks": [ + { + "bbox": [ + 67, + 352, + 524, + 375 + ], + "lines": [ + { + "bbox": [ + 67, + 352, + 524, + 375 + ], + "spans": [ + { + "bbox": [ + 67, + 352, + 524, + 375 + ], + "type": "text", + "content": "Table 2: The automatic results of PLATO+GOLD (up) and PLATO+ACK-DEF (down) on four test sets with unseen knowledge." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 73, + 394, + 285, + 505 + ], + "lines": [ + { + "bbox": [ + 73, + 394, + 285, + 505 + ], + "spans": [ + { + "bbox": [ + 73, + 394, + 285, + 505 + ], + "type": "table", + "html": "
test setRecallPrecisionF1avg. Dec.
normal0.36070.70090.4546-
ak-less0.28830.55850.3618∇ 0.1026
ak-more0.17520.36320.2228∇ 0.2517
ck0.31930.61330.4003∇ 0.0611
normal0.36950.65380.4520-
ak-less0.32510.56360.3927∇ 0.0647
ak-more0.23350.39830.2775∇ 0.1887
ck0.10650.20410.1337∇ 0.3437
", + "image_path": "a87608cbd02b6c4e4d49c9a8af9a6556762e9af95f42b39c1c2d58fb13c5a0fc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 74, + 565, + 285, + 675 + ], + "blocks": [ + { + "bbox": [ + 67, + 513, + 290, + 550 + ], + "lines": [ + { + "bbox": [ + 67, + 513, + 290, + 550 + ], + "spans": [ + { + "bbox": [ + 67, + 513, + 290, + 550 + ], + "type": "text", + "content": "Table 3: The knowledge correlation results of PLATO+GOLD (up) and PLATO+ACK-DEF (down) on four test sets with seen knowledge." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 74, + 565, + 285, + 675 + ], + "lines": [ + { + "bbox": [ + 74, + 565, + 285, + 675 + ], + "spans": [ + { + "bbox": [ + 74, + 565, + 285, + 675 + ], + "type": "table", + "html": "
test setRecallPrecisionF1avg. Dec.
normal0.37320.74420.4736-
ak-less0.27280.54750.3452∇ 0.1418
ak-more0.16650.36270.2152∇ 0.2822
ck0.30280.60680.3830∇ 0.0995
normal0.36550.68820.4535-
ak-less0.29380.53480.3579∇ 0.1069
ak-more0.20460.37140.2481∇ 0.2277
ck0.08700.18470.1116∇ 0.3747
", + "image_path": "a03261e80b612cb54fd6234ecea69baca569b4fe3c58a8a91e06df4c1a9e1370.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 683, + 290, + 719 + ], + "lines": [ + { + "bbox": [ + 67, + 683, + 290, + 719 + ], + "spans": [ + { + "bbox": [ + 67, + 683, + 290, + 719 + ], + "type": "text", + "content": "Table 4: The knowledge correlation results of PLATO+GOLD (up) and PLATO+ACK-DEF (down) on four test sets with unseen knowledge." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "type": "text", + "content": "Even if the knowledge information only changes by " + }, + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "type": "text", + "content": ", the performance of the model will" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 397, + 525, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 397, + 525, + 491 + ], + "spans": [ + { + "bbox": [ + 302, + 397, + 525, + 491 + ], + "type": "text", + "content": "significantly decline, especially consistency metrics (i.e. BLEU, GRE, AVG and EXT). (3) the PLATO+GOLD achieve better Distinct scores but weaker BLEU and embedding-based scores, which means that the PLATO+GOLD is easy to generate responses that are very different from ground-truth responses, that is, the hallucinations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 508, + 482, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 508, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 302, + 508, + 482, + 521 + ], + "type": "text", + "content": "3.3 Knowledge Correlation Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 529, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 529, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 529, + 526, + 772 + ], + "type": "text", + "content": "Table 3 and Table 4 report the knowledge correlation result of PLATO+GOLD and PLATO+ACK-DEF on four test sets and four test unseen sets, respectively. From these table, we can observe that the performance of PLATO+GOLD is reduced when the given knowledge changed, which illustrates the danger that the model generate responses based on incorrect knowledge. In addition to the above findings, we also observed that the recall, precision and f1 scores of PLATO+ACK-DEF are better than PLATO+GOLD on ak-less and ak-more sets, which demonstrates that using ACK-DEF effectively enhance the model's capability for the similar knowledge information. Moreover, the result of PLATO+ACK-DEF on the ck set is significantly reduced, which shows that the model distinguishes the wrong knowledge constructed with antonyms and gives an appropriate response with" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "1744" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 68, + 289, + 149 + ], + "blocks": [ + { + "bbox": [ + 69, + 68, + 289, + 149 + ], + "lines": [ + { + "bbox": [ + 69, + 68, + 289, + 149 + ], + "spans": [ + { + "bbox": [ + 69, + 68, + 289, + 149 + ], + "type": "table", + "html": "
test setw. GOLD (%)w. ACK-DEF (%)kappa
normal13.0014.000.481
ak-less23.6717.330.513
ak-more33.6724.330.479
ck21.675.670.597
total23.0015.330.552
", + "image_path": "aa53b36205c9ac306af3b3709673d3aa885824f3643542f0bb83eecc70b81481.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 190, + 290, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 190, + 290, + 271 + ], + "spans": [ + { + "bbox": [ + 67, + 190, + 290, + 271 + ], + "type": "text", + "content": "out knowledge (see Table 1 and Table 2 for the effect). These results are inline with our exception that incorporating noised knowledge dialogue samples in training stages can smooth the polarized optimization objective, and mitigate the hallucination problem." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 273, + 291, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 273, + 291, + 516 + ], + "spans": [ + { + "bbox": [ + 69, + 273, + 291, + 516 + ], + "type": "text", + "content": "According to the results of test seen sets and unseen sets), we can notice that the PLATO+ACK-DEF achieves a good performance on ground-truth seen knowledge and a weak performance on ground-truth unseen knowledge. This illustrates that the PLATO+ACK-DEF may doubt the authenticity of unseen given knowledge (even if the knowledge is the ground-truth), and will not fully use it to generate responses. This may alleviate the hallucination, and we believe it is caused by (1) the Augmentative knowledge dialogue introduces similar knowledge to improve the generalization of the model; (2) the Contrastive knowledge dialogue introduces knowledge independent responses, which tell the model to generate responses without knowledge; (3) the ACK-DEF smooths the polarized optimization, which ensures the model not to directly use the given knowledge." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 526, + 185, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 526, + 185, + 538 + ], + "spans": [ + { + "bbox": [ + 67, + 526, + 185, + 538 + ], + "type": "text", + "content": "3.4 Human Evaluation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 543, + 291, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 543, + 291, + 677 + ], + "spans": [ + { + "bbox": [ + 67, + 543, + 291, + 677 + ], + "type": "text", + "content": "To further evaluation the ability of our ACK-DEF on reducing the hallucination problem, we randomly select 400 samples form four test sets, and hire three annotators to do human evaluations by assessing whether the responses generated by PLATO +GOLD and +ACK-DEL have hallucinations. Table 5 reports the results of human evaluation, from which we can notice that the PLATO+ACK-DEF generate less hallucinations than PLATO+GOLD. This shows the effectiveness of our ACK-DEF." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 687, + 148, + 700 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 687, + 148, + 700 + ], + "spans": [ + { + "bbox": [ + 67, + 687, + 148, + 700 + ], + "type": "text", + "content": "3.5 Case Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 705, + 290, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 705, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 705, + 290, + 772 + ], + "type": "text", + "content": "Table 6 shows a generated case of PLATO+GOLD (GOLD) and PLATO+ACK-DEF (ACK-DEF) based on different knowledge information. We can observed that the GOLD is convinced of the given knowledge, regardless of whether the knowledge" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 305, + 68, + 525, + 409 + ], + "blocks": [ + { + "bbox": [ + 99, + 158, + 257, + 169 + ], + "lines": [ + { + "bbox": [ + 99, + 158, + 257, + 169 + ], + "spans": [ + { + "bbox": [ + 99, + 158, + 257, + 169 + ], + "type": "text", + "content": "Table 5: The human evaluation results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 305, + 68, + 525, + 409 + ], + "lines": [ + { + "bbox": [ + 305, + 68, + 525, + 409 + ], + "spans": [ + { + "bbox": [ + 305, + 68, + 525, + 409 + ], + "type": "table", + "html": "
QueryThey really are super cool looking dogs\nthey are so amazing
Normal\nKnowledgegolden retriever are well suited to resi-\ndency in suburban or country environ-\nments.
GOLDI love golden retrievers! They are well\nsuited to residency in suburban or country.
ACK-DEFThey are! They are well suited to suburban\nor country environments.
AK-Less\nKnowledgegolden retriever are good suited to resi-\ndency in suburban operating elbow room\nland environments.
GOLDGolden retriever are great suited to resi-\ndency in suburban operating elbow
ACK-DEFGolden retriever are great suited to subur-\nban environments.
AK-More\nKnowledgegild retriever make up considerably suit\nto residency in suburban surgery common-\nwealth environments.
GOLDgild retrievers make up considerably suit\nto residency in suburban areas.
ACK-DEFNot very clear, maybe they are! They can\nbe found in suburban areas. If it is wrong,\nplease tell me.
CK Knowl-\nedgegolden retriever unequal disadvantage-\ngeously suited to residency in suburban or\ncountry environments.
GOLDThey sure are! I have a golden retriever\nwho is unequally suited to residency in sub-\nurban environments.
ACK-DEFThat sounds pretty good.
", + "image_path": "111b9f5f83d9f38c8bc93b98e00d9c5f0b4f838d443cf77e47a60d152e7e7fd8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 417, + 521, + 428 + ], + "lines": [ + { + "bbox": [ + 306, + 417, + 521, + 428 + ], + "spans": [ + { + "bbox": [ + 306, + 417, + 521, + 428 + ], + "type": "text", + "content": "Table 6: A case of PLATO +GOLD and +ACK-DEF." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 302, + 453, + 525, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 453, + 525, + 534 + ], + "spans": [ + { + "bbox": [ + 302, + 453, + 525, + 534 + ], + "type": "text", + "content": "is appropriate or not, and more easily to copy the knowledge information into responses. Even the GOLD has seen the knowledge topic, it could not remember the knowledge in their parameters. On the contrary, the ACK-DEF has good resistance to incomplete correct knowledge." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 547, + 381, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 547, + 381, + 559 + ], + "spans": [ + { + "bbox": [ + 302, + 547, + 381, + 559 + ], + "type": "text", + "content": "4 Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 301, + 570, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 570, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 301, + 570, + 526, + 772 + ], + "type": "text", + "content": "This paper focuses on the hallucinations caused by polarized optimization objective in knowledge-grounded dialogue generation (KGDG), and proposes an augmentative and contrastive knowledge dialogue expansion framework (ACK-DEF) to mitigate it. The optimization objective of KGDG is to train the model could generate proper response with or without knowledge, which inevitably weaken the model's ability on unrecognized knowledge and lead hallucinations. Therefore, ACK-DEF constructs multiple level knowledge-dialogue samples to soften the optimization objective of KGDG. Extension experimental results show the superior performance of using our methods on dialogue metrics and knowledge correlations." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "1745" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 71, + 131, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 71, + 131, + 84 + ], + "spans": [ + { + "bbox": [ + 68, + 71, + 131, + 84 + ], + "type": "text", + "content": "Limitations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 91, + 199, + 104 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 91, + 199, + 104 + ], + "spans": [ + { + "bbox": [ + 68, + 91, + 199, + 104 + ], + "type": "text", + "content": "Our limitations are as follow:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 81, + 111, + 289, + 715 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 81, + 111, + 289, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 111, + 289, + 177 + ], + "spans": [ + { + "bbox": [ + 81, + 111, + 289, + 177 + ], + "type": "text", + "content": "- Data Scale: This paper only employs the Wikipedia of Wizard dataset, a small scale and well-established knowledge conversation dataset, and lack of the validation on large-scale dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 81, + 186, + 289, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 186, + 289, + 348 + ], + "spans": [ + { + "bbox": [ + 81, + 186, + 289, + 348 + ], + "type": "text", + "content": "- **Backbones:** This paper lacks the evaluating of other knowledge dialogue model on the proposed method. Actually, we have two reasons to employ the PLATO. First, the PLATO can better handle the one-to-many phenomenon, which is suitable for learning our expansion samples. Second, the PLATO is a pre-trained dialogue model, and its performance on knowledge dialogue generation task has been proved. We will evaluating the performance of other knowledge dialogue model on our method for our future work." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 81, + 356, + 289, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 356, + 289, + 518 + ], + "spans": [ + { + "bbox": [ + 81, + 356, + 289, + 518 + ], + "type": "text", + "content": "- Knowledge Expansion Methods: This paper only uses the synonym and antonym to construct the noised knowledge, which lacks of the comparison of using other data augmentation method. Indeed, we use two token-level data augmentation methods (synonym and antonym augmentation) to prove our statements on hallucination problem in knowledge dialogue generation task. Based on this study, we believe that incorporating other data augmentation methods will also mitigate the hallucinations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 81, + 527, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 527, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 81, + 527, + 289, + 715 + ], + "type": "text", + "content": "- Manual Prompts and Responses: This paper designed five prefix prompts, four post-prompts and nineteen euphemistic responses. For " + }, + { + "bbox": [ + 81, + 527, + 289, + 715 + ], + "type": "inline_equation", + "content": "AK" + }, + { + "bbox": [ + 81, + 527, + 289, + 715 + ], + "type": "text", + "content": "-More method, we simply randomly choose one prefix-prompt and one post-prompt and concatenate them with the ground-truth response. This leads to some irregular responses. As for " + }, + { + "bbox": [ + 81, + 527, + 289, + 715 + ], + "type": "inline_equation", + "content": "CK" + }, + { + "bbox": [ + 81, + 527, + 289, + 715 + ], + "type": "text", + "content": " method, we randomly select one euphemistic response for the incorrect knowledge. However, we found that the response may not coherent with the query. We will design more smooth expansion ways to construct more human-like training samples for our future work." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 68, + 725, + 158, + 737 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 725, + 158, + 737 + ], + "spans": [ + { + "bbox": [ + 68, + 725, + 158, + 737 + ], + "type": "text", + "content": "Ethics Statement" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "type": "text", + "content": "We acknowledge and ensure that our study is compatible with the provided Code of Ethics." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 71, + 525, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 525, + 179 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 525, + 179 + ], + "type": "text", + "content": "Knowledge-grounded open-domain dialogue generation is crucial for building a knowledgeable dialogue system, which is beyond the wildest dreams in natural language process field. All our experiments are conducted on public available datasets to avoid ethical concerns. All terms for using these datasets are strictly followed in our study. There are no direct ethical concerns in our research." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 303, + 189, + 400, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 189, + 400, + 202 + ], + "spans": [ + { + "bbox": [ + 303, + 189, + 400, + 202 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 209, + 525, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 209, + 525, + 302 + ], + "spans": [ + { + "bbox": [ + 302, + 209, + 525, + 302 + ], + "type": "text", + "content": "We would like to thank the anonymous reviewers for their constructive comments. This research is supported by Beijing Natural Science Foundation (No.4222037 and L181010) and BIT Research and Innovation Promoting Project (Grant No.2022YCXY021). Kan Li is the corresponding author." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 327, + 362, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 362, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 362, + 338 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 344, + 526, + 772 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 304, + 344, + 524, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 344, + 524, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 344, + 524, + 390 + ], + "type": "text", + "content": "Leonard Adolphs, Kurt Shuster, Jack Urbanek, Arthur Szlam, and Jason Weston. 2021. Reason first, then respond: Modular generation for knowledge-infused dialogue. CoRR, abs/2111.05204." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 396, + 525, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 525, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 525, + 441 + ], + "type": "text", + "content": "Siqi Bao, Huang He, Fan Wang, Hua Wu, and Haifeng Wang. 2020. PLATO: pre-trained dialogue generation model with discrete latent variable. In ACL, pages 85-96. ACL." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 448, + 526, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 448, + 526, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 448, + 526, + 502 + ], + "type": "text", + "content": "Siqi Bao, Huang He, Jun Xu, Hua Lu, Fan Wang, Hua Wu, Han Zhou, Wenquan Wu, Zheng-Yu Niu, and Haifeng Wang. 2022. PLATO-K: internal and external knowledge enhanced dialogue generation. CoRR, abs/2211.00910." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 510, + 525, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 525, + 566 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 525, + 566 + ], + "type": "text", + "content": "Hengyi Cai, Hongshen Chen, Yonghao Song, Zhuoye Ding, Yongjun Bao, Weipeng Yan, and Xiaofang Zhao. 2020. Group-wise contrastive learning for neural dialogue generation. In EMNLP, pages 793-802. Association for Computational Linguistics." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 573, + 525, + 629 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 573, + 525, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 573, + 525, + 629 + ], + "type": "text", + "content": "Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. 2020a. A simple framework for contrastive learning of visual representations. In ICML, volume 119 of Proceedings of Machine Learning Research, pages 1597-1607. PMLR." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 635, + 525, + 670 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 635, + 525, + 670 + ], + "spans": [ + { + "bbox": [ + 304, + 635, + 525, + 670 + ], + "type": "text", + "content": "Xinlei Chen, Haoqi Fan, Ross B. Girshick, and Kaiming He. 2020b. Improved baselines with momentum contrastive learning. CoRR, abs/2003.04297." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 676, + 525, + 720 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 676, + 525, + 720 + ], + "spans": [ + { + "bbox": [ + 304, + 676, + 525, + 720 + ], + "type": "text", + "content": "Richard Csaky, Patrik Purgai, and Gábor Recski. 2019. Improving neural conversational models with entropy-based data filtering. In ACL (1), pages 5650-5669." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 728, + 525, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 728, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 728, + 525, + 772 + ], + "type": "text", + "content": "Emily Dinan, Stephen Roller, Kurt Shuster, Angela Fan, Michael Auli, and Jason Weston. 2019. Wizard of wikipedia: Knowledge-powered conversational agents. In ICLR. OpenReview.net." + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "text", + "content": "1746" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 289, + 772 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 69, + 72, + 289, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 289, + 127 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 289, + 127 + ], + "type": "text", + "content": "Nouha Dziri, Sivan Milton, Mo Yu, Osmar R. Zaïane, and Siva Reddy. 2022. On the origin of hallucinations in conversational models: Is it the datasets or the models? In *NAACL*, pages 5271-5285. Association for Computational Linguistics." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 136, + 289, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 136, + 289, + 200 + ], + "spans": [ + { + "bbox": [ + 69, + 136, + 289, + 200 + ], + "type": "text", + "content": "Karthik Gopalakrishnan, Behnam Hedayatnia, Qinglang Chen, Anna Gottardi, Sanjeev Kwatra, Anu Venkatesh, Raefer Gabriel, and Dilek Hakkani-Tur. 2019. Topical-chat: Towards knowledge-grounded open-domain conversations. In Interspeech, pages 1891-1895. ISCA." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 210, + 289, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 210, + 289, + 277 + ], + "spans": [ + { + "bbox": [ + 69, + 210, + 289, + 277 + ], + "type": "text", + "content": "Behnam Hedayatnia, Karthik Gopalakrishnan, Seokhwan Kim, Yang Liu, Mihail Eric, and Dilek Hakkani-Tur. 2020. Policy-driven neural response generation for knowledge-grounded dialog systems. In INLG, pages 412-421. Association for Computational Linguistics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 285, + 289, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 285, + 289, + 328 + ], + "spans": [ + { + "bbox": [ + 69, + 285, + 289, + 328 + ], + "type": "text", + "content": "Byeongchang Kim, Jaewoo Ahn, and Gunhee Kim. 2020. Sequential latent knowledge selection for knowledge-grounded dialogue. In ICLR. OpenReview.net." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 338, + 289, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 338, + 289, + 381 + ], + "spans": [ + { + "bbox": [ + 69, + 338, + 289, + 381 + ], + "type": "text", + "content": "Mojtaba Komeili, Kurt Shuster, and Jason Weston. 2022. Internet-augmented dialogue generation. In ACL, pages 8460-8478. Association for Computational Linguistics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 391, + 289, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 391, + 289, + 434 + ], + "spans": [ + { + "bbox": [ + 69, + 391, + 289, + 434 + ], + "type": "text", + "content": "Jiwei Li, Michel Galley, Chris Brockett, Jianfeng Gao, and Bill Dolan. 2016. A diversity-promoting objective function for neural conversation models. In HLT-NAACL, pages 110-119." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 443, + 289, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 443, + 289, + 487 + ], + "spans": [ + { + "bbox": [ + 69, + 443, + 289, + 487 + ], + "type": "text", + "content": "Rongzhong Lian, Min Xie, Fan Wang, Jinhua Peng, and Hua Wu. 2019. Learning to select knowledge for response generation in dialog systems. In *IJCAI*, pages 5081-5087. ijcai.org." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 495, + 289, + 560 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 495, + 289, + 560 + ], + "spans": [ + { + "bbox": [ + 69, + 495, + 289, + 560 + ], + "type": "text", + "content": "Chia-Wei Liu, Ryan Lowe, Iulian Serban, Michael Noseworthy, Laurent Charlin, and Joelle Pineau. 2016. How NOT to evaluate your dialogue system: An empirical study of unsupervised evaluation metrics for dialogue response generation. In EMNLP, pages 2122-2132." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 570, + 289, + 636 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 570, + 289, + 636 + ], + "spans": [ + { + "bbox": [ + 69, + 570, + 289, + 636 + ], + "type": "text", + "content": "Zihan Liu, Mostofa Patwary, Ryan Prenger, Shrimai Prabhumoye, Wei Ping, Mohammad Shoeybi, and Bryan Catanzaro. 2022. Multi-stage prompting for knowledgeable dialogue generation. In *Findings of ACL*, pages 1317-1337. Association for Computational Linguistics." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 645, + 289, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 645, + 289, + 666 + ], + "spans": [ + { + "bbox": [ + 69, + 645, + 289, + 666 + ], + "type": "text", + "content": "George A. Miller. 1995. Wordnet: A lexical database for english. Commun. ACM, 38(11):39-41." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 675, + 289, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 675, + 289, + 708 + ], + "spans": [ + { + "bbox": [ + 69, + 675, + 289, + 708 + ], + "type": "text", + "content": "Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In ACL, pages 311-318." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 69, + 717, + 289, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 717, + 289, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 717, + 289, + 772 + ], + "type": "text", + "content": "Hannah Rashkin, David Reitter, Gaurav Singh Tomar, and Dipanjan Das. 2021. Increasing faithfulness in knowledge-grounded dialogue with controllable features. In ACL/IJCNLP, pages 704-718. Association for Computational Linguistics." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 525, + 772 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 304, + 72, + 524, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 524, + 116 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 524, + 116 + ], + "type": "text", + "content": "Nils Reimers and Iryna Gurevych. 2021. The curse of dense low-dimensional information retrieval for large index sizes. In ACL/IJCNLP, pages 605-611. Association for Computational Linguistics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 126, + 525, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 126, + 525, + 182 + ], + "spans": [ + { + "bbox": [ + 304, + 126, + 525, + 182 + ], + "type": "text", + "content": "Kurt Shuster, Spencer Poff, Moya Chen, Douwe Kiela, and Jason Weston. 2021. Retrieval augmentation reduces hallucination in conversation. In *Findings of EMNLP*, pages 3784–3803. Association for Computational Linguistics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 192, + 525, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 192, + 525, + 248 + ], + "spans": [ + { + "bbox": [ + 304, + 192, + 525, + 248 + ], + "type": "text", + "content": "Bin Sun, Shaoxiong Feng, Yiwei Li, Jiamou Liu, and Kan Li. 2021. Generating relevant and coherent dialogue responses using self-separated conditional variational autoencoders. In ACL/IJCNLP, pages 5624-5637. Association for Computational Linguistics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 258, + 525, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 525, + 301 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 525, + 301 + ], + "type": "text", + "content": "Weiwei Sun, Zhengliang Shi, Shen Gao, Pengjie Ren, Maarten de Rijke, and Zhaochun Ren. 2022. Contrastive learning reduces hallucination in conversations. CoRR, abs/2212.10400." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 312, + 525, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 312, + 525, + 379 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 525, + 379 + ], + "type": "text", + "content": "Josef Valvoda, Yimai Fang, and David Vandyke. 2022. Prompting for a conversation: How to control a dialog model? In Proceedings of the Second Workshop on When Creative AI Meets Conversational AI, pages 1-8, Gyeongju, Republic of Korea. Association for Computational Linguistics." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 389, + 525, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 389, + 525, + 444 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 525, + 444 + ], + "type": "text", + "content": "Wenquan Wu, Zhen Guo, Xiangyang Zhou, Hua Wu, Xiyuan Zhang, Rongzhong Lian, and Haifeng Wang. 2019. Proactive human-machine conversation with explicit conversation goal. In ACL, pages 3794-3804. Association for Computational Linguistics." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 454, + 525, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 454, + 525, + 499 + ], + "spans": [ + { + "bbox": [ + 304, + 454, + 525, + 499 + ], + "type": "text", + "content": "Xinnuo Xu, Ondrej Dusek, Ioannis Konstas, and Verena Rieser. 2018. Better conversations by modeling, filtering, and optimizing for coherence and diversity. In EMNLP, pages 3981-3991." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 509, + 525, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 525, + 587 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 525, + 587 + ], + "type": "text", + "content": "Yuan Yao, Bowen Dong, Ao Zhang, Zhengyan Zhang, Ruobing Xie, Zhiyuan Liu, Leyu Lin, Maosong Sun, and Jianyong Wang. 2022. Prompt tuning for discriminative pre-trained language models. In *Findings of the Association for Computational Linguistics: ACL* 2022, pages 3468-3473, Dublin, Ireland. Association for Computational Linguistics." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 597, + 525, + 650 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 597, + 525, + 650 + ], + "spans": [ + { + "bbox": [ + 304, + 597, + 525, + 650 + ], + "type": "text", + "content": "Xueliang Zhao, Tingchen Fu, Chongyang Tao, and Rui Yan. 2022a. There is no standard answer: Knowledge-grounded dialogue generation with adversarial activated multi-reference learning. CoRR, abs/2210.12459." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 662, + 525, + 706 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 662, + 525, + 706 + ], + "spans": [ + { + "bbox": [ + 304, + 662, + 525, + 706 + ], + "type": "text", + "content": "Xueliang Zhao, Wei Wu, Chongyang Tao, Can Xu, Dongyan Zhao, and Rui Yan. 2020a. Low-resource knowledge-grounded dialogue generation. In ICLR. OpenReview.net." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 717, + 525, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 717, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 717, + 525, + 772 + ], + "type": "text", + "content": "Xueliang Zhao, Wei Wu, Can Xu, Chongyang Tao, Dongyan Zhao, and Rui Yan. 2020b. Knowledge-grounded dialogue generation with pre-trained language models. In EMNLP, pages 3377-3390. Association for Computational Linguistics." + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1747" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 290, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 290, + 127 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 290, + 127 + ], + "type": "text", + "content": "Yingxiu Zhao, Yinhe Zheng, Zhiliang Tian, Chang Gao, Bowen Yu, Haiyang Yu, Yongbin Li, Jian Sun, and Nevin L. Zhang. 2022b. Prompt conditioned VAE: enhancing generative replay for lifelong learning in task-oriented dialogue. CoRR, abs/2210.07783." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 137, + 291, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 137, + 291, + 202 + ], + "spans": [ + { + "bbox": [ + 69, + 137, + 291, + 202 + ], + "type": "text", + "content": "Wen Zheng, Natasa Milic-Frayling, and Ke Zhou. 2021. Knowledge-grounded dialogue generation with term-level de-noising. In Findings of ACL/IJCNLP, volume ACL/IJCNLP 2021 of Findings of ACL, pages 2972-2983. Association for Computational Linguistics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 213, + 291, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 213, + 291, + 259 + ], + "spans": [ + { + "bbox": [ + 69, + 213, + 291, + 259 + ], + "type": "text", + "content": "Kangyan Zhou, Shrimai Prabhumoye, and Alan W. Black. 2018. A dataset for document grounded conversations. In EMNLP, pages 708-713. Association for Computational Linguistics." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 69, + 280, + 289, + 364 + ], + "blocks": [ + { + "bbox": [ + 69, + 280, + 289, + 364 + ], + "lines": [ + { + "bbox": [ + 69, + 280, + 289, + 364 + ], + "spans": [ + { + "bbox": [ + 69, + 280, + 289, + 364 + ], + "type": "table", + "html": "
Prefix PromptsPost Prompts
I was thinking that perhaps I am not sure, maybe that Not very clear, maybe Not very clear, perhaps I was thinking that maybeMaybe i am wrong. If I am wrong, please correct me. If I am wrong, please for-give me. If it is wrong, please tell me.
", + "image_path": "b08d270a0120c4730f3e02269fa6eb542e02689cca965b90e7cb474a6f182ffa.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 69, + 410, + 289, + 623 + ], + "blocks": [ + { + "bbox": [ + 83, + 373, + 273, + 386 + ], + "lines": [ + { + "bbox": [ + 83, + 373, + 273, + 386 + ], + "spans": [ + { + "bbox": [ + 83, + 373, + 273, + 386 + ], + "type": "text", + "content": "Table 7: The designed prefix and post prompts." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 410, + 289, + 623 + ], + "lines": [ + { + "bbox": [ + 69, + 410, + 289, + 623 + ], + "spans": [ + { + "bbox": [ + 69, + 410, + 289, + 623 + ], + "type": "table", + "html": "
Euphemistic Responses
Interesting, do you know that? \nThat sounds pretty good. Are there any way to visit? \nOh, I had not heard. \nHmm, I have never heard of that. What is that one about? \nI have never heard. Can you tell me more about it? \nOh, wow, that is remarkable. \nI have never played those, are they fun? \nCan I ask you about it? \nPlease tell me more about that. \nCan you tell me more about that? \nI have never had that. Anything else you can tell me? \nThat's really interesting! But I have never heard of that. \nI literally know nothing about that! \nI have no idea about that. \nI have not heard that one. I will have to check it out. \nHuh, maybe I will need to check that out then. \nOh, I misunderstood then. \nOh, i do not know about that. \nWow, that's a lot! I haven't heard of those.
", + "image_path": "e84dbfc0526ffd29c63be0d7023df44e5b0fb52d5743ea9dd3c01d1e8c378a7d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 85, + 631, + 271, + 644 + ], + "lines": [ + { + "bbox": [ + 85, + 631, + 271, + 644 + ], + "spans": [ + { + "bbox": [ + 85, + 631, + 271, + 644 + ], + "type": "text", + "content": "Table 8: The designed euphemistic responses." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 68, + 669, + 216, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 669, + 216, + 683 + ], + "spans": [ + { + "bbox": [ + 68, + 669, + 216, + 683 + ], + "type": "text", + "content": "A Prefix and Post Prompts" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 692, + 290, + 731 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 692, + 290, + 731 + ], + "spans": [ + { + "bbox": [ + 67, + 692, + 290, + 731 + ], + "type": "text", + "content": "We manually design five prefix prompts and four post prompts, which are shown in Table 7. We discuss below about the prefixes and posts." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 733, + 290, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 733, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 733, + 290, + 772 + ], + "type": "text", + "content": "We designed the prefixes and posts based on the WoW dataset and our daily conversation habits. In WoW dataset, one role is \"0_Wizard\", and the other" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 71, + 526, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 192 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 192 + ], + "type": "text", + "content": "is \"1_Aprentice\". We noticed that the 1_Aprentice will give the sentences such as \"correct my if I am wrong ...\", which is also easy to appear in our daily conversation. Taking inspiration of this, we manually designed the prefixes and posts. Moreover, since the PLATO is pre-trained on conversation datasets, these prefixes may introduce the pre-knowledge that the model learned during the pre-training process." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 193, + 527, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 193, + 527, + 314 + ], + "spans": [ + { + "bbox": [ + 302, + 193, + 527, + 314 + ], + "type": "text", + "content": "In fact, we declare the weakness of our manual prefixes and posts, i.e. direct connections of prefixes, responses, and posts do not fit all contexts. Therefore, we are exploring a new way of constructing replies, such as passing the design prefix, response, post, and context into the large-language-model to rewrite the appropriate response. We believe that better prefixes and posts will lead to more benefits in solving the hallucination problem." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 324, + 446, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 324, + 446, + 338 + ], + "spans": [ + { + "bbox": [ + 302, + 324, + 446, + 338 + ], + "type": "text", + "content": "B Euphemistic Responses" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 346, + 525, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 346, + 525, + 372 + ], + "spans": [ + { + "bbox": [ + 302, + 346, + 525, + 372 + ], + "type": "text", + "content": "We manually design nineteen euphemistic responses, which are shown in Table 8." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 382, + 485, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 382, + 485, + 409 + ], + "spans": [ + { + "bbox": [ + 302, + 382, + 485, + 409 + ], + "type": "text", + "content": "C Dissuasion about the boundary between ak-less and ak-more" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 303, + 418, + 502, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 418, + 502, + 431 + ], + "spans": [ + { + "bbox": [ + 303, + 418, + 502, + 431 + ], + "type": "text", + "content": "Below we provide an example in our dataset:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 440, + 525, + 551 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 316, + 440, + 524, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 440, + 524, + 467 + ], + "spans": [ + { + "bbox": [ + 316, + 440, + 524, + 467 + ], + "type": "text", + "content": "- Ground-truth Knowledge: laziness | thesis (\"thesis\") is a 1996 spanish thriller film." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 476, + 524, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 476, + 524, + 503 + ], + "spans": [ + { + "bbox": [ + 316, + 476, + 524, + 503 + ], + "type": "text", + "content": "- AK-Less Knowledge: acedia | thesis (\"thesis\") is a 1996 spanish thriller film." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 513, + 525, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 513, + 525, + 551 + ], + "spans": [ + { + "bbox": [ + 316, + 513, + 525, + 551 + ], + "type": "text", + "content": "- AK_More Knowledge: laziness | thesis (\"thesis\") personate a 1996 spanish thriller picture show." + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 302, + 562, + 526, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 562, + 526, + 656 + ], + "spans": [ + { + "bbox": [ + 302, + 562, + 526, + 656 + ], + "type": "text", + "content": "It can be noted that the more synonyms are introduced into a sentence, the semantics of the sentence will become more and more different from the original semantics. Therefore, we suppose that replacing at least " + }, + { + "bbox": [ + 302, + 562, + 526, + 656 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 302, + 562, + 526, + 656 + ], + "type": "text", + "content": " of words at once will make a big difference in sentence semantics. Then, we decided the boundary between ak-less and ak-more." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "1748" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "spans": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "type": "text", + "content": "A For every submission:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 106, + 317, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 106, + 317, + 121 + ], + "spans": [ + { + "bbox": [ + 77, + 106, + 317, + 121 + ], + "type": "text", + "content": "A1. Did you describe the limitations of your work?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 121, + 479, + 134 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 121, + 479, + 134 + ], + "spans": [ + { + "bbox": [ + 89, + 121, + 479, + 134 + ], + "type": "text", + "content": "We provide a section of Limitations after the Conclusion and before the Ethics Statement" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 143, + 329, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 143, + 329, + 157 + ], + "spans": [ + { + "bbox": [ + 77, + 143, + 329, + 157 + ], + "type": "text", + "content": "A2. Did you discuss any potential risks of your work?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 158, + 208, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 158, + 208, + 169 + ], + "spans": [ + { + "bbox": [ + 89, + 158, + 208, + 169 + ], + "type": "text", + "content": "Not applicable. Left blank." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 77, + 179, + 414, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 179, + 414, + 192 + ], + "spans": [ + { + "bbox": [ + 77, + 179, + 414, + 192 + ], + "type": "text", + "content": "A3. Do the abstract and introduction summarize the paper's main claims?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 91, + 194, + 97, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 194, + 97, + 204 + ], + "spans": [ + { + "bbox": [ + 91, + 194, + 97, + 204 + ], + "type": "text", + "content": "1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 77, + 215, + 398, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 215, + 398, + 229 + ], + "spans": [ + { + "bbox": [ + 77, + 215, + 398, + 229 + ], + "type": "text", + "content": "A4. Have you used AI writing assistants when working on this paper?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 91, + 230, + 138, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 230, + 138, + 242 + ], + "spans": [ + { + "bbox": [ + 91, + 230, + 138, + 242 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 253, + 291, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 253, + 291, + 266 + ], + "spans": [ + { + "bbox": [ + 69, + 253, + 291, + 266 + ], + "type": "text", + "content": "B Did you use or create scientific artifacts?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 79, + 270, + 127, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 270, + 127, + 283 + ], + "spans": [ + { + "bbox": [ + 79, + 270, + 127, + 283 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 77, + 292, + 315, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 292, + 315, + 306 + ], + "spans": [ + { + "bbox": [ + 77, + 292, + 315, + 306 + ], + "type": "text", + "content": "B1. Did you cite the creators of artifacts you used?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 89, + 306, + 208, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 306, + 208, + 319 + ], + "spans": [ + { + "bbox": [ + 89, + 306, + 208, + 319 + ], + "type": "text", + "content": "Not applicable. Left blank." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 77, + 328, + 463, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 328, + 463, + 342 + ], + "spans": [ + { + "bbox": [ + 77, + 328, + 463, + 342 + ], + "type": "text", + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 90, + 343, + 208, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 343, + 208, + 355 + ], + "spans": [ + { + "bbox": [ + 90, + 343, + 208, + 355 + ], + "type": "text", + "content": "Not applicable. Left blank." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 77, + 364, + 524, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 364, + 524, + 417 + ], + "spans": [ + { + "bbox": [ + 77, + 364, + 524, + 417 + ], + "type": "text", + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 90, + 419, + 208, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 419, + 208, + 432 + ], + "spans": [ + { + "bbox": [ + 90, + 419, + 208, + 432 + ], + "type": "text", + "content": "Not applicable. Left blank." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 77, + 441, + 524, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 441, + 524, + 481 + ], + "spans": [ + { + "bbox": [ + 77, + 441, + 524, + 481 + ], + "type": "text", + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 90, + 482, + 278, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 482, + 278, + 495 + ], + "spans": [ + { + "bbox": [ + 90, + 482, + 278, + 495 + ], + "type": "text", + "content": "We use a publicly well-established dataset." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 77, + 504, + 524, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 504, + 524, + 531 + ], + "spans": [ + { + "bbox": [ + 77, + 504, + 524, + 531 + ], + "type": "text", + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 90, + 532, + 208, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 532, + 208, + 544 + ], + "spans": [ + { + "bbox": [ + 90, + 532, + 208, + 544 + ], + "type": "text", + "content": "Not applicable. Left blank." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 77, + 554, + 524, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 554, + 524, + 622 + ], + "spans": [ + { + "bbox": [ + 77, + 554, + 524, + 622 + ], + "type": "text", + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 91, + 623, + 138, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 623, + 138, + 634 + ], + "spans": [ + { + "bbox": [ + 91, + 623, + 138, + 634 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 69, + 643, + 293, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 643, + 293, + 657 + ], + "spans": [ + { + "bbox": [ + 69, + 643, + 293, + 657 + ], + "type": "text", + "content": "C Did you run computational experiments?" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 80, + 662, + 87, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 662, + 87, + 672 + ], + "spans": [ + { + "bbox": [ + 80, + 662, + 87, + 672 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 77, + 684, + 524, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 684, + 524, + 711 + ], + "spans": [ + { + "bbox": [ + 77, + 684, + 524, + 711 + ], + "type": "text", + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 89, + 712, + 419, + 724 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 712, + 419, + 724 + ], + "spans": [ + { + "bbox": [ + 89, + 712, + 419, + 724 + ], + "type": "text", + "content": "We use the released code and checkpoints. We cite the source of our model." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "spans": [ + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "type": "text", + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "spans": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "text", + "content": "ACL 2023 Responsible NLP Checklist" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1749" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 76, + 71, + 524, + 236 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 76, + 71, + 523, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 71, + 523, + 111 + ], + "spans": [ + { + "bbox": [ + 76, + 71, + 523, + 111 + ], + "type": "text", + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Not applicable. Left blank." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 76, + 121, + 524, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 121, + 524, + 174 + ], + "spans": [ + { + "bbox": [ + 76, + 121, + 524, + 174 + ], + "type": "text", + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Not applicable. Left blank." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 183, + 524, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 183, + 524, + 236 + ], + "spans": [ + { + "bbox": [ + 77, + 183, + 524, + 236 + ], + "type": "text", + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? 3" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 68, + 246, + 522, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 246, + 522, + 275 + ], + "spans": [ + { + "bbox": [ + 68, + 246, + 522, + 275 + ], + "type": "text", + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 76, + 287, + 524, + 539 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 76, + 287, + 524, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 287, + 524, + 327 + ], + "spans": [ + { + "bbox": [ + 76, + 287, + 524, + 327 + ], + "type": "text", + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Not applicable. Left blank." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 76, + 337, + 524, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 337, + 524, + 390 + ], + "spans": [ + { + "bbox": [ + 76, + 337, + 524, + 390 + ], + "type": "text", + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Not applicable. Left blank." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 77, + 399, + 524, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 399, + 524, + 452 + ], + "spans": [ + { + "bbox": [ + 77, + 399, + 524, + 452 + ], + "type": "text", + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 76, + 463, + 519, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 463, + 519, + 489 + ], + "spans": [ + { + "bbox": [ + 76, + 463, + 519, + 489 + ], + "type": "text", + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Not applicable. Left blank." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 76, + 498, + 524, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 498, + 524, + 539 + ], + "spans": [ + { + "bbox": [ + 76, + 498, + 524, + 539 + ], + "type": "text", + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Not applicable. Left blank." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "text", + "content": "1750" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Tracing Linguistic Markers of Influence in a Large Online Organisation/e643dbbe-47ab-405d-a70d-46f0528045a3_content_list.json b/2023/Tracing Linguistic Markers of Influence in a Large Online Organisation/e643dbbe-47ab-405d-a70d-46f0528045a3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..91d4638de3e142078d4ffc9297261d3d7fc3c533 --- /dev/null +++ b/2023/Tracing Linguistic Markers of Influence in a Large Online Organisation/e643dbbe-47ab-405d-a70d-46f0528045a3_content_list.json @@ -0,0 +1,1185 @@ +[ + { + "type": "text", + "text": "Tracing Linguistic Markers of Influence in a Large Online Organisation", + "text_level": 1, + "bbox": [ + 131, + 89, + 878, + 110 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Prashant Khare*, Ravi Shekhar†, Vanja Mladen Karan*, Stephen McQuistin‡, Colin Perkins‡, Ignacio Castro*, Gareth Tyson*§, Patrick G.T. Healey*, Matthew Purver*¶", + "bbox": [ + 117, + 124, + 887, + 156 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "*Queen Mary University of London, †University of Essex, ‡University of Glasgow", + "bbox": [ + 166, + 159, + 833, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{\\S}$ Hong Kong University of Science & Technology, $^{\\ddagger}$ Jožef Stefan Institute", + "bbox": [ + 203, + 175, + 796, + 192 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{p.khare, m.karan, i.castro, g.tyson, p.healey, m.purver}@qmul.ac.uk, r.shekhar@essex.ac.uk, sm@smcquistin.uk, csp@csperkins.org", + "bbox": [ + 156, + 193, + 845, + 225 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 252, + 339, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Social science and psycholinguistic research have shown that power and status affect how people use language in a range of domains. Here, we investigate a similar question in a large, distributed, consensus-driven community – the Internet Engineering Task Force (IETF), a collaborative organisation that develops technical standards for the Internet. Our analysis, based on lexical categories (LIWC) and BERT, shows that participants' levels of influence can be predicted from their email text, and identifies key linguistic differences (e.g., certain LIWC categories, such as WE are positively correlated with high-influence). We also identify the differences in language use for the same person before and after becoming influential1.", + "bbox": [ + 139, + 282, + 460, + 510 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction and Related Work", + "text_level": 1, + "bbox": [ + 114, + 524, + 421, + 539 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Motivation Online communities are rapidly growing. It is imperative to study them to gain a better understanding of online dynamics and important processes such as decision-making. Prior work has shown that influence is an important aspect to consider while analysing online community dynamics (Bapna and Umyarov, 2015; Vega et al., 2021). Social and psycholinguistic research has also revealed that a person's power and status (i.e., influence) is reflected in their usage of language (Nguyen et al., 2016; Guinote, 2017). In this paper, we focus on linguistic traits exhibited by influential people in a large online community.", + "bbox": [ + 112, + 551, + 489, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Detecting meaningful domain-independent indicators of influence is difficult (Danescu-Niculescu-Mizil et al., 2012). Instead, we focus on the Internet Engineering Task Force $^{2}$ (IETF) - a large, open, voluntary, standards developing organisation with over 2M emails between 56k participants over", + "bbox": [ + 112, + 760, + 489, + 857 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "20 years. The decentralised, consensus-oriented nature of the IETF makes it an interesting case study for two reasons. First, compared to the social media data commonly used in similar studies (e.g. Tchokni et al., 2014; Prabhakaran, 2015), IETF emails are usually longer and goal-oriented. Second, the IETF is a decentralised organisation where the decision-making is collaborative and consensus-driven (Bradner, 1996; Resnick, 2014). Hence, the resulting social interactions are very different to alternative email-based datasets such as the Enron Corpus (Klimt and Yang, 2004), or interactions with more rigidly defined power distinctions e.g., admin/users, judges/lawyers (Danescu-Niculescu-Mizil et al., 2012).", + "bbox": [ + 507, + 252, + 885, + 493 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Related Work Most studies of influence either focus on community structure rather than language, or use language indirectly. Urena et al. (2019) give a survey of the former approach. In an example of the latter, Prabhakaran et al. (2014) compare users with different influence in terms of their linguistic similarity or co-adaptation, the increasing similarity of interlocutors to each other in how they use language (see also Danescu-Niculescu-Mizil et al., 2012; Ver Steeg and Galstyan, 2013; Noble and Fernandez, 2015; Kawabata et al., 2016; Buske, 2019; Healey et al., 2023). Some studies (Bramsen et al., 2011; Gilbert, 2012) do focus on modelling influence from text of Enron emails by identifying keywords/phrases that indicate influence. Rosenthal (2014) and Tchokni et al. (2014) extend this approach to other domains, including Twitter, Wikipedia talk pages, and debates, and include a wider range of linguistic markers.", + "bbox": [ + 507, + 505, + 884, + 810 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Goals We focus on discovering linguistic markers of influence in a large consensus-driven standards developing organisation, where the consensus is based on elaborate discussions between participants on mailing lists. To complement this analysis, we also study the linguistic behaviour", + "bbox": [ + 507, + 822, + 884, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Code: https://github.com/sodestream/acl2023-tracing-linguistic-markers", + "bbox": [ + 112, + 866, + 487, + 891 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{2}$ IETF is responsible for producing technical standards for internet infrastructure. https://www.ietf.org/", + "bbox": [ + 112, + 892, + 485, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "82", + "bbox": [ + 489, + 928, + 510, + 939 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics", + "bbox": [ + 226, + 945, + 769, + 957 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Volume 2: Short Papers, pages 82-90", + "bbox": [ + 384, + 958, + 613, + 971 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "July 9-14, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 295, + 972, + 700, + 984 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of participants at different hierarchical levels in IETF, as well as participants in different periods of their participation, similar to Danescu-Niculescu-Mizil et al. (2013), who considered the behaviour of participants as a measure of influence and claim that participants tend to echo the linguistic style of influential individuals. We map this to three research questions: RQ1: How do linguistic traits differ between more and less influential participants? RQ2: How do linguistic traits vary for participants at different levels of the organisation hierarchy? RQ3: How does linguistic behaviour of participants change as they gain influence?", + "bbox": [ + 112, + 84, + 492, + 294 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Methodology", + "text_level": 1, + "bbox": [ + 112, + 311, + 265, + 329 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We aim to understand the correlation between influence, as defined by either network-based centrality metrics (mail-based) or organisational role influence (role-based), and language usage in terms of linguistic traits. For each participant, we consider the emails they sent in a given time period and investigate correlations of certain features of their email text with two different measures of influence.", + "bbox": [ + 112, + 341, + 489, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LIWC Representation Linguistic Inquiry and Word Count (LIWC, Pennebaker et al., 2015) is a well-recognised psycholinguistic lexicon; it provides word counts for 85 different linguistic, psychological, personal concern, and informal language marker categories. Here, we aggregate the word counts within each linguistic category for each participant using the LIWC 2015 dictionary (academic license), and normalise by the total number of emails sent by that participant. Such a normalisation is more appropriate here than normalising by total number of words written, as many IETF emails include long technical sections. This generates a representation of a participant as their mean usage of each LIWC category; while this is a relatively reduced, low-dimensional representation of a person's language, it has the advantage of being interpretable and psychologically well-motivated.", + "bbox": [ + 112, + 485, + 489, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "BERT Representation The LIWC representation ignores context. To allow comparison to more advanced methods, we use the context-dependent representations from BERT (Devlin et al., 2019) via the open-source HuggingFace library (Wolf et al., 2019). The participant-specific BERT representation is calculated by averaging the text representations (last layer CLS vectors) over all their emails.", + "bbox": [ + 112, + 790, + 489, + 919 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3 Experimental Set-up", + "text_level": 1, + "bbox": [ + 507, + 83, + 724, + 101 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dataset The IETF is organised in Working Groups (WGs). Each WG has a technical focus (e.g., HTTP WG for the HTTP protocol) and one or more WG chairs. We use data from two public sources: the IETF mail archives3 and the Datatracker4. The mail archives cover WG activities, meetings, and administration. We gathered 2,106,804 emails from 56,733 email addresses spanning 2000-2019.", + "bbox": [ + 507, + 108, + 884, + 253 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To determine mail-based influence, we use a social graph based on mailing list interactions (messages from one person to another) as built by Khare et al. (2022). We rank participants by their eigenvector centrality, a measure of a node's influence in a graph, and transform rank to a percentile. To determine role-based influence, we used Datatracker for information about WG chairs and their tenure.", + "bbox": [ + 507, + 254, + 884, + 384 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "RQ1 (mail-based influence) We used a 5-year subset of the data for RQ1 due to the computation cost, still giving a reasonable period to observe the participation consistency in the IETF community (McQuistin et al., 2021; Khare et al., 2022). We took data from 2015-2019 with 300,806 emails from 5,363 unique participants. This subset has 212,253 unique tokens, as opposed to 735,605 unique tokens in the whole dataset, and the median length of emails is 504. We calculate the mail-based influence score and LIWC representation for each participant as described. We fit a linear regression model using LIWC representations to predict influence percentile and observe the magnitude and directions of significant coefficients.", + "bbox": [ + 507, + 391, + 884, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "RQ2 (role-based influence) While mail-based influence was crucial to consider the activities of the participants based on the email network, role-based influence is equally crucial as they are involved in organisational decision making. We use the same time period as in RQ1, but here we predict organisational role-based influence. We split the data into two categories: (a) WG chairs and (b) participants who have never been WG chair. We", + "bbox": [ + 507, + 640, + 885, + 787 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "3https://mailarchive.ietf.org/", + "bbox": [ + 532, + 794, + 761, + 808 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "4https://datattracker.ietf.org/ - the administrative database of the IETF, containing metadata about participants and their roles, working groups, document status, etc.", + "bbox": [ + 509, + 808, + 882, + 844 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "5We filter out 104 ambiguous words that are present in LIWC but have technology, security, and network context meaning in IETF, using manually curated lists, for e.g., attack, argument, secure etc. We do this across all RQs.", + "bbox": [ + 509, + 844, + 882, + 892 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "In the top $10\\%$ mail-based influential participants, less than $30\\%$ are WG chairs with significant role-based influence.", + "bbox": [ + 509, + 892, + 882, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "83", + "bbox": [ + 490, + 928, + 510, + 940 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "calculate the LIWC representations for each person, train a logistic regression model to predict category, and observe the LIWC category coefficients.", + "bbox": [ + 112, + 84, + 489, + 134 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "RQ3 (changes in influence) We look at participants who went from low to high influence over time: individuals who had a mail-based influence below the 50th percentile when they joined the IETF, and reached the top 10th percentile at some point. For each participant, we generate two different representations based on two periods — the year of joining and year of reaching the top 10th percentile for the first time — and assign these to two different classes. As in RQ2, we then train a logistic regression model to predict these classes, and examine the coefficients of the LIWC categories.", + "bbox": [ + 112, + 142, + 489, + 336 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "BERT-based variants Our primary purpose is not to assess the predictive power of LIWC representations, but to use them as a tool to characterise linguistic variations in a meaningful way. However, in order to understand their predictive potential, given their relatively simple nature, we compare them to BERT. For these comparisons, we use the BERT representations described in Section 2.", + "bbox": [ + 112, + 346, + 489, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For each RQ we use the same experimental setup as described above. We split the data 80:20 into train and test set and train a prediction model (regression for RQ1 and classification for RQ2 & RQ3). To experiment with both linear and nonlinear models, we include linear and logistic regression and multi layer perceptrons, using implementations from scikit-learn (Pedregosa et al., 2011) with default parameters. As evaluation metrics we used Pearson's $\\rho$ and macro-F1 score.", + "bbox": [ + 112, + 476, + 489, + 637 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4 Results & Discussion", + "text_level": 1, + "bbox": [ + 112, + 650, + 331, + 665 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We now explore the results (see Table 1 for all experiments) and answer our research questions.", + "bbox": [ + 112, + 676, + 487, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1 Answers to RQs", + "text_level": 1, + "bbox": [ + 112, + 720, + 289, + 736 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "RQ1 — The following LIWC categories are significantly correlated $(p < 0.05)$ with higher mail-based influence: WE, INFORMAL, RISK, ADJECTIVE, ANGER, THEY, and BIO. Categories such as NETSPEAK, SEXUAL, HEALTH, DEATH, BODY are correlated with lower influence. This suggests that influential people tend to indicate a collaborative and community-oriented approach with first-person plural (WE) and third-person plural category (THEY) usage. This is consistent with Kacewicz et al. (2014) and Guinote (2017), who show that in", + "bbox": [ + 112, + 741, + 489, + 919 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "fluential people use more first-person plural. They also use more organisational language, which is shown by the negative correlation of informal slang language categories (NETSPEAK, SEXUAL, BODY). We see some unexpected hidden trends due to word ambiguity (e.g., words like 'trust' and 'live'), which are investigated in Section 4.2.", + "bbox": [ + 507, + 84, + 884, + 197 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "RQ2 - From 1, we see that working group (WG) chairs are more social and collaborative, as is shown by WE and SOCIAL categories. This is in line with similar findings from RQ1 and also about leadership engagements from previous works (Strzalkowski et al., 2012; Liu, 2022; Kacewicz et al., 2014; Guinote, 2017; Akstinaite et al., 2020). Also, WG chairs use tentative statements (TENTAT) in discussions, primarily focused on technical feedback and revisions, or suggesting alternatives. Examples showcasing the use of words such as 'or' and 'seems'-", + "bbox": [ + 507, + 197, + 885, + 388 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "seems': \"With the risk of disturbing with statements, but avoiding too many questions: This seems against the goal of reducing headers.\"", + "- 'or': \"Question is do we need to carry around an outer IP-in-IP header for that or not?\"" + ], + "bbox": [ + 531, + 401, + 882, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "RQ3 — From Table 1, we observe that when participants become mail-based influential they are likely to be more descriptive and engaged in immediate state of issues and situations as seen from the correlation of auxiliary verbs (AUXVERB), adverb, risk, and present focus (FOCUSPRESENT). They are also more involved in cognitive processes (COGPROC) as compared to their previous self when they were new to IETF and had little influence.", + "bbox": [ + 507, + 501, + 884, + 645 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.2 Discussion", + "text_level": 1, + "bbox": [ + 507, + 657, + 638, + 671 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To better understand these LIWC categories and what kind of words play a role in the behaviour of individual categories, we calculate the frequency of words in each LIWC category as they appear in the emails. Next, we consider the top 30 most frequent words in each LIWC category and perform regression analysis on mail-based influence for participants, but using only these 30 words as features to generate the participant representation. We conducted this experiment separately for each LIWC category that was significant in the first experiment.", + "bbox": [ + 507, + 677, + 884, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "From the word based analysis we make multiple observations. E.g., words like 'we' imply a collective approach and is strongly correlated with the higher influence. Similarly, the use of word 'well'", + "bbox": [ + 507, + 854, + 884, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "84", + "bbox": [ + 489, + 928, + 510, + 940 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/ba1d152cd34d91f38965aaf03eda1e38dab812747aff9a0568141d4a62fca86e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
RQ1High influenceBIO, WE, INFORMAL, THEY, NEGEMO, ANGER, RISK, ADJECTIVE
Low influenceSEXUAL, DEATH, INGEST, NETSPEAK, HEALTH, FEMALE, BODY, AFFILIATION, CONJ
RQ2WG Chair influenceTENTAT, IPRON, SOCIAL, SEE, FEEL, WE
non-WG ChairCOGPROC, RELativ, AFFILIATION, I, REWARD
RQ3Top 10 percentileADVERB, PREP, ANGER, AUXVERB, MALE, COGPROC, ACHIEV, RISK, FOCUSPRESENT
Below 50th percentileFUNCTION, PPRON, SHEHE, IPRON, NUMBER, CERTAIN, SEXUAL, INFORMAL
", + "bbox": [ + 115, + 80, + 878, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "is standard, such as politely resuming the conversation (e.g., 'well, I agree') or providing an approval over something (e.g., 'this works as well'). These words are well associated with the influential participants. Otherwise, influential participants are generally not observed to be informal and other frequent words (other than 'well') within INFORMAL category do not demonstrate a strong correlation with the growing influence. Also, 'well' is the most frequent word in the INFORMAL category.", + "bbox": [ + 110, + 216, + 487, + 376 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "More influential people (both mail-based and role-based) are also observed to engage more in IETF communities. The conversations can often reflect situations where, as a part of review and feedback process, more influential people highlight limitations in protocol standards, stress on specifics, and compare with existing protocols or previous versions. Several words across different LIWC categories (RISK, NEGEMO, and ADJ) highlight such behaviour, e.g., 'problems', 'before', 'particular', 'specific', 'different', 'most', and 'than'.", + "bbox": [ + 110, + 378, + 489, + 554 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, there are many words with dual sense, like 'trust' which has a very technology specific usage related to network security instead of conversations involving trust issues between individuals or trust in any given situation. Similarly, the word 'live' is related with an application or network being live, instead of its conventional meaning. We also observed that some of the LIWC categories, such as BIO, did not have specific terms that could clearly establish its significance in favour of influential participants (e.g., word 'problems' and 'trust' reflecting the significance for the category RISK), instead such categories had several words with quite weak correlation with influential participants. Such words collectively drifted the weight of the category towards influential participants.", + "bbox": [ + 110, + 557, + 489, + 815 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3 BERT-based results", + "text_level": 1, + "bbox": [ + 112, + 831, + 315, + 845 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We compared the performance of the LIWC- and BERT-based models. Results in Table 2 indicate our LIWC approach is better than an intuitive BERT-based baseline. We hypothesize that the", + "bbox": [ + 112, + 854, + 487, + 919 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/300b79aeb577c665823b70a6c034889c7f191dcec3018e809690c9ddd2614200.jpg", + "table_caption": [ + "Table 1: LIWC categories where $p < {0.05}$ ." + ], + "table_footnote": [], + "table_body": "
LIWCBERT
LRMLPLRMLP
RQ1 (Pearson ρ)0.850*0.852*-0.0180.015
RQ2 (Micro F1)91.2192.4687.6992.21
RQ3 (Micro F1)88.8990.7451.8555.56
", + "bbox": [ + 514, + 212, + 858, + 275 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 2: LIWC vs BERT(\\* $p < 0.0001$", + "bbox": [ + 556, + 285, + 833, + 299 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "reason for this is that LIWC is specialised to detect linguistic markers relevant for this task. Also, to ensure fair comparison, BERT representations were not fine-tuned for the tasks. We believe combining LIWC and BERT might give better representations, especially when dealing with ambiguous words. Curiously, when observing t-SNE (Van der Maaten and Hinton, 2008) projections of participants' BERT representations (Appendix A), we find that low-influence users show a much bigger variation for relevant categories such as WE, NETS-PEAK and INFORMAL. We will investigate this in future.", + "bbox": [ + 507, + 326, + 884, + 533 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5 Conclusions & Future Directions", + "text_level": 1, + "bbox": [ + 507, + 552, + 828, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This paper explores the linguistic patterns of influence in an online collaborative organisation, by analysing the differences between high- and lowinfluence participants. Using two aspects of influence — mail-based, derived from the email network, and organisational role-based — we were able to unfold several traits that differentiate influential participants from others. Many of our findings seem corroborated by studies in organisational theory. We observed that influential people exhibit more collaborative and community-oriented traits, and also stronger signs of engagement in discussions. We also observed that as people go on to become influential participants, they evolve in their communication and are seen to be more engaging and descriptive in their linguistic style. An interesting practical application of our research is identifying and analyzing groups that are dysfunctional in terms of participant roles and their communication patterns (e.g., where the chair is not performing their role). In future work, we will", + "bbox": [ + 507, + 581, + 884, + 919 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "85", + "bbox": [ + 490, + 928, + 510, + 940 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "extend the experiments to study these patterns of interaction in more linguistic depth, between more different roles within an organisation (possibly for multiple collaborative organisations). We will attempt to go beyond lexical count and account for word context.", + "bbox": [ + 112, + 84, + 489, + 180 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6 Limitations", + "text_level": 1, + "bbox": [ + 112, + 193, + 250, + 209 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "One of the main limitations is that we used the standard LIWC-based analysis approach, which is purely lexical and does not take into account the context in which a word appears. Consequently, many words that have very specific senses in the context of the IETF get miscounted as occurrences of LIWC categories. This could be addressed by a more advanced method of mapping to LIWC categories that would account for context. Another limitation is that we manually generated a filtering list containing words specific to the IETF. This list might not be exhaustive enough. Also, we were limited by not conducting an exhaustive hyper-parameter search on our models. We also understand that many emails are longer than 512 tokens (the input limit of the BERT model we used) and might have not been captured completely by our BERT model. However, most of the emails do fit into this BERT sequence length limit. We did not fine tune BERT on the IETF data; this might have given better performance, although it is not clear if it would have given more insight: our main goal is not performance but analyzing/comparing characteristics of existing models. It is also worth highlighting that the data used in this work is strictly in English, and the psycholinguistic categories in LIWC are also based on English language. Hence, this study may be biased and not fully capture variations in linguistic traits that are culturally agnostic.", + "bbox": [ + 112, + 219, + 490, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Ethical considerations — Participation in the IETF is bound by agreements and policies explicitly stating that mailing list discussions and Data-tracker metadata will be made publicly available.7 We use only this publicly available data in our analysis. We have discussed our work with the IETF leadership to confirm that it fits their acceptable use policies. We have also made provisions to manage the data securely, and retain it only as necessary for our work.", + "bbox": [ + 112, + 713, + 489, + 873 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "See both https://www.ieft.org/about/note-well/ and the IETF privacy policy available at https://www.ieft.org/privacy-statement/.", + "bbox": [ + 112, + 879, + 489, + 917 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 509, + 84, + 684, + 99 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We thank the anonymous reviewers for their helpful comments. This work was supported by the UK EPSRC under grants EP/S033564/1 and EP/S036075/1 (Sodestream: Streamlining Social Decision Making for Enhanced Internet Standards). Purver was also supported by the Slovenian Research Agency via research core funding for the programme Knowledge Technologies (P2-0103).", + "bbox": [ + 507, + 109, + 885, + 237 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 510, + 265, + 608, + 279 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Vita Akstinaite, Graham Robinson, and Eugene Sadler-Smith. 2020. Linguistic markers of ceo hubris. Journal of Business Ethics, 167:687-705.", + "Ravi Bapna and Akhmed Umyarov. 2015. Do your online friends make you pay? a randomized field experiment on peer influence in online social networks. Management Science, 61(8):1902-1920.", + "Scott O. Bradner. 1996. The Internet Standards Process - Revision 3. RFC 2026.", + "Philip Bramsen, Martha Escobar-Molano, Ami Patel, and Rafael Alonso. 2011. Extracting social power relationships from natural language. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 773-782.", + "Jakob A Buske. 2019. Linguistic accommodation between leaders and followers. B.S. thesis, University of Twente.", + "Cristian Danescu-Niculescu-Mizil, Lillian Lee, Bo Pang, and Jon Kleinberg. 2012. Echoes of power: Language effects and power differences in social interaction. In Proceedings of the 21st international conference on World Wide Web, pages 699-708.", + "Cristian Danescu-Niculescu-Mizil, Robert West, Dan Jurafsky, Jure Leskovec, and Christopher Potts. 2013. No country for old members: User lifecycle and linguistic change in online communities. In Proceedings of the 22nd international conference on World Wide Web, pages 307-318.", + "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics.", + "Eric Gilbert. 2012. Phrases that signal workplace hierarchy. In Proceedings of the ACM 2012 conference on Computer Supported Cooperative Work, pages 1037-1046." + ], + "bbox": [ + 509, + 286, + 885, + 917 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "86", + "bbox": [ + 490, + 928, + 512, + 940 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Ana Guinote. 2017. How power affects people: Activating, wanting and goal seeking. Annual review of psychology, 68:353-381.", + "Patrick Healey, Prashant Khare, Ignacio Castro, Gareth Tyson, Mladen Karan, Ravi Shekhar, Stephen McGuistin, Colin Perkins, and Matthew Purver. 2023. Power and vulnerability: Managing sensitive language in organisational communication (extended abstract). In ST&D 2023: Annual Meeting of the Society for Text and Discourse, June 28 – June 30, 2023, Oslo, Norway.", + "Ewa Kacewicz, James W Pennebaker, Matthew Davis, Moongee Jeon, and Arthur C Graesser. 2014. Pronoun use reflects standings in social hierarchies. Journal of Language and Social Psychology, 33(2):125-143.", + "Kan Kawabata, Visar Berisha, Anna Scaglione, and Amy LaCross. 2016. A convex model for linguistic influence in group conversations. In *INTERSPEECH*, pages 1442-1446.", + "Prashant Khare, Mladen Karan, Stephen McQuistin, Colin Perkins, Gareth Tyson, Matthew Purver, Patrick Healey, and Ignacio Castro. 2022. The web we weave: Untangling the social graph of the IETF. In Proceedings of the International AAAI Conference on Web and Social Media, volume 16, pages 500-511.", + "Bryan Klimt and Yiming Yang. 2004. The enron corpus: A new dataset for email classification research. In European conference on machine learning, pages 217-226. Springer.", + "Amy H Liu. 2022. Pronoun usage as a measure of power personalization: A general theory with evidence from the chinese-speaking world. British Journal of Political Science, 52(3):1258-1275.", + "Stephen McQuistin, Mladen Karan, Prashant Khare, Colin Perkins, Gareth Tyson, Matthew Purver, Patrick Healey, Waleed Iqbal, Junaid Qadir, and Ignacio Castro. 2021. Characterising the IETF through the lens of RFC deployment. In Proceedings of the 21st ACM Internet Measurement Conference, pages 137-149.", + "Dong Nguyen, A Seza Dogruoz, Carolyn P Rose, and Franciska De Jong. 2016. Computational sociolinguistics: A survey. Computational linguistics, 42(3):537-593.", + "Bill Noble and Raquel Fernandez. 2015. Centre stage: How social network position shapes linguistic coordination. In Proceedings of the 6th workshop on cognitive modeling and computational linguistics, pages 29-38.", + "F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. 2011. Scikit-learn: Machine learning in" + ], + "bbox": [ + 115, + 85, + 489, + 917 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Python. Journal of Machine Learning Research, 12:2825-2830.", + "James W Pennebaker, Ryan L Boyd, Kayla Jordan, and Kate Blackburn. 2015. The development and psychometric properties of LIWC2015. Technical report.", + "Vinodkumar Prabhakaran. 2015. Social power in interactions: Computational analysis and detection of power relations. Ph.D. thesis, Columbia University.", + "Vinodkumar Prabhakaran, Ashima Arora, and Owen Rambow. 2014. Staying on topic: An indicator of power in political debates. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1481-1486.", + "Pete Resnick. 2014. On Consensus and Humming in the IETF. RFC 7282.", + "Sara Rosenthal. 2014. Detecting influencers in social media discussions. XRDS: Crossroads, The ACM Magazine for Students, 21(1):40-45.", + "Tomek Strzalkowski, Samira Shaikh, Ting Liu, George Aaron Broadwell, Jenny Stromer-Galley, Sarah Taylor, Umit Boz, Veena Ravishankar, and Xiaoai Ren. 2012. Modeling leadership and influence in multi-party online discourse. In Proceedings of COLING 2012, pages 2535-2552.", + "Simo Editha Tchokni, Diarmuid O Seaghdha, and Daniele Quercia. 2014. Emoticons and phrases: Status symbols in social media. In Eighth International AAAI Conference on Weblogs and Social Media.", + "Raquel Urena, Gang Kou, Yucheng Dong, Francisco Chiclana, and Enrique Herrera-Viedma. 2019. A review on trust propagation and opinion dynamics in social networks and group decision making frameworks. Information Sciences, 478:461-475.", + "Laurens Van der Maaten and Geoffrey Hinton. 2008. Visualizing data using t-sne. Journal of machine learning research, 9(11).", + "Lea Vega, Andres Mendez-Vazquez, and Armando López-Cuevas. 2021. Probabilistic reasoning system for social influence analysis in online social networks. Social Network Analysis and Mining, 11(1):1-20.", + "Greg Ver Steeg and Aram Galstyan. 2013. Information-theoretic measures of influence based on content dynamics. In Proceedings of the sixth ACM international conference on Web search and data mining, pages 3-12.", + "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. 2019. Huggingface's transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771." + ], + "bbox": [ + 510, + 85, + 880, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "87", + "bbox": [ + 489, + 928, + 510, + 939 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A Appendix A: BERT-based results", + "text_level": 1, + "bbox": [ + 114, + 84, + 438, + 99 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We investigated how BERT representations vary for participants, as per influence, across different significant LIWC categories. For each participant, we calculated the LIWC category representation by averaging the BERT representation of the words in that LIWC category and then projected using t-SNE. As Figures 1, 2 and 3 show, high-influence participants show less variation in their BERT representations compared to lower-influence participants, for the LIWC categories WE, NETSPEAK and INFORMAL respectively.", + "bbox": [ + 112, + 109, + 492, + 287 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5a752473ca314eac08147331e77521d5b8578bc8bbd72742f47aa1a8085d321a.jpg", + "image_caption": [ + "Figure 1: WE category representation" + ], + "image_footnote": [], + "bbox": [ + 137, + 309, + 445, + 481 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/902aca23094a75daff91ba0d8694da8fafb63a22c39a235129ee4b9af051714f.jpg", + "image_caption": [ + "Figure 2: NETSPEAK category representation" + ], + "image_footnote": [], + "bbox": [ + 137, + 552, + 442, + 721 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2f6cb6cddf76e36a4ea8b467d7bac1fe32e2956854c8362d275e5eec60e14255.jpg", + "image_caption": [ + "Figure 3: INFORMAL category representation" + ], + "image_footnote": [], + "bbox": [ + 532, + 403, + 838, + 571 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "88", + "bbox": [ + 490, + 928, + 510, + 940 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A For every submission:", + "bbox": [ + 115, + 107, + 322, + 122 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "A1. Did you describe the limitations of your work? Section 6", + "A2. Did you discuss any potential risks of your work? Section 6 in Limitations section", + "A3. Do the abstract and introduction summarize the paper's main claims? Abstract and Section 1", + "A4. Have you used AI writing assistants when working on this paper? Left blank." + ], + "bbox": [ + 129, + 126, + 695, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "B Did you use or create scientific artifacts?", + "bbox": [ + 115, + 299, + 487, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Section 2", + "bbox": [ + 132, + 321, + 205, + 335 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "B1. Did you cite the creators of artifacts you used? Section 2", + "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Section 2", + "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?" + ], + "bbox": [ + 129, + 346, + 880, + 499 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Section 2 - we used artifact(s) as they they were intended to without any modifications.", + "bbox": [ + 149, + 499, + 786, + 512 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Not applicable. We have used a publicly available dataset as allowed by IETF's privacy statement https://www.ieft.org/privacy-statement/", + "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Section 2 LIWC Representation", + "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Section 3" + ], + "bbox": [ + 129, + 524, + 880, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "C Did you run computational experiments?", + "bbox": [ + 115, + 780, + 492, + 797 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Section 3", + "bbox": [ + 132, + 803, + 205, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? We used default parameters for experiments without parameter tuning.", + "bbox": [ + 129, + 829, + 880, + 876 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "ACL 2023 Responsible NLP Checklist", + "bbox": [ + 132, + 84, + 433, + 99 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.", + "bbox": [ + 112, + 884, + 877, + 908 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "89", + "bbox": [ + 489, + 928, + 510, + 940 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Section 3 (default parameters)", + "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Section 4", + "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Section 2 and Section 3" + ], + "bbox": [ + 129, + 83, + 878, + 281 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank.", + "bbox": [ + 112, + 293, + 877, + 330 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response.", + "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response.", + "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response.", + "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response.", + "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response." + ], + "bbox": [ + 127, + 340, + 878, + 640 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "90", + "bbox": [ + 489, + 928, + 510, + 940 + ], + "page_idx": 8 + } +] \ No newline at end of file diff --git a/2023/Tracing Linguistic Markers of Influence in a Large Online Organisation/e643dbbe-47ab-405d-a70d-46f0528045a3_model.json b/2023/Tracing Linguistic Markers of Influence in a Large Online Organisation/e643dbbe-47ab-405d-a70d-46f0528045a3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0babb72ebb04fdee2f86b0fe869d73c3b5aae4b2 --- /dev/null +++ b/2023/Tracing Linguistic Markers of Influence in a Large Online Organisation/e643dbbe-47ab-405d-a70d-46f0528045a3_model.json @@ -0,0 +1,1747 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.132, + 0.09, + 0.88, + 0.111 + ], + "angle": 0, + "content": "Tracing Linguistic Markers of Influence in a Large Online Organisation" + }, + { + "type": "text", + "bbox": [ + 0.119, + 0.125, + 0.888, + 0.157 + ], + "angle": 0, + "content": "Prashant Khare*, Ravi Shekhar†, Vanja Mladen Karan*, Stephen McQuistin‡, Colin Perkins‡, Ignacio Castro*, Gareth Tyson*§, Patrick G.T. Healey*, Matthew Purver*¶" + }, + { + "type": "text", + "bbox": [ + 0.167, + 0.16, + 0.835, + 0.176 + ], + "angle": 0, + "content": "*Queen Mary University of London, †University of Essex, ‡University of Glasgow" + }, + { + "type": "text", + "bbox": [ + 0.205, + 0.177, + 0.798, + 0.193 + ], + "angle": 0, + "content": "\\(^{\\S}\\)Hong Kong University of Science & Technology, \\(^{\\ddagger}\\)Jožef Stefan Institute" + }, + { + "type": "text", + "bbox": [ + 0.157, + 0.194, + 0.846, + 0.227 + ], + "angle": 0, + "content": "{p.khare, m.karan, i.castro, g.tyson, p.healey, m.purver}@qmul.ac.uk, r.shekhar@essex.ac.uk, sm@smcquistin.uk, csp@csperkins.org" + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.253, + 0.341, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.14, + 0.283, + 0.461, + 0.511 + ], + "angle": 0, + "content": "Social science and psycholinguistic research have shown that power and status affect how people use language in a range of domains. Here, we investigate a similar question in a large, distributed, consensus-driven community – the Internet Engineering Task Force (IETF), a collaborative organisation that develops technical standards for the Internet. Our analysis, based on lexical categories (LIWC) and BERT, shows that participants' levels of influence can be predicted from their email text, and identifies key linguistic differences (e.g., certain LIWC categories, such as WE are positively correlated with high-influence). We also identify the differences in language use for the same person before and after becoming influential1." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.525, + 0.422, + 0.54 + ], + "angle": 0, + "content": "1 Introduction and Related Work" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.552, + 0.49, + 0.761 + ], + "angle": 0, + "content": "Motivation Online communities are rapidly growing. It is imperative to study them to gain a better understanding of online dynamics and important processes such as decision-making. Prior work has shown that influence is an important aspect to consider while analysing online community dynamics (Bapna and Umyarov, 2015; Vega et al., 2021). Social and psycholinguistic research has also revealed that a person's power and status (i.e., influence) is reflected in their usage of language (Nguyen et al., 2016; Guinote, 2017). In this paper, we focus on linguistic traits exhibited by influential people in a large online community." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.762, + 0.49, + 0.858 + ], + "angle": 0, + "content": "Detecting meaningful domain-independent indicators of influence is difficult (Danescu-Niculescu-Mizil et al., 2012). Instead, we focus on the Internet Engineering Task Force\\(^{2}\\) (IETF) - a large, open, voluntary, standards developing organisation with over 2M emails between 56k participants over" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.253, + 0.886, + 0.494 + ], + "angle": 0, + "content": "20 years. The decentralised, consensus-oriented nature of the IETF makes it an interesting case study for two reasons. First, compared to the social media data commonly used in similar studies (e.g. Tchokni et al., 2014; Prabhakaran, 2015), IETF emails are usually longer and goal-oriented. Second, the IETF is a decentralised organisation where the decision-making is collaborative and consensus-driven (Bradner, 1996; Resnick, 2014). Hence, the resulting social interactions are very different to alternative email-based datasets such as the Enron Corpus (Klimt and Yang, 2004), or interactions with more rigidly defined power distinctions e.g., admin/users, judges/lawyers (Danescu-Niculescu-Mizil et al., 2012)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.506, + 0.885, + 0.812 + ], + "angle": 0, + "content": "Related Work Most studies of influence either focus on community structure rather than language, or use language indirectly. Urena et al. (2019) give a survey of the former approach. In an example of the latter, Prabhakaran et al. (2014) compare users with different influence in terms of their linguistic similarity or co-adaptation, the increasing similarity of interlocutors to each other in how they use language (see also Danescu-Niculescu-Mizil et al., 2012; Ver Steeg and Galstyan, 2013; Noble and Fernandez, 2015; Kawabata et al., 2016; Buske, 2019; Healey et al., 2023). Some studies (Bramsen et al., 2011; Gilbert, 2012) do focus on modelling influence from text of Enron emails by identifying keywords/phrases that indicate influence. Rosenthal (2014) and Tchokni et al. (2014) extend this approach to other domains, including Twitter, Wikipedia talk pages, and debates, and include a wider range of linguistic markers." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.823, + 0.885, + 0.919 + ], + "angle": 0, + "content": "Goals We focus on discovering linguistic markers of influence in a large consensus-driven standards developing organisation, where the consensus is based on elaborate discussions between participants on mailing lists. To complement this analysis, we also study the linguistic behaviour" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.868, + 0.488, + 0.892 + ], + "angle": 0, + "content": "\\(^{1}\\)Code: https://github.com/sodestream/acl2023-tracing-linguistic-markers" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.893, + 0.487, + 0.918 + ], + "angle": 0, + "content": "\\(^{2}\\)IETF is responsible for producing technical standards for internet infrastructure. https://www.ietf.org/" + }, + { + "type": "list", + "bbox": [ + 0.114, + 0.868, + 0.488, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.49, + 0.929, + 0.512, + 0.94 + ], + "angle": 0, + "content": "82" + }, + { + "type": "footer", + "bbox": [ + 0.227, + 0.946, + 0.77, + 0.958 + ], + "angle": 0, + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + }, + { + "type": "footer", + "bbox": [ + 0.385, + 0.959, + 0.614, + 0.972 + ], + "angle": 0, + "content": "Volume 2: Short Papers, pages 82-90" + }, + { + "type": "footer", + "bbox": [ + 0.297, + 0.973, + 0.702, + 0.985 + ], + "angle": 0, + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.493, + 0.295 + ], + "angle": 0, + "content": "of participants at different hierarchical levels in IETF, as well as participants in different periods of their participation, similar to Danescu-Niculescu-Mizil et al. (2013), who considered the behaviour of participants as a measure of influence and claim that participants tend to echo the linguistic style of influential individuals. We map this to three research questions: RQ1: How do linguistic traits differ between more and less influential participants? RQ2: How do linguistic traits vary for participants at different levels of the organisation hierarchy? RQ3: How does linguistic behaviour of participants change as they gain influence?" + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.312, + 0.266, + 0.33 + ], + "angle": 0, + "content": "2 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.342, + 0.49, + 0.471 + ], + "angle": 0, + "content": "We aim to understand the correlation between influence, as defined by either network-based centrality metrics (mail-based) or organisational role influence (role-based), and language usage in terms of linguistic traits. For each participant, we consider the emails they sent in a given time period and investigate correlations of certain features of their email text with two different measures of influence." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.486, + 0.49, + 0.777 + ], + "angle": 0, + "content": "LIWC Representation Linguistic Inquiry and Word Count (LIWC, Pennebaker et al., 2015) is a well-recognised psycholinguistic lexicon; it provides word counts for 85 different linguistic, psychological, personal concern, and informal language marker categories. Here, we aggregate the word counts within each linguistic category for each participant using the LIWC 2015 dictionary (academic license), and normalise by the total number of emails sent by that participant. Such a normalisation is more appropriate here than normalising by total number of words written, as many IETF emails include long technical sections. This generates a representation of a participant as their mean usage of each LIWC category; while this is a relatively reduced, low-dimensional representation of a person's language, it has the advantage of being interpretable and psychologically well-motivated." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.791, + 0.49, + 0.92 + ], + "angle": 0, + "content": "BERT Representation The LIWC representation ignores context. To allow comparison to more advanced methods, we use the context-dependent representations from BERT (Devlin et al., 2019) via the open-source HuggingFace library (Wolf et al., 2019). The participant-specific BERT representation is calculated by averaging the text representations (last layer CLS vectors) over all their emails." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.084, + 0.725, + 0.102 + ], + "angle": 0, + "content": "3 Experimental Set-up" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.109, + 0.885, + 0.254 + ], + "angle": 0, + "content": "Dataset The IETF is organised in Working Groups (WGs). Each WG has a technical focus (e.g., HTTP WG for the HTTP protocol) and one or more WG chairs. We use data from two public sources: the IETF mail archives3 and the Datatracker4. The mail archives cover WG activities, meetings, and administration. We gathered 2,106,804 emails from 56,733 email addresses spanning 2000-2019." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.255, + 0.885, + 0.385 + ], + "angle": 0, + "content": "To determine mail-based influence, we use a social graph based on mailing list interactions (messages from one person to another) as built by Khare et al. (2022). We rank participants by their eigenvector centrality, a measure of a node's influence in a graph, and transform rank to a percentile. To determine role-based influence, we used Datatracker for information about WG chairs and their tenure." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.392, + 0.885, + 0.634 + ], + "angle": 0, + "content": "RQ1 (mail-based influence) We used a 5-year subset of the data for RQ1 due to the computation cost, still giving a reasonable period to observe the participation consistency in the IETF community (McQuistin et al., 2021; Khare et al., 2022). We took data from 2015-2019 with 300,806 emails from 5,363 unique participants. This subset has 212,253 unique tokens, as opposed to 735,605 unique tokens in the whole dataset, and the median length of emails is 504. We calculate the mail-based influence score and LIWC representation for each participant as described. We fit a linear regression model using LIWC representations to predict influence percentile and observe the magnitude and directions of significant coefficients." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.642, + 0.887, + 0.788 + ], + "angle": 0, + "content": "RQ2 (role-based influence) While mail-based influence was crucial to consider the activities of the participants based on the email network, role-based influence is equally crucial as they are involved in organisational decision making. We use the same time period as in RQ1, but here we predict organisational role-based influence. We split the data into two categories: (a) WG chairs and (b) participants who have never been WG chair. We" + }, + { + "type": "page_footnote", + "bbox": [ + 0.533, + 0.795, + 0.762, + 0.809 + ], + "angle": 0, + "content": "3https://mailarchive.ietf.org/" + }, + { + "type": "page_footnote", + "bbox": [ + 0.51, + 0.809, + 0.883, + 0.845 + ], + "angle": 0, + "content": "4https://datattracker.ietf.org/ - the administrative database of the IETF, containing metadata about participants and their roles, working groups, document status, etc." + }, + { + "type": "page_footnote", + "bbox": [ + 0.51, + 0.845, + 0.883, + 0.893 + ], + "angle": 0, + "content": "5We filter out 104 ambiguous words that are present in LIWC but have technology, security, and network context meaning in IETF, using manually curated lists, for e.g., attack, argument, secure etc. We do this across all RQs." + }, + { + "type": "page_footnote", + "bbox": [ + 0.51, + 0.894, + 0.883, + 0.919 + ], + "angle": 0, + "content": "In the top \\(10\\%\\) mail-based influential participants, less than \\(30\\%\\) are WG chairs with significant role-based influence." + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.795, + 0.883, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.491, + 0.929, + 0.511, + 0.941 + ], + "angle": 0, + "content": "83" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.114, + 0.085, + 0.49, + 0.135 + ], + "angle": 0, + "content": "calculate the LIWC representations for each person, train a logistic regression model to predict category, and observe the LIWC category coefficients." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.143, + 0.49, + 0.337 + ], + "angle": 0, + "content": "RQ3 (changes in influence) We look at participants who went from low to high influence over time: individuals who had a mail-based influence below the 50th percentile when they joined the IETF, and reached the top 10th percentile at some point. For each participant, we generate two different representations based on two periods — the year of joining and year of reaching the top 10th percentile for the first time — and assign these to two different classes. As in RQ2, we then train a logistic regression model to predict these classes, and examine the coefficients of the LIWC categories." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.347, + 0.49, + 0.475 + ], + "angle": 0, + "content": "BERT-based variants Our primary purpose is not to assess the predictive power of LIWC representations, but to use them as a tool to characterise linguistic variations in a meaningful way. However, in order to understand their predictive potential, given their relatively simple nature, we compare them to BERT. For these comparisons, we use the BERT representations described in Section 2." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.477, + 0.49, + 0.638 + ], + "angle": 0, + "content": "For each RQ we use the same experimental setup as described above. We split the data 80:20 into train and test set and train a prediction model (regression for RQ1 and classification for RQ2 & RQ3). To experiment with both linear and nonlinear models, we include linear and logistic regression and multi layer perceptrons, using implementations from scikit-learn (Pedregosa et al., 2011) with default parameters. As evaluation metrics we used Pearson's \\(\\rho\\) and macro-F1 score." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.651, + 0.332, + 0.667 + ], + "angle": 0, + "content": "4 Results & Discussion" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.677, + 0.489, + 0.71 + ], + "angle": 0, + "content": "We now explore the results (see Table 1 for all experiments) and answer our research questions." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.721, + 0.29, + 0.737 + ], + "angle": 0, + "content": "4.1 Answers to RQs" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.743, + 0.49, + 0.92 + ], + "angle": 0, + "content": "RQ1 — The following LIWC categories are significantly correlated \\((p < 0.05)\\) with higher mail-based influence: WE, INFORMAL, RISK, ADJECTIVE, ANGER, THEY, and BIO. Categories such as NETSPEAK, SEXUAL, HEALTH, DEATH, BODY are correlated with lower influence. This suggests that influential people tend to indicate a collaborative and community-oriented approach with first-person plural (WE) and third-person plural category (THEY) usage. This is consistent with Kacewicz et al. (2014) and Guinote (2017), who show that in" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.198 + ], + "angle": 0, + "content": "fluential people use more first-person plural. They also use more organisational language, which is shown by the negative correlation of informal slang language categories (NETSPEAK, SEXUAL, BODY). We see some unexpected hidden trends due to word ambiguity (e.g., words like 'trust' and 'live'), which are investigated in Section 4.2." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.198, + 0.887, + 0.39 + ], + "angle": 0, + "content": "RQ2 - From 1, we see that working group (WG) chairs are more social and collaborative, as is shown by WE and SOCIAL categories. This is in line with similar findings from RQ1 and also about leadership engagements from previous works (Strzalkowski et al., 2012; Liu, 2022; Kacewicz et al., 2014; Guinote, 2017; Akstinaite et al., 2020). Also, WG chairs use tentative statements (TENTAT) in discussions, primarily focused on technical feedback and revisions, or suggesting alternatives. Examples showcasing the use of words such as 'or' and 'seems'-" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.402, + 0.883, + 0.451 + ], + "angle": 0, + "content": "seems': \"With the risk of disturbing with statements, but avoiding too many questions: This seems against the goal of reducing headers.\"" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.46, + 0.884, + 0.492 + ], + "angle": 0, + "content": "- 'or': \"Question is do we need to carry around an outer IP-in-IP header for that or not?\"" + }, + { + "type": "list", + "bbox": [ + 0.532, + 0.402, + 0.884, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.502, + 0.885, + 0.646 + ], + "angle": 0, + "content": "RQ3 — From Table 1, we observe that when participants become mail-based influential they are likely to be more descriptive and engaged in immediate state of issues and situations as seen from the correlation of auxiliary verbs (AUXVERB), adverb, risk, and present focus (FOCUSPRESENT). They are also more involved in cognitive processes (COGPROC) as compared to their previous self when they were new to IETF and had little influence." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.658, + 0.64, + 0.672 + ], + "angle": 0, + "content": "4.2 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.678, + 0.885, + 0.856 + ], + "angle": 0, + "content": "To better understand these LIWC categories and what kind of words play a role in the behaviour of individual categories, we calculate the frequency of words in each LIWC category as they appear in the emails. Next, we consider the top 30 most frequent words in each LIWC category and perform regression analysis on mail-based influence for participants, but using only these 30 words as features to generate the participant representation. We conducted this experiment separately for each LIWC category that was significant in the first experiment." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.856, + 0.885, + 0.919 + ], + "angle": 0, + "content": "From the word based analysis we make multiple observations. E.g., words like 'we' imply a collective approach and is strongly correlated with the higher influence. Similarly, the use of word 'well'" + }, + { + "type": "page_number", + "bbox": [ + 0.49, + 0.929, + 0.512, + 0.941 + ], + "angle": 0, + "content": "84" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.117, + 0.082, + 0.88, + 0.161 + ], + "angle": 0, + "content": "
RQ1High influenceBIO, WE, INFORMAL, THEY, NEGEMO, ANGER, RISK, ADJECTIVE
Low influenceSEXUAL, DEATH, INGEST, NETSPEAK, HEALTH, FEMALE, BODY, AFFILIATION, CONJ
RQ2WG Chair influenceTENTAT, IPRON, SOCIAL, SEE, FEEL, WE
non-WG ChairCOGPROC, RELativ, AFFILIATION, I, REWARD
RQ3Top 10 percentileADVERB, PREP, ANGER, AUXVERB, MALE, COGPROC, ACHIEV, RISK, FOCUSPRESENT
Below 50th percentileFUNCTION, PPRON, SHEHE, IPRON, NUMBER, CERTAIN, SEXUAL, INFORMAL
" + }, + { + "type": "table_caption", + "bbox": [ + 0.351, + 0.177, + 0.645, + 0.192 + ], + "angle": 0, + "content": "Table 1: LIWC categories where \\( p < {0.05} \\) ." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.217, + 0.489, + 0.378 + ], + "angle": 0, + "content": "is standard, such as politely resuming the conversation (e.g., 'well, I agree') or providing an approval over something (e.g., 'this works as well'). These words are well associated with the influential participants. Otherwise, influential participants are generally not observed to be informal and other frequent words (other than 'well') within INFORMAL category do not demonstrate a strong correlation with the growing influence. Also, 'well' is the most frequent word in the INFORMAL category." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.379, + 0.49, + 0.555 + ], + "angle": 0, + "content": "More influential people (both mail-based and role-based) are also observed to engage more in IETF communities. The conversations can often reflect situations where, as a part of review and feedback process, more influential people highlight limitations in protocol standards, stress on specifics, and compare with existing protocols or previous versions. Several words across different LIWC categories (RISK, NEGEMO, and ADJ) highlight such behaviour, e.g., 'problems', 'before', 'particular', 'specific', 'different', 'most', and 'than'." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.558, + 0.49, + 0.816 + ], + "angle": 0, + "content": "However, there are many words with dual sense, like 'trust' which has a very technology specific usage related to network security instead of conversations involving trust issues between individuals or trust in any given situation. Similarly, the word 'live' is related with an application or network being live, instead of its conventional meaning. We also observed that some of the LIWC categories, such as BIO, did not have specific terms that could clearly establish its significance in favour of influential participants (e.g., word 'problems' and 'trust' reflecting the significance for the category RISK), instead such categories had several words with quite weak correlation with influential participants. Such words collectively drifted the weight of the category towards influential participants." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.832, + 0.317, + 0.846 + ], + "angle": 0, + "content": "4.3 BERT-based results" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.855, + 0.489, + 0.92 + ], + "angle": 0, + "content": "We compared the performance of the LIWC- and BERT-based models. Results in Table 2 indicate our LIWC approach is better than an intuitive BERT-based baseline. We hypothesize that the" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.213, + 0.86, + 0.276 + ], + "angle": 0, + "content": "
LIWCBERT
LRMLPLRMLP
RQ1 (Pearson ρ)0.850*0.852*-0.0180.015
RQ2 (Micro F1)91.2192.4687.6992.21
RQ3 (Micro F1)88.8990.7451.8555.56
" + }, + { + "type": "table_caption", + "bbox": [ + 0.557, + 0.286, + 0.835, + 0.3 + ], + "angle": 0, + "content": "Table 2: LIWC vs BERT(\\* \\(p < 0.0001\\)" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.328, + 0.885, + 0.535 + ], + "angle": 0, + "content": "reason for this is that LIWC is specialised to detect linguistic markers relevant for this task. Also, to ensure fair comparison, BERT representations were not fine-tuned for the tasks. We believe combining LIWC and BERT might give better representations, especially when dealing with ambiguous words. Curiously, when observing t-SNE (Van der Maaten and Hinton, 2008) projections of participants' BERT representations (Appendix A), we find that low-influence users show a much bigger variation for relevant categories such as WE, NETS-PEAK and INFORMAL. We will investigate this in future." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.553, + 0.829, + 0.569 + ], + "angle": 0, + "content": "5 Conclusions & Future Directions" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.582, + 0.885, + 0.92 + ], + "angle": 0, + "content": "This paper explores the linguistic patterns of influence in an online collaborative organisation, by analysing the differences between high- and lowinfluence participants. Using two aspects of influence — mail-based, derived from the email network, and organisational role-based — we were able to unfold several traits that differentiate influential participants from others. Many of our findings seem corroborated by studies in organisational theory. We observed that influential people exhibit more collaborative and community-oriented traits, and also stronger signs of engagement in discussions. We also observed that as people go on to become influential participants, they evolve in their communication and are seen to be more engaging and descriptive in their linguistic style. An interesting practical application of our research is identifying and analyzing groups that are dysfunctional in terms of participant roles and their communication patterns (e.g., where the chair is not performing their role). In future work, we will" + }, + { + "type": "page_number", + "bbox": [ + 0.491, + 0.929, + 0.511, + 0.941 + ], + "angle": 0, + "content": "85" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.49, + 0.181 + ], + "angle": 0, + "content": "extend the experiments to study these patterns of interaction in more linguistic depth, between more different roles within an organisation (possibly for multiple collaborative organisations). We will attempt to go beyond lexical count and account for word context." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.195, + 0.251, + 0.21 + ], + "angle": 0, + "content": "6 Limitations" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.221, + 0.492, + 0.689 + ], + "angle": 0, + "content": "One of the main limitations is that we used the standard LIWC-based analysis approach, which is purely lexical and does not take into account the context in which a word appears. Consequently, many words that have very specific senses in the context of the IETF get miscounted as occurrences of LIWC categories. This could be addressed by a more advanced method of mapping to LIWC categories that would account for context. Another limitation is that we manually generated a filtering list containing words specific to the IETF. This list might not be exhaustive enough. Also, we were limited by not conducting an exhaustive hyper-parameter search on our models. We also understand that many emails are longer than 512 tokens (the input limit of the BERT model we used) and might have not been captured completely by our BERT model. However, most of the emails do fit into this BERT sequence length limit. We did not fine tune BERT on the IETF data; this might have given better performance, although it is not clear if it would have given more insight: our main goal is not performance but analyzing/comparing characteristics of existing models. It is also worth highlighting that the data used in this work is strictly in English, and the psycholinguistic categories in LIWC are also based on English language. Hence, this study may be biased and not fully capture variations in linguistic traits that are culturally agnostic." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.714, + 0.49, + 0.874 + ], + "angle": 0, + "content": "Ethical considerations — Participation in the IETF is bound by agreements and policies explicitly stating that mailing list discussions and Data-tracker metadata will be made publicly available.7 We use only this publicly available data in our analysis. We have discussed our work with the IETF leadership to confirm that it fits their acceptable use policies. We have also made provisions to manage the data securely, and retain it only as necessary for our work." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.881, + 0.49, + 0.919 + ], + "angle": 0, + "content": "See both https://www.ieft.org/about/note-well/ and the IETF privacy policy available at https://www.ieft.org/privacy-statement/." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.085, + 0.685, + 0.101 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.11, + 0.886, + 0.239 + ], + "angle": 0, + "content": "We thank the anonymous reviewers for their helpful comments. This work was supported by the UK EPSRC under grants EP/S033564/1 and EP/S036075/1 (Sodestream: Streamlining Social Decision Making for Enhanced Internet Standards). Purver was also supported by the Slovenian Research Agency via research core funding for the programme Knowledge Technologies (P2-0103)." + }, + { + "type": "title", + "bbox": [ + 0.511, + 0.266, + 0.61, + 0.28 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.287, + 0.886, + 0.328 + ], + "angle": 0, + "content": "Vita Akstinaite, Graham Robinson, and Eugene Sadler-Smith. 2020. Linguistic markers of ceo hubris. Journal of Business Ethics, 167:687-705." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.337, + 0.887, + 0.391 + ], + "angle": 0, + "content": "Ravi Bapna and Akhmed Umyarov. 2015. Do your online friends make you pay? a randomized field experiment on peer influence in online social networks. Management Science, 61(8):1902-1920." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.4, + 0.884, + 0.427 + ], + "angle": 0, + "content": "Scott O. Bradner. 1996. The Internet Standards Process - Revision 3. RFC 2026." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.436, + 0.886, + 0.515 + ], + "angle": 0, + "content": "Philip Bramsen, Martha Escobar-Molano, Ami Patel, and Rafael Alonso. 2011. Extracting social power relationships from natural language. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 773-782." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.525, + 0.886, + 0.565 + ], + "angle": 0, + "content": "Jakob A Buske. 2019. Linguistic accommodation between leaders and followers. B.S. thesis, University of Twente." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.574, + 0.886, + 0.641 + ], + "angle": 0, + "content": "Cristian Danescu-Niculescu-Mizil, Lillian Lee, Bo Pang, and Jon Kleinberg. 2012. Echoes of power: Language effects and power differences in social interaction. In Proceedings of the 21st international conference on World Wide Web, pages 699-708." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.649, + 0.886, + 0.73 + ], + "angle": 0, + "content": "Cristian Danescu-Niculescu-Mizil, Robert West, Dan Jurafsky, Jure Leskovec, and Christopher Potts. 2013. No country for old members: User lifecycle and linguistic change in online communities. In Proceedings of the 22nd international conference on World Wide Web, pages 307-318." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.738, + 0.886, + 0.858 + ], + "angle": 0, + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.866, + 0.886, + 0.919 + ], + "angle": 0, + "content": "Eric Gilbert. 2012. Phrases that signal workplace hierarchy. In Proceedings of the ACM 2012 conference on Computer Supported Cooperative Work, pages 1037-1046." + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.287, + 0.887, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.491, + 0.929, + 0.513, + 0.941 + ], + "angle": 0, + "content": "86" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.126 + ], + "angle": 0, + "content": "Ana Guinote. 2017. How power affects people: Activating, wanting and goal seeking. Annual review of psychology, 68:353-381." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.136, + 0.488, + 0.241 + ], + "angle": 0, + "content": "Patrick Healey, Prashant Khare, Ignacio Castro, Gareth Tyson, Mladen Karan, Ravi Shekhar, Stephen McGuistin, Colin Perkins, and Matthew Purver. 2023. Power and vulnerability: Managing sensitive language in organisational communication (extended abstract). In ST&D 2023: Annual Meeting of the Society for Text and Discourse, June 28 – June 30, 2023, Oslo, Norway." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.25, + 0.488, + 0.315 + ], + "angle": 0, + "content": "Ewa Kacewicz, James W Pennebaker, Matthew Davis, Moongee Jeon, and Arthur C Graesser. 2014. Pronoun use reflects standings in social hierarchies. Journal of Language and Social Psychology, 33(2):125-143." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.326, + 0.488, + 0.378 + ], + "angle": 0, + "content": "Kan Kawabata, Visar Berisha, Anna Scaglione, and Amy LaCross. 2016. A convex model for linguistic influence in group conversations. In *INTERSPEECH*, pages 1442-1446." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.388, + 0.488, + 0.478 + ], + "angle": 0, + "content": "Prashant Khare, Mladen Karan, Stephen McQuistin, Colin Perkins, Gareth Tyson, Matthew Purver, Patrick Healey, and Ignacio Castro. 2022. The web we weave: Untangling the social graph of the IETF. In Proceedings of the International AAAI Conference on Web and Social Media, volume 16, pages 500-511." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.489, + 0.488, + 0.542 + ], + "angle": 0, + "content": "Bryan Klimt and Yiming Yang. 2004. The enron corpus: A new dataset for email classification research. In European conference on machine learning, pages 217-226. Springer." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.551, + 0.488, + 0.604 + ], + "angle": 0, + "content": "Amy H Liu. 2022. Pronoun usage as a measure of power personalization: A general theory with evidence from the chinese-speaking world. British Journal of Political Science, 52(3):1258-1275." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.613, + 0.488, + 0.705 + ], + "angle": 0, + "content": "Stephen McQuistin, Mladen Karan, Prashant Khare, Colin Perkins, Gareth Tyson, Matthew Purver, Patrick Healey, Waleed Iqbal, Junaid Qadir, and Ignacio Castro. 2021. Characterising the IETF through the lens of RFC deployment. In Proceedings of the 21st ACM Internet Measurement Conference, pages 137-149." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.715, + 0.488, + 0.767 + ], + "angle": 0, + "content": "Dong Nguyen, A Seza Dogruoz, Carolyn P Rose, and Franciska De Jong. 2016. Computational sociolinguistics: A survey. Computational linguistics, 42(3):537-593." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.777, + 0.488, + 0.843 + ], + "angle": 0, + "content": "Bill Noble and Raquel Fernandez. 2015. Centre stage: How social network position shapes linguistic coordination. In Proceedings of the 6th workshop on cognitive modeling and computational linguistics, pages 29-38." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.853, + 0.488, + 0.919 + ], + "angle": 0, + "content": "F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. 2011. Scikit-learn: Machine learning in" + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.53, + 0.086, + 0.882, + 0.112 + ], + "angle": 0, + "content": "Python. Journal of Machine Learning Research, 12:2825-2830." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.126, + 0.882, + 0.166 + ], + "angle": 0, + "content": "James W Pennebaker, Ryan L Boyd, Kayla Jordan, and Kate Blackburn. 2015. The development and psychometric properties of LIWC2015. Technical report." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.179, + 0.882, + 0.219 + ], + "angle": 0, + "content": "Vinodkumar Prabhakaran. 2015. Social power in interactions: Computational analysis and detection of power relations. Ph.D. thesis, Columbia University." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.232, + 0.882, + 0.298 + ], + "angle": 0, + "content": "Vinodkumar Prabhakaran, Ashima Arora, and Owen Rambow. 2014. Staying on topic: An indicator of power in political debates. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1481-1486." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.311, + 0.882, + 0.337 + ], + "angle": 0, + "content": "Pete Resnick. 2014. On Consensus and Humming in the IETF. RFC 7282." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.351, + 0.882, + 0.391 + ], + "angle": 0, + "content": "Sara Rosenthal. 2014. Detecting influencers in social media discussions. XRDS: Crossroads, The ACM Magazine for Students, 21(1):40-45." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.404, + 0.882, + 0.483 + ], + "angle": 0, + "content": "Tomek Strzalkowski, Samira Shaikh, Ting Liu, George Aaron Broadwell, Jenny Stromer-Galley, Sarah Taylor, Umit Boz, Veena Ravishankar, and Xiaoai Ren. 2012. Modeling leadership and influence in multi-party online discourse. In Proceedings of COLING 2012, pages 2535-2552." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.496, + 0.882, + 0.549 + ], + "angle": 0, + "content": "Simo Editha Tchokni, Diarmuid O Seaghdha, and Daniele Quercia. 2014. Emoticons and phrases: Status symbols in social media. In Eighth International AAAI Conference on Weblogs and Social Media." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.562, + 0.882, + 0.628 + ], + "angle": 0, + "content": "Raquel Urena, Gang Kou, Yucheng Dong, Francisco Chiclana, and Enrique Herrera-Viedma. 2019. A review on trust propagation and opinion dynamics in social networks and group decision making frameworks. Information Sciences, 478:461-475." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.641, + 0.882, + 0.681 + ], + "angle": 0, + "content": "Laurens Van der Maaten and Geoffrey Hinton. 2008. Visualizing data using t-sne. Journal of machine learning research, 9(11)." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.694, + 0.882, + 0.748 + ], + "angle": 0, + "content": "Lea Vega, Andres Mendez-Vazquez, and Armando López-Cuevas. 2021. Probabilistic reasoning system for social influence analysis in online social networks. Social Network Analysis and Mining, 11(1):1-20." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.761, + 0.882, + 0.827 + ], + "angle": 0, + "content": "Greg Ver Steeg and Aram Galstyan. 2013. Information-theoretic measures of influence based on content dynamics. In Proceedings of the sixth ACM international conference on Web search and data mining, pages 3-12." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.84, + 0.882, + 0.918 + ], + "angle": 0, + "content": "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. 2019. Huggingface's transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.882, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.49, + 0.929, + 0.511, + 0.94 + ], + "angle": 0, + "content": "87" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.115, + 0.085, + 0.44, + 0.101 + ], + "angle": 0, + "content": "A Appendix A: BERT-based results" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.11, + 0.493, + 0.288 + ], + "angle": 0, + "content": "We investigated how BERT representations vary for participants, as per influence, across different significant LIWC categories. For each participant, we calculated the LIWC category representation by averaging the BERT representation of the words in that LIWC category and then projected using t-SNE. As Figures 1, 2 and 3 show, high-influence participants show less variation in their BERT representations compared to lower-influence participants, for the LIWC categories WE, NETSPEAK and INFORMAL respectively." + }, + { + "type": "image", + "bbox": [ + 0.138, + 0.31, + 0.446, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.174, + 0.498, + 0.428, + 0.513 + ], + "angle": 0, + "content": "Figure 1: WE category representation" + }, + { + "type": "image", + "bbox": [ + 0.139, + 0.553, + 0.443, + 0.722 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.147, + 0.741, + 0.454, + 0.756 + ], + "angle": 0, + "content": "Figure 2: NETSPEAK category representation" + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.404, + 0.839, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.542, + 0.592, + 0.851, + 0.607 + ], + "angle": 0, + "content": "Figure 3: INFORMAL category representation" + }, + { + "type": "page_number", + "bbox": [ + 0.491, + 0.929, + 0.511, + 0.941 + ], + "angle": 0, + "content": "88" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.133, + 0.085, + 0.435, + 0.1 + ], + "angle": 0, + "content": "ACL 2023 Responsible NLP Checklist" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.108, + 0.323, + 0.123 + ], + "angle": 0, + "content": "A For every submission:" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.127, + 0.533, + 0.158 + ], + "angle": 0, + "content": "A1. Did you describe the limitations of your work? Section 6" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.169, + 0.553, + 0.201 + ], + "angle": 0, + "content": "A2. Did you discuss any potential risks of your work? Section 6 in Limitations section" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.213, + 0.696, + 0.244 + ], + "angle": 0, + "content": "A3. Do the abstract and introduction summarize the paper's main claims? Abstract and Section 1" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.256, + 0.669, + 0.288 + ], + "angle": 0, + "content": "A4. Have you used AI writing assistants when working on this paper? Left blank." + }, + { + "type": "list", + "bbox": [ + 0.13, + 0.127, + 0.696, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.3, + 0.489, + 0.317 + ], + "angle": 0, + "content": "B Did you use or create scientific artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.322, + 0.206, + 0.336 + ], + "angle": 0, + "content": "Section 2" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.347, + 0.53, + 0.378 + ], + "angle": 0, + "content": "B1. Did you cite the creators of artifacts you used? Section 2" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.39, + 0.779, + 0.421 + ], + "angle": 0, + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Section 2" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.433, + 0.881, + 0.5 + ], + "angle": 0, + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?" + }, + { + "type": "list", + "bbox": [ + 0.131, + 0.347, + 0.881, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.5, + 0.788, + 0.513 + ], + "angle": 0, + "content": "Section 2 - we used artifact(s) as they they were intended to without any modifications." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.525, + 0.881, + 0.605 + ], + "angle": 0, + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Not applicable. We have used a publicly available dataset as allowed by IETF's privacy statement https://www.ieft.org/privacy-statement/" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.615, + 0.881, + 0.663 + ], + "angle": 0, + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Section 2 LIWC Representation" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.674, + 0.881, + 0.769 + ], + "angle": 0, + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Section 3" + }, + { + "type": "list", + "bbox": [ + 0.13, + 0.525, + 0.881, + 0.769 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.781, + 0.494, + 0.798 + ], + "angle": 0, + "content": "C Did you run computational experiments?" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.804, + 0.206, + 0.817 + ], + "angle": 0, + "content": "Section 3" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.83, + 0.881, + 0.877 + ], + "angle": 0, + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? We used default parameters for experiments without parameter tuning." + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.885, + 0.878, + 0.909 + ], + "angle": 0, + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + }, + { + "type": "page_number", + "bbox": [ + 0.49, + 0.929, + 0.511, + 0.941 + ], + "angle": 0, + "content": "89" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.13, + 0.084, + 0.88, + 0.133 + ], + "angle": 0, + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Section 3 (default parameters)" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.143, + 0.88, + 0.207 + ], + "angle": 0, + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Section 4" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.218, + 0.88, + 0.282 + ], + "angle": 0, + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Section 2 and Section 3" + }, + { + "type": "list", + "bbox": [ + 0.13, + 0.084, + 0.88, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.294, + 0.878, + 0.331 + ], + "angle": 0, + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.341, + 0.88, + 0.388 + ], + "angle": 0, + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.4, + 0.88, + 0.464 + ], + "angle": 0, + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.476, + 0.88, + 0.54 + ], + "angle": 0, + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.55, + 0.873, + 0.582 + ], + "angle": 0, + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.593, + 0.88, + 0.641 + ], + "angle": 0, + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response." + }, + { + "type": "list", + "bbox": [ + 0.128, + 0.341, + 0.88, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.49, + 0.929, + 0.512, + 0.941 + ], + "angle": 0, + "content": "90" + } + ] +] \ No newline at end of file diff --git a/2023/Tracing Linguistic Markers of Influence in a Large Online Organisation/e643dbbe-47ab-405d-a70d-46f0528045a3_origin.pdf b/2023/Tracing Linguistic Markers of Influence in a Large Online Organisation/e643dbbe-47ab-405d-a70d-46f0528045a3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bd967d376a5d6071a3a778101cfc1e3b0eb6bd27 --- /dev/null +++ b/2023/Tracing Linguistic Markers of Influence in a Large Online Organisation/e643dbbe-47ab-405d-a70d-46f0528045a3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0de11b1762ce708d859f3fec14ef57da9d0f64a65dd28d4c54744cb56fdf9146 +size 302372 diff --git a/2023/Tracing Linguistic Markers of Influence in a Large Online Organisation/full.md b/2023/Tracing Linguistic Markers of Influence in a Large Online Organisation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ca749be5980edfcd32747a5350f81d34ba37e29a --- /dev/null +++ b/2023/Tracing Linguistic Markers of Influence in a Large Online Organisation/full.md @@ -0,0 +1,204 @@ +# Tracing Linguistic Markers of Influence in a Large Online Organisation + +Prashant Khare*, Ravi Shekhar†, Vanja Mladen Karan*, Stephen McQuistin‡, Colin Perkins‡, Ignacio Castro*, Gareth Tyson*§, Patrick G.T. Healey*, Matthew Purver*¶ + +*Queen Mary University of London, †University of Essex, ‡University of Glasgow + +$^{\S}$ Hong Kong University of Science & Technology, $^{\ddagger}$ Jožef Stefan Institute + +{p.khare, m.karan, i.castro, g.tyson, p.healey, m.purver}@qmul.ac.uk, r.shekhar@essex.ac.uk, sm@smcquistin.uk, csp@csperkins.org + +# Abstract + +Social science and psycholinguistic research have shown that power and status affect how people use language in a range of domains. Here, we investigate a similar question in a large, distributed, consensus-driven community – the Internet Engineering Task Force (IETF), a collaborative organisation that develops technical standards for the Internet. Our analysis, based on lexical categories (LIWC) and BERT, shows that participants' levels of influence can be predicted from their email text, and identifies key linguistic differences (e.g., certain LIWC categories, such as WE are positively correlated with high-influence). We also identify the differences in language use for the same person before and after becoming influential1. + +# 1 Introduction and Related Work + +Motivation Online communities are rapidly growing. It is imperative to study them to gain a better understanding of online dynamics and important processes such as decision-making. Prior work has shown that influence is an important aspect to consider while analysing online community dynamics (Bapna and Umyarov, 2015; Vega et al., 2021). Social and psycholinguistic research has also revealed that a person's power and status (i.e., influence) is reflected in their usage of language (Nguyen et al., 2016; Guinote, 2017). In this paper, we focus on linguistic traits exhibited by influential people in a large online community. + +Detecting meaningful domain-independent indicators of influence is difficult (Danescu-Niculescu-Mizil et al., 2012). Instead, we focus on the Internet Engineering Task Force $^{2}$ (IETF) - a large, open, voluntary, standards developing organisation with over 2M emails between 56k participants over + +20 years. The decentralised, consensus-oriented nature of the IETF makes it an interesting case study for two reasons. First, compared to the social media data commonly used in similar studies (e.g. Tchokni et al., 2014; Prabhakaran, 2015), IETF emails are usually longer and goal-oriented. Second, the IETF is a decentralised organisation where the decision-making is collaborative and consensus-driven (Bradner, 1996; Resnick, 2014). Hence, the resulting social interactions are very different to alternative email-based datasets such as the Enron Corpus (Klimt and Yang, 2004), or interactions with more rigidly defined power distinctions e.g., admin/users, judges/lawyers (Danescu-Niculescu-Mizil et al., 2012). + +Related Work Most studies of influence either focus on community structure rather than language, or use language indirectly. Urena et al. (2019) give a survey of the former approach. In an example of the latter, Prabhakaran et al. (2014) compare users with different influence in terms of their linguistic similarity or co-adaptation, the increasing similarity of interlocutors to each other in how they use language (see also Danescu-Niculescu-Mizil et al., 2012; Ver Steeg and Galstyan, 2013; Noble and Fernandez, 2015; Kawabata et al., 2016; Buske, 2019; Healey et al., 2023). Some studies (Bramsen et al., 2011; Gilbert, 2012) do focus on modelling influence from text of Enron emails by identifying keywords/phrases that indicate influence. Rosenthal (2014) and Tchokni et al. (2014) extend this approach to other domains, including Twitter, Wikipedia talk pages, and debates, and include a wider range of linguistic markers. + +Goals We focus on discovering linguistic markers of influence in a large consensus-driven standards developing organisation, where the consensus is based on elaborate discussions between participants on mailing lists. To complement this analysis, we also study the linguistic behaviour + +of participants at different hierarchical levels in IETF, as well as participants in different periods of their participation, similar to Danescu-Niculescu-Mizil et al. (2013), who considered the behaviour of participants as a measure of influence and claim that participants tend to echo the linguistic style of influential individuals. We map this to three research questions: RQ1: How do linguistic traits differ between more and less influential participants? RQ2: How do linguistic traits vary for participants at different levels of the organisation hierarchy? RQ3: How does linguistic behaviour of participants change as they gain influence? + +# 2 Methodology + +We aim to understand the correlation between influence, as defined by either network-based centrality metrics (mail-based) or organisational role influence (role-based), and language usage in terms of linguistic traits. For each participant, we consider the emails they sent in a given time period and investigate correlations of certain features of their email text with two different measures of influence. + +LIWC Representation Linguistic Inquiry and Word Count (LIWC, Pennebaker et al., 2015) is a well-recognised psycholinguistic lexicon; it provides word counts for 85 different linguistic, psychological, personal concern, and informal language marker categories. Here, we aggregate the word counts within each linguistic category for each participant using the LIWC 2015 dictionary (academic license), and normalise by the total number of emails sent by that participant. Such a normalisation is more appropriate here than normalising by total number of words written, as many IETF emails include long technical sections. This generates a representation of a participant as their mean usage of each LIWC category; while this is a relatively reduced, low-dimensional representation of a person's language, it has the advantage of being interpretable and psychologically well-motivated. + +BERT Representation The LIWC representation ignores context. To allow comparison to more advanced methods, we use the context-dependent representations from BERT (Devlin et al., 2019) via the open-source HuggingFace library (Wolf et al., 2019). The participant-specific BERT representation is calculated by averaging the text representations (last layer CLS vectors) over all their emails. + +# 3 Experimental Set-up + +Dataset The IETF is organised in Working Groups (WGs). Each WG has a technical focus (e.g., HTTP WG for the HTTP protocol) and one or more WG chairs. We use data from two public sources: the IETF mail archives3 and the Datatracker4. The mail archives cover WG activities, meetings, and administration. We gathered 2,106,804 emails from 56,733 email addresses spanning 2000-2019. + +To determine mail-based influence, we use a social graph based on mailing list interactions (messages from one person to another) as built by Khare et al. (2022). We rank participants by their eigenvector centrality, a measure of a node's influence in a graph, and transform rank to a percentile. To determine role-based influence, we used Datatracker for information about WG chairs and their tenure. + +RQ1 (mail-based influence) We used a 5-year subset of the data for RQ1 due to the computation cost, still giving a reasonable period to observe the participation consistency in the IETF community (McQuistin et al., 2021; Khare et al., 2022). We took data from 2015-2019 with 300,806 emails from 5,363 unique participants. This subset has 212,253 unique tokens, as opposed to 735,605 unique tokens in the whole dataset, and the median length of emails is 504. We calculate the mail-based influence score and LIWC representation for each participant as described. We fit a linear regression model using LIWC representations to predict influence percentile and observe the magnitude and directions of significant coefficients. + +RQ2 (role-based influence) While mail-based influence was crucial to consider the activities of the participants based on the email network, role-based influence is equally crucial as they are involved in organisational decision making. We use the same time period as in RQ1, but here we predict organisational role-based influence. We split the data into two categories: (a) WG chairs and (b) participants who have never been WG chair. We + +calculate the LIWC representations for each person, train a logistic regression model to predict category, and observe the LIWC category coefficients. + +RQ3 (changes in influence) We look at participants who went from low to high influence over time: individuals who had a mail-based influence below the 50th percentile when they joined the IETF, and reached the top 10th percentile at some point. For each participant, we generate two different representations based on two periods — the year of joining and year of reaching the top 10th percentile for the first time — and assign these to two different classes. As in RQ2, we then train a logistic regression model to predict these classes, and examine the coefficients of the LIWC categories. + +BERT-based variants Our primary purpose is not to assess the predictive power of LIWC representations, but to use them as a tool to characterise linguistic variations in a meaningful way. However, in order to understand their predictive potential, given their relatively simple nature, we compare them to BERT. For these comparisons, we use the BERT representations described in Section 2. + +For each RQ we use the same experimental setup as described above. We split the data 80:20 into train and test set and train a prediction model (regression for RQ1 and classification for RQ2 & RQ3). To experiment with both linear and nonlinear models, we include linear and logistic regression and multi layer perceptrons, using implementations from scikit-learn (Pedregosa et al., 2011) with default parameters. As evaluation metrics we used Pearson's $\rho$ and macro-F1 score. + +# 4 Results & Discussion + +We now explore the results (see Table 1 for all experiments) and answer our research questions. + +# 4.1 Answers to RQs + +RQ1 — The following LIWC categories are significantly correlated $(p < 0.05)$ with higher mail-based influence: WE, INFORMAL, RISK, ADJECTIVE, ANGER, THEY, and BIO. Categories such as NETSPEAK, SEXUAL, HEALTH, DEATH, BODY are correlated with lower influence. This suggests that influential people tend to indicate a collaborative and community-oriented approach with first-person plural (WE) and third-person plural category (THEY) usage. This is consistent with Kacewicz et al. (2014) and Guinote (2017), who show that in + +fluential people use more first-person plural. They also use more organisational language, which is shown by the negative correlation of informal slang language categories (NETSPEAK, SEXUAL, BODY). We see some unexpected hidden trends due to word ambiguity (e.g., words like 'trust' and 'live'), which are investigated in Section 4.2. + +RQ2 - From 1, we see that working group (WG) chairs are more social and collaborative, as is shown by WE and SOCIAL categories. This is in line with similar findings from RQ1 and also about leadership engagements from previous works (Strzalkowski et al., 2012; Liu, 2022; Kacewicz et al., 2014; Guinote, 2017; Akstinaite et al., 2020). Also, WG chairs use tentative statements (TENTAT) in discussions, primarily focused on technical feedback and revisions, or suggesting alternatives. Examples showcasing the use of words such as 'or' and 'seems'- + +seems': "With the risk of disturbing with statements, but avoiding too many questions: This seems against the goal of reducing headers." +- 'or': "Question is do we need to carry around an outer IP-in-IP header for that or not?" + +RQ3 — From Table 1, we observe that when participants become mail-based influential they are likely to be more descriptive and engaged in immediate state of issues and situations as seen from the correlation of auxiliary verbs (AUXVERB), adverb, risk, and present focus (FOCUSPRESENT). They are also more involved in cognitive processes (COGPROC) as compared to their previous self when they were new to IETF and had little influence. + +# 4.2 Discussion + +To better understand these LIWC categories and what kind of words play a role in the behaviour of individual categories, we calculate the frequency of words in each LIWC category as they appear in the emails. Next, we consider the top 30 most frequent words in each LIWC category and perform regression analysis on mail-based influence for participants, but using only these 30 words as features to generate the participant representation. We conducted this experiment separately for each LIWC category that was significant in the first experiment. + +From the word based analysis we make multiple observations. E.g., words like 'we' imply a collective approach and is strongly correlated with the higher influence. Similarly, the use of word 'well' + +
RQ1High influenceBIO, WE, INFORMAL, THEY, NEGEMO, ANGER, RISK, ADJECTIVE
Low influenceSEXUAL, DEATH, INGEST, NETSPEAK, HEALTH, FEMALE, BODY, AFFILIATION, CONJ
RQ2WG Chair influenceTENTAT, IPRON, SOCIAL, SEE, FEEL, WE
non-WG ChairCOGPROC, RELativ, AFFILIATION, I, REWARD
RQ3Top 10 percentileADVERB, PREP, ANGER, AUXVERB, MALE, COGPROC, ACHIEV, RISK, FOCUSPRESENT
Below 50th percentileFUNCTION, PPRON, SHEHE, IPRON, NUMBER, CERTAIN, SEXUAL, INFORMAL
+ +is standard, such as politely resuming the conversation (e.g., 'well, I agree') or providing an approval over something (e.g., 'this works as well'). These words are well associated with the influential participants. Otherwise, influential participants are generally not observed to be informal and other frequent words (other than 'well') within INFORMAL category do not demonstrate a strong correlation with the growing influence. Also, 'well' is the most frequent word in the INFORMAL category. + +More influential people (both mail-based and role-based) are also observed to engage more in IETF communities. The conversations can often reflect situations where, as a part of review and feedback process, more influential people highlight limitations in protocol standards, stress on specifics, and compare with existing protocols or previous versions. Several words across different LIWC categories (RISK, NEGEMO, and ADJ) highlight such behaviour, e.g., 'problems', 'before', 'particular', 'specific', 'different', 'most', and 'than'. + +However, there are many words with dual sense, like 'trust' which has a very technology specific usage related to network security instead of conversations involving trust issues between individuals or trust in any given situation. Similarly, the word 'live' is related with an application or network being live, instead of its conventional meaning. We also observed that some of the LIWC categories, such as BIO, did not have specific terms that could clearly establish its significance in favour of influential participants (e.g., word 'problems' and 'trust' reflecting the significance for the category RISK), instead such categories had several words with quite weak correlation with influential participants. Such words collectively drifted the weight of the category towards influential participants. + +# 4.3 BERT-based results + +We compared the performance of the LIWC- and BERT-based models. Results in Table 2 indicate our LIWC approach is better than an intuitive BERT-based baseline. We hypothesize that the + +Table 1: LIWC categories where $p < {0.05}$ . + +
LIWCBERT
LRMLPLRMLP
RQ1 (Pearson ρ)0.850*0.852*-0.0180.015
RQ2 (Micro F1)91.2192.4687.6992.21
RQ3 (Micro F1)88.8990.7451.8555.56
+ +Table 2: LIWC vs BERT(\* $p < 0.0001$ + +reason for this is that LIWC is specialised to detect linguistic markers relevant for this task. Also, to ensure fair comparison, BERT representations were not fine-tuned for the tasks. We believe combining LIWC and BERT might give better representations, especially when dealing with ambiguous words. Curiously, when observing t-SNE (Van der Maaten and Hinton, 2008) projections of participants' BERT representations (Appendix A), we find that low-influence users show a much bigger variation for relevant categories such as WE, NETS-PEAK and INFORMAL. We will investigate this in future. + +# 5 Conclusions & Future Directions + +This paper explores the linguistic patterns of influence in an online collaborative organisation, by analysing the differences between high- and lowinfluence participants. Using two aspects of influence — mail-based, derived from the email network, and organisational role-based — we were able to unfold several traits that differentiate influential participants from others. Many of our findings seem corroborated by studies in organisational theory. We observed that influential people exhibit more collaborative and community-oriented traits, and also stronger signs of engagement in discussions. We also observed that as people go on to become influential participants, they evolve in their communication and are seen to be more engaging and descriptive in their linguistic style. An interesting practical application of our research is identifying and analyzing groups that are dysfunctional in terms of participant roles and their communication patterns (e.g., where the chair is not performing their role). In future work, we will + +extend the experiments to study these patterns of interaction in more linguistic depth, between more different roles within an organisation (possibly for multiple collaborative organisations). We will attempt to go beyond lexical count and account for word context. + +# 6 Limitations + +One of the main limitations is that we used the standard LIWC-based analysis approach, which is purely lexical and does not take into account the context in which a word appears. Consequently, many words that have very specific senses in the context of the IETF get miscounted as occurrences of LIWC categories. This could be addressed by a more advanced method of mapping to LIWC categories that would account for context. Another limitation is that we manually generated a filtering list containing words specific to the IETF. This list might not be exhaustive enough. Also, we were limited by not conducting an exhaustive hyper-parameter search on our models. We also understand that many emails are longer than 512 tokens (the input limit of the BERT model we used) and might have not been captured completely by our BERT model. However, most of the emails do fit into this BERT sequence length limit. We did not fine tune BERT on the IETF data; this might have given better performance, although it is not clear if it would have given more insight: our main goal is not performance but analyzing/comparing characteristics of existing models. It is also worth highlighting that the data used in this work is strictly in English, and the psycholinguistic categories in LIWC are also based on English language. Hence, this study may be biased and not fully capture variations in linguistic traits that are culturally agnostic. + +Ethical considerations — Participation in the IETF is bound by agreements and policies explicitly stating that mailing list discussions and Data-tracker metadata will be made publicly available.7 We use only this publicly available data in our analysis. We have discussed our work with the IETF leadership to confirm that it fits their acceptable use policies. We have also made provisions to manage the data securely, and retain it only as necessary for our work. + +See both https://www.ieft.org/about/note-well/ and the IETF privacy policy available at https://www.ieft.org/privacy-statement/. + +# Acknowledgements + +We thank the anonymous reviewers for their helpful comments. This work was supported by the UK EPSRC under grants EP/S033564/1 and EP/S036075/1 (Sodestream: Streamlining Social Decision Making for Enhanced Internet Standards). Purver was also supported by the Slovenian Research Agency via research core funding for the programme Knowledge Technologies (P2-0103). + +# References + +Vita Akstinaite, Graham Robinson, and Eugene Sadler-Smith. 2020. Linguistic markers of ceo hubris. Journal of Business Ethics, 167:687-705. +Ravi Bapna and Akhmed Umyarov. 2015. Do your online friends make you pay? a randomized field experiment on peer influence in online social networks. Management Science, 61(8):1902-1920. +Scott O. Bradner. 1996. The Internet Standards Process - Revision 3. RFC 2026. +Philip Bramsen, Martha Escobar-Molano, Ami Patel, and Rafael Alonso. 2011. Extracting social power relationships from natural language. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 773-782. +Jakob A Buske. 2019. Linguistic accommodation between leaders and followers. B.S. thesis, University of Twente. +Cristian Danescu-Niculescu-Mizil, Lillian Lee, Bo Pang, and Jon Kleinberg. 2012. Echoes of power: Language effects and power differences in social interaction. In Proceedings of the 21st international conference on World Wide Web, pages 699-708. +Cristian Danescu-Niculescu-Mizil, Robert West, Dan Jurafsky, Jure Leskovec, and Christopher Potts. 2013. No country for old members: User lifecycle and linguistic change in online communities. In Proceedings of the 22nd international conference on World Wide Web, pages 307-318. +Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics. +Eric Gilbert. 2012. Phrases that signal workplace hierarchy. In Proceedings of the ACM 2012 conference on Computer Supported Cooperative Work, pages 1037-1046. + +Ana Guinote. 2017. How power affects people: Activating, wanting and goal seeking. Annual review of psychology, 68:353-381. +Patrick Healey, Prashant Khare, Ignacio Castro, Gareth Tyson, Mladen Karan, Ravi Shekhar, Stephen McGuistin, Colin Perkins, and Matthew Purver. 2023. Power and vulnerability: Managing sensitive language in organisational communication (extended abstract). In ST&D 2023: Annual Meeting of the Society for Text and Discourse, June 28 – June 30, 2023, Oslo, Norway. +Ewa Kacewicz, James W Pennebaker, Matthew Davis, Moongee Jeon, and Arthur C Graesser. 2014. Pronoun use reflects standings in social hierarchies. Journal of Language and Social Psychology, 33(2):125-143. +Kan Kawabata, Visar Berisha, Anna Scaglione, and Amy LaCross. 2016. A convex model for linguistic influence in group conversations. In *INTERSPEECH*, pages 1442-1446. +Prashant Khare, Mladen Karan, Stephen McQuistin, Colin Perkins, Gareth Tyson, Matthew Purver, Patrick Healey, and Ignacio Castro. 2022. The web we weave: Untangling the social graph of the IETF. In Proceedings of the International AAAI Conference on Web and Social Media, volume 16, pages 500-511. +Bryan Klimt and Yiming Yang. 2004. The enron corpus: A new dataset for email classification research. In European conference on machine learning, pages 217-226. Springer. +Amy H Liu. 2022. Pronoun usage as a measure of power personalization: A general theory with evidence from the chinese-speaking world. British Journal of Political Science, 52(3):1258-1275. +Stephen McQuistin, Mladen Karan, Prashant Khare, Colin Perkins, Gareth Tyson, Matthew Purver, Patrick Healey, Waleed Iqbal, Junaid Qadir, and Ignacio Castro. 2021. Characterising the IETF through the lens of RFC deployment. In Proceedings of the 21st ACM Internet Measurement Conference, pages 137-149. +Dong Nguyen, A Seza Dogruoz, Carolyn P Rose, and Franciska De Jong. 2016. Computational sociolinguistics: A survey. Computational linguistics, 42(3):537-593. +Bill Noble and Raquel Fernandez. 2015. Centre stage: How social network position shapes linguistic coordination. In Proceedings of the 6th workshop on cognitive modeling and computational linguistics, pages 29-38. +F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. 2011. Scikit-learn: Machine learning in + +Python. Journal of Machine Learning Research, 12:2825-2830. +James W Pennebaker, Ryan L Boyd, Kayla Jordan, and Kate Blackburn. 2015. The development and psychometric properties of LIWC2015. Technical report. +Vinodkumar Prabhakaran. 2015. Social power in interactions: Computational analysis and detection of power relations. Ph.D. thesis, Columbia University. +Vinodkumar Prabhakaran, Ashima Arora, and Owen Rambow. 2014. Staying on topic: An indicator of power in political debates. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1481-1486. +Pete Resnick. 2014. On Consensus and Humming in the IETF. RFC 7282. +Sara Rosenthal. 2014. Detecting influencers in social media discussions. XRDS: Crossroads, The ACM Magazine for Students, 21(1):40-45. +Tomek Strzalkowski, Samira Shaikh, Ting Liu, George Aaron Broadwell, Jenny Stromer-Galley, Sarah Taylor, Umit Boz, Veena Ravishankar, and Xiaoai Ren. 2012. Modeling leadership and influence in multi-party online discourse. In Proceedings of COLING 2012, pages 2535-2552. +Simo Editha Tchokni, Diarmuid O Seaghdha, and Daniele Quercia. 2014. Emoticons and phrases: Status symbols in social media. In Eighth International AAAI Conference on Weblogs and Social Media. +Raquel Urena, Gang Kou, Yucheng Dong, Francisco Chiclana, and Enrique Herrera-Viedma. 2019. A review on trust propagation and opinion dynamics in social networks and group decision making frameworks. Information Sciences, 478:461-475. +Laurens Van der Maaten and Geoffrey Hinton. 2008. Visualizing data using t-sne. Journal of machine learning research, 9(11). +Lea Vega, Andres Mendez-Vazquez, and Armando López-Cuevas. 2021. Probabilistic reasoning system for social influence analysis in online social networks. Social Network Analysis and Mining, 11(1):1-20. +Greg Ver Steeg and Aram Galstyan. 2013. Information-theoretic measures of influence based on content dynamics. In Proceedings of the sixth ACM international conference on Web search and data mining, pages 3-12. +Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. 2019. Huggingface's transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771. + +# A Appendix A: BERT-based results + +We investigated how BERT representations vary for participants, as per influence, across different significant LIWC categories. For each participant, we calculated the LIWC category representation by averaging the BERT representation of the words in that LIWC category and then projected using t-SNE. As Figures 1, 2 and 3 show, high-influence participants show less variation in their BERT representations compared to lower-influence participants, for the LIWC categories WE, NETSPEAK and INFORMAL respectively. + +![](images/5a752473ca314eac08147331e77521d5b8578bc8bbd72742f47aa1a8085d321a.jpg) +Figure 1: WE category representation + +![](images/902aca23094a75daff91ba0d8694da8fafb63a22c39a235129ee4b9af051714f.jpg) +Figure 2: NETSPEAK category representation + +![](images/2f6cb6cddf76e36a4ea8b467d7bac1fe32e2956854c8362d275e5eec60e14255.jpg) +Figure 3: INFORMAL category representation + +A For every submission: + +A1. Did you describe the limitations of your work? Section 6 +A2. Did you discuss any potential risks of your work? Section 6 in Limitations section +A3. Do the abstract and introduction summarize the paper's main claims? Abstract and Section 1 +A4. Have you used AI writing assistants when working on this paper? Left blank. + +B Did you use or create scientific artifacts? + +Section 2 + +B1. Did you cite the creators of artifacts you used? Section 2 +B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Section 2 +B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? + +Section 2 - we used artifact(s) as they they were intended to without any modifications. + +B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Not applicable. We have used a publicly available dataset as allowed by IETF's privacy statement https://www.ieft.org/privacy-statement/ +B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Section 2 LIWC Representation +B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Section 3 + +C Did you run computational experiments? + +Section 3 + +C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? We used default parameters for experiments without parameter tuning. + +C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Section 3 (default parameters) +C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Section 4 +C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Section 2 and Section 3 + +D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank. + +D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response. +D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response. +D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response. +D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response. +D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response. \ No newline at end of file diff --git a/2023/Tracing Linguistic Markers of Influence in a Large Online Organisation/images.zip b/2023/Tracing Linguistic Markers of Influence in a Large Online Organisation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2f654c1906cbfeb254d79f75ab6f185e33bf6ca4 --- /dev/null +++ b/2023/Tracing Linguistic Markers of Influence in a Large Online Organisation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7b40d2170ba970388cc263a0f245b7d7873745f95a1a2f2007d135af8961e70e +size 94530 diff --git a/2023/Tracing Linguistic Markers of Influence in a Large Online Organisation/layout.json b/2023/Tracing Linguistic Markers of Influence in a Large Online Organisation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a3f789d7085ae27a2a33576149b2244cfa784cd8 --- /dev/null +++ b/2023/Tracing Linguistic Markers of Influence in a Large Online Organisation/layout.json @@ -0,0 +1,5291 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 78, + 75, + 523, + 93 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 75, + 523, + 93 + ], + "spans": [ + { + "bbox": [ + 78, + 75, + 523, + 93 + ], + "type": "text", + "content": "Tracing Linguistic Markers of Influence in a Large Online Organisation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 70, + 105, + 528, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 105, + 528, + 132 + ], + "spans": [ + { + "bbox": [ + 70, + 105, + 528, + 132 + ], + "type": "text", + "content": "Prashant Khare*, Ravi Shekhar†, Vanja Mladen Karan*, Stephen McQuistin‡, Colin Perkins‡, Ignacio Castro*, Gareth Tyson*§, Patrick G.T. Healey*, Matthew Purver*¶" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 99, + 134, + 496, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 134, + 496, + 148 + ], + "spans": [ + { + "bbox": [ + 99, + 134, + 496, + 148 + ], + "type": "text", + "content": "*Queen Mary University of London, †University of Essex, ‡University of Glasgow" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 121, + 148, + 474, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 148, + 474, + 162 + ], + "spans": [ + { + "bbox": [ + 121, + 148, + 474, + 162 + ], + "type": "inline_equation", + "content": "^{\\S}" + }, + { + "bbox": [ + 121, + 148, + 474, + 162 + ], + "type": "text", + "content": "Hong Kong University of Science & Technology, " + }, + { + "bbox": [ + 121, + 148, + 474, + 162 + ], + "type": "inline_equation", + "content": "^{\\ddagger}" + }, + { + "bbox": [ + 121, + 148, + 474, + 162 + ], + "type": "text", + "content": "Jožef Stefan Institute" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 93, + 163, + 503, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 163, + 503, + 190 + ], + "spans": [ + { + "bbox": [ + 93, + 163, + 503, + 190 + ], + "type": "text", + "content": "{p.khare, m.karan, i.castro, g.tyson, p.healey, m.purver}@qmul.ac.uk, r.shekhar@essex.ac.uk, sm@smcquistin.uk, csp@csperkins.org" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 83, + 238, + 274, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 238, + 274, + 429 + ], + "spans": [ + { + "bbox": [ + 83, + 238, + 274, + 429 + ], + "type": "text", + "content": "Social science and psycholinguistic research have shown that power and status affect how people use language in a range of domains. Here, we investigate a similar question in a large, distributed, consensus-driven community – the Internet Engineering Task Force (IETF), a collaborative organisation that develops technical standards for the Internet. Our analysis, based on lexical categories (LIWC) and BERT, shows that participants' levels of influence can be predicted from their email text, and identifies key linguistic differences (e.g., certain LIWC categories, such as WE are positively correlated with high-influence). We also identify the differences in language use for the same person before and after becoming influential1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 68, + 441, + 251, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 441, + 251, + 454 + ], + "spans": [ + { + "bbox": [ + 68, + 441, + 251, + 454 + ], + "type": "text", + "content": "1 Introduction and Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 464, + 291, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 464, + 291, + 640 + ], + "spans": [ + { + "bbox": [ + 67, + 464, + 291, + 640 + ], + "type": "text", + "content": "Motivation Online communities are rapidly growing. It is imperative to study them to gain a better understanding of online dynamics and important processes such as decision-making. Prior work has shown that influence is an important aspect to consider while analysing online community dynamics (Bapna and Umyarov, 2015; Vega et al., 2021). Social and psycholinguistic research has also revealed that a person's power and status (i.e., influence) is reflected in their usage of language (Nguyen et al., 2016; Guinote, 2017). In this paper, we focus on linguistic traits exhibited by influential people in a large online community." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 640, + 291, + 721 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 640, + 291, + 721 + ], + "spans": [ + { + "bbox": [ + 67, + 640, + 291, + 721 + ], + "type": "text", + "content": "Detecting meaningful domain-independent indicators of influence is difficult (Danescu-Niculescu-Mizil et al., 2012). Instead, we focus on the Internet Engineering Task Force" + }, + { + "bbox": [ + 67, + 640, + 291, + 721 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 640, + 291, + 721 + ], + "type": "text", + "content": " (IETF) - a large, open, voluntary, standards developing organisation with over 2M emails between 56k participants over" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 212, + 527, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 212, + 527, + 415 + ], + "spans": [ + { + "bbox": [ + 302, + 212, + 527, + 415 + ], + "type": "text", + "content": "20 years. The decentralised, consensus-oriented nature of the IETF makes it an interesting case study for two reasons. First, compared to the social media data commonly used in similar studies (e.g. Tchokni et al., 2014; Prabhakaran, 2015), IETF emails are usually longer and goal-oriented. Second, the IETF is a decentralised organisation where the decision-making is collaborative and consensus-driven (Bradner, 1996; Resnick, 2014). Hence, the resulting social interactions are very different to alternative email-based datasets such as the Enron Corpus (Klimt and Yang, 2004), or interactions with more rigidly defined power distinctions e.g., admin/users, judges/lawyers (Danescu-Niculescu-Mizil et al., 2012)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 425, + 526, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 425, + 526, + 682 + ], + "spans": [ + { + "bbox": [ + 302, + 425, + 526, + 682 + ], + "type": "text", + "content": "Related Work Most studies of influence either focus on community structure rather than language, or use language indirectly. Urena et al. (2019) give a survey of the former approach. In an example of the latter, Prabhakaran et al. (2014) compare users with different influence in terms of their linguistic similarity or co-adaptation, the increasing similarity of interlocutors to each other in how they use language (see also Danescu-Niculescu-Mizil et al., 2012; Ver Steeg and Galstyan, 2013; Noble and Fernandez, 2015; Kawabata et al., 2016; Buske, 2019; Healey et al., 2023). Some studies (Bramsen et al., 2011; Gilbert, 2012) do focus on modelling influence from text of Enron emails by identifying keywords/phrases that indicate influence. Rosenthal (2014) and Tchokni et al. (2014) extend this approach to other domains, including Twitter, Wikipedia talk pages, and debates, and include a wider range of linguistic markers." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 692, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 692, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 692, + 526, + 772 + ], + "type": "text", + "content": "Goals We focus on discovering linguistic markers of influence in a large consensus-driven standards developing organisation, where the consensus is based on elaborate discussions between participants on mailing lists. To complement this analysis, we also study the linguistic behaviour" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 729, + 290, + 750 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 729, + 290, + 750 + ], + "spans": [ + { + "bbox": [ + 67, + 729, + 290, + 750 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 729, + 290, + 750 + ], + "type": "text", + "content": "Code: https://github.com/sodestream/acl2023-tracing-linguistic-markers" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 67, + 751, + 289, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 751, + 289, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 751, + 289, + 772 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 751, + 289, + 772 + ], + "type": "text", + "content": "IETF is responsible for producing technical standards for internet infrastructure. https://www.ietf.org/" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 291, + 781, + 304, + 790 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 781, + 304, + 790 + ], + "spans": [ + { + "bbox": [ + 291, + 781, + 304, + 790 + ], + "type": "text", + "content": "82" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "spans": [ + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "type": "text", + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 229, + 806, + 365, + 817 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 806, + 365, + 817 + ], + "spans": [ + { + "bbox": [ + 229, + 806, + 365, + 817 + ], + "type": "text", + "content": "Volume 2: Short Papers, pages 82-90" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 176, + 818, + 417, + 828 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 818, + 417, + 828 + ], + "spans": [ + { + "bbox": [ + 176, + 818, + 417, + 828 + ], + "type": "text", + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 293, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 293, + 248 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 293, + 248 + ], + "type": "text", + "content": "of participants at different hierarchical levels in IETF, as well as participants in different periods of their participation, similar to Danescu-Niculescu-Mizil et al. (2013), who considered the behaviour of participants as a measure of influence and claim that participants tend to echo the linguistic style of influential individuals. We map this to three research questions: RQ1: How do linguistic traits differ between more and less influential participants? RQ2: How do linguistic traits vary for participants at different levels of the organisation hierarchy? RQ3: How does linguistic behaviour of participants change as they gain influence?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 262, + 158, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 262, + 158, + 277 + ], + "spans": [ + { + "bbox": [ + 67, + 262, + 158, + 277 + ], + "type": "text", + "content": "2 Methodology" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 287, + 291, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 287, + 291, + 396 + ], + "spans": [ + { + "bbox": [ + 67, + 287, + 291, + 396 + ], + "type": "text", + "content": "We aim to understand the correlation between influence, as defined by either network-based centrality metrics (mail-based) or organisational role influence (role-based), and language usage in terms of linguistic traits. For each participant, we consider the emails they sent in a given time period and investigate correlations of certain features of their email text with two different measures of influence." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 408, + 291, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 408, + 291, + 653 + ], + "spans": [ + { + "bbox": [ + 67, + 408, + 291, + 653 + ], + "type": "text", + "content": "LIWC Representation Linguistic Inquiry and Word Count (LIWC, Pennebaker et al., 2015) is a well-recognised psycholinguistic lexicon; it provides word counts for 85 different linguistic, psychological, personal concern, and informal language marker categories. Here, we aggregate the word counts within each linguistic category for each participant using the LIWC 2015 dictionary (academic license), and normalise by the total number of emails sent by that participant. Such a normalisation is more appropriate here than normalising by total number of words written, as many IETF emails include long technical sections. This generates a representation of a participant as their mean usage of each LIWC category; while this is a relatively reduced, low-dimensional representation of a person's language, it has the advantage of being interpretable and psychologically well-motivated." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 665, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 665, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 665, + 291, + 773 + ], + "type": "text", + "content": "BERT Representation The LIWC representation ignores context. To allow comparison to more advanced methods, we use the context-dependent representations from BERT (Devlin et al., 2019) via the open-source HuggingFace library (Wolf et al., 2019). The participant-specific BERT representation is calculated by averaging the text representations (last layer CLS vectors) over all their emails." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 302, + 70, + 431, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 70, + 431, + 85 + ], + "spans": [ + { + "bbox": [ + 302, + 70, + 431, + 85 + ], + "type": "text", + "content": "3 Experimental Set-up" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 91, + 526, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 91, + 526, + 213 + ], + "spans": [ + { + "bbox": [ + 302, + 91, + 526, + 213 + ], + "type": "text", + "content": "Dataset The IETF is organised in Working Groups (WGs). Each WG has a technical focus (e.g., HTTP WG for the HTTP protocol) and one or more WG chairs. We use data from two public sources: the IETF mail archives3 and the Datatracker4. The mail archives cover WG activities, meetings, and administration. We gathered 2,106,804 emails from 56,733 email addresses spanning 2000-2019." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 214, + 526, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 214, + 526, + 323 + ], + "spans": [ + { + "bbox": [ + 302, + 214, + 526, + 323 + ], + "type": "text", + "content": "To determine mail-based influence, we use a social graph based on mailing list interactions (messages from one person to another) as built by Khare et al. (2022). We rank participants by their eigenvector centrality, a measure of a node's influence in a graph, and transform rank to a percentile. To determine role-based influence, we used Datatracker for information about WG chairs and their tenure." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 329, + 526, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 329, + 526, + 533 + ], + "spans": [ + { + "bbox": [ + 302, + 329, + 526, + 533 + ], + "type": "text", + "content": "RQ1 (mail-based influence) We used a 5-year subset of the data for RQ1 due to the computation cost, still giving a reasonable period to observe the participation consistency in the IETF community (McQuistin et al., 2021; Khare et al., 2022). We took data from 2015-2019 with 300,806 emails from 5,363 unique participants. This subset has 212,253 unique tokens, as opposed to 735,605 unique tokens in the whole dataset, and the median length of emails is 504. We calculate the mail-based influence score and LIWC representation for each participant as described. We fit a linear regression model using LIWC representations to predict influence percentile and observe the magnitude and directions of significant coefficients." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 539, + 527, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 539, + 527, + 662 + ], + "spans": [ + { + "bbox": [ + 302, + 539, + 527, + 662 + ], + "type": "text", + "content": "RQ2 (role-based influence) While mail-based influence was crucial to consider the activities of the participants based on the email network, role-based influence is equally crucial as they are involved in organisational decision making. We use the same time period as in RQ1, but here we predict organisational role-based influence. We split the data into two categories: (a) WG chairs and (b) participants who have never been WG chair. We" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 668, + 453, + 680 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 668, + 453, + 680 + ], + "spans": [ + { + "bbox": [ + 317, + 668, + 453, + 680 + ], + "type": "text", + "content": "3https://mailarchive.ietf.org/" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 303, + 680, + 525, + 710 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 680, + 525, + 710 + ], + "spans": [ + { + "bbox": [ + 303, + 680, + 525, + 710 + ], + "type": "text", + "content": "4https://datattracker.ietf.org/ - the administrative database of the IETF, containing metadata about participants and their roles, working groups, document status, etc." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 303, + 710, + 525, + 751 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 710, + 525, + 751 + ], + "spans": [ + { + "bbox": [ + 303, + 710, + 525, + 751 + ], + "type": "text", + "content": "5We filter out 104 ambiguous words that are present in LIWC but have technology, security, and network context meaning in IETF, using manually curated lists, for e.g., attack, argument, secure etc. We do this across all RQs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 303, + 751, + 525, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 751, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 303, + 751, + 525, + 772 + ], + "type": "text", + "content": "In the top " + }, + { + "bbox": [ + 303, + 751, + 525, + 772 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 303, + 751, + 525, + 772 + ], + "type": "text", + "content": " mail-based influential participants, less than " + }, + { + "bbox": [ + 303, + 751, + 525, + 772 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 303, + 751, + 525, + 772 + ], + "type": "text", + "content": " are WG chairs with significant role-based influence." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 292, + 781, + 304, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 292, + 781, + 304, + 791 + ], + "spans": [ + { + "bbox": [ + 292, + 781, + 304, + 791 + ], + "type": "text", + "content": "83" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 291, + 113 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 291, + 113 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 291, + 113 + ], + "type": "text", + "content": "calculate the LIWC representations for each person, train a logistic regression model to predict category, and observe the LIWC category coefficients." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 120, + 291, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 120, + 291, + 283 + ], + "spans": [ + { + "bbox": [ + 67, + 120, + 291, + 283 + ], + "type": "text", + "content": "RQ3 (changes in influence) We look at participants who went from low to high influence over time: individuals who had a mail-based influence below the 50th percentile when they joined the IETF, and reached the top 10th percentile at some point. For each participant, we generate two different representations based on two periods — the year of joining and year of reaching the top 10th percentile for the first time — and assign these to two different classes. As in RQ2, we then train a logistic regression model to predict these classes, and examine the coefficients of the LIWC categories." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 291, + 291, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 291, + 291, + 399 + ], + "spans": [ + { + "bbox": [ + 67, + 291, + 291, + 399 + ], + "type": "text", + "content": "BERT-based variants Our primary purpose is not to assess the predictive power of LIWC representations, but to use them as a tool to characterise linguistic variations in a meaningful way. However, in order to understand their predictive potential, given their relatively simple nature, we compare them to BERT. For these comparisons, we use the BERT representations described in Section 2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 401, + 291, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 401, + 291, + 536 + ], + "spans": [ + { + "bbox": [ + 67, + 401, + 291, + 536 + ], + "type": "text", + "content": "For each RQ we use the same experimental setup as described above. We split the data 80:20 into train and test set and train a prediction model (regression for RQ1 and classification for RQ2 & RQ3). To experiment with both linear and nonlinear models, we include linear and logistic regression and multi layer perceptrons, using implementations from scikit-learn (Pedregosa et al., 2011) with default parameters. As evaluation metrics we used Pearson's " + }, + { + "bbox": [ + 67, + 401, + 291, + 536 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 67, + 401, + 291, + 536 + ], + "type": "text", + "content": " and macro-F1 score." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 547, + 197, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 547, + 197, + 560 + ], + "spans": [ + { + "bbox": [ + 67, + 547, + 197, + 560 + ], + "type": "text", + "content": "4 Results & Discussion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 569, + 290, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 569, + 290, + 597 + ], + "spans": [ + { + "bbox": [ + 67, + 569, + 290, + 597 + ], + "type": "text", + "content": "We now explore the results (see Table 1 for all experiments) and answer our research questions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 606, + 172, + 619 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 606, + 172, + 619 + ], + "spans": [ + { + "bbox": [ + 67, + 606, + 172, + 619 + ], + "type": "text", + "content": "4.1 Answers to RQs" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 624, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 624, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 624, + 291, + 773 + ], + "type": "text", + "content": "RQ1 — The following LIWC categories are significantly correlated " + }, + { + "bbox": [ + 67, + 624, + 291, + 773 + ], + "type": "inline_equation", + "content": "(p < 0.05)" + }, + { + "bbox": [ + 67, + 624, + 291, + 773 + ], + "type": "text", + "content": " with higher mail-based influence: WE, INFORMAL, RISK, ADJECTIVE, ANGER, THEY, and BIO. Categories such as NETSPEAK, SEXUAL, HEALTH, DEATH, BODY are correlated with lower influence. This suggests that influential people tend to indicate a collaborative and community-oriented approach with first-person plural (WE) and third-person plural category (THEY) usage. This is consistent with Kacewicz et al. (2014) and Guinote (2017), who show that in" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 71, + 526, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 166 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 166 + ], + "type": "text", + "content": "fluential people use more first-person plural. They also use more organisational language, which is shown by the negative correlation of informal slang language categories (NETSPEAK, SEXUAL, BODY). We see some unexpected hidden trends due to word ambiguity (e.g., words like 'trust' and 'live'), which are investigated in Section 4.2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 166, + 527, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 166, + 527, + 327 + ], + "spans": [ + { + "bbox": [ + 302, + 166, + 527, + 327 + ], + "type": "text", + "content": "RQ2 - From 1, we see that working group (WG) chairs are more social and collaborative, as is shown by WE and SOCIAL categories. This is in line with similar findings from RQ1 and also about leadership engagements from previous works (Strzalkowski et al., 2012; Liu, 2022; Kacewicz et al., 2014; Guinote, 2017; Akstinaite et al., 2020). Also, WG chairs use tentative statements (TENTAT) in discussions, primarily focused on technical feedback and revisions, or suggesting alternatives. Examples showcasing the use of words such as 'or' and 'seems'-" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 316, + 338, + 525, + 413 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 316, + 338, + 525, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 525, + 379 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 525, + 379 + ], + "type": "text", + "content": "seems': \"With the risk of disturbing with statements, but avoiding too many questions: This seems against the goal of reducing headers.\"" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 316, + 386, + 525, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 386, + 525, + 413 + ], + "spans": [ + { + "bbox": [ + 316, + 386, + 525, + 413 + ], + "type": "text", + "content": "- 'or': \"Question is do we need to carry around an outer IP-in-IP header for that or not?\"" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 302, + 422, + 526, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 422, + 526, + 543 + ], + "spans": [ + { + "bbox": [ + 302, + 422, + 526, + 543 + ], + "type": "text", + "content": "RQ3 — From Table 1, we observe that when participants become mail-based influential they are likely to be more descriptive and engaged in immediate state of issues and situations as seen from the correlation of auxiliary verbs (AUXVERB), adverb, risk, and present focus (FOCUSPRESENT). They are also more involved in cognitive processes (COGPROC) as compared to their previous self when they were new to IETF and had little influence." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 553, + 380, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 553, + 380, + 565 + ], + "spans": [ + { + "bbox": [ + 302, + 553, + 380, + 565 + ], + "type": "text", + "content": "4.2 Discussion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 570, + 526, + 719 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 570, + 526, + 719 + ], + "spans": [ + { + "bbox": [ + 302, + 570, + 526, + 719 + ], + "type": "text", + "content": "To better understand these LIWC categories and what kind of words play a role in the behaviour of individual categories, we calculate the frequency of words in each LIWC category as they appear in the emails. Next, we consider the top 30 most frequent words in each LIWC category and perform regression analysis on mail-based influence for participants, but using only these 30 words as features to generate the participant representation. We conducted this experiment separately for each LIWC category that was significant in the first experiment." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 719, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 719, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 719, + 526, + 772 + ], + "type": "text", + "content": "From the word based analysis we make multiple observations. E.g., words like 'we' imply a collective approach and is strongly correlated with the higher influence. Similarly, the use of word 'well'" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 291, + 781, + 304, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 781, + 304, + 791 + ], + "spans": [ + { + "bbox": [ + 291, + 781, + 304, + 791 + ], + "type": "text", + "content": "84" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 68, + 523, + 135 + ], + "blocks": [ + { + "bbox": [ + 69, + 68, + 523, + 135 + ], + "lines": [ + { + "bbox": [ + 69, + 68, + 523, + 135 + ], + "spans": [ + { + "bbox": [ + 69, + 68, + 523, + 135 + ], + "type": "table", + "html": "
RQ1High influenceBIO, WE, INFORMAL, THEY, NEGEMO, ANGER, RISK, ADJECTIVE
Low influenceSEXUAL, DEATH, INGEST, NETSPEAK, HEALTH, FEMALE, BODY, AFFILIATION, CONJ
RQ2WG Chair influenceTENTAT, IPRON, SOCIAL, SEE, FEEL, WE
non-WG ChairCOGPROC, RELativ, AFFILIATION, I, REWARD
RQ3Top 10 percentileADVERB, PREP, ANGER, AUXVERB, MALE, COGPROC, ACHIEV, RISK, FOCUSPRESENT
Below 50th percentileFUNCTION, PPRON, SHEHE, IPRON, NUMBER, CERTAIN, SEXUAL, INFORMAL
", + "image_path": "ba1d152cd34d91f38965aaf03eda1e38dab812747aff9a0568141d4a62fca86e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 182, + 290, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 182, + 290, + 317 + ], + "spans": [ + { + "bbox": [ + 66, + 182, + 290, + 317 + ], + "type": "text", + "content": "is standard, such as politely resuming the conversation (e.g., 'well, I agree') or providing an approval over something (e.g., 'this works as well'). These words are well associated with the influential participants. Otherwise, influential participants are generally not observed to be informal and other frequent words (other than 'well') within INFORMAL category do not demonstrate a strong correlation with the growing influence. Also, 'well' is the most frequent word in the INFORMAL category." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 318, + 291, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 318, + 291, + 466 + ], + "spans": [ + { + "bbox": [ + 66, + 318, + 291, + 466 + ], + "type": "text", + "content": "More influential people (both mail-based and role-based) are also observed to engage more in IETF communities. The conversations can often reflect situations where, as a part of review and feedback process, more influential people highlight limitations in protocol standards, stress on specifics, and compare with existing protocols or previous versions. Several words across different LIWC categories (RISK, NEGEMO, and ADJ) highlight such behaviour, e.g., 'problems', 'before', 'particular', 'specific', 'different', 'most', and 'than'." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 469, + 291, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 469, + 291, + 686 + ], + "spans": [ + { + "bbox": [ + 66, + 469, + 291, + 686 + ], + "type": "text", + "content": "However, there are many words with dual sense, like 'trust' which has a very technology specific usage related to network security instead of conversations involving trust issues between individuals or trust in any given situation. Similarly, the word 'live' is related with an application or network being live, instead of its conventional meaning. We also observed that some of the LIWC categories, such as BIO, did not have specific terms that could clearly establish its significance in favour of influential participants (e.g., word 'problems' and 'trust' reflecting the significance for the category RISK), instead such categories had several words with quite weak correlation with influential participants. Such words collectively drifted the weight of the category towards influential participants." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 699, + 188, + 711 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 699, + 188, + 711 + ], + "spans": [ + { + "bbox": [ + 67, + 699, + 188, + 711 + ], + "type": "text", + "content": "4.3 BERT-based results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 719, + 290, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 719, + 290, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 719, + 290, + 773 + ], + "type": "text", + "content": "We compared the performance of the LIWC- and BERT-based models. Results in Table 2 indicate our LIWC approach is better than an intuitive BERT-based baseline. We hypothesize that the" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 306, + 179, + 511, + 232 + ], + "blocks": [ + { + "bbox": [ + 208, + 148, + 383, + 161 + ], + "lines": [ + { + "bbox": [ + 208, + 148, + 383, + 161 + ], + "spans": [ + { + "bbox": [ + 208, + 148, + 383, + 161 + ], + "type": "text", + "content": "Table 1: LIWC categories where " + }, + { + "bbox": [ + 208, + 148, + 383, + 161 + ], + "type": "inline_equation", + "content": "p < {0.05}" + }, + { + "bbox": [ + 208, + 148, + 383, + 161 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 179, + 511, + 232 + ], + "lines": [ + { + "bbox": [ + 306, + 179, + 511, + 232 + ], + "spans": [ + { + "bbox": [ + 306, + 179, + 511, + 232 + ], + "type": "table", + "html": "
LIWCBERT
LRMLPLRMLP
RQ1 (Pearson ρ)0.850*0.852*-0.0180.015
RQ2 (Micro F1)91.2192.4687.6992.21
RQ3 (Micro F1)88.8990.7451.8555.56
", + "image_path": "300b79aeb577c665823b70a6c034889c7f191dcec3018e809690c9ddd2614200.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 331, + 240, + 496, + 252 + ], + "lines": [ + { + "bbox": [ + 331, + 240, + 496, + 252 + ], + "spans": [ + { + "bbox": [ + 331, + 240, + 496, + 252 + ], + "type": "text", + "content": "Table 2: LIWC vs BERT(\\* " + }, + { + "bbox": [ + 331, + 240, + 496, + 252 + ], + "type": "inline_equation", + "content": "p < 0.0001" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 302, + 275, + 526, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 275, + 526, + 449 + ], + "spans": [ + { + "bbox": [ + 302, + 275, + 526, + 449 + ], + "type": "text", + "content": "reason for this is that LIWC is specialised to detect linguistic markers relevant for this task. Also, to ensure fair comparison, BERT representations were not fine-tuned for the tasks. We believe combining LIWC and BERT might give better representations, especially when dealing with ambiguous words. Curiously, when observing t-SNE (Van der Maaten and Hinton, 2008) projections of participants' BERT representations (Appendix A), we find that low-influence users show a much bigger variation for relevant categories such as WE, NETS-PEAK and INFORMAL. We will investigate this in future." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 465, + 493, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 465, + 493, + 478 + ], + "spans": [ + { + "bbox": [ + 302, + 465, + 493, + 478 + ], + "type": "text", + "content": "5 Conclusions & Future Directions" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 489, + 526, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 489, + 526, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 489, + 526, + 773 + ], + "type": "text", + "content": "This paper explores the linguistic patterns of influence in an online collaborative organisation, by analysing the differences between high- and lowinfluence participants. Using two aspects of influence — mail-based, derived from the email network, and organisational role-based — we were able to unfold several traits that differentiate influential participants from others. Many of our findings seem corroborated by studies in organisational theory. We observed that influential people exhibit more collaborative and community-oriented traits, and also stronger signs of engagement in discussions. We also observed that as people go on to become influential participants, they evolve in their communication and are seen to be more engaging and descriptive in their linguistic style. An interesting practical application of our research is identifying and analyzing groups that are dysfunctional in terms of participant roles and their communication patterns (e.g., where the chair is not performing their role). In future work, we will" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 292, + 781, + 304, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 292, + 781, + 304, + 791 + ], + "spans": [ + { + "bbox": [ + 292, + 781, + 304, + 791 + ], + "type": "text", + "content": "85" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 291, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 291, + 152 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 291, + 152 + ], + "type": "text", + "content": "extend the experiments to study these patterns of interaction in more linguistic depth, between more different roles within an organisation (possibly for multiple collaborative organisations). We will attempt to go beyond lexical count and account for word context." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 163, + 149, + 176 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 163, + 149, + 176 + ], + "spans": [ + { + "bbox": [ + 67, + 163, + 149, + 176 + ], + "type": "text", + "content": "6 Limitations" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 185, + 292, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 185, + 292, + 579 + ], + "spans": [ + { + "bbox": [ + 67, + 185, + 292, + 579 + ], + "type": "text", + "content": "One of the main limitations is that we used the standard LIWC-based analysis approach, which is purely lexical and does not take into account the context in which a word appears. Consequently, many words that have very specific senses in the context of the IETF get miscounted as occurrences of LIWC categories. This could be addressed by a more advanced method of mapping to LIWC categories that would account for context. Another limitation is that we manually generated a filtering list containing words specific to the IETF. This list might not be exhaustive enough. Also, we were limited by not conducting an exhaustive hyper-parameter search on our models. We also understand that many emails are longer than 512 tokens (the input limit of the BERT model we used) and might have not been captured completely by our BERT model. However, most of the emails do fit into this BERT sequence length limit. We did not fine tune BERT on the IETF data; this might have given better performance, although it is not clear if it would have given more insight: our main goal is not performance but analyzing/comparing characteristics of existing models. It is also worth highlighting that the data used in this work is strictly in English, and the psycholinguistic categories in LIWC are also based on English language. Hence, this study may be biased and not fully capture variations in linguistic traits that are culturally agnostic." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 600, + 291, + 735 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 600, + 291, + 735 + ], + "spans": [ + { + "bbox": [ + 67, + 600, + 291, + 735 + ], + "type": "text", + "content": "Ethical considerations — Participation in the IETF is bound by agreements and policies explicitly stating that mailing list discussions and Data-tracker metadata will be made publicly available.7 We use only this publicly available data in our analysis. We have discussed our work with the IETF leadership to confirm that it fits their acceptable use policies. We have also made provisions to manage the data securely, and retain it only as necessary for our work." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 740, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 740, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 740, + 291, + 772 + ], + "type": "text", + "content": "See both https://www.ieft.org/about/note-well/ and the IETF privacy policy available at https://www.ieft.org/privacy-statement/." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 303, + 71, + 407, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 71, + 407, + 84 + ], + "spans": [ + { + "bbox": [ + 303, + 71, + 407, + 84 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 92, + 527, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 92, + 527, + 200 + ], + "spans": [ + { + "bbox": [ + 302, + 92, + 527, + 200 + ], + "type": "text", + "content": "We thank the anonymous reviewers for their helpful comments. This work was supported by the UK EPSRC under grants EP/S033564/1 and EP/S036075/1 (Sodestream: Streamlining Social Decision Making for Enhanced Internet Standards). Purver was also supported by the Slovenian Research Agency via research core funding for the programme Knowledge Technologies (P2-0103)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 223, + 362, + 235 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 223, + 362, + 235 + ], + "spans": [ + { + "bbox": [ + 304, + 223, + 362, + 235 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 303, + 241, + 527, + 772 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 304, + 241, + 527, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 241, + 527, + 275 + ], + "spans": [ + { + "bbox": [ + 304, + 241, + 527, + 275 + ], + "type": "text", + "content": "Vita Akstinaite, Graham Robinson, and Eugene Sadler-Smith. 2020. Linguistic markers of ceo hubris. Journal of Business Ethics, 167:687-705." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 283, + 527, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 283, + 527, + 328 + ], + "spans": [ + { + "bbox": [ + 304, + 283, + 527, + 328 + ], + "type": "text", + "content": "Ravi Bapna and Akhmed Umyarov. 2015. Do your online friends make you pay? a randomized field experiment on peer influence in online social networks. Management Science, 61(8):1902-1920." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 336, + 525, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 336, + 525, + 359 + ], + "spans": [ + { + "bbox": [ + 304, + 336, + 525, + 359 + ], + "type": "text", + "content": "Scott O. Bradner. 1996. The Internet Standards Process - Revision 3. RFC 2026." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 366, + 527, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 527, + 433 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 527, + 433 + ], + "type": "text", + "content": "Philip Bramsen, Martha Escobar-Molano, Ami Patel, and Rafael Alonso. 2011. Extracting social power relationships from natural language. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 773-782." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 303, + 441, + 527, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 441, + 527, + 475 + ], + "spans": [ + { + "bbox": [ + 303, + 441, + 527, + 475 + ], + "type": "text", + "content": "Jakob A Buske. 2019. Linguistic accommodation between leaders and followers. B.S. thesis, University of Twente." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 303, + 482, + 527, + 539 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 482, + 527, + 539 + ], + "spans": [ + { + "bbox": [ + 303, + 482, + 527, + 539 + ], + "type": "text", + "content": "Cristian Danescu-Niculescu-Mizil, Lillian Lee, Bo Pang, and Jon Kleinberg. 2012. Echoes of power: Language effects and power differences in social interaction. In Proceedings of the 21st international conference on World Wide Web, pages 699-708." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 303, + 545, + 527, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 545, + 527, + 613 + ], + "spans": [ + { + "bbox": [ + 303, + 545, + 527, + 613 + ], + "type": "text", + "content": "Cristian Danescu-Niculescu-Mizil, Robert West, Dan Jurafsky, Jure Leskovec, and Christopher Potts. 2013. No country for old members: User lifecycle and linguistic change in online communities. In Proceedings of the 22nd international conference on World Wide Web, pages 307-318." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 303, + 620, + 527, + 721 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 620, + 527, + 721 + ], + "spans": [ + { + "bbox": [ + 303, + 620, + 527, + 721 + ], + "type": "text", + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 303, + 728, + 527, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 728, + 527, + 772 + ], + "spans": [ + { + "bbox": [ + 303, + 728, + 527, + 772 + ], + "type": "text", + "content": "Eric Gilbert. 2012. Phrases that signal workplace hierarchy. In Proceedings of the ACM 2012 conference on Computer Supported Cooperative Work, pages 1037-1046." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 292, + 781, + 305, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 292, + 781, + 305, + 791 + ], + "spans": [ + { + "bbox": [ + 292, + 781, + 305, + 791 + ], + "type": "text", + "content": "86" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 291, + 105 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 291, + 105 + ], + "type": "text", + "content": "Ana Guinote. 2017. How power affects people: Activating, wanting and goal seeking. Annual review of psychology, 68:353-381." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 114, + 290, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 114, + 290, + 202 + ], + "spans": [ + { + "bbox": [ + 69, + 114, + 290, + 202 + ], + "type": "text", + "content": "Patrick Healey, Prashant Khare, Ignacio Castro, Gareth Tyson, Mladen Karan, Ravi Shekhar, Stephen McGuistin, Colin Perkins, and Matthew Purver. 2023. Power and vulnerability: Managing sensitive language in organisational communication (extended abstract). In ST&D 2023: Annual Meeting of the Society for Text and Discourse, June 28 – June 30, 2023, Oslo, Norway." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 210, + 290, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 210, + 290, + 264 + ], + "spans": [ + { + "bbox": [ + 69, + 210, + 290, + 264 + ], + "type": "text", + "content": "Ewa Kacewicz, James W Pennebaker, Matthew Davis, Moongee Jeon, and Arthur C Graesser. 2014. Pronoun use reflects standings in social hierarchies. Journal of Language and Social Psychology, 33(2):125-143." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 274, + 290, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 274, + 290, + 317 + ], + "spans": [ + { + "bbox": [ + 69, + 274, + 290, + 317 + ], + "type": "text", + "content": "Kan Kawabata, Visar Berisha, Anna Scaglione, and Amy LaCross. 2016. A convex model for linguistic influence in group conversations. In *INTERSPEECH*, pages 1442-1446." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 326, + 290, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 326, + 290, + 401 + ], + "spans": [ + { + "bbox": [ + 69, + 326, + 290, + 401 + ], + "type": "text", + "content": "Prashant Khare, Mladen Karan, Stephen McQuistin, Colin Perkins, Gareth Tyson, Matthew Purver, Patrick Healey, and Ignacio Castro. 2022. The web we weave: Untangling the social graph of the IETF. In Proceedings of the International AAAI Conference on Web and Social Media, volume 16, pages 500-511." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 411, + 290, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 411, + 290, + 455 + ], + "spans": [ + { + "bbox": [ + 69, + 411, + 290, + 455 + ], + "type": "text", + "content": "Bryan Klimt and Yiming Yang. 2004. The enron corpus: A new dataset for email classification research. In European conference on machine learning, pages 217-226. Springer." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 463, + 290, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 463, + 290, + 507 + ], + "spans": [ + { + "bbox": [ + 69, + 463, + 290, + 507 + ], + "type": "text", + "content": "Amy H Liu. 2022. Pronoun usage as a measure of power personalization: A general theory with evidence from the chinese-speaking world. British Journal of Political Science, 52(3):1258-1275." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 515, + 290, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 515, + 290, + 592 + ], + "spans": [ + { + "bbox": [ + 69, + 515, + 290, + 592 + ], + "type": "text", + "content": "Stephen McQuistin, Mladen Karan, Prashant Khare, Colin Perkins, Gareth Tyson, Matthew Purver, Patrick Healey, Waleed Iqbal, Junaid Qadir, and Ignacio Castro. 2021. Characterising the IETF through the lens of RFC deployment. In Proceedings of the 21st ACM Internet Measurement Conference, pages 137-149." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 601, + 290, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 601, + 290, + 645 + ], + "spans": [ + { + "bbox": [ + 69, + 601, + 290, + 645 + ], + "type": "text", + "content": "Dong Nguyen, A Seza Dogruoz, Carolyn P Rose, and Franciska De Jong. 2016. Computational sociolinguistics: A survey. Computational linguistics, 42(3):537-593." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 653, + 290, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 653, + 290, + 708 + ], + "spans": [ + { + "bbox": [ + 69, + 653, + 290, + 708 + ], + "type": "text", + "content": "Bill Noble and Raquel Fernandez. 2015. Centre stage: How social network position shapes linguistic coordination. In Proceedings of the 6th workshop on cognitive modeling and computational linguistics, pages 29-38." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 717, + 290, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 717, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 717, + 290, + 772 + ], + "type": "text", + "content": "F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. 2011. Scikit-learn: Machine learning in" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 524, + 772 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 315, + 72, + 524, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 524, + 94 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 524, + 94 + ], + "type": "text", + "content": "Python. Journal of Machine Learning Research, 12:2825-2830." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 105, + 524, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 105, + 524, + 139 + ], + "spans": [ + { + "bbox": [ + 304, + 105, + 524, + 139 + ], + "type": "text", + "content": "James W Pennebaker, Ryan L Boyd, Kayla Jordan, and Kate Blackburn. 2015. The development and psychometric properties of LIWC2015. Technical report." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 150, + 524, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 150, + 524, + 184 + ], + "spans": [ + { + "bbox": [ + 304, + 150, + 524, + 184 + ], + "type": "text", + "content": "Vinodkumar Prabhakaran. 2015. Social power in interactions: Computational analysis and detection of power relations. Ph.D. thesis, Columbia University." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 195, + 524, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 195, + 524, + 250 + ], + "spans": [ + { + "bbox": [ + 304, + 195, + 524, + 250 + ], + "type": "text", + "content": "Vinodkumar Prabhakaran, Ashima Arora, and Owen Rambow. 2014. Staying on topic: An indicator of power in political debates. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1481-1486." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 261, + 524, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 261, + 524, + 283 + ], + "spans": [ + { + "bbox": [ + 304, + 261, + 524, + 283 + ], + "type": "text", + "content": "Pete Resnick. 2014. On Consensus and Humming in the IETF. RFC 7282." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 295, + 524, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 295, + 524, + 328 + ], + "spans": [ + { + "bbox": [ + 304, + 295, + 524, + 328 + ], + "type": "text", + "content": "Sara Rosenthal. 2014. Detecting influencers in social media discussions. XRDS: Crossroads, The ACM Magazine for Students, 21(1):40-45." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 339, + 524, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 339, + 524, + 406 + ], + "spans": [ + { + "bbox": [ + 304, + 339, + 524, + 406 + ], + "type": "text", + "content": "Tomek Strzalkowski, Samira Shaikh, Ting Liu, George Aaron Broadwell, Jenny Stromer-Galley, Sarah Taylor, Umit Boz, Veena Ravishankar, and Xiaoai Ren. 2012. Modeling leadership and influence in multi-party online discourse. In Proceedings of COLING 2012, pages 2535-2552." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 417, + 524, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 417, + 524, + 461 + ], + "spans": [ + { + "bbox": [ + 304, + 417, + 524, + 461 + ], + "type": "text", + "content": "Simo Editha Tchokni, Diarmuid O Seaghdha, and Daniele Quercia. 2014. Emoticons and phrases: Status symbols in social media. In Eighth International AAAI Conference on Weblogs and Social Media." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 472, + 524, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 524, + 528 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 524, + 528 + ], + "type": "text", + "content": "Raquel Urena, Gang Kou, Yucheng Dong, Francisco Chiclana, and Enrique Herrera-Viedma. 2019. A review on trust propagation and opinion dynamics in social networks and group decision making frameworks. Information Sciences, 478:461-475." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 539, + 524, + 572 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 539, + 524, + 572 + ], + "spans": [ + { + "bbox": [ + 304, + 539, + 524, + 572 + ], + "type": "text", + "content": "Laurens Van der Maaten and Geoffrey Hinton. 2008. Visualizing data using t-sne. Journal of machine learning research, 9(11)." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 583, + 524, + 629 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 583, + 524, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 583, + 524, + 629 + ], + "type": "text", + "content": "Lea Vega, Andres Mendez-Vazquez, and Armando López-Cuevas. 2021. Probabilistic reasoning system for social influence analysis in online social networks. Social Network Analysis and Mining, 11(1):1-20." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 640, + 524, + 695 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 640, + 524, + 695 + ], + "spans": [ + { + "bbox": [ + 304, + 640, + 524, + 695 + ], + "type": "text", + "content": "Greg Ver Steeg and Aram Galstyan. 2013. Information-theoretic measures of influence based on content dynamics. In Proceedings of the sixth ACM international conference on Web search and data mining, pages 3-12." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 706, + 524, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 706, + 524, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 706, + 524, + 772 + ], + "type": "text", + "content": "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. 2019. Huggingface's transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771." + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 291, + 781, + 304, + 790 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 781, + 304, + 790 + ], + "spans": [ + { + "bbox": [ + 291, + 781, + 304, + 790 + ], + "type": "text", + "content": "87" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 71, + 261, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 71, + 261, + 84 + ], + "spans": [ + { + "bbox": [ + 68, + 71, + 261, + 84 + ], + "type": "text", + "content": "A Appendix A: BERT-based results" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 92, + 293, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 92, + 293, + 242 + ], + "spans": [ + { + "bbox": [ + 67, + 92, + 293, + 242 + ], + "type": "text", + "content": "We investigated how BERT representations vary for participants, as per influence, across different significant LIWC categories. For each participant, we calculated the LIWC category representation by averaging the BERT representation of the words in that LIWC category and then projected using t-SNE. As Figures 1, 2 and 3 show, high-influence participants show less variation in their BERT representations compared to lower-influence participants, for the LIWC categories WE, NETSPEAK and INFORMAL respectively." + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 82, + 260, + 265, + 405 + ], + "blocks": [ + { + "bbox": [ + 82, + 260, + 265, + 405 + ], + "lines": [ + { + "bbox": [ + 82, + 260, + 265, + 405 + ], + "spans": [ + { + "bbox": [ + 82, + 260, + 265, + 405 + ], + "type": "image", + "image_path": "5a752473ca314eac08147331e77521d5b8578bc8bbd72742f47aa1a8085d321a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 103, + 418, + 254, + 431 + ], + "lines": [ + { + "bbox": [ + 103, + 418, + 254, + 431 + ], + "spans": [ + { + "bbox": [ + 103, + 418, + 254, + 431 + ], + "type": "text", + "content": "Figure 1: WE category representation" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 82, + 465, + 263, + 607 + ], + "blocks": [ + { + "bbox": [ + 82, + 465, + 263, + 607 + ], + "lines": [ + { + "bbox": [ + 82, + 465, + 263, + 607 + ], + "spans": [ + { + "bbox": [ + 82, + 465, + 263, + 607 + ], + "type": "image", + "image_path": "902aca23094a75daff91ba0d8694da8fafb63a22c39a235129ee4b9af051714f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 87, + 623, + 270, + 635 + ], + "lines": [ + { + "bbox": [ + 87, + 623, + 270, + 635 + ], + "spans": [ + { + "bbox": [ + 87, + 623, + 270, + 635 + ], + "type": "text", + "content": "Figure 2: NETSPEAK category representation" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 317, + 339, + 499, + 481 + ], + "blocks": [ + { + "bbox": [ + 317, + 339, + 499, + 481 + ], + "lines": [ + { + "bbox": [ + 317, + 339, + 499, + 481 + ], + "spans": [ + { + "bbox": [ + 317, + 339, + 499, + 481 + ], + "type": "image", + "image_path": "2f6cb6cddf76e36a4ea8b467d7bac1fe32e2956854c8362d275e5eec60e14255.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 497, + 506, + 510 + ], + "lines": [ + { + "bbox": [ + 322, + 497, + 506, + 510 + ], + "spans": [ + { + "bbox": [ + 322, + 497, + 506, + 510 + ], + "type": "text", + "content": "Figure 3: INFORMAL category representation" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 292, + 781, + 304, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 292, + 781, + 304, + 791 + ], + "spans": [ + { + "bbox": [ + 292, + 781, + 304, + 791 + ], + "type": "text", + "content": "88" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "spans": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "type": "text", + "content": "A For every submission:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 106, + 414, + 242 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "spans": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "type": "text", + "content": "A1. Did you describe the limitations of your work? Section 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 142, + 329, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 142, + 329, + 169 + ], + "spans": [ + { + "bbox": [ + 77, + 142, + 329, + 169 + ], + "type": "text", + "content": "A2. Did you discuss any potential risks of your work? Section 6 in Limitations section" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 179, + 414, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 179, + 414, + 205 + ], + "spans": [ + { + "bbox": [ + 77, + 179, + 414, + 205 + ], + "type": "text", + "content": "A3. Do the abstract and introduction summarize the paper's main claims? Abstract and Section 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 77, + 215, + 398, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 215, + 398, + 242 + ], + "spans": [ + { + "bbox": [ + 77, + 215, + 398, + 242 + ], + "type": "text", + "content": "A4. Have you used AI writing assistants when working on this paper? Left blank." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 69, + 252, + 290, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 252, + 290, + 266 + ], + "spans": [ + { + "bbox": [ + 69, + 252, + 290, + 266 + ], + "type": "text", + "content": "B Did you use or create scientific artifacts?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 79, + 270, + 122, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 270, + 122, + 282 + ], + "spans": [ + { + "bbox": [ + 79, + 270, + 122, + 282 + ], + "type": "text", + "content": "Section 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 77, + 291, + 524, + 420 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 77, + 291, + 315, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 291, + 315, + 317 + ], + "spans": [ + { + "bbox": [ + 77, + 291, + 315, + 317 + ], + "type": "text", + "content": "B1. Did you cite the creators of artifacts you used? Section 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 77, + 327, + 463, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 327, + 463, + 354 + ], + "spans": [ + { + "bbox": [ + 77, + 327, + 463, + 354 + ], + "type": "text", + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Section 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 77, + 364, + 524, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 364, + 524, + 420 + ], + "spans": [ + { + "bbox": [ + 77, + 364, + 524, + 420 + ], + "type": "text", + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 89, + 420, + 468, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 420, + 468, + 431 + ], + "spans": [ + { + "bbox": [ + 89, + 420, + 468, + 431 + ], + "type": "text", + "content": "Section 2 - we used artifact(s) as they they were intended to without any modifications." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 77, + 441, + 524, + 646 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 77, + 441, + 524, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 441, + 524, + 508 + ], + "spans": [ + { + "bbox": [ + 77, + 441, + 524, + 508 + ], + "type": "text", + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Not applicable. We have used a publicly available dataset as allowed by IETF's privacy statement https://www.ieft.org/privacy-statement/" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 77, + 517, + 524, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 517, + 524, + 557 + ], + "spans": [ + { + "bbox": [ + 77, + 517, + 524, + 557 + ], + "type": "text", + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Section 2 LIWC Representation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 77, + 566, + 524, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 566, + 524, + 646 + ], + "spans": [ + { + "bbox": [ + 77, + 566, + 524, + 646 + ], + "type": "text", + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Section 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 69, + 656, + 293, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 656, + 293, + 671 + ], + "spans": [ + { + "bbox": [ + 69, + 656, + 293, + 671 + ], + "type": "text", + "content": "C Did you run computational experiments?" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 79, + 676, + 122, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 676, + 122, + 687 + ], + "spans": [ + { + "bbox": [ + 79, + 676, + 122, + 687 + ], + "type": "text", + "content": "Section 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 77, + 698, + 524, + 737 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 698, + 524, + 737 + ], + "spans": [ + { + "bbox": [ + 77, + 698, + 524, + 737 + ], + "type": "text", + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? We used default parameters for experiments without parameter tuning." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "spans": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "text", + "content": "ACL 2023 Responsible NLP Checklist" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 744, + 522, + 764 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 744, + 522, + 764 + ], + "spans": [ + { + "bbox": [ + 67, + 744, + 522, + 764 + ], + "type": "text", + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 291, + 781, + 304, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 781, + 304, + 791 + ], + "spans": [ + { + "bbox": [ + 291, + 781, + 304, + 791 + ], + "type": "text", + "content": "89" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 77, + 70, + 523, + 237 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 77, + 70, + 523, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 70, + 523, + 111 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 523, + 111 + ], + "type": "text", + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Section 3 (default parameters)" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 77, + 120, + 523, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 120, + 523, + 174 + ], + "spans": [ + { + "bbox": [ + 77, + 120, + 523, + 174 + ], + "type": "text", + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Section 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 183, + 523, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 183, + 523, + 237 + ], + "spans": [ + { + "bbox": [ + 77, + 183, + 523, + 237 + ], + "type": "text", + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Section 2 and Section 3" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 67, + 247, + 522, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 247, + 522, + 278 + ], + "spans": [ + { + "bbox": [ + 67, + 247, + 522, + 278 + ], + "type": "text", + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 76, + 286, + 523, + 539 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 76, + 286, + 523, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 286, + 523, + 326 + ], + "spans": [ + { + "bbox": [ + 76, + 286, + 523, + 326 + ], + "type": "text", + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 76, + 336, + 523, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 336, + 523, + 390 + ], + "spans": [ + { + "bbox": [ + 76, + 336, + 523, + 390 + ], + "type": "text", + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 76, + 400, + 523, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 400, + 523, + 454 + ], + "spans": [ + { + "bbox": [ + 76, + 400, + 523, + 454 + ], + "type": "text", + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 76, + 462, + 519, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 462, + 519, + 489 + ], + "spans": [ + { + "bbox": [ + 76, + 462, + 519, + 489 + ], + "type": "text", + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 76, + 498, + 523, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 498, + 523, + 539 + ], + "spans": [ + { + "bbox": [ + 76, + 498, + 523, + 539 + ], + "type": "text", + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 291, + 781, + 304, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 781, + 304, + 791 + ], + "spans": [ + { + "bbox": [ + 291, + 781, + 304, + 791 + ], + "type": "text", + "content": "90" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Trading Syntax Trees for Wordpieces_ Target-oriented Opinion Words Extraction with Wordpieces and Aspect Enhancement/3b3076af-e14d-4a09-b446-5edad95e9345_content_list.json b/2023/Trading Syntax Trees for Wordpieces_ Target-oriented Opinion Words Extraction with Wordpieces and Aspect Enhancement/3b3076af-e14d-4a09-b446-5edad95e9345_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..90fbafb3ffec44ff6d67d737f2f6f586ea626469 --- /dev/null +++ b/2023/Trading Syntax Trees for Wordpieces_ Target-oriented Opinion Words Extraction with Wordpieces and Aspect Enhancement/3b3076af-e14d-4a09-b446-5edad95e9345_content_list.json @@ -0,0 +1,1228 @@ +[ + { + "type": "text", + "text": "Trading Syntax Trees for Wordpieces: Target-oriented Opinion Words Extraction with Wordpieces and Aspect Enhancement", + "text_level": 1, + "bbox": [ + 127, + 87, + 872, + 145 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Samuel Mensah \nComputer Science Department \nUniversity of Sheffield, UK \ns.mensah@sheffield.ac.uk", + "bbox": [ + 124, + 160, + 374, + 225 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kai Sun \nBDBC and SKLSDE \nBeihang University, China \nsunkai@buaa.edu.cn", + "bbox": [ + 389, + 160, + 606, + 225 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nikolaos Aletras \nComputer Science Department \nUniversity of Sheffield, UK \nn.aletras@sheffield.ac.uk", + "bbox": [ + 623, + 160, + 875, + 225 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 252, + 339, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "State-of-the-art target-oriented opinion word extraction (TOWE) models typically use BERT-based text encoders that operate on the word level, along with graph convolutional networks (GCNs) that incorporate syntactic information extracted from syntax trees. These methods achieve limited gains with GCNs and have difficulty using BERT wordpieces. Meanwhile, BERT wordpieces are known to be effective at representing rare words or words with insufficient context information. To address this issue, this work trades syntax trees for BERT wordpieces by entirely removing the GCN component from the methods' architectures. To enhance TOWE performance, we tackle the issue of aspect representation loss during encoding. Instead of solely utilizing a sentence as the input, we use a sentence-aspect pair. Our relatively simple approach achieves state-of-the-art results on benchmark datasets and should serve as a strong baseline for further research.", + "bbox": [ + 141, + 284, + 460, + 583 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 114, + 600, + 258, + 615 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Target-oriented opinion word extraction (TOWE; Fan et al. (2019)) is a subtask in aspect-based sentiment analysis (ABSA; Pontiki et al. (2014b)), which aims to identify words that express an opinion about a specific target (or aspect) in a sentence. For instance, in the sentence \"Such an awesome surfboard.\", a TOWE model should identify \"awesome\" as the opinion word for the given aspect surfboard. TOWE provides explicit aspect-opinion pairs which can be used to improve results in downstream tasks such as opinion summarization (Kim et al., 2011) and information extraction (Pontiki et al., 2014b; Tang et al., 2016; Sun et al., 2023).", + "bbox": [ + 112, + 627, + 489, + 851 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Currently, many TOWE methods (Veyseh et al., 2020; Chen et al., 2020; Jiang et al., 2021; Feng et al., 2021; Mensah et al., 2021) use pretrained BERT (Devlin et al., 2018), to encode the input", + "bbox": [ + 112, + 854, + 489, + 919 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/bfeec2e2e654b44679e1596245e4b1fb970a59881a66ba60524f71e393f61d3c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
1. Sentence:Wordpieces:Such an awesome surfboard‘such’, ‘an’, ‘awesome’, ‘surf’,‘##board’
2. Sentence:Wordpieces:A great snowboard which holds edges well when riding on snow.A, ‘great’, ‘snow’, ‘#’board’, ‘which’, ‘holds’, ‘edges’, ‘well’, ‘when’, ‘riding’, ‘on’, ‘snow’.
", + "bbox": [ + 512, + 250, + 880, + 376 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Table 1: Sentences demonstrating contextual understanding through shared wordpieces. The table shows each sentence and its corresponding BERT wordpiece sequence. Aspect words are bold-typed and opinion words are italicized. The shared wordpiece '##board' helps in decoding the meaning of \"surfboard\".", + "bbox": [ + 507, + 386, + 884, + 473 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sentence. BERT has the ability to effectively capture context, which can improve TOWE performance. However, many of these methods are rather complex, as they often incorporate syntax tree information using a graph convolutional network (GCN) (Kipf and Welling, 2017). For instance, Veyseh et al. (2020) uses an ordered-neuron LSTM (Shen et al., 2018) encoder with a GCN while Jiang et al. (2021) applies an attention-based relational GCN on the syntax tree. Mensah et al. (2021) applies a BiLSTM (Hochreiter and Schmidhuber, 1997) on BERT embeddings and incorporate syntax information via a GCN.", + "bbox": [ + 507, + 500, + 884, + 708 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While incorporating syntax information through GCNs has been shown to provide some performance gains in TOWE, these are usually limited (Mensah et al., 2021). Moreover, modeling subword tokens with a GCN can be challenging because the syntax tree consists of whole words rather than subword tokens like wordpieces (Schuster and Nakajima, 2012; Devlin et al., 2018). Models based on subword tokens strike a good balance between character- and word-based encoders. They are able to effectively learn representations of rare words or words with insufficient context information. Consider the example in Table 1. The context", + "bbox": [ + 507, + 709, + 884, + 919 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "999", + "bbox": [ + 485, + 927, + 515, + 940 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics", + "bbox": [ + 226, + 945, + 769, + 957 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Volume 2: Short Papers, pages 999-1007", + "bbox": [ + 371, + 958, + 623, + 971 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "July 9-14, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 295, + 972, + 700, + 985 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "information for \"surfboard\" is limited, making it difficult to understand its meaning without additional context. However, both aspects share the wordpiece \"#board\", which allows the meaning of \"surfboard\" to be partially understood by using information from the context of \"snowboard\". In this case, \"riding\" is related to both aspects through the shared wordpiece, enabling the representation of \"surfboard\" to be improved.", + "bbox": [ + 110, + 84, + 487, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a substantial simplification for syntax-aware TOWE models (Veyseh et al., 2020; Jiang et al., 2021; Mensah et al., 2021) by replacing the syntax tree with subword information while maintaining good prediction performance. This is accomplished by removing the GCN from these architectures and using BERT wordpieces instead. Additionally, we address the issue of aspect representation degradation during encoding. This degradation negatively affects TOWE performance by reducing the availability of semantic information about the aspect for determining the opinion words to extract. To solve this problem, we propose using a sentence-aspect pair as input rather than just a sentence, similar to the approach used by Tian et al. (2021) for aspect-based sentiment classification. Through extensive experimentation, we found that our simple approach achieves state-of-the-art (SOTA) results by outperforming the method proposed by Mensah et al. (2021) without the need of a GCN component.", + "bbox": [ + 115, + 230, + 489, + 567 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Task Formalization", + "text_level": 1, + "bbox": [ + 112, + 580, + 317, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The TOWE task aims to identify an opinion word in a sentence $S = \\{w_{1},\\ldots ,w_{n_{s}}\\}$ with respect to an aspect $w_{a}\\in S$ . The sentence is typically tokenized into a sequence of tokens at different levels of granularity (e.g. subwords or whole words), $T = \\{t_1,\\dots ,t_{n_t}\\}$ , with $t_a\\in T$ denoting a subsequence of the aspect $w_{a}$ and $n_s\\leq n_t$ . The goal is to assign one of three tags (I, O, or B) to each token using the IOB format (Ramshaw and Marcus, 1995), which indicates whether the word is at the Inside, Outside or Beginning of the opinion word relative to the aspect.", + "bbox": [ + 112, + 606, + 489, + 800 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3 Syntax-aware Approaches to TOWE", + "text_level": 1, + "bbox": [ + 112, + 812, + 463, + 829 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Typically, syntax-aware approaches to TOWE (Veyseh et al., 2020; Jiang et al., 2021; Mensah et al., 2021) employ a text encoder that utilizes pretrained BERT (Devlin et al., 2018) and position embeddings (Zeng et al., 2014) (or category em", + "bbox": [ + 112, + 838, + 489, + 917 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "beddings (Jiang et al., 2021)) to learn whole word representations that are aware of the aspect's location in text. These approaches also include a GCN that operates on a syntax tree in order to incorporate syntactic information into the model.", + "bbox": [ + 507, + 84, + 882, + 164 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Ordered-Neuron LSTM GCN (ONG): Veyseh et al. (2020) combine an ordered neuron LSTM (ON-LSTM; Shen et al. (2018)) and a GCN for TOWE. The ON-LSTM layer is an LSTM variant that considers the order of elements in the input sequence (including BERT and position embeddings) when modeling dependencies between them. The GCN encodes syntactic structural information into the representations obtained by the ON-LSTM layer.", + "bbox": [ + 507, + 175, + 884, + 336 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "BERT+BiLSTM+GCN: Mensah et al. (2021) replaces the ON-LSTM of the ONG model with a BiLSTM to better capture short-term dependencies between aspect and opinion words.", + "bbox": [ + 507, + 348, + 882, + 412 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Attention-based Relational GCN (ARGCN): Jiang et al. (2021) combine contextualized embedding obtained using BERT with a category embedding (i.e., IOB tag embedding) to incorporate aspect information. They subsequently use a relational GCN (Schlichtkrull et al., 2018) and BiLSTM to respectively incorporate syntactic and sequential information for TOWE classification.", + "bbox": [ + 507, + 424, + 884, + 552 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "4 Trading Syntax Trees for Wordpieces", + "text_level": 1, + "bbox": [ + 507, + 565, + 865, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Mensah et al. (2021) have recently demonstrated that the use of a GCN to incorporate syntax tree information has little impact in TOWE model performance. Meanwhile, the GCN presents challenges when using subword tokens, as previously mentioned. Therefore, we propose a simplified version of the TOWE model that omits the GCN component from syntax-aware approaches and instead uses subword tokens as the input to the BERT component. In this work, we use BERT's Wordpieces (Devlin et al., 2018) as the subword representation because they are highly informative, having been derived from the BERT pretraining process. However, methods such as Byte-Pair Encoding (BPE) (Sennrich et al., 2016) can also be used, as we will see later in the experiments.", + "bbox": [ + 507, + 594, + 882, + 851 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "4.1 Formatting BERT Input", + "text_level": 1, + "bbox": [ + 507, + 865, + 747, + 879 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given sentence $S$ , the BERT wordpiece tokenizer segments $S$ into a sequence of wordpieces $T =$", + "bbox": [ + 507, + 887, + 882, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1000", + "bbox": [ + 482, + 927, + 519, + 940 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/5b04d04e5154db6b28bcd673063f30635301c7817b7ac6932b42d561a4f13312.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelsGranularityLap14Res14Res15Res16Avg
ONGword75.7782.3378.8186.0180.73
ONG w/o GCNword74.1784.1078.3384.8780.37
ONG(S) w/o GCNwordpiece79.7986.6380.7288.3083.86
ONG(S,A) w/o GCNwordpiece81.7088.7082.5591.1886.03
ARGCNword76.3685.4278.2486.6981.68
ARGCN w/o R-GCNword76.3884.3678.4184.6180.94
ARGCN(S) w/o R-GCNwordpiece80.0885.9281.3689.7284.27
ARGCN(S,A) w/o R-GCNwordpiece81.3788.1882.4990.8285.72
BERT+BiLSTM+GCNword78.8285.7480.5487.3583.11
BERT+BiLSTMword78.2585.6080.4186.9482.80
BERT+BiLSTM(S)wordpiece80.4586.2780.8989.8084.35
BERT+BiLSTM(S,A)wordpiece82.5988.6082.3791.2586.20
", + "bbox": [ + 208, + 80, + 789, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 2: F1 performance of syntax-aware methods and their variants. \"Avg\" refers to the average F1 score calculated across all of the datasets. \"Granularity\" highlights the level of granularity at which input tokens are represented.", + "bbox": [ + 112, + 296, + 882, + 326 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\{t_1,t_2,\\ldots ,t_{nt}\\}$ . The BERT input for $S$ is then formatted as follows:", + "bbox": [ + 112, + 350, + 485, + 381 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nT ^ {(S)} = \\left\\{\\left[ \\mathrm {C L S} \\right], T, [ \\mathrm {S E P} ] \\right\\} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 391, + 487, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where [CLS] and [SEP] are special tokens that mark the boundaries of the sentence.", + "bbox": [ + 112, + 420, + 485, + 451 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While this format may be adequate for some NLP tasks, it can be problematic for learning good aspect representations in aspect-based sentiment classifica- tion (Tian et al., 2021). To mitigate this issue, we adopt the approach of Tian et al. (2021) and reformat the BERT input by using a sentence-aspect pair $T^{(S,A)}$ , which combines $T^{(S)}$ and $t_a$ (i.e. the aspect subsequence) along with special tokens.", + "bbox": [ + 112, + 454, + 487, + 596 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nT ^ {(S, A)} = \\left\\{\\left[ \\mathrm {C L S} \\right], T, [ \\mathrm {S E P} ], t _ {a}, [ \\mathrm {S E P} ] \\right\\} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 607, + 487, + 626 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.2 Classification and Optimization", + "text_level": 1, + "bbox": [ + 112, + 637, + 410, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The input $T^{(S,A)}$ consists of two parts: $T^{(S)}$ and $t_a$ . Since $t_a$ only serves to enhance the aspect representation in $T^{(S)}$ , sequence labeling is done on $T^{(S)}$ only. During sequence labeling, we follow the common approach of predicting based on the first wordpiece representation of a word. For instance, given the word \"surfboard\" that consists of the wordpieces \"surf\" and \"board\" which both are learned during encoding, only the representation of \"surf\" is fed to a softmax classifier to predict the tag for the whole word. The cross-entropy function is minimized for each word in the training set.", + "bbox": [ + 112, + 656, + 489, + 850 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "5 Experiments and Results", + "text_level": 1, + "bbox": [ + 112, + 862, + 366, + 878 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We experiment with the following baselines: ARGCN, BERT+BiLSTM+GCN and ONG. We", + "bbox": [ + 112, + 887, + 487, + 917 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "use the suffixes (S) or (S,A) to indicate whether the modified versions of these methods use a wordpiece sentence or wordpiece sentence-aspect pair as input, respectively. We used the publicly available code and optimal hyperparameter settings from the authors of $\\mathrm{ARGCN}^1$ and BERT+BiLSTM+GCN.2 We have implemented ONG model variants ourselves using the suggested hyperparameter configurations from the authors.3 Following previous work (Fan et al., 2019), we use the same experimental setup and evaluate on the Laptop dataset (Lap14) and the Restaurant datasets (Res14, Res15, Res16) (Pontiki et al., 2014a, 2015, 2016). The result reported for each dataset is the average over Micro F1 scores obtained from five different runs. Each run uses a different random seed to ensure the stability of our results.", + "bbox": [ + 507, + 350, + 884, + 623 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "5.1 F1 Performance Comparison", + "text_level": 1, + "bbox": [ + 507, + 636, + 784, + 651 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The results, shown in Table 2, indicate that removing the GCN component from syntax-aware approaches does not substantially impact their performance, with average decreases in performance of 0.36, 0.74, and 0.31, respectively. However, we observed a large improvement in model performance when using wordpieces, as indicated by the models with the (S) suffix. It is possible that BERT captures enough syntax information already (Clark et al., 2019) and, therefore, using GCNs to exploit syntax trees does not substantially improve", + "bbox": [ + 507, + 657, + 882, + 834 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "$^{1}$ https://github.com/samensah/encoders_towe_emnlp2021", + "bbox": [ + 507, + 843, + 878, + 868 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "$^{2}$ https://github.com/wcwowwww/towe-eacl", + "bbox": [ + 509, + 868, + 803, + 892 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "3https://github.com/samensah/Towe-TradeSyntax4WP", + "bbox": [ + 510, + 892, + 794, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1001", + "bbox": [ + 482, + 928, + 517, + 940 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/83b3f0f294de2f829e0bfaa894e97ee83c219b4780fd29b02450bfb86374e58d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelLap14Res14Res15Res16Avg
BERT-BiLSTM(S)80.4586.2780.8989.8084.35
-Mask Aspect80.0186.1180.4288.5983.78
", + "bbox": [ + 117, + 80, + 482, + 122 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 3: F1 performance of BERT-BiLSTM(S) with and without aspect masking.", + "bbox": [ + 112, + 131, + 485, + 161 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "performance on the task. This suggests that it may be beneficial to prioritize wordpieces over syntax trees to allow BERT to fully utilize rare and out-of-vocabulary words. We also discovered that using a sentence-aspect pair as input resulted in better performance than using only the sentence for the models, as indicated by the results of models with the (S,A) suffix. We believe that this may be due to the aspect information being lost or degraded during the encoding process for models with the (S) suffix. Among the methods, BERT+BiLSTM(S,A) had the highest average F1 score of 86.2.", + "bbox": [ + 112, + 187, + 489, + 381 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.2 Influence of Aspect Representation", + "text_level": 1, + "bbox": [ + 112, + 395, + 433, + 410 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To determine if the aspect representation is degraded during encoding, we evaluate BERT+BiLSTM(S) with and without aspect masking. The results, shown in Table 3, show that masking the aspect representation had only a minimal impact on performance, with a decrease in performance of 0.44 (Lap14), 0.16 (Res14), 0.47 (Res15), and 1.2 (Res16). These findings suggest that the aspect information has limited contribution and requires enhancement to improve performance, as demonstrated by the improved results of BERT+BiLSTM(S,A).", + "bbox": [ + 112, + 417, + 489, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.3 Qualitative Analysis", + "text_level": 1, + "bbox": [ + 112, + 624, + 319, + 639 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We examined the performance of BERT+BiLSTM, BERT+BiLSTM(S), and BERT+BiLSTM(S,A) on three case examples, as shown in Table 4. The results show that the BERT+BiLSTM and BERT+BiLSTM(S) models struggled to identify opinion words that were farther away from the aspect, particularly in the first and second cases where the opinion words \"beautiful\" and \"fresh\" were missed. Upon further investigation, we discovered that these opinion words were closer to the aspect's co-referential term \"it\". The model struggled to determine what \"it\" referred to due to degradation of the aspect representation, leading to the missed identification of the opinion words. However, BERT+BiLSTM(S,A) was able to recover these opinion words due to its ability to enhance the aspect representation. In the third", + "bbox": [ + 112, + 645, + 489, + 917 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "case example, the use of wordpieces was beneficial as the opinion word \"minimally\" was not present in the training set, but its wordpiece \"#lly,\" was associated with 15 opinion words in the training set. BERT+BiLSTM(S) and BERT+BiLSTM(S,A) were able to identify the opinion word \"minimally\" in the test set by leveraging the context of \"#lly\".", + "bbox": [ + 507, + 84, + 884, + 197 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "6 Impact of BPE Subword Representations", + "text_level": 1, + "bbox": [ + 507, + 211, + 756, + 243 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We previously examined the use of wordpiece representations derived from pretrained BERT for TOWE models. In this section, we look into using Byte Pair Encoding (BPE) (Sennrich et al., 2016) as an alternative method for subword representation, which is inspired by data compression techniques (Gage, 1994). It is worth noting that BPE representations are generally not obtained from pretrained BERT. However, since RoBERTa is pretrained using BPE, and RoBERTa is a variant of BERT, we can still explore the impact of using BPE representations in TOWE models. To do this, we replace the BERT component in our best model, BERT+BiLSTM(S,A), with RoBERTa, developing the model RoBERTa+BiLSTM(S,A). The results of RoBERTa+BiLSTM(S,A) and its variations are shown in Table 5.", + "bbox": [ + 507, + 255, + 884, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note, while RoBERTa+BiLSTM(S,A) and RoBERTa+BiLSTM(S) use BPE subword token representations as input, RoBERTa+BiLSTM and RoBERTa+BiLSTM+GCN operate on the word-level. Our findings support the notion that GCNs have a limited impact on performance, as demonstrated by a relatively small decrease in average F1 score when comparing RoBERTa+BiLSTM+GCN to RoBERTa+BiLSTM. On the other hand, using BPE representations instead of GCN resulted in a substantial improvement in model performance of +5.27 when comparing RoBERTa+BiLSTM and RoBERTa+BiLSTM(S). The results indicate that syntax trees via GCNs may not be necessary and can be replaced by subword representations such as BPE for better performance in TOWE. Additionally, the performance of RoBERTa+BiLSTM(S) can be further improved by using BPE-based sentence-aspect pairs, as seen by the +1.75 performance gain in RoBERTa+BiLSTM(S,A).", + "bbox": [ + 507, + 529, + 884, + 851 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "6.1 State-of-the-art Models", + "text_level": 1, + "bbox": [ + 507, + 865, + 741, + 879 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we compare the performance of BERT+BiLSTM(S,A) with recent methods,", + "bbox": [ + 507, + 887, + 882, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1002", + "bbox": [ + 482, + 927, + 521, + 940 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/1e088fbb4d21a80d601576729bb4a13f3759f0a0968743f2ec52db45e4575041.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SentenceBERT+BiLSTMBERT+BiLSTM(S)BERT+BiLSTM(S,A)
The OS is fast and fluid, everything is organi-zed and it's just beautiful.fast, fluidfast, fluidfast, fluid, beautiful
Certainly not the best sushi in new york, however, it is always fresh, and the place is very clean, sterile.freshnot the bestnot the best, fresh
Although somewhat load, the noise was min-imally intrusiveloud, intrusiveloud, minimally in-trusiveloud, minimally in-trusive.
", + "bbox": [ + 114, + 80, + 823, + 195 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/3974679c18548a37b390ecada54a8ffd85c15a7dcc00aef74e4559b9b2a5599b.jpg", + "table_caption": [ + "Table 4: Case Study: Evaluating the model performance on different case examples. Aspect words are bold-typed and opinion words are italicized." + ], + "table_footnote": [], + "table_body": "
ModelLap14Res14Res15Res16Avg
RoBERTa-BiLSTM(S,A)82.7788.2783.8491.0686.49
RoBERTa-BiLSTM(S)81.1086.9582.2188.7084.74
RoBERTa-BiLSTM75.8781.3875.9484.7079.47
RoBERTa-BiLSTM+GCN77.5782.0977.8585.3780.72
", + "bbox": [ + 115, + 256, + 497, + 322 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d2bdddeb4c3540ae4018cb7885865bf5f65c2cca0c956338d4e9141fc01b9842.jpg", + "table_caption": [ + "Table 5: F1 Performance of RoBERTa models to investigate the use of BPE subword representations." + ], + "table_footnote": [], + "table_body": "
ModelLap14Res14Res15Res16Avg
IOG71.3580.0273.2581.6976.58
LOTN72.0282.2173.2983.6277.79
SDRN+BERT*73.6983.1076.3885.4079.64
ONG75.7782.3378.8186.0180.73
ARGCN76.3685.4278.2486.6981.68
BERT+BiLSTM+GCN78.8285.7480.5487.3583.11
QD-OWSE80.3587.2380.7188.1484.11
TSMSA82.1886.3781.6489.2084.85
BERT-BiLSTM(S,A)82.5988.6082.3791.2586.20
", + "bbox": [ + 115, + 406, + 500, + 533 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 6: Performance of TOWE methods. Results for the method marked “*” are from (Jiang et al., 2021).", + "bbox": [ + 112, + 543, + 487, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "including IOG (Fan et al., 2019), LOTN (Wu et al., 2020), SDRN+BERT (Chen et al., 2020), BERT+BiLSTM+GCN (Mensah et al., 2021), QD-OWSE (Gao et al., 2021), TSMSA (Feng et al., 2021). The results of this comparison are shown in Table 6. Among these methods, the recent proposed methods QD-OWSE and TSMSA, which both use BERT as a basis for their approach, achieved competitive results with ours. QD-OWSE uses a generated question-answer pair as BERT input, while TSMSA uses multi-head attention to identify opinion words. These methods go on to demonstrate that BERT can capture sufficient syntax information for this task, even without the use of syntax trees. However, BERT+BiLSTM(S,A) achieved the best results, with F1 scores 82.59 (Lap14), 88.6 (Res14), 82.37 (Res15) and 91.25 (Res16), setting a new SOTA for the task.", + "bbox": [ + 112, + 629, + 489, + 917 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "7 Conclusion", + "text_level": 1, + "bbox": [ + 509, + 259, + 640, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We demonstrated that replacing GCNs with BERT wordpieces while enhancing the aspect representation achieves SOTA results in syntax-aware TOWE approaches. The aspect enhancement method serves as a \"prompt\" for the model. We intend to explore prompt-based learning (Brown et al., 2020) to further improve the aspect representation.", + "bbox": [ + 507, + 284, + 884, + 397 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "8 Limitations", + "text_level": 1, + "bbox": [ + 509, + 409, + 645, + 423 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Currently, our approach does not effectively leverage syntax tree information via GCNs, a commonly used method for incorporating syntax trees in this task. Further research is required to determine the most effective way to integrate syntax tree information into TOWE models.", + "bbox": [ + 507, + 432, + 882, + 529 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 509, + 542, + 680, + 558 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This work was supported by the Leverhulme Trust under Grant Number: RPG#2020#148.", + "bbox": [ + 507, + 567, + 880, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 510, + 626, + 608, + 640 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901.", + "bbox": [ + 510, + 648, + 880, + 727 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Shaowei Chen, Jie Liu, Yu Wang, Wenzheng Zhang, and Ziming Chi. 2020. Synchronous double-channel recurrent network for aspect-opinion pair extraction. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6515-6524.", + "bbox": [ + 509, + 737, + 882, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Kevin Clark, Urvashi Khandelwal, Omer Levy, and Christopher D. Manning. 2019. What does BERT look at? an analysis of BERT's attention. In Proceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 276-286, Florence, Italy. Association for Computational Linguistics.", + "bbox": [ + 509, + 825, + 884, + 917 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1003", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina N. Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186.", + "Zhifang Fan, Zhen Wu, Xinyu Dai, Shujian Huang, and Jiajun Chen. 2019. Target-oriented opinion words extraction with target-fused neural sequence labeling. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2509-2518.", + "Yuhao Feng, Yanghui Rao, Yuyao Tang, Ninghua Wang, and He Liu. 2021. Target-specified sequence labeling with multi-head self-attention for target-oriented opinion words extraction. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1805–1815.", + "Philip Gage. 1994. A new algorithm for data compression. C Users Journal, 12(2):23-38.", + "Lei Gao, Yulong Wang, Tongcun Liu, Jingyu Wang, Lei Zhang, and Jianxin Liao. 2021. Question-driven span labeling model for aspect-opinion pair extraction. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 12875-12883.", + "Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory. *Neural Computation*, 9(8):1735-1780.", + "Junfeng Jiang, An Wang, and Akiko Aizawa. 2021. Attention-based relational graph convolutional network for target-oriented opinion words extraction. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 1986-1997.", + "Hyun Duk Kim, Kavita Ganesan, Parikshit Sondhi, and ChengXiang Zhai. 2011. Comprehensive review of opinion summarization.", + "Thomas N. Kipf and Max Welling. 2017. Semi-supervised classification with graph convolutional networks. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net.", + "Samuel Mensah, Kai Sun, and Nikolaos Aletras. 2021. An empirical study on leveraging position embeddings for target-oriented opinion words extraction. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 9174-9179, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.", + "Maria Pontiki, Dimitrios Galanis, Haris Papageorgiou, Ion Androutsopoulos, Suresh Manandhar, Mohammad Al-Smadi, Mahmoud Al-Ayyoub, Yanyan Zhao," + ], + "bbox": [ + 115, + 85, + 489, + 919 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Bing Qin, Orphée De Clercq, et al. 2016. Semeval-2016 task 5: Aspect based sentiment analysis. In International workshop on semantic evaluation, pages 19-30.", + "Maria Pontiki, Dimitrios Galanis, Harris Papageorgiou, Suresh Manandhar, and Ion Androutsopoulos. 2015. Semeval-2015 task 12: Aspect based sentiment analysis. In Proceedings of the 9th international workshop on semantic evaluation (SemEval 2015), pages 486-495.", + "Maria Pontiki, Dimitrios Galanis, John Pavlopoulos, Harris Papageorgiou, Ion Androutsopoulos, and Suresh Manandhar. 2014a. Semeval-2014 task 4: Aspect based sentiment analysis. In Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), page 27-35.", + "Maria Pontiki, Dimitris Galanis, John Pavlopoulos, Harris Papageorgiou, Ion Androutsopoulos, and Suresh Manandhar. 2014b. Semeval-2014 task 4: Aspect based sentiment analysis. In Proceedings of the 8th International Workshop on Semantic Evaluation, SemEval@COLING 2014, Dublin, Ireland, August 23-24, 2014, pages 27-35. The Association for Computer Linguistics.", + "Lance A. Ramshaw and Mitch Marcus. 1995. Text chunking using transformation-based learning. In Third Workshop on Very Large Corpora, VLC@ACL 1995, Cambridge, Massachusetts, USA, June 30, 1995.", + "Michael Schlichtkrull, Thomas N. Kipf, Peter Bloem, Rianne van den Berg, Ivan Titov, and Max Welling. 2018. Modeling relational data with graph convolutional networks. In *The Semantic Web*, pages 593–607, Cham. Springer International Publishing.", + "Mike Schuster and Kaisuke Nakajima. 2012. Japanese and korean voice search. In 2012 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 5149-5152. IEEE.", + "Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Neural machine translation of rare words with subword units. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, ACL 2016, August 7-12, 2016, Berlin, Germany, Volume 1: Long Papers. The Association for Computer Linguistics.", + "Yikang Shen, Shawn Tan, Alessandro Sordoni, and Aaron C. Courville. 2018. Ordered neurons: Integrating tree structures into recurrent neural networks. In International Conference on Learning Representations.", + "Kai Sun, Richong Zhang, Mensah Samuel, Aletras Nikolaos, Yongyi Mao, and Xudong Liu. 2023. Self-training through classifier disagreement for cross-domain opinion target extraction. In Proceedings of the ACM Web Conference 2023, pages 1594-1603." + ], + "bbox": [ + 510, + 85, + 882, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1004", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Duyu Tang, Bing Qin, and Ting Liu. 2016. Aspect level sentiment classification with deep memory network. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 214-224.", + "Yuanhe Tian, Guimin Chen, and Yan Song. 2021. Aspect-based sentiment analysis with type-aware graph convolutional networks and layer ensemble. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2910-2922.", + "Amir Pouran Ben Veyseh, Nasim Nouri, Franck Der-noncourt, Dejing Dou, and Thien Huu Nguyen. 2020. Introducing syntactic structures into target opinion word extraction with deep learning. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online," + ], + "bbox": [ + 115, + 85, + 489, + 342 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "November 16-20, 2020, pages 8947-8956. Association for Computational Linguistics.", + "Zhen Wu, Fei Zhao, Xin-Yu Dai, Shujian Huang, and Jiajun Chen. 2020. Latent opinions transfer network for target-oriented opinion words extraction. In The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, The Thirty-Second Innovative Applications of Artificial Intelligence Conference, IAAI 2020, The Tenth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2020, New York, NY, USA, February 7-12, 2020, pages 9298-9305. AAAI Press.", + "Daojian Zeng, Kang Liu, Siwei Lai, Guangyou Zhou, and Jun Zhao. 2014. Relation classification via convolutional deep neural network. In Proceedings of COLING 2014, the 25th international conference on computational linguistics: technical papers, pages 2335-2344." + ], + "bbox": [ + 510, + 85, + 884, + 341 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1005", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A For every submission:", + "bbox": [ + 115, + 107, + 322, + 122 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "A1. Did you describe the limitations of your work? 7", + "A2. Did you discuss any potential risks of your work? There are no risks", + "A3. Do the abstract and introduction summarize the paper's main claims? Abstract and Section 1", + "A4. Have you used AI writing assistants when working on this paper? Left blank." + ], + "bbox": [ + 129, + 126, + 695, + 288 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "B Did you use or create scientific artifacts?", + "bbox": [ + 114, + 299, + 489, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5", + "bbox": [ + 134, + 322, + 146, + 332 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "B1. Did you cite the creators of artifacts you used? No response.", + "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? No response.", + "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? No response.", + "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? No response.", + "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? No response.", + "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. No response." + ], + "bbox": [ + 127, + 347, + 880, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "C Did you run computational experiments?", + "bbox": [ + 114, + 764, + 492, + 781 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5", + "bbox": [ + 134, + 787, + 146, + 799 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? No response.", + "bbox": [ + 127, + 813, + 880, + 860 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "ACL 2023 Responsible NLP Checklist", + "bbox": [ + 132, + 84, + 433, + 99 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.", + "bbox": [ + 112, + 868, + 877, + 892 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1006", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? No response.", + "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? No response.", + "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? No response." + ], + "bbox": [ + 127, + 84, + 880, + 282 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?", + "text_level": 1, + "bbox": [ + 112, + 293, + 877, + 310 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 132, + 313, + 213, + 329 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response.", + "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response.", + "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response.", + "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response.", + "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response." + ], + "bbox": [ + 127, + 340, + 880, + 640 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1007", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 8 + } +] \ No newline at end of file diff --git a/2023/Trading Syntax Trees for Wordpieces_ Target-oriented Opinion Words Extraction with Wordpieces and Aspect Enhancement/3b3076af-e14d-4a09-b446-5edad95e9345_model.json b/2023/Trading Syntax Trees for Wordpieces_ Target-oriented Opinion Words Extraction with Wordpieces and Aspect Enhancement/3b3076af-e14d-4a09-b446-5edad95e9345_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3b42adf9830ad024ae77379dde63186c30f0a553 --- /dev/null +++ b/2023/Trading Syntax Trees for Wordpieces_ Target-oriented Opinion Words Extraction with Wordpieces and Aspect Enhancement/3b3076af-e14d-4a09-b446-5edad95e9345_model.json @@ -0,0 +1,1670 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.128, + 0.088, + 0.873, + 0.146 + ], + "angle": 0, + "content": "Trading Syntax Trees for Wordpieces: Target-oriented Opinion Words Extraction with Wordpieces and Aspect Enhancement" + }, + { + "type": "text", + "bbox": [ + 0.125, + 0.161, + 0.376, + 0.226 + ], + "angle": 0, + "content": "Samuel Mensah \nComputer Science Department \nUniversity of Sheffield, UK \ns.mensah@sheffield.ac.uk" + }, + { + "type": "text", + "bbox": [ + 0.391, + 0.161, + 0.608, + 0.226 + ], + "angle": 0, + "content": "Kai Sun \nBDBC and SKLSDE \nBeihang University, China \nsunkai@buaa.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.624, + 0.161, + 0.877, + 0.226 + ], + "angle": 0, + "content": "Nikolaos Aletras \nComputer Science Department \nUniversity of Sheffield, UK \nn.aletras@sheffield.ac.uk" + }, + { + "type": "title", + "bbox": [ + 0.262, + 0.253, + 0.341, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.285, + 0.461, + 0.584 + ], + "angle": 0, + "content": "State-of-the-art target-oriented opinion word extraction (TOWE) models typically use BERT-based text encoders that operate on the word level, along with graph convolutional networks (GCNs) that incorporate syntactic information extracted from syntax trees. These methods achieve limited gains with GCNs and have difficulty using BERT wordpieces. Meanwhile, BERT wordpieces are known to be effective at representing rare words or words with insufficient context information. To address this issue, this work trades syntax trees for BERT wordpieces by entirely removing the GCN component from the methods' architectures. To enhance TOWE performance, we tackle the issue of aspect representation loss during encoding. Instead of solely utilizing a sentence as the input, we use a sentence-aspect pair. Our relatively simple approach achieves state-of-the-art results on benchmark datasets and should serve as a strong baseline for further research." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.601, + 0.26, + 0.616 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.629, + 0.49, + 0.852 + ], + "angle": 0, + "content": "Target-oriented opinion word extraction (TOWE; Fan et al. (2019)) is a subtask in aspect-based sentiment analysis (ABSA; Pontiki et al. (2014b)), which aims to identify words that express an opinion about a specific target (or aspect) in a sentence. For instance, in the sentence \"Such an awesome surfboard.\", a TOWE model should identify \"awesome\" as the opinion word for the given aspect surfboard. TOWE provides explicit aspect-opinion pairs which can be used to improve results in downstream tasks such as opinion summarization (Kim et al., 2011) and information extraction (Pontiki et al., 2014b; Tang et al., 2016; Sun et al., 2023)." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.855, + 0.49, + 0.92 + ], + "angle": 0, + "content": "Currently, many TOWE methods (Veyseh et al., 2020; Chen et al., 2020; Jiang et al., 2021; Feng et al., 2021; Mensah et al., 2021) use pretrained BERT (Devlin et al., 2018), to encode the input" + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.251, + 0.881, + 0.378 + ], + "angle": 0, + "content": "
1. Sentence:Wordpieces:Such an awesome surfboard‘such’, ‘an’, ‘awesome’, ‘surf’,‘##board’
2. Sentence:Wordpieces:A great snowboard which holds edges well when riding on snow.A, ‘great’, ‘snow’, ‘#’board’, ‘which’, ‘holds’, ‘edges’, ‘well’, ‘when’, ‘riding’, ‘on’, ‘snow’.
" + }, + { + "type": "table_caption", + "bbox": [ + 0.508, + 0.387, + 0.885, + 0.474 + ], + "angle": 0, + "content": "Table 1: Sentences demonstrating contextual understanding through shared wordpieces. The table shows each sentence and its corresponding BERT wordpiece sequence. Aspect words are bold-typed and opinion words are italicized. The shared wordpiece '##board' helps in decoding the meaning of \"surfboard\"." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.501, + 0.885, + 0.709 + ], + "angle": 0, + "content": "sentence. BERT has the ability to effectively capture context, which can improve TOWE performance. However, many of these methods are rather complex, as they often incorporate syntax tree information using a graph convolutional network (GCN) (Kipf and Welling, 2017). For instance, Veyseh et al. (2020) uses an ordered-neuron LSTM (Shen et al., 2018) encoder with a GCN while Jiang et al. (2021) applies an attention-based relational GCN on the syntax tree. Mensah et al. (2021) applies a BiLSTM (Hochreiter and Schmidhuber, 1997) on BERT embeddings and incorporate syntax information via a GCN." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.71, + 0.885, + 0.92 + ], + "angle": 0, + "content": "While incorporating syntax information through GCNs has been shown to provide some performance gains in TOWE, these are usually limited (Mensah et al., 2021). Moreover, modeling subword tokens with a GCN can be challenging because the syntax tree consists of whole words rather than subword tokens like wordpieces (Schuster and Nakajima, 2012; Devlin et al., 2018). Models based on subword tokens strike a good balance between character- and word-based encoders. They are able to effectively learn representations of rare words or words with insufficient context information. Consider the example in Table 1. The context" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.928, + 0.516, + 0.941 + ], + "angle": 0, + "content": "999" + }, + { + "type": "footer", + "bbox": [ + 0.227, + 0.946, + 0.77, + 0.958 + ], + "angle": 0, + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + }, + { + "type": "footer", + "bbox": [ + 0.373, + 0.959, + 0.625, + 0.972 + ], + "angle": 0, + "content": "Volume 2: Short Papers, pages 999-1007" + }, + { + "type": "footer", + "bbox": [ + 0.297, + 0.973, + 0.702, + 0.986 + ], + "angle": 0, + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.112, + 0.085, + 0.489, + 0.228 + ], + "angle": 0, + "content": "information for \"surfboard\" is limited, making it difficult to understand its meaning without additional context. However, both aspects share the wordpiece \"#board\", which allows the meaning of \"surfboard\" to be partially understood by using information from the context of \"snowboard\". In this case, \"riding\" is related to both aspects through the shared wordpiece, enabling the representation of \"surfboard\" to be improved." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.231, + 0.49, + 0.568 + ], + "angle": 0, + "content": "In this paper, we propose a substantial simplification for syntax-aware TOWE models (Veyseh et al., 2020; Jiang et al., 2021; Mensah et al., 2021) by replacing the syntax tree with subword information while maintaining good prediction performance. This is accomplished by removing the GCN from these architectures and using BERT wordpieces instead. Additionally, we address the issue of aspect representation degradation during encoding. This degradation negatively affects TOWE performance by reducing the availability of semantic information about the aspect for determining the opinion words to extract. To solve this problem, we propose using a sentence-aspect pair as input rather than just a sentence, similar to the approach used by Tian et al. (2021) for aspect-based sentiment classification. Through extensive experimentation, we found that our simple approach achieves state-of-the-art (SOTA) results by outperforming the method proposed by Mensah et al. (2021) without the need of a GCN component." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.581, + 0.318, + 0.595 + ], + "angle": 0, + "content": "2 Task Formalization" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.607, + 0.49, + 0.801 + ], + "angle": 0, + "content": "The TOWE task aims to identify an opinion word in a sentence \\( S = \\{w_{1},\\ldots ,w_{n_{s}}\\} \\) with respect to an aspect \\( w_{a}\\in S \\). The sentence is typically tokenized into a sequence of tokens at different levels of granularity (e.g. subwords or whole words), \\( T = \\{t_1,\\dots ,t_{n_t}\\} \\), with \\( t_a\\in T \\) denoting a subsequence of the aspect \\( w_{a} \\) and \\( n_s\\leq n_t \\). The goal is to assign one of three tags (I, O, or B) to each token using the IOB format (Ramshaw and Marcus, 1995), which indicates whether the word is at the Inside, Outside or Beginning of the opinion word relative to the aspect." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.813, + 0.465, + 0.831 + ], + "angle": 0, + "content": "3 Syntax-aware Approaches to TOWE" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.839, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Typically, syntax-aware approaches to TOWE (Veyseh et al., 2020; Jiang et al., 2021; Mensah et al., 2021) employ a text encoder that utilizes pretrained BERT (Devlin et al., 2018) and position embeddings (Zeng et al., 2014) (or category em" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.883, + 0.165 + ], + "angle": 0, + "content": "beddings (Jiang et al., 2021)) to learn whole word representations that are aware of the aspect's location in text. These approaches also include a GCN that operates on a syntax tree in order to incorporate syntactic information into the model." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.177, + 0.885, + 0.337 + ], + "angle": 0, + "content": "Ordered-Neuron LSTM GCN (ONG): Veyseh et al. (2020) combine an ordered neuron LSTM (ON-LSTM; Shen et al. (2018)) and a GCN for TOWE. The ON-LSTM layer is an LSTM variant that considers the order of elements in the input sequence (including BERT and position embeddings) when modeling dependencies between them. The GCN encodes syntactic structural information into the representations obtained by the ON-LSTM layer." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.349, + 0.884, + 0.413 + ], + "angle": 0, + "content": "BERT+BiLSTM+GCN: Mensah et al. (2021) replaces the ON-LSTM of the ONG model with a BiLSTM to better capture short-term dependencies between aspect and opinion words." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.425, + 0.885, + 0.553 + ], + "angle": 0, + "content": "Attention-based Relational GCN (ARGCN): Jiang et al. (2021) combine contextualized embedding obtained using BERT with a category embedding (i.e., IOB tag embedding) to incorporate aspect information. They subsequently use a relational GCN (Schlichtkrull et al., 2018) and BiLSTM to respectively incorporate syntactic and sequential information for TOWE classification." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.567, + 0.866, + 0.584 + ], + "angle": 0, + "content": "4 Trading Syntax Trees for Wordpieces" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.595, + 0.884, + 0.852 + ], + "angle": 0, + "content": "Mensah et al. (2021) have recently demonstrated that the use of a GCN to incorporate syntax tree information has little impact in TOWE model performance. Meanwhile, the GCN presents challenges when using subword tokens, as previously mentioned. Therefore, we propose a simplified version of the TOWE model that omits the GCN component from syntax-aware approaches and instead uses subword tokens as the input to the BERT component. In this work, we use BERT's Wordpieces (Devlin et al., 2018) as the subword representation because they are highly informative, having been derived from the BERT pretraining process. However, methods such as Byte-Pair Encoding (BPE) (Sennrich et al., 2016) can also be used, as we will see later in the experiments." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.866, + 0.749, + 0.881 + ], + "angle": 0, + "content": "4.1 Formatting BERT Input" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.888, + 0.883, + 0.919 + ], + "angle": 0, + "content": "Given sentence \\( S \\), the BERT wordpiece tokenizer segments \\( S \\) into a sequence of wordpieces \\( T = \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1000" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.209, + 0.082, + 0.791, + 0.288 + ], + "angle": 0, + "content": "
ModelsGranularityLap14Res14Res15Res16Avg
ONGword75.7782.3378.8186.0180.73
ONG w/o GCNword74.1784.1078.3384.8780.37
ONG(S) w/o GCNwordpiece79.7986.6380.7288.3083.86
ONG(S,A) w/o GCNwordpiece81.7088.7082.5591.1886.03
ARGCNword76.3685.4278.2486.6981.68
ARGCN w/o R-GCNword76.3884.3678.4184.6180.94
ARGCN(S) w/o R-GCNwordpiece80.0885.9281.3689.7284.27
ARGCN(S,A) w/o R-GCNwordpiece81.3788.1882.4990.8285.72
BERT+BiLSTM+GCNword78.8285.7480.5487.3583.11
BERT+BiLSTMword78.2585.6080.4186.9482.80
BERT+BiLSTM(S)wordpiece80.4586.2780.8989.8084.35
BERT+BiLSTM(S,A)wordpiece82.5988.6082.3791.2586.20
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.297, + 0.884, + 0.327 + ], + "angle": 0, + "content": "Table 2: F1 performance of syntax-aware methods and their variants. \"Avg\" refers to the average F1 score calculated across all of the datasets. \"Granularity\" highlights the level of granularity at which input tokens are represented." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.351, + 0.486, + 0.382 + ], + "angle": 0, + "content": "\\(\\{t_1,t_2,\\ldots ,t_{nt}\\}\\) . The BERT input for \\(S\\) is then formatted as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.392, + 0.488, + 0.412 + ], + "angle": 0, + "content": "\\[\nT ^ {(S)} = \\left\\{\\left[ \\mathrm {C L S} \\right], T, [ \\mathrm {S E P} ] \\right\\} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.421, + 0.487, + 0.452 + ], + "angle": 0, + "content": "where [CLS] and [SEP] are special tokens that mark the boundaries of the sentence." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.455, + 0.489, + 0.598 + ], + "angle": 0, + "content": "While this format may be adequate for some NLP tasks, it can be problematic for learning good aspect representations in aspect-based sentiment classifica- tion (Tian et al., 2021). To mitigate this issue, we adopt the approach of Tian et al. (2021) and reformat the BERT input by using a sentence-aspect pair \\( T^{(S,A)} \\), which combines \\( T^{(S)} \\) and \\( t_a \\) (i.e. the aspect subsequence) along with special tokens." + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.608, + 0.488, + 0.627 + ], + "angle": 0, + "content": "\\[\nT ^ {(S, A)} = \\left\\{\\left[ \\mathrm {C L S} \\right], T, [ \\mathrm {S E P} ], t _ {a}, [ \\mathrm {S E P} ] \\right\\} \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.638, + 0.411, + 0.653 + ], + "angle": 0, + "content": "4.2 Classification and Optimization" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.657, + 0.49, + 0.851 + ], + "angle": 0, + "content": "The input \\( T^{(S,A)} \\) consists of two parts: \\( T^{(S)} \\) and \\( t_a \\). Since \\( t_a \\) only serves to enhance the aspect representation in \\( T^{(S)} \\), sequence labeling is done on \\( T^{(S)} \\) only. During sequence labeling, we follow the common approach of predicting based on the first wordpiece representation of a word. For instance, given the word \"surfboard\" that consists of the wordpieces \"surf\" and \"board\" which both are learned during encoding, only the representation of \"surf\" is fed to a softmax classifier to predict the tag for the whole word. The cross-entropy function is minimized for each word in the training set." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.863, + 0.367, + 0.879 + ], + "angle": 0, + "content": "5 Experiments and Results" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.888, + 0.489, + 0.918 + ], + "angle": 0, + "content": "We experiment with the following baselines: ARGCN, BERT+BiLSTM+GCN and ONG. We" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.351, + 0.885, + 0.624 + ], + "angle": 0, + "content": "use the suffixes (S) or (S,A) to indicate whether the modified versions of these methods use a wordpiece sentence or wordpiece sentence-aspect pair as input, respectively. We used the publicly available code and optimal hyperparameter settings from the authors of \\(\\mathrm{ARGCN}^1\\) and BERT+BiLSTM+GCN.2 We have implemented ONG model variants ourselves using the suggested hyperparameter configurations from the authors.3 Following previous work (Fan et al., 2019), we use the same experimental setup and evaluate on the Laptop dataset (Lap14) and the Restaurant datasets (Res14, Res15, Res16) (Pontiki et al., 2014a, 2015, 2016). The result reported for each dataset is the average over Micro F1 scores obtained from five different runs. Each run uses a different random seed to ensure the stability of our results." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.637, + 0.785, + 0.652 + ], + "angle": 0, + "content": "5.1 F1 Performance Comparison" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.658, + 0.884, + 0.835 + ], + "angle": 0, + "content": "The results, shown in Table 2, indicate that removing the GCN component from syntax-aware approaches does not substantially impact their performance, with average decreases in performance of 0.36, 0.74, and 0.31, respectively. However, we observed a large improvement in model performance when using wordpieces, as indicated by the models with the (S) suffix. It is possible that BERT captures enough syntax information already (Clark et al., 2019) and, therefore, using GCNs to exploit syntax trees does not substantially improve" + }, + { + "type": "page_footnote", + "bbox": [ + 0.509, + 0.844, + 0.879, + 0.869 + ], + "angle": 0, + "content": "\\(^{1}\\)https://github.com/samensah/encoders_towe_emnlp2021" + }, + { + "type": "page_footnote", + "bbox": [ + 0.51, + 0.869, + 0.804, + 0.893 + ], + "angle": 0, + "content": "\\(^{2}\\)https://github.com/wcwowwww/towe-eacl" + }, + { + "type": "page_footnote", + "bbox": [ + 0.511, + 0.894, + 0.796, + 0.919 + ], + "angle": 0, + "content": "3https://github.com/samensah/Towe-TradeSyntax4WP" + }, + { + "type": "list", + "bbox": [ + 0.509, + 0.844, + 0.879, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.519, + 0.941 + ], + "angle": 0, + "content": "1001" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.118, + 0.082, + 0.484, + 0.123 + ], + "angle": 0, + "content": "
ModelLap14Res14Res15Res16Avg
BERT-BiLSTM(S)80.4586.2780.8989.8084.35
-Mask Aspect80.0186.1180.4288.5983.78
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.133, + 0.486, + 0.162 + ], + "angle": 0, + "content": "Table 3: F1 performance of BERT-BiLSTM(S) with and without aspect masking." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.189, + 0.49, + 0.382 + ], + "angle": 0, + "content": "performance on the task. This suggests that it may be beneficial to prioritize wordpieces over syntax trees to allow BERT to fully utilize rare and out-of-vocabulary words. We also discovered that using a sentence-aspect pair as input resulted in better performance than using only the sentence for the models, as indicated by the results of models with the (S,A) suffix. We believe that this may be due to the aspect information being lost or degraded during the encoding process for models with the (S) suffix. Among the methods, BERT+BiLSTM(S,A) had the highest average F1 score of 86.2." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.396, + 0.435, + 0.411 + ], + "angle": 0, + "content": "5.2 Influence of Aspect Representation" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.418, + 0.49, + 0.61 + ], + "angle": 0, + "content": "To determine if the aspect representation is degraded during encoding, we evaluate BERT+BiLSTM(S) with and without aspect masking. The results, shown in Table 3, show that masking the aspect representation had only a minimal impact on performance, with a decrease in performance of 0.44 (Lap14), 0.16 (Res14), 0.47 (Res15), and 1.2 (Res16). These findings suggest that the aspect information has limited contribution and requires enhancement to improve performance, as demonstrated by the improved results of BERT+BiLSTM(S,A)." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.625, + 0.32, + 0.64 + ], + "angle": 0, + "content": "5.3 Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.646, + 0.49, + 0.919 + ], + "angle": 0, + "content": "We examined the performance of BERT+BiLSTM, BERT+BiLSTM(S), and BERT+BiLSTM(S,A) on three case examples, as shown in Table 4. The results show that the BERT+BiLSTM and BERT+BiLSTM(S) models struggled to identify opinion words that were farther away from the aspect, particularly in the first and second cases where the opinion words \"beautiful\" and \"fresh\" were missed. Upon further investigation, we discovered that these opinion words were closer to the aspect's co-referential term \"it\". The model struggled to determine what \"it\" referred to due to degradation of the aspect representation, leading to the missed identification of the opinion words. However, BERT+BiLSTM(S,A) was able to recover these opinion words due to its ability to enhance the aspect representation. In the third" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.198 + ], + "angle": 0, + "content": "case example, the use of wordpieces was beneficial as the opinion word \"minimally\" was not present in the training set, but its wordpiece \"#lly,\" was associated with 15 opinion words in the training set. BERT+BiLSTM(S) and BERT+BiLSTM(S,A) were able to identify the opinion word \"minimally\" in the test set by leveraging the context of \"#lly\"." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.212, + 0.757, + 0.244 + ], + "angle": 0, + "content": "6 Impact of BPE Subword Representations" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.256, + 0.885, + 0.528 + ], + "angle": 0, + "content": "We previously examined the use of wordpiece representations derived from pretrained BERT for TOWE models. In this section, we look into using Byte Pair Encoding (BPE) (Sennrich et al., 2016) as an alternative method for subword representation, which is inspired by data compression techniques (Gage, 1994). It is worth noting that BPE representations are generally not obtained from pretrained BERT. However, since RoBERTa is pretrained using BPE, and RoBERTa is a variant of BERT, we can still explore the impact of using BPE representations in TOWE models. To do this, we replace the BERT component in our best model, BERT+BiLSTM(S,A), with RoBERTa, developing the model RoBERTa+BiLSTM(S,A). The results of RoBERTa+BiLSTM(S,A) and its variations are shown in Table 5." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.53, + 0.885, + 0.852 + ], + "angle": 0, + "content": "Note, while RoBERTa+BiLSTM(S,A) and RoBERTa+BiLSTM(S) use BPE subword token representations as input, RoBERTa+BiLSTM and RoBERTa+BiLSTM+GCN operate on the word-level. Our findings support the notion that GCNs have a limited impact on performance, as demonstrated by a relatively small decrease in average F1 score when comparing RoBERTa+BiLSTM+GCN to RoBERTa+BiLSTM. On the other hand, using BPE representations instead of GCN resulted in a substantial improvement in model performance of +5.27 when comparing RoBERTa+BiLSTM and RoBERTa+BiLSTM(S). The results indicate that syntax trees via GCNs may not be necessary and can be replaced by subword representations such as BPE for better performance in TOWE. Additionally, the performance of RoBERTa+BiLSTM(S) can be further improved by using BPE-based sentence-aspect pairs, as seen by the +1.75 performance gain in RoBERTa+BiLSTM(S,A)." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.866, + 0.742, + 0.88 + ], + "angle": 0, + "content": "6.1 State-of-the-art Models" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.888, + 0.884, + 0.919 + ], + "angle": 0, + "content": "Finally, we compare the performance of BERT+BiLSTM(S,A) with recent methods," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.928, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1002" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.115, + 0.082, + 0.825, + 0.196 + ], + "angle": 0, + "content": "
SentenceBERT+BiLSTMBERT+BiLSTM(S)BERT+BiLSTM(S,A)
The OS is fast and fluid, everything is organi-zed and it's just beautiful.fast, fluidfast, fluidfast, fluid, beautiful
Certainly not the best sushi in new york, however, it is always fresh, and the place is very clean, sterile.freshnot the bestnot the best, fresh
Although somewhat load, the noise was min-imally intrusiveloud, intrusiveloud, minimally in-trusiveloud, minimally in-trusive.
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.206, + 0.884, + 0.234 + ], + "angle": 0, + "content": "Table 4: Case Study: Evaluating the model performance on different case examples. Aspect words are bold-typed and opinion words are italicized." + }, + { + "type": "table", + "bbox": [ + 0.116, + 0.257, + 0.499, + 0.323 + ], + "angle": 0, + "content": "
ModelLap14Res14Res15Res16Avg
RoBERTa-BiLSTM(S,A)82.7788.2783.8491.0686.49
RoBERTa-BiLSTM(S)81.1086.9582.2188.7084.74
RoBERTa-BiLSTM75.8781.3875.9484.7079.47
RoBERTa-BiLSTM+GCN77.5782.0977.8585.3780.72
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.333, + 0.49, + 0.362 + ], + "angle": 0, + "content": "Table 5: F1 Performance of RoBERTa models to investigate the use of BPE subword representations." + }, + { + "type": "table", + "bbox": [ + 0.116, + 0.407, + 0.501, + 0.535 + ], + "angle": 0, + "content": "
ModelLap14Res14Res15Res16Avg
IOG71.3580.0273.2581.6976.58
LOTN72.0282.2173.2983.6277.79
SDRN+BERT*73.6983.1076.3885.4079.64
ONG75.7782.3378.8186.0180.73
ARGCN76.3685.4278.2486.6981.68
BERT+BiLSTM+GCN78.8285.7480.5487.3583.11
QD-OWSE80.3587.2380.7188.1484.11
TSMSA82.1886.3781.6489.2084.85
BERT-BiLSTM(S,A)82.5988.6082.3791.2586.20
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.544, + 0.488, + 0.573 + ], + "angle": 0, + "content": "Table 6: Performance of TOWE methods. Results for the method marked “*” are from (Jiang et al., 2021)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.63, + 0.49, + 0.919 + ], + "angle": 0, + "content": "including IOG (Fan et al., 2019), LOTN (Wu et al., 2020), SDRN+BERT (Chen et al., 2020), BERT+BiLSTM+GCN (Mensah et al., 2021), QD-OWSE (Gao et al., 2021), TSMSA (Feng et al., 2021). The results of this comparison are shown in Table 6. Among these methods, the recent proposed methods QD-OWSE and TSMSA, which both use BERT as a basis for their approach, achieved competitive results with ours. QD-OWSE uses a generated question-answer pair as BERT input, while TSMSA uses multi-head attention to identify opinion words. These methods go on to demonstrate that BERT can capture sufficient syntax information for this task, even without the use of syntax trees. However, BERT+BiLSTM(S,A) achieved the best results, with F1 scores 82.59 (Lap14), 88.6 (Res14), 82.37 (Res15) and 91.25 (Res16), setting a new SOTA for the task." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.26, + 0.642, + 0.274 + ], + "angle": 0, + "content": "7 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.285, + 0.885, + 0.398 + ], + "angle": 0, + "content": "We demonstrated that replacing GCNs with BERT wordpieces while enhancing the aspect representation achieves SOTA results in syntax-aware TOWE approaches. The aspect enhancement method serves as a \"prompt\" for the model. We intend to explore prompt-based learning (Brown et al., 2020) to further improve the aspect representation." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.41, + 0.646, + 0.424 + ], + "angle": 0, + "content": "8 Limitations" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.434, + 0.884, + 0.53 + ], + "angle": 0, + "content": "Currently, our approach does not effectively leverage syntax tree information via GCNs, a commonly used method for incorporating syntax trees in this task. Further research is required to determine the most effective way to integrate syntax tree information into TOWE models." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.543, + 0.682, + 0.559 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.568, + 0.882, + 0.599 + ], + "angle": 0, + "content": "This work was supported by the Leverhulme Trust under Grant Number: RPG#2020#148." + }, + { + "type": "title", + "bbox": [ + 0.511, + 0.627, + 0.61, + 0.642 + ], + "angle": 0, + "content": "References" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.649, + 0.882, + 0.728 + ], + "angle": 0, + "content": "Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901." + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.738, + 0.884, + 0.817 + ], + "angle": 0, + "content": "Shaowei Chen, Jie Liu, Yu Wang, Wenzheng Zhang, and Ziming Chi. 2020. Synchronous double-channel recurrent network for aspect-opinion pair extraction. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6515-6524." + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.826, + 0.885, + 0.919 + ], + "angle": 0, + "content": "Kevin Clark, Urvashi Khandelwal, Omer Levy, and Christopher D. Manning. 2019. What does BERT look at? an analysis of BERT's attention. In Proceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 276-286, Florence, Italy. Association for Computational Linguistics." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1003" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.19 + ], + "angle": 0, + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina N. Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.2, + 0.49, + 0.292 + ], + "angle": 0, + "content": "Zhifang Fan, Zhen Wu, Xinyu Dai, Shujian Huang, and Jiajun Chen. 2019. Target-oriented opinion words extraction with target-fused neural sequence labeling. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2509-2518." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.3, + 0.488, + 0.393 + ], + "angle": 0, + "content": "Yuhao Feng, Yanghui Rao, Yuyao Tang, Ninghua Wang, and He Liu. 2021. Target-specified sequence labeling with multi-head self-attention for target-oriented opinion words extraction. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1805–1815." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.401, + 0.488, + 0.427 + ], + "angle": 0, + "content": "Philip Gage. 1994. A new algorithm for data compression. C Users Journal, 12(2):23-38." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.436, + 0.488, + 0.502 + ], + "angle": 0, + "content": "Lei Gao, Yulong Wang, Tongcun Liu, Jingyu Wang, Lei Zhang, and Jianxin Liao. 2021. Question-driven span labeling model for aspect-opinion pair extraction. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 12875-12883." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.511, + 0.488, + 0.55 + ], + "angle": 0, + "content": "Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory. *Neural Computation*, 9(8):1735-1780." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.56, + 0.488, + 0.638 + ], + "angle": 0, + "content": "Junfeng Jiang, An Wang, and Akiko Aizawa. 2021. Attention-based relational graph convolutional network for target-oriented opinion words extraction. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 1986-1997." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.647, + 0.488, + 0.686 + ], + "angle": 0, + "content": "Hyun Duk Kim, Kavita Ganesan, Parikshit Sondhi, and ChengXiang Zhai. 2011. Comprehensive review of opinion summarization." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.695, + 0.488, + 0.774 + ], + "angle": 0, + "content": "Thomas N. Kipf and Max Welling. 2017. Semi-supervised classification with graph convolutional networks. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.783, + 0.488, + 0.875 + ], + "angle": 0, + "content": "Samuel Mensah, Kai Sun, and Nikolaos Aletras. 2021. An empirical study on leveraging position embeddings for target-oriented opinion words extraction. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 9174-9179, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.881, + 0.488, + 0.92 + ], + "angle": 0, + "content": "Maria Pontiki, Dimitrios Galanis, Haris Papageorgiou, Ion Androutsopoulos, Suresh Manandhar, Mohammad Al-Smadi, Mahmoud Al-Ayyoub, Yanyan Zhao," + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.92 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.529, + 0.086, + 0.884, + 0.138 + ], + "angle": 0, + "content": "Bing Qin, Orphée De Clercq, et al. 2016. Semeval-2016 task 5: Aspect based sentiment analysis. In International workshop on semantic evaluation, pages 19-30." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.149, + 0.884, + 0.227 + ], + "angle": 0, + "content": "Maria Pontiki, Dimitrios Galanis, Harris Papageorgiou, Suresh Manandhar, and Ion Androutsopoulos. 2015. Semeval-2015 task 12: Aspect based sentiment analysis. In Proceedings of the 9th international workshop on semantic evaluation (SemEval 2015), pages 486-495." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.238, + 0.883, + 0.317 + ], + "angle": 0, + "content": "Maria Pontiki, Dimitrios Galanis, John Pavlopoulos, Harris Papageorgiou, Ion Androutsopoulos, and Suresh Manandhar. 2014a. Semeval-2014 task 4: Aspect based sentiment analysis. In Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), page 27-35." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.327, + 0.883, + 0.431 + ], + "angle": 0, + "content": "Maria Pontiki, Dimitris Galanis, John Pavlopoulos, Harris Papageorgiou, Ion Androutsopoulos, and Suresh Manandhar. 2014b. Semeval-2014 task 4: Aspect based sentiment analysis. In Proceedings of the 8th International Workshop on Semantic Evaluation, SemEval@COLING 2014, Dublin, Ireland, August 23-24, 2014, pages 27-35. The Association for Computer Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.442, + 0.883, + 0.506 + ], + "angle": 0, + "content": "Lance A. Ramshaw and Mitch Marcus. 1995. Text chunking using transformation-based learning. In Third Workshop on Very Large Corpora, VLC@ACL 1995, Cambridge, Massachusetts, USA, June 30, 1995." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.517, + 0.883, + 0.583 + ], + "angle": 0, + "content": "Michael Schlichtkrull, Thomas N. Kipf, Peter Bloem, Rianne van den Berg, Ivan Titov, and Max Welling. 2018. Modeling relational data with graph convolutional networks. In *The Semantic Web*, pages 593–607, Cham. Springer International Publishing." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.593, + 0.883, + 0.646 + ], + "angle": 0, + "content": "Mike Schuster and Kaisuke Nakajima. 2012. Japanese and korean voice search. In 2012 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 5149-5152. IEEE." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.656, + 0.883, + 0.747 + ], + "angle": 0, + "content": "Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Neural machine translation of rare words with subword units. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, ACL 2016, August 7-12, 2016, Berlin, Germany, Volume 1: Long Papers. The Association for Computer Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.757, + 0.883, + 0.822 + ], + "angle": 0, + "content": "Yikang Shen, Shawn Tan, Alessandro Sordoni, and Aaron C. Courville. 2018. Ordered neurons: Integrating tree structures into recurrent neural networks. In International Conference on Learning Representations." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.833, + 0.883, + 0.899 + ], + "angle": 0, + "content": "Kai Sun, Richong Zhang, Mensah Samuel, Aletras Nikolaos, Yongyi Mao, and Xudong Liu. 2023. Self-training through classifier disagreement for cross-domain opinion target extraction. In Proceedings of the ACM Web Conference 2023, pages 1594-1603." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.884, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1004" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.151 + ], + "angle": 0, + "content": "Duyu Tang, Bing Qin, and Ting Liu. 2016. Aspect level sentiment classification with deep memory network. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 214-224." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.162, + 0.49, + 0.254 + ], + "angle": 0, + "content": "Yuanhe Tian, Guimin Chen, and Yan Song. 2021. Aspect-based sentiment analysis with type-aware graph convolutional networks and layer ensemble. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2910-2922." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.264, + 0.49, + 0.343 + ], + "angle": 0, + "content": "Amir Pouran Ben Veyseh, Nasim Nouri, Franck Der-noncourt, Dejing Dou, and Thien Huu Nguyen. 2020. Introducing syntactic structures into target opinion word extraction with deep learning. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online," + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.529, + 0.086, + 0.884, + 0.113 + ], + "angle": 0, + "content": "November 16-20, 2020, pages 8947-8956. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.123, + 0.885, + 0.253 + ], + "angle": 0, + "content": "Zhen Wu, Fei Zhao, Xin-Yu Dai, Shujian Huang, and Jiajun Chen. 2020. Latent opinions transfer network for target-oriented opinion words extraction. In The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, The Thirty-Second Innovative Applications of Artificial Intelligence Conference, IAAI 2020, The Tenth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2020, New York, NY, USA, February 7-12, 2020, pages 9298-9305. AAAI Press." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.264, + 0.885, + 0.342 + ], + "angle": 0, + "content": "Daojian Zeng, Kang Liu, Siwei Lai, Guangyou Zhou, and Jun Zhao. 2014. Relation classification via convolutional deep neural network. In Proceedings of COLING 2014, the 25th international conference on computational linguistics: technical papers, pages 2335-2344." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.885, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1005" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.133, + 0.085, + 0.435, + 0.1 + ], + "angle": 0, + "content": "ACL 2023 Responsible NLP Checklist" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.108, + 0.323, + 0.123 + ], + "angle": 0, + "content": "A For every submission:" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.127, + 0.533, + 0.158 + ], + "angle": 0, + "content": "A1. Did you describe the limitations of your work? 7" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.171, + 0.553, + 0.201 + ], + "angle": 0, + "content": "A2. Did you discuss any potential risks of your work? There are no risks" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.212, + 0.697, + 0.245 + ], + "angle": 0, + "content": "A3. Do the abstract and introduction summarize the paper's main claims? Abstract and Section 1" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.256, + 0.669, + 0.289 + ], + "angle": 0, + "content": "A4. Have you used AI writing assistants when working on this paper? Left blank." + }, + { + "type": "list", + "bbox": [ + 0.13, + 0.127, + 0.697, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.3, + 0.49, + 0.317 + ], + "angle": 0, + "content": "B Did you use or create scientific artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.323, + 0.147, + 0.334 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.348, + 0.531, + 0.38 + ], + "angle": 0, + "content": "B1. Did you cite the creators of artifacts you used? No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.391, + 0.779, + 0.424 + ], + "angle": 0, + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.434, + 0.881, + 0.514 + ], + "angle": 0, + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.525, + 0.881, + 0.588 + ], + "angle": 0, + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.6, + 0.881, + 0.648 + ], + "angle": 0, + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.659, + 0.881, + 0.755 + ], + "angle": 0, + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. No response." + }, + { + "type": "list", + "bbox": [ + 0.129, + 0.348, + 0.881, + 0.755 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.765, + 0.494, + 0.782 + ], + "angle": 0, + "content": "C Did you run computational experiments?" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.788, + 0.147, + 0.8 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.814, + 0.881, + 0.861 + ], + "angle": 0, + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? No response." + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.869, + 0.878, + 0.894 + ], + "angle": 0, + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1006" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.129, + 0.085, + 0.88, + 0.133 + ], + "angle": 0, + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.144, + 0.881, + 0.207 + ], + "angle": 0, + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.219, + 0.881, + 0.283 + ], + "angle": 0, + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? No response." + }, + { + "type": "list", + "bbox": [ + 0.129, + 0.085, + 0.881, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.294, + 0.878, + 0.311 + ], + "angle": 0, + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.315, + 0.215, + 0.33 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.341, + 0.881, + 0.388 + ], + "angle": 0, + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.4, + 0.881, + 0.464 + ], + "angle": 0, + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.475, + 0.881, + 0.54 + ], + "angle": 0, + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.55, + 0.873, + 0.582 + ], + "angle": 0, + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.593, + 0.881, + 0.641 + ], + "angle": 0, + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response." + }, + { + "type": "list", + "bbox": [ + 0.129, + 0.341, + 0.881, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1007" + } + ] +] \ No newline at end of file diff --git a/2023/Trading Syntax Trees for Wordpieces_ Target-oriented Opinion Words Extraction with Wordpieces and Aspect Enhancement/3b3076af-e14d-4a09-b446-5edad95e9345_origin.pdf b/2023/Trading Syntax Trees for Wordpieces_ Target-oriented Opinion Words Extraction with Wordpieces and Aspect Enhancement/3b3076af-e14d-4a09-b446-5edad95e9345_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..04e2ea58a7d6b60b27f2e4de6585864a22b9311c --- /dev/null +++ b/2023/Trading Syntax Trees for Wordpieces_ Target-oriented Opinion Words Extraction with Wordpieces and Aspect Enhancement/3b3076af-e14d-4a09-b446-5edad95e9345_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f59899c520b22ab412d8053a405b3382d2985eb44b03d69a2676d70eb2909930 +size 228954 diff --git a/2023/Trading Syntax Trees for Wordpieces_ Target-oriented Opinion Words Extraction with Wordpieces and Aspect Enhancement/full.md b/2023/Trading Syntax Trees for Wordpieces_ Target-oriented Opinion Words Extraction with Wordpieces and Aspect Enhancement/full.md new file mode 100644 index 0000000000000000000000000000000000000000..65ef57697e6de02201b44da00e070a72f9ebdc6c --- /dev/null +++ b/2023/Trading Syntax Trees for Wordpieces_ Target-oriented Opinion Words Extraction with Wordpieces and Aspect Enhancement/full.md @@ -0,0 +1,223 @@ +# Trading Syntax Trees for Wordpieces: Target-oriented Opinion Words Extraction with Wordpieces and Aspect Enhancement + +Samuel Mensah +Computer Science Department +University of Sheffield, UK +s.mensah@sheffield.ac.uk + +Kai Sun +BDBC and SKLSDE +Beihang University, China +sunkai@buaa.edu.cn + +Nikolaos Aletras +Computer Science Department +University of Sheffield, UK +n.aletras@sheffield.ac.uk + +# Abstract + +State-of-the-art target-oriented opinion word extraction (TOWE) models typically use BERT-based text encoders that operate on the word level, along with graph convolutional networks (GCNs) that incorporate syntactic information extracted from syntax trees. These methods achieve limited gains with GCNs and have difficulty using BERT wordpieces. Meanwhile, BERT wordpieces are known to be effective at representing rare words or words with insufficient context information. To address this issue, this work trades syntax trees for BERT wordpieces by entirely removing the GCN component from the methods' architectures. To enhance TOWE performance, we tackle the issue of aspect representation loss during encoding. Instead of solely utilizing a sentence as the input, we use a sentence-aspect pair. Our relatively simple approach achieves state-of-the-art results on benchmark datasets and should serve as a strong baseline for further research. + +# 1 Introduction + +Target-oriented opinion word extraction (TOWE; Fan et al. (2019)) is a subtask in aspect-based sentiment analysis (ABSA; Pontiki et al. (2014b)), which aims to identify words that express an opinion about a specific target (or aspect) in a sentence. For instance, in the sentence "Such an awesome surfboard.", a TOWE model should identify "awesome" as the opinion word for the given aspect surfboard. TOWE provides explicit aspect-opinion pairs which can be used to improve results in downstream tasks such as opinion summarization (Kim et al., 2011) and information extraction (Pontiki et al., 2014b; Tang et al., 2016; Sun et al., 2023). + +Currently, many TOWE methods (Veyseh et al., 2020; Chen et al., 2020; Jiang et al., 2021; Feng et al., 2021; Mensah et al., 2021) use pretrained BERT (Devlin et al., 2018), to encode the input + +
1. Sentence:Wordpieces:Such an awesome surfboard‘such’, ‘an’, ‘awesome’, ‘surf’,‘##board’
2. Sentence:Wordpieces:A great snowboard which holds edges well when riding on snow.A, ‘great’, ‘snow’, ‘#’board’, ‘which’, ‘holds’, ‘edges’, ‘well’, ‘when’, ‘riding’, ‘on’, ‘snow’.
+ +Table 1: Sentences demonstrating contextual understanding through shared wordpieces. The table shows each sentence and its corresponding BERT wordpiece sequence. Aspect words are bold-typed and opinion words are italicized. The shared wordpiece '##board' helps in decoding the meaning of "surfboard". + +sentence. BERT has the ability to effectively capture context, which can improve TOWE performance. However, many of these methods are rather complex, as they often incorporate syntax tree information using a graph convolutional network (GCN) (Kipf and Welling, 2017). For instance, Veyseh et al. (2020) uses an ordered-neuron LSTM (Shen et al., 2018) encoder with a GCN while Jiang et al. (2021) applies an attention-based relational GCN on the syntax tree. Mensah et al. (2021) applies a BiLSTM (Hochreiter and Schmidhuber, 1997) on BERT embeddings and incorporate syntax information via a GCN. + +While incorporating syntax information through GCNs has been shown to provide some performance gains in TOWE, these are usually limited (Mensah et al., 2021). Moreover, modeling subword tokens with a GCN can be challenging because the syntax tree consists of whole words rather than subword tokens like wordpieces (Schuster and Nakajima, 2012; Devlin et al., 2018). Models based on subword tokens strike a good balance between character- and word-based encoders. They are able to effectively learn representations of rare words or words with insufficient context information. Consider the example in Table 1. The context + +information for "surfboard" is limited, making it difficult to understand its meaning without additional context. However, both aspects share the wordpiece "#board", which allows the meaning of "surfboard" to be partially understood by using information from the context of "snowboard". In this case, "riding" is related to both aspects through the shared wordpiece, enabling the representation of "surfboard" to be improved. + +In this paper, we propose a substantial simplification for syntax-aware TOWE models (Veyseh et al., 2020; Jiang et al., 2021; Mensah et al., 2021) by replacing the syntax tree with subword information while maintaining good prediction performance. This is accomplished by removing the GCN from these architectures and using BERT wordpieces instead. Additionally, we address the issue of aspect representation degradation during encoding. This degradation negatively affects TOWE performance by reducing the availability of semantic information about the aspect for determining the opinion words to extract. To solve this problem, we propose using a sentence-aspect pair as input rather than just a sentence, similar to the approach used by Tian et al. (2021) for aspect-based sentiment classification. Through extensive experimentation, we found that our simple approach achieves state-of-the-art (SOTA) results by outperforming the method proposed by Mensah et al. (2021) without the need of a GCN component. + +# 2 Task Formalization + +The TOWE task aims to identify an opinion word in a sentence $S = \{w_{1},\ldots ,w_{n_{s}}\}$ with respect to an aspect $w_{a}\in S$ . The sentence is typically tokenized into a sequence of tokens at different levels of granularity (e.g. subwords or whole words), $T = \{t_1,\dots ,t_{n_t}\}$ , with $t_a\in T$ denoting a subsequence of the aspect $w_{a}$ and $n_s\leq n_t$ . The goal is to assign one of three tags (I, O, or B) to each token using the IOB format (Ramshaw and Marcus, 1995), which indicates whether the word is at the Inside, Outside or Beginning of the opinion word relative to the aspect. + +# 3 Syntax-aware Approaches to TOWE + +Typically, syntax-aware approaches to TOWE (Veyseh et al., 2020; Jiang et al., 2021; Mensah et al., 2021) employ a text encoder that utilizes pretrained BERT (Devlin et al., 2018) and position embeddings (Zeng et al., 2014) (or category em + +beddings (Jiang et al., 2021)) to learn whole word representations that are aware of the aspect's location in text. These approaches also include a GCN that operates on a syntax tree in order to incorporate syntactic information into the model. + +Ordered-Neuron LSTM GCN (ONG): Veyseh et al. (2020) combine an ordered neuron LSTM (ON-LSTM; Shen et al. (2018)) and a GCN for TOWE. The ON-LSTM layer is an LSTM variant that considers the order of elements in the input sequence (including BERT and position embeddings) when modeling dependencies between them. The GCN encodes syntactic structural information into the representations obtained by the ON-LSTM layer. + +BERT+BiLSTM+GCN: Mensah et al. (2021) replaces the ON-LSTM of the ONG model with a BiLSTM to better capture short-term dependencies between aspect and opinion words. + +Attention-based Relational GCN (ARGCN): Jiang et al. (2021) combine contextualized embedding obtained using BERT with a category embedding (i.e., IOB tag embedding) to incorporate aspect information. They subsequently use a relational GCN (Schlichtkrull et al., 2018) and BiLSTM to respectively incorporate syntactic and sequential information for TOWE classification. + +# 4 Trading Syntax Trees for Wordpieces + +Mensah et al. (2021) have recently demonstrated that the use of a GCN to incorporate syntax tree information has little impact in TOWE model performance. Meanwhile, the GCN presents challenges when using subword tokens, as previously mentioned. Therefore, we propose a simplified version of the TOWE model that omits the GCN component from syntax-aware approaches and instead uses subword tokens as the input to the BERT component. In this work, we use BERT's Wordpieces (Devlin et al., 2018) as the subword representation because they are highly informative, having been derived from the BERT pretraining process. However, methods such as Byte-Pair Encoding (BPE) (Sennrich et al., 2016) can also be used, as we will see later in the experiments. + +# 4.1 Formatting BERT Input + +Given sentence $S$ , the BERT wordpiece tokenizer segments $S$ into a sequence of wordpieces $T =$ + +
ModelsGranularityLap14Res14Res15Res16Avg
ONGword75.7782.3378.8186.0180.73
ONG w/o GCNword74.1784.1078.3384.8780.37
ONG(S) w/o GCNwordpiece79.7986.6380.7288.3083.86
ONG(S,A) w/o GCNwordpiece81.7088.7082.5591.1886.03
ARGCNword76.3685.4278.2486.6981.68
ARGCN w/o R-GCNword76.3884.3678.4184.6180.94
ARGCN(S) w/o R-GCNwordpiece80.0885.9281.3689.7284.27
ARGCN(S,A) w/o R-GCNwordpiece81.3788.1882.4990.8285.72
BERT+BiLSTM+GCNword78.8285.7480.5487.3583.11
BERT+BiLSTMword78.2585.6080.4186.9482.80
BERT+BiLSTM(S)wordpiece80.4586.2780.8989.8084.35
BERT+BiLSTM(S,A)wordpiece82.5988.6082.3791.2586.20
+ +Table 2: F1 performance of syntax-aware methods and their variants. "Avg" refers to the average F1 score calculated across all of the datasets. "Granularity" highlights the level of granularity at which input tokens are represented. + +$\{t_1,t_2,\ldots ,t_{nt}\}$ . The BERT input for $S$ is then formatted as follows: + +$$ +T ^ {(S)} = \left\{\left[ \mathrm {C L S} \right], T, [ \mathrm {S E P} ] \right\} \tag {1} +$$ + +where [CLS] and [SEP] are special tokens that mark the boundaries of the sentence. + +While this format may be adequate for some NLP tasks, it can be problematic for learning good aspect representations in aspect-based sentiment classifica- tion (Tian et al., 2021). To mitigate this issue, we adopt the approach of Tian et al. (2021) and reformat the BERT input by using a sentence-aspect pair $T^{(S,A)}$ , which combines $T^{(S)}$ and $t_a$ (i.e. the aspect subsequence) along with special tokens. + +$$ +T ^ {(S, A)} = \left\{\left[ \mathrm {C L S} \right], T, [ \mathrm {S E P} ], t _ {a}, [ \mathrm {S E P} ] \right\} \tag {2} +$$ + +# 4.2 Classification and Optimization + +The input $T^{(S,A)}$ consists of two parts: $T^{(S)}$ and $t_a$ . Since $t_a$ only serves to enhance the aspect representation in $T^{(S)}$ , sequence labeling is done on $T^{(S)}$ only. During sequence labeling, we follow the common approach of predicting based on the first wordpiece representation of a word. For instance, given the word "surfboard" that consists of the wordpieces "surf" and "board" which both are learned during encoding, only the representation of "surf" is fed to a softmax classifier to predict the tag for the whole word. The cross-entropy function is minimized for each word in the training set. + +# 5 Experiments and Results + +We experiment with the following baselines: ARGCN, BERT+BiLSTM+GCN and ONG. We + +use the suffixes (S) or (S,A) to indicate whether the modified versions of these methods use a wordpiece sentence or wordpiece sentence-aspect pair as input, respectively. We used the publicly available code and optimal hyperparameter settings from the authors of $\mathrm{ARGCN}^1$ and BERT+BiLSTM+GCN.2 We have implemented ONG model variants ourselves using the suggested hyperparameter configurations from the authors.3 Following previous work (Fan et al., 2019), we use the same experimental setup and evaluate on the Laptop dataset (Lap14) and the Restaurant datasets (Res14, Res15, Res16) (Pontiki et al., 2014a, 2015, 2016). The result reported for each dataset is the average over Micro F1 scores obtained from five different runs. Each run uses a different random seed to ensure the stability of our results. + +# 5.1 F1 Performance Comparison + +The results, shown in Table 2, indicate that removing the GCN component from syntax-aware approaches does not substantially impact their performance, with average decreases in performance of 0.36, 0.74, and 0.31, respectively. However, we observed a large improvement in model performance when using wordpieces, as indicated by the models with the (S) suffix. It is possible that BERT captures enough syntax information already (Clark et al., 2019) and, therefore, using GCNs to exploit syntax trees does not substantially improve + +
ModelLap14Res14Res15Res16Avg
BERT-BiLSTM(S)80.4586.2780.8989.8084.35
-Mask Aspect80.0186.1180.4288.5983.78
+ +Table 3: F1 performance of BERT-BiLSTM(S) with and without aspect masking. + +performance on the task. This suggests that it may be beneficial to prioritize wordpieces over syntax trees to allow BERT to fully utilize rare and out-of-vocabulary words. We also discovered that using a sentence-aspect pair as input resulted in better performance than using only the sentence for the models, as indicated by the results of models with the (S,A) suffix. We believe that this may be due to the aspect information being lost or degraded during the encoding process for models with the (S) suffix. Among the methods, BERT+BiLSTM(S,A) had the highest average F1 score of 86.2. + +# 5.2 Influence of Aspect Representation + +To determine if the aspect representation is degraded during encoding, we evaluate BERT+BiLSTM(S) with and without aspect masking. The results, shown in Table 3, show that masking the aspect representation had only a minimal impact on performance, with a decrease in performance of 0.44 (Lap14), 0.16 (Res14), 0.47 (Res15), and 1.2 (Res16). These findings suggest that the aspect information has limited contribution and requires enhancement to improve performance, as demonstrated by the improved results of BERT+BiLSTM(S,A). + +# 5.3 Qualitative Analysis + +We examined the performance of BERT+BiLSTM, BERT+BiLSTM(S), and BERT+BiLSTM(S,A) on three case examples, as shown in Table 4. The results show that the BERT+BiLSTM and BERT+BiLSTM(S) models struggled to identify opinion words that were farther away from the aspect, particularly in the first and second cases where the opinion words "beautiful" and "fresh" were missed. Upon further investigation, we discovered that these opinion words were closer to the aspect's co-referential term "it". The model struggled to determine what "it" referred to due to degradation of the aspect representation, leading to the missed identification of the opinion words. However, BERT+BiLSTM(S,A) was able to recover these opinion words due to its ability to enhance the aspect representation. In the third + +case example, the use of wordpieces was beneficial as the opinion word "minimally" was not present in the training set, but its wordpiece "#lly," was associated with 15 opinion words in the training set. BERT+BiLSTM(S) and BERT+BiLSTM(S,A) were able to identify the opinion word "minimally" in the test set by leveraging the context of "#lly". + +# 6 Impact of BPE Subword Representations + +We previously examined the use of wordpiece representations derived from pretrained BERT for TOWE models. In this section, we look into using Byte Pair Encoding (BPE) (Sennrich et al., 2016) as an alternative method for subword representation, which is inspired by data compression techniques (Gage, 1994). It is worth noting that BPE representations are generally not obtained from pretrained BERT. However, since RoBERTa is pretrained using BPE, and RoBERTa is a variant of BERT, we can still explore the impact of using BPE representations in TOWE models. To do this, we replace the BERT component in our best model, BERT+BiLSTM(S,A), with RoBERTa, developing the model RoBERTa+BiLSTM(S,A). The results of RoBERTa+BiLSTM(S,A) and its variations are shown in Table 5. + +Note, while RoBERTa+BiLSTM(S,A) and RoBERTa+BiLSTM(S) use BPE subword token representations as input, RoBERTa+BiLSTM and RoBERTa+BiLSTM+GCN operate on the word-level. Our findings support the notion that GCNs have a limited impact on performance, as demonstrated by a relatively small decrease in average F1 score when comparing RoBERTa+BiLSTM+GCN to RoBERTa+BiLSTM. On the other hand, using BPE representations instead of GCN resulted in a substantial improvement in model performance of +5.27 when comparing RoBERTa+BiLSTM and RoBERTa+BiLSTM(S). The results indicate that syntax trees via GCNs may not be necessary and can be replaced by subword representations such as BPE for better performance in TOWE. Additionally, the performance of RoBERTa+BiLSTM(S) can be further improved by using BPE-based sentence-aspect pairs, as seen by the +1.75 performance gain in RoBERTa+BiLSTM(S,A). + +# 6.1 State-of-the-art Models + +Finally, we compare the performance of BERT+BiLSTM(S,A) with recent methods, + +
SentenceBERT+BiLSTMBERT+BiLSTM(S)BERT+BiLSTM(S,A)
The OS is fast and fluid, everything is organi-zed and it's just beautiful.fast, fluidfast, fluidfast, fluid, beautiful
Certainly not the best sushi in new york, however, it is always fresh, and the place is very clean, sterile.freshnot the bestnot the best, fresh
Although somewhat load, the noise was min-imally intrusiveloud, intrusiveloud, minimally in-trusiveloud, minimally in-trusive.
+ +Table 4: Case Study: Evaluating the model performance on different case examples. Aspect words are bold-typed and opinion words are italicized. + +
ModelLap14Res14Res15Res16Avg
RoBERTa-BiLSTM(S,A)82.7788.2783.8491.0686.49
RoBERTa-BiLSTM(S)81.1086.9582.2188.7084.74
RoBERTa-BiLSTM75.8781.3875.9484.7079.47
RoBERTa-BiLSTM+GCN77.5782.0977.8585.3780.72
+ +Table 5: F1 Performance of RoBERTa models to investigate the use of BPE subword representations. + +
ModelLap14Res14Res15Res16Avg
IOG71.3580.0273.2581.6976.58
LOTN72.0282.2173.2983.6277.79
SDRN+BERT*73.6983.1076.3885.4079.64
ONG75.7782.3378.8186.0180.73
ARGCN76.3685.4278.2486.6981.68
BERT+BiLSTM+GCN78.8285.7480.5487.3583.11
QD-OWSE80.3587.2380.7188.1484.11
TSMSA82.1886.3781.6489.2084.85
BERT-BiLSTM(S,A)82.5988.6082.3791.2586.20
+ +Table 6: Performance of TOWE methods. Results for the method marked “*” are from (Jiang et al., 2021). + +including IOG (Fan et al., 2019), LOTN (Wu et al., 2020), SDRN+BERT (Chen et al., 2020), BERT+BiLSTM+GCN (Mensah et al., 2021), QD-OWSE (Gao et al., 2021), TSMSA (Feng et al., 2021). The results of this comparison are shown in Table 6. Among these methods, the recent proposed methods QD-OWSE and TSMSA, which both use BERT as a basis for their approach, achieved competitive results with ours. QD-OWSE uses a generated question-answer pair as BERT input, while TSMSA uses multi-head attention to identify opinion words. These methods go on to demonstrate that BERT can capture sufficient syntax information for this task, even without the use of syntax trees. However, BERT+BiLSTM(S,A) achieved the best results, with F1 scores 82.59 (Lap14), 88.6 (Res14), 82.37 (Res15) and 91.25 (Res16), setting a new SOTA for the task. + +# 7 Conclusion + +We demonstrated that replacing GCNs with BERT wordpieces while enhancing the aspect representation achieves SOTA results in syntax-aware TOWE approaches. The aspect enhancement method serves as a "prompt" for the model. We intend to explore prompt-based learning (Brown et al., 2020) to further improve the aspect representation. + +# 8 Limitations + +Currently, our approach does not effectively leverage syntax tree information via GCNs, a commonly used method for incorporating syntax trees in this task. Further research is required to determine the most effective way to integrate syntax tree information into TOWE models. + +# Acknowledgements + +This work was supported by the Leverhulme Trust under Grant Number: RPG#2020#148. + +# References + +Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901. + +Shaowei Chen, Jie Liu, Yu Wang, Wenzheng Zhang, and Ziming Chi. 2020. Synchronous double-channel recurrent network for aspect-opinion pair extraction. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6515-6524. + +Kevin Clark, Urvashi Khandelwal, Omer Levy, and Christopher D. Manning. 2019. What does BERT look at? an analysis of BERT's attention. In Proceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 276-286, Florence, Italy. Association for Computational Linguistics. + +Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina N. Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186. +Zhifang Fan, Zhen Wu, Xinyu Dai, Shujian Huang, and Jiajun Chen. 2019. Target-oriented opinion words extraction with target-fused neural sequence labeling. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2509-2518. +Yuhao Feng, Yanghui Rao, Yuyao Tang, Ninghua Wang, and He Liu. 2021. Target-specified sequence labeling with multi-head self-attention for target-oriented opinion words extraction. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1805–1815. +Philip Gage. 1994. A new algorithm for data compression. C Users Journal, 12(2):23-38. +Lei Gao, Yulong Wang, Tongcun Liu, Jingyu Wang, Lei Zhang, and Jianxin Liao. 2021. Question-driven span labeling model for aspect-opinion pair extraction. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 12875-12883. +Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory. *Neural Computation*, 9(8):1735-1780. +Junfeng Jiang, An Wang, and Akiko Aizawa. 2021. Attention-based relational graph convolutional network for target-oriented opinion words extraction. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 1986-1997. +Hyun Duk Kim, Kavita Ganesan, Parikshit Sondhi, and ChengXiang Zhai. 2011. Comprehensive review of opinion summarization. +Thomas N. Kipf and Max Welling. 2017. Semi-supervised classification with graph convolutional networks. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net. +Samuel Mensah, Kai Sun, and Nikolaos Aletras. 2021. An empirical study on leveraging position embeddings for target-oriented opinion words extraction. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 9174-9179, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics. +Maria Pontiki, Dimitrios Galanis, Haris Papageorgiou, Ion Androutsopoulos, Suresh Manandhar, Mohammad Al-Smadi, Mahmoud Al-Ayyoub, Yanyan Zhao, + +Bing Qin, Orphée De Clercq, et al. 2016. Semeval-2016 task 5: Aspect based sentiment analysis. In International workshop on semantic evaluation, pages 19-30. +Maria Pontiki, Dimitrios Galanis, Harris Papageorgiou, Suresh Manandhar, and Ion Androutsopoulos. 2015. Semeval-2015 task 12: Aspect based sentiment analysis. In Proceedings of the 9th international workshop on semantic evaluation (SemEval 2015), pages 486-495. +Maria Pontiki, Dimitrios Galanis, John Pavlopoulos, Harris Papageorgiou, Ion Androutsopoulos, and Suresh Manandhar. 2014a. Semeval-2014 task 4: Aspect based sentiment analysis. In Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), page 27-35. +Maria Pontiki, Dimitris Galanis, John Pavlopoulos, Harris Papageorgiou, Ion Androutsopoulos, and Suresh Manandhar. 2014b. Semeval-2014 task 4: Aspect based sentiment analysis. In Proceedings of the 8th International Workshop on Semantic Evaluation, SemEval@COLING 2014, Dublin, Ireland, August 23-24, 2014, pages 27-35. The Association for Computer Linguistics. +Lance A. Ramshaw and Mitch Marcus. 1995. Text chunking using transformation-based learning. In Third Workshop on Very Large Corpora, VLC@ACL 1995, Cambridge, Massachusetts, USA, June 30, 1995. +Michael Schlichtkrull, Thomas N. Kipf, Peter Bloem, Rianne van den Berg, Ivan Titov, and Max Welling. 2018. Modeling relational data with graph convolutional networks. In *The Semantic Web*, pages 593–607, Cham. Springer International Publishing. +Mike Schuster and Kaisuke Nakajima. 2012. Japanese and korean voice search. In 2012 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 5149-5152. IEEE. +Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Neural machine translation of rare words with subword units. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, ACL 2016, August 7-12, 2016, Berlin, Germany, Volume 1: Long Papers. The Association for Computer Linguistics. +Yikang Shen, Shawn Tan, Alessandro Sordoni, and Aaron C. Courville. 2018. Ordered neurons: Integrating tree structures into recurrent neural networks. In International Conference on Learning Representations. +Kai Sun, Richong Zhang, Mensah Samuel, Aletras Nikolaos, Yongyi Mao, and Xudong Liu. 2023. Self-training through classifier disagreement for cross-domain opinion target extraction. In Proceedings of the ACM Web Conference 2023, pages 1594-1603. + +Duyu Tang, Bing Qin, and Ting Liu. 2016. Aspect level sentiment classification with deep memory network. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 214-224. +Yuanhe Tian, Guimin Chen, and Yan Song. 2021. Aspect-based sentiment analysis with type-aware graph convolutional networks and layer ensemble. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2910-2922. +Amir Pouran Ben Veyseh, Nasim Nouri, Franck Der-noncourt, Dejing Dou, and Thien Huu Nguyen. 2020. Introducing syntactic structures into target opinion word extraction with deep learning. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online, + +November 16-20, 2020, pages 8947-8956. Association for Computational Linguistics. +Zhen Wu, Fei Zhao, Xin-Yu Dai, Shujian Huang, and Jiajun Chen. 2020. Latent opinions transfer network for target-oriented opinion words extraction. In The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, The Thirty-Second Innovative Applications of Artificial Intelligence Conference, IAAI 2020, The Tenth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2020, New York, NY, USA, February 7-12, 2020, pages 9298-9305. AAAI Press. +Daojian Zeng, Kang Liu, Siwei Lai, Guangyou Zhou, and Jun Zhao. 2014. Relation classification via convolutional deep neural network. In Proceedings of COLING 2014, the 25th international conference on computational linguistics: technical papers, pages 2335-2344. + +A For every submission: + +A1. Did you describe the limitations of your work? 7 +A2. Did you discuss any potential risks of your work? There are no risks +A3. Do the abstract and introduction summarize the paper's main claims? Abstract and Section 1 +A4. Have you used AI writing assistants when working on this paper? Left blank. + +B Did you use or create scientific artifacts? + +5 + +B1. Did you cite the creators of artifacts you used? No response. +B2. Did you discuss the license or terms for use and / or distribution of any artifacts? No response. +B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? No response. +B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? No response. +B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? No response. +B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. No response. + +C Did you run computational experiments? + +5 + +C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? No response. + +C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? No response. +C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? No response. +C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? No response. + +# D Did you use human annotators (e.g., crowdworkers) or research with human participants? + +Left blank. + +D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response. +D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response. +D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response. +D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response. +D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response. \ No newline at end of file diff --git a/2023/Trading Syntax Trees for Wordpieces_ Target-oriented Opinion Words Extraction with Wordpieces and Aspect Enhancement/images.zip b/2023/Trading Syntax Trees for Wordpieces_ Target-oriented Opinion Words Extraction with Wordpieces and Aspect Enhancement/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d5aa52db084874e0bf8700f9e00b6f42f1658fe7 --- /dev/null +++ b/2023/Trading Syntax Trees for Wordpieces_ Target-oriented Opinion Words Extraction with Wordpieces and Aspect Enhancement/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8430b7962f3a3682fe7c6effd573024400d1e5137d17cbbe04520a31569eed4d +size 240903 diff --git a/2023/Trading Syntax Trees for Wordpieces_ Target-oriented Opinion Words Extraction with Wordpieces and Aspect Enhancement/layout.json b/2023/Trading Syntax Trees for Wordpieces_ Target-oriented Opinion Words Extraction with Wordpieces and Aspect Enhancement/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1e390b2bc33c30b0f8d85f4383ff35d2d464aa59 --- /dev/null +++ b/2023/Trading Syntax Trees for Wordpieces_ Target-oriented Opinion Words Extraction with Wordpieces and Aspect Enhancement/layout.json @@ -0,0 +1,5370 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 76, + 74, + 519, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 74, + 519, + 122 + ], + "spans": [ + { + "bbox": [ + 76, + 74, + 519, + 122 + ], + "type": "text", + "content": "Trading Syntax Trees for Wordpieces: Target-oriented Opinion Words Extraction with Wordpieces and Aspect Enhancement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 74, + 135, + 223, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 135, + 223, + 190 + ], + "spans": [ + { + "bbox": [ + 74, + 135, + 223, + 190 + ], + "type": "text", + "content": "Samuel Mensah \nComputer Science Department \nUniversity of Sheffield, UK \ns.mensah@sheffield.ac.uk" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 232, + 135, + 361, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 135, + 361, + 190 + ], + "spans": [ + { + "bbox": [ + 232, + 135, + 361, + 190 + ], + "type": "text", + "content": "Kai Sun \nBDBC and SKLSDE \nBeihang University, China \nsunkai@buaa.edu.cn" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 371, + 135, + 521, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 135, + 521, + 190 + ], + "spans": [ + { + "bbox": [ + 371, + 135, + 521, + 190 + ], + "type": "text", + "content": "Nikolaos Aletras \nComputer Science Department \nUniversity of Sheffield, UK \nn.aletras@sheffield.ac.uk" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 84, + 239, + 274, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 239, + 274, + 491 + ], + "spans": [ + { + "bbox": [ + 84, + 239, + 274, + 491 + ], + "type": "text", + "content": "State-of-the-art target-oriented opinion word extraction (TOWE) models typically use BERT-based text encoders that operate on the word level, along with graph convolutional networks (GCNs) that incorporate syntactic information extracted from syntax trees. These methods achieve limited gains with GCNs and have difficulty using BERT wordpieces. Meanwhile, BERT wordpieces are known to be effective at representing rare words or words with insufficient context information. To address this issue, this work trades syntax trees for BERT wordpieces by entirely removing the GCN component from the methods' architectures. To enhance TOWE performance, we tackle the issue of aspect representation loss during encoding. Instead of solely utilizing a sentence as the input, we use a sentence-aspect pair. Our relatively simple approach achieves state-of-the-art results on benchmark datasets and should serve as a strong baseline for further research." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 505, + 154, + 518 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 505, + 154, + 518 + ], + "spans": [ + { + "bbox": [ + 68, + 505, + 154, + 518 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 528, + 291, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 528, + 291, + 716 + ], + "spans": [ + { + "bbox": [ + 67, + 528, + 291, + 716 + ], + "type": "text", + "content": "Target-oriented opinion word extraction (TOWE; Fan et al. (2019)) is a subtask in aspect-based sentiment analysis (ABSA; Pontiki et al. (2014b)), which aims to identify words that express an opinion about a specific target (or aspect) in a sentence. For instance, in the sentence \"Such an awesome surfboard.\", a TOWE model should identify \"awesome\" as the opinion word for the given aspect surfboard. TOWE provides explicit aspect-opinion pairs which can be used to improve results in downstream tasks such as opinion summarization (Kim et al., 2011) and information extraction (Pontiki et al., 2014b; Tang et al., 2016; Sun et al., 2023)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 719, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 719, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 719, + 291, + 773 + ], + "type": "text", + "content": "Currently, many TOWE methods (Veyseh et al., 2020; Chen et al., 2020; Jiang et al., 2021; Feng et al., 2021; Mensah et al., 2021) use pretrained BERT (Devlin et al., 2018), to encode the input" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 305, + 211, + 524, + 317 + ], + "blocks": [ + { + "bbox": [ + 305, + 211, + 524, + 317 + ], + "lines": [ + { + "bbox": [ + 305, + 211, + 524, + 317 + ], + "spans": [ + { + "bbox": [ + 305, + 211, + 524, + 317 + ], + "type": "table", + "html": "
1. Sentence:Wordpieces:Such an awesome surfboard‘such’, ‘an’, ‘awesome’, ‘surf’,‘##board’
2. Sentence:Wordpieces:A great snowboard which holds edges well when riding on snow.A, ‘great’, ‘snow’, ‘#’board’, ‘which’, ‘holds’, ‘edges’, ‘well’, ‘when’, ‘riding’, ‘on’, ‘snow’.
", + "image_path": "bfeec2e2e654b44679e1596245e4b1fb970a59881a66ba60524f71e393f61d3c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 325, + 526, + 398 + ], + "lines": [ + { + "bbox": [ + 302, + 325, + 526, + 398 + ], + "spans": [ + { + "bbox": [ + 302, + 325, + 526, + 398 + ], + "type": "text", + "content": "Table 1: Sentences demonstrating contextual understanding through shared wordpieces. The table shows each sentence and its corresponding BERT wordpiece sequence. Aspect words are bold-typed and opinion words are italicized. The shared wordpiece '##board' helps in decoding the meaning of \"surfboard\"." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 302, + 421, + 526, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 421, + 526, + 596 + ], + "spans": [ + { + "bbox": [ + 302, + 421, + 526, + 596 + ], + "type": "text", + "content": "sentence. BERT has the ability to effectively capture context, which can improve TOWE performance. However, many of these methods are rather complex, as they often incorporate syntax tree information using a graph convolutional network (GCN) (Kipf and Welling, 2017). For instance, Veyseh et al. (2020) uses an ordered-neuron LSTM (Shen et al., 2018) encoder with a GCN while Jiang et al. (2021) applies an attention-based relational GCN on the syntax tree. Mensah et al. (2021) applies a BiLSTM (Hochreiter and Schmidhuber, 1997) on BERT embeddings and incorporate syntax information via a GCN." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 597, + 526, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 597, + 526, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 597, + 526, + 773 + ], + "type": "text", + "content": "While incorporating syntax information through GCNs has been shown to provide some performance gains in TOWE, these are usually limited (Mensah et al., 2021). Moreover, modeling subword tokens with a GCN can be challenging because the syntax tree consists of whole words rather than subword tokens like wordpieces (Schuster and Nakajima, 2012; Devlin et al., 2018). Models based on subword tokens strike a good balance between character- and word-based encoders. They are able to effectively learn representations of rare words or words with insufficient context information. Consider the example in Table 1. The context" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "text", + "content": "999" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "spans": [ + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "type": "text", + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 221, + 806, + 371, + 817 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 806, + 371, + 817 + ], + "spans": [ + { + "bbox": [ + 221, + 806, + 371, + 817 + ], + "type": "text", + "content": "Volume 2: Short Papers, pages 999-1007" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "spans": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "text", + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 66, + 71, + 290, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 71, + 290, + 191 + ], + "spans": [ + { + "bbox": [ + 66, + 71, + 290, + 191 + ], + "type": "text", + "content": "information for \"surfboard\" is limited, making it difficult to understand its meaning without additional context. However, both aspects share the wordpiece \"#board\", which allows the meaning of \"surfboard\" to be partially understood by using information from the context of \"snowboard\". In this case, \"riding\" is related to both aspects through the shared wordpiece, enabling the representation of \"surfboard\" to be improved." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 194, + 291, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 194, + 291, + 477 + ], + "spans": [ + { + "bbox": [ + 69, + 194, + 291, + 477 + ], + "type": "text", + "content": "In this paper, we propose a substantial simplification for syntax-aware TOWE models (Veyseh et al., 2020; Jiang et al., 2021; Mensah et al., 2021) by replacing the syntax tree with subword information while maintaining good prediction performance. This is accomplished by removing the GCN from these architectures and using BERT wordpieces instead. Additionally, we address the issue of aspect representation degradation during encoding. This degradation negatively affects TOWE performance by reducing the availability of semantic information about the aspect for determining the opinion words to extract. To solve this problem, we propose using a sentence-aspect pair as input rather than just a sentence, similar to the approach used by Tian et al. (2021) for aspect-based sentiment classification. Through extensive experimentation, we found that our simple approach achieves state-of-the-art (SOTA) results by outperforming the method proposed by Mensah et al. (2021) without the need of a GCN component." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 488, + 189, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 488, + 189, + 500 + ], + "spans": [ + { + "bbox": [ + 67, + 488, + 189, + 500 + ], + "type": "text", + "content": "2 Task Formalization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 510, + 291, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 510, + 291, + 673 + ], + "spans": [ + { + "bbox": [ + 67, + 510, + 291, + 673 + ], + "type": "text", + "content": "The TOWE task aims to identify an opinion word in a sentence " + }, + { + "bbox": [ + 67, + 510, + 291, + 673 + ], + "type": "inline_equation", + "content": "S = \\{w_{1},\\ldots ,w_{n_{s}}\\}" + }, + { + "bbox": [ + 67, + 510, + 291, + 673 + ], + "type": "text", + "content": " with respect to an aspect " + }, + { + "bbox": [ + 67, + 510, + 291, + 673 + ], + "type": "inline_equation", + "content": "w_{a}\\in S" + }, + { + "bbox": [ + 67, + 510, + 291, + 673 + ], + "type": "text", + "content": ". The sentence is typically tokenized into a sequence of tokens at different levels of granularity (e.g. subwords or whole words), " + }, + { + "bbox": [ + 67, + 510, + 291, + 673 + ], + "type": "inline_equation", + "content": "T = \\{t_1,\\dots ,t_{n_t}\\}" + }, + { + "bbox": [ + 67, + 510, + 291, + 673 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 67, + 510, + 291, + 673 + ], + "type": "inline_equation", + "content": "t_a\\in T" + }, + { + "bbox": [ + 67, + 510, + 291, + 673 + ], + "type": "text", + "content": " denoting a subsequence of the aspect " + }, + { + "bbox": [ + 67, + 510, + 291, + 673 + ], + "type": "inline_equation", + "content": "w_{a}" + }, + { + "bbox": [ + 67, + 510, + 291, + 673 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 510, + 291, + 673 + ], + "type": "inline_equation", + "content": "n_s\\leq n_t" + }, + { + "bbox": [ + 67, + 510, + 291, + 673 + ], + "type": "text", + "content": ". The goal is to assign one of three tags (I, O, or B) to each token using the IOB format (Ramshaw and Marcus, 1995), which indicates whether the word is at the Inside, Outside or Beginning of the opinion word relative to the aspect." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 683, + 276, + 698 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 683, + 276, + 698 + ], + "spans": [ + { + "bbox": [ + 67, + 683, + 276, + 698 + ], + "type": "text", + "content": "3 Syntax-aware Approaches to TOWE" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 705, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 705, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 705, + 291, + 772 + ], + "type": "text", + "content": "Typically, syntax-aware approaches to TOWE (Veyseh et al., 2020; Jiang et al., 2021; Mensah et al., 2021) employ a text encoder that utilizes pretrained BERT (Devlin et al., 2018) and position embeddings (Zeng et al., 2014) (or category em" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 71, + 525, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 525, + 138 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 525, + 138 + ], + "type": "text", + "content": "beddings (Jiang et al., 2021)) to learn whole word representations that are aware of the aspect's location in text. These approaches also include a GCN that operates on a syntax tree in order to incorporate syntactic information into the model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 148, + 526, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 148, + 526, + 283 + ], + "spans": [ + { + "bbox": [ + 302, + 148, + 526, + 283 + ], + "type": "text", + "content": "Ordered-Neuron LSTM GCN (ONG): Veyseh et al. (2020) combine an ordered neuron LSTM (ON-LSTM; Shen et al. (2018)) and a GCN for TOWE. The ON-LSTM layer is an LSTM variant that considers the order of elements in the input sequence (including BERT and position embeddings) when modeling dependencies between them. The GCN encodes syntactic structural information into the representations obtained by the ON-LSTM layer." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 293, + 525, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 293, + 525, + 347 + ], + "spans": [ + { + "bbox": [ + 302, + 293, + 525, + 347 + ], + "type": "text", + "content": "BERT+BiLSTM+GCN: Mensah et al. (2021) replaces the ON-LSTM of the ONG model with a BiLSTM to better capture short-term dependencies between aspect and opinion words." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 357, + 526, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 357, + 526, + 465 + ], + "spans": [ + { + "bbox": [ + 302, + 357, + 526, + 465 + ], + "type": "text", + "content": "Attention-based Relational GCN (ARGCN): Jiang et al. (2021) combine contextualized embedding obtained using BERT with a category embedding (i.e., IOB tag embedding) to incorporate aspect information. They subsequently use a relational GCN (Schlichtkrull et al., 2018) and BiLSTM to respectively incorporate syntactic and sequential information for TOWE classification." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 476, + 515, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 476, + 515, + 491 + ], + "spans": [ + { + "bbox": [ + 302, + 476, + 515, + 491 + ], + "type": "text", + "content": "4 Trading Syntax Trees for Wordpieces" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 500, + 525, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 500, + 525, + 716 + ], + "spans": [ + { + "bbox": [ + 302, + 500, + 525, + 716 + ], + "type": "text", + "content": "Mensah et al. (2021) have recently demonstrated that the use of a GCN to incorporate syntax tree information has little impact in TOWE model performance. Meanwhile, the GCN presents challenges when using subword tokens, as previously mentioned. Therefore, we propose a simplified version of the TOWE model that omits the GCN component from syntax-aware approaches and instead uses subword tokens as the input to the BERT component. In this work, we use BERT's Wordpieces (Devlin et al., 2018) as the subword representation because they are highly informative, having been derived from the BERT pretraining process. However, methods such as Byte-Pair Encoding (BPE) (Sennrich et al., 2016) can also be used, as we will see later in the experiments." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 728, + 445, + 740 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 728, + 445, + 740 + ], + "spans": [ + { + "bbox": [ + 302, + 728, + 445, + 740 + ], + "type": "text", + "content": "4.1 Formatting BERT Input" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "type": "text", + "content": "Given sentence " + }, + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "type": "text", + "content": ", the BERT wordpiece tokenizer segments " + }, + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "type": "text", + "content": " into a sequence of wordpieces " + }, + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "type": "inline_equation", + "content": "T =" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 780, + 309, + 791 + ], + "type": "text", + "content": "1000" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 124, + 68, + 470, + 242 + ], + "blocks": [ + { + "bbox": [ + 124, + 68, + 470, + 242 + ], + "lines": [ + { + "bbox": [ + 124, + 68, + 470, + 242 + ], + "spans": [ + { + "bbox": [ + 124, + 68, + 470, + 242 + ], + "type": "table", + "html": "
ModelsGranularityLap14Res14Res15Res16Avg
ONGword75.7782.3378.8186.0180.73
ONG w/o GCNword74.1784.1078.3384.8780.37
ONG(S) w/o GCNwordpiece79.7986.6380.7288.3083.86
ONG(S,A) w/o GCNwordpiece81.7088.7082.5591.1886.03
ARGCNword76.3685.4278.2486.6981.68
ARGCN w/o R-GCNword76.3884.3678.4184.6180.94
ARGCN(S) w/o R-GCNwordpiece80.0885.9281.3689.7284.27
ARGCN(S,A) w/o R-GCNwordpiece81.3788.1882.4990.8285.72
BERT+BiLSTM+GCNword78.8285.7480.5487.3583.11
BERT+BiLSTMword78.2585.6080.4186.9482.80
BERT+BiLSTM(S)wordpiece80.4586.2780.8989.8084.35
BERT+BiLSTM(S,A)wordpiece82.5988.6082.3791.2586.20
", + "image_path": "5b04d04e5154db6b28bcd673063f30635301c7817b7ac6932b42d561a4f13312.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 249, + 525, + 275 + ], + "lines": [ + { + "bbox": [ + 67, + 249, + 525, + 275 + ], + "spans": [ + { + "bbox": [ + 67, + 249, + 525, + 275 + ], + "type": "text", + "content": "Table 2: F1 performance of syntax-aware methods and their variants. \"Avg\" refers to the average F1 score calculated across all of the datasets. \"Granularity\" highlights the level of granularity at which input tokens are represented." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 295, + 289, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 295, + 289, + 321 + ], + "spans": [ + { + "bbox": [ + 67, + 295, + 289, + 321 + ], + "type": "inline_equation", + "content": "\\{t_1,t_2,\\ldots ,t_{nt}\\}" + }, + { + "bbox": [ + 67, + 295, + 289, + 321 + ], + "type": "text", + "content": " . The BERT input for " + }, + { + "bbox": [ + 67, + 295, + 289, + 321 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 67, + 295, + 289, + 321 + ], + "type": "text", + "content": " is then formatted as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 329, + 290, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 329, + 290, + 346 + ], + "spans": [ + { + "bbox": [ + 116, + 329, + 290, + 346 + ], + "type": "interline_equation", + "content": "T ^ {(S)} = \\left\\{\\left[ \\mathrm {C L S} \\right], T, [ \\mathrm {S E P} ] \\right\\} \\tag {1}", + "image_path": "24f5a9619c3a5d90790796be0f892867d50058b85dcce29af28b494c890ec323.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 354, + 289, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 354, + 289, + 380 + ], + "spans": [ + { + "bbox": [ + 67, + 354, + 289, + 380 + ], + "type": "text", + "content": "where [CLS] and [SEP] are special tokens that mark the boundaries of the sentence." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 382, + 290, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 382, + 290, + 502 + ], + "spans": [ + { + "bbox": [ + 67, + 382, + 290, + 502 + ], + "type": "text", + "content": "While this format may be adequate for some NLP tasks, it can be problematic for learning good aspect representations in aspect-based sentiment classifica- tion (Tian et al., 2021). To mitigate this issue, we adopt the approach of Tian et al. (2021) and reformat the BERT input by using a sentence-aspect pair " + }, + { + "bbox": [ + 67, + 382, + 290, + 502 + ], + "type": "inline_equation", + "content": "T^{(S,A)}" + }, + { + "bbox": [ + 67, + 382, + 290, + 502 + ], + "type": "text", + "content": ", which combines " + }, + { + "bbox": [ + 67, + 382, + 290, + 502 + ], + "type": "inline_equation", + "content": "T^{(S)}" + }, + { + "bbox": [ + 67, + 382, + 290, + 502 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 382, + 290, + 502 + ], + "type": "inline_equation", + "content": "t_a" + }, + { + "bbox": [ + 67, + 382, + 290, + 502 + ], + "type": "text", + "content": " (i.e. the aspect subsequence) along with special tokens." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 82, + 511, + 290, + 527 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 511, + 290, + 527 + ], + "spans": [ + { + "bbox": [ + 82, + 511, + 290, + 527 + ], + "type": "interline_equation", + "content": "T ^ {(S, A)} = \\left\\{\\left[ \\mathrm {C L S} \\right], T, [ \\mathrm {S E P} ], t _ {a}, [ \\mathrm {S E P} ] \\right\\} \\tag {2}", + "image_path": "58c0619db0c01a7787a8818821802051cd8b8060ff955401f67702c1dfad6458.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 536, + 244, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 536, + 244, + 549 + ], + "spans": [ + { + "bbox": [ + 67, + 536, + 244, + 549 + ], + "type": "text", + "content": "4.2 Classification and Optimization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 552, + 291, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 552, + 291, + 715 + ], + "spans": [ + { + "bbox": [ + 67, + 552, + 291, + 715 + ], + "type": "text", + "content": "The input " + }, + { + "bbox": [ + 67, + 552, + 291, + 715 + ], + "type": "inline_equation", + "content": "T^{(S,A)}" + }, + { + "bbox": [ + 67, + 552, + 291, + 715 + ], + "type": "text", + "content": " consists of two parts: " + }, + { + "bbox": [ + 67, + 552, + 291, + 715 + ], + "type": "inline_equation", + "content": "T^{(S)}" + }, + { + "bbox": [ + 67, + 552, + 291, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 552, + 291, + 715 + ], + "type": "inline_equation", + "content": "t_a" + }, + { + "bbox": [ + 67, + 552, + 291, + 715 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 67, + 552, + 291, + 715 + ], + "type": "inline_equation", + "content": "t_a" + }, + { + "bbox": [ + 67, + 552, + 291, + 715 + ], + "type": "text", + "content": " only serves to enhance the aspect representation in " + }, + { + "bbox": [ + 67, + 552, + 291, + 715 + ], + "type": "inline_equation", + "content": "T^{(S)}" + }, + { + "bbox": [ + 67, + 552, + 291, + 715 + ], + "type": "text", + "content": ", sequence labeling is done on " + }, + { + "bbox": [ + 67, + 552, + 291, + 715 + ], + "type": "inline_equation", + "content": "T^{(S)}" + }, + { + "bbox": [ + 67, + 552, + 291, + 715 + ], + "type": "text", + "content": " only. During sequence labeling, we follow the common approach of predicting based on the first wordpiece representation of a word. For instance, given the word \"surfboard\" that consists of the wordpieces \"surf\" and \"board\" which both are learned during encoding, only the representation of \"surf\" is fed to a softmax classifier to predict the tag for the whole word. The cross-entropy function is minimized for each word in the training set." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 725, + 218, + 739 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 725, + 218, + 739 + ], + "spans": [ + { + "bbox": [ + 67, + 725, + 218, + 739 + ], + "type": "text", + "content": "5 Experiments and Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "type": "text", + "content": "We experiment with the following baselines: ARGCN, BERT+BiLSTM+GCN and ONG. We" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 295, + 526, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 295, + 526, + 524 + ], + "spans": [ + { + "bbox": [ + 302, + 295, + 526, + 524 + ], + "type": "text", + "content": "use the suffixes (S) or (S,A) to indicate whether the modified versions of these methods use a wordpiece sentence or wordpiece sentence-aspect pair as input, respectively. We used the publicly available code and optimal hyperparameter settings from the authors of " + }, + { + "bbox": [ + 302, + 295, + 526, + 524 + ], + "type": "inline_equation", + "content": "\\mathrm{ARGCN}^1" + }, + { + "bbox": [ + 302, + 295, + 526, + 524 + ], + "type": "text", + "content": " and BERT+BiLSTM+GCN.2 We have implemented ONG model variants ourselves using the suggested hyperparameter configurations from the authors.3 Following previous work (Fan et al., 2019), we use the same experimental setup and evaluate on the Laptop dataset (Lap14) and the Restaurant datasets (Res14, Res15, Res16) (Pontiki et al., 2014a, 2015, 2016). The result reported for each dataset is the average over Micro F1 scores obtained from five different runs. Each run uses a different random seed to ensure the stability of our results." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 535, + 467, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 535, + 467, + 548 + ], + "spans": [ + { + "bbox": [ + 302, + 535, + 467, + 548 + ], + "type": "text", + "content": "5.1 F1 Performance Comparison" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 553, + 525, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 553, + 525, + 702 + ], + "spans": [ + { + "bbox": [ + 302, + 553, + 525, + 702 + ], + "type": "text", + "content": "The results, shown in Table 2, indicate that removing the GCN component from syntax-aware approaches does not substantially impact their performance, with average decreases in performance of 0.36, 0.74, and 0.31, respectively. However, we observed a large improvement in model performance when using wordpieces, as indicated by the models with the (S) suffix. It is possible that BERT captures enough syntax information already (Clark et al., 2019) and, therefore, using GCNs to exploit syntax trees does not substantially improve" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 709, + 523, + 730 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 709, + 523, + 730 + ], + "spans": [ + { + "bbox": [ + 302, + 709, + 523, + 730 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 302, + 709, + 523, + 730 + ], + "type": "text", + "content": "https://github.com/samensah/encoders_towe_emnlp2021" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 303, + 730, + 478, + 751 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 730, + 478, + 751 + ], + "spans": [ + { + "bbox": [ + 303, + 730, + 478, + 751 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 303, + 730, + 478, + 751 + ], + "type": "text", + "content": "https://github.com/wcwowwww/towe-eacl" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 751, + 473, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 751, + 473, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 751, + 473, + 772 + ], + "type": "text", + "content": "3https://github.com/samensah/Towe-TradeSyntax4WP" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 287, + 781, + 308, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 308, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 308, + 791 + ], + "type": "text", + "content": "1001" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 68, + 287, + 103 + ], + "blocks": [ + { + "bbox": [ + 70, + 68, + 287, + 103 + ], + "lines": [ + { + "bbox": [ + 70, + 68, + 287, + 103 + ], + "spans": [ + { + "bbox": [ + 70, + 68, + 287, + 103 + ], + "type": "table", + "html": "
ModelLap14Res14Res15Res16Avg
BERT-BiLSTM(S)80.4586.2780.8989.8084.35
-Mask Aspect80.0186.1180.4288.5983.78
", + "image_path": "83b3f0f294de2f829e0bfaa894e97ee83c219b4780fd29b02450bfb86374e58d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 111, + 289, + 136 + ], + "lines": [ + { + "bbox": [ + 67, + 111, + 289, + 136 + ], + "spans": [ + { + "bbox": [ + 67, + 111, + 289, + 136 + ], + "type": "text", + "content": "Table 3: F1 performance of BERT-BiLSTM(S) with and without aspect masking." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 158, + 291, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 158, + 291, + 321 + ], + "spans": [ + { + "bbox": [ + 67, + 158, + 291, + 321 + ], + "type": "text", + "content": "performance on the task. This suggests that it may be beneficial to prioritize wordpieces over syntax trees to allow BERT to fully utilize rare and out-of-vocabulary words. We also discovered that using a sentence-aspect pair as input resulted in better performance than using only the sentence for the models, as indicated by the results of models with the (S,A) suffix. We believe that this may be due to the aspect information being lost or degraded during the encoding process for models with the (S) suffix. Among the methods, BERT+BiLSTM(S,A) had the highest average F1 score of 86.2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 333, + 258, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 333, + 258, + 345 + ], + "spans": [ + { + "bbox": [ + 67, + 333, + 258, + 345 + ], + "type": "text", + "content": "5.2 Influence of Aspect Representation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 351, + 291, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 351, + 291, + 513 + ], + "spans": [ + { + "bbox": [ + 67, + 351, + 291, + 513 + ], + "type": "text", + "content": "To determine if the aspect representation is degraded during encoding, we evaluate BERT+BiLSTM(S) with and without aspect masking. The results, shown in Table 3, show that masking the aspect representation had only a minimal impact on performance, with a decrease in performance of 0.44 (Lap14), 0.16 (Res14), 0.47 (Res15), and 1.2 (Res16). These findings suggest that the aspect information has limited contribution and requires enhancement to improve performance, as demonstrated by the improved results of BERT+BiLSTM(S,A)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 525, + 190, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 525, + 190, + 538 + ], + "spans": [ + { + "bbox": [ + 67, + 525, + 190, + 538 + ], + "type": "text", + "content": "5.3 Qualitative Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 543, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 543, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 543, + 291, + 772 + ], + "type": "text", + "content": "We examined the performance of BERT+BiLSTM, BERT+BiLSTM(S), and BERT+BiLSTM(S,A) on three case examples, as shown in Table 4. The results show that the BERT+BiLSTM and BERT+BiLSTM(S) models struggled to identify opinion words that were farther away from the aspect, particularly in the first and second cases where the opinion words \"beautiful\" and \"fresh\" were missed. Upon further investigation, we discovered that these opinion words were closer to the aspect's co-referential term \"it\". The model struggled to determine what \"it\" referred to due to degradation of the aspect representation, leading to the missed identification of the opinion words. However, BERT+BiLSTM(S,A) was able to recover these opinion words due to its ability to enhance the aspect representation. In the third" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 71, + 526, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 166 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 166 + ], + "type": "text", + "content": "case example, the use of wordpieces was beneficial as the opinion word \"minimally\" was not present in the training set, but its wordpiece \"#lly,\" was associated with 15 opinion words in the training set. BERT+BiLSTM(S) and BERT+BiLSTM(S,A) were able to identify the opinion word \"minimally\" in the test set by leveraging the context of \"#lly\"." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 178, + 450, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 178, + 450, + 205 + ], + "spans": [ + { + "bbox": [ + 302, + 178, + 450, + 205 + ], + "type": "text", + "content": "6 Impact of BPE Subword Representations" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 215, + 526, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 215, + 526, + 444 + ], + "spans": [ + { + "bbox": [ + 302, + 215, + 526, + 444 + ], + "type": "text", + "content": "We previously examined the use of wordpiece representations derived from pretrained BERT for TOWE models. In this section, we look into using Byte Pair Encoding (BPE) (Sennrich et al., 2016) as an alternative method for subword representation, which is inspired by data compression techniques (Gage, 1994). It is worth noting that BPE representations are generally not obtained from pretrained BERT. However, since RoBERTa is pretrained using BPE, and RoBERTa is a variant of BERT, we can still explore the impact of using BPE representations in TOWE models. To do this, we replace the BERT component in our best model, BERT+BiLSTM(S,A), with RoBERTa, developing the model RoBERTa+BiLSTM(S,A). The results of RoBERTa+BiLSTM(S,A) and its variations are shown in Table 5." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 445, + 526, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 445, + 526, + 716 + ], + "spans": [ + { + "bbox": [ + 302, + 445, + 526, + 716 + ], + "type": "text", + "content": "Note, while RoBERTa+BiLSTM(S,A) and RoBERTa+BiLSTM(S) use BPE subword token representations as input, RoBERTa+BiLSTM and RoBERTa+BiLSTM+GCN operate on the word-level. Our findings support the notion that GCNs have a limited impact on performance, as demonstrated by a relatively small decrease in average F1 score when comparing RoBERTa+BiLSTM+GCN to RoBERTa+BiLSTM. On the other hand, using BPE representations instead of GCN resulted in a substantial improvement in model performance of +5.27 when comparing RoBERTa+BiLSTM and RoBERTa+BiLSTM(S). The results indicate that syntax trees via GCNs may not be necessary and can be replaced by subword representations such as BPE for better performance in TOWE. Additionally, the performance of RoBERTa+BiLSTM(S) can be further improved by using BPE-based sentence-aspect pairs, as seen by the +1.75 performance gain in RoBERTa+BiLSTM(S,A)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 728, + 441, + 740 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 728, + 441, + 740 + ], + "spans": [ + { + "bbox": [ + 302, + 728, + 441, + 740 + ], + "type": "text", + "content": "6.1 State-of-the-art Models" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "type": "text", + "content": "Finally, we compare the performance of BERT+BiLSTM(S,A) with recent methods," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 780, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 780, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 780, + 310, + 791 + ], + "type": "text", + "content": "1002" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 68, + 68, + 490, + 164 + ], + "blocks": [ + { + "bbox": [ + 68, + 68, + 490, + 164 + ], + "lines": [ + { + "bbox": [ + 68, + 68, + 490, + 164 + ], + "spans": [ + { + "bbox": [ + 68, + 68, + 490, + 164 + ], + "type": "table", + "html": "
SentenceBERT+BiLSTMBERT+BiLSTM(S)BERT+BiLSTM(S,A)
The OS is fast and fluid, everything is organi-zed and it's just beautiful.fast, fluidfast, fluidfast, fluid, beautiful
Certainly not the best sushi in new york, however, it is always fresh, and the place is very clean, sterile.freshnot the bestnot the best, fresh
Although somewhat load, the noise was min-imally intrusiveloud, intrusiveloud, minimally in-trusiveloud, minimally in-trusive.
", + "image_path": "1e088fbb4d21a80d601576729bb4a13f3759f0a0968743f2ec52db45e4575041.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 69, + 216, + 296, + 271 + ], + "blocks": [ + { + "bbox": [ + 67, + 173, + 525, + 196 + ], + "lines": [ + { + "bbox": [ + 67, + 173, + 525, + 196 + ], + "spans": [ + { + "bbox": [ + 67, + 173, + 525, + 196 + ], + "type": "text", + "content": "Table 4: Case Study: Evaluating the model performance on different case examples. Aspect words are bold-typed and opinion words are italicized." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 216, + 296, + 271 + ], + "lines": [ + { + "bbox": [ + 69, + 216, + 296, + 271 + ], + "spans": [ + { + "bbox": [ + 69, + 216, + 296, + 271 + ], + "type": "table", + "html": "
ModelLap14Res14Res15Res16Avg
RoBERTa-BiLSTM(S,A)82.7788.2783.8491.0686.49
RoBERTa-BiLSTM(S)81.1086.9582.2188.7084.74
RoBERTa-BiLSTM75.8781.3875.9484.7079.47
RoBERTa-BiLSTM+GCN77.5782.0977.8585.3780.72
", + "image_path": "3974679c18548a37b390ecada54a8ffd85c15a7dcc00aef74e4559b9b2a5599b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 69, + 342, + 298, + 449 + ], + "blocks": [ + { + "bbox": [ + 67, + 280, + 291, + 304 + ], + "lines": [ + { + "bbox": [ + 67, + 280, + 291, + 304 + ], + "spans": [ + { + "bbox": [ + 67, + 280, + 291, + 304 + ], + "type": "text", + "content": "Table 5: F1 Performance of RoBERTa models to investigate the use of BPE subword representations." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 342, + 298, + 449 + ], + "lines": [ + { + "bbox": [ + 69, + 342, + 298, + 449 + ], + "spans": [ + { + "bbox": [ + 69, + 342, + 298, + 449 + ], + "type": "table", + "html": "
ModelLap14Res14Res15Res16Avg
IOG71.3580.0273.2581.6976.58
LOTN72.0282.2173.2983.6277.79
SDRN+BERT*73.6983.1076.3885.4079.64
ONG75.7782.3378.8186.0180.73
ARGCN76.3685.4278.2486.6981.68
BERT+BiLSTM+GCN78.8285.7480.5487.3583.11
QD-OWSE80.3587.2380.7188.1484.11
TSMSA82.1886.3781.6489.2084.85
BERT-BiLSTM(S,A)82.5988.6082.3791.2586.20
", + "image_path": "d2bdddeb4c3540ae4018cb7885865bf5f65c2cca0c956338d4e9141fc01b9842.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 457, + 290, + 481 + ], + "lines": [ + { + "bbox": [ + 67, + 457, + 290, + 481 + ], + "spans": [ + { + "bbox": [ + 67, + 457, + 290, + 481 + ], + "type": "text", + "content": "Table 6: Performance of TOWE methods. Results for the method marked “*” are from (Jiang et al., 2021)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 529, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 529, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 529, + 291, + 772 + ], + "type": "text", + "content": "including IOG (Fan et al., 2019), LOTN (Wu et al., 2020), SDRN+BERT (Chen et al., 2020), BERT+BiLSTM+GCN (Mensah et al., 2021), QD-OWSE (Gao et al., 2021), TSMSA (Feng et al., 2021). The results of this comparison are shown in Table 6. Among these methods, the recent proposed methods QD-OWSE and TSMSA, which both use BERT as a basis for their approach, achieved competitive results with ours. QD-OWSE uses a generated question-answer pair as BERT input, while TSMSA uses multi-head attention to identify opinion words. These methods go on to demonstrate that BERT can capture sufficient syntax information for this task, even without the use of syntax trees. However, BERT+BiLSTM(S,A) achieved the best results, with F1 scores 82.59 (Lap14), 88.6 (Res14), 82.37 (Res15) and 91.25 (Res16), setting a new SOTA for the task." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 303, + 218, + 381, + 230 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 218, + 381, + 230 + ], + "spans": [ + { + "bbox": [ + 303, + 218, + 381, + 230 + ], + "type": "text", + "content": "7 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 239, + 526, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 239, + 526, + 334 + ], + "spans": [ + { + "bbox": [ + 302, + 239, + 526, + 334 + ], + "type": "text", + "content": "We demonstrated that replacing GCNs with BERT wordpieces while enhancing the aspect representation achieves SOTA results in syntax-aware TOWE approaches. The aspect enhancement method serves as a \"prompt\" for the model. We intend to explore prompt-based learning (Brown et al., 2020) to further improve the aspect representation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 303, + 344, + 384, + 356 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 344, + 384, + 356 + ], + "spans": [ + { + "bbox": [ + 303, + 344, + 384, + 356 + ], + "type": "text", + "content": "8 Limitations" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 364, + 525, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 364, + 525, + 445 + ], + "spans": [ + { + "bbox": [ + 302, + 364, + 525, + 445 + ], + "type": "text", + "content": "Currently, our approach does not effectively leverage syntax tree information via GCNs, a commonly used method for incorporating syntax trees in this task. Further research is required to determine the most effective way to integrate syntax tree information into TOWE models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 303, + 456, + 405, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 456, + 405, + 470 + ], + "spans": [ + { + "bbox": [ + 303, + 456, + 405, + 470 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 477, + 524, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 477, + 524, + 503 + ], + "spans": [ + { + "bbox": [ + 302, + 477, + 524, + 503 + ], + "type": "text", + "content": "This work was supported by the Leverhulme Trust under Grant Number: RPG#2020#148." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 527, + 362, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 527, + 362, + 539 + ], + "spans": [ + { + "bbox": [ + 304, + 527, + 362, + 539 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 545, + 524, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 524, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 524, + 612 + ], + "type": "text", + "content": "Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 303, + 620, + 525, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 620, + 525, + 687 + ], + "spans": [ + { + "bbox": [ + 303, + 620, + 525, + 687 + ], + "type": "text", + "content": "Shaowei Chen, Jie Liu, Yu Wang, Wenzheng Zhang, and Ziming Chi. 2020. Synchronous double-channel recurrent network for aspect-opinion pair extraction. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6515-6524." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 303, + 694, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 694, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 303, + 694, + 526, + 772 + ], + "type": "text", + "content": "Kevin Clark, Urvashi Khandelwal, Omer Levy, and Christopher D. Manning. 2019. What does BERT look at? an analysis of BERT's attention. In Proceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 276-286, Florence, Italy. Association for Computational Linguistics." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1003" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 773 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 291, + 159 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 291, + 159 + ], + "type": "text", + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina N. Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 168, + 291, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 168, + 291, + 245 + ], + "spans": [ + { + "bbox": [ + 69, + 168, + 291, + 245 + ], + "type": "text", + "content": "Zhifang Fan, Zhen Wu, Xinyu Dai, Shujian Huang, and Jiajun Chen. 2019. Target-oriented opinion words extraction with target-fused neural sequence labeling. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2509-2518." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 252, + 290, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 252, + 290, + 330 + ], + "spans": [ + { + "bbox": [ + 69, + 252, + 290, + 330 + ], + "type": "text", + "content": "Yuhao Feng, Yanghui Rao, Yuyao Tang, Ninghua Wang, and He Liu. 2021. Target-specified sequence labeling with multi-head self-attention for target-oriented opinion words extraction. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1805–1815." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 337, + 290, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 337, + 290, + 359 + ], + "spans": [ + { + "bbox": [ + 69, + 337, + 290, + 359 + ], + "type": "text", + "content": "Philip Gage. 1994. A new algorithm for data compression. C Users Journal, 12(2):23-38." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 366, + 290, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 366, + 290, + 422 + ], + "spans": [ + { + "bbox": [ + 69, + 366, + 290, + 422 + ], + "type": "text", + "content": "Lei Gao, Yulong Wang, Tongcun Liu, Jingyu Wang, Lei Zhang, and Jianxin Liao. 2021. Question-driven span labeling model for aspect-opinion pair extraction. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 12875-12883." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 429, + 290, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 429, + 290, + 462 + ], + "spans": [ + { + "bbox": [ + 69, + 429, + 290, + 462 + ], + "type": "text", + "content": "Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory. *Neural Computation*, 9(8):1735-1780." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 470, + 290, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 470, + 290, + 536 + ], + "spans": [ + { + "bbox": [ + 69, + 470, + 290, + 536 + ], + "type": "text", + "content": "Junfeng Jiang, An Wang, and Akiko Aizawa. 2021. Attention-based relational graph convolutional network for target-oriented opinion words extraction. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 1986-1997." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 544, + 290, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 544, + 290, + 576 + ], + "spans": [ + { + "bbox": [ + 69, + 544, + 290, + 576 + ], + "type": "text", + "content": "Hyun Duk Kim, Kavita Ganesan, Parikshit Sondhi, and ChengXiang Zhai. 2011. Comprehensive review of opinion summarization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 584, + 290, + 650 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 584, + 290, + 650 + ], + "spans": [ + { + "bbox": [ + 69, + 584, + 290, + 650 + ], + "type": "text", + "content": "Thomas N. Kipf and Max Welling. 2017. Semi-supervised classification with graph convolutional networks. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 658, + 290, + 735 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 658, + 290, + 735 + ], + "spans": [ + { + "bbox": [ + 69, + 658, + 290, + 735 + ], + "type": "text", + "content": "Samuel Mensah, Kai Sun, and Nikolaos Aletras. 2021. An empirical study on leveraging position embeddings for target-oriented opinion words extraction. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 9174-9179, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 740, + 290, + 773 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 740, + 290, + 773 + ], + "spans": [ + { + "bbox": [ + 69, + 740, + 290, + 773 + ], + "type": "text", + "content": "Maria Pontiki, Dimitrios Galanis, Haris Papageorgiou, Ion Androutsopoulos, Suresh Manandhar, Mohammad Al-Smadi, Mahmoud Al-Ayyoub, Yanyan Zhao," + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 525, + 756 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 314, + 72, + 525, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 525, + 116 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 525, + 116 + ], + "type": "text", + "content": "Bing Qin, Orphée De Clercq, et al. 2016. Semeval-2016 task 5: Aspect based sentiment analysis. In International workshop on semantic evaluation, pages 19-30." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 125, + 525, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 125, + 525, + 190 + ], + "spans": [ + { + "bbox": [ + 304, + 125, + 525, + 190 + ], + "type": "text", + "content": "Maria Pontiki, Dimitrios Galanis, Harris Papageorgiou, Suresh Manandhar, and Ion Androutsopoulos. 2015. Semeval-2015 task 12: Aspect based sentiment analysis. In Proceedings of the 9th international workshop on semantic evaluation (SemEval 2015), pages 486-495." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 200, + 525, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 200, + 525, + 266 + ], + "spans": [ + { + "bbox": [ + 304, + 200, + 525, + 266 + ], + "type": "text", + "content": "Maria Pontiki, Dimitrios Galanis, John Pavlopoulos, Harris Papageorgiou, Ion Androutsopoulos, and Suresh Manandhar. 2014a. Semeval-2014 task 4: Aspect based sentiment analysis. In Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), page 27-35." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 275, + 525, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 275, + 525, + 362 + ], + "spans": [ + { + "bbox": [ + 304, + 275, + 525, + 362 + ], + "type": "text", + "content": "Maria Pontiki, Dimitris Galanis, John Pavlopoulos, Harris Papageorgiou, Ion Androutsopoulos, and Suresh Manandhar. 2014b. Semeval-2014 task 4: Aspect based sentiment analysis. In Proceedings of the 8th International Workshop on Semantic Evaluation, SemEval@COLING 2014, Dublin, Ireland, August 23-24, 2014, pages 27-35. The Association for Computer Linguistics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 371, + 525, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 371, + 525, + 425 + ], + "spans": [ + { + "bbox": [ + 304, + 371, + 525, + 425 + ], + "type": "text", + "content": "Lance A. Ramshaw and Mitch Marcus. 1995. Text chunking using transformation-based learning. In Third Workshop on Very Large Corpora, VLC@ACL 1995, Cambridge, Massachusetts, USA, June 30, 1995." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 434, + 525, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 525, + 490 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 525, + 490 + ], + "type": "text", + "content": "Michael Schlichtkrull, Thomas N. Kipf, Peter Bloem, Rianne van den Berg, Ivan Titov, and Max Welling. 2018. Modeling relational data with graph convolutional networks. In *The Semantic Web*, pages 593–607, Cham. Springer International Publishing." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 498, + 525, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 525, + 543 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 525, + 543 + ], + "type": "text", + "content": "Mike Schuster and Kaisuke Nakajima. 2012. Japanese and korean voice search. In 2012 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 5149-5152. IEEE." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 551, + 525, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 551, + 525, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 551, + 525, + 628 + ], + "type": "text", + "content": "Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Neural machine translation of rare words with subword units. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, ACL 2016, August 7-12, 2016, Berlin, Germany, Volume 1: Long Papers. The Association for Computer Linguistics." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 636, + 525, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 636, + 525, + 691 + ], + "spans": [ + { + "bbox": [ + 304, + 636, + 525, + 691 + ], + "type": "text", + "content": "Yikang Shen, Shawn Tan, Alessandro Sordoni, and Aaron C. Courville. 2018. Ordered neurons: Integrating tree structures into recurrent neural networks. In International Conference on Learning Representations." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 700, + 525, + 756 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 700, + 525, + 756 + ], + "spans": [ + { + "bbox": [ + 304, + 700, + 525, + 756 + ], + "type": "text", + "content": "Kai Sun, Richong Zhang, Mensah Samuel, Aletras Nikolaos, Yongyi Mao, and Xudong Liu. 2023. Self-training through classifier disagreement for cross-domain opinion target extraction. In Proceedings of the ACM Web Conference 2023, pages 1594-1603." + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1004" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 288 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 291, + 126 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 291, + 126 + ], + "type": "text", + "content": "Duyu Tang, Bing Qin, and Ting Liu. 2016. Aspect level sentiment classification with deep memory network. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 214-224." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 136, + 291, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 136, + 291, + 213 + ], + "spans": [ + { + "bbox": [ + 69, + 136, + 291, + 213 + ], + "type": "text", + "content": "Yuanhe Tian, Guimin Chen, and Yan Song. 2021. Aspect-based sentiment analysis with type-aware graph convolutional networks and layer ensemble. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2910-2922." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 222, + 291, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 222, + 291, + 288 + ], + "spans": [ + { + "bbox": [ + 69, + 222, + 291, + 288 + ], + "type": "text", + "content": "Amir Pouran Ben Veyseh, Nasim Nouri, Franck Der-noncourt, Dejing Dou, and Thien Huu Nguyen. 2020. Introducing syntactic structures into target opinion word extraction with deep learning. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online," + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 526, + 287 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 314, + 72, + 525, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 525, + 95 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 525, + 95 + ], + "type": "text", + "content": "November 16-20, 2020, pages 8947-8956. Association for Computational Linguistics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 103, + 526, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 103, + 526, + 212 + ], + "spans": [ + { + "bbox": [ + 304, + 103, + 526, + 212 + ], + "type": "text", + "content": "Zhen Wu, Fei Zhao, Xin-Yu Dai, Shujian Huang, and Jiajun Chen. 2020. Latent opinions transfer network for target-oriented opinion words extraction. In The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, The Thirty-Second Innovative Applications of Artificial Intelligence Conference, IAAI 2020, The Tenth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2020, New York, NY, USA, February 7-12, 2020, pages 9298-9305. AAAI Press." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 222, + 526, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 222, + 526, + 287 + ], + "spans": [ + { + "bbox": [ + 304, + 222, + 526, + 287 + ], + "type": "text", + "content": "Daojian Zeng, Kang Liu, Siwei Lai, Guangyou Zhou, and Jun Zhao. 2014. Relation classification via convolutional deep neural network. In Proceedings of COLING 2014, the 25th international conference on computational linguistics: technical papers, pages 2335-2344." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1005" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "spans": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "type": "text", + "content": "A For every submission:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 106, + 414, + 243 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "spans": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "type": "text", + "content": "A1. Did you describe the limitations of your work? 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 143, + 329, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 143, + 329, + 169 + ], + "spans": [ + { + "bbox": [ + 77, + 143, + 329, + 169 + ], + "type": "text", + "content": "A2. Did you discuss any potential risks of your work? There are no risks" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 178, + 414, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 178, + 414, + 206 + ], + "spans": [ + { + "bbox": [ + 77, + 178, + 414, + 206 + ], + "type": "text", + "content": "A3. Do the abstract and introduction summarize the paper's main claims? Abstract and Section 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 77, + 215, + 398, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 215, + 398, + 243 + ], + "spans": [ + { + "bbox": [ + 77, + 215, + 398, + 243 + ], + "type": "text", + "content": "A4. Have you used AI writing assistants when working on this paper? Left blank." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 68, + 252, + 291, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 252, + 291, + 266 + ], + "spans": [ + { + "bbox": [ + 68, + 252, + 291, + 266 + ], + "type": "text", + "content": "B Did you use or create scientific artifacts?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 80, + 271, + 87, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 271, + 87, + 280 + ], + "spans": [ + { + "bbox": [ + 80, + 271, + 87, + 280 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 76, + 292, + 524, + 634 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 77, + 292, + 315, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 292, + 315, + 319 + ], + "spans": [ + { + "bbox": [ + 77, + 292, + 315, + 319 + ], + "type": "text", + "content": "B1. Did you cite the creators of artifacts you used? No response." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 76, + 328, + 463, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 328, + 463, + 356 + ], + "spans": [ + { + "bbox": [ + 76, + 328, + 463, + 356 + ], + "type": "text", + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? No response." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 76, + 364, + 524, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 364, + 524, + 432 + ], + "spans": [ + { + "bbox": [ + 76, + 364, + 524, + 432 + ], + "type": "text", + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? No response." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 76, + 441, + 524, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 441, + 524, + 494 + ], + "spans": [ + { + "bbox": [ + 76, + 441, + 524, + 494 + ], + "type": "text", + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? No response." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 76, + 504, + 524, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 504, + 524, + 544 + ], + "spans": [ + { + "bbox": [ + 76, + 504, + 524, + 544 + ], + "type": "text", + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? No response." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 76, + 554, + 524, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 554, + 524, + 634 + ], + "spans": [ + { + "bbox": [ + 76, + 554, + 524, + 634 + ], + "type": "text", + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. No response." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "spans": [ + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "type": "text", + "content": "C Did you run computational experiments?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 80, + 662, + 87, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 662, + 87, + 672 + ], + "spans": [ + { + "bbox": [ + 80, + 662, + 87, + 672 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 76, + 684, + 524, + 724 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 684, + 524, + 724 + ], + "spans": [ + { + "bbox": [ + 76, + 684, + 524, + 724 + ], + "type": "text", + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? No response." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "spans": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "text", + "content": "ACL 2023 Responsible NLP Checklist" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "spans": [ + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "type": "text", + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1006" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 76, + 71, + 524, + 238 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 76, + 71, + 523, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 71, + 523, + 111 + ], + "spans": [ + { + "bbox": [ + 76, + 71, + 523, + 111 + ], + "type": "text", + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? No response." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 76, + 121, + 524, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 121, + 524, + 174 + ], + "spans": [ + { + "bbox": [ + 76, + 121, + 524, + 174 + ], + "type": "text", + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? No response." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 76, + 184, + 524, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 184, + 524, + 238 + ], + "spans": [ + { + "bbox": [ + 76, + 184, + 524, + 238 + ], + "type": "text", + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? No response." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 67, + 247, + 522, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 247, + 522, + 261 + ], + "spans": [ + { + "bbox": [ + 67, + 247, + 522, + 261 + ], + "type": "text", + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 79, + 264, + 127, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 264, + 127, + 277 + ], + "spans": [ + { + "bbox": [ + 79, + 264, + 127, + 277 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 76, + 286, + 524, + 539 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 76, + 286, + 524, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 286, + 524, + 326 + ], + "spans": [ + { + "bbox": [ + 76, + 286, + 524, + 326 + ], + "type": "text", + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 76, + 336, + 524, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 336, + 524, + 390 + ], + "spans": [ + { + "bbox": [ + 76, + 336, + 524, + 390 + ], + "type": "text", + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 76, + 399, + 524, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 399, + 524, + 454 + ], + "spans": [ + { + "bbox": [ + 76, + 399, + 524, + 454 + ], + "type": "text", + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 76, + 462, + 519, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 462, + 519, + 489 + ], + "spans": [ + { + "bbox": [ + 76, + 462, + 519, + 489 + ], + "type": "text", + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 76, + 498, + 524, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 498, + 524, + 539 + ], + "spans": [ + { + "bbox": [ + 76, + 498, + 524, + 539 + ], + "type": "text", + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1007" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Transformed Protoform Reconstruction/244a271d-f768-4ade-bdfe-e0da000e90a0_content_list.json b/2023/Transformed Protoform Reconstruction/244a271d-f768-4ade-bdfe-e0da000e90a0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4008133b28c85304a5119b729ea9143340333f78 --- /dev/null +++ b/2023/Transformed Protoform Reconstruction/244a271d-f768-4ade-bdfe-e0da000e90a0_content_list.json @@ -0,0 +1,1721 @@ +[ + { + "type": "text", + "text": "Transformed Protoform Reconstruction", + "text_level": 1, + "bbox": [ + 290, + 90, + 707, + 109 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Young Min Kim* and Kalvin Chang* and Chenxuan Cui and David Mortensen \nLanguage Technologies Institute, Carnegie Mellon University \n{youngmik, kalvinc, cxcui, dmortens}@cs.cmu.edu", + "bbox": [ + 159, + 141, + 843, + 193 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 252, + 341, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Protoform reconstruction is the task of inferring how morphemes or words sounded in ancestral languages of a set of daughter languages. Meloni et al. (2021) achieved the state-of-the-art on Latin protoform reconstruction with an RNN-based encoder-decoder with attention model. We update their model with the state-of-the-art seq2seq model—the Transformer. Our model outperforms their model on a suite of different metrics on two different datasets: Meloni et al.'s Romance data of $8,000+$ cognates (spanning 5 languages) and a Chinese dataset (Hóu, 2004) of $800+$ cognates (spanning 39 varieties). We also probe our model for potential phylogenetic signal contained in the model. Our code is publicly available1.", + "bbox": [ + 141, + 275, + 460, + 517 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 114, + 527, + 260, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Languages change over time and sometimes diverge into multiple daughter languages. The common ancestor of a set of genetically related languages is their proto-language. While there are proto-languages such as Latin that are attested, they are the exception2. Reconstructed words and morphemes in proto-languages are called protoforms. The task of reconstructing unattested protolanguages is called protoform reconstruction.", + "bbox": [ + 112, + 552, + 490, + 697 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Historical linguists reconstruct proto-languages by identifying systematic sound changes that can be inferred from correspondences between attested daughter languages (see Table 1). They compare the sounds between a set of cognates, or words with a common ancestor, to develop hypotheses about the types and chronologies of sound changes.", + "bbox": [ + 112, + 697, + 489, + 826 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/6fdf9d698a1840ce766ff34c1ee4d7151dfd84876cddb589635e6829b2f3e63d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
‘tooth’‘two’‘ten’
Englishtoothtwotent
Dutchtandtweetient
GermanZahnzweizehnz
PWG*tanþ*twai-*tehun*t
", + "bbox": [ + 527, + 249, + 867, + 335 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Table 1: Sound correspondences in West Germanic Languages and Proto-West-Germanic (PWG).", + "bbox": [ + 507, + 344, + 885, + 375 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This task is inherently data-constrained, especially for under-documented languages. Such data scarcity makes it a particularly difficult task for contemporary neural network architectures such as the Transformer (Vaswani et al., 2017), which are data hungry.", + "bbox": [ + 507, + 400, + 884, + 495 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The contributions of this paper are as follows:", + "bbox": [ + 527, + 498, + 870, + 514 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Application of the Transformer architecture to the protoform reconstruction task, achieving state of the art performance, contrary to expectation.", + "- Expansion of prior digital versions of Hóu (2004)'s Chinese dataset to include a total of 804 cognate sets across 39 modern varieties and Middle Chinese." + ], + "bbox": [ + 531, + 527, + 884, + 668 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 507, + 683, + 665, + 697 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Applying machine learning to protoform reconstruction is not new. Bouchard-Côté et al. (2013) learn an unsupervised protoform reconstruction model for the large Oceanic language family using Monte Carlo Expectation Maximization (Dempster et al., 1977; Bouchard-Côté et al., 2008), supervising the model with a gold phylogeny and using a probabilistic, generative model of sound change. He et al. (2022) modernize an earlier version of Bouchard-Côté et al. (2013)'s model with RNNs for a 4 language subset of Romance, but they rely on a bigram language model of Latin, making their model technically not unsupervised.", + "bbox": [ + 505, + 709, + 884, + 919 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution", + "bbox": [ + 139, + 831, + 268, + 845 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "", + "bbox": [ + 137, + 846, + 433, + 857 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "2 In fact, the proto-language from which Romance languages like Spanish and Italian are descended is not identical to Classical Latin but is, rather, a closely related and sparsely attested language sometimes called Proto-Romance or Vulgar Latin.", + "bbox": [ + 115, + 857, + 487, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24", + "bbox": [ + 489, + 928, + 510, + 939 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics", + "bbox": [ + 226, + 945, + 769, + 957 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Volume 2: Short Papers, pages 24-38", + "bbox": [ + 384, + 958, + 613, + 971 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "July 9-14, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 295, + 972, + 700, + 985 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "List et al. (2022) apply an SVM classifier to supervised reconstruction by treating sound correspondences as training examples. Note that there were no word boundaries in the input matrix; that is, all sound correspondences across the training set are flattened into one matrix. Furthermore, each language has an independent phonemic inventory. To learn contextual information, the authors experiment with adding features encoding the position of phonemes, among others.", + "bbox": [ + 112, + 84, + 489, + 243 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Ciobanu and Dinu (2018) learn a conditional random field (Lafferty et al., 2001) using n-gram features for supervised reconstruction and ensemble 5 daughter-to-protoform models. They use a dataset of 3,218 complete cognate sets spanning Latin (the proto-language) and 5 Romance languages: Romanian, French, Italian, Spanish, Portuguese.", + "bbox": [ + 112, + 244, + 489, + 372 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Meloni et al. (2021) employ a GRU-based seq2seq approach (Cho et al., 2014) to Latin protoform reconstruction and achieve state-of-the-art character edit distances. They extend Dinu and Ciobanu (2014)'s Romance data using data from Wiktionary—for a total of 8,799 cognate sets across 5 Romance languages plus Latin—in both orthographic and phonetic (IPA) representations. In their model, all entries comprising the cognate set are concatenated together in a fixed order to form a training example. Chang et al. (2022) applied Meloni et al. (2021)'s architecture to the reconstruction of Middle Chinese on a dataset of $5000+$ cognate sets spanning 8 languages they compiled from Wiktionary. $^3$", + "bbox": [ + 115, + 374, + 489, + 615 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fourrier (2022) compares statistical machine translation, RNN, and Transformer architectures for protoform reconstruction, but they evaluate their results using BLEU scores (Papineni et al., 2002) instead of edit distance. They find that their Transformer model did not outperform the RNN models on protoform reconstruction. In addition, their multilingual NMT (neural machine translation) model predicts many languages instead of one target language and is trained on bilingual pairs for protoform reconstruction (e.g. Italian-Latin and Spanish-Latin), unlike comparative reconstruction. In contrast, we encode the entire cognate set consisting of multiple daughter languages (5 for the Romance dataset; 39 for Chinese) and predict the corresponding protoform.", + "bbox": [ + 112, + 615, + 489, + 872 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3 Datasets", + "text_level": 1, + "bbox": [ + 509, + 84, + 618, + 98 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We train and test our model on Romance and Sinitic (Chinese) language datasets. For Romance languages, we use Meloni et al. (2021)'s dataset which consists of 8,799 cognate sets of Romanian, French, Italian, Spanish, Portuguese words and the corresponding Latin form (approximately, a protoform). There are two versions of this dataset: phonetic and orthographic. The phonetic dataset (Rom-phon) represents words with IPA symbols whereas the orthographic dataset (Rom-orth) represents words in the orthographic form of each language. We preserved all diacritics, except for vowel length. This dataset is an extension of Dinu and Ciobanu (2014)'s original dataset of 3,218 cognate sets, which is not publicly available. Refer to Table 2 for more information.", + "bbox": [ + 507, + 123, + 884, + 380 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1 Expanding digital versions of Hóu (2004)", + "text_level": 1, + "bbox": [ + 507, + 412, + 877, + 429 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For Sinitic languages, we created a dataset of Middle Chinese and its modern daughter languages. Middle Chinese is an unattested language, and we thus have to rely on Baxter and Sagart (2014)'s reconstructions of forms corresponding to 4,967 Chinese characters. We scraped Wiktionary to obtain Hóu (2004)'s phonetic representations of their modern reflexes. The resulting dataset contains 804 cognate sets of 39 modern Sinitic languages and the corresponding reconstructed Middle Chinese word. List (2021)'s version previously had 894 cognate sets across 15 varieties.", + "bbox": [ + 507, + 445, + 884, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "4 Model", + "text_level": 1, + "bbox": [ + 507, + 670, + 601, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose a Transformer-based encoder-decoder architecture (Vaswani et al., 2017) because such models have produced state-of-the-art results on many sequence processing tasks. Transformers are by reputation data hungry, though, which poses a challenge to our problem setting, where the number of available training examples is often very small.", + "bbox": [ + 507, + 711, + 884, + 838 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "3The original dataset contains 21,000 cognate sets, but only $5000+$ had at least 3 daughter entries and were used as input to the model.", + "bbox": [ + 112, + 879, + 487, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "4https://en.wiktionary.org/wiki/Module: zh/data/dial-pron/documentation originally had 1,023 characters, but only 804 had reconstructions from Baxter and Sagart (2014).", + "bbox": [ + 507, + 869, + 882, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25", + "bbox": [ + 489, + 928, + 510, + 940 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f7b8fb97a67b65f9e747cee35450c06455cec12b94f2336c6353635c555db3c3.jpg", + "image_caption": [ + "Figure 1: Diagram of our encoder-decoder architecture. Additive positional encoding and language embedding are applied to each daughter sequence before all daughter sequences are concatenated into a single sequence." + ], + "image_footnote": [], + "bbox": [ + 152, + 84, + 448, + 265 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We modify the standard encoder-decoder architecture to accommodate the structure of our datasets, where multiple daughter sequences correspond to a single protoform sequence. Like Meloni et al. (2021), the daughter sequences are concatenated into a single sequence before being fed into the encoder. Because we only care about the relative position between tokens within each daughter sequence but not across daughter sequences, positional encoding is applied to each individual daughter sequence before concatenation. Along with positional encoding, an additive language embedding is applied to the token embeddings to differentiate between input tokens of different daughter languages.", + "bbox": [ + 112, + 354, + 489, + 596 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 114, + 611, + 258, + 627 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "5.1 Baselines", + "text_level": 1, + "bbox": [ + 114, + 639, + 233, + 653 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We compare our Transformer model to a variety of baselines. For Meloni et al. (2021), we use Chang et al. (2022)'s PyTorch re-implementation and reran a Bayesian hyperparameter search using WandB (Biewald, 2020) to ensure a more fair comparison (since our model is tuned with WandB as well). We also include the random daughter (randomly designate a daughter form as the protoform and assume no sound change) and the majority constituent baselines (predict the most common phoneme in each syllable constituent) from Chang et al. (2022). For the SVM and CoRPaR classifiers (List et al., 2022), we experiment with different contextual features, such as Pos (position), Str (prosodic structure), and Ini (whether or not the phoneme appears word-initially or wordfinally).", + "bbox": [ + 112, + 661, + 489, + 919 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We publish results on Meloni et al. (2021)'s full set of 8,799 cognates but cannot redistribute this set due to Dinu and Ciobanu (2014)'s restrictions. For reproducibility, we include results on Meloni et al. (2021)'s public subset of 5,419 cognates in the Appendix (Table 7), both of which include vowel length. Observe that these results are worse than those obtained on the full set, suggesting that the RNN and Transformer are dependent on a wealth of training data.", + "bbox": [ + 507, + 84, + 884, + 244 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "5.2 Preprocessing", + "text_level": 1, + "bbox": [ + 507, + 256, + 665, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In all our datasets, we merge diacritics to their base segments to form a multi-character token. For instance, the sequence $\\left[\\mathrm{t},\\mathrm{h}\\right]$ is concatenated to $\\left[\\mathrm{th}\\right]$ . This ensures that phonemes are treated as one token. For Chinese, tone contours (a sequence of tones) are treated as one token. When multiple pronunciation variants are listed for a single Chinese character, we arbitrarily pick the first one.", + "bbox": [ + 507, + 278, + 884, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "6 Results and Discussion", + "text_level": 1, + "bbox": [ + 507, + 420, + 741, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "6.1 Evaluation criteria", + "text_level": 1, + "bbox": [ + 507, + 445, + 705, + 458 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We evaluate the predicted protoforms using edit distance (Levenshtein et al., 1966), normalized edit distance (edit distance normalized by the length of the target) and accuracy (the percentage of protoforms that are reconstructed without any mistakes). Like Chang et al. (2022), we also use feature error rate calculated using articulatory feature vectors from PanPhon (Mortensen et al., 2016) because it reflects the phonetic similarity between the prediction and the gold protoform. For datasets with phonetic transcriptions (Romance-phonetic and Chinese), we use phoneme edit distance and normalized phoneme edit distance. As List (2019) suggests, we use B-Cubed F Scores (Amigo et al., 2009) to capture the structural similarity between the gold and predicted protoforms (0: structurally dissimilar, 1: similar). With the exception of character and phoneme edit distance, the metrics enable fair comparison across different language families, which will differ in the average word length.", + "bbox": [ + 507, + 467, + 884, + 804 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "6.2 Results", + "text_level": 1, + "bbox": [ + 507, + 816, + 613, + 831 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 3 shows that our model consistently has the best performance on all datasets with regards to most metrics. The results were averaged across 5 runs. Out of all datasets, our model performs best on the Rom-orth dataset, where we achieve a $7.0\\%$", + "bbox": [ + 507, + 838, + 882, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26", + "bbox": [ + 489, + 928, + 510, + 940 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/91fe4f67e27b10ef3b25b979e13b9b20905a6133441ae6d279a76cb8e5935bc9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Language FamilySource# varietiesCognate setsProto-language
Rom-phonDinu and Ciobanu (2014), Meloni et al. (2021)58,799Latin
Rom-orthDinu and Ciobanu (2014), Meloni et al. (2021)58,799Latin
Sinitic (Chinese)Hóu (2004)39804Middle Chinese
", + "bbox": [ + 114, + 80, + 885, + 186 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 2: Statistics on both datasets used in our experiments. # varieties refers to the number of daughter varieties.", + "bbox": [ + 112, + 193, + 880, + 208 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c592223a9b2834446e15eea3269624c5d3a859092d8336de5af115ff8da6ccec.jpg", + "image_caption": [ + "Figure 2: A gold phylogeny of Romance (left) compared with those derived by probing the RNN model (middle) and the Transformer model (right) on Rom-phon." + ], + "image_footnote": [], + "bbox": [ + 119, + 219, + 371, + 338 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5bc091c6f6ab7c6395d4be007dba35a798343972ad3762c35667220fa1c358bc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 389, + 219, + 640, + 337 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/eb5a89ca965ce80dc9ef2e39ce40c63b323765447d15d389f4faf61b53fe0f46.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 219, + 878, + 338 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "decrease in phoneme edit distance and a 1.43p.p improvement in accuracy relative to the RNN baseline. We observe the most dramatic performance difference with the RNN baseline on the Sinitic dataset: a $10.48\\%$ decrease in phoneme edit distance and a 5.47p.p increase in accuracy. For reproducibility, results on the publicly available portion of the Rom-phon and Rom-orth datasets are provided in Table 7 in the Appendix.", + "bbox": [ + 112, + 401, + 489, + 548 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "6.3 Analysis", + "text_level": 1, + "bbox": [ + 112, + 558, + 226, + 574 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We observe that the BCFS is relatively high for the Romance non-neural baselines compared to those of the Chinese ones. This suggests that the sound changes in the Romance datasets are more regular than that of Chinese, which corroborates List et al. (2014)'s results that more than half of the Chinese characters in their dataset could not be explained by a tree model.", + "bbox": [ + 112, + 580, + 487, + 708 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We examine the errors made by the Transformer model on the Rom-phon dataset. Substitutions constitute around $61\\%$ of the errors made by the Transformer; deletions, $21\\%$ , and insertions, $18\\%$ . The highest number of substitution errors occur between $[\\mathrm{i}, \\mathrm{I}]$ , $[\\mathrm{e}, \\varepsilon]$ , $[\\mathrm{o}, \\mathfrak{o}]$ and $[\\mathrm{u}, \\mathrm{v}]$ —vowel pairs that contrast only in tenseness. This is consistent with the analysis of Meloni et al. (2021), where substitutions between tense-lax vowel pairs take up the largest portion of errors.", + "bbox": [ + 112, + 709, + 487, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We observe that other common substitution errors also happen between phonemes that share major phonetic features. This demonstrates that al", + "bbox": [ + 112, + 870, + 489, + 917 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "though no explicit phonetic information is fed directly into the model, the model makes mistakes motivated by phonetic similarity, like Meloni et al. (2021).", + "bbox": [ + 507, + 401, + 884, + 466 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We do not observe notable differences in the error statistics between the Transformer and the RNN.", + "bbox": [ + 507, + 468, + 882, + 514 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "6.4 Language relatedness", + "text_level": 1, + "bbox": [ + 507, + 527, + 726, + 543 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by Fourrier (2022), we probe our model for diachronic information on how genetically related each Romance language is to each other. We create a distance matrix between every pair of languages in a dataset by taking the cosine similarity between a pair's language embeddings. We then use sklearnn (Pedregosa et al., 2011)'s implementation of the Ward variance minimization algorithm (Ward Jr, 1963) to perform hierarchical clustering on the distance matrix. We take a consensus of the dendrograms from 5 different runs using the consense program from PHYLIP (Felsenstein, 2013).", + "bbox": [ + 505, + 548, + 885, + 756 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As we see in Figure 2, the Transformer captures more of the phylogenetic relationships among the languages correctly for the Rom-phon dataset. Indeed, the Generalized Quartet Distance (GQD) (Sand et al., 2013; Pompei et al., 2011; Rama et al., 2018) between the gold and predicted tree, calculated using quartetDist from the tqDist library (Sand et al., 2014), is 0.4 for the Transformer but 0.8 for the RNN. See Figure 5 in the Appendix for the results of the orthographic dataset.", + "bbox": [ + 507, + 758, + 885, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27", + "bbox": [ + 489, + 928, + 510, + 940 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/bc087ad415dada9f2e0679c05062b4785a9224f4fe5fba1aed8c77ffedcddbe8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetModelPED ↓NPED ↓Acc % ↑FER ↓BCFS ↑
SiniticRandom daughter (Chang et al., 2022)3.77020.84050%0.28930.2748
Majority constituent (Chang et al., 2022)3.50310.78060%0.20130.3695
CorPaR (List et al., 2022)3.27950.72780%0.39720.3332
SVM + PosStr (List et al., 2022)1.68940.369215.52%0.16690.5418
RNN (Meloni et al., 2021)1.06710.242135.65%0.08990.6781
Transformer (present work)0.95530.215041.12%0.08420.7033
Rom-phonRandom daughter (Chang et al., 2022)6.15340.69140.06%0.62640.4016
CorPaR + PosIni (List et al., 2022)1.68470.197822.18%0.07280.7403
SVM + PosStrIni (List et al., 2022)1.57870.186124.69%0.07130.7610
RNN (Meloni et al., 2021)0.96550.122452.31%0.03840.8296
Transformer (present work)0.89260.113753.75%0.03730.8435
Rom-orthRandom daughter (Chang et al., 2022)4.25670.48542.97%-0.5147
CorPaR + Ini (List et al., 2022)0.95310.116047.23%-0.8400
SVM + PosStr (List et al., 2022)0.89880.110550.43%-0.8501
RNN (Meloni et al., 2021)0.59410.077069.80%-0.8916
Transformer (present work)0.55250.072071.23%-0.9002
", + "bbox": [ + 114, + 80, + 882, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 3: Evaluation of models and baselines using various metrics, averaged across 5 runs (same hyperparameters, different seeds). Because Rom-orth is not in IPA, character edit distance is used instead of PED, and we cannot accurately calculate FER. See Section 6.1 for an explanation of each evaluation metric. See Table 4 for the standard deviation values.", + "bbox": [ + 112, + 476, + 882, + 533 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the Romance dataset only includes 5 daughter languages, our results are insufficient to corroborate or contradict Cathcart and Wandl (2020)'s findings: the more accurate the protoforms, the less accurate the phylogeny will be. It is not clear if the model's language embeddings are learning information that reflects shared innovations (sound changes that if shared among a set of daughter languages, would be acceptable justification for grouping them)—the only acceptable criterion for phylogenetic inference in historical linguistics (Campbell, 2013)—or if the model is learning superficial phonetic similarity.", + "bbox": [ + 112, + 558, + 489, + 769 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "7 Conclusion", + "text_level": 1, + "bbox": [ + 114, + 780, + 247, + 796 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By showing that Transformers can outperform previous architectures in protoform reconstruction despite the inherent data scarcity of the task, our work motivates future research in this area to take full advantage of the recent advancements in the Transformer space.", + "bbox": [ + 112, + 806, + 489, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Accurate supervised reconstruction can help pre", + "bbox": [ + 132, + 903, + 489, + 919 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "dict protoforms for cognate sets where linguists have not reconstructed one yet. Future work could reconstruct proto-languages whose linguist reconstructions are not available, by transferring knowledge learned from languages with already reconstructed protoforms. Furthermore, future work can leverage the abundance of work in unsupervised NMT to adapt our Transformer model for the unsupervised setting, a more realistic scenario for the historical linguist.", + "bbox": [ + 507, + 558, + 884, + 720 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Limitations", + "text_level": 1, + "bbox": [ + 509, + 741, + 613, + 757 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "One limitation of our work is that the RNN (Meloni et al., 2021) actually outperforms our Transformer on the Chinese dataset in Chang et al. (2022). In addition, as with other neural approaches, our model requires significant amounts of data, which is often not available to historical linguists researching less well-studied language families based on field reports. Romance and Chinese have relatively many cognate sets because the protoforms", + "bbox": [ + 507, + 774, + 884, + 919 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28", + "bbox": [ + 489, + 928, + 510, + 940 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "are documented5, but a low resource setup with 200 cognate sets would not fare well on our data-hungrier Transformer model. Furthermore, concatenating the entire cognate set may not work on language families with hundreds of languages such as Oceanic because the input sequence would be too long compared to the output protoform sequence.", + "bbox": [ + 112, + 83, + 492, + 211 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, we obtain our Chinese gold protoforms from Baxter and Sagart (2014)'s Middle Chinese reconstruction, which was actually a transcription of the Qieyun, a rhyme dictionary. Norman and Coblin (1995) disagree with relying on such a philological source and prefer comparative reconstructions that begin from daughter data. However, there is no available comparative reconstruction of Middle Chinese with protoforms corresponding to thousands of characters to use as a gold standard. Be that as it may, it seems clear that Middle Chinese as recorded in the Qieyun is not identical to the most recent ancestor of the Chinese languages. Its preface concedes that it is a compromise between Tang Dynasty dialects. The situation with Romance is, in some ways, comparable. Classical Latin—the variety on which we train—is not the direct ancestor of modern Romance languages. Instead, they are descended from Vulgar Latin or Proto-Romance, which is not well-attested and is primarily through graffiti and other informal inscriptions. Proto-Romance reconstructions are also not exhaustive. As a result, it is difficult to find a dataset like Meloni et al. (2021) with thousands of such ancestor forms. We are also limited to the faithfulness of espeak-ng's Latin G2P, from which Meloni et al. (2021) obtain their phonetic Romance dataset.", + "bbox": [ + 115, + 214, + 490, + 661 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For most language families, protoforms are not attested. In fact, as the term is often used, protoform refers to a form that is inferred only through linguists' comparative method. We adopt the other usage for simplicity. In practice, our approach would require reconstructions made by a linguist to serve as training labels for cognate sets.", + "bbox": [ + 112, + 665, + 489, + 778 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 114, + 791, + 287, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We would like to thank Liang (Leon) Lu for finding a bug in our implementation, Ying Chen for writing the code for the baselines, and Brendon Boldt and Graham Neubig for providing useful feedback", + "bbox": [ + 112, + 816, + 489, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "for the first iteration of our paper.", + "bbox": [ + 509, + 84, + 759, + 99 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 510, + 127, + 608, + 142 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Enrique Amigo, Julio Gonzalo, Javier Artiles, and Felisa Verdejo. 2009. A comparison of extrinsic clustering evaluation metrics based on formal constraints. Information retrieval, 12(4):461-486.", + "William H Baxter and Laurent Sagart. 2014. Old Chinese: A new reconstruction. Oxford University Press.", + "Lukas Biewald. 2020. Experiment tracking with weights and biases. Software available from wandb.com.", + "Alexandre Bouchard-Côté, Dan Klein, and Michael Jordan. 2008. Efficient inference in phylogenetic indel trees. In Advances in Neural Information Processing Systems, volume 21. Curran Associates, Inc.", + "Alexandre Bouchard-Côté, David Hall, Thomas L. Griffiths, and Dan Klein. 2013. Automated reconstruction of ancient languages using probabilistic models of sound change. Proceedings of the National Academy of Sciences, 110(11):4224-4229.", + "Lyle Campbell. 2013. *Historical Linguistics: an Introduction*. Edinburgh University Press.", + "Chundra Cathcart and Florian Wandl. 2020. In search of isoglosses: continuous and discrete language embeddings in Slavic historical phonology. In Proceedings of the 17th SIGMORPHON Workshop on Computational Research in Phonetics, Phonology, and Morphology, pages 233-244, Online. Association for Computational Linguistics.", + "Kalvin Chang, Chenxuan Cui, Youngmin Kim, and David R. Mortensen. 2022. WikiHan: A new comparative dataset for Chinese languages. In Proceedings of the 29th International Conference on Computational Linguistics (COLING 2022).", + "Kyunghyun Cho, Bart van Merrienboer, Dzmitry Bah-danau, and Yoshua Bengio. 2014. On the properties of neural machine translation: Encoder-decoder approaches. In Proceedings of SSST-8, Eighth Workshop on Syntax, Semantics and Structure in Statistical Translation, pages 103-111, Doha, Qatar. Association for Computational Linguistics.", + "Alina Maria Ciobanu and Liviu P. Dinu. 2018. Ab initio: Automatic Latin proto-word reconstruction. In Proceedings of the 27th International Conference on Computational Linguistics, pages 1604-1614, Santa Fe, New Mexico, USA. Association for Computational Linguistics.", + "Arthur P Dempster, Nan M Laird, and Donald B Rubin. 1977. Maximum likelihood from incomplete data via the em algorithm. Journal of the Royal Statistical Society: Series B (Methodological), 39(1):1-22." + ], + "bbox": [ + 509, + 149, + 885, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "In the case of Chinese, only equivalence classes of pronunciations and not exact pronunciations are recorded.", + "bbox": [ + 112, + 891, + 489, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "29", + "bbox": [ + 490, + 928, + 510, + 939 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Liviu Dinu and Alina Maria Ciobanu. 2014. Building a dataset of multilingual cognates for the Romanian lexicon. In Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC'14), pages 1038-1043, Reykjavik, Iceland. European Language Resources Association (ELRA).", + "Joseph Felsenstein. 2013. Phylip (phylogeny inference package), version 3.695. Department of Genome Sciences, University of Washington, Seattle.", + "Clémentine Fourrier. 2022. Neural Approaches to Historical Word Reconstruction. Ph.D. thesis, Université PSL (Paris Sciences & Lettres).", + "Andre He, Nicholas Tomlin, and Dan Klein. 2022. Neural unsupervised reconstruction of protolanguage word forms. arXiv preprint arXiv:2211.08684.", + "侯精一 Jingyi Hóu, editor. 2004. Xiandai Hanyu fangyan yinku 现代汉语方言音库 [Phonological database of Chinese dialects]. Shanghai Jiaoyu 上海教育, Shanghai 上海.", + "John D. Lafferty, Andrew McCallum, and Fernando C. N. Pereira. 2001. Conditional random fields: Probabilistic models for segmenting and labeling sequence data. In Proceedings of the Eighteenth International Conference on Machine Learning, ICML '01, page 282-289, San Francisco, CA, USA. Morgan Kaufmann Publishers Inc.", + "Vladimir I Levenshtein et al. 1966. Binary codes capable of correcting deletions, insertions, and reversals. Soviet physics doklady, 10(8):707-710.", + "Johann-Mattis List. 2019. Beyond edit distances: Comparing linguistic reconstruction systems. Theoretical Linguistics, 45(3-4):247-258.", + "Johann-Mattis List. 2021. CLDF dataset derived from Hóu's \"Phonological Database of Chinese Dialects\" from 2004. Zenodo.", + "Johann-Mattis List, Robert Forkel, and Nathan Hill. 2022. A new framework for fast automated phonological reconstruction using trimmed alignments and sound correspondence patterns. In Proceedings of the 3rd Workshop on Computational Approaches to Historical Language Change, pages 89-96, Dublin, Ireland. Association for Computational Linguistics.", + "Johann-Mattis List, Nelson-Sathi Shijulal, William Martin, and Hans Geisler. 2014. Using phylogenetic networks to model chinese dialect history. Language Dynamics and Change, 4(2):222-252.", + "Carlo Meloni, Shauli Ravfogel, and Yoav Goldberg. 2021. Ab antiquo: Neural proto-language reconstruction. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4460-4473, Online. Association for Computational Linguistics." + ], + "bbox": [ + 115, + 85, + 485, + 917 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "David R. Mortensen, Patrick Littell, Akash Bharadwaj, Kartik Goyal, Chris Dyer, and Lori S. Levin. 2016. Panphon: A resource for mapping IPA segments to articulatory feature vectors. In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 3475-3484.", + "Jerry L. Norman and W. South Coblin. 1995. A new approach to Chinese historical linguistics. Journal of the American Oriental Society, 115(4):576-584.", + "Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318.", + "F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. 2011. Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12:2825-2830.", + "Simone Pompei, Vittorio Loreto, and Francesca Tria. 2011. On the accuracy of language trees. PloS one, 6(6):e20109.", + "Taraka Rama, Johann-Mattis List, Johannes Wahle, and Gerhard Jäger. 2018. Are automatic methods for cognate detection good enough for phylogenetic reconstruction in historical linguistics? arXiv preprint arXiv:1804.05416.", + "Andreas Sand, Morten K Holt, Jens Johansen, Rolf Fagerberg, Gerth Stolting Brodal, Christian NS Pedersen, and Thomas Mailund. 2013. Algorithms for computing the triplet and quartet distances for binary general trees. *Biology*, 2(4):1189–1209.", + "Andreas Sand, Morten Kragelund Holt, Jens Johansen, Rolf Fagerberg, Gerth Stolting Brodal, Thomas Mailund, and Christian N. S. Pedersen. 2014. tqdist: A library for computing the quartet and triplet distances between binary or general trees. BMC Bioinformatics, yy(xx):ii-jj.", + "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30.", + "Joe H Ward Jr. 1963. Hierarchical grouping to optimize an objective function. Journal of the American statistical association, 58(301):236-244." + ], + "bbox": [ + 510, + 85, + 880, + 819 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "30", + "bbox": [ + 490, + 928, + 510, + 939 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A Training", + "text_level": 1, + "bbox": [ + 115, + 84, + 228, + 99 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We split $70\\%$ , $10\\%$ , and $20\\%$ of our dataset into train, validation, and test sets, respectively. We conduct hyperparameter searches using WandB (Biewald, 2020) and use early stopping, picking the epoch with lowest edit distance on validation data. All experiments are performed on a Ubuntu server with 4 GPUs and 20 CPUs. For both the RNN and the Transformer, Meloni et al. (2021)'s dataset takes less than 7 GPU hours to run, while Hóu (2004) takes less than 1 GPU hour. For the large Romance orthographic dataset, the RNN model has around 480,000 parameters, while the Transformer has around 800,000 parameters.", + "bbox": [ + 115, + 109, + 485, + 317 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "B Hyper-parameters", + "text_level": 1, + "bbox": [ + 115, + 330, + 310, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Refer to Table 5 and Table 6 for the best hyperparameters we found during hyperparameter search via WandB.", + "bbox": [ + 115, + 355, + 487, + 401 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "C Supplementary Results", + "text_level": 1, + "bbox": [ + 115, + 414, + 352, + 432 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In order to compare our model to earlier work, we used the Rom-phon and Rom-orth datasets from Meloni et al. (2021). However, this set includes a subset from Ciobanu and Dinu (2018) which is not freely redistributable. So that our results can be reproduced, we also computed them on the publicly available subset of Meloni et al. (2021)'s dataset, which is presented in Table 7.", + "bbox": [ + 115, + 439, + 485, + 567 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Phylogenetic trees for Chinese were also extracted from the RNN and Transformer models. These are shown in Figures 3 and 4.", + "bbox": [ + 115, + 569, + 487, + 615 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also plot the dendrograms derived from the Rom-ortho dataset in Figure 5.", + "bbox": [ + 115, + 617, + 485, + 648 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "31", + "bbox": [ + 490, + 928, + 507, + 940 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5248f64ba74abe05eb22a117ff6b520cd8f1b60dec47d4820d7c71d6254699a2.jpg", + "image_caption": [ + "Figure 3: Consensus tree of the dendrograms from the 5 runs of the Transformer for the Chinese dataset" + ], + "image_footnote": [], + "bbox": [ + 124, + 143, + 875, + 428 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/8dcd73a8ee4ee6b081a940c4a2652505f191cb10ce1052f0a7021a0b8dcc613c.jpg", + "image_caption": [ + "Figure 4: Consensus tree of the dendrograms from the 5 runs of the RNN for the Chinese dataset" + ], + "image_footnote": [], + "bbox": [ + 124, + 576, + 878, + 834 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "32", + "bbox": [ + 490, + 928, + 512, + 940 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/4eae6397ed120cb2ac347a487c7648b8b1e88f7345c1f3ca11567d35a24f1b24.jpg", + "image_caption": [ + "Figure 5: A gold phylogeny of Romance (left) compared with those derived by probing the RNN model (middle) and the Transformer model (right) on Rom-ortho. GQD is 0.4 for both models." + ], + "image_footnote": [], + "bbox": [ + 127, + 418, + 378, + 536 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/06d23af2c68209392b2831f214ca29d7fd3ecd61bfb0d986f2b4e4c4e9b31e68.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 419, + 613, + 537 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/c9ca5665eeebeb9fad4556f807fdefcdbc81b6e0eb60ee9cf4c64b6a02809d6c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 621, + 419, + 872, + 537 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "33", + "bbox": [ + 489, + 928, + 509, + 940 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/e796477de49899d75b81676a57f77e560288bc93fda55a4d5e0bfc446453ceec.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetModelPED ↓NPED ↓Acc % ↑FER ↓BCFS ↑
SiniticRandom daughter3.77020.84050%0.28930.2748
Majority constituent3.50310.78060%0.20130.3695
CorPaR3.27950.72780%0.39720.3332
SVM +PosStr1.68940.369215.52%0.16690.5418
RNN1.0671 ± 0.06190.2421 ± 0.014035.65% ± 1.60%0.0899 ± 0.00480.6781 ± 0.0174±
Transformer (present work)0.9553 ± 0.03920.2150 ± 0.007541.12% ± 2.3%0.0842 ± 0.00700.7033 ± 0.0087±
Rom-phonRandom daughter6.15340.69140.06%0.62640.4016
CorPaR +PosIni1.68470.197822.18%0.07280.7403
SVM +PosStrIni1.57870.186124.69%0.07130.7610
RNN0.9655 ± 0.01890.1224 ± 0.002252.31% ± 0.63%0.0384 ± 0.00110.8296 ± 0.0029±
Transformer (present work)0.8926 ± 0.01660.1137 ± 0.001753.75% ± 0.40%0.0373 ± 0.00090.8435 ± 0.0026±
Rom-orthRandom daughter4.25670.48542.97%-0.5147
CorPaR +Ini0.95310.116047.23%-0.8400
SVM +PosStr0.89880.110550.43%-0.8501
RNN0.5941 ± 0.01000.0770 ± 0.001569.80% ±0.22%-0.8916 ± 0.0019±
Transformer (present work)0.5525 ± 0.01040.0720 ± 0.001771.23% ± 0.52%-0.9002 ± 0.0017±
", + "bbox": [ + 114, + 165, + 882, + 778 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 4: Evaluation of models and baselines using various metrics, averaged across 5 runs (same hyperparameters, different seeds), with standard deviations. Because Rom-orth is not in IPA, character edit distance is used instead of PED, and we cannot accurately calculate FER. See Section 6.1 for an explanation of each evaluation metric.", + "bbox": [ + 112, + 785, + 882, + 831 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "34", + "bbox": [ + 490, + 928, + 510, + 940 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/204d40ab64127de14e24da13d38979c8fe0ea1098c77b9bb0a814d23ef15f193.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Romance (phon & orth)Sinitic
learning rate0.000130.0007487
num Encoder_layers32
num Decoder_layers35
embedding size128128
n_head88
dim_feedforward128647
dropout0.2020.1708861
training epochs200200
warmup epochs5032
weight decay00.0000001
batch size132
", + "bbox": [ + 115, + 175, + 588, + 375 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/fdaa1806f0633968c5c4bf81644afdad8e38ac450a67352e93d07d8cd58573e3.jpg", + "table_caption": [ + "Table 5: Hyper-parameters used in training the Transformer" + ], + "table_footnote": [], + "table_body": "
Romance-phonRomance-orthSinitic
learning rate0.000557390.0009640.000864
num Encoder_layers111
num Decoder_layers111
embedding size1075178
hidden size18513073
dim_feedforward147111136
dropout0.18080.3237940.321639
training epochs181193237
warmup epochs151515
batch size884
", + "bbox": [ + 115, + 612, + 636, + 795 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 6: Hyper-parameters used in training the RNN", + "bbox": [ + 121, + 804, + 478, + 819 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "35", + "bbox": [ + 490, + 928, + 510, + 940 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/7c65b04eb523de4fc6c0d2f08a75b4456743aed3581e9fc4c03daa01101f7786.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetModelPED ↓NPED ↓Acc % ↑FER ↓BCFS ↑
Rom-phonRandom daughter (Chang et al., 2022)7.18800.82010%1.13960.3406
CorPaR + Ini (List et al., 2022)2.08850.249114.29%0.08740.6799
SVM + PosStrIni (List et al., 2022)1.90050.227617.05%0.08830.7039
RNN (Meloni et al., 2021)1.45810.181536.68 %0.05920.7435
Transformer (present work)1.25160.157341.38%0.05500.7790
Rom-orthRandom daughter (Chang et al., 2022)6.32720.65420.55%-0.4023
CorPaR + PosStrIni (List et al., 2022)1.83130.200118.89%-0.7227
SVM + PosStr (List et al., 2022)1.69950.186721.66%-0.7454
RNN (Meloni et al., 2021)1.31890.150538.89%-0.7742
Transformer (present work)1.16220.134345.53%-0.7989
", + "bbox": [ + 115, + 185, + 882, + 469 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/a6afb984b87dcd51da4e9edc19a73e10023ca7868fa2041a3d966628c00fa39b.jpg", + "table_caption": [ + "Table 7: Evaluation of models and baselines with various metrics on Meloni et al. (2021)'s Romance datasets, where all entries from Dinu and Ciobanu (2014) are removed, for 1 run (using the hyperparameters of the best run on the full dataset)" + ], + "table_footnote": [], + "table_body": "
LatinRomanianFrenchItalianSpanishPortuguese
[kɔlle:ktio:nεm][kolektsie][kɔlɛksjɔ][kolletsoine][kolekθjon][kulisțu]
", + "bbox": [ + 169, + 739, + 826, + 785 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 8: One cognate set, with Latin as the protoform and all columns to its right as the daughter cognates", + "bbox": [ + 139, + 795, + 855, + 810 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "36", + "bbox": [ + 490, + 928, + 512, + 940 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A For every submission:", + "bbox": [ + 115, + 107, + 322, + 122 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "A1. Did you describe the limitations of your work? Section 8", + "A2. Did you discuss any potential risks of your work? Section 8", + "A3. Do the abstract and introduction summarize the paper's main claims? Section 1", + "□ A4. Have you used AI writing assistants when working on this paper? Not applicable. Left blank." + ], + "bbox": [ + 129, + 126, + 695, + 288 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "B Did you use or create scientific artifacts?", + "bbox": [ + 114, + 299, + 487, + 316 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Sections 3, 4", + "bbox": [ + 132, + 321, + 231, + 335 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "B1. Did you cite the creators of artifacts you used? Sections 3,4,5,6", + "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Sections 3, 5.1", + "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Not applicable. Left blank.", + "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Not applicable. Left blank.", + "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Section 3", + "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Table 2 and Appendix Section A" + ], + "bbox": [ + 129, + 346, + 880, + 753 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "C Did you run computational experiments?", + "bbox": [ + 114, + 764, + 492, + 781 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Section 4", + "bbox": [ + 132, + 785, + 206, + 800 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Appendix A", + "bbox": [ + 129, + 812, + 880, + 860 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.", + "bbox": [ + 112, + 868, + 877, + 892 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "ACL 2023 Responsible NLP Checklist", + "bbox": [ + 132, + 84, + 433, + 99 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "37", + "bbox": [ + 489, + 928, + 510, + 939 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?", + "bbox": [ + 129, + 83, + 878, + 115 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Hyperparameter search: 5.1 Hyperparameter values: Appendix Section B", + "bbox": [ + 149, + 117, + 697, + 131 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?", + "bbox": [ + 129, + 142, + 882, + 190 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Table 3", + "bbox": [ + 149, + 191, + 208, + 205 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?", + "bbox": [ + 129, + 217, + 882, + 265 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "5.1, 6.1, 6.3", + "bbox": [ + 149, + 267, + 242, + 280 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?", + "bbox": [ + 112, + 293, + 877, + 309 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 132, + 313, + 213, + 329 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?", + "bbox": [ + 127, + 340, + 882, + 372 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 374, + 248, + 388 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?", + "bbox": [ + 127, + 399, + 882, + 447 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 449, + 248, + 463 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?", + "bbox": [ + 127, + 475, + 882, + 521 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 524, + 248, + 539 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?", + "bbox": [ + 127, + 549, + 873, + 564 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 565, + 248, + 581 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?", + "bbox": [ + 127, + 592, + 880, + 623 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 626, + 248, + 640 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "38", + "bbox": [ + 490, + 928, + 510, + 940 + ], + "page_idx": 14 + } +] \ No newline at end of file diff --git a/2023/Transformed Protoform Reconstruction/244a271d-f768-4ade-bdfe-e0da000e90a0_model.json b/2023/Transformed Protoform Reconstruction/244a271d-f768-4ade-bdfe-e0da000e90a0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..494047441f7a97ea886bc3cf13bdc07841b2aa30 --- /dev/null +++ b/2023/Transformed Protoform Reconstruction/244a271d-f768-4ade-bdfe-e0da000e90a0_model.json @@ -0,0 +1,2199 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.291, + 0.091, + 0.709, + 0.11 + ], + "angle": 0, + "content": "Transformed Protoform Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.142, + 0.844, + 0.194 + ], + "angle": 0, + "content": "Young Min Kim* and Kalvin Chang* and Chenxuan Cui and David Mortensen \nLanguage Technologies Institute, Carnegie Mellon University \n{youngmik, kalvinc, cxcui, dmortens}@cs.cmu.edu" + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.253, + 0.342, + 0.269 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.277, + 0.462, + 0.518 + ], + "angle": 0, + "content": "Protoform reconstruction is the task of inferring how morphemes or words sounded in ancestral languages of a set of daughter languages. Meloni et al. (2021) achieved the state-of-the-art on Latin protoform reconstruction with an RNN-based encoder-decoder with attention model. We update their model with the state-of-the-art seq2seq model—the Transformer. Our model outperforms their model on a suite of different metrics on two different datasets: Meloni et al.'s Romance data of \\(8,000+\\) cognates (spanning 5 languages) and a Chinese dataset (Hóu, 2004) of \\(800+\\) cognates (spanning 39 varieties). We also probe our model for potential phylogenetic signal contained in the model. Our code is publicly available1." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.529, + 0.262, + 0.543 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.553, + 0.491, + 0.698 + ], + "angle": 0, + "content": "Languages change over time and sometimes diverge into multiple daughter languages. The common ancestor of a set of genetically related languages is their proto-language. While there are proto-languages such as Latin that are attested, they are the exception2. Reconstructed words and morphemes in proto-languages are called protoforms. The task of reconstructing unattested protolanguages is called protoform reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.698, + 0.49, + 0.827 + ], + "angle": 0, + "content": "Historical linguists reconstruct proto-languages by identifying systematic sound changes that can be inferred from correspondences between attested daughter languages (see Table 1). They compare the sounds between a set of cognates, or words with a common ancestor, to develop hypotheses about the types and chronologies of sound changes." + }, + { + "type": "table", + "bbox": [ + 0.528, + 0.25, + 0.868, + 0.336 + ], + "angle": 0, + "content": "
‘tooth’‘two’‘ten’
Englishtoothtwotent
Dutchtandtweetient
GermanZahnzweizehnz
PWG*tanþ*twai-*tehun*t
" + }, + { + "type": "table_caption", + "bbox": [ + 0.509, + 0.346, + 0.886, + 0.376 + ], + "angle": 0, + "content": "Table 1: Sound correspondences in West Germanic Languages and Proto-West-Germanic (PWG)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.401, + 0.885, + 0.497 + ], + "angle": 0, + "content": "This task is inherently data-constrained, especially for under-documented languages. Such data scarcity makes it a particularly difficult task for contemporary neural network architectures such as the Transformer (Vaswani et al., 2017), which are data hungry." + }, + { + "type": "text", + "bbox": [ + 0.529, + 0.499, + 0.871, + 0.515 + ], + "angle": 0, + "content": "The contributions of this paper are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.528, + 0.884, + 0.592 + ], + "angle": 0, + "content": "- Application of the Transformer architecture to the protoform reconstruction task, achieving state of the art performance, contrary to expectation." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.605, + 0.885, + 0.669 + ], + "angle": 0, + "content": "- Expansion of prior digital versions of Hóu (2004)'s Chinese dataset to include a total of 804 cognate sets across 39 modern varieties and Middle Chinese." + }, + { + "type": "list", + "bbox": [ + 0.532, + 0.528, + 0.885, + 0.669 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.684, + 0.667, + 0.699 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.71, + 0.885, + 0.92 + ], + "angle": 0, + "content": "Applying machine learning to protoform reconstruction is not new. Bouchard-Côté et al. (2013) learn an unsupervised protoform reconstruction model for the large Oceanic language family using Monte Carlo Expectation Maximization (Dempster et al., 1977; Bouchard-Côté et al., 2008), supervising the model with a gold phylogeny and using a probabilistic, generative model of sound change. He et al. (2022) modernize an earlier version of Bouchard-Côté et al. (2013)'s model with RNNs for a 4 language subset of Romance, but they rely on a bigram language model of Latin, making their model technically not unsupervised." + }, + { + "type": "page_footnote", + "bbox": [ + 0.14, + 0.832, + 0.269, + 0.846 + ], + "angle": 0, + "content": "* Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.139, + 0.847, + 0.434, + 0.858 + ], + "angle": 0, + "content": "" + }, + { + "type": "page_footnote", + "bbox": [ + 0.116, + 0.858, + 0.488, + 0.918 + ], + "angle": 0, + "content": "2 In fact, the proto-language from which Romance languages like Spanish and Italian are descended is not identical to Classical Latin but is, rather, a closely related and sparsely attested language sometimes called Proto-Romance or Vulgar Latin." + }, + { + "type": "list", + "bbox": [ + 0.116, + 0.832, + 0.488, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.49, + 0.929, + 0.512, + 0.94 + ], + "angle": 0, + "content": "24" + }, + { + "type": "footer", + "bbox": [ + 0.227, + 0.946, + 0.771, + 0.958 + ], + "angle": 0, + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + }, + { + "type": "footer", + "bbox": [ + 0.385, + 0.959, + 0.614, + 0.972 + ], + "angle": 0, + "content": "Volume 2: Short Papers, pages 24-38" + }, + { + "type": "footer", + "bbox": [ + 0.297, + 0.973, + 0.702, + 0.986 + ], + "angle": 0, + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.49, + 0.244 + ], + "angle": 0, + "content": "List et al. (2022) apply an SVM classifier to supervised reconstruction by treating sound correspondences as training examples. Note that there were no word boundaries in the input matrix; that is, all sound correspondences across the training set are flattened into one matrix. Furthermore, each language has an independent phonemic inventory. To learn contextual information, the authors experiment with adding features encoding the position of phonemes, among others." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.246, + 0.49, + 0.373 + ], + "angle": 0, + "content": "Ciobanu and Dinu (2018) learn a conditional random field (Lafferty et al., 2001) using n-gram features for supervised reconstruction and ensemble 5 daughter-to-protoform models. They use a dataset of 3,218 complete cognate sets spanning Latin (the proto-language) and 5 Romance languages: Romanian, French, Italian, Spanish, Portuguese." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.375, + 0.49, + 0.616 + ], + "angle": 0, + "content": "Meloni et al. (2021) employ a GRU-based seq2seq approach (Cho et al., 2014) to Latin protoform reconstruction and achieve state-of-the-art character edit distances. They extend Dinu and Ciobanu (2014)'s Romance data using data from Wiktionary—for a total of 8,799 cognate sets across 5 Romance languages plus Latin—in both orthographic and phonetic (IPA) representations. In their model, all entries comprising the cognate set are concatenated together in a fixed order to form a training example. Chang et al. (2022) applied Meloni et al. (2021)'s architecture to the reconstruction of Middle Chinese on a dataset of \\(5000+\\) cognate sets spanning 8 languages they compiled from Wiktionary.\\(^3\\)" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.617, + 0.49, + 0.873 + ], + "angle": 0, + "content": "Fourrier (2022) compares statistical machine translation, RNN, and Transformer architectures for protoform reconstruction, but they evaluate their results using BLEU scores (Papineni et al., 2002) instead of edit distance. They find that their Transformer model did not outperform the RNN models on protoform reconstruction. In addition, their multilingual NMT (neural machine translation) model predicts many languages instead of one target language and is trained on bilingual pairs for protoform reconstruction (e.g. Italian-Latin and Spanish-Latin), unlike comparative reconstruction. In contrast, we encode the entire cognate set consisting of multiple daughter languages (5 for the Romance dataset; 39 for Chinese) and predict the corresponding protoform." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.085, + 0.619, + 0.099 + ], + "angle": 0, + "content": "3 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.124, + 0.885, + 0.381 + ], + "angle": 0, + "content": "We train and test our model on Romance and Sinitic (Chinese) language datasets. For Romance languages, we use Meloni et al. (2021)'s dataset which consists of 8,799 cognate sets of Romanian, French, Italian, Spanish, Portuguese words and the corresponding Latin form (approximately, a protoform). There are two versions of this dataset: phonetic and orthographic. The phonetic dataset (Rom-phon) represents words with IPA symbols whereas the orthographic dataset (Rom-orth) represents words in the orthographic form of each language. We preserved all diacritics, except for vowel length. This dataset is an extension of Dinu and Ciobanu (2014)'s original dataset of 3,218 cognate sets, which is not publicly available. Refer to Table 2 for more information." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.413, + 0.878, + 0.43 + ], + "angle": 0, + "content": "3.1 Expanding digital versions of Hóu (2004)" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.446, + 0.885, + 0.639 + ], + "angle": 0, + "content": "For Sinitic languages, we created a dataset of Middle Chinese and its modern daughter languages. Middle Chinese is an unattested language, and we thus have to rely on Baxter and Sagart (2014)'s reconstructions of forms corresponding to 4,967 Chinese characters. We scraped Wiktionary to obtain Hóu (2004)'s phonetic representations of their modern reflexes. The resulting dataset contains 804 cognate sets of 39 modern Sinitic languages and the corresponding reconstructed Middle Chinese word. List (2021)'s version previously had 894 cognate sets across 15 varieties." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.671, + 0.602, + 0.687 + ], + "angle": 0, + "content": "4 Model" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.712, + 0.885, + 0.839 + ], + "angle": 0, + "content": "We propose a Transformer-based encoder-decoder architecture (Vaswani et al., 2017) because such models have produced state-of-the-art results on many sequence processing tasks. Transformers are by reputation data hungry, though, which poses a challenge to our problem setting, where the number of available training examples is often very small." + }, + { + "type": "page_footnote", + "bbox": [ + 0.113, + 0.881, + 0.489, + 0.919 + ], + "angle": 0, + "content": "3The original dataset contains 21,000 cognate sets, but only \\(5000+\\) had at least 3 daughter entries and were used as input to the model." + }, + { + "type": "page_footnote", + "bbox": [ + 0.509, + 0.87, + 0.883, + 0.919 + ], + "angle": 0, + "content": "4https://en.wiktionary.org/wiki/Module: zh/data/dial-pron/documentation originally had 1,023 characters, but only 804 had reconstructions from Baxter and Sagart (2014)." + }, + { + "type": "page_number", + "bbox": [ + 0.49, + 0.929, + 0.511, + 0.941 + ], + "angle": 0, + "content": "25" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.154, + 0.085, + 0.45, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.275, + 0.49, + 0.334 + ], + "angle": 0, + "content": "Figure 1: Diagram of our encoder-decoder architecture. Additive positional encoding and language embedding are applied to each daughter sequence before all daughter sequences are concatenated into a single sequence." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.355, + 0.49, + 0.597 + ], + "angle": 0, + "content": "We modify the standard encoder-decoder architecture to accommodate the structure of our datasets, where multiple daughter sequences correspond to a single protoform sequence. Like Meloni et al. (2021), the daughter sequences are concatenated into a single sequence before being fed into the encoder. Because we only care about the relative position between tokens within each daughter sequence but not across daughter sequences, positional encoding is applied to each individual daughter sequence before concatenation. Along with positional encoding, an additive language embedding is applied to the token embeddings to differentiate between input tokens of different daughter languages." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.612, + 0.26, + 0.628 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.64, + 0.234, + 0.654 + ], + "angle": 0, + "content": "5.1 Baselines" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.662, + 0.49, + 0.92 + ], + "angle": 0, + "content": "We compare our Transformer model to a variety of baselines. For Meloni et al. (2021), we use Chang et al. (2022)'s PyTorch re-implementation and reran a Bayesian hyperparameter search using WandB (Biewald, 2020) to ensure a more fair comparison (since our model is tuned with WandB as well). We also include the random daughter (randomly designate a daughter form as the protoform and assume no sound change) and the majority constituent baselines (predict the most common phoneme in each syllable constituent) from Chang et al. (2022). For the SVM and CoRPaR classifiers (List et al., 2022), we experiment with different contextual features, such as Pos (position), Str (prosodic structure), and Ini (whether or not the phoneme appears word-initially or wordfinally)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.246 + ], + "angle": 0, + "content": "We publish results on Meloni et al. (2021)'s full set of 8,799 cognates but cannot redistribute this set due to Dinu and Ciobanu (2014)'s restrictions. For reproducibility, we include results on Meloni et al. (2021)'s public subset of 5,419 cognates in the Appendix (Table 7), both of which include vowel length. Observe that these results are worse than those obtained on the full set, suggesting that the RNN and Transformer are dependent on a wealth of training data." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.258, + 0.667, + 0.273 + ], + "angle": 0, + "content": "5.2 Preprocessing" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.279, + 0.885, + 0.408 + ], + "angle": 0, + "content": "In all our datasets, we merge diacritics to their base segments to form a multi-character token. For instance, the sequence \\(\\left[\\mathrm{t},\\mathrm{h}\\right]\\) is concatenated to \\(\\left[\\mathrm{th}\\right]\\). This ensures that phonemes are treated as one token. For Chinese, tone contours (a sequence of tones) are treated as one token. When multiple pronunciation variants are listed for a single Chinese character, we arbitrarily pick the first one." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.421, + 0.742, + 0.435 + ], + "angle": 0, + "content": "6 Results and Discussion" + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.447, + 0.706, + 0.46 + ], + "angle": 0, + "content": "6.1 Evaluation criteria" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.468, + 0.885, + 0.806 + ], + "angle": 0, + "content": "We evaluate the predicted protoforms using edit distance (Levenshtein et al., 1966), normalized edit distance (edit distance normalized by the length of the target) and accuracy (the percentage of protoforms that are reconstructed without any mistakes). Like Chang et al. (2022), we also use feature error rate calculated using articulatory feature vectors from PanPhon (Mortensen et al., 2016) because it reflects the phonetic similarity between the prediction and the gold protoform. For datasets with phonetic transcriptions (Romance-phonetic and Chinese), we use phoneme edit distance and normalized phoneme edit distance. As List (2019) suggests, we use B-Cubed F Scores (Amigo et al., 2009) to capture the structural similarity between the gold and predicted protoforms (0: structurally dissimilar, 1: similar). With the exception of character and phoneme edit distance, the metrics enable fair comparison across different language families, which will differ in the average word length." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.818, + 0.614, + 0.832 + ], + "angle": 0, + "content": "6.2 Results" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.839, + 0.884, + 0.919 + ], + "angle": 0, + "content": "Table 3 shows that our model consistently has the best performance on all datasets with regards to most metrics. The results were averaged across 5 runs. Out of all datasets, our model performs best on the Rom-orth dataset, where we achieve a \\(7.0\\%\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.49, + 0.929, + 0.512, + 0.941 + ], + "angle": 0, + "content": "26" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.115, + 0.081, + 0.886, + 0.187 + ], + "angle": 0, + "content": "
Language FamilySource# varietiesCognate setsProto-language
Rom-phonDinu and Ciobanu (2014), Meloni et al. (2021)58,799Latin
Rom-orthDinu and Ciobanu (2014), Meloni et al. (2021)58,799Latin
Sinitic (Chinese)Hóu (2004)39804Middle Chinese
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.194, + 0.882, + 0.209 + ], + "angle": 0, + "content": "Table 2: Statistics on both datasets used in our experiments. # varieties refers to the number of daughter varieties." + }, + { + "type": "image", + "bbox": [ + 0.12, + 0.22, + 0.372, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.22, + 0.642, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.22, + 0.88, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.349, + 0.884, + 0.38 + ], + "angle": 0, + "content": "Figure 2: A gold phylogeny of Romance (left) compared with those derived by probing the RNN model (middle) and the Transformer model (right) on Rom-phon." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.403, + 0.49, + 0.549 + ], + "angle": 0, + "content": "decrease in phoneme edit distance and a 1.43p.p improvement in accuracy relative to the RNN baseline. We observe the most dramatic performance difference with the RNN baseline on the Sinitic dataset: a \\(10.48\\%\\) decrease in phoneme edit distance and a 5.47p.p increase in accuracy. For reproducibility, results on the publicly available portion of the Rom-phon and Rom-orth datasets are provided in Table 7 in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.56, + 0.228, + 0.575 + ], + "angle": 0, + "content": "6.3 Analysis" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.581, + 0.489, + 0.709 + ], + "angle": 0, + "content": "We observe that the BCFS is relatively high for the Romance non-neural baselines compared to those of the Chinese ones. This suggests that the sound changes in the Romance datasets are more regular than that of Chinese, which corroborates List et al. (2014)'s results that more than half of the Chinese characters in their dataset could not be explained by a tree model." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.71, + 0.489, + 0.87 + ], + "angle": 0, + "content": "We examine the errors made by the Transformer model on the Rom-phon dataset. Substitutions constitute around \\(61\\%\\) of the errors made by the Transformer; deletions, \\(21\\%\\), and insertions, \\(18\\%\\). The highest number of substitution errors occur between \\([\\mathrm{i}, \\mathrm{I}]\\), \\([\\mathrm{e}, \\varepsilon]\\), \\([\\mathrm{o}, \\mathfrak{o}]\\) and \\([\\mathrm{u}, \\mathrm{v}]\\)—vowel pairs that contrast only in tenseness. This is consistent with the analysis of Meloni et al. (2021), where substitutions between tense-lax vowel pairs take up the largest portion of errors." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.871, + 0.49, + 0.919 + ], + "angle": 0, + "content": "We observe that other common substitution errors also happen between phonemes that share major phonetic features. This demonstrates that al" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.403, + 0.885, + 0.467 + ], + "angle": 0, + "content": "though no explicit phonetic information is fed directly into the model, the model makes mistakes motivated by phonetic similarity, like Meloni et al. (2021)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.469, + 0.884, + 0.515 + ], + "angle": 0, + "content": "We do not observe notable differences in the error statistics between the Transformer and the RNN." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.529, + 0.727, + 0.544 + ], + "angle": 0, + "content": "6.4 Language relatedness" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.549, + 0.886, + 0.757 + ], + "angle": 0, + "content": "Inspired by Fourrier (2022), we probe our model for diachronic information on how genetically related each Romance language is to each other. We create a distance matrix between every pair of languages in a dataset by taking the cosine similarity between a pair's language embeddings. We then use sklearnn (Pedregosa et al., 2011)'s implementation of the Ward variance minimization algorithm (Ward Jr, 1963) to perform hierarchical clustering on the distance matrix. We take a consensus of the dendrograms from 5 different runs using the consense program from PHYLIP (Felsenstein, 2013)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.759, + 0.886, + 0.919 + ], + "angle": 0, + "content": "As we see in Figure 2, the Transformer captures more of the phylogenetic relationships among the languages correctly for the Rom-phon dataset. Indeed, the Generalized Quartet Distance (GQD) (Sand et al., 2013; Pompei et al., 2011; Rama et al., 2018) between the gold and predicted tree, calculated using quartetDist from the tqDist library (Sand et al., 2014), is 0.4 for the Transformer but 0.8 for the RNN. See Figure 5 in the Appendix for the results of the orthographic dataset." + }, + { + "type": "page_number", + "bbox": [ + 0.49, + 0.929, + 0.511, + 0.941 + ], + "angle": 0, + "content": "27" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.115, + 0.082, + 0.884, + 0.468 + ], + "angle": 0, + "content": "
DatasetModelPED ↓NPED ↓Acc % ↑FER ↓BCFS ↑
SiniticRandom daughter (Chang et al., 2022)3.77020.84050%0.28930.2748
Majority constituent (Chang et al., 2022)3.50310.78060%0.20130.3695
CorPaR (List et al., 2022)3.27950.72780%0.39720.3332
SVM + PosStr (List et al., 2022)1.68940.369215.52%0.16690.5418
RNN (Meloni et al., 2021)1.06710.242135.65%0.08990.6781
Transformer (present work)0.95530.215041.12%0.08420.7033
Rom-phonRandom daughter (Chang et al., 2022)6.15340.69140.06%0.62640.4016
CorPaR + PosIni (List et al., 2022)1.68470.197822.18%0.07280.7403
SVM + PosStrIni (List et al., 2022)1.57870.186124.69%0.07130.7610
RNN (Meloni et al., 2021)0.96550.122452.31%0.03840.8296
Transformer (present work)0.89260.113753.75%0.03730.8435
Rom-orthRandom daughter (Chang et al., 2022)4.25670.48542.97%-0.5147
CorPaR + Ini (List et al., 2022)0.95310.116047.23%-0.8400
SVM + PosStr (List et al., 2022)0.89880.110550.43%-0.8501
RNN (Meloni et al., 2021)0.59410.077069.80%-0.8916
Transformer (present work)0.55250.072071.23%-0.9002
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.477, + 0.884, + 0.534 + ], + "angle": 0, + "content": "Table 3: Evaluation of models and baselines using various metrics, averaged across 5 runs (same hyperparameters, different seeds). Because Rom-orth is not in IPA, character edit distance is used instead of PED, and we cannot accurately calculate FER. See Section 6.1 for an explanation of each evaluation metric. See Table 4 for the standard deviation values." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.56, + 0.49, + 0.77 + ], + "angle": 0, + "content": "Since the Romance dataset only includes 5 daughter languages, our results are insufficient to corroborate or contradict Cathcart and Wandl (2020)'s findings: the more accurate the protoforms, the less accurate the phylogeny will be. It is not clear if the model's language embeddings are learning information that reflects shared innovations (sound changes that if shared among a set of daughter languages, would be acceptable justification for grouping them)—the only acceptable criterion for phylogenetic inference in historical linguistics (Campbell, 2013)—or if the model is learning superficial phonetic similarity." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.781, + 0.248, + 0.797 + ], + "angle": 0, + "content": "7 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.807, + 0.49, + 0.902 + ], + "angle": 0, + "content": "By showing that Transformers can outperform previous architectures in protoform reconstruction despite the inherent data scarcity of the task, our work motivates future research in this area to take full advantage of the recent advancements in the Transformer space." + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.904, + 0.49, + 0.92 + ], + "angle": 0, + "content": "Accurate supervised reconstruction can help pre" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.56, + 0.885, + 0.721 + ], + "angle": 0, + "content": "dict protoforms for cognate sets where linguists have not reconstructed one yet. Future work could reconstruct proto-languages whose linguist reconstructions are not available, by transferring knowledge learned from languages with already reconstructed protoforms. Furthermore, future work can leverage the abundance of work in unsupervised NMT to adapt our Transformer model for the unsupervised setting, a more realistic scenario for the historical linguist." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.742, + 0.615, + 0.758 + ], + "angle": 0, + "content": "Limitations" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.775, + 0.885, + 0.92 + ], + "angle": 0, + "content": "One limitation of our work is that the RNN (Meloni et al., 2021) actually outperforms our Transformer on the Chinese dataset in Chang et al. (2022). In addition, as with other neural approaches, our model requires significant amounts of data, which is often not available to historical linguists researching less well-studied language families based on field reports. Romance and Chinese have relatively many cognate sets because the protoforms" + }, + { + "type": "page_number", + "bbox": [ + 0.49, + 0.929, + 0.512, + 0.941 + ], + "angle": 0, + "content": "28" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.084, + 0.493, + 0.212 + ], + "angle": 0, + "content": "are documented5, but a low resource setup with 200 cognate sets would not fare well on our data-hungrier Transformer model. Furthermore, concatenating the entire cognate set may not work on language families with hundreds of languages such as Oceanic because the input sequence would be too long compared to the output protoform sequence." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.215, + 0.492, + 0.662 + ], + "angle": 0, + "content": "Finally, we obtain our Chinese gold protoforms from Baxter and Sagart (2014)'s Middle Chinese reconstruction, which was actually a transcription of the Qieyun, a rhyme dictionary. Norman and Coblin (1995) disagree with relying on such a philological source and prefer comparative reconstructions that begin from daughter data. However, there is no available comparative reconstruction of Middle Chinese with protoforms corresponding to thousands of characters to use as a gold standard. Be that as it may, it seems clear that Middle Chinese as recorded in the Qieyun is not identical to the most recent ancestor of the Chinese languages. Its preface concedes that it is a compromise between Tang Dynasty dialects. The situation with Romance is, in some ways, comparable. Classical Latin—the variety on which we train—is not the direct ancestor of modern Romance languages. Instead, they are descended from Vulgar Latin or Proto-Romance, which is not well-attested and is primarily through graffiti and other informal inscriptions. Proto-Romance reconstructions are also not exhaustive. As a result, it is difficult to find a dataset like Meloni et al. (2021) with thousands of such ancestor forms. We are also limited to the faithfulness of espeak-ng's Latin G2P, from which Meloni et al. (2021) obtain their phonetic Romance dataset." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.666, + 0.49, + 0.779 + ], + "angle": 0, + "content": "For most language families, protoforms are not attested. In fact, as the term is often used, protoform refers to a form that is inferred only through linguists' comparative method. We adopt the other usage for simplicity. In practice, our approach would require reconstructions made by a linguist to serve as training labels for cognate sets." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.792, + 0.288, + 0.809 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.818, + 0.49, + 0.884 + ], + "angle": 0, + "content": "We would like to thank Liang (Leon) Lu for finding a bug in our implementation, Ying Chen for writing the code for the baselines, and Brendon Boldt and Graham Neubig for providing useful feedback" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.085, + 0.761, + 0.101 + ], + "angle": 0, + "content": "for the first iteration of our paper." + }, + { + "type": "title", + "bbox": [ + 0.511, + 0.128, + 0.61, + 0.143 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.151, + 0.886, + 0.205 + ], + "angle": 0, + "content": "Enrique Amigo, Julio Gonzalo, Javier Artiles, and Felisa Verdejo. 2009. A comparison of extrinsic clustering evaluation metrics based on formal constraints. Information retrieval, 12(4):461-486." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.215, + 0.886, + 0.255 + ], + "angle": 0, + "content": "William H Baxter and Laurent Sagart. 2014. Old Chinese: A new reconstruction. Oxford University Press." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.265, + 0.885, + 0.305 + ], + "angle": 0, + "content": "Lukas Biewald. 2020. Experiment tracking with weights and biases. Software available from wandb.com." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.316, + 0.885, + 0.37 + ], + "angle": 0, + "content": "Alexandre Bouchard-Côté, Dan Klein, and Michael Jordan. 2008. Efficient inference in phylogenetic indel trees. In Advances in Neural Information Processing Systems, volume 21. Curran Associates, Inc." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.38, + 0.886, + 0.446 + ], + "angle": 0, + "content": "Alexandre Bouchard-Côté, David Hall, Thomas L. Griffiths, and Dan Klein. 2013. Automated reconstruction of ancient languages using probabilistic models of sound change. Proceedings of the National Academy of Sciences, 110(11):4224-4229." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.456, + 0.885, + 0.484 + ], + "angle": 0, + "content": "Lyle Campbell. 2013. *Historical Linguistics: an Introduction*. Edinburgh University Press." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.494, + 0.886, + 0.587 + ], + "angle": 0, + "content": "Chundra Cathcart and Florian Wandl. 2020. In search of isoglosses: continuous and discrete language embeddings in Slavic historical phonology. In Proceedings of the 17th SIGMORPHON Workshop on Computational Research in Phonetics, Phonology, and Morphology, pages 233-244, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.596, + 0.886, + 0.663 + ], + "angle": 0, + "content": "Kalvin Chang, Chenxuan Cui, Youngmin Kim, and David R. Mortensen. 2022. WikiHan: A new comparative dataset for Chinese languages. In Proceedings of the 29th International Conference on Computational Linguistics (COLING 2022)." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.673, + 0.886, + 0.766 + ], + "angle": 0, + "content": "Kyunghyun Cho, Bart van Merrienboer, Dzmitry Bah-danau, and Yoshua Bengio. 2014. On the properties of neural machine translation: Encoder-decoder approaches. In Proceedings of SSST-8, Eighth Workshop on Syntax, Semantics and Structure in Statistical Translation, pages 103-111, Doha, Qatar. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.776, + 0.886, + 0.855 + ], + "angle": 0, + "content": "Alina Maria Ciobanu and Liviu P. Dinu. 2018. Ab initio: Automatic Latin proto-word reconstruction. In Proceedings of the 27th International Conference on Computational Linguistics, pages 1604-1614, Santa Fe, New Mexico, USA. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.866, + 0.886, + 0.919 + ], + "angle": 0, + "content": "Arthur P Dempster, Nan M Laird, and Donald B Rubin. 1977. Maximum likelihood from incomplete data via the em algorithm. Journal of the Royal Statistical Society: Series B (Methodological), 39(1):1-22." + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.151, + 0.886, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.892, + 0.49, + 0.919 + ], + "angle": 0, + "content": "In the case of Chinese, only equivalence classes of pronunciations and not exact pronunciations are recorded." + }, + { + "type": "page_number", + "bbox": [ + 0.491, + 0.929, + 0.511, + 0.94 + ], + "angle": 0, + "content": "29" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.086, + 0.487, + 0.166 + ], + "angle": 0, + "content": "Liviu Dinu and Alina Maria Ciobanu. 2014. Building a dataset of multilingual cognates for the Romanian lexicon. In Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC'14), pages 1038-1043, Reykjavik, Iceland. European Language Resources Association (ELRA)." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.177, + 0.486, + 0.217 + ], + "angle": 0, + "content": "Joseph Felsenstein. 2013. Phylip (phylogeny inference package), version 3.695. Department of Genome Sciences, University of Washington, Seattle." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.229, + 0.487, + 0.268 + ], + "angle": 0, + "content": "Clémentine Fourrier. 2022. Neural Approaches to Historical Word Reconstruction. Ph.D. thesis, Université PSL (Paris Sciences & Lettres)." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.281, + 0.487, + 0.32 + ], + "angle": 0, + "content": "Andre He, Nicholas Tomlin, and Dan Klein. 2022. Neural unsupervised reconstruction of protolanguage word forms. arXiv preprint arXiv:2211.08684." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.332, + 0.487, + 0.386 + ], + "angle": 0, + "content": "侯精一 Jingyi Hóu, editor. 2004. Xiandai Hanyu fangyan yinku 现代汉语方言音库 [Phonological database of Chinese dialects]. Shanghai Jiaoyu 上海教育, Shanghai 上海." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.398, + 0.487, + 0.489 + ], + "angle": 0, + "content": "John D. Lafferty, Andrew McCallum, and Fernando C. N. Pereira. 2001. Conditional random fields: Probabilistic models for segmenting and labeling sequence data. In Proceedings of the Eighteenth International Conference on Machine Learning, ICML '01, page 282-289, San Francisco, CA, USA. Morgan Kaufmann Publishers Inc." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.502, + 0.487, + 0.542 + ], + "angle": 0, + "content": "Vladimir I Levenshtein et al. 1966. Binary codes capable of correcting deletions, insertions, and reversals. Soviet physics doklady, 10(8):707-710." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.554, + 0.487, + 0.593 + ], + "angle": 0, + "content": "Johann-Mattis List. 2019. Beyond edit distances: Comparing linguistic reconstruction systems. Theoretical Linguistics, 45(3-4):247-258." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.606, + 0.487, + 0.645 + ], + "angle": 0, + "content": "Johann-Mattis List. 2021. CLDF dataset derived from Hóu's \"Phonological Database of Chinese Dialects\" from 2004. Zenodo." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.657, + 0.487, + 0.749 + ], + "angle": 0, + "content": "Johann-Mattis List, Robert Forkel, and Nathan Hill. 2022. A new framework for fast automated phonological reconstruction using trimmed alignments and sound correspondence patterns. In Proceedings of the 3rd Workshop on Computational Approaches to Historical Language Change, pages 89-96, Dublin, Ireland. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.762, + 0.487, + 0.814 + ], + "angle": 0, + "content": "Johann-Mattis List, Nelson-Sathi Shijulal, William Martin, and Hans Geisler. 2014. Using phylogenetic networks to model chinese dialect history. Language Dynamics and Change, 4(2):222-252." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.826, + 0.487, + 0.918 + ], + "angle": 0, + "content": "Carlo Meloni, Shauli Ravfogel, and Yoav Goldberg. 2021. Ab antiquo: Neural proto-language reconstruction. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4460-4473, Online. Association for Computational Linguistics." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.487, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.086, + 0.882, + 0.177 + ], + "angle": 0, + "content": "David R. Mortensen, Patrick Littell, Akash Bharadwaj, Kartik Goyal, Chris Dyer, and Lori S. Levin. 2016. Panphon: A resource for mapping IPA segments to articulatory feature vectors. In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 3475-3484." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.188, + 0.882, + 0.227 + ], + "angle": 0, + "content": "Jerry L. Norman and W. South Coblin. 1995. A new approach to Chinese historical linguistics. Journal of the American Oriental Society, 115(4):576-584." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.238, + 0.882, + 0.303 + ], + "angle": 0, + "content": "Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.313, + 0.882, + 0.403 + ], + "angle": 0, + "content": "F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. 2011. Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12:2825-2830." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.416, + 0.882, + 0.454 + ], + "angle": 0, + "content": "Simone Pompei, Vittorio Loreto, and Francesca Tria. 2011. On the accuracy of language trees. PloS one, 6(6):e20109." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.465, + 0.882, + 0.529 + ], + "angle": 0, + "content": "Taraka Rama, Johann-Mattis List, Johannes Wahle, and Gerhard Jäger. 2018. Are automatic methods for cognate detection good enough for phylogenetic reconstruction in historical linguistics? arXiv preprint arXiv:1804.05416." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.541, + 0.882, + 0.606 + ], + "angle": 0, + "content": "Andreas Sand, Morten K Holt, Jens Johansen, Rolf Fagerberg, Gerth Stolting Brodal, Christian NS Pedersen, and Thomas Mailund. 2013. Algorithms for computing the triplet and quartet distances for binary general trees. *Biology*, 2(4):1189–1209." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.616, + 0.882, + 0.696 + ], + "angle": 0, + "content": "Andreas Sand, Morten Kragelund Holt, Jens Johansen, Rolf Fagerberg, Gerth Stolting Brodal, Thomas Mailund, and Christian N. S. Pedersen. 2014. tqdist: A library for computing the quartet and triplet distances between binary or general trees. BMC Bioinformatics, yy(xx):ii-jj." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.705, + 0.882, + 0.771 + ], + "angle": 0, + "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.781, + 0.882, + 0.82 + ], + "angle": 0, + "content": "Joe H Ward Jr. 1963. Hierarchical grouping to optimize an objective function. Journal of the American statistical association, 58(301):236-244." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.882, + 0.82 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.491, + 0.929, + 0.511, + 0.94 + ], + "angle": 0, + "content": "30" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.116, + 0.085, + 0.23, + 0.101 + ], + "angle": 0, + "content": "A Training" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.11, + 0.486, + 0.318 + ], + "angle": 0, + "content": "We split \\(70\\%\\), \\(10\\%\\), and \\(20\\%\\) of our dataset into train, validation, and test sets, respectively. We conduct hyperparameter searches using WandB (Biewald, 2020) and use early stopping, picking the epoch with lowest edit distance on validation data. All experiments are performed on a Ubuntu server with 4 GPUs and 20 CPUs. For both the RNN and the Transformer, Meloni et al. (2021)'s dataset takes less than 7 GPU hours to run, while Hóu (2004) takes less than 1 GPU hour. For the large Romance orthographic dataset, the RNN model has around 480,000 parameters, while the Transformer has around 800,000 parameters." + }, + { + "type": "title", + "bbox": [ + 0.116, + 0.331, + 0.312, + 0.348 + ], + "angle": 0, + "content": "B Hyper-parameters" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.356, + 0.488, + 0.402 + ], + "angle": 0, + "content": "Refer to Table 5 and Table 6 for the best hyperparameters we found during hyperparameter search via WandB." + }, + { + "type": "title", + "bbox": [ + 0.117, + 0.416, + 0.353, + 0.433 + ], + "angle": 0, + "content": "C Supplementary Results" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.441, + 0.486, + 0.568 + ], + "angle": 0, + "content": "In order to compare our model to earlier work, we used the Rom-phon and Rom-orth datasets from Meloni et al. (2021). However, this set includes a subset from Ciobanu and Dinu (2018) which is not freely redistributable. So that our results can be reproduced, we also computed them on the publicly available subset of Meloni et al. (2021)'s dataset, which is presented in Table 7." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.57, + 0.488, + 0.617 + ], + "angle": 0, + "content": "Phylogenetic trees for Chinese were also extracted from the RNN and Transformer models. These are shown in Figures 3 and 4." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.618, + 0.486, + 0.649 + ], + "angle": 0, + "content": "We also plot the dendrograms derived from the Rom-ortho dataset in Figure 5." + }, + { + "type": "page_number", + "bbox": [ + 0.492, + 0.929, + 0.509, + 0.941 + ], + "angle": 0, + "content": "31" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.125, + 0.145, + 0.877, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.147, + 0.44, + 0.851, + 0.457 + ], + "angle": 0, + "content": "Figure 3: Consensus tree of the dendrograms from the 5 runs of the Transformer for the Chinese dataset" + }, + { + "type": "image", + "bbox": [ + 0.125, + 0.577, + 0.88, + 0.835 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.171, + 0.85, + 0.828, + 0.866 + ], + "angle": 0, + "content": "Figure 4: Consensus tree of the dendrograms from the 5 runs of the RNN for the Chinese dataset" + }, + { + "type": "page_number", + "bbox": [ + 0.491, + 0.929, + 0.513, + 0.941 + ], + "angle": 0, + "content": "32" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.129, + 0.419, + 0.379, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.42, + 0.614, + 0.538 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.42, + 0.873, + 0.538 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.552, + 0.883, + 0.582 + ], + "angle": 0, + "content": "Figure 5: A gold phylogeny of Romance (left) compared with those derived by probing the RNN model (middle) and the Transformer model (right) on Rom-ortho. GQD is 0.4 for both models." + }, + { + "type": "page_number", + "bbox": [ + 0.49, + 0.929, + 0.51, + 0.941 + ], + "angle": 0, + "content": "33" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.115, + 0.166, + 0.884, + 0.779 + ], + "angle": 0, + "content": "
DatasetModelPED ↓NPED ↓Acc % ↑FER ↓BCFS ↑
SiniticRandom daughter3.77020.84050%0.28930.2748
Majority constituent3.50310.78060%0.20130.3695
CorPaR3.27950.72780%0.39720.3332
SVM +PosStr1.68940.369215.52%0.16690.5418
RNN1.0671 ± 0.06190.2421 ± 0.014035.65% ± 1.60%0.0899 ± 0.00480.6781 ± 0.0174±
Transformer (present work)0.9553 ± 0.03920.2150 ± 0.007541.12% ± 2.3%0.0842 ± 0.00700.7033 ± 0.0087±
Rom-phonRandom daughter6.15340.69140.06%0.62640.4016
CorPaR +PosIni1.68470.197822.18%0.07280.7403
SVM +PosStrIni1.57870.186124.69%0.07130.7610
RNN0.9655 ± 0.01890.1224 ± 0.002252.31% ± 0.63%0.0384 ± 0.00110.8296 ± 0.0029±
Transformer (present work)0.8926 ± 0.01660.1137 ± 0.001753.75% ± 0.40%0.0373 ± 0.00090.8435 ± 0.0026±
Rom-orthRandom daughter4.25670.48542.97%-0.5147
CorPaR +Ini0.95310.116047.23%-0.8400
SVM +PosStr0.89880.110550.43%-0.8501
RNN0.5941 ± 0.01000.0770 ± 0.001569.80% ±0.22%-0.8916 ± 0.0019±
Transformer (present work)0.5525 ± 0.01040.0720 ± 0.001771.23% ± 0.52%-0.9002 ± 0.0017±
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.787, + 0.884, + 0.832 + ], + "angle": 0, + "content": "Table 4: Evaluation of models and baselines using various metrics, averaged across 5 runs (same hyperparameters, different seeds), with standard deviations. Because Rom-orth is not in IPA, character edit distance is used instead of PED, and we cannot accurately calculate FER. See Section 6.1 for an explanation of each evaluation metric." + }, + { + "type": "page_number", + "bbox": [ + 0.491, + 0.929, + 0.511, + 0.941 + ], + "angle": 0, + "content": "34" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.117, + 0.177, + 0.589, + 0.376 + ], + "angle": 0, + "content": "
Romance (phon & orth)Sinitic
learning rate0.000130.0007487
num Encoder_layers32
num Decoder_layers35
embedding size128128
n_head88
dim_feedforward128647
dropout0.2020.1708861
training epochs200200
warmup epochs5032
weight decay00.0000001
batch size132
" + }, + { + "type": "table_caption", + "bbox": [ + 0.115, + 0.385, + 0.49, + 0.412 + ], + "angle": 0, + "content": "Table 5: Hyper-parameters used in training the Transformer" + }, + { + "type": "table", + "bbox": [ + 0.117, + 0.613, + 0.638, + 0.796 + ], + "angle": 0, + "content": "
Romance-phonRomance-orthSinitic
learning rate0.000557390.0009640.000864
num Encoder_layers111
num Decoder_layers111
embedding size1075178
hidden size18513073
dim_feedforward147111136
dropout0.18080.3237940.321639
training epochs181193237
warmup epochs151515
batch size884
" + }, + { + "type": "table_caption", + "bbox": [ + 0.122, + 0.805, + 0.48, + 0.82 + ], + "angle": 0, + "content": "Table 6: Hyper-parameters used in training the RNN" + }, + { + "type": "page_number", + "bbox": [ + 0.491, + 0.929, + 0.511, + 0.941 + ], + "angle": 0, + "content": "35" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.116, + 0.186, + 0.884, + 0.47 + ], + "angle": 0, + "content": "
DatasetModelPED ↓NPED ↓Acc % ↑FER ↓BCFS ↑
Rom-phonRandom daughter (Chang et al., 2022)7.18800.82010%1.13960.3406
CorPaR + Ini (List et al., 2022)2.08850.249114.29%0.08740.6799
SVM + PosStrIni (List et al., 2022)1.90050.227617.05%0.08830.7039
RNN (Meloni et al., 2021)1.45810.181536.68 %0.05920.7435
Transformer (present work)1.25160.157341.38%0.05500.7790
Rom-orthRandom daughter (Chang et al., 2022)6.32720.65420.55%-0.4023
CorPaR + PosStrIni (List et al., 2022)1.83130.200118.89%-0.7227
SVM + PosStr (List et al., 2022)1.69950.186721.66%-0.7454
RNN (Meloni et al., 2021)1.31890.150538.89%-0.7742
Transformer (present work)1.16220.134345.53%-0.7989
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.479, + 0.884, + 0.524 + ], + "angle": 0, + "content": "Table 7: Evaluation of models and baselines with various metrics on Meloni et al. (2021)'s Romance datasets, where all entries from Dinu and Ciobanu (2014) are removed, for 1 run (using the hyperparameters of the best run on the full dataset)" + }, + { + "type": "table", + "bbox": [ + 0.17, + 0.74, + 0.828, + 0.786 + ], + "angle": 0, + "content": "
LatinRomanianFrenchItalianSpanishPortuguese
[kɔlle:ktio:nεm][kolektsie][kɔlɛksjɔ][kolletsoine][kolekθjon][kulisțu]
" + }, + { + "type": "table_caption", + "bbox": [ + 0.14, + 0.796, + 0.856, + 0.811 + ], + "angle": 0, + "content": "Table 8: One cognate set, with Latin as the protoform and all columns to its right as the daughter cognates" + }, + { + "type": "page_number", + "bbox": [ + 0.491, + 0.929, + 0.513, + 0.941 + ], + "angle": 0, + "content": "36" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.133, + 0.085, + 0.435, + 0.1 + ], + "angle": 0, + "content": "ACL 2023 Responsible NLP Checklist" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.108, + 0.323, + 0.123 + ], + "angle": 0, + "content": "A For every submission:" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.127, + 0.533, + 0.158 + ], + "angle": 0, + "content": "A1. Did you describe the limitations of your work? Section 8" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.169, + 0.553, + 0.201 + ], + "angle": 0, + "content": "A2. Did you discuss any potential risks of your work? Section 8" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.213, + 0.696, + 0.244 + ], + "angle": 0, + "content": "A3. Do the abstract and introduction summarize the paper's main claims? Section 1" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.256, + 0.669, + 0.289 + ], + "angle": 0, + "content": "□ A4. Have you used AI writing assistants when working on this paper? Not applicable. Left blank." + }, + { + "type": "list", + "bbox": [ + 0.13, + 0.127, + 0.696, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.3, + 0.489, + 0.317 + ], + "angle": 0, + "content": "B Did you use or create scientific artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.322, + 0.232, + 0.336 + ], + "angle": 0, + "content": "Sections 3, 4" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.347, + 0.53, + 0.379 + ], + "angle": 0, + "content": "B1. Did you cite the creators of artifacts you used? Sections 3,4,5,6" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.39, + 0.779, + 0.422 + ], + "angle": 0, + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Sections 3, 5.1" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.434, + 0.881, + 0.514 + ], + "angle": 0, + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.525, + 0.881, + 0.589 + ], + "angle": 0, + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.599, + 0.881, + 0.647 + ], + "angle": 0, + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Section 3" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.658, + 0.881, + 0.755 + ], + "angle": 0, + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Table 2 and Appendix Section A" + }, + { + "type": "list", + "bbox": [ + 0.13, + 0.347, + 0.881, + 0.755 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.765, + 0.494, + 0.782 + ], + "angle": 0, + "content": "C Did you run computational experiments?" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.787, + 0.207, + 0.801 + ], + "angle": 0, + "content": "Section 4" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.813, + 0.881, + 0.861 + ], + "angle": 0, + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Appendix A" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.869, + 0.878, + 0.893 + ], + "angle": 0, + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + }, + { + "type": "page_number", + "bbox": [ + 0.49, + 0.929, + 0.511, + 0.94 + ], + "angle": 0, + "content": "37" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.13, + 0.084, + 0.88, + 0.116 + ], + "angle": 0, + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.118, + 0.699, + 0.133 + ], + "angle": 0, + "content": "Hyperparameter search: 5.1 Hyperparameter values: Appendix Section B" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.143, + 0.884, + 0.191 + ], + "angle": 0, + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.192, + 0.21, + 0.206 + ], + "angle": 0, + "content": "Table 3" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.218, + 0.884, + 0.266 + ], + "angle": 0, + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.268, + 0.243, + 0.281 + ], + "angle": 0, + "content": "5.1, 6.1, 6.3" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.294, + 0.878, + 0.31 + ], + "angle": 0, + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.315, + 0.215, + 0.33 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.341, + 0.883, + 0.373 + ], + "angle": 0, + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.375, + 0.249, + 0.389 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.4, + 0.884, + 0.448 + ], + "angle": 0, + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.45, + 0.249, + 0.464 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.476, + 0.884, + 0.523 + ], + "angle": 0, + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.525, + 0.25, + 0.54 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.55, + 0.875, + 0.565 + ], + "angle": 0, + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.567, + 0.249, + 0.582 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.593, + 0.881, + 0.624 + ], + "angle": 0, + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.627, + 0.25, + 0.641 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "page_number", + "bbox": [ + 0.491, + 0.929, + 0.511, + 0.941 + ], + "angle": 0, + "content": "38" + } + ] +] \ No newline at end of file diff --git a/2023/Transformed Protoform Reconstruction/244a271d-f768-4ade-bdfe-e0da000e90a0_origin.pdf b/2023/Transformed Protoform Reconstruction/244a271d-f768-4ade-bdfe-e0da000e90a0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..49410b637079b8a85510880b5b6159e5eb9acc7d --- /dev/null +++ b/2023/Transformed Protoform Reconstruction/244a271d-f768-4ade-bdfe-e0da000e90a0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2e86917e438e93ebabe7ebada8435535cd4fe96e80b4921b151148cf09d0b051 +size 667694 diff --git a/2023/Transformed Protoform Reconstruction/full.md b/2023/Transformed Protoform Reconstruction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4264ab470ce281747f86030d35b691dc2289e15e --- /dev/null +++ b/2023/Transformed Protoform Reconstruction/full.md @@ -0,0 +1,287 @@ +# Transformed Protoform Reconstruction + +Young Min Kim* and Kalvin Chang* and Chenxuan Cui and David Mortensen +Language Technologies Institute, Carnegie Mellon University +{youngmik, kalvinc, cxcui, dmortens}@cs.cmu.edu + +# Abstract + +Protoform reconstruction is the task of inferring how morphemes or words sounded in ancestral languages of a set of daughter languages. Meloni et al. (2021) achieved the state-of-the-art on Latin protoform reconstruction with an RNN-based encoder-decoder with attention model. We update their model with the state-of-the-art seq2seq model—the Transformer. Our model outperforms their model on a suite of different metrics on two different datasets: Meloni et al.'s Romance data of $8,000+$ cognates (spanning 5 languages) and a Chinese dataset (Hóu, 2004) of $800+$ cognates (spanning 39 varieties). We also probe our model for potential phylogenetic signal contained in the model. Our code is publicly available1. + +# 1 Introduction + +Languages change over time and sometimes diverge into multiple daughter languages. The common ancestor of a set of genetically related languages is their proto-language. While there are proto-languages such as Latin that are attested, they are the exception2. Reconstructed words and morphemes in proto-languages are called protoforms. The task of reconstructing unattested protolanguages is called protoform reconstruction. + +Historical linguists reconstruct proto-languages by identifying systematic sound changes that can be inferred from correspondences between attested daughter languages (see Table 1). They compare the sounds between a set of cognates, or words with a common ancestor, to develop hypotheses about the types and chronologies of sound changes. + +
‘tooth’‘two’‘ten’
Englishtoothtwotent
Dutchtandtweetient
GermanZahnzweizehnz
PWG*tanþ*twai-*tehun*t
+ +Table 1: Sound correspondences in West Germanic Languages and Proto-West-Germanic (PWG). + +This task is inherently data-constrained, especially for under-documented languages. Such data scarcity makes it a particularly difficult task for contemporary neural network architectures such as the Transformer (Vaswani et al., 2017), which are data hungry. + +The contributions of this paper are as follows: + +- Application of the Transformer architecture to the protoform reconstruction task, achieving state of the art performance, contrary to expectation. +- Expansion of prior digital versions of Hóu (2004)'s Chinese dataset to include a total of 804 cognate sets across 39 modern varieties and Middle Chinese. + +# 2 Related Work + +Applying machine learning to protoform reconstruction is not new. Bouchard-Côté et al. (2013) learn an unsupervised protoform reconstruction model for the large Oceanic language family using Monte Carlo Expectation Maximization (Dempster et al., 1977; Bouchard-Côté et al., 2008), supervising the model with a gold phylogeny and using a probabilistic, generative model of sound change. He et al. (2022) modernize an earlier version of Bouchard-Côté et al. (2013)'s model with RNNs for a 4 language subset of Romance, but they rely on a bigram language model of Latin, making their model technically not unsupervised. + +List et al. (2022) apply an SVM classifier to supervised reconstruction by treating sound correspondences as training examples. Note that there were no word boundaries in the input matrix; that is, all sound correspondences across the training set are flattened into one matrix. Furthermore, each language has an independent phonemic inventory. To learn contextual information, the authors experiment with adding features encoding the position of phonemes, among others. + +Ciobanu and Dinu (2018) learn a conditional random field (Lafferty et al., 2001) using n-gram features for supervised reconstruction and ensemble 5 daughter-to-protoform models. They use a dataset of 3,218 complete cognate sets spanning Latin (the proto-language) and 5 Romance languages: Romanian, French, Italian, Spanish, Portuguese. + +Meloni et al. (2021) employ a GRU-based seq2seq approach (Cho et al., 2014) to Latin protoform reconstruction and achieve state-of-the-art character edit distances. They extend Dinu and Ciobanu (2014)'s Romance data using data from Wiktionary—for a total of 8,799 cognate sets across 5 Romance languages plus Latin—in both orthographic and phonetic (IPA) representations. In their model, all entries comprising the cognate set are concatenated together in a fixed order to form a training example. Chang et al. (2022) applied Meloni et al. (2021)'s architecture to the reconstruction of Middle Chinese on a dataset of $5000+$ cognate sets spanning 8 languages they compiled from Wiktionary. $^3$ + +Fourrier (2022) compares statistical machine translation, RNN, and Transformer architectures for protoform reconstruction, but they evaluate their results using BLEU scores (Papineni et al., 2002) instead of edit distance. They find that their Transformer model did not outperform the RNN models on protoform reconstruction. In addition, their multilingual NMT (neural machine translation) model predicts many languages instead of one target language and is trained on bilingual pairs for protoform reconstruction (e.g. Italian-Latin and Spanish-Latin), unlike comparative reconstruction. In contrast, we encode the entire cognate set consisting of multiple daughter languages (5 for the Romance dataset; 39 for Chinese) and predict the corresponding protoform. + +# 3 Datasets + +We train and test our model on Romance and Sinitic (Chinese) language datasets. For Romance languages, we use Meloni et al. (2021)'s dataset which consists of 8,799 cognate sets of Romanian, French, Italian, Spanish, Portuguese words and the corresponding Latin form (approximately, a protoform). There are two versions of this dataset: phonetic and orthographic. The phonetic dataset (Rom-phon) represents words with IPA symbols whereas the orthographic dataset (Rom-orth) represents words in the orthographic form of each language. We preserved all diacritics, except for vowel length. This dataset is an extension of Dinu and Ciobanu (2014)'s original dataset of 3,218 cognate sets, which is not publicly available. Refer to Table 2 for more information. + +# 3.1 Expanding digital versions of Hóu (2004) + +For Sinitic languages, we created a dataset of Middle Chinese and its modern daughter languages. Middle Chinese is an unattested language, and we thus have to rely on Baxter and Sagart (2014)'s reconstructions of forms corresponding to 4,967 Chinese characters. We scraped Wiktionary to obtain Hóu (2004)'s phonetic representations of their modern reflexes. The resulting dataset contains 804 cognate sets of 39 modern Sinitic languages and the corresponding reconstructed Middle Chinese word. List (2021)'s version previously had 894 cognate sets across 15 varieties. + +# 4 Model + +We propose a Transformer-based encoder-decoder architecture (Vaswani et al., 2017) because such models have produced state-of-the-art results on many sequence processing tasks. Transformers are by reputation data hungry, though, which poses a challenge to our problem setting, where the number of available training examples is often very small. + +![](images/f7b8fb97a67b65f9e747cee35450c06455cec12b94f2336c6353635c555db3c3.jpg) +Figure 1: Diagram of our encoder-decoder architecture. Additive positional encoding and language embedding are applied to each daughter sequence before all daughter sequences are concatenated into a single sequence. + +We modify the standard encoder-decoder architecture to accommodate the structure of our datasets, where multiple daughter sequences correspond to a single protoform sequence. Like Meloni et al. (2021), the daughter sequences are concatenated into a single sequence before being fed into the encoder. Because we only care about the relative position between tokens within each daughter sequence but not across daughter sequences, positional encoding is applied to each individual daughter sequence before concatenation. Along with positional encoding, an additive language embedding is applied to the token embeddings to differentiate between input tokens of different daughter languages. + +# 5 Experiments + +# 5.1 Baselines + +We compare our Transformer model to a variety of baselines. For Meloni et al. (2021), we use Chang et al. (2022)'s PyTorch re-implementation and reran a Bayesian hyperparameter search using WandB (Biewald, 2020) to ensure a more fair comparison (since our model is tuned with WandB as well). We also include the random daughter (randomly designate a daughter form as the protoform and assume no sound change) and the majority constituent baselines (predict the most common phoneme in each syllable constituent) from Chang et al. (2022). For the SVM and CoRPaR classifiers (List et al., 2022), we experiment with different contextual features, such as Pos (position), Str (prosodic structure), and Ini (whether or not the phoneme appears word-initially or wordfinally). + +We publish results on Meloni et al. (2021)'s full set of 8,799 cognates but cannot redistribute this set due to Dinu and Ciobanu (2014)'s restrictions. For reproducibility, we include results on Meloni et al. (2021)'s public subset of 5,419 cognates in the Appendix (Table 7), both of which include vowel length. Observe that these results are worse than those obtained on the full set, suggesting that the RNN and Transformer are dependent on a wealth of training data. + +# 5.2 Preprocessing + +In all our datasets, we merge diacritics to their base segments to form a multi-character token. For instance, the sequence $\left[\mathrm{t},\mathrm{h}\right]$ is concatenated to $\left[\mathrm{th}\right]$ . This ensures that phonemes are treated as one token. For Chinese, tone contours (a sequence of tones) are treated as one token. When multiple pronunciation variants are listed for a single Chinese character, we arbitrarily pick the first one. + +# 6 Results and Discussion + +# 6.1 Evaluation criteria + +We evaluate the predicted protoforms using edit distance (Levenshtein et al., 1966), normalized edit distance (edit distance normalized by the length of the target) and accuracy (the percentage of protoforms that are reconstructed without any mistakes). Like Chang et al. (2022), we also use feature error rate calculated using articulatory feature vectors from PanPhon (Mortensen et al., 2016) because it reflects the phonetic similarity between the prediction and the gold protoform. For datasets with phonetic transcriptions (Romance-phonetic and Chinese), we use phoneme edit distance and normalized phoneme edit distance. As List (2019) suggests, we use B-Cubed F Scores (Amigo et al., 2009) to capture the structural similarity between the gold and predicted protoforms (0: structurally dissimilar, 1: similar). With the exception of character and phoneme edit distance, the metrics enable fair comparison across different language families, which will differ in the average word length. + +# 6.2 Results + +Table 3 shows that our model consistently has the best performance on all datasets with regards to most metrics. The results were averaged across 5 runs. Out of all datasets, our model performs best on the Rom-orth dataset, where we achieve a $7.0\%$ + +
Language FamilySource# varietiesCognate setsProto-language
Rom-phonDinu and Ciobanu (2014), Meloni et al. (2021)58,799Latin
Rom-orthDinu and Ciobanu (2014), Meloni et al. (2021)58,799Latin
Sinitic (Chinese)Hóu (2004)39804Middle Chinese
+ +Table 2: Statistics on both datasets used in our experiments. # varieties refers to the number of daughter varieties. + +![](images/c592223a9b2834446e15eea3269624c5d3a859092d8336de5af115ff8da6ccec.jpg) +Figure 2: A gold phylogeny of Romance (left) compared with those derived by probing the RNN model (middle) and the Transformer model (right) on Rom-phon. + +![](images/5bc091c6f6ab7c6395d4be007dba35a798343972ad3762c35667220fa1c358bc.jpg) + +![](images/eb5a89ca965ce80dc9ef2e39ce40c63b323765447d15d389f4faf61b53fe0f46.jpg) + +decrease in phoneme edit distance and a 1.43p.p improvement in accuracy relative to the RNN baseline. We observe the most dramatic performance difference with the RNN baseline on the Sinitic dataset: a $10.48\%$ decrease in phoneme edit distance and a 5.47p.p increase in accuracy. For reproducibility, results on the publicly available portion of the Rom-phon and Rom-orth datasets are provided in Table 7 in the Appendix. + +# 6.3 Analysis + +We observe that the BCFS is relatively high for the Romance non-neural baselines compared to those of the Chinese ones. This suggests that the sound changes in the Romance datasets are more regular than that of Chinese, which corroborates List et al. (2014)'s results that more than half of the Chinese characters in their dataset could not be explained by a tree model. + +We examine the errors made by the Transformer model on the Rom-phon dataset. Substitutions constitute around $61\%$ of the errors made by the Transformer; deletions, $21\%$ , and insertions, $18\%$ . The highest number of substitution errors occur between $[\mathrm{i}, \mathrm{I}]$ , $[\mathrm{e}, \varepsilon]$ , $[\mathrm{o}, \mathfrak{o}]$ and $[\mathrm{u}, \mathrm{v}]$ —vowel pairs that contrast only in tenseness. This is consistent with the analysis of Meloni et al. (2021), where substitutions between tense-lax vowel pairs take up the largest portion of errors. + +We observe that other common substitution errors also happen between phonemes that share major phonetic features. This demonstrates that al + +though no explicit phonetic information is fed directly into the model, the model makes mistakes motivated by phonetic similarity, like Meloni et al. (2021). + +We do not observe notable differences in the error statistics between the Transformer and the RNN. + +# 6.4 Language relatedness + +Inspired by Fourrier (2022), we probe our model for diachronic information on how genetically related each Romance language is to each other. We create a distance matrix between every pair of languages in a dataset by taking the cosine similarity between a pair's language embeddings. We then use sklearnn (Pedregosa et al., 2011)'s implementation of the Ward variance minimization algorithm (Ward Jr, 1963) to perform hierarchical clustering on the distance matrix. We take a consensus of the dendrograms from 5 different runs using the consense program from PHYLIP (Felsenstein, 2013). + +As we see in Figure 2, the Transformer captures more of the phylogenetic relationships among the languages correctly for the Rom-phon dataset. Indeed, the Generalized Quartet Distance (GQD) (Sand et al., 2013; Pompei et al., 2011; Rama et al., 2018) between the gold and predicted tree, calculated using quartetDist from the tqDist library (Sand et al., 2014), is 0.4 for the Transformer but 0.8 for the RNN. See Figure 5 in the Appendix for the results of the orthographic dataset. + +
DatasetModelPED ↓NPED ↓Acc % ↑FER ↓BCFS ↑
SiniticRandom daughter (Chang et al., 2022)3.77020.84050%0.28930.2748
Majority constituent (Chang et al., 2022)3.50310.78060%0.20130.3695
CorPaR (List et al., 2022)3.27950.72780%0.39720.3332
SVM + PosStr (List et al., 2022)1.68940.369215.52%0.16690.5418
RNN (Meloni et al., 2021)1.06710.242135.65%0.08990.6781
Transformer (present work)0.95530.215041.12%0.08420.7033
Rom-phonRandom daughter (Chang et al., 2022)6.15340.69140.06%0.62640.4016
CorPaR + PosIni (List et al., 2022)1.68470.197822.18%0.07280.7403
SVM + PosStrIni (List et al., 2022)1.57870.186124.69%0.07130.7610
RNN (Meloni et al., 2021)0.96550.122452.31%0.03840.8296
Transformer (present work)0.89260.113753.75%0.03730.8435
Rom-orthRandom daughter (Chang et al., 2022)4.25670.48542.97%-0.5147
CorPaR + Ini (List et al., 2022)0.95310.116047.23%-0.8400
SVM + PosStr (List et al., 2022)0.89880.110550.43%-0.8501
RNN (Meloni et al., 2021)0.59410.077069.80%-0.8916
Transformer (present work)0.55250.072071.23%-0.9002
+ +Table 3: Evaluation of models and baselines using various metrics, averaged across 5 runs (same hyperparameters, different seeds). Because Rom-orth is not in IPA, character edit distance is used instead of PED, and we cannot accurately calculate FER. See Section 6.1 for an explanation of each evaluation metric. See Table 4 for the standard deviation values. + +Since the Romance dataset only includes 5 daughter languages, our results are insufficient to corroborate or contradict Cathcart and Wandl (2020)'s findings: the more accurate the protoforms, the less accurate the phylogeny will be. It is not clear if the model's language embeddings are learning information that reflects shared innovations (sound changes that if shared among a set of daughter languages, would be acceptable justification for grouping them)—the only acceptable criterion for phylogenetic inference in historical linguistics (Campbell, 2013)—or if the model is learning superficial phonetic similarity. + +# 7 Conclusion + +By showing that Transformers can outperform previous architectures in protoform reconstruction despite the inherent data scarcity of the task, our work motivates future research in this area to take full advantage of the recent advancements in the Transformer space. + +Accurate supervised reconstruction can help pre + +dict protoforms for cognate sets where linguists have not reconstructed one yet. Future work could reconstruct proto-languages whose linguist reconstructions are not available, by transferring knowledge learned from languages with already reconstructed protoforms. Furthermore, future work can leverage the abundance of work in unsupervised NMT to adapt our Transformer model for the unsupervised setting, a more realistic scenario for the historical linguist. + +# Limitations + +One limitation of our work is that the RNN (Meloni et al., 2021) actually outperforms our Transformer on the Chinese dataset in Chang et al. (2022). In addition, as with other neural approaches, our model requires significant amounts of data, which is often not available to historical linguists researching less well-studied language families based on field reports. Romance and Chinese have relatively many cognate sets because the protoforms + +are documented5, but a low resource setup with 200 cognate sets would not fare well on our data-hungrier Transformer model. Furthermore, concatenating the entire cognate set may not work on language families with hundreds of languages such as Oceanic because the input sequence would be too long compared to the output protoform sequence. + +Finally, we obtain our Chinese gold protoforms from Baxter and Sagart (2014)'s Middle Chinese reconstruction, which was actually a transcription of the Qieyun, a rhyme dictionary. Norman and Coblin (1995) disagree with relying on such a philological source and prefer comparative reconstructions that begin from daughter data. However, there is no available comparative reconstruction of Middle Chinese with protoforms corresponding to thousands of characters to use as a gold standard. Be that as it may, it seems clear that Middle Chinese as recorded in the Qieyun is not identical to the most recent ancestor of the Chinese languages. Its preface concedes that it is a compromise between Tang Dynasty dialects. The situation with Romance is, in some ways, comparable. Classical Latin—the variety on which we train—is not the direct ancestor of modern Romance languages. Instead, they are descended from Vulgar Latin or Proto-Romance, which is not well-attested and is primarily through graffiti and other informal inscriptions. Proto-Romance reconstructions are also not exhaustive. As a result, it is difficult to find a dataset like Meloni et al. (2021) with thousands of such ancestor forms. We are also limited to the faithfulness of espeak-ng's Latin G2P, from which Meloni et al. (2021) obtain their phonetic Romance dataset. + +For most language families, protoforms are not attested. In fact, as the term is often used, protoform refers to a form that is inferred only through linguists' comparative method. We adopt the other usage for simplicity. In practice, our approach would require reconstructions made by a linguist to serve as training labels for cognate sets. + +# Acknowledgements + +We would like to thank Liang (Leon) Lu for finding a bug in our implementation, Ying Chen for writing the code for the baselines, and Brendon Boldt and Graham Neubig for providing useful feedback + +for the first iteration of our paper. + +# References + +Enrique Amigo, Julio Gonzalo, Javier Artiles, and Felisa Verdejo. 2009. A comparison of extrinsic clustering evaluation metrics based on formal constraints. Information retrieval, 12(4):461-486. +William H Baxter and Laurent Sagart. 2014. Old Chinese: A new reconstruction. Oxford University Press. +Lukas Biewald. 2020. Experiment tracking with weights and biases. Software available from wandb.com. +Alexandre Bouchard-Côté, Dan Klein, and Michael Jordan. 2008. Efficient inference in phylogenetic indel trees. In Advances in Neural Information Processing Systems, volume 21. Curran Associates, Inc. +Alexandre Bouchard-Côté, David Hall, Thomas L. Griffiths, and Dan Klein. 2013. Automated reconstruction of ancient languages using probabilistic models of sound change. Proceedings of the National Academy of Sciences, 110(11):4224-4229. +Lyle Campbell. 2013. *Historical Linguistics: an Introduction*. Edinburgh University Press. +Chundra Cathcart and Florian Wandl. 2020. In search of isoglosses: continuous and discrete language embeddings in Slavic historical phonology. In Proceedings of the 17th SIGMORPHON Workshop on Computational Research in Phonetics, Phonology, and Morphology, pages 233-244, Online. Association for Computational Linguistics. +Kalvin Chang, Chenxuan Cui, Youngmin Kim, and David R. Mortensen. 2022. WikiHan: A new comparative dataset for Chinese languages. In Proceedings of the 29th International Conference on Computational Linguistics (COLING 2022). +Kyunghyun Cho, Bart van Merrienboer, Dzmitry Bah-danau, and Yoshua Bengio. 2014. On the properties of neural machine translation: Encoder-decoder approaches. In Proceedings of SSST-8, Eighth Workshop on Syntax, Semantics and Structure in Statistical Translation, pages 103-111, Doha, Qatar. Association for Computational Linguistics. +Alina Maria Ciobanu and Liviu P. Dinu. 2018. Ab initio: Automatic Latin proto-word reconstruction. In Proceedings of the 27th International Conference on Computational Linguistics, pages 1604-1614, Santa Fe, New Mexico, USA. Association for Computational Linguistics. +Arthur P Dempster, Nan M Laird, and Donald B Rubin. 1977. Maximum likelihood from incomplete data via the em algorithm. Journal of the Royal Statistical Society: Series B (Methodological), 39(1):1-22. + +Liviu Dinu and Alina Maria Ciobanu. 2014. Building a dataset of multilingual cognates for the Romanian lexicon. In Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC'14), pages 1038-1043, Reykjavik, Iceland. European Language Resources Association (ELRA). +Joseph Felsenstein. 2013. Phylip (phylogeny inference package), version 3.695. Department of Genome Sciences, University of Washington, Seattle. +Clémentine Fourrier. 2022. Neural Approaches to Historical Word Reconstruction. Ph.D. thesis, Université PSL (Paris Sciences & Lettres). +Andre He, Nicholas Tomlin, and Dan Klein. 2022. Neural unsupervised reconstruction of protolanguage word forms. arXiv preprint arXiv:2211.08684. +侯精一 Jingyi Hóu, editor. 2004. Xiandai Hanyu fangyan yinku 现代汉语方言音库 [Phonological database of Chinese dialects]. Shanghai Jiaoyu 上海教育, Shanghai 上海. +John D. Lafferty, Andrew McCallum, and Fernando C. N. Pereira. 2001. Conditional random fields: Probabilistic models for segmenting and labeling sequence data. In Proceedings of the Eighteenth International Conference on Machine Learning, ICML '01, page 282-289, San Francisco, CA, USA. Morgan Kaufmann Publishers Inc. +Vladimir I Levenshtein et al. 1966. Binary codes capable of correcting deletions, insertions, and reversals. Soviet physics doklady, 10(8):707-710. +Johann-Mattis List. 2019. Beyond edit distances: Comparing linguistic reconstruction systems. Theoretical Linguistics, 45(3-4):247-258. +Johann-Mattis List. 2021. CLDF dataset derived from Hóu's "Phonological Database of Chinese Dialects" from 2004. Zenodo. +Johann-Mattis List, Robert Forkel, and Nathan Hill. 2022. A new framework for fast automated phonological reconstruction using trimmed alignments and sound correspondence patterns. In Proceedings of the 3rd Workshop on Computational Approaches to Historical Language Change, pages 89-96, Dublin, Ireland. Association for Computational Linguistics. +Johann-Mattis List, Nelson-Sathi Shijulal, William Martin, and Hans Geisler. 2014. Using phylogenetic networks to model chinese dialect history. Language Dynamics and Change, 4(2):222-252. +Carlo Meloni, Shauli Ravfogel, and Yoav Goldberg. 2021. Ab antiquo: Neural proto-language reconstruction. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4460-4473, Online. Association for Computational Linguistics. + +David R. Mortensen, Patrick Littell, Akash Bharadwaj, Kartik Goyal, Chris Dyer, and Lori S. Levin. 2016. Panphon: A resource for mapping IPA segments to articulatory feature vectors. In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 3475-3484. +Jerry L. Norman and W. South Coblin. 1995. A new approach to Chinese historical linguistics. Journal of the American Oriental Society, 115(4):576-584. +Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318. +F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. 2011. Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12:2825-2830. +Simone Pompei, Vittorio Loreto, and Francesca Tria. 2011. On the accuracy of language trees. PloS one, 6(6):e20109. +Taraka Rama, Johann-Mattis List, Johannes Wahle, and Gerhard Jäger. 2018. Are automatic methods for cognate detection good enough for phylogenetic reconstruction in historical linguistics? arXiv preprint arXiv:1804.05416. +Andreas Sand, Morten K Holt, Jens Johansen, Rolf Fagerberg, Gerth Stolting Brodal, Christian NS Pedersen, and Thomas Mailund. 2013. Algorithms for computing the triplet and quartet distances for binary general trees. *Biology*, 2(4):1189–1209. +Andreas Sand, Morten Kragelund Holt, Jens Johansen, Rolf Fagerberg, Gerth Stolting Brodal, Thomas Mailund, and Christian N. S. Pedersen. 2014. tqdist: A library for computing the quartet and triplet distances between binary or general trees. BMC Bioinformatics, yy(xx):ii-jj. +Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30. +Joe H Ward Jr. 1963. Hierarchical grouping to optimize an objective function. Journal of the American statistical association, 58(301):236-244. + +# A Training + +We split $70\%$ , $10\%$ , and $20\%$ of our dataset into train, validation, and test sets, respectively. We conduct hyperparameter searches using WandB (Biewald, 2020) and use early stopping, picking the epoch with lowest edit distance on validation data. All experiments are performed on a Ubuntu server with 4 GPUs and 20 CPUs. For both the RNN and the Transformer, Meloni et al. (2021)'s dataset takes less than 7 GPU hours to run, while Hóu (2004) takes less than 1 GPU hour. For the large Romance orthographic dataset, the RNN model has around 480,000 parameters, while the Transformer has around 800,000 parameters. + +# B Hyper-parameters + +Refer to Table 5 and Table 6 for the best hyperparameters we found during hyperparameter search via WandB. + +# C Supplementary Results + +In order to compare our model to earlier work, we used the Rom-phon and Rom-orth datasets from Meloni et al. (2021). However, this set includes a subset from Ciobanu and Dinu (2018) which is not freely redistributable. So that our results can be reproduced, we also computed them on the publicly available subset of Meloni et al. (2021)'s dataset, which is presented in Table 7. + +Phylogenetic trees for Chinese were also extracted from the RNN and Transformer models. These are shown in Figures 3 and 4. + +We also plot the dendrograms derived from the Rom-ortho dataset in Figure 5. + +![](images/5248f64ba74abe05eb22a117ff6b520cd8f1b60dec47d4820d7c71d6254699a2.jpg) +Figure 3: Consensus tree of the dendrograms from the 5 runs of the Transformer for the Chinese dataset + +![](images/8dcd73a8ee4ee6b081a940c4a2652505f191cb10ce1052f0a7021a0b8dcc613c.jpg) +Figure 4: Consensus tree of the dendrograms from the 5 runs of the RNN for the Chinese dataset + +![](images/4eae6397ed120cb2ac347a487c7648b8b1e88f7345c1f3ca11567d35a24f1b24.jpg) +Figure 5: A gold phylogeny of Romance (left) compared with those derived by probing the RNN model (middle) and the Transformer model (right) on Rom-ortho. GQD is 0.4 for both models. + +![](images/06d23af2c68209392b2831f214ca29d7fd3ecd61bfb0d986f2b4e4c4e9b31e68.jpg) + +![](images/c9ca5665eeebeb9fad4556f807fdefcdbc81b6e0eb60ee9cf4c64b6a02809d6c.jpg) + +
DatasetModelPED ↓NPED ↓Acc % ↑FER ↓BCFS ↑
SiniticRandom daughter3.77020.84050%0.28930.2748
Majority constituent3.50310.78060%0.20130.3695
CorPaR3.27950.72780%0.39720.3332
SVM +PosStr1.68940.369215.52%0.16690.5418
RNN1.0671 ± 0.06190.2421 ± 0.014035.65% ± 1.60%0.0899 ± 0.00480.6781 ± 0.0174±
Transformer (present work)0.9553 ± 0.03920.2150 ± 0.007541.12% ± 2.3%0.0842 ± 0.00700.7033 ± 0.0087±
Rom-phonRandom daughter6.15340.69140.06%0.62640.4016
CorPaR +PosIni1.68470.197822.18%0.07280.7403
SVM +PosStrIni1.57870.186124.69%0.07130.7610
RNN0.9655 ± 0.01890.1224 ± 0.002252.31% ± 0.63%0.0384 ± 0.00110.8296 ± 0.0029±
Transformer (present work)0.8926 ± 0.01660.1137 ± 0.001753.75% ± 0.40%0.0373 ± 0.00090.8435 ± 0.0026±
Rom-orthRandom daughter4.25670.48542.97%-0.5147
CorPaR +Ini0.95310.116047.23%-0.8400
SVM +PosStr0.89880.110550.43%-0.8501
RNN0.5941 ± 0.01000.0770 ± 0.001569.80% ±0.22%-0.8916 ± 0.0019±
Transformer (present work)0.5525 ± 0.01040.0720 ± 0.001771.23% ± 0.52%-0.9002 ± 0.0017±
+ +Table 4: Evaluation of models and baselines using various metrics, averaged across 5 runs (same hyperparameters, different seeds), with standard deviations. Because Rom-orth is not in IPA, character edit distance is used instead of PED, and we cannot accurately calculate FER. See Section 6.1 for an explanation of each evaluation metric. + +
Romance (phon & orth)Sinitic
learning rate0.000130.0007487
num Encoder_layers32
num Decoder_layers35
embedding size128128
n_head88
dim_feedforward128647
dropout0.2020.1708861
training epochs200200
warmup epochs5032
weight decay00.0000001
batch size132
+ +Table 5: Hyper-parameters used in training the Transformer + +
Romance-phonRomance-orthSinitic
learning rate0.000557390.0009640.000864
num Encoder_layers111
num Decoder_layers111
embedding size1075178
hidden size18513073
dim_feedforward147111136
dropout0.18080.3237940.321639
training epochs181193237
warmup epochs151515
batch size884
+ +Table 6: Hyper-parameters used in training the RNN + +
DatasetModelPED ↓NPED ↓Acc % ↑FER ↓BCFS ↑
Rom-phonRandom daughter (Chang et al., 2022)7.18800.82010%1.13960.3406
CorPaR + Ini (List et al., 2022)2.08850.249114.29%0.08740.6799
SVM + PosStrIni (List et al., 2022)1.90050.227617.05%0.08830.7039
RNN (Meloni et al., 2021)1.45810.181536.68 %0.05920.7435
Transformer (present work)1.25160.157341.38%0.05500.7790
Rom-orthRandom daughter (Chang et al., 2022)6.32720.65420.55%-0.4023
CorPaR + PosStrIni (List et al., 2022)1.83130.200118.89%-0.7227
SVM + PosStr (List et al., 2022)1.69950.186721.66%-0.7454
RNN (Meloni et al., 2021)1.31890.150538.89%-0.7742
Transformer (present work)1.16220.134345.53%-0.7989
+ +Table 7: Evaluation of models and baselines with various metrics on Meloni et al. (2021)'s Romance datasets, where all entries from Dinu and Ciobanu (2014) are removed, for 1 run (using the hyperparameters of the best run on the full dataset) + +
LatinRomanianFrenchItalianSpanishPortuguese
[kɔlle:ktio:nεm][kolektsie][kɔlɛksjɔ][kolletsoine][kolekθjon][kulisțu]
+ +Table 8: One cognate set, with Latin as the protoform and all columns to its right as the daughter cognates + +A For every submission: + +A1. Did you describe the limitations of your work? Section 8 +A2. Did you discuss any potential risks of your work? Section 8 +A3. Do the abstract and introduction summarize the paper's main claims? Section 1 +□ A4. Have you used AI writing assistants when working on this paper? Not applicable. Left blank. + +B Did you use or create scientific artifacts? + +Sections 3, 4 + +B1. Did you cite the creators of artifacts you used? Sections 3,4,5,6 +B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Sections 3, 5.1 +B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Not applicable. Left blank. +B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Not applicable. Left blank. +B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Section 3 +B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Table 2 and Appendix Section A + +C Did you run computational experiments? + +Section 4 + +C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Appendix A + +The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance. + +C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? + +Hyperparameter search: 5.1 Hyperparameter values: Appendix Section B + +C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? + +Table 3 + +C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? + +5.1, 6.1, 6.3 + +D Did you use human annotators (e.g., crowdworkers) or research with human participants? + +Left blank. + +D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? + +No response. + +D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? + +No response. + +D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? + +No response. + +D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? + +No response. + +D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? + +No response. \ No newline at end of file diff --git a/2023/Transformed Protoform Reconstruction/images.zip b/2023/Transformed Protoform Reconstruction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..80d8ee69e20af20b5ea301886ed2ddbcae66f757 --- /dev/null +++ b/2023/Transformed Protoform Reconstruction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:54f94b88e7527dba4c4d7cadaa519815b57ff4100cfcd99962c79fe4b474c613 +size 707856 diff --git a/2023/Transformed Protoform Reconstruction/layout.json b/2023/Transformed Protoform Reconstruction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..62b64008487a84d275fb04550daff7ccdee5d218 --- /dev/null +++ b/2023/Transformed Protoform Reconstruction/layout.json @@ -0,0 +1,7097 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 173, + 76, + 421, + 92 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 76, + 421, + 92 + ], + "spans": [ + { + "bbox": [ + 173, + 76, + 421, + 92 + ], + "type": "text", + "content": "Transformed Protoform Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 95, + 119, + 502, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 119, + 502, + 163 + ], + "spans": [ + { + "bbox": [ + 95, + 119, + 502, + 163 + ], + "type": "text", + "content": "Young Min Kim* and Kalvin Chang* and Chenxuan Cui and David Mortensen \nLanguage Technologies Institute, Carnegie Mellon University \n{youngmik, kalvinc, cxcui, dmortens}@cs.cmu.edu" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 155, + 212, + 203, + 226 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 212, + 203, + 226 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 203, + 226 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 84, + 232, + 274, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 232, + 274, + 435 + ], + "spans": [ + { + "bbox": [ + 84, + 232, + 274, + 435 + ], + "type": "text", + "content": "Protoform reconstruction is the task of inferring how morphemes or words sounded in ancestral languages of a set of daughter languages. Meloni et al. (2021) achieved the state-of-the-art on Latin protoform reconstruction with an RNN-based encoder-decoder with attention model. We update their model with the state-of-the-art seq2seq model—the Transformer. Our model outperforms their model on a suite of different metrics on two different datasets: Meloni et al.'s Romance data of " + }, + { + "bbox": [ + 84, + 232, + 274, + 435 + ], + "type": "inline_equation", + "content": "8,000+" + }, + { + "bbox": [ + 84, + 232, + 274, + 435 + ], + "type": "text", + "content": " cognates (spanning 5 languages) and a Chinese dataset (Hóu, 2004) of " + }, + { + "bbox": [ + 84, + 232, + 274, + 435 + ], + "type": "inline_equation", + "content": "800+" + }, + { + "bbox": [ + 84, + 232, + 274, + 435 + ], + "type": "text", + "content": " cognates (spanning 39 varieties). We also probe our model for potential phylogenetic signal contained in the model. Our code is publicly available1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 444, + 155, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 444, + 155, + 456 + ], + "spans": [ + { + "bbox": [ + 68, + 444, + 155, + 456 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 465, + 292, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 465, + 292, + 587 + ], + "spans": [ + { + "bbox": [ + 67, + 465, + 292, + 587 + ], + "type": "text", + "content": "Languages change over time and sometimes diverge into multiple daughter languages. The common ancestor of a set of genetically related languages is their proto-language. While there are proto-languages such as Latin that are attested, they are the exception2. Reconstructed words and morphemes in proto-languages are called protoforms. The task of reconstructing unattested protolanguages is called protoform reconstruction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 587, + 291, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 587, + 291, + 695 + ], + "spans": [ + { + "bbox": [ + 67, + 587, + 291, + 695 + ], + "type": "text", + "content": "Historical linguists reconstruct proto-languages by identifying systematic sound changes that can be inferred from correspondences between attested daughter languages (see Table 1). They compare the sounds between a set of cognates, or words with a common ancestor, to develop hypotheses about the types and chronologies of sound changes." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 314, + 210, + 516, + 282 + ], + "blocks": [ + { + "bbox": [ + 314, + 210, + 516, + 282 + ], + "lines": [ + { + "bbox": [ + 314, + 210, + 516, + 282 + ], + "spans": [ + { + "bbox": [ + 314, + 210, + 516, + 282 + ], + "type": "table", + "html": "
‘tooth’‘two’‘ten’
Englishtoothtwotent
Dutchtandtweetient
GermanZahnzweizehnz
PWG*tanþ*twai-*tehun*t
", + "image_path": "6fdf9d698a1840ce766ff34c1ee4d7151dfd84876cddb589635e6829b2f3e63d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 290, + 527, + 316 + ], + "lines": [ + { + "bbox": [ + 302, + 290, + 527, + 316 + ], + "spans": [ + { + "bbox": [ + 302, + 290, + 527, + 316 + ], + "type": "text", + "content": "Table 1: Sound correspondences in West Germanic Languages and Proto-West-Germanic (PWG)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 302, + 337, + 526, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 337, + 526, + 417 + ], + "spans": [ + { + "bbox": [ + 302, + 337, + 526, + 417 + ], + "type": "text", + "content": "This task is inherently data-constrained, especially for under-documented languages. Such data scarcity makes it a particularly difficult task for contemporary neural network architectures such as the Transformer (Vaswani et al., 2017), which are data hungry." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 419, + 518, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 419, + 518, + 433 + ], + "spans": [ + { + "bbox": [ + 314, + 419, + 518, + 433 + ], + "type": "text", + "content": "The contributions of this paper are as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 316, + 444, + 526, + 562 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 316, + 444, + 525, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 525, + 497 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 525, + 497 + ], + "type": "text", + "content": "- Application of the Transformer architecture to the protoform reconstruction task, achieving state of the art performance, contrary to expectation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 508, + 526, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 508, + 526, + 562 + ], + "spans": [ + { + "bbox": [ + 316, + 508, + 526, + 562 + ], + "type": "text", + "content": "- Expansion of prior digital versions of Hóu (2004)'s Chinese dataset to include a total of 804 cognate sets across 39 modern varieties and Middle Chinese." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 302, + 575, + 396, + 587 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 575, + 396, + 587 + ], + "spans": [ + { + "bbox": [ + 302, + 575, + 396, + 587 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 301, + 597, + 526, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 597, + 526, + 773 + ], + "spans": [ + { + "bbox": [ + 301, + 597, + 526, + 773 + ], + "type": "text", + "content": "Applying machine learning to protoform reconstruction is not new. Bouchard-Côté et al. (2013) learn an unsupervised protoform reconstruction model for the large Oceanic language family using Monte Carlo Expectation Maximization (Dempster et al., 1977; Bouchard-Côté et al., 2008), supervising the model with a gold phylogeny and using a probabilistic, generative model of sound change. He et al. (2022) modernize an earlier version of Bouchard-Côté et al. (2013)'s model with RNNs for a 4 language subset of Romance, but they rely on a bigram language model of Latin, making their model technically not unsupervised." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 83, + 699, + 160, + 711 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 699, + 160, + 711 + ], + "spans": [ + { + "bbox": [ + 83, + 699, + 160, + 711 + ], + "type": "text", + "content": "* Equal contribution" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 82, + 712, + 258, + 721 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 712, + 258, + 721 + ], + "spans": [ + { + "bbox": [ + 82, + 712, + 258, + 721 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 69, + 721, + 290, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 721, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 721, + 290, + 772 + ], + "type": "text", + "content": "2 In fact, the proto-language from which Romance languages like Spanish and Italian are descended is not identical to Classical Latin but is, rather, a closely related and sparsely attested language sometimes called Proto-Romance or Vulgar Latin." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 291, + 781, + 304, + 790 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 781, + 304, + 790 + ], + "spans": [ + { + "bbox": [ + 291, + 781, + 304, + 790 + ], + "type": "text", + "content": "24" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "spans": [ + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "type": "text", + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 229, + 806, + 365, + 817 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 806, + 365, + 817 + ], + "spans": [ + { + "bbox": [ + 229, + 806, + 365, + 817 + ], + "type": "text", + "content": "Volume 2: Short Papers, pages 24-38" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "spans": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "text", + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 291, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 291, + 205 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 291, + 205 + ], + "type": "text", + "content": "List et al. (2022) apply an SVM classifier to supervised reconstruction by treating sound correspondences as training examples. Note that there were no word boundaries in the input matrix; that is, all sound correspondences across the training set are flattened into one matrix. Furthermore, each language has an independent phonemic inventory. To learn contextual information, the authors experiment with adding features encoding the position of phonemes, among others." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 206, + 291, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 206, + 291, + 313 + ], + "spans": [ + { + "bbox": [ + 67, + 206, + 291, + 313 + ], + "type": "text", + "content": "Ciobanu and Dinu (2018) learn a conditional random field (Lafferty et al., 2001) using n-gram features for supervised reconstruction and ensemble 5 daughter-to-protoform models. They use a dataset of 3,218 complete cognate sets spanning Latin (the proto-language) and 5 Romance languages: Romanian, French, Italian, Spanish, Portuguese." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 315, + 291, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 315, + 291, + 518 + ], + "spans": [ + { + "bbox": [ + 69, + 315, + 291, + 518 + ], + "type": "text", + "content": "Meloni et al. (2021) employ a GRU-based seq2seq approach (Cho et al., 2014) to Latin protoform reconstruction and achieve state-of-the-art character edit distances. They extend Dinu and Ciobanu (2014)'s Romance data using data from Wiktionary—for a total of 8,799 cognate sets across 5 Romance languages plus Latin—in both orthographic and phonetic (IPA) representations. In their model, all entries comprising the cognate set are concatenated together in a fixed order to form a training example. Chang et al. (2022) applied Meloni et al. (2021)'s architecture to the reconstruction of Middle Chinese on a dataset of " + }, + { + "bbox": [ + 69, + 315, + 291, + 518 + ], + "type": "inline_equation", + "content": "5000+" + }, + { + "bbox": [ + 69, + 315, + 291, + 518 + ], + "type": "text", + "content": " cognate sets spanning 8 languages they compiled from Wiktionary." + }, + { + "bbox": [ + 69, + 315, + 291, + 518 + ], + "type": "inline_equation", + "content": "^3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 518, + 291, + 734 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 518, + 291, + 734 + ], + "spans": [ + { + "bbox": [ + 67, + 518, + 291, + 734 + ], + "type": "text", + "content": "Fourrier (2022) compares statistical machine translation, RNN, and Transformer architectures for protoform reconstruction, but they evaluate their results using BLEU scores (Papineni et al., 2002) instead of edit distance. They find that their Transformer model did not outperform the RNN models on protoform reconstruction. In addition, their multilingual NMT (neural machine translation) model predicts many languages instead of one target language and is trained on bilingual pairs for protoform reconstruction (e.g. Italian-Latin and Spanish-Latin), unlike comparative reconstruction. In contrast, we encode the entire cognate set consisting of multiple daughter languages (5 for the Romance dataset; 39 for Chinese) and predict the corresponding protoform." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 303, + 71, + 368, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 71, + 368, + 83 + ], + "spans": [ + { + "bbox": [ + 303, + 71, + 368, + 83 + ], + "type": "text", + "content": "3 Datasets" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 302, + 104, + 526, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 104, + 526, + 320 + ], + "spans": [ + { + "bbox": [ + 302, + 104, + 526, + 320 + ], + "type": "text", + "content": "We train and test our model on Romance and Sinitic (Chinese) language datasets. For Romance languages, we use Meloni et al. (2021)'s dataset which consists of 8,799 cognate sets of Romanian, French, Italian, Spanish, Portuguese words and the corresponding Latin form (approximately, a protoform). There are two versions of this dataset: phonetic and orthographic. The phonetic dataset (Rom-phon) represents words with IPA symbols whereas the orthographic dataset (Rom-orth) represents words in the orthographic form of each language. We preserved all diacritics, except for vowel length. This dataset is an extension of Dinu and Ciobanu (2014)'s original dataset of 3,218 cognate sets, which is not publicly available. Refer to Table 2 for more information." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 347, + 522, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 347, + 522, + 361 + ], + "spans": [ + { + "bbox": [ + 302, + 347, + 522, + 361 + ], + "type": "text", + "content": "3.1 Expanding digital versions of Hóu (2004)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 375, + 526, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 375, + 526, + 537 + ], + "spans": [ + { + "bbox": [ + 302, + 375, + 526, + 537 + ], + "type": "text", + "content": "For Sinitic languages, we created a dataset of Middle Chinese and its modern daughter languages. Middle Chinese is an unattested language, and we thus have to rely on Baxter and Sagart (2014)'s reconstructions of forms corresponding to 4,967 Chinese characters. We scraped Wiktionary to obtain Hóu (2004)'s phonetic representations of their modern reflexes. The resulting dataset contains 804 cognate sets of 39 modern Sinitic languages and the corresponding reconstructed Middle Chinese word. List (2021)'s version previously had 894 cognate sets across 15 varieties." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 564, + 358, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 564, + 358, + 577 + ], + "spans": [ + { + "bbox": [ + 302, + 564, + 358, + 577 + ], + "type": "text", + "content": "4 Model" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 598, + 526, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 598, + 526, + 705 + ], + "spans": [ + { + "bbox": [ + 302, + 598, + 526, + 705 + ], + "type": "text", + "content": "We propose a Transformer-based encoder-decoder architecture (Vaswani et al., 2017) because such models have produced state-of-the-art results on many sequence processing tasks. Transformers are by reputation data hungry, though, which poses a challenge to our problem setting, where the number of available training examples is often very small." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 740, + 290, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 740, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 740, + 290, + 772 + ], + "type": "text", + "content": "3The original dataset contains 21,000 cognate sets, but only " + }, + { + "bbox": [ + 67, + 740, + 290, + 772 + ], + "type": "inline_equation", + "content": "5000+" + }, + { + "bbox": [ + 67, + 740, + 290, + 772 + ], + "type": "text", + "content": " had at least 3 daughter entries and were used as input to the model." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 731, + 525, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 731, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 731, + 525, + 772 + ], + "type": "text", + "content": "4https://en.wiktionary.org/wiki/Module: zh/data/dial-pron/documentation originally had 1,023 characters, but only 804 had reconstructions from Baxter and Sagart (2014)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 291, + 781, + 304, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 781, + 304, + 791 + ], + "spans": [ + { + "bbox": [ + 291, + 781, + 304, + 791 + ], + "type": "text", + "content": "25" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 91, + 71, + 267, + 223 + ], + "blocks": [ + { + "bbox": [ + 91, + 71, + 267, + 223 + ], + "lines": [ + { + "bbox": [ + 91, + 71, + 267, + 223 + ], + "spans": [ + { + "bbox": [ + 91, + 71, + 267, + 223 + ], + "type": "image", + "image_path": "f7b8fb97a67b65f9e747cee35450c06455cec12b94f2336c6353635c555db3c3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 231, + 291, + 280 + ], + "lines": [ + { + "bbox": [ + 67, + 231, + 291, + 280 + ], + "spans": [ + { + "bbox": [ + 67, + 231, + 291, + 280 + ], + "type": "text", + "content": "Figure 1: Diagram of our encoder-decoder architecture. Additive positional encoding and language embedding are applied to each daughter sequence before all daughter sequences are concatenated into a single sequence." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 298, + 291, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 298, + 291, + 502 + ], + "spans": [ + { + "bbox": [ + 67, + 298, + 291, + 502 + ], + "type": "text", + "content": "We modify the standard encoder-decoder architecture to accommodate the structure of our datasets, where multiple daughter sequences correspond to a single protoform sequence. Like Meloni et al. (2021), the daughter sequences are concatenated into a single sequence before being fed into the encoder. Because we only care about the relative position between tokens within each daughter sequence but not across daughter sequences, positional encoding is applied to each individual daughter sequence before concatenation. Along with positional encoding, an additive language embedding is applied to the token embeddings to differentiate between input tokens of different daughter languages." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 514, + 154, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 514, + 154, + 528 + ], + "spans": [ + { + "bbox": [ + 68, + 514, + 154, + 528 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 538, + 139, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 538, + 139, + 550 + ], + "spans": [ + { + "bbox": [ + 68, + 538, + 139, + 550 + ], + "type": "text", + "content": "5.1 Baselines" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 556, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 556, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 556, + 291, + 773 + ], + "type": "text", + "content": "We compare our Transformer model to a variety of baselines. For Meloni et al. (2021), we use Chang et al. (2022)'s PyTorch re-implementation and reran a Bayesian hyperparameter search using WandB (Biewald, 2020) to ensure a more fair comparison (since our model is tuned with WandB as well). We also include the random daughter (randomly designate a daughter form as the protoform and assume no sound change) and the majority constituent baselines (predict the most common phoneme in each syllable constituent) from Chang et al. (2022). For the SVM and CoRPaR classifiers (List et al., 2022), we experiment with different contextual features, such as Pos (position), Str (prosodic structure), and Ini (whether or not the phoneme appears word-initially or wordfinally)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 71, + 526, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 206 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 206 + ], + "type": "text", + "content": "We publish results on Meloni et al. (2021)'s full set of 8,799 cognates but cannot redistribute this set due to Dinu and Ciobanu (2014)'s restrictions. For reproducibility, we include results on Meloni et al. (2021)'s public subset of 5,419 cognates in the Appendix (Table 7), both of which include vowel length. Observe that these results are worse than those obtained on the full set, suggesting that the RNN and Transformer are dependent on a wealth of training data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 216, + 396, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 216, + 396, + 229 + ], + "spans": [ + { + "bbox": [ + 302, + 216, + 396, + 229 + ], + "type": "text", + "content": "5.2 Preprocessing" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 234, + 526, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 234, + 526, + 343 + ], + "spans": [ + { + "bbox": [ + 302, + 234, + 526, + 343 + ], + "type": "text", + "content": "In all our datasets, we merge diacritics to their base segments to form a multi-character token. For instance, the sequence " + }, + { + "bbox": [ + 302, + 234, + 526, + 343 + ], + "type": "inline_equation", + "content": "\\left[\\mathrm{t},\\mathrm{h}\\right]" + }, + { + "bbox": [ + 302, + 234, + 526, + 343 + ], + "type": "text", + "content": " is concatenated to " + }, + { + "bbox": [ + 302, + 234, + 526, + 343 + ], + "type": "inline_equation", + "content": "\\left[\\mathrm{th}\\right]" + }, + { + "bbox": [ + 302, + 234, + 526, + 343 + ], + "type": "text", + "content": ". This ensures that phonemes are treated as one token. For Chinese, tone contours (a sequence of tones) are treated as one token. When multiple pronunciation variants are listed for a single Chinese character, we arbitrarily pick the first one." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 354, + 441, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 354, + 441, + 365 + ], + "spans": [ + { + "bbox": [ + 302, + 354, + 441, + 365 + ], + "type": "text", + "content": "6 Results and Discussion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 375, + 420, + 386 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 375, + 420, + 386 + ], + "spans": [ + { + "bbox": [ + 302, + 375, + 420, + 386 + ], + "type": "text", + "content": "6.1 Evaluation criteria" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 393, + 526, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 393, + 526, + 677 + ], + "spans": [ + { + "bbox": [ + 302, + 393, + 526, + 677 + ], + "type": "text", + "content": "We evaluate the predicted protoforms using edit distance (Levenshtein et al., 1966), normalized edit distance (edit distance normalized by the length of the target) and accuracy (the percentage of protoforms that are reconstructed without any mistakes). Like Chang et al. (2022), we also use feature error rate calculated using articulatory feature vectors from PanPhon (Mortensen et al., 2016) because it reflects the phonetic similarity between the prediction and the gold protoform. For datasets with phonetic transcriptions (Romance-phonetic and Chinese), we use phoneme edit distance and normalized phoneme edit distance. As List (2019) suggests, we use B-Cubed F Scores (Amigo et al., 2009) to capture the structural similarity between the gold and predicted protoforms (0: structurally dissimilar, 1: similar). With the exception of character and phoneme edit distance, the metrics enable fair comparison across different language families, which will differ in the average word length." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 687, + 365, + 699 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 687, + 365, + 699 + ], + "spans": [ + { + "bbox": [ + 302, + 687, + 365, + 699 + ], + "type": "text", + "content": "6.2 Results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 705, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 705, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 705, + 525, + 772 + ], + "type": "text", + "content": "Table 3 shows that our model consistently has the best performance on all datasets with regards to most metrics. The results were averaged across 5 runs. Out of all datasets, our model performs best on the Rom-orth dataset, where we achieve a " + }, + { + "bbox": [ + 302, + 705, + 525, + 772 + ], + "type": "inline_equation", + "content": "7.0\\%" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 291, + 781, + 304, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 781, + 304, + 791 + ], + "spans": [ + { + "bbox": [ + 291, + 781, + 304, + 791 + ], + "type": "text", + "content": "26" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 68, + 68, + 527, + 157 + ], + "blocks": [ + { + "bbox": [ + 68, + 68, + 527, + 157 + ], + "lines": [ + { + "bbox": [ + 68, + 68, + 527, + 157 + ], + "spans": [ + { + "bbox": [ + 68, + 68, + 527, + 157 + ], + "type": "table", + "html": "
Language FamilySource# varietiesCognate setsProto-language
Rom-phonDinu and Ciobanu (2014), Meloni et al. (2021)58,799Latin
Rom-orthDinu and Ciobanu (2014), Meloni et al. (2021)58,799Latin
Sinitic (Chinese)Hóu (2004)39804Middle Chinese
", + "image_path": "91fe4f67e27b10ef3b25b979e13b9b20905a6133441ae6d279a76cb8e5935bc9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 163, + 524, + 175 + ], + "lines": [ + { + "bbox": [ + 67, + 163, + 524, + 175 + ], + "spans": [ + { + "bbox": [ + 67, + 163, + 524, + 175 + ], + "type": "text", + "content": "Table 2: Statistics on both datasets used in our experiments. # varieties refers to the number of daughter varieties." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 71, + 185, + 221, + 285 + ], + "blocks": [ + { + "bbox": [ + 71, + 185, + 221, + 285 + ], + "lines": [ + { + "bbox": [ + 71, + 185, + 221, + 285 + ], + "spans": [ + { + "bbox": [ + 71, + 185, + 221, + 285 + ], + "type": "image", + "image_path": "c592223a9b2834446e15eea3269624c5d3a859092d8336de5af115ff8da6ccec.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 293, + 525, + 319 + ], + "lines": [ + { + "bbox": [ + 67, + 293, + 525, + 319 + ], + "spans": [ + { + "bbox": [ + 67, + 293, + 525, + 319 + ], + "type": "text", + "content": "Figure 2: A gold phylogeny of Romance (left) compared with those derived by probing the RNN model (middle) and the Transformer model (right) on Rom-phon." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 232, + 185, + 381, + 284 + ], + "blocks": [ + { + "bbox": [ + 232, + 185, + 381, + 284 + ], + "lines": [ + { + "bbox": [ + 232, + 185, + 381, + 284 + ], + "spans": [ + { + "bbox": [ + 232, + 185, + 381, + 284 + ], + "type": "image", + "image_path": "5bc091c6f6ab7c6395d4be007dba35a798343972ad3762c35667220fa1c358bc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 384, + 185, + 523, + 285 + ], + "blocks": [ + { + "bbox": [ + 384, + 185, + 523, + 285 + ], + "lines": [ + { + "bbox": [ + 384, + 185, + 523, + 285 + ], + "spans": [ + { + "bbox": [ + 384, + 185, + 523, + 285 + ], + "type": "image", + "image_path": "eb5a89ca965ce80dc9ef2e39ce40c63b323765447d15d389f4faf61b53fe0f46.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 338, + 291, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 338, + 291, + 461 + ], + "spans": [ + { + "bbox": [ + 67, + 338, + 291, + 461 + ], + "type": "text", + "content": "decrease in phoneme edit distance and a 1.43p.p improvement in accuracy relative to the RNN baseline. We observe the most dramatic performance difference with the RNN baseline on the Sinitic dataset: a " + }, + { + "bbox": [ + 67, + 338, + 291, + 461 + ], + "type": "inline_equation", + "content": "10.48\\%" + }, + { + "bbox": [ + 67, + 338, + 291, + 461 + ], + "type": "text", + "content": " decrease in phoneme edit distance and a 5.47p.p increase in accuracy. For reproducibility, results on the publicly available portion of the Rom-phon and Rom-orth datasets are provided in Table 7 in the Appendix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 470, + 135, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 470, + 135, + 483 + ], + "spans": [ + { + "bbox": [ + 67, + 470, + 135, + 483 + ], + "type": "text", + "content": "6.3 Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 488, + 290, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 488, + 290, + 596 + ], + "spans": [ + { + "bbox": [ + 67, + 488, + 290, + 596 + ], + "type": "text", + "content": "We observe that the BCFS is relatively high for the Romance non-neural baselines compared to those of the Chinese ones. This suggests that the sound changes in the Romance datasets are more regular than that of Chinese, which corroborates List et al. (2014)'s results that more than half of the Chinese characters in their dataset could not be explained by a tree model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 597, + 290, + 731 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 597, + 290, + 731 + ], + "spans": [ + { + "bbox": [ + 67, + 597, + 290, + 731 + ], + "type": "text", + "content": "We examine the errors made by the Transformer model on the Rom-phon dataset. Substitutions constitute around " + }, + { + "bbox": [ + 67, + 597, + 290, + 731 + ], + "type": "inline_equation", + "content": "61\\%" + }, + { + "bbox": [ + 67, + 597, + 290, + 731 + ], + "type": "text", + "content": " of the errors made by the Transformer; deletions, " + }, + { + "bbox": [ + 67, + 597, + 290, + 731 + ], + "type": "inline_equation", + "content": "21\\%" + }, + { + "bbox": [ + 67, + 597, + 290, + 731 + ], + "type": "text", + "content": ", and insertions, " + }, + { + "bbox": [ + 67, + 597, + 290, + 731 + ], + "type": "inline_equation", + "content": "18\\%" + }, + { + "bbox": [ + 67, + 597, + 290, + 731 + ], + "type": "text", + "content": ". The highest number of substitution errors occur between " + }, + { + "bbox": [ + 67, + 597, + 290, + 731 + ], + "type": "inline_equation", + "content": "[\\mathrm{i}, \\mathrm{I}]" + }, + { + "bbox": [ + 67, + 597, + 290, + 731 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 67, + 597, + 290, + 731 + ], + "type": "inline_equation", + "content": "[\\mathrm{e}, \\varepsilon]" + }, + { + "bbox": [ + 67, + 597, + 290, + 731 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 67, + 597, + 290, + 731 + ], + "type": "inline_equation", + "content": "[\\mathrm{o}, \\mathfrak{o}]" + }, + { + "bbox": [ + 67, + 597, + 290, + 731 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 597, + 290, + 731 + ], + "type": "inline_equation", + "content": "[\\mathrm{u}, \\mathrm{v}]" + }, + { + "bbox": [ + 67, + 597, + 290, + 731 + ], + "type": "text", + "content": "—vowel pairs that contrast only in tenseness. This is consistent with the analysis of Meloni et al. (2021), where substitutions between tense-lax vowel pairs take up the largest portion of errors." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 732, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 732, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 732, + 291, + 772 + ], + "type": "text", + "content": "We observe that other common substitution errors also happen between phonemes that share major phonetic features. This demonstrates that al" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 338, + 526, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 338, + 526, + 392 + ], + "spans": [ + { + "bbox": [ + 302, + 338, + 526, + 392 + ], + "type": "text", + "content": "though no explicit phonetic information is fed directly into the model, the model makes mistakes motivated by phonetic similarity, like Meloni et al. (2021)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 394, + 525, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 394, + 525, + 433 + ], + "spans": [ + { + "bbox": [ + 302, + 394, + 525, + 433 + ], + "type": "text", + "content": "We do not observe notable differences in the error statistics between the Transformer and the RNN." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 444, + 432, + 457 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 444, + 432, + 457 + ], + "spans": [ + { + "bbox": [ + 302, + 444, + 432, + 457 + ], + "type": "text", + "content": "6.4 Language relatedness" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 301, + 461, + 527, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 461, + 527, + 636 + ], + "spans": [ + { + "bbox": [ + 301, + 461, + 527, + 636 + ], + "type": "text", + "content": "Inspired by Fourrier (2022), we probe our model for diachronic information on how genetically related each Romance language is to each other. We create a distance matrix between every pair of languages in a dataset by taking the cosine similarity between a pair's language embeddings. We then use sklearnn (Pedregosa et al., 2011)'s implementation of the Ward variance minimization algorithm (Ward Jr, 1963) to perform hierarchical clustering on the distance matrix. We take a consensus of the dendrograms from 5 different runs using the consense program from PHYLIP (Felsenstein, 2013)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 638, + 527, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 638, + 527, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 638, + 527, + 772 + ], + "type": "text", + "content": "As we see in Figure 2, the Transformer captures more of the phylogenetic relationships among the languages correctly for the Rom-phon dataset. Indeed, the Generalized Quartet Distance (GQD) (Sand et al., 2013; Pompei et al., 2011; Rama et al., 2018) between the gold and predicted tree, calculated using quartetDist from the tqDist library (Sand et al., 2014), is 0.4 for the Transformer but 0.8 for the RNN. See Figure 5 in the Appendix for the results of the orthographic dataset." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 291, + 781, + 304, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 781, + 304, + 791 + ], + "spans": [ + { + "bbox": [ + 291, + 781, + 304, + 791 + ], + "type": "text", + "content": "27" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 68, + 68, + 525, + 393 + ], + "blocks": [ + { + "bbox": [ + 68, + 68, + 525, + 393 + ], + "lines": [ + { + "bbox": [ + 68, + 68, + 525, + 393 + ], + "spans": [ + { + "bbox": [ + 68, + 68, + 525, + 393 + ], + "type": "table", + "html": "
DatasetModelPED ↓NPED ↓Acc % ↑FER ↓BCFS ↑
SiniticRandom daughter (Chang et al., 2022)3.77020.84050%0.28930.2748
Majority constituent (Chang et al., 2022)3.50310.78060%0.20130.3695
CorPaR (List et al., 2022)3.27950.72780%0.39720.3332
SVM + PosStr (List et al., 2022)1.68940.369215.52%0.16690.5418
RNN (Meloni et al., 2021)1.06710.242135.65%0.08990.6781
Transformer (present work)0.95530.215041.12%0.08420.7033
Rom-phonRandom daughter (Chang et al., 2022)6.15340.69140.06%0.62640.4016
CorPaR + PosIni (List et al., 2022)1.68470.197822.18%0.07280.7403
SVM + PosStrIni (List et al., 2022)1.57870.186124.69%0.07130.7610
RNN (Meloni et al., 2021)0.96550.122452.31%0.03840.8296
Transformer (present work)0.89260.113753.75%0.03730.8435
Rom-orthRandom daughter (Chang et al., 2022)4.25670.48542.97%-0.5147
CorPaR + Ini (List et al., 2022)0.95310.116047.23%-0.8400
SVM + PosStr (List et al., 2022)0.89880.110550.43%-0.8501
RNN (Meloni et al., 2021)0.59410.077069.80%-0.8916
Transformer (present work)0.55250.072071.23%-0.9002
", + "image_path": "bc087ad415dada9f2e0679c05062b4785a9224f4fe5fba1aed8c77ffedcddbe8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 401, + 525, + 449 + ], + "lines": [ + { + "bbox": [ + 67, + 401, + 525, + 449 + ], + "spans": [ + { + "bbox": [ + 67, + 401, + 525, + 449 + ], + "type": "text", + "content": "Table 3: Evaluation of models and baselines using various metrics, averaged across 5 runs (same hyperparameters, different seeds). Because Rom-orth is not in IPA, character edit distance is used instead of PED, and we cannot accurately calculate FER. See Section 6.1 for an explanation of each evaluation metric. See Table 4 for the standard deviation values." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 470, + 291, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 470, + 291, + 647 + ], + "spans": [ + { + "bbox": [ + 67, + 470, + 291, + 647 + ], + "type": "text", + "content": "Since the Romance dataset only includes 5 daughter languages, our results are insufficient to corroborate or contradict Cathcart and Wandl (2020)'s findings: the more accurate the protoforms, the less accurate the phylogeny will be. It is not clear if the model's language embeddings are learning information that reflects shared innovations (sound changes that if shared among a set of daughter languages, would be acceptable justification for grouping them)—the only acceptable criterion for phylogenetic inference in historical linguistics (Campbell, 2013)—or if the model is learning superficial phonetic similarity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 656, + 147, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 656, + 147, + 670 + ], + "spans": [ + { + "bbox": [ + 68, + 656, + 147, + 670 + ], + "type": "text", + "content": "7 Conclusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 678, + 291, + 758 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 678, + 291, + 758 + ], + "spans": [ + { + "bbox": [ + 67, + 678, + 291, + 758 + ], + "type": "text", + "content": "By showing that Transformers can outperform previous architectures in protoform reconstruction despite the inherent data scarcity of the task, our work motivates future research in this area to take full advantage of the recent advancements in the Transformer space." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 79, + 760, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 760, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 79, + 760, + 291, + 773 + ], + "type": "text", + "content": "Accurate supervised reconstruction can help pre" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 470, + 526, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 470, + 526, + 606 + ], + "spans": [ + { + "bbox": [ + 302, + 470, + 526, + 606 + ], + "type": "text", + "content": "dict protoforms for cognate sets where linguists have not reconstructed one yet. Future work could reconstruct proto-languages whose linguist reconstructions are not available, by transferring knowledge learned from languages with already reconstructed protoforms. Furthermore, future work can leverage the abundance of work in unsupervised NMT to adapt our Transformer model for the unsupervised setting, a more realistic scenario for the historical linguist." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 303, + 624, + 365, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 624, + 365, + 637 + ], + "spans": [ + { + "bbox": [ + 303, + 624, + 365, + 637 + ], + "type": "text", + "content": "Limitations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 651, + 526, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 651, + 526, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 651, + 526, + 773 + ], + "type": "text", + "content": "One limitation of our work is that the RNN (Meloni et al., 2021) actually outperforms our Transformer on the Chinese dataset in Chang et al. (2022). In addition, as with other neural approaches, our model requires significant amounts of data, which is often not available to historical linguists researching less well-studied language families based on field reports. Romance and Chinese have relatively many cognate sets because the protoforms" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 291, + 781, + 304, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 781, + 304, + 791 + ], + "spans": [ + { + "bbox": [ + 291, + 781, + 304, + 791 + ], + "type": "text", + "content": "28" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 70, + 293, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 70, + 293, + 178 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 293, + 178 + ], + "type": "text", + "content": "are documented5, but a low resource setup with 200 cognate sets would not fare well on our data-hungrier Transformer model. Furthermore, concatenating the entire cognate set may not work on language families with hundreds of languages such as Oceanic because the input sequence would be too long compared to the output protoform sequence." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 180, + 292, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 180, + 292, + 556 + ], + "spans": [ + { + "bbox": [ + 69, + 180, + 292, + 556 + ], + "type": "text", + "content": "Finally, we obtain our Chinese gold protoforms from Baxter and Sagart (2014)'s Middle Chinese reconstruction, which was actually a transcription of the Qieyun, a rhyme dictionary. Norman and Coblin (1995) disagree with relying on such a philological source and prefer comparative reconstructions that begin from daughter data. However, there is no available comparative reconstruction of Middle Chinese with protoforms corresponding to thousands of characters to use as a gold standard. Be that as it may, it seems clear that Middle Chinese as recorded in the Qieyun is not identical to the most recent ancestor of the Chinese languages. Its preface concedes that it is a compromise between Tang Dynasty dialects. The situation with Romance is, in some ways, comparable. Classical Latin—the variety on which we train—is not the direct ancestor of modern Romance languages. Instead, they are descended from Vulgar Latin or Proto-Romance, which is not well-attested and is primarily through graffiti and other informal inscriptions. Proto-Romance reconstructions are also not exhaustive. As a result, it is difficult to find a dataset like Meloni et al. (2021) with thousands of such ancestor forms. We are also limited to the faithfulness of espeak-ng's Latin G2P, from which Meloni et al. (2021) obtain their phonetic Romance dataset." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 560, + 291, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 560, + 291, + 655 + ], + "spans": [ + { + "bbox": [ + 67, + 560, + 291, + 655 + ], + "type": "text", + "content": "For most language families, protoforms are not attested. In fact, as the term is often used, protoform refers to a form that is inferred only through linguists' comparative method. We adopt the other usage for simplicity. In practice, our approach would require reconstructions made by a linguist to serve as training labels for cognate sets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 666, + 171, + 680 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 666, + 171, + 680 + ], + "spans": [ + { + "bbox": [ + 68, + 666, + 171, + 680 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 687, + 291, + 743 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 687, + 291, + 743 + ], + "spans": [ + { + "bbox": [ + 67, + 687, + 291, + 743 + ], + "type": "text", + "content": "We would like to thank Liang (Leon) Lu for finding a bug in our implementation, Ying Chen for writing the code for the baselines, and Brendon Boldt and Graham Neubig for providing useful feedback" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 303, + 71, + 452, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 71, + 452, + 84 + ], + "spans": [ + { + "bbox": [ + 303, + 71, + 452, + 84 + ], + "type": "text", + "content": "for the first iteration of our paper." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 107, + 362, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 107, + 362, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 107, + 362, + 120 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 303, + 126, + 527, + 772 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 304, + 126, + 527, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 126, + 527, + 172 + ], + "spans": [ + { + "bbox": [ + 304, + 126, + 527, + 172 + ], + "type": "text", + "content": "Enrique Amigo, Julio Gonzalo, Javier Artiles, and Felisa Verdejo. 2009. A comparison of extrinsic clustering evaluation metrics based on formal constraints. Information retrieval, 12(4):461-486." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 180, + 527, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 527, + 214 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 527, + 214 + ], + "type": "text", + "content": "William H Baxter and Laurent Sagart. 2014. Old Chinese: A new reconstruction. Oxford University Press." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 222, + 526, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 222, + 526, + 256 + ], + "spans": [ + { + "bbox": [ + 304, + 222, + 526, + 256 + ], + "type": "text", + "content": "Lukas Biewald. 2020. Experiment tracking with weights and biases. Software available from wandb.com." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 265, + 526, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 265, + 526, + 311 + ], + "spans": [ + { + "bbox": [ + 304, + 265, + 526, + 311 + ], + "type": "text", + "content": "Alexandre Bouchard-Côté, Dan Klein, and Michael Jordan. 2008. Efficient inference in phylogenetic indel trees. In Advances in Neural Information Processing Systems, volume 21. Curran Associates, Inc." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 319, + 527, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 527, + 375 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 527, + 375 + ], + "type": "text", + "content": "Alexandre Bouchard-Côté, David Hall, Thomas L. Griffiths, and Dan Klein. 2013. Automated reconstruction of ancient languages using probabilistic models of sound change. Proceedings of the National Academy of Sciences, 110(11):4224-4229." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 303, + 383, + 526, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 383, + 526, + 407 + ], + "spans": [ + { + "bbox": [ + 303, + 383, + 526, + 407 + ], + "type": "text", + "content": "Lyle Campbell. 2013. *Historical Linguistics: an Introduction*. Edinburgh University Press." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 415, + 527, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 527, + 493 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 527, + 493 + ], + "type": "text", + "content": "Chundra Cathcart and Florian Wandl. 2020. In search of isoglosses: continuous and discrete language embeddings in Slavic historical phonology. In Proceedings of the 17th SIGMORPHON Workshop on Computational Research in Phonetics, Phonology, and Morphology, pages 233-244, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 501, + 527, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 501, + 527, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 501, + 527, + 557 + ], + "type": "text", + "content": "Kalvin Chang, Chenxuan Cui, Youngmin Kim, and David R. Mortensen. 2022. WikiHan: A new comparative dataset for Chinese languages. In Proceedings of the 29th International Conference on Computational Linguistics (COLING 2022)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 565, + 527, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 565, + 527, + 644 + ], + "spans": [ + { + "bbox": [ + 304, + 565, + 527, + 644 + ], + "type": "text", + "content": "Kyunghyun Cho, Bart van Merrienboer, Dzmitry Bah-danau, and Yoshua Bengio. 2014. On the properties of neural machine translation: Encoder-decoder approaches. In Proceedings of SSST-8, Eighth Workshop on Syntax, Semantics and Structure in Statistical Translation, pages 103-111, Doha, Qatar. Association for Computational Linguistics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 652, + 527, + 719 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 652, + 527, + 719 + ], + "spans": [ + { + "bbox": [ + 304, + 652, + 527, + 719 + ], + "type": "text", + "content": "Alina Maria Ciobanu and Liviu P. Dinu. 2018. Ab initio: Automatic Latin proto-word reconstruction. In Proceedings of the 27th International Conference on Computational Linguistics, pages 1604-1614, Santa Fe, New Mexico, USA. Association for Computational Linguistics." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 728, + 527, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 728, + 527, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 728, + 527, + 772 + ], + "type": "text", + "content": "Arthur P Dempster, Nan M Laird, and Donald B Rubin. 1977. Maximum likelihood from incomplete data via the em algorithm. Journal of the Royal Statistical Society: Series B (Methodological), 39(1):1-22." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 750, + 291, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 750, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 750, + 291, + 772 + ], + "type": "text", + "content": "In the case of Chinese, only equivalence classes of pronunciations and not exact pronunciations are recorded." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 292, + 781, + 304, + 790 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 292, + 781, + 304, + 790 + ], + "spans": [ + { + "bbox": [ + 292, + 781, + 304, + 790 + ], + "type": "text", + "content": "29" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 289, + 772 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 69, + 72, + 289, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 289, + 139 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 289, + 139 + ], + "type": "text", + "content": "Liviu Dinu and Alina Maria Ciobanu. 2014. Building a dataset of multilingual cognates for the Romanian lexicon. In Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC'14), pages 1038-1043, Reykjavik, Iceland. European Language Resources Association (ELRA)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 148, + 289, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 148, + 289, + 182 + ], + "spans": [ + { + "bbox": [ + 69, + 148, + 289, + 182 + ], + "type": "text", + "content": "Joseph Felsenstein. 2013. Phylip (phylogeny inference package), version 3.695. Department of Genome Sciences, University of Washington, Seattle." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 192, + 289, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 192, + 289, + 225 + ], + "spans": [ + { + "bbox": [ + 69, + 192, + 289, + 225 + ], + "type": "text", + "content": "Clémentine Fourrier. 2022. Neural Approaches to Historical Word Reconstruction. Ph.D. thesis, Université PSL (Paris Sciences & Lettres)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 236, + 289, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 236, + 289, + 269 + ], + "spans": [ + { + "bbox": [ + 69, + 236, + 289, + 269 + ], + "type": "text", + "content": "Andre He, Nicholas Tomlin, and Dan Klein. 2022. Neural unsupervised reconstruction of protolanguage word forms. arXiv preprint arXiv:2211.08684." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 279, + 289, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 279, + 289, + 324 + ], + "spans": [ + { + "bbox": [ + 69, + 279, + 289, + 324 + ], + "type": "text", + "content": "侯精一 Jingyi Hóu, editor. 2004. Xiandai Hanyu fangyan yinku 现代汉语方言音库 [Phonological database of Chinese dialects]. Shanghai Jiaoyu 上海教育, Shanghai 上海." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 334, + 289, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 334, + 289, + 411 + ], + "spans": [ + { + "bbox": [ + 69, + 334, + 289, + 411 + ], + "type": "text", + "content": "John D. Lafferty, Andrew McCallum, and Fernando C. N. Pereira. 2001. Conditional random fields: Probabilistic models for segmenting and labeling sequence data. In Proceedings of the Eighteenth International Conference on Machine Learning, ICML '01, page 282-289, San Francisco, CA, USA. Morgan Kaufmann Publishers Inc." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 422, + 289, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 422, + 289, + 455 + ], + "spans": [ + { + "bbox": [ + 69, + 422, + 289, + 455 + ], + "type": "text", + "content": "Vladimir I Levenshtein et al. 1966. Binary codes capable of correcting deletions, insertions, and reversals. Soviet physics doklady, 10(8):707-710." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 465, + 289, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 465, + 289, + 498 + ], + "spans": [ + { + "bbox": [ + 69, + 465, + 289, + 498 + ], + "type": "text", + "content": "Johann-Mattis List. 2019. Beyond edit distances: Comparing linguistic reconstruction systems. Theoretical Linguistics, 45(3-4):247-258." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 509, + 289, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 509, + 289, + 542 + ], + "spans": [ + { + "bbox": [ + 69, + 509, + 289, + 542 + ], + "type": "text", + "content": "Johann-Mattis List. 2021. CLDF dataset derived from Hóu's \"Phonological Database of Chinese Dialects\" from 2004. Zenodo." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 552, + 289, + 629 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 552, + 289, + 629 + ], + "spans": [ + { + "bbox": [ + 69, + 552, + 289, + 629 + ], + "type": "text", + "content": "Johann-Mattis List, Robert Forkel, and Nathan Hill. 2022. A new framework for fast automated phonological reconstruction using trimmed alignments and sound correspondence patterns. In Proceedings of the 3rd Workshop on Computational Approaches to Historical Language Change, pages 89-96, Dublin, Ireland. Association for Computational Linguistics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 640, + 289, + 684 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 640, + 289, + 684 + ], + "spans": [ + { + "bbox": [ + 69, + 640, + 289, + 684 + ], + "type": "text", + "content": "Johann-Mattis List, Nelson-Sathi Shijulal, William Martin, and Hans Geisler. 2014. Using phylogenetic networks to model chinese dialect history. Language Dynamics and Change, 4(2):222-252." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 69, + 694, + 289, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 694, + 289, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 694, + 289, + 772 + ], + "type": "text", + "content": "Carlo Meloni, Shauli Ravfogel, and Yoav Goldberg. 2021. Ab antiquo: Neural proto-language reconstruction. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4460-4473, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 524, + 689 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 304, + 72, + 524, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 524, + 148 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 524, + 148 + ], + "type": "text", + "content": "David R. Mortensen, Patrick Littell, Akash Bharadwaj, Kartik Goyal, Chris Dyer, and Lori S. Levin. 2016. Panphon: A resource for mapping IPA segments to articulatory feature vectors. In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 3475-3484." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 158, + 524, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 158, + 524, + 190 + ], + "spans": [ + { + "bbox": [ + 304, + 158, + 524, + 190 + ], + "type": "text", + "content": "Jerry L. Norman and W. South Coblin. 1995. A new approach to Chinese historical linguistics. Journal of the American Oriental Society, 115(4):576-584." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 200, + 524, + 254 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 200, + 524, + 254 + ], + "spans": [ + { + "bbox": [ + 304, + 200, + 524, + 254 + ], + "type": "text", + "content": "Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 263, + 524, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 263, + 524, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 263, + 524, + 338 + ], + "type": "text", + "content": "F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. 2011. Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12:2825-2830." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 349, + 524, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 349, + 524, + 381 + ], + "spans": [ + { + "bbox": [ + 304, + 349, + 524, + 381 + ], + "type": "text", + "content": "Simone Pompei, Vittorio Loreto, and Francesca Tria. 2011. On the accuracy of language trees. PloS one, 6(6):e20109." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 391, + 524, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 391, + 524, + 444 + ], + "spans": [ + { + "bbox": [ + 304, + 391, + 524, + 444 + ], + "type": "text", + "content": "Taraka Rama, Johann-Mattis List, Johannes Wahle, and Gerhard Jäger. 2018. Are automatic methods for cognate detection good enough for phylogenetic reconstruction in historical linguistics? arXiv preprint arXiv:1804.05416." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 454, + 524, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 454, + 524, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 454, + 524, + 509 + ], + "type": "text", + "content": "Andreas Sand, Morten K Holt, Jens Johansen, Rolf Fagerberg, Gerth Stolting Brodal, Christian NS Pedersen, and Thomas Mailund. 2013. Algorithms for computing the triplet and quartet distances for binary general trees. *Biology*, 2(4):1189–1209." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 518, + 524, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 524, + 585 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 524, + 585 + ], + "type": "text", + "content": "Andreas Sand, Morten Kragelund Holt, Jens Johansen, Rolf Fagerberg, Gerth Stolting Brodal, Thomas Mailund, and Christian N. S. Pedersen. 2014. tqdist: A library for computing the quartet and triplet distances between binary or general trees. BMC Bioinformatics, yy(xx):ii-jj." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 592, + 524, + 648 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 592, + 524, + 648 + ], + "spans": [ + { + "bbox": [ + 304, + 592, + 524, + 648 + ], + "type": "text", + "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 656, + 524, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 656, + 524, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 656, + 524, + 689 + ], + "type": "text", + "content": "Joe H Ward Jr. 1963. Hierarchical grouping to optimize an objective function. Journal of the American statistical association, 58(301):236-244." + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 292, + 781, + 304, + 790 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 292, + 781, + 304, + 790 + ], + "spans": [ + { + "bbox": [ + 292, + 781, + 304, + 790 + ], + "type": "text", + "content": "30" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 71, + 136, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 71, + 136, + 84 + ], + "spans": [ + { + "bbox": [ + 69, + 71, + 136, + 84 + ], + "type": "text", + "content": "A Training" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 92, + 289, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 92, + 289, + 267 + ], + "spans": [ + { + "bbox": [ + 69, + 92, + 289, + 267 + ], + "type": "text", + "content": "We split " + }, + { + "bbox": [ + 69, + 92, + 289, + 267 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 69, + 92, + 289, + 267 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 69, + 92, + 289, + 267 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 69, + 92, + 289, + 267 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 69, + 92, + 289, + 267 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 69, + 92, + 289, + 267 + ], + "type": "text", + "content": " of our dataset into train, validation, and test sets, respectively. We conduct hyperparameter searches using WandB (Biewald, 2020) and use early stopping, picking the epoch with lowest edit distance on validation data. All experiments are performed on a Ubuntu server with 4 GPUs and 20 CPUs. For both the RNN and the Transformer, Meloni et al. (2021)'s dataset takes less than 7 GPU hours to run, while Hóu (2004) takes less than 1 GPU hour. For the large Romance orthographic dataset, the RNN model has around 480,000 parameters, while the Transformer has around 800,000 parameters." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 278, + 185, + 292 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 278, + 185, + 292 + ], + "spans": [ + { + "bbox": [ + 69, + 278, + 185, + 292 + ], + "type": "text", + "content": "B Hyper-parameters" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 299, + 290, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 299, + 290, + 338 + ], + "spans": [ + { + "bbox": [ + 69, + 299, + 290, + 338 + ], + "type": "text", + "content": "Refer to Table 5 and Table 6 for the best hyperparameters we found during hyperparameter search via WandB." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 349, + 210, + 364 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 349, + 210, + 364 + ], + "spans": [ + { + "bbox": [ + 69, + 349, + 210, + 364 + ], + "type": "text", + "content": "C Supplementary Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 370, + 289, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 370, + 289, + 477 + ], + "spans": [ + { + "bbox": [ + 69, + 370, + 289, + 477 + ], + "type": "text", + "content": "In order to compare our model to earlier work, we used the Rom-phon and Rom-orth datasets from Meloni et al. (2021). However, this set includes a subset from Ciobanu and Dinu (2018) which is not freely redistributable. So that our results can be reproduced, we also computed them on the publicly available subset of Meloni et al. (2021)'s dataset, which is presented in Table 7." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 479, + 290, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 479, + 290, + 518 + ], + "spans": [ + { + "bbox": [ + 69, + 479, + 290, + 518 + ], + "type": "text", + "content": "Phylogenetic trees for Chinese were also extracted from the RNN and Transformer models. These are shown in Figures 3 and 4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 519, + 289, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 519, + 289, + 545 + ], + "spans": [ + { + "bbox": [ + 69, + 519, + 289, + 545 + ], + "type": "text", + "content": "We also plot the dendrograms derived from the Rom-ortho dataset in Figure 5." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 292, + 781, + 302, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 292, + 781, + 302, + 791 + ], + "spans": [ + { + "bbox": [ + 292, + 781, + 302, + 791 + ], + "type": "text", + "content": "31" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 74, + 121, + 521, + 360 + ], + "blocks": [ + { + "bbox": [ + 74, + 121, + 521, + 360 + ], + "lines": [ + { + "bbox": [ + 74, + 121, + 521, + 360 + ], + "spans": [ + { + "bbox": [ + 74, + 121, + 521, + 360 + ], + "type": "image", + "image_path": "5248f64ba74abe05eb22a117ff6b520cd8f1b60dec47d4820d7c71d6254699a2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 87, + 370, + 506, + 384 + ], + "lines": [ + { + "bbox": [ + 87, + 370, + 506, + 384 + ], + "spans": [ + { + "bbox": [ + 87, + 370, + 506, + 384 + ], + "type": "text", + "content": "Figure 3: Consensus tree of the dendrograms from the 5 runs of the Transformer for the Chinese dataset" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 74, + 485, + 523, + 702 + ], + "blocks": [ + { + "bbox": [ + 74, + 485, + 523, + 702 + ], + "lines": [ + { + "bbox": [ + 74, + 485, + 523, + 702 + ], + "spans": [ + { + "bbox": [ + 74, + 485, + 523, + 702 + ], + "type": "image", + "image_path": "8dcd73a8ee4ee6b081a940c4a2652505f191cb10ce1052f0a7021a0b8dcc613c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 101, + 714, + 492, + 728 + ], + "lines": [ + { + "bbox": [ + 101, + 714, + 492, + 728 + ], + "spans": [ + { + "bbox": [ + 101, + 714, + 492, + 728 + ], + "type": "text", + "content": "Figure 4: Consensus tree of the dendrograms from the 5 runs of the RNN for the Chinese dataset" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 292, + 781, + 305, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 292, + 781, + 305, + 791 + ], + "spans": [ + { + "bbox": [ + 292, + 781, + 305, + 791 + ], + "type": "text", + "content": "32" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 76, + 352, + 225, + 451 + ], + "blocks": [ + { + "bbox": [ + 76, + 352, + 225, + 451 + ], + "lines": [ + { + "bbox": [ + 76, + 352, + 225, + 451 + ], + "spans": [ + { + "bbox": [ + 76, + 352, + 225, + 451 + ], + "type": "image", + "image_path": "4eae6397ed120cb2ac347a487c7648b8b1e88f7345c1f3ca11567d35a24f1b24.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 464, + 525, + 489 + ], + "lines": [ + { + "bbox": [ + 67, + 464, + 525, + 489 + ], + "spans": [ + { + "bbox": [ + 67, + 464, + 525, + 489 + ], + "type": "text", + "content": "Figure 5: A gold phylogeny of Romance (left) compared with those derived by probing the RNN model (middle) and the Transformer model (right) on Rom-ortho. GQD is 0.4 for both models." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 228, + 353, + 365, + 452 + ], + "blocks": [ + { + "bbox": [ + 228, + 353, + 365, + 452 + ], + "lines": [ + { + "bbox": [ + 228, + 353, + 365, + 452 + ], + "spans": [ + { + "bbox": [ + 228, + 353, + 365, + 452 + ], + "type": "image", + "image_path": "06d23af2c68209392b2831f214ca29d7fd3ecd61bfb0d986f2b4e4c4e9b31e68.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 370, + 353, + 519, + 452 + ], + "blocks": [ + { + "bbox": [ + 370, + 353, + 519, + 452 + ], + "lines": [ + { + "bbox": [ + 370, + 353, + 519, + 452 + ], + "spans": [ + { + "bbox": [ + 370, + 353, + 519, + 452 + ], + "type": "image", + "image_path": "c9ca5665eeebeb9fad4556f807fdefcdbc81b6e0eb60ee9cf4c64b6a02809d6c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 291, + 781, + 303, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 781, + 303, + 791 + ], + "spans": [ + { + "bbox": [ + 291, + 781, + 303, + 791 + ], + "type": "text", + "content": "33" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 68, + 139, + 525, + 655 + ], + "blocks": [ + { + "bbox": [ + 68, + 139, + 525, + 655 + ], + "lines": [ + { + "bbox": [ + 68, + 139, + 525, + 655 + ], + "spans": [ + { + "bbox": [ + 68, + 139, + 525, + 655 + ], + "type": "table", + "html": "
DatasetModelPED ↓NPED ↓Acc % ↑FER ↓BCFS ↑
SiniticRandom daughter3.77020.84050%0.28930.2748
Majority constituent3.50310.78060%0.20130.3695
CorPaR3.27950.72780%0.39720.3332
SVM +PosStr1.68940.369215.52%0.16690.5418
RNN1.0671 ± 0.06190.2421 ± 0.014035.65% ± 1.60%0.0899 ± 0.00480.6781 ± 0.0174±
Transformer (present work)0.9553 ± 0.03920.2150 ± 0.007541.12% ± 2.3%0.0842 ± 0.00700.7033 ± 0.0087±
Rom-phonRandom daughter6.15340.69140.06%0.62640.4016
CorPaR +PosIni1.68470.197822.18%0.07280.7403
SVM +PosStrIni1.57870.186124.69%0.07130.7610
RNN0.9655 ± 0.01890.1224 ± 0.002252.31% ± 0.63%0.0384 ± 0.00110.8296 ± 0.0029±
Transformer (present work)0.8926 ± 0.01660.1137 ± 0.001753.75% ± 0.40%0.0373 ± 0.00090.8435 ± 0.0026±
Rom-orthRandom daughter4.25670.48542.97%-0.5147
CorPaR +Ini0.95310.116047.23%-0.8400
SVM +PosStr0.89880.110550.43%-0.8501
RNN0.5941 ± 0.01000.0770 ± 0.001569.80% ±0.22%-0.8916 ± 0.0019±
Transformer (present work)0.5525 ± 0.01040.0720 ± 0.001771.23% ± 0.52%-0.9002 ± 0.0017±
", + "image_path": "e796477de49899d75b81676a57f77e560288bc93fda55a4d5e0bfc446453ceec.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 661, + 525, + 699 + ], + "lines": [ + { + "bbox": [ + 67, + 661, + 525, + 699 + ], + "spans": [ + { + "bbox": [ + 67, + 661, + 525, + 699 + ], + "type": "text", + "content": "Table 4: Evaluation of models and baselines using various metrics, averaged across 5 runs (same hyperparameters, different seeds), with standard deviations. Because Rom-orth is not in IPA, character edit distance is used instead of PED, and we cannot accurately calculate FER. See Section 6.1 for an explanation of each evaluation metric." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 292, + 781, + 304, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 292, + 781, + 304, + 791 + ], + "spans": [ + { + "bbox": [ + 292, + 781, + 304, + 791 + ], + "type": "text", + "content": "34" + } + ] + } + ], + "index": 2 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 148, + 350, + 316 + ], + "blocks": [ + { + "bbox": [ + 69, + 148, + 350, + 316 + ], + "lines": [ + { + "bbox": [ + 69, + 148, + 350, + 316 + ], + "spans": [ + { + "bbox": [ + 69, + 148, + 350, + 316 + ], + "type": "table", + "html": "
Romance (phon & orth)Sinitic
learning rate0.000130.0007487
num Encoder_layers32
num Decoder_layers35
embedding size128128
n_head88
dim_feedforward128647
dropout0.2020.1708861
training epochs200200
warmup epochs5032
weight decay00.0000001
batch size132
", + "image_path": "204d40ab64127de14e24da13d38979c8fe0ea1098c77b9bb0a814d23ef15f193.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 69, + 515, + 379, + 669 + ], + "blocks": [ + { + "bbox": [ + 68, + 323, + 291, + 346 + ], + "lines": [ + { + "bbox": [ + 68, + 323, + 291, + 346 + ], + "spans": [ + { + "bbox": [ + 68, + 323, + 291, + 346 + ], + "type": "text", + "content": "Table 5: Hyper-parameters used in training the Transformer" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 515, + 379, + 669 + ], + "lines": [ + { + "bbox": [ + 69, + 515, + 379, + 669 + ], + "spans": [ + { + "bbox": [ + 69, + 515, + 379, + 669 + ], + "type": "table", + "html": "
Romance-phonRomance-orthSinitic
learning rate0.000557390.0009640.000864
num Encoder_layers111
num Decoder_layers111
embedding size1075178
hidden size18513073
dim_feedforward147111136
dropout0.18080.3237940.321639
training epochs181193237
warmup epochs151515
batch size884
", + "image_path": "fdaa1806f0633968c5c4bf81644afdad8e38ac450a67352e93d07d8cd58573e3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 72, + 677, + 285, + 689 + ], + "lines": [ + { + "bbox": [ + 72, + 677, + 285, + 689 + ], + "spans": [ + { + "bbox": [ + 72, + 677, + 285, + 689 + ], + "type": "text", + "content": "Table 6: Hyper-parameters used in training the RNN" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 292, + 781, + 304, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 292, + 781, + 304, + 791 + ], + "spans": [ + { + "bbox": [ + 292, + 781, + 304, + 791 + ], + "type": "text", + "content": "35" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 156, + 525, + 395 + ], + "blocks": [ + { + "bbox": [ + 69, + 156, + 525, + 395 + ], + "lines": [ + { + "bbox": [ + 69, + 156, + 525, + 395 + ], + "spans": [ + { + "bbox": [ + 69, + 156, + 525, + 395 + ], + "type": "table", + "html": "
DatasetModelPED ↓NPED ↓Acc % ↑FER ↓BCFS ↑
Rom-phonRandom daughter (Chang et al., 2022)7.18800.82010%1.13960.3406
CorPaR + Ini (List et al., 2022)2.08850.249114.29%0.08740.6799
SVM + PosStrIni (List et al., 2022)1.90050.227617.05%0.08830.7039
RNN (Meloni et al., 2021)1.45810.181536.68 %0.05920.7435
Transformer (present work)1.25160.157341.38%0.05500.7790
Rom-orthRandom daughter (Chang et al., 2022)6.32720.65420.55%-0.4023
CorPaR + PosStrIni (List et al., 2022)1.83130.200118.89%-0.7227
SVM + PosStr (List et al., 2022)1.69950.186721.66%-0.7454
RNN (Meloni et al., 2021)1.31890.150538.89%-0.7742
Transformer (present work)1.16220.134345.53%-0.7989
", + "image_path": "7c65b04eb523de4fc6c0d2f08a75b4456743aed3581e9fc4c03daa01101f7786.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 101, + 622, + 492, + 661 + ], + "blocks": [ + { + "bbox": [ + 67, + 402, + 525, + 440 + ], + "lines": [ + { + "bbox": [ + 67, + 402, + 525, + 440 + ], + "spans": [ + { + "bbox": [ + 67, + 402, + 525, + 440 + ], + "type": "text", + "content": "Table 7: Evaluation of models and baselines with various metrics on Meloni et al. (2021)'s Romance datasets, where all entries from Dinu and Ciobanu (2014) are removed, for 1 run (using the hyperparameters of the best run on the full dataset)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 101, + 622, + 492, + 661 + ], + "lines": [ + { + "bbox": [ + 101, + 622, + 492, + 661 + ], + "spans": [ + { + "bbox": [ + 101, + 622, + 492, + 661 + ], + "type": "table", + "html": "
LatinRomanianFrenchItalianSpanishPortuguese
[kɔlle:ktio:nεm][kolektsie][kɔlɛksjɔ][kolletsoine][kolekθjon][kulisțu]
", + "image_path": "a6afb984b87dcd51da4e9edc19a73e10023ca7868fa2041a3d966628c00fa39b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 669, + 509, + 682 + ], + "lines": [ + { + "bbox": [ + 83, + 669, + 509, + 682 + ], + "spans": [ + { + "bbox": [ + 83, + 669, + 509, + 682 + ], + "type": "text", + "content": "Table 8: One cognate set, with Latin as the protoform and all columns to its right as the daughter cognates" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 292, + 781, + 305, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 292, + 781, + 305, + 791 + ], + "spans": [ + { + "bbox": [ + 292, + 781, + 305, + 791 + ], + "type": "text", + "content": "36" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "spans": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "type": "text", + "content": "A For every submission:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 106, + 414, + 243 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "spans": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "type": "text", + "content": "A1. Did you describe the limitations of your work? Section 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 142, + 329, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 142, + 329, + 169 + ], + "spans": [ + { + "bbox": [ + 77, + 142, + 329, + 169 + ], + "type": "text", + "content": "A2. Did you discuss any potential risks of your work? Section 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 179, + 414, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 179, + 414, + 205 + ], + "spans": [ + { + "bbox": [ + 77, + 179, + 414, + 205 + ], + "type": "text", + "content": "A3. Do the abstract and introduction summarize the paper's main claims? Section 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 77, + 215, + 398, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 215, + 398, + 243 + ], + "spans": [ + { + "bbox": [ + 77, + 215, + 398, + 243 + ], + "type": "text", + "content": "□ A4. Have you used AI writing assistants when working on this paper? Not applicable. Left blank." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 68, + 252, + 290, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 252, + 290, + 266 + ], + "spans": [ + { + "bbox": [ + 68, + 252, + 290, + 266 + ], + "type": "text", + "content": "B Did you use or create scientific artifacts?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 79, + 270, + 138, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 270, + 138, + 282 + ], + "spans": [ + { + "bbox": [ + 79, + 270, + 138, + 282 + ], + "type": "text", + "content": "Sections 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 77, + 291, + 524, + 634 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 77, + 291, + 315, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 291, + 315, + 318 + ], + "spans": [ + { + "bbox": [ + 77, + 291, + 315, + 318 + ], + "type": "text", + "content": "B1. Did you cite the creators of artifacts you used? Sections 3,4,5,6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 77, + 327, + 463, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 327, + 463, + 354 + ], + "spans": [ + { + "bbox": [ + 77, + 327, + 463, + 354 + ], + "type": "text", + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Sections 3, 5.1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 77, + 364, + 524, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 364, + 524, + 432 + ], + "spans": [ + { + "bbox": [ + 77, + 364, + 524, + 432 + ], + "type": "text", + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Not applicable. Left blank." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 77, + 441, + 524, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 441, + 524, + 495 + ], + "spans": [ + { + "bbox": [ + 77, + 441, + 524, + 495 + ], + "type": "text", + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Not applicable. Left blank." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 77, + 503, + 524, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 503, + 524, + 544 + ], + "spans": [ + { + "bbox": [ + 77, + 503, + 524, + 544 + ], + "type": "text", + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Section 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 77, + 553, + 524, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 553, + 524, + 634 + ], + "spans": [ + { + "bbox": [ + 77, + 553, + 524, + 634 + ], + "type": "text", + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Table 2 and Appendix Section A" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "spans": [ + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "type": "text", + "content": "C Did you run computational experiments?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 79, + 661, + 123, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 661, + 123, + 673 + ], + "spans": [ + { + "bbox": [ + 79, + 661, + 123, + 673 + ], + "type": "text", + "content": "Section 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 77, + 683, + 524, + 724 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 683, + 524, + 724 + ], + "spans": [ + { + "bbox": [ + 77, + 683, + 524, + 724 + ], + "type": "text", + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Appendix A" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "spans": [ + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "type": "text", + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "spans": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "text", + "content": "ACL 2023 Responsible NLP Checklist" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 291, + 781, + 304, + 790 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 781, + 304, + 790 + ], + "spans": [ + { + "bbox": [ + 291, + 781, + 304, + 790 + ], + "type": "text", + "content": "37" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 77, + 70, + 523, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 70, + 523, + 97 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 523, + 97 + ], + "type": "text", + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 89, + 99, + 415, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 99, + 415, + 111 + ], + "spans": [ + { + "bbox": [ + 89, + 99, + 415, + 111 + ], + "type": "text", + "content": "Hyperparameter search: 5.1 Hyperparameter values: Appendix Section B" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 120, + 525, + 160 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 120, + 525, + 160 + ], + "spans": [ + { + "bbox": [ + 77, + 120, + 525, + 160 + ], + "type": "text", + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 161, + 124, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 161, + 124, + 173 + ], + "spans": [ + { + "bbox": [ + 89, + 161, + 124, + 173 + ], + "type": "text", + "content": "Table 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 183, + 525, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 183, + 525, + 223 + ], + "spans": [ + { + "bbox": [ + 77, + 183, + 525, + 223 + ], + "type": "text", + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 225, + 144, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 225, + 144, + 236 + ], + "spans": [ + { + "bbox": [ + 89, + 225, + 144, + 236 + ], + "type": "text", + "content": "5.1, 6.1, 6.3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 247, + 522, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 247, + 522, + 260 + ], + "spans": [ + { + "bbox": [ + 67, + 247, + 522, + 260 + ], + "type": "text", + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 79, + 264, + 127, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 264, + 127, + 277 + ], + "spans": [ + { + "bbox": [ + 79, + 264, + 127, + 277 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 76, + 286, + 525, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 286, + 525, + 313 + ], + "spans": [ + { + "bbox": [ + 76, + 286, + 525, + 313 + ], + "type": "text", + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 89, + 315, + 148, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 315, + 148, + 327 + ], + "spans": [ + { + "bbox": [ + 89, + 315, + 148, + 327 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 76, + 336, + 525, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 336, + 525, + 376 + ], + "spans": [ + { + "bbox": [ + 76, + 336, + 525, + 376 + ], + "type": "text", + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 89, + 378, + 148, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 378, + 148, + 390 + ], + "spans": [ + { + "bbox": [ + 89, + 378, + 148, + 390 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 76, + 400, + 525, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 400, + 525, + 439 + ], + "spans": [ + { + "bbox": [ + 76, + 400, + 525, + 439 + ], + "type": "text", + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 89, + 441, + 148, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 441, + 148, + 454 + ], + "spans": [ + { + "bbox": [ + 89, + 441, + 148, + 454 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 76, + 462, + 520, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 462, + 520, + 475 + ], + "spans": [ + { + "bbox": [ + 76, + 462, + 520, + 475 + ], + "type": "text", + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 89, + 476, + 148, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 476, + 148, + 489 + ], + "spans": [ + { + "bbox": [ + 89, + 476, + 148, + 489 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 76, + 498, + 524, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 498, + 524, + 524 + ], + "spans": [ + { + "bbox": [ + 76, + 498, + 524, + 524 + ], + "type": "text", + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 89, + 527, + 148, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 527, + 148, + 539 + ], + "spans": [ + { + "bbox": [ + 89, + 527, + 148, + 539 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 292, + 781, + 304, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 292, + 781, + 304, + 791 + ], + "spans": [ + { + "bbox": [ + 292, + 781, + 304, + 791 + ], + "type": "text", + "content": "38" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 14 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/TwistList_ Resources and Baselines for Tongue Twister Generation/450a8be3-d9ba-4e62-a952-97c7e24dcba6_content_list.json b/2023/TwistList_ Resources and Baselines for Tongue Twister Generation/450a8be3-d9ba-4e62-a952-97c7e24dcba6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0c5cb835e0d8494d07fcb13b32105491b4b74d19 --- /dev/null +++ b/2023/TwistList_ Resources and Baselines for Tongue Twister Generation/450a8be3-d9ba-4e62-a952-97c7e24dcba6_content_list.json @@ -0,0 +1,2535 @@ +[ + { + "type": "text", + "text": "TwistList: Resources and Baselines for Tongue Twister Generation", + "text_level": 1, + "bbox": [ + 169, + 90, + 825, + 110 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tyler Loakman $^{1*}$ , Chen Tang $^{2*}$ and Chenghua Lin $^{1\\dagger}$", + "bbox": [ + 292, + 130, + 712, + 147 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Department of Computer Science, The University of Sheffield, UK", + "bbox": [ + 240, + 148, + 759, + 164 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Department of Computer Science, The University of Surrey, UK", + "bbox": [ + 250, + 165, + 751, + 181 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{tcloakman1,c.lin}@sheffield.ac.uk", + "bbox": [ + 337, + 181, + 665, + 197 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "chen.tang@surrey.ac.uk", + "bbox": [ + 394, + 199, + 608, + 212 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 262, + 252, + 337, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Previous work in phonetically-grounded language generation has mainly focused on domains such as lyrics and poetry. In this paper, we present work on the generation of tongue twisters - a form of language that is required to be phonetically conditioned to maximise sound overlap, whilst maintaining semantic consistency with an input topic, and still being grammatically correct. We present TwistList, a large annotated dataset of tongue twisters, consisting of $2.1\\mathrm{K}+$ human-authored examples. We additionally present several benchmark systems (referred to as TwisterMisters) for the proposed task of tongue twister generation, including models that both do and do not require training on in-domain data. We present the results of automatic and human evaluation to demonstrate the performance of existing mainstream pre-trained models in this task with limited (or no) task specific training and data, and no explicit phonetic knowledge. We find that the task of tongue twister generation is challenging for models under these conditions, yet some models are still capable of generating acceptable examples of this language type.", + "bbox": [ + 139, + 284, + 460, + 627 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 114, + 643, + 253, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Phonetically constrained language generation is a primary subarea of computational creativity in natural language generation (NLG), primarily encompassing lyric and poetry generation (Tian and Peng, 2022; Wöckener et al., 2021; Xue et al., 2021; Zhang et al., 2020a; Agarwal and Kann, 2020), as well as pun generation (Sun et al., 2022; He et al., 2019; Yu et al., 2018), and continues to prove challenging for myriad reasons. Primarily, such works require the inclusion of phonetic factors such as metre and rhyme, which involves careful consideration of candidate vocabulary on the syllable level, leading to a reduced pool of allowable vocabulary once these constraints are in place.", + "bbox": [ + 112, + 671, + 489, + 881 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/df922e0bf87dd9a82794c843e2cecb8fc093ef3f8eecea4a89ad98ced56b52a3.jpg", + "image_caption": [ + "Figure 1: Tongue Twister Generation aims to generate an utterance with high levels of phonetic overlap, requiring understanding of semantics, grammar, and phonetics." + ], + "image_footnote": [], + "bbox": [ + 510, + 249, + 884, + 406 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we present work on the generation of tongue twisters, a type of phonetically constrained language that is rarely explored in the NLG community. As a form of creative generation, tongue twisters can facilitate numerous useful applications, including: (1) being used as a pedagogical tool (Sugiharto et al., 2022; Somoff, 2014; Wilshire, 1999); (2) as a source of humorous entertainment stemming from unintentional mispronunciations; (3) as a stylistic device for engaging children in reading (e.g. Dr. Seuss stories (Geisel, 1965)); (4) as a method of designing memorable slogans and tag lines (Guerini et al., 2015); and (5) as stimuli in neuroscience/physiology research (Wong et al., 2019; O'Halloran, 2020; Kember et al., 2017).", + "bbox": [ + 507, + 495, + 882, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tongue twister generation posits unique challenges compared to other generation tasks. One of the most pertinent features of tongue twisters is the presence of high levels of phonetic overlap across tokens (Wilshire, 1999). Consequently, whilst other types of creative generation may require only some output tokens to consider phonetics (such as rhyme or syllable counts), tongue twisters present an extreme version of this problem where the phonetics of almost all generated tokens must be considered. This leads to a very small vocabulary from which to choose", + "bbox": [ + 507, + 741, + 884, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 136, + 892, + 253, + 904 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding author.", + "bbox": [ + 136, + 904, + 272, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "579", + "bbox": [ + 485, + 927, + 515, + 940 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics", + "bbox": [ + 226, + 945, + 769, + 957 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Volume 2: Short Papers, pages 579-589", + "bbox": [ + 376, + 958, + 620, + 971 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "July 9-14, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 295, + 972, + 699, + 984 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "semantically relevant words, and presents further challenges with maintaining grammatical validity.", + "bbox": [ + 112, + 84, + 487, + 116 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The only work that we are aware of on tongue twister generation at the time of conducting this research is by Keh et al. (2022), who present models that train on graphemes and phonemes, and take either a starting prompt to be continued, or keywords around which to theme an output. They release TT-Corp, a dataset of 644 tongue twisters with parallel non-twister equivalents. We differentiate our work through the release of a dataset that is over $3\\mathrm{x}$ larger and which has undergone substantial human quality control. Furthermore, we assess the results of a wider range of popular pre-trained models on this task, including ChatGPT, without explicit injection of phonetic knowledge due to the difficulty in encoding phonetics and the expertise required to utilise phonetic characteristics appropriately. Our experimental results show that most popular pretrained language models (PLMs) rely on pure word repetition to generate tongue twisters, whilst some (i.e. BART) are able to generate more sophisticated examples. Additionally, very large zero-shot models (i.e. ChatGPT) are able to generate convincing tongue twisters almost on-par with human equivalents. $^{1}$", + "bbox": [ + 115, + 118, + 489, + 486 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarise our contributions, we present:", + "bbox": [ + 131, + 488, + 448, + 502 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- TwistList, a large annotated dataset of human-authored tongue twisters, containing $2.1\\mathrm{K}+$ examples with human evaluation of their quality.", + "- TwisterMisters, a series of baseline models for tongue twister generation using the most popular state-of-the-art PLMs.", + "- Extensive automatic and human evaluation to assess the ability of PLMs to implicitly model the complex phonetic phenomena in tongue twisters." + ], + "bbox": [ + 114, + 505, + 487, + 649 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 114, + 662, + 272, + 678 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous work in phonetically constrained generation has taken one of two approaches: 1) train a generation model on a collection of in-domain texts, or 2) train a generation model on prosaic out-of-domain text, with constraints imposed at decoding time. For example, Lau et al. (2018) collect 3,355 sonnets to produce novel poetry and train models to generate text in iambic pentameter, whilst Xue et al. (2021) train a rap generation model on 272,839 in-domain examples, infusing knowledge of rhythm afterwards. On the other hand, Van de Cruys (2020) train on a subset of CommonCrawl, imposing constraints on topic and", + "bbox": [ + 112, + 689, + 489, + 883 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/79ffdcc132cb6409a323d26bc731e27151df5d2b265c4e7f99818bda248f8a81.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetTrainValTestTotal
# Tongue Twisters19121061072128
Vocabulary Size955694688010358
# Total Phonemes55434656
# RAKE Keywords33333162883567
# BERTopic Keywords250132160250
Avg. # Input Keywords (RAKE)3.163.323.013.16
Avg. # Input Phonemes5.575.835.165.56
Avg. Tongue Twister Length (Words)15.0116.5913.5415.01
Avg. # Input Phonemes26.0628.2523.5026.04
", + "bbox": [ + 515, + 80, + 878, + 199 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1: The Statistics of TwistList.", + "bbox": [ + 576, + 208, + 811, + 221 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "rhyme as a priori distributions, whilst Tian and Peng (2022) train a title-to-keyword module on narrative texts in addition to a sonnet generation model trained on news articles and short stories from Reddit. They imposed literary techniques (simile/metaphor) and metre/rhyme constraints at decoding time, owing to the lack of sufficient training data. $^2$", + "bbox": [ + 507, + 250, + 882, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3 Tongue Twister Generation", + "text_level": 1, + "bbox": [ + 507, + 380, + 766, + 395 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1 Task Definition", + "text_level": 1, + "bbox": [ + 507, + 409, + 670, + 423 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We formulate the task of tongue twister generation as follows: for a given set of keywords, we aim to generate a tongue twister $T$ , whereby $T$ comprises a sequence of words $\\{w_1, w_2, \\dots, w_n\\}$ . The generated output must satisfy the following constraints: (1) the output should be semantically related to the input keywords; (2) the output should show maximal levels of phonetic overlap across tokens; and (3) the output should be grammatically valid (Wilshire, 1999). Of these requirements, phonetic overlap is the most central to defining text as a \"tongue twister\".", + "bbox": [ + 507, + 432, + 882, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.2 TwistList Dataset", + "text_level": 1, + "bbox": [ + 507, + 625, + 685, + 639 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dataset Construction. We present TwistList, an annotated dataset of $2.1\\mathrm{K}+$ human-authored tongue twisters for use by the community. The examples contained therein come from a variety of sources available on the web. For each tongue twister, phonetic transcription is provided using the g2p-en package, in addition to keywords extracted with RAKE and BERTopic to represent the topic of the tongue twister. Following experimentation with both RAKE and BERTopic, only RAKE keywords are used in training due to human preference and issues regarding the use of BERTopic on short texts (where", + "bbox": [ + 507, + 648, + 882, + 841 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "2Additionally, there is often a reluctance in computational creativity to train on examples, under the assumption that the newly generated content will be overly derivative.", + "bbox": [ + 507, + 854, + 882, + 891 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "3The source of each tongue twister is stated for each entry.", + "bbox": [ + 532, + 891, + 868, + 904 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "4https://pypi.org/project/g2p-en/", + "bbox": [ + 532, + 904, + 769, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1Our code and resources can be accessed at https://github.com/tangg555/TwistList", + "bbox": [ + 112, + 891, + 487, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "580", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "frequently no keywords are extracted). The main statistics of the dataset are presented in Table 1.", + "bbox": [ + 114, + 84, + 487, + 116 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/80ed8d4600a98e49a2b4dfdb7896e4a66e2e302d1bc905de00d367d9e68711d2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
RAKE:sells thick socks
BERTopic:short shorts socks sock
Twister:Seth at Sainsbury's sells thick socks.
Phonetics:[S EH1 TH] [AE1 T] [S EY1 N S B ER0 IY0 Z] [S EH1 L Z] [TH IH1 K] [S AA1 K S]
", + "bbox": [ + 119, + 139, + 485, + 246 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 2: Example from TwistList", + "bbox": [ + 191, + 255, + 410, + 269 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Quality Control. Quality control on our dataset was performed in multiple ways. Firstly, it was ensured that only sufficiently unique tongue twisters were kept in the dataset, as determined by removing examples with over $90\\%$ word overlap (rather than keeping variants of the same tongue twister, such as \"Peter Piper picked a pickled pepper\" versus \"Peter the Piper picked...\"). Additionally, non-standard spellings were manually converted to standard US English5 to avoid G2P issues.6 Similarly, tongue-twisters containing obscure vocabulary (such as medicine and dinosaur names) were excluded to further minimise errors. An annotation platform was developed (see Appendix A.1), with which 3 human evaluators, who are native speakers of English, were provided with 100 sampled instances from the dataset to rate the quality of the resulting tongue twisters and the associated extracted keywords. The full dataset contains $2,500+$ tongue twisters, of which 2,128 are kept for training/development/testing after filtering examples with insufficient extracted keywords and excessive similarity to existing entries.", + "bbox": [ + 110, + 319, + 489, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To summarise, 3 annotators evaluated the quality of the dataset, where $88\\%$ of assessed tongue twisters were considered high quality, and $6\\%$ considered \"suitable\" (Kappa $= 0.321$ ). An example from TwistList is provided in Table 2. As Table 4 shows, the final dataset can be considered high quality, owing to fair/moderate levels of approval and agreement across evaluators. Demographic information of the evaluators can be found in Appendix A.2.", + "bbox": [ + 110, + 680, + 489, + 826 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3 Baseline Models", + "text_level": 1, + "bbox": [ + 507, + 84, + 675, + 98 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We present the following baseline models (dubbed TwisterMisters) for the task of tongue twister generation on our TwistList dataset:", + "bbox": [ + 507, + 105, + 882, + 154 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finetuned Baselines. For the finetuned baselines, we chose popular models for language generation, including GPT-2 (Radford et al., 2019), DialogGPT (Zhang et al., 2020c), T5 (Raffel et al., 2020), and BART (Lewis et al., 2020). These were finetuned with RAKE keywords extracted from human-authored tongue twisters as the input and the tongue twister text from TwistList as the target. This was in order to represent our baselines training on in-domain data. At inference time, the prompt \"Generate tongue twisters about the keyword(s): X\" is used, where X refers to the input consisting of one or more RAKE keywords extracted from tongue twisters. The full training details are given in Appendix A.3. We also conducted experiments on all aforementioned baselines without finetuning (i.e., a zero-shot setting), and the results were very poor. Therefore, we did not include these results in the paper.", + "bbox": [ + 507, + 165, + 884, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training-Free Baseline We additionally provide a TwisterMister baseline that does not require any training. We utilise OpenAI's ChatGPT7 with the same prompt as a zero-shot setting for generation.8 Each request to ChatGPT was submitted as part of a separate session, to avoid the effects of extended dialogue influencing outputs. ChatGPT has been utilised in order to set a practical upper-bound of what may be expected from models without explicit phonetic knowledge, owing to its wealth of training data and 175B parameter architecture.9 It is assumed that ChatGPT's training data contains tongue twisters, and therefore it is able to abstract away the general patterns of such language in order to provide novel examples (though most likely based on graphemes rather than phonemes).", + "bbox": [ + 507, + 466, + 882, + 708 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 507, + 721, + 650, + 738 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Automatic Evaluation. We present the results of automatic evaluation on generated outputs and golden examples in Table 3 for the following metrics: Perplexity (PPL), BLEU (B-1/B-2) (Papineni et al., 2002), ROUGE (R-1/R-2/R-L) (Lin, 2004), and BERTScore Precision, Recall, and F-Measure (Zhang", + "bbox": [ + 507, + 747, + 882, + 844 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "For example, where phonetic spellings or letter substitutions such as \"k\" for \"c\" were used for literary and visual effect, such as \"kwik\" for \"quick\".", + "bbox": [ + 112, + 854, + 487, + 891 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "$^6 g2p$ -en uses the CMU Pronouncing Dictionary to retrieve transcriptions, which is an American English resource.", + "bbox": [ + 112, + 892, + 485, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "7https://chat.openai.com/chat", + "bbox": [ + 532, + 854, + 742, + 866 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "8No direct comparison is made to PANCETTA (Keh et al., 2022) as no code has been publicly released at the time of writing, and essential implementation details are absent from the paper.", + "bbox": [ + 509, + 866, + 882, + 904 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "9ModelPPL↓B-1↑B-2↑R-1↑R-2↑R-L↑PO↓Init-PO↓BS-P↑BS-R↑BS-F↑GPT-28.400.0070.0031.3010.1231.3150.0220.0200.6900.8100.744DialoGPT3.830.0380.0257.7243.6107.6400.0690.0890.7540.8310.790T510.160.0570.0389.7014.5739.5740.6890.7270.7950.8180.806BART1.650.0730.05111.8836.10910.3530.0750.1200.7950.8450.819ChatGPTN/A0.2000.13736.76520.65933.4370.0930.1570.8880.8940.883", + "bbox": [ + 137, + 80, + 858, + 174 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/e6c3fa3ffb3747a02c714049eda8f7e2c7b40a1eb7868abf9c6028ce05620605.jpg", + "table_caption": [ + "Table 3: Results of Automatic Evaluation. Golden PO = 0.385 and Golden Init-PO = 0.417. Due to the one-to-many issue in creative language generation, we acknowledge that the referenced metrics are imperfect." + ], + "table_footnote": [], + "table_body": "
Choices (%)Sample Quality
High.Suitable.Bad.Kappa
RAKE keywords82.018.00.00.321
BERTopic keywords15.085.00.00.445
Tongue Twisters88.06.04.00.321
", + "bbox": [ + 127, + 236, + 473, + 312 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 4: Kappa refers to Fleiss' Kappa (Fleiss, 1971). All results achieve fair or moderate agreement. Good tongue twisters that are deemed a bit longer (3%) or shorter (3%) than expected are considered \"suitable\".", + "bbox": [ + 112, + 321, + 487, + 379 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "et al., 2020b) (BS-P/BS-R/BS-F). PPL, BLEU and ROUGE are standard metrics in language generation to assess quality, whilst BERTScore assesses semantic similarity to a gold reference. Additionally, we propose two new metrics, Phonetic Overlap (PO) and Initial Phonetic Overlap (Init-PO). PO refers to the average overlap of all phonemes across tokens (# unique phonemes/#totalphonemes),whereas Init-PO is the ratio of unique word-initial phonemes to the number of words (# unique word-initial phonemes/#words).", + "bbox": [ + 112, + 411, + 487, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "These phonetic metrics reward longer outputs. We argue that, all things equal, a longer tongue twister is better than a shorter one as it provides more entertainment and more opportunities for mispronunciation. Perfect scores on PO and Init-PO can be achieved by repetition of a single word. Whilst this does not lead to high quality outputs, these metrics are intended exclusively to be indicators of the phonetics, rather than an overall guide to quality. In both cases, higher levels of overlap results in lower (\"better\") scores, and the highest (\"worst\") achievable score is 1.", + "bbox": [ + 112, + 577, + 487, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The results in Table 3 show rather clear scaling, with the performance ranking on most metrics (except Perplexity and phoneme overlap) being identical. On the models explicitly finetuned for this task, GPT-2 is shown to be the worst, whilst BART performs the best. We hypothesise that GPT-2's poor performance is in part due to its simple causal language modelling objective alongside its decoder-only architecture (which is also in DialogGPT). Furthermore, whilst T5 performed well on the automatic metrics, manual", + "bbox": [ + 112, + 758, + 487, + 917 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "inspection revealed that T5 often misinterpreted the task from the prompt, choosing to select its own keywords from the entire prompt, rather than using only the provided keyword list. On the other hand, the training-free zero-shot model, ChatGPT, was shown to perform best on all metrics. This is to be expected as ChatGPT has over 50x more parameters than any other tested PLM, with various pre-training objectives and reinforcement learning, leading to performant zero-shot capabilities. This further demonstrates that PLMs struggle to learn phonetic patterns implicitly from text, especially in English, which has high levels of irregular orthography. Furthermore, with limited data, PLMs struggle to learn the unusual probability distributions underlying tongue twisters, where word choices are intentionally \"twisted\", obscure, and anti-euphonious. Additionally, due to the wealth of training data seen by ChatGPT, it is likely that many examples have been seen during training.", + "bbox": [ + 505, + 237, + 882, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Human Evaluation. Due to tongue twisters being a creative domain where articulation abilities are tested, we also perform human evaluation. 3 evaluators were asked to rate 100 outputs from the best performing standard baseline (BART), in addition to ChatGPT outputs and gold examples from TwistList on the following criteria: Relevance (how relevant the tongue twister is given the keyword inputs), Fluency (how grammatically valid the output is), Difficulty of Articulation (how difficult a tongue twister is to say), Cohesion (how much sense the output makes), and Entertainment Value (how entertaining the output is, considering sounds and semantics). All ratings were on a 5-point Likert scale. Evaluator demographics and training materials are in Appendix A.2.", + "bbox": [ + 507, + 557, + 882, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The mean scores of human evaluation (Table 5) fall in line with expectations, with golden examples performing best on all metrics, and ChatGPT placing second on all but Difficulty of Articulation.[10] BART is able to produce outputs that are deemed to be the", + "bbox": [ + 507, + 799, + 882, + 879 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "10We exclude relevance for the golden examples as these were collected from the web, not elicited with keyword prompts.", + "bbox": [ + 507, + 892, + 880, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "582", + "bbox": [ + 485, + 927, + 515, + 940 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/65969a47912de5bc50d6f38177701cc9162d5660f9aad3bfc7d543af72c03022.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Score (1 to 5)Human Evaluation
BARTChatGPTGolden
Relevance4.667*4.971†N/A
Difficulty of Articulation4.143*4.102*4.291*
Fluency3.028**4.915**4.938**
Coherence3.217*4.798*4.909*
Entertainment Value3.269*4.070*4.254*
", + "bbox": [ + 127, + 80, + 475, + 183 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "second most difficult to articulate, which we infer may be the result of slight morphological variants of input keywords being used repeatedly, making distinguishing between them during articulation quite challenging (whilst not being able to exploit deeper phonetic relations). The moderate score on Fluency (3.028) suggests instances of poor grammar may also hinder articulation abilities when expected grammatical structures are not found, leading to an interaction between grammatical validity and articulatory difficulty. Additionally, ChatGPT scoring the lowest for articulatory difficulty may be due to occasionally misunderstanding the requirements of a tongue twister, sometimes producing rhymes or standard prose (see Appendix A.4). However, ChatGPT scores well for Relevance and Fluency, highlighting its capability in producing high-quality coherent language. Perhaps most interestingly, none of the BART score averages on any human evaluation criteria fall below 3 (\"neither agree nor disagree\"). This performance is therefore quite good for a model finetuned on only 2128 examples, with no additional phonetic knowledge.", + "bbox": [ + 112, + 282, + 489, + 636 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/9c108b07e72f1da819ac8f0ed8a7050c7a2490c45b4f84cab83601b30b897d9f.jpg", + "table_caption": [ + "Table 5: Results of Human Evaluation. The best scores are in bold, and the second best are underlined. We calculate Fleiss' Kappa for each metric, and mark the agreement fair*, moderate** and substantial†." + ], + "table_footnote": [], + "table_body": "
Inputassistant assistant assist
GPT-2assistant assistant assist assistant assist assistant
DialogGPTassistant assistant assistant assistant assistant assistant assistant assistant assistant assistant assistant assistant
T5assistant assistant assist assistant
BARTA assistant assist is an assistant assist, assistants assist to assist assistants.
ChatGPTAssistant ants assist ants in carrying leaves to the ant hill.
GoldenIf I assist a sister-assistant, will the sister's sister-assistant assist me?
", + "bbox": [ + 119, + 656, + 485, + 858 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 6: Example outputs for the input \"assistant assist\". \"Golden\" refers to the human-authored tongue twisters.", + "bbox": [ + 112, + 866, + 487, + 896 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5 Case Study", + "text_level": 1, + "bbox": [ + 509, + 83, + 636, + 99 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Within the example in Table 6, GPT-2 resorts to simply repeating the input, successfully achieving phonetic overlap, but failing to be grammatically valid or particularly sophisticated. This pattern is also demonstrated by DialogGPT and T5. Conversely, BART is able to introduce tokens unseen in the input to create an almost grammatically valid output (the primary mistake being indefinite article agreement, where in the first instance \"an\" would have been correct, rather than \"a\"). BART's output is also semantically and logically coherent, with \"A assistant assist is an assistant assist\" being valid (yet redundant), and \"assistants assist to assist assistants\" also being comprehensible. This example demonstrates why evaluators with high English proficiency and language/linguistics education were selected, as the same word may have different parts of speech, creating outputs that seem grammatically invalid, but do actually follow the rules of English.[11]", + "bbox": [ + 505, + 118, + 885, + 423 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Further investigation is needed to ascertain whether the models are intentionally exploiting this lexical ambiguity, or if human evaluators are demonstrating apophobia, where patterns are found in what is effectively noise (Brugger, 2001). Finally, ChatGPT utilises morphology to exploit the similarity of the plural noun \"assistants\" and the phrase \"assist ants\", and provides a continuation that is in line with the expected behaviour of ants. In comparison to the golden example, ChatGPT's output may be considered more interesting topic-wise, at the expense of not being as phonetically complex (\"carrying leaves to the ant hill\" contributes heavily to semantics, whilst not being recognisable as part of a tongue twister). For further analysis, please see Appendix A.4.", + "bbox": [ + 507, + 424, + 885, + 665 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 507, + 689, + 636, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We present work on the topic of tongue twister generation, a form of phonetically-constrained language generation that aims to maximise phonetic overlap, whilst conveying meaningful semantics. We motivate the potential application domains for such generated language, and provide a large annotated dataset of tongue twisters, TwistList, to encourage further work. Finally, we present a series of benchmark models alongside automatic/human evaluation to assess generation quality.", + "bbox": [ + 507, + 724, + 885, + 869 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "11https://en.wikipedia.org/wiki/Buffalo_buffalo_Buffalo_buffalo_buffalo_buffalo_buffalo_buffalo", + "bbox": [ + 507, + 891, + 870, + 917 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "583", + "bbox": [ + 485, + 927, + 515, + 940 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Limitations", + "text_level": 1, + "bbox": [ + 114, + 83, + 213, + 98 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Whilst the system presented within this paper is capable of allowing human-in-the-loop contributions (via selecting the input keywords on which to condition the output), it is not able to produce tongue-twisters that take advantage of particular features of speech sounds such as place and manner of articulation, in order to create more advanced outputs that exploit phonetic relatedness (rather than exact matches). The same can be said of our proposed metrics, PO and Init-PO, which do not account for phonetic similarity across sounds that share manner/place of articulation (e.g. \"she sells sea shells\"). Additionally, whilst commonly known tongue twisters may follow a particular format (e.g. rhyme schemes), such schemes and templates have not been enforced here. We also do not demonstrate the capabilities of these systems if they were trained on phonetic transcriptions explicitly, as we only aim to assess their performance when training on graphemes in standard orthography.", + "bbox": [ + 115, + 114, + 489, + 419 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ethics Statement", + "text_level": 1, + "bbox": [ + 114, + 438, + 257, + 453 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All use of human participants in this study has been approved by the Ethics Board of the primary author's institution, including the disclosure of demographic information. Regarding the generation of tongue twisters, language generation is a necessarily creative domain that has the ability to reproduce content that some individuals may find offensive. Care was taken to check outputs in the human evaluation set for any such materials, and if they had been produced, they would have been removed from the evaluation set. Additionally, no egregiously offensive material has been provided in the TwistList dataset. However, the distinction between offensive and humorous content is a highly complex topic, and therefore some examples within the dataset may not be suitable for all individuals (e.g. suggestive content and swearing, such as \"I'm not the pheasant plucker, I'm the pheasant plucker's son\", and the clear relation to common expletives).", + "bbox": [ + 115, + 468, + 487, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 114, + 776, + 278, + 791 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tyler Loakman is supported by the Centre for Doctoral Training in Speech and Language Technologies (SLT) and their Applications funded by UK Research and Innovation [grant number EP/S023062/1]. Chen Tang is supported by the China Scholarship Council (CSC) for his doctoral study (File No.202006120039).", + "bbox": [ + 112, + 806, + 487, + 903 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 510, + 83, + 603, + 98 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Rajat Agarwal and Katharina Kann. 2020. Acrostic poem generation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1230-1240, Online. Association for Computational Linguistics.", + "Peter Brugger. 2001. From haunted brain to haunted science: A cognitive neuroscience view of paranormal and pseudoscientific thought. In James Hournan and RenseEditors Lange, editors, *Hauntings and Poltergeists: Multidisciplinary Perspectives*, page 195-213. McFarland.", + "Joseph L Fleiss. 1971. Measuring nominal scale agreement among many raters. Psychological bulletin, 76(5):378.", + "Theodore Seuss Geisel. 1965. *Fox in socks: Dr. Seuss's book of tongue tanglers*. Random House.", + "Marco Guerini, Gözde Özbal, and Carlo Strapparava. 2015. Echoes of persuasion: The effect of euphony in persuasive communication. In Proceedings of the 2015 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1483-1493, Denver, Colorado. Association for Computational Linguistics.", + "He He, Nanyun Peng, and Percy Liang. 2019. Pun generation with surprise. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 1734-1744, Minneapolis, Minnesota. Association for Computational Linguistics.", + "Henglin Huang, Chen Tang, Tyler Loakman, Frank Guerin, and Chenghua Lin. 2022. Improving Chinese story generation via awareness of syntactic dependencies and semantics. In Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 2: Short Papers).", + "Sedrick Scott Keh, Steven Y. Feng, Varun Gangal, Malihe Alikhani, and Eduard Hovy. 2022. Pancetta: Phoneme aware neural completion to elicit tongue twisters automatically.", + "Heather Kember, Kathryn Connaghan, and Rupal Patel. 2017. Inducing speech errors in dysarthria using tongue twisters. International journal of language & communication disorders, 52(4):469-478.", + "Jey Han Lau, Trevor Cohn, Timothy Baldwin, Julian Brooke, and Adam Hammond. 2018. Deep-spare: A joint neural model of poetic language, meter and rhyme. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1948-1958, Melbourne, Australia. Association for Computational Linguistics.", + "Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy," + ], + "bbox": [ + 510, + 105, + 884, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "584", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871-7880, Online. Association for Computational Linguistics.", + "Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81, Barcelona, Spain. Association for Computational Linguistics.", + "Ken D. O'Halloran. 2020. A tongue-twister to translation? increased complexity of genioglossus movement during wakefulness in persons with obstructive sleep apnoea. The Journal of Physiology, 598(3):435-436.", + "Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311-318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics.", + "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9.", + "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(140):1-67.", + "Victoria Somoff. 2014. Four is not fourteen: Tongue twister patterns and the unmastery of language. Western Folklore, 73(2/3):195-215.", + "Prasetyawan Sugiharto, Yan Santoso, and Maila Shofyana. 2022. Teaching english pronunciation using tongue twister. Acitya: Journal of Teaching and Education, 4(1):189-197.", + "Jiao Sun, Anjali Narayan-Chen, Shereen Oraby, Shuyang Gao, Tagyoung Chung, Jing Huang, Yang Liu, and Nanyun Peng. 2022. Context-situated pun generation. In EMNLP 2022.", + "Chen Tang, Chenghua Lin, Henglin Huang, Frank Guerin, and Zhihao Zhang. 2022a. EtrICA: Event-triggered context-aware story generation augmented by cross attention. In *Findings of the Association for Computational Linguistics: EMNLP* 2022.", + "Chen Tang, Hongbo Zhang, Tyler Loakman, Chenghua Lin, and Frank Guerin. 2022b. Terminology-aware medical dialogue generation. arXiv preprint arXiv:2210.15551.", + "Chen Tang, Zhihao Zhang, Tyler Loakman, Chenghua Lin, and Frank Guerin. 2022c. NGEP: A graph-based event planning framework for story generation. In Proceedings of AACL-IJCNLP, Online." + ], + "bbox": [ + 115, + 85, + 487, + 917 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Yufei Tian and Nanyun Peng. 2022. Zero-shot sonnet generation with discourse-level planning and aesthetics features. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3587-3597, Seattle, United States. Association for Computational Linguistics.", + "Tim Van de Cruys. 2020. Automatic poetry generation from prosaic text. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2471-2480, Online. Association for Computational Linguistics.", + "Carolyn E. Wilshire. 1999. The \"tongue twister\" paradigm as a technique for studying phonological encoding. Language and Speech, 42(1):57-82.", + "Jörg Wöckener, Thomas Haider, Tristan Miller, The-Khang Nguyen, Thanh Tung Linh Nguyen, Minh Vu Pham, Jonas Belouadi, and Steffen Eger. 2021. End-to-end style-conditioned poetry generation: What does it take to learn from examples alone? In Proceedings of the 5th Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature, pages 57-66, Punta Cana, Dominican Republic (online). Association for Computational Linguistics.", + "Min Ney Wong, Yanky Chan, Manwa L. Ng, and Frank F. Zhu. 2019. Effects of transcranial direct current stimulation over the broca's area on tongue twister production. International Journal of Speech-Language Pathology, 21(2):182-188. PMID: 29642741.", + "Lanqing Xue, Kaitao Song, Duocai Wu, Xu Tan, Nevin L. Zhang, Tao Qin, Wei-Qiang Zhang, and Tie-Yan Liu. 2021. DeepRapper: Neural rap generation with rhyme and rhythm modeling. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 69-81, Online. Association for Computational Linguistics.", + "Zhiwei Yu, Jiwei Tan, and Xiaojun Wan. 2018. A neural approach to pun generation. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1650-1660, Melbourne, Australia. Association for Computational Linguistics.", + "Rongsheng Zhang, Xiaoxi Mao, Le Li, Lin Jiang, Lin Chen, Zhiwei Hu, Yadong Xi, Changjie Fan, and Minlie Huang. 2020a. Youling: an AI-assisted lyrics creation system. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 85-91, Online. Association for Computational Linguistics.", + "Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi. 2020b. *Bertscore: Evaluating text generation with bert*. In International Conference on Learning Representations." + ], + "bbox": [ + 510, + 85, + 880, + 917 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "585", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, and Bill Dolan. 2020c. DIALOGPT: Large-scale generative pre-training for conversational response generation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 270-278, Online. Association for Computational Linguistics.", + "bbox": [ + 115, + 85, + 489, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A Appendices", + "text_level": 1, + "bbox": [ + 114, + 203, + 248, + 219 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A.1 Dataset Quality Control", + "text_level": 1, + "bbox": [ + 114, + 229, + 342, + 243 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "An annotation platform was developed as shown in (Figure 2).", + "bbox": [ + 112, + 250, + 487, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A.2 Human Participants", + "text_level": 1, + "bbox": [ + 114, + 292, + 315, + 307 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Due to tongue twisters being highly reliant on articulation abilities, the demographics of the human participants used within this work are highly important. Additionally, tongue twisters are also a form of humour and entertainment, where individual perceptions of what may or may not be considered humorous or entertaining differ according to numerous factors. In an effort to remain as transparent as possible, and follow best practices for human evaluation, relevant demographic information of participants are outlined below (with the necessary requisite permission and ethical approval).", + "bbox": [ + 112, + 313, + 485, + 505 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Dataset Evaluation All evaluators involved in the quality control process of the TwistList dataset are native speakers of English, and either have or are working towards University level qualifications. Additionally, 2 of the 3 evaluators have extensive education in linguistics or modern languages. No monetary incentive was provided.", + "bbox": [ + 112, + 514, + 487, + 627 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Generation Evaluation All evaluators involved in the evaluation of the quality of generated tongue twisters are native speakers of English, and either hold or are working towards University level qualifications in Linguistics, Modern Languages or NLP. Additionally, all evaluators cited the United Kingdom as their country of socialisation, and no participants reported language processing difficulties that could affect results. No monetary incentive was provided.", + "bbox": [ + 112, + 636, + 489, + 781 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Materials Provided to Human Participants Additionally, all evaluators for both the dataset and generation outputs were presented with calibration examples to demonstrate the sort of outputs that would be presented, and the logic behind particular scores, in order to minimise individual interpretations of the scoring criteria. All evaluation was performed on a custom made online annotation platform (Figure 3).", + "bbox": [ + 112, + 790, + 487, + 919 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A.3 Training Details", + "text_level": 1, + "bbox": [ + 509, + 84, + 680, + 99 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "All pre-trained models used (naturally excluding ChatGPT) are based on publicly available checkpoints from Hugging Face.12 Models are trained for up to 5 epochs on a Tesla A5000 machine with the best checkpoints selected based on the validation loss. The batch size is set to 32, and the learning rate is $8e^{-5}$ , with the Adam optimiser selected for training. To help the loss curve converge on our small few-shot dataset, we limit the generation length to 100 (covering all test tongue twisters). Meanwhile, the source length is limited to 150. The training and testing steps are set up with the implementation of the PyTorch Lightning13 framework to guarantee the reliability of the experiment. All language models are fairly trained and tested with the same steps.", + "bbox": [ + 507, + 105, + 882, + 348 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A.4 Further Qualitative Comments", + "text_level": 1, + "bbox": [ + 507, + 360, + 789, + 374 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Whilst the pattern of extreme word repetition is seen in many of the finetuned models (often with the exception of BART, which is demonstrated to be capable of producing slightly more sophisticated outputs), overall assessment of the tongue twisters produced at inference time reveals interesting patterns, particularly in regard to ChatGPT outputs. Firstly, the limits of ChatGPT are made apparent in a few examples such as the input \"silver shiny ship sank\" generating \"How much wood would a woodchuck chuck if a woodchuck could chuck silver shiny ships?\", a clear derivation of a famous woodchuck related tongue twister that it is rather safe to assume appears multiple times in ChatGPTs training material. Additionally, comments from evaluators also reveal that ChatGPT's output is often considered more of a rhyme or general literary text, rather than specifically a tongue twister. However, examples such as these are also found in the human-authored golden examples, demonstrating that there is no steadfast consistent opinion as to what constitutes a (good) tongue twister. Likewise, some examples may contain large amounts of sound repetition, but not in a way that necessarily presents articulatory difficulty.", + "bbox": [ + 507, + 381, + 882, + 751 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A.5 Future Works", + "text_level": 1, + "bbox": [ + 509, + 764, + 663, + 778 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we mainly analyse the performance of large-scale pretrained language models (PLMs) on Tongue Twister Generation, and propose a corresponding dataset for further investigation. In further works, we aim to propose novel models which can better leverage phonetic symbols. There", + "bbox": [ + 507, + 785, + 880, + 881 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "12https://huggingface.co/models", + "bbox": [ + 526, + 890, + 749, + 904 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "13https://www.pytorchlightning.ai/", + "bbox": [ + 527, + 904, + 769, + 917 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "586", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Introduction", + "text_level": 1, + "bbox": [ + 137, + 103, + 184, + 110 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Individually read the tongue twister, phonetics, and key words on the left side.", + "Select the options on the right side to evaluate the data quality from the following perspectives:", + "- The quality of the RAKE Keywords: Do these suitably represent the topic of the tongue twister?", + "- The quality of the BERTopic Keywords: Do these suitably represent the topic of the tongue twister?", + "- The quality of the Tongue Twister: Is it a good tongue twister, or too short/long, or generally bad quality?" + ], + "bbox": [ + 144, + 120, + 433, + 154 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Tongue Twisters", + "text_level": 1, + "bbox": [ + 144, + 167, + 194, + 174 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Keywords for Tongue Twisters", + "bbox": [ + 152, + 187, + 242, + 193 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "RAKE Keywords:", + "bbox": [ + 157, + 206, + 206, + 212 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "pickled peppers peter piper picked", + "bbox": [ + 221, + 206, + 317, + 212 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "[P I H1 K A H0 L D] [P E H1 P E R O Z] [P I Y1 T E R O] [P A Y1 P E R O] [P I H1 K T]", + "bbox": [ + 221, + 212, + 411, + 219 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "BERTopic", + "bbox": [ + 157, + 227, + 186, + 233 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Keywords:", + "bbox": [ + 157, + 234, + 189, + 240 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "pink peter piper peck", + "bbox": [ + 221, + 228, + 282, + 233 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "[P I H1NGK][P I Y1TERO][P AY1PER0][P E H1K]", + "bbox": [ + 221, + 234, + 352, + 241 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Tongue Twister", + "bbox": [ + 152, + 258, + 200, + 265 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Tongue Twister", + "bbox": [ + 157, + 279, + 203, + 286 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Text:", + "bbox": [ + 157, + 287, + 174, + 292 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Peter Piper picked a peck of pickled peppers. A peck of pickled peppers Peter Piper picked. If Peter Piper picked a peck of pickled peppers, Where's the peck of pickled peppers Peter Piper picked?", + "bbox": [ + 221, + 278, + 473, + 298 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Tongue Twister", + "bbox": [ + 157, + 306, + 203, + 312 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Phonetics:", + "bbox": [ + 157, + 313, + 189, + 318 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "[PI Y1 ER0] [PY A1 ER0] [PI H1 K T] [AOH] [PE H1 K] [AHV] [PI I H1 K AI O L D] [PE H1 ER0 Z] [JI AHO] [PE H1 K] [AHV] [PI I H1 K AI O L D] [PE H1 ER0 Z] [PI Y1 ER0] [PY A1 ER0] [PI I H1 K T] [JI (IH1 F) [PI Y1 T ER0] [PY A1 ER0] [PI I H1 K T] [AOH] [PE H1 K] [AHV] [PI I H1 K AI O L D] [PE H1 ER0 Z] [WEH1 K R2] [DH AHO] [PE H1 K] [AHV] [PI I H1 K AI O L D] [PE H1 ER0 Z] [PI Y1 T ER0] [PY A1 ER0] [PI I H1 K T] [Q]:", + "bbox": [ + 221, + 306, + 473, + 340 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Sample Quality", + "text_level": 1, + "bbox": [ + 509, + 104, + 556, + 111 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "A1: Are the RAKE Keywords highly suitable for the Tongue Twister?", + "bbox": [ + 515, + 123, + 712, + 131 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Yes", + "bbox": [ + 519, + 140, + 539, + 145 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "No", + "bbox": [ + 519, + 146, + 539, + 152 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "A2: Are the BERTopic Keywords highly suitable for the Tongue Twister?", + "bbox": [ + 515, + 172, + 724, + 179 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Yes", + "bbox": [ + 519, + 187, + 537, + 193 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "No", + "bbox": [ + 519, + 195, + 537, + 200 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "A3: Is it a good tongue twister?", + "bbox": [ + 515, + 219, + 608, + 227 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "Good", + "A Bit Too Short", + "A Bit Too Long", + "$\\bigcirc$ Bad Quality" + ], + "bbox": [ + 519, + 236, + 569, + 262 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "A4: Please input up to 5 manual keywords for the tongue twister (these can either be extracted from the example, or any other words).", + "bbox": [ + 515, + 282, + 821, + 296 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Keywords", + "bbox": [ + 519, + 305, + 549, + 311 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Quit", + "bbox": [ + 512, + 349, + 529, + 356 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Submit", + "bbox": [ + 825, + 349, + 848, + 356 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Figure 2: TwistList Quality Control Annotation Platform", + "bbox": [ + 315, + 374, + 680, + 388 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Tongue Twister Evaluation", + "text_level": 1, + "bbox": [ + 137, + 403, + 265, + 413 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Home / Sample Annotation", + "bbox": [ + 779, + 404, + 858, + 411 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Introduction", + "text_level": 1, + "bbox": [ + 142, + 422, + 181, + 429 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Individually read the input keywords and the tongue twister on the left side.", + "2. Give a score for each metric on the right to evaluate the quality of generated tongue twisters:", + "- Relevance: The extent to which the tongue twister is remantically/topically related to the input keywords.", + "- Difficulty of Articulation: The extent to which the tongue twister is hard to say (aka. how much your tongue twists).", + "- Fluency: The extent to which the tongue twister can be considered grammatically acceptable.", + "- Coherence: The extent to which the tongue twister can be considered logically and semantically coherent.", + "Entertainment: The extent to which the tongue twister is considered enteratining (primarily relating to phonetics +", + "semantics)" + ], + "bbox": [ + 142, + 438, + 455, + 492 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Tongue Twisters", + "text_level": 1, + "bbox": [ + 142, + 505, + 191, + 512 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Input Keywords", + "bbox": [ + 151, + 526, + 200, + 532 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "RAKE Keywords:", + "bbox": [ + 156, + 545, + 203, + 551 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Generate tongue twisters about key words: nope", + "bbox": [ + 221, + 545, + 351, + 551 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Tongue Twister", + "bbox": [ + 151, + 569, + 196, + 575 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Tongue Twister Text:", + "bbox": [ + 156, + 589, + 201, + 601 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Nope, an antelope can't elope with a cantaloupe.", + "bbox": [ + 221, + 588, + 352, + 594 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "", + "bbox": [ + 221, + 595, + 352, + 601 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluate the Quality of Tongue Twisters by giving a Score (1 to 5)", + "text_level": 1, + "bbox": [ + 505, + 420, + 697, + 430 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Please select all options before submission.", + "bbox": [ + 515, + 442, + 645, + 449 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "O1:Relevance.", + "bbox": [ + 519, + 458, + 561, + 463 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "#", + "bbox": [ + 519, + 464, + 537, + 470 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "O2: Difficulty of articulation.", + "bbox": [ + 519, + 482, + 601, + 487 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "0", + "bbox": [ + 519, + 493, + 537, + 498 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Q3:FLuency.", + "bbox": [ + 519, + 507, + 556, + 512 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "0", + "bbox": [ + 519, + 520, + 537, + 525 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Q4: Coherence.", + "bbox": [ + 519, + 532, + 561, + 537 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5", + "bbox": [ + 519, + 545, + 537, + 550 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Q5: Entertainment.", + "bbox": [ + 519, + 556, + 574, + 561 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "。", + "bbox": [ + 519, + 570, + 537, + 575 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Quit", + "bbox": [ + 512, + 600, + 527, + 606 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Submit", + "bbox": [ + 823, + 600, + 848, + 606 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Figure 3: Human Evaluation Platform for Generated Outputs", + "bbox": [ + 300, + 626, + 694, + 640 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "are numerous existing works (Huang et al., 2022; Tang et al., 2022a,b) that provide approaches for injecting such knowledge into PLMs. However, the phonetic features differ from these text-format knowledge items, as phonemes are hard to encode with input text tokens when feeding into PLM encoders. Another promising approach is to explicitly model the phonetic features into text sequences (Tang et al., 2022c), though there is no observed method for transforming phonetic notation. We intend to perform further research based on these existing approaches.", + "bbox": [ + 112, + 665, + 487, + 843 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Tongue Twister Dataset Evaluation", + "bbox": [ + 137, + 84, + 307, + 93 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Home / Sample Annotation", + "bbox": [ + 779, + 85, + 858, + 91 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "587", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "A For every submission:", + "text_level": 1, + "bbox": [ + 114, + 107, + 322, + 122 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "A1. Did you describe the limitations of your work?", + "bbox": [ + 129, + 126, + 532, + 143 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Yes, in the required Limitations section as well as Section 4 (concerning our proposed metrics)", + "bbox": [ + 149, + 143, + 848, + 159 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "A2. Did you discuss any potential risks of your work?", + "bbox": [ + 129, + 168, + 552, + 186 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Ethics Statement", + "bbox": [ + 149, + 187, + 278, + 200 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "A3. Do the abstract and introduction summarize the paper's main claims?", + "bbox": [ + 129, + 212, + 695, + 229 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Abstract (all) and contribution summary at the end of the introduction.", + "bbox": [ + 149, + 230, + 672, + 244 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "A4. Have you used AI writing assistants when working on this paper?", + "bbox": [ + 129, + 256, + 668, + 272 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 149, + 273, + 231, + 287 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "B Did you use or create scientific artifacts?", + "text_level": 1, + "bbox": [ + 114, + 299, + 487, + 316 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "TwistList dataset (Section 3.2)", + "bbox": [ + 132, + 321, + 359, + 336 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "B1. Did you cite the creators of artifacts you used?", + "bbox": [ + 129, + 346, + 529, + 363 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Sources of all entries in the dataset are credited in the .json file for each entry.", + "bbox": [ + 149, + 363, + 727, + 379 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?", + "bbox": [ + 129, + 390, + 779, + 406 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We did not discuss the licensing around our dataset. The dataset uses works that are freely available on the web and come from various sources such as websites, blogs, and eBooks. Many of these cases are Public Domain, and for those that are not, we believe we are in accordance with Fair Use, as the dataset does not encroach on the use case of the original works (no graphic design/other elements are maintained) and the dataset is for use as a research tool only. We will also reply promptly to any cases of copyright infringement that relevant copyright holders make us aware of.", + "bbox": [ + 149, + 407, + 880, + 502 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?", + "bbox": [ + 129, + 513, + 880, + 577 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "See answer to B2.", + "bbox": [ + 149, + 579, + 285, + 592 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?", + "bbox": [ + 129, + 604, + 880, + 652 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "See the Ethics Statement regarding the potential for tongue twisters to be offensive. Additionally, all tongue twisters are believed to be about fictional characters, rather than individuals.", + "bbox": [ + 149, + 653, + 880, + 684 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?", + "bbox": [ + 129, + 696, + 880, + 728 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Such details are not explicitly stated. However, it can be easily ascertained from the paper that the tongue twisters we focus on are entirely in English (and the range of domains the tongue twisters were taken from can be seen in the \"source\" entry for each example).", + "bbox": [ + 149, + 728, + 880, + 776 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be.", + "bbox": [ + 129, + 785, + 880, + 866 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "See Table 1 for dataset statistics.", + "bbox": [ + 149, + 868, + 394, + 883 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "ACL 2023 Responsible NLP Checklist", + "bbox": [ + 132, + 84, + 433, + 99 + ], + "page_idx": 9 + }, + { + "type": "page_footnote", + "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.", + "bbox": [ + 112, + 887, + 877, + 912 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "588", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "C Did you run computational experiments?", + "text_level": 1, + "bbox": [ + 114, + 83, + 494, + 99 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Section 4 (page 3)", + "bbox": [ + 132, + 105, + 272, + 121 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?", + "bbox": [ + 129, + 130, + 878, + 162 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Appendix A.3", + "bbox": [ + 149, + 164, + 253, + 180 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?", + "bbox": [ + 129, + 189, + 880, + 222 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Appendix A.3", + "bbox": [ + 149, + 223, + 253, + 239 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?", + "bbox": [ + 129, + 248, + 880, + 296 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Tables 3/5. Scores are the mean, as is standard.", + "bbox": [ + 149, + 298, + 502, + 313 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?", + "bbox": [ + 129, + 324, + 880, + 370 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Exact details of evaluation implementations (except Phonetic Overlap) were not detailed. This is in part due to these metrics (BLEU/ROUGE/BERTScore) not being very reliable for creative language generation, and therefore the exact values from different implementations are not likely to be of use.", + "bbox": [ + 147, + 373, + 880, + 420 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?", + "text_level": 1, + "bbox": [ + 114, + 430, + 875, + 448 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Section 3.2 and Section 4. In addition to Appendix A.2", + "bbox": [ + 132, + 453, + 537, + 468 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?", + "bbox": [ + 129, + 478, + 880, + 510 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Screenshot of the annotation platforms can be found in Figures 2 and 3 in the Appendix", + "bbox": [ + 149, + 512, + 803, + 527 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?", + "bbox": [ + 129, + 537, + 880, + 585 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We declared that no monetary incentive was given to participants. We did not specify the recruitment process, but due to participants all holding or working towards university level qualifications, it can be inferred that they are colleagues.", + "bbox": [ + 147, + 587, + 880, + 634 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?", + "bbox": [ + 129, + 645, + 880, + 692 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "This information was not deemed necessary in the submitted paper (due to the limited risk of the data we were working with). However, it is stated in the Ethical Statement and Appendix A.2 that all shared information about human demographics was collected with the necessary permissions and approval.", + "bbox": [ + 147, + 694, + 880, + 757 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?", + "bbox": [ + 129, + 766, + 875, + 784 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ethical approval was gained for human evaluation of the dataset and generated outputs from the relevant institution", + "bbox": [ + 149, + 785, + 880, + 815 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?", + "bbox": [ + 129, + 827, + 880, + 858 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We provide demographic information for human participants in Appendix A.2", + "bbox": [ + 149, + 860, + 722, + 876 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "589", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/TwistList_ Resources and Baselines for Tongue Twister Generation/450a8be3-d9ba-4e62-a952-97c7e24dcba6_model.json b/2023/TwistList_ Resources and Baselines for Tongue Twister Generation/450a8be3-d9ba-4e62-a952-97c7e24dcba6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a85cdd4e50f11321962afdc8e09028bac67036df --- /dev/null +++ b/2023/TwistList_ Resources and Baselines for Tongue Twister Generation/450a8be3-d9ba-4e62-a952-97c7e24dcba6_model.json @@ -0,0 +1,3093 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.171, + 0.091, + 0.826, + 0.111 + ], + "angle": 0, + "content": "TwistList: Resources and Baselines for Tongue Twister Generation" + }, + { + "type": "text", + "bbox": [ + 0.293, + 0.131, + 0.714, + 0.148 + ], + "angle": 0, + "content": "Tyler Loakman\\(^{1*}\\), Chen Tang\\(^{2*}\\) and Chenghua Lin\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.149, + 0.761, + 0.165 + ], + "angle": 0, + "content": "\\(^{1}\\)Department of Computer Science, The University of Sheffield, UK" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.166, + 0.752, + 0.182 + ], + "angle": 0, + "content": "\\(^{2}\\)Department of Computer Science, The University of Surrey, UK" + }, + { + "type": "text", + "bbox": [ + 0.339, + 0.183, + 0.666, + 0.198 + ], + "angle": 0, + "content": "{tcloakman1,c.lin}@sheffield.ac.uk" + }, + { + "type": "text", + "bbox": [ + 0.395, + 0.2, + 0.609, + 0.214 + ], + "angle": 0, + "content": "chen.tang@surrey.ac.uk" + }, + { + "type": "title", + "bbox": [ + 0.263, + 0.253, + 0.339, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.141, + 0.285, + 0.461, + 0.628 + ], + "angle": 0, + "content": "Previous work in phonetically-grounded language generation has mainly focused on domains such as lyrics and poetry. In this paper, we present work on the generation of tongue twisters - a form of language that is required to be phonetically conditioned to maximise sound overlap, whilst maintaining semantic consistency with an input topic, and still being grammatically correct. We present TwistList, a large annotated dataset of tongue twisters, consisting of \\(2.1\\mathrm{K}+\\) human-authored examples. We additionally present several benchmark systems (referred to as TwisterMisters) for the proposed task of tongue twister generation, including models that both do and do not require training on in-domain data. We present the results of automatic and human evaluation to demonstrate the performance of existing mainstream pre-trained models in this task with limited (or no) task specific training and data, and no explicit phonetic knowledge. We find that the task of tongue twister generation is challenging for models under these conditions, yet some models are still capable of generating acceptable examples of this language type." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.644, + 0.254, + 0.658 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.672, + 0.49, + 0.882 + ], + "angle": 0, + "content": "Phonetically constrained language generation is a primary subarea of computational creativity in natural language generation (NLG), primarily encompassing lyric and poetry generation (Tian and Peng, 2022; Wöckener et al., 2021; Xue et al., 2021; Zhang et al., 2020a; Agarwal and Kann, 2020), as well as pun generation (Sun et al., 2022; He et al., 2019; Yu et al., 2018), and continues to prove challenging for myriad reasons. Primarily, such works require the inclusion of phonetic factors such as metre and rhyme, which involves careful consideration of candidate vocabulary on the syllable level, leading to a reduced pool of allowable vocabulary once these constraints are in place." + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.25, + 0.885, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.416, + 0.883, + 0.46 + ], + "angle": 0, + "content": "Figure 1: Tongue Twister Generation aims to generate an utterance with high levels of phonetic overlap, requiring understanding of semantics, grammar, and phonetics." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.496, + 0.884, + 0.736 + ], + "angle": 0, + "content": "In this paper, we present work on the generation of tongue twisters, a type of phonetically constrained language that is rarely explored in the NLG community. As a form of creative generation, tongue twisters can facilitate numerous useful applications, including: (1) being used as a pedagogical tool (Sugiharto et al., 2022; Somoff, 2014; Wilshire, 1999); (2) as a source of humorous entertainment stemming from unintentional mispronunciations; (3) as a stylistic device for engaging children in reading (e.g. Dr. Seuss stories (Geisel, 1965)); (4) as a method of designing memorable slogans and tag lines (Guerini et al., 2015); and (5) as stimuli in neuroscience/physiology research (Wong et al., 2019; O'Halloran, 2020; Kember et al., 2017)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.743, + 0.885, + 0.919 + ], + "angle": 0, + "content": "Tongue twister generation posits unique challenges compared to other generation tasks. One of the most pertinent features of tongue twisters is the presence of high levels of phonetic overlap across tokens (Wilshire, 1999). Consequently, whilst other types of creative generation may require only some output tokens to consider phonetics (such as rhyme or syllable counts), tongue twisters present an extreme version of this problem where the phonetics of almost all generated tokens must be considered. This leads to a very small vocabulary from which to choose" + }, + { + "type": "page_footnote", + "bbox": [ + 0.137, + 0.893, + 0.255, + 0.905 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.137, + 0.905, + 0.273, + 0.918 + ], + "angle": 0, + "content": "†Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.137, + 0.893, + 0.273, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.928, + 0.516, + 0.941 + ], + "angle": 0, + "content": "579" + }, + { + "type": "footer", + "bbox": [ + 0.227, + 0.946, + 0.77, + 0.958 + ], + "angle": 0, + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + }, + { + "type": "footer", + "bbox": [ + 0.377, + 0.959, + 0.621, + 0.972 + ], + "angle": 0, + "content": "Volume 2: Short Papers, pages 579-589" + }, + { + "type": "footer", + "bbox": [ + 0.297, + 0.973, + 0.7, + 0.985 + ], + "angle": 0, + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.489, + 0.117 + ], + "angle": 0, + "content": "semantically relevant words, and presents further challenges with maintaining grammatical validity." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.119, + 0.49, + 0.487 + ], + "angle": 0, + "content": "The only work that we are aware of on tongue twister generation at the time of conducting this research is by Keh et al. (2022), who present models that train on graphemes and phonemes, and take either a starting prompt to be continued, or keywords around which to theme an output. They release TT-Corp, a dataset of 644 tongue twisters with parallel non-twister equivalents. We differentiate our work through the release of a dataset that is over \\(3\\mathrm{x}\\) larger and which has undergone substantial human quality control. Furthermore, we assess the results of a wider range of popular pre-trained models on this task, including ChatGPT, without explicit injection of phonetic knowledge due to the difficulty in encoding phonetics and the expertise required to utilise phonetic characteristics appropriately. Our experimental results show that most popular pretrained language models (PLMs) rely on pure word repetition to generate tongue twisters, whilst some (i.e. BART) are able to generate more sophisticated examples. Additionally, very large zero-shot models (i.e. ChatGPT) are able to generate convincing tongue twisters almost on-par with human equivalents.\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.132, + 0.489, + 0.449, + 0.504 + ], + "angle": 0, + "content": "To summarise our contributions, we present:" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.506, + 0.489, + 0.553 + ], + "angle": 0, + "content": "- TwistList, a large annotated dataset of human-authored tongue twisters, containing \\(2.1\\mathrm{K}+\\) examples with human evaluation of their quality." + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.554, + 0.488, + 0.6 + ], + "angle": 0, + "content": "- TwisterMisters, a series of baseline models for tongue twister generation using the most popular state-of-the-art PLMs." + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.602, + 0.488, + 0.65 + ], + "angle": 0, + "content": "- Extensive automatic and human evaluation to assess the ability of PLMs to implicitly model the complex phonetic phenomena in tongue twisters." + }, + { + "type": "list", + "bbox": [ + 0.115, + 0.506, + 0.489, + 0.65 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.663, + 0.273, + 0.679 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.69, + 0.49, + 0.884 + ], + "angle": 0, + "content": "Previous work in phonetically constrained generation has taken one of two approaches: 1) train a generation model on a collection of in-domain texts, or 2) train a generation model on prosaic out-of-domain text, with constraints imposed at decoding time. For example, Lau et al. (2018) collect 3,355 sonnets to produce novel poetry and train models to generate text in iambic pentameter, whilst Xue et al. (2021) train a rap generation model on 272,839 in-domain examples, infusing knowledge of rhythm afterwards. On the other hand, Van de Cruys (2020) train on a subset of CommonCrawl, imposing constraints on topic and" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.082, + 0.88, + 0.2 + ], + "angle": 0, + "content": "
DatasetTrainValTestTotal
# Tongue Twisters19121061072128
Vocabulary Size955694688010358
# Total Phonemes55434656
# RAKE Keywords33333162883567
# BERTopic Keywords250132160250
Avg. # Input Keywords (RAKE)3.163.323.013.16
Avg. # Input Phonemes5.575.835.165.56
Avg. Tongue Twister Length (Words)15.0116.5913.5415.01
Avg. # Input Phonemes26.0628.2523.5026.04
" + }, + { + "type": "table_caption", + "bbox": [ + 0.578, + 0.209, + 0.813, + 0.222 + ], + "angle": 0, + "content": "Table 1: The Statistics of TwistList." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.252, + 0.883, + 0.365 + ], + "angle": 0, + "content": "rhyme as a priori distributions, whilst Tian and Peng (2022) train a title-to-keyword module on narrative texts in addition to a sonnet generation model trained on news articles and short stories from Reddit. They imposed literary techniques (simile/metaphor) and metre/rhyme constraints at decoding time, owing to the lack of sufficient training data.\\(^2\\)" + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.381, + 0.768, + 0.397 + ], + "angle": 0, + "content": "3 Tongue Twister Generation" + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.41, + 0.671, + 0.424 + ], + "angle": 0, + "content": "3.1 Task Definition" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.433, + 0.884, + 0.61 + ], + "angle": 0, + "content": "We formulate the task of tongue twister generation as follows: for a given set of keywords, we aim to generate a tongue twister \\( T \\), whereby \\( T \\) comprises a sequence of words \\( \\{w_1, w_2, \\dots, w_n\\} \\). The generated output must satisfy the following constraints: (1) the output should be semantically related to the input keywords; (2) the output should show maximal levels of phonetic overlap across tokens; and (3) the output should be grammatically valid (Wilshire, 1999). Of these requirements, phonetic overlap is the most central to defining text as a \"tongue twister\"." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.626, + 0.686, + 0.64 + ], + "angle": 0, + "content": "3.2 TwistList Dataset" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.649, + 0.884, + 0.842 + ], + "angle": 0, + "content": "Dataset Construction. We present TwistList, an annotated dataset of \\(2.1\\mathrm{K}+\\) human-authored tongue twisters for use by the community. The examples contained therein come from a variety of sources available on the web. For each tongue twister, phonetic transcription is provided using the g2p-en package, in addition to keywords extracted with RAKE and BERTopic to represent the topic of the tongue twister. Following experimentation with both RAKE and BERTopic, only RAKE keywords are used in training due to human preference and issues regarding the use of BERTopic on short texts (where" + }, + { + "type": "page_footnote", + "bbox": [ + 0.508, + 0.855, + 0.883, + 0.892 + ], + "angle": 0, + "content": "2Additionally, there is often a reluctance in computational creativity to train on examples, under the assumption that the newly generated content will be overly derivative." + }, + { + "type": "page_footnote", + "bbox": [ + 0.533, + 0.892, + 0.869, + 0.905 + ], + "angle": 0, + "content": "3The source of each tongue twister is stated for each entry." + }, + { + "type": "page_footnote", + "bbox": [ + 0.533, + 0.905, + 0.771, + 0.918 + ], + "angle": 0, + "content": "4https://pypi.org/project/g2p-en/" + }, + { + "type": "list", + "bbox": [ + 0.508, + 0.855, + 0.883, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.892, + 0.488, + 0.919 + ], + "angle": 0, + "content": "1Our code and resources can be accessed at https://github.com/tangg555/TwistList" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.517, + 0.941 + ], + "angle": 0, + "content": "580" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.115, + 0.085, + 0.488, + 0.117 + ], + "angle": 0, + "content": "frequently no keywords are extracted). The main statistics of the dataset are presented in Table 1." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.14, + 0.486, + 0.247 + ], + "angle": 0, + "content": "
RAKE:sells thick socks
BERTopic:short shorts socks sock
Twister:Seth at Sainsbury's sells thick socks.
Phonetics:[S EH1 TH] [AE1 T] [S EY1 N S B ER0 IY0 Z] [S EH1 L Z] [TH IH1 K] [S AA1 K S]
" + }, + { + "type": "table_caption", + "bbox": [ + 0.193, + 0.256, + 0.411, + 0.27 + ], + "angle": 0, + "content": "Table 2: Example from TwistList" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.32, + 0.49, + 0.675 + ], + "angle": 0, + "content": "Quality Control. Quality control on our dataset was performed in multiple ways. Firstly, it was ensured that only sufficiently unique tongue twisters were kept in the dataset, as determined by removing examples with over \\(90\\%\\) word overlap (rather than keeping variants of the same tongue twister, such as \"Peter Piper picked a pickled pepper\" versus \"Peter the Piper picked...\"). Additionally, non-standard spellings were manually converted to standard US English5 to avoid G2P issues.6 Similarly, tongue-twisters containing obscure vocabulary (such as medicine and dinosaur names) were excluded to further minimise errors. An annotation platform was developed (see Appendix A.1), with which 3 human evaluators, who are native speakers of English, were provided with 100 sampled instances from the dataset to rate the quality of the resulting tongue twisters and the associated extracted keywords. The full dataset contains \\(2,500+\\) tongue twisters, of which 2,128 are kept for training/development/testing after filtering examples with insufficient extracted keywords and excessive similarity to existing entries." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.681, + 0.49, + 0.827 + ], + "angle": 0, + "content": "To summarise, 3 annotators evaluated the quality of the dataset, where \\(88\\%\\) of assessed tongue twisters were considered high quality, and \\(6\\%\\) considered \"suitable\" (Kappa \\(= 0.321\\)). An example from TwistList is provided in Table 2. As Table 4 shows, the final dataset can be considered high quality, owing to fair/moderate levels of approval and agreement across evaluators. Demographic information of the evaluators can be found in Appendix A.2." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.085, + 0.677, + 0.099 + ], + "angle": 0, + "content": "3.3 Baseline Models" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.107, + 0.883, + 0.155 + ], + "angle": 0, + "content": "We present the following baseline models (dubbed TwisterMisters) for the task of tongue twister generation on our TwistList dataset:" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.166, + 0.885, + 0.456 + ], + "angle": 0, + "content": "Finetuned Baselines. For the finetuned baselines, we chose popular models for language generation, including GPT-2 (Radford et al., 2019), DialogGPT (Zhang et al., 2020c), T5 (Raffel et al., 2020), and BART (Lewis et al., 2020). These were finetuned with RAKE keywords extracted from human-authored tongue twisters as the input and the tongue twister text from TwistList as the target. This was in order to represent our baselines training on in-domain data. At inference time, the prompt \"Generate tongue twisters about the keyword(s): X\" is used, where X refers to the input consisting of one or more RAKE keywords extracted from tongue twisters. The full training details are given in Appendix A.3. We also conducted experiments on all aforementioned baselines without finetuning (i.e., a zero-shot setting), and the results were very poor. Therefore, we did not include these results in the paper." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.467, + 0.884, + 0.709 + ], + "angle": 0, + "content": "Training-Free Baseline We additionally provide a TwisterMister baseline that does not require any training. We utilise OpenAI's ChatGPT7 with the same prompt as a zero-shot setting for generation.8 Each request to ChatGPT was submitted as part of a separate session, to avoid the effects of extended dialogue influencing outputs. ChatGPT has been utilised in order to set a practical upper-bound of what may be expected from models without explicit phonetic knowledge, owing to its wealth of training data and 175B parameter architecture.9 It is assumed that ChatGPT's training data contains tongue twisters, and therefore it is able to abstract away the general patterns of such language in order to provide novel examples (though most likely based on graphemes rather than phonemes)." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.722, + 0.652, + 0.739 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.749, + 0.884, + 0.845 + ], + "angle": 0, + "content": "Automatic Evaluation. We present the results of automatic evaluation on generated outputs and golden examples in Table 3 for the following metrics: Perplexity (PPL), BLEU (B-1/B-2) (Papineni et al., 2002), ROUGE (R-1/R-2/R-L) (Lin, 2004), and BERTScore Precision, Recall, and F-Measure (Zhang" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.856, + 0.488, + 0.892 + ], + "angle": 0, + "content": "For example, where phonetic spellings or letter substitutions such as \"k\" for \"c\" were used for literary and visual effect, such as \"kwik\" for \"quick\"." + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.893, + 0.487, + 0.919 + ], + "angle": 0, + "content": "\\( ^6 g2p \\)-en uses the CMU Pronouncing Dictionary to retrieve transcriptions, which is an American English resource." + }, + { + "type": "list", + "bbox": [ + 0.114, + 0.856, + 0.488, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.533, + 0.855, + 0.744, + 0.868 + ], + "angle": 0, + "content": "7https://chat.openai.com/chat" + }, + { + "type": "page_footnote", + "bbox": [ + 0.51, + 0.868, + 0.883, + 0.905 + ], + "angle": 0, + "content": "8No direct comparison is made to PANCETTA (Keh et al., 2022) as no code has been publicly released at the time of writing, and essential implementation details are absent from the paper." + }, + { + "type": "page_footnote", + "bbox": [ + 0.533, + 0.905, + 0.816, + 0.918 + ], + "angle": 0, + "content": "9ModelPPL↓B-1↑B-2↑R-1↑R-2↑R-L↑PO↓Init-PO↓BS-P↑BS-R↑BS-F↑GPT-28.400.0070.0031.3010.1231.3150.0220.0200.6900.8100.744DialoGPT3.830.0380.0257.7243.6107.6400.0690.0890.7540.8310.790T510.160.0570.0389.7014.5739.5740.6890.7270.7950.8180.806BART1.650.0730.05111.8836.10910.3530.0750.1200.7950.8450.819ChatGPTN/A0.2000.13736.76520.65933.4370.0930.1570.8880.8940.883" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.185, + 0.882, + 0.214 + ], + "angle": 0, + "content": "Table 3: Results of Automatic Evaluation. Golden PO = 0.385 and Golden Init-PO = 0.417. Due to the one-to-many issue in creative language generation, we acknowledge that the referenced metrics are imperfect." + }, + { + "type": "table", + "bbox": [ + 0.129, + 0.237, + 0.475, + 0.313 + ], + "angle": 0, + "content": "
Choices (%)Sample Quality
High.Suitable.Bad.Kappa
RAKE keywords82.018.00.00.321
BERTopic keywords15.085.00.00.445
Tongue Twisters88.06.04.00.321
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.322, + 0.489, + 0.38 + ], + "angle": 0, + "content": "Table 4: Kappa refers to Fleiss' Kappa (Fleiss, 1971). All results achieve fair or moderate agreement. Good tongue twisters that are deemed a bit longer (3%) or shorter (3%) than expected are considered \"suitable\"." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.412, + 0.489, + 0.573 + ], + "angle": 0, + "content": "et al., 2020b) (BS-P/BS-R/BS-F). PPL, BLEU and ROUGE are standard metrics in language generation to assess quality, whilst BERTScore assesses semantic similarity to a gold reference. Additionally, we propose two new metrics, Phonetic Overlap (PO) and Initial Phonetic Overlap (Init-PO). PO refers to the average overlap of all phonemes across tokens (# unique phonemes/#totalphonemes),whereas Init-PO is the ratio of unique word-initial phonemes to the number of words (# unique word-initial phonemes/#words)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.578, + 0.489, + 0.755 + ], + "angle": 0, + "content": "These phonetic metrics reward longer outputs. We argue that, all things equal, a longer tongue twister is better than a shorter one as it provides more entertainment and more opportunities for mispronunciation. Perfect scores on PO and Init-PO can be achieved by repetition of a single word. Whilst this does not lead to high quality outputs, these metrics are intended exclusively to be indicators of the phonetics, rather than an overall guide to quality. In both cases, higher levels of overlap results in lower (\"better\") scores, and the highest (\"worst\") achievable score is 1." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.759, + 0.489, + 0.919 + ], + "angle": 0, + "content": "The results in Table 3 show rather clear scaling, with the performance ranking on most metrics (except Perplexity and phoneme overlap) being identical. On the models explicitly finetuned for this task, GPT-2 is shown to be the worst, whilst BART performs the best. We hypothesise that GPT-2's poor performance is in part due to its simple causal language modelling objective alongside its decoder-only architecture (which is also in DialogGPT). Furthermore, whilst T5 performed well on the automatic metrics, manual" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.239, + 0.883, + 0.546 + ], + "angle": 0, + "content": "inspection revealed that T5 often misinterpreted the task from the prompt, choosing to select its own keywords from the entire prompt, rather than using only the provided keyword list. On the other hand, the training-free zero-shot model, ChatGPT, was shown to perform best on all metrics. This is to be expected as ChatGPT has over 50x more parameters than any other tested PLM, with various pre-training objectives and reinforcement learning, leading to performant zero-shot capabilities. This further demonstrates that PLMs struggle to learn phonetic patterns implicitly from text, especially in English, which has high levels of irregular orthography. Furthermore, with limited data, PLMs struggle to learn the unusual probability distributions underlying tongue twisters, where word choices are intentionally \"twisted\", obscure, and anti-euphonious. Additionally, due to the wealth of training data seen by ChatGPT, it is likely that many examples have been seen during training." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.558, + 0.884, + 0.799 + ], + "angle": 0, + "content": "Human Evaluation. Due to tongue twisters being a creative domain where articulation abilities are tested, we also perform human evaluation. 3 evaluators were asked to rate 100 outputs from the best performing standard baseline (BART), in addition to ChatGPT outputs and gold examples from TwistList on the following criteria: Relevance (how relevant the tongue twister is given the keyword inputs), Fluency (how grammatically valid the output is), Difficulty of Articulation (how difficult a tongue twister is to say), Cohesion (how much sense the output makes), and Entertainment Value (how entertaining the output is, considering sounds and semantics). All ratings were on a 5-point Likert scale. Evaluator demographics and training materials are in Appendix A.2." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.8, + 0.883, + 0.881 + ], + "angle": 0, + "content": "The mean scores of human evaluation (Table 5) fall in line with expectations, with golden examples performing best on all metrics, and ChatGPT placing second on all but Difficulty of Articulation.[10] BART is able to produce outputs that are deemed to be the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.509, + 0.893, + 0.882, + 0.919 + ], + "angle": 0, + "content": "10We exclude relevance for the golden examples as these were collected from the web, not elicited with keyword prompts." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.928, + 0.516, + 0.941 + ], + "angle": 0, + "content": "582" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.129, + 0.082, + 0.477, + 0.184 + ], + "angle": 0, + "content": "
Score (1 to 5)Human Evaluation
BARTChatGPTGolden
Relevance4.667*4.971†N/A
Difficulty of Articulation4.143*4.102*4.291*
Fluency3.028**4.915**4.938**
Coherence3.217*4.798*4.909*
Entertainment Value3.269*4.070*4.254*
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.192, + 0.49, + 0.25 + ], + "angle": 0, + "content": "Table 5: Results of Human Evaluation. The best scores are in bold, and the second best are underlined. We calculate Fleiss' Kappa for each metric, and mark the agreement fair*, moderate** and substantial†." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.284, + 0.49, + 0.637 + ], + "angle": 0, + "content": "second most difficult to articulate, which we infer may be the result of slight morphological variants of input keywords being used repeatedly, making distinguishing between them during articulation quite challenging (whilst not being able to exploit deeper phonetic relations). The moderate score on Fluency (3.028) suggests instances of poor grammar may also hinder articulation abilities when expected grammatical structures are not found, leading to an interaction between grammatical validity and articulatory difficulty. Additionally, ChatGPT scoring the lowest for articulatory difficulty may be due to occasionally misunderstanding the requirements of a tongue twister, sometimes producing rhymes or standard prose (see Appendix A.4). However, ChatGPT scores well for Relevance and Fluency, highlighting its capability in producing high-quality coherent language. Perhaps most interestingly, none of the BART score averages on any human evaluation criteria fall below 3 (\"neither agree nor disagree\"). This performance is therefore quite good for a model finetuned on only 2128 examples, with no additional phonetic knowledge." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.657, + 0.486, + 0.859 + ], + "angle": 0, + "content": "
Inputassistant assistant assist
GPT-2assistant assistant assist assistant assist assistant
DialogGPTassistant assistant assistant assistant assistant assistant assistant assistant assistant assistant assistant assistant
T5assistant assistant assist assistant
BARTA assistant assist is an assistant assist, assistants assist to assist assistants.
ChatGPTAssistant ants assist ants in carrying leaves to the ant hill.
GoldenIf I assist a sister-assistant, will the sister's sister-assistant assist me?
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.867, + 0.489, + 0.897 + ], + "angle": 0, + "content": "Table 6: Example outputs for the input \"assistant assist\". \"Golden\" refers to the human-authored tongue twisters." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.084, + 0.638, + 0.101 + ], + "angle": 0, + "content": "5 Case Study" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.119, + 0.886, + 0.424 + ], + "angle": 0, + "content": "Within the example in Table 6, GPT-2 resorts to simply repeating the input, successfully achieving phonetic overlap, but failing to be grammatically valid or particularly sophisticated. This pattern is also demonstrated by DialogGPT and T5. Conversely, BART is able to introduce tokens unseen in the input to create an almost grammatically valid output (the primary mistake being indefinite article agreement, where in the first instance \"an\" would have been correct, rather than \"a\"). BART's output is also semantically and logically coherent, with \"A assistant assist is an assistant assist\" being valid (yet redundant), and \"assistants assist to assist assistants\" also being comprehensible. This example demonstrates why evaluators with high English proficiency and language/linguistics education were selected, as the same word may have different parts of speech, creating outputs that seem grammatically invalid, but do actually follow the rules of English.[11]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.425, + 0.886, + 0.667 + ], + "angle": 0, + "content": "Further investigation is needed to ascertain whether the models are intentionally exploiting this lexical ambiguity, or if human evaluators are demonstrating apophobia, where patterns are found in what is effectively noise (Brugger, 2001). Finally, ChatGPT utilises morphology to exploit the similarity of the plural noun \"assistants\" and the phrase \"assist ants\", and provides a continuation that is in line with the expected behaviour of ants. In comparison to the golden example, ChatGPT's output may be considered more interesting topic-wise, at the expense of not being as phonetically complex (\"carrying leaves to the ant hill\" contributes heavily to semantics, whilst not being recognisable as part of a tongue twister). For further analysis, please see Appendix A.4." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.69, + 0.638, + 0.706 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.725, + 0.886, + 0.87 + ], + "angle": 0, + "content": "We present work on the topic of tongue twister generation, a form of phonetically-constrained language generation that aims to maximise phonetic overlap, whilst conveying meaningful semantics. We motivate the potential application domains for such generated language, and provide a large annotated dataset of tongue twisters, TwistList, to encourage further work. Finally, we present a series of benchmark models alongside automatic/human evaluation to assess generation quality." + }, + { + "type": "page_footnote", + "bbox": [ + 0.509, + 0.892, + 0.871, + 0.918 + ], + "angle": 0, + "content": "11https://en.wikipedia.org/wiki/Buffalo_buffalo_Buffalo_buffalo_buffalo_buffalo_buffalo_buffalo" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.928, + 0.516, + 0.941 + ], + "angle": 0, + "content": "583" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.115, + 0.084, + 0.214, + 0.099 + ], + "angle": 0, + "content": "Limitations" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.115, + 0.49, + 0.42 + ], + "angle": 0, + "content": "Whilst the system presented within this paper is capable of allowing human-in-the-loop contributions (via selecting the input keywords on which to condition the output), it is not able to produce tongue-twisters that take advantage of particular features of speech sounds such as place and manner of articulation, in order to create more advanced outputs that exploit phonetic relatedness (rather than exact matches). The same can be said of our proposed metrics, PO and Init-PO, which do not account for phonetic similarity across sounds that share manner/place of articulation (e.g. \"she sells sea shells\"). Additionally, whilst commonly known tongue twisters may follow a particular format (e.g. rhyme schemes), such schemes and templates have not been enforced here. We also do not demonstrate the capabilities of these systems if they were trained on phonetic transcriptions explicitly, as we only aim to assess their performance when training on graphemes in standard orthography." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.439, + 0.258, + 0.454 + ], + "angle": 0, + "content": "Ethics Statement" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.469, + 0.489, + 0.758 + ], + "angle": 0, + "content": "All use of human participants in this study has been approved by the Ethics Board of the primary author's institution, including the disclosure of demographic information. Regarding the generation of tongue twisters, language generation is a necessarily creative domain that has the ability to reproduce content that some individuals may find offensive. Care was taken to check outputs in the human evaluation set for any such materials, and if they had been produced, they would have been removed from the evaluation set. Additionally, no egregiously offensive material has been provided in the TwistList dataset. However, the distinction between offensive and humorous content is a highly complex topic, and therefore some examples within the dataset may not be suitable for all individuals (e.g. suggestive content and swearing, such as \"I'm not the pheasant plucker, I'm the pheasant plucker's son\", and the clear relation to common expletives)." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.777, + 0.279, + 0.793 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.807, + 0.489, + 0.904 + ], + "angle": 0, + "content": "Tyler Loakman is supported by the Centre for Doctoral Training in Speech and Language Technologies (SLT) and their Applications funded by UK Research and Innovation [grant number EP/S023062/1]. Chen Tang is supported by the China Scholarship Council (CSC) for his doctoral study (File No.202006120039)." + }, + { + "type": "title", + "bbox": [ + 0.511, + 0.084, + 0.605, + 0.099 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.107, + 0.883, + 0.174 + ], + "angle": 0, + "content": "Rajat Agarwal and Katharina Kann. 2020. Acrostic poem generation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1230-1240, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.183, + 0.885, + 0.262 + ], + "angle": 0, + "content": "Peter Brugger. 2001. From haunted brain to haunted science: A cognitive neuroscience view of paranormal and pseudoscientific thought. In James Hournan and RenseEditors Lange, editors, *Hauntings and Poltergeists: Multidisciplinary Perspectives*, page 195-213. McFarland." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.272, + 0.883, + 0.299 + ], + "angle": 0, + "content": "Joseph L Fleiss. 1971. Measuring nominal scale agreement among many raters. Psychological bulletin, 76(5):378." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.309, + 0.883, + 0.336 + ], + "angle": 0, + "content": "Theodore Seuss Geisel. 1965. *Fox in socks: Dr. Seuss's book of tongue tanglers*. Random House." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.345, + 0.885, + 0.438 + ], + "angle": 0, + "content": "Marco Guerini, Gözde Özbal, and Carlo Strapparava. 2015. Echoes of persuasion: The effect of euphony in persuasive communication. In Proceedings of the 2015 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1483-1493, Denver, Colorado. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.447, + 0.884, + 0.54 + ], + "angle": 0, + "content": "He He, Nanyun Peng, and Percy Liang. 2019. Pun generation with surprise. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 1734-1744, Minneapolis, Minnesota. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.549, + 0.884, + 0.655 + ], + "angle": 0, + "content": "Henglin Huang, Chen Tang, Tyler Loakman, Frank Guerin, and Chenghua Lin. 2022. Improving Chinese story generation via awareness of syntactic dependencies and semantics. In Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 2: Short Papers)." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.664, + 0.883, + 0.717 + ], + "angle": 0, + "content": "Sedrick Scott Keh, Steven Y. Feng, Varun Gangal, Malihe Alikhani, and Eduard Hovy. 2022. Pancetta: Phoneme aware neural completion to elicit tongue twisters automatically." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.727, + 0.883, + 0.78 + ], + "angle": 0, + "content": "Heather Kember, Kathryn Connaghan, and Rupal Patel. 2017. Inducing speech errors in dysarthria using tongue twisters. International journal of language & communication disorders, 52(4):469-478." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.79, + 0.883, + 0.882 + ], + "angle": 0, + "content": "Jey Han Lau, Trevor Cohn, Timothy Baldwin, Julian Brooke, and Adam Hammond. 2018. Deep-spare: A joint neural model of poetic language, meter and rhyme. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1948-1958, Melbourne, Australia. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.892, + 0.883, + 0.919 + ], + "angle": 0, + "content": "Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy," + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.107, + 0.885, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.516, + 0.941 + ], + "angle": 0, + "content": "584" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.135, + 0.086, + 0.489, + 0.166 + ], + "angle": 0, + "content": "Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871-7880, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.178, + 0.487, + 0.231 + ], + "angle": 0, + "content": "Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81, Barcelona, Spain. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.244, + 0.487, + 0.297 + ], + "angle": 0, + "content": "Ken D. O'Halloran. 2020. A tongue-twister to translation? increased complexity of genioglossus movement during wakefulness in persons with obstructive sleep apnoea. The Journal of Physiology, 598(3):435-436." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.31, + 0.487, + 0.39 + ], + "angle": 0, + "content": "Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311-318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.403, + 0.487, + 0.455 + ], + "angle": 0, + "content": "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.469, + 0.487, + 0.535 + ], + "angle": 0, + "content": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(140):1-67." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.548, + 0.486, + 0.587 + ], + "angle": 0, + "content": "Victoria Somoff. 2014. Four is not fourteen: Tongue twister patterns and the unmastery of language. Western Folklore, 73(2/3):195-215." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.602, + 0.487, + 0.653 + ], + "angle": 0, + "content": "Prasetyawan Sugiharto, Yan Santoso, and Maila Shofyana. 2022. Teaching english pronunciation using tongue twister. Acitya: Journal of Teaching and Education, 4(1):189-197." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.667, + 0.487, + 0.719 + ], + "angle": 0, + "content": "Jiao Sun, Anjali Narayan-Chen, Shereen Oraby, Shuyang Gao, Tagyoung Chung, Jing Huang, Yang Liu, and Nanyun Peng. 2022. Context-situated pun generation. In EMNLP 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.733, + 0.487, + 0.799 + ], + "angle": 0, + "content": "Chen Tang, Chenghua Lin, Henglin Huang, Frank Guerin, and Zhihao Zhang. 2022a. EtrICA: Event-triggered context-aware story generation augmented by cross attention. In *Findings of the Association for Computational Linguistics: EMNLP* 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.813, + 0.487, + 0.853 + ], + "angle": 0, + "content": "Chen Tang, Hongbo Zhang, Tyler Loakman, Chenghua Lin, and Frank Guerin. 2022b. Terminology-aware medical dialogue generation. arXiv preprint arXiv:2210.15551." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.866, + 0.487, + 0.918 + ], + "angle": 0, + "content": "Chen Tang, Zhihao Zhang, Tyler Loakman, Chenghua Lin, and Frank Guerin. 2022c. NGEP: A graph-based event planning framework for story generation. In Proceedings of AACL-IJCNLP, Online." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.489, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.513, + 0.086, + 0.882, + 0.178 + ], + "angle": 0, + "content": "Yufei Tian and Nanyun Peng. 2022. Zero-shot sonnet generation with discourse-level planning and aesthetics features. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3587-3597, Seattle, United States. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.192, + 0.882, + 0.258 + ], + "angle": 0, + "content": "Tim Van de Cruys. 2020. Automatic poetry generation from prosaic text. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2471-2480, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.271, + 0.882, + 0.311 + ], + "angle": 0, + "content": "Carolyn E. Wilshire. 1999. The \"tongue twister\" paradigm as a technique for studying phonological encoding. Language and Speech, 42(1):57-82." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.325, + 0.882, + 0.443 + ], + "angle": 0, + "content": "Jörg Wöckener, Thomas Haider, Tristan Miller, The-Khang Nguyen, Thanh Tung Linh Nguyen, Minh Vu Pham, Jonas Belouadi, and Steffen Eger. 2021. End-to-end style-conditioned poetry generation: What does it take to learn from examples alone? In Proceedings of the 5th Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature, pages 57-66, Punta Cana, Dominican Republic (online). Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.456, + 0.882, + 0.522 + ], + "angle": 0, + "content": "Min Ney Wong, Yanky Chan, Manwa L. Ng, and Frank F. Zhu. 2019. Effects of transcranial direct current stimulation over the broca's area on tongue twister production. International Journal of Speech-Language Pathology, 21(2):182-188. PMID: 29642741." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.536, + 0.882, + 0.654 + ], + "angle": 0, + "content": "Lanqing Xue, Kaitao Song, Duocai Wu, Xu Tan, Nevin L. Zhang, Tao Qin, Wei-Qiang Zhang, and Tie-Yan Liu. 2021. DeepRapper: Neural rap generation with rhyme and rhythm modeling. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 69-81, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.667, + 0.882, + 0.747 + ], + "angle": 0, + "content": "Zhiwei Yu, Jiwei Tan, and Xiaojun Wan. 2018. A neural approach to pun generation. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1650-1660, Melbourne, Australia. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.76, + 0.882, + 0.853 + ], + "angle": 0, + "content": "Rongsheng Zhang, Xiaoxi Mao, Le Li, Lin Jiang, Lin Chen, Zhiwei Hu, Yadong Xi, Changjie Fan, and Minlie Huang. 2020a. Youling: an AI-assisted lyrics creation system. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 85-91, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.866, + 0.882, + 0.918 + ], + "angle": 0, + "content": "Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi. 2020b. *Bertscore: Evaluating text generation with bert*. In International Conference on Learning Representations." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.882, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.516, + 0.941 + ], + "angle": 0, + "content": "585" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.116, + 0.086, + 0.49, + 0.192 + ], + "angle": 0, + "content": "Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, and Bill Dolan. 2020c. DIALOGPT: Large-scale generative pre-training for conversational response generation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 270-278, Online. Association for Computational Linguistics." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.204, + 0.25, + 0.22 + ], + "angle": 0, + "content": "A Appendices" + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.23, + 0.343, + 0.244 + ], + "angle": 0, + "content": "A.1 Dataset Quality Control" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.251, + 0.489, + 0.283 + ], + "angle": 0, + "content": "An annotation platform was developed as shown in (Figure 2)." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.293, + 0.317, + 0.309 + ], + "angle": 0, + "content": "A.2 Human Participants" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.314, + 0.486, + 0.506 + ], + "angle": 0, + "content": "Due to tongue twisters being highly reliant on articulation abilities, the demographics of the human participants used within this work are highly important. Additionally, tongue twisters are also a form of humour and entertainment, where individual perceptions of what may or may not be considered humorous or entertaining differ according to numerous factors. In an effort to remain as transparent as possible, and follow best practices for human evaluation, relevant demographic information of participants are outlined below (with the necessary requisite permission and ethical approval)." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.516, + 0.489, + 0.629 + ], + "angle": 0, + "content": "Dataset Evaluation All evaluators involved in the quality control process of the TwistList dataset are native speakers of English, and either have or are working towards University level qualifications. Additionally, 2 of the 3 evaluators have extensive education in linguistics or modern languages. No monetary incentive was provided." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.637, + 0.49, + 0.782 + ], + "angle": 0, + "content": "Generation Evaluation All evaluators involved in the evaluation of the quality of generated tongue twisters are native speakers of English, and either hold or are working towards University level qualifications in Linguistics, Modern Languages or NLP. Additionally, all evaluators cited the United Kingdom as their country of socialisation, and no participants reported language processing difficulties that could affect results. No monetary incentive was provided." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.791, + 0.489, + 0.92 + ], + "angle": 0, + "content": "Materials Provided to Human Participants Additionally, all evaluators for both the dataset and generation outputs were presented with calibration examples to demonstrate the sort of outputs that would be presented, and the logic behind particular scores, in order to minimise individual interpretations of the scoring criteria. All evaluation was performed on a custom made online annotation platform (Figure 3)." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.085, + 0.681, + 0.1 + ], + "angle": 0, + "content": "A.3 Training Details" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.107, + 0.884, + 0.349 + ], + "angle": 0, + "content": "All pre-trained models used (naturally excluding ChatGPT) are based on publicly available checkpoints from Hugging Face.12 Models are trained for up to 5 epochs on a Tesla A5000 machine with the best checkpoints selected based on the validation loss. The batch size is set to 32, and the learning rate is \\(8e^{-5}\\), with the Adam optimiser selected for training. To help the loss curve converge on our small few-shot dataset, we limit the generation length to 100 (covering all test tongue twisters). Meanwhile, the source length is limited to 150. The training and testing steps are set up with the implementation of the PyTorch Lightning13 framework to guarantee the reliability of the experiment. All language models are fairly trained and tested with the same steps." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.361, + 0.791, + 0.375 + ], + "angle": 0, + "content": "A.4 Further Qualitative Comments" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.382, + 0.884, + 0.752 + ], + "angle": 0, + "content": "Whilst the pattern of extreme word repetition is seen in many of the finetuned models (often with the exception of BART, which is demonstrated to be capable of producing slightly more sophisticated outputs), overall assessment of the tongue twisters produced at inference time reveals interesting patterns, particularly in regard to ChatGPT outputs. Firstly, the limits of ChatGPT are made apparent in a few examples such as the input \"silver shiny ship sank\" generating \"How much wood would a woodchuck chuck if a woodchuck could chuck silver shiny ships?\", a clear derivation of a famous woodchuck related tongue twister that it is rather safe to assume appears multiple times in ChatGPTs training material. Additionally, comments from evaluators also reveal that ChatGPT's output is often considered more of a rhyme or general literary text, rather than specifically a tongue twister. However, examples such as these are also found in the human-authored golden examples, demonstrating that there is no steadfast consistent opinion as to what constitutes a (good) tongue twister. Likewise, some examples may contain large amounts of sound repetition, but not in a way that necessarily presents articulatory difficulty." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.765, + 0.665, + 0.779 + ], + "angle": 0, + "content": "A.5 Future Works" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.786, + 0.882, + 0.882 + ], + "angle": 0, + "content": "In this paper, we mainly analyse the performance of large-scale pretrained language models (PLMs) on Tongue Twister Generation, and propose a corresponding dataset for further investigation. In further works, we aim to propose novel models which can better leverage phonetic symbols. There" + }, + { + "type": "page_footnote", + "bbox": [ + 0.527, + 0.891, + 0.751, + 0.905 + ], + "angle": 0, + "content": "12https://huggingface.co/models" + }, + { + "type": "page_footnote", + "bbox": [ + 0.529, + 0.905, + 0.771, + 0.918 + ], + "angle": 0, + "content": "13https://www.pytorchlightning.ai/" + }, + { + "type": "list", + "bbox": [ + 0.527, + 0.891, + 0.771, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.517, + 0.941 + ], + "angle": 0, + "content": "586" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.139, + 0.085, + 0.308, + 0.095 + ], + "angle": 0, + "content": "Tongue Twister Dataset Evaluation" + }, + { + "type": "header", + "bbox": [ + 0.78, + 0.086, + 0.86, + 0.092 + ], + "angle": 0, + "content": "Home / Sample Annotation" + }, + { + "type": "title", + "bbox": [ + 0.139, + 0.104, + 0.185, + 0.111 + ], + "angle": 0, + "content": "Introduction" + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.121, + 0.362, + 0.127 + ], + "angle": 0, + "content": "1. Individually read the tongue twister, phonetics, and key words on the left side." + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.129, + 0.409, + 0.135 + ], + "angle": 0, + "content": "Select the options on the right side to evaluate the data quality from the following perspectives:" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.136, + 0.408, + 0.141 + ], + "angle": 0, + "content": "- The quality of the RAKE Keywords: Do these suitably represent the topic of the tongue twister?" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.142, + 0.419, + 0.147 + ], + "angle": 0, + "content": "- The quality of the BERTopic Keywords: Do these suitably represent the topic of the tongue twister?" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.148, + 0.435, + 0.155 + ], + "angle": 0, + "content": "- The quality of the Tongue Twister: Is it a good tongue twister, or too short/long, or generally bad quality?" + }, + { + "type": "list", + "bbox": [ + 0.145, + 0.121, + 0.435, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.146, + 0.168, + 0.195, + 0.175 + ], + "angle": 0, + "content": "Tongue Twisters" + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.188, + 0.243, + 0.195 + ], + "angle": 0, + "content": "Keywords for Tongue Twisters" + }, + { + "type": "text", + "bbox": [ + 0.158, + 0.208, + 0.207, + 0.214 + ], + "angle": 0, + "content": "RAKE Keywords:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.207, + 0.319, + 0.213 + ], + "angle": 0, + "content": "pickled peppers peter piper picked" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.214, + 0.412, + 0.221 + ], + "angle": 0, + "content": "[P I H1 K A H0 L D] [P E H1 P E R O Z] [P I Y1 T E R O] [P A Y1 P E R O] [P I H1 K T]" + }, + { + "type": "text", + "bbox": [ + 0.158, + 0.228, + 0.188, + 0.234 + ], + "angle": 0, + "content": "BERTopic" + }, + { + "type": "text", + "bbox": [ + 0.158, + 0.235, + 0.19, + 0.241 + ], + "angle": 0, + "content": "Keywords:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.229, + 0.283, + 0.234 + ], + "angle": 0, + "content": "pink peter piper peck" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.235, + 0.354, + 0.242 + ], + "angle": 0, + "content": "[P I H1NGK][P I Y1TERO][P AY1PER0][P E H1K]" + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.259, + 0.2, + 0.266 + ], + "angle": 0, + "content": "Tongue Twister" + }, + { + "type": "text", + "bbox": [ + 0.158, + 0.28, + 0.204, + 0.287 + ], + "angle": 0, + "content": "Tongue Twister" + }, + { + "type": "text", + "bbox": [ + 0.158, + 0.288, + 0.175, + 0.293 + ], + "angle": 0, + "content": "Text:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.279, + 0.474, + 0.299 + ], + "angle": 0, + "content": "Peter Piper picked a peck of pickled peppers. A peck of pickled peppers Peter Piper picked. If Peter Piper picked a peck of pickled peppers, Where's the peck of pickled peppers Peter Piper picked?" + }, + { + "type": "text", + "bbox": [ + 0.158, + 0.307, + 0.204, + 0.313 + ], + "angle": 0, + "content": "Tongue Twister" + }, + { + "type": "text", + "bbox": [ + 0.158, + 0.314, + 0.19, + 0.319 + ], + "angle": 0, + "content": "Phonetics:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.307, + 0.475, + 0.341 + ], + "angle": 0, + "content": "[PI Y1 ER0] [PY A1 ER0] [PI H1 K T] [AOH] [PE H1 K] [AHV] [PI I H1 K AI O L D] [PE H1 ER0 Z] [JI AHO] [PE H1 K] [AHV] [PI I H1 K AI O L D] [PE H1 ER0 Z] [PI Y1 ER0] [PY A1 ER0] [PI I H1 K T] [JI (IH1 F) [PI Y1 T ER0] [PY A1 ER0] [PI I H1 K T] [AOH] [PE H1 K] [AHV] [PI I H1 K AI O L D] [PE H1 ER0 Z] [WEH1 K R2] [DH AHO] [PE H1 K] [AHV] [PI I H1 K AI O L D] [PE H1 ER0 Z] [PI Y1 T ER0] [PY A1 ER0] [PI I H1 K T] [Q]:" + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.105, + 0.557, + 0.112 + ], + "angle": 0, + "content": "Sample Quality" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.124, + 0.713, + 0.132 + ], + "angle": 0, + "content": "A1: Are the RAKE Keywords highly suitable for the Tongue Twister?" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.141, + 0.54, + 0.146 + ], + "angle": 0, + "content": "Yes" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.147, + 0.54, + 0.153 + ], + "angle": 0, + "content": "No" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.173, + 0.726, + 0.18 + ], + "angle": 0, + "content": "A2: Are the BERTopic Keywords highly suitable for the Tongue Twister?" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.189, + 0.539, + 0.195 + ], + "angle": 0, + "content": "Yes" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.196, + 0.539, + 0.201 + ], + "angle": 0, + "content": "No" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.22, + 0.61, + 0.228 + ], + "angle": 0, + "content": "A3: Is it a good tongue twister?" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.237, + 0.546, + 0.242 + ], + "angle": 0, + "content": "Good" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.244, + 0.57, + 0.249 + ], + "angle": 0, + "content": "A Bit Too Short" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.25, + 0.568, + 0.255 + ], + "angle": 0, + "content": "A Bit Too Long" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.256, + 0.562, + 0.263 + ], + "angle": 0, + "content": "\\(\\bigcirc\\) Bad Quality" + }, + { + "type": "list", + "bbox": [ + 0.521, + 0.237, + 0.57, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.283, + 0.823, + 0.297 + ], + "angle": 0, + "content": "A4: Please input up to 5 manual keywords for the tongue twister (these can either be extracted from the example, or any other words)." + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.306, + 0.551, + 0.312 + ], + "angle": 0, + "content": "Keywords" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.35, + 0.53, + 0.357 + ], + "angle": 0, + "content": "Quit" + }, + { + "type": "text", + "bbox": [ + 0.826, + 0.35, + 0.849, + 0.357 + ], + "angle": 0, + "content": "Submit" + }, + { + "type": "image_caption", + "bbox": [ + 0.316, + 0.375, + 0.681, + 0.389 + ], + "angle": 0, + "content": "Figure 2: TwistList Quality Control Annotation Platform" + }, + { + "type": "title", + "bbox": [ + 0.139, + 0.404, + 0.266, + 0.414 + ], + "angle": 0, + "content": "Tongue Twister Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.781, + 0.405, + 0.86, + 0.412 + ], + "angle": 0, + "content": "Home / Sample Annotation" + }, + { + "type": "title", + "bbox": [ + 0.144, + 0.423, + 0.182, + 0.43 + ], + "angle": 0, + "content": "Introduction" + }, + { + "type": "text", + "bbox": [ + 0.144, + 0.439, + 0.351, + 0.445 + ], + "angle": 0, + "content": "1. Individually read the input keywords and the tongue twister on the left side." + }, + { + "type": "text", + "bbox": [ + 0.144, + 0.446, + 0.396, + 0.453 + ], + "angle": 0, + "content": "2. Give a score for each metric on the right to evaluate the quality of generated tongue twisters:" + }, + { + "type": "text", + "bbox": [ + 0.144, + 0.454, + 0.432, + 0.46 + ], + "angle": 0, + "content": "- Relevance: The extent to which the tongue twister is remantically/topically related to the input keywords." + }, + { + "type": "text", + "bbox": [ + 0.144, + 0.461, + 0.456, + 0.466 + ], + "angle": 0, + "content": "- Difficulty of Articulation: The extent to which the tongue twister is hard to say (aka. how much your tongue twists)." + }, + { + "type": "text", + "bbox": [ + 0.144, + 0.467, + 0.4, + 0.473 + ], + "angle": 0, + "content": "- Fluency: The extent to which the tongue twister can be considered grammatically acceptable." + }, + { + "type": "text", + "bbox": [ + 0.144, + 0.474, + 0.434, + 0.48 + ], + "angle": 0, + "content": "- Coherence: The extent to which the tongue twister can be considered logically and semantically coherent." + }, + { + "type": "text", + "bbox": [ + 0.144, + 0.481, + 0.457, + 0.487 + ], + "angle": 0, + "content": "Entertainment: The extent to which the tongue twister is considered enteratining (primarily relating to phonetics +" + }, + { + "type": "text", + "bbox": [ + 0.144, + 0.488, + 0.174, + 0.493 + ], + "angle": 0, + "content": "semantics)" + }, + { + "type": "list", + "bbox": [ + 0.144, + 0.439, + 0.457, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.144, + 0.506, + 0.192, + 0.513 + ], + "angle": 0, + "content": "Tongue Twisters" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.527, + 0.2, + 0.533 + ], + "angle": 0, + "content": "Input Keywords" + }, + { + "type": "text", + "bbox": [ + 0.157, + 0.546, + 0.204, + 0.552 + ], + "angle": 0, + "content": "RAKE Keywords:" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.546, + 0.352, + 0.552 + ], + "angle": 0, + "content": "Generate tongue twisters about key words: nope" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.57, + 0.198, + 0.576 + ], + "angle": 0, + "content": "Tongue Twister" + }, + { + "type": "text", + "bbox": [ + 0.157, + 0.59, + 0.202, + 0.602 + ], + "angle": 0, + "content": "Tongue Twister Text:" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.589, + 0.353, + 0.595 + ], + "angle": 0, + "content": "Nope, an antelope can't elope with a cantaloupe." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.596, + 0.353, + 0.602 + ], + "angle": 0, + "content": "" + }, + { + "type": "title", + "bbox": [ + 0.507, + 0.422, + 0.698, + 0.431 + ], + "angle": 0, + "content": "Evaluate the Quality of Tongue Twisters by giving a Score (1 to 5)" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.443, + 0.646, + 0.45 + ], + "angle": 0, + "content": "Please select all options before submission." + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.459, + 0.562, + 0.464 + ], + "angle": 0, + "content": "O1:Relevance." + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.466, + 0.539, + 0.472 + ], + "angle": 0, + "content": "#" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.483, + 0.603, + 0.488 + ], + "angle": 0, + "content": "O2: Difficulty of articulation." + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.494, + 0.539, + 0.499 + ], + "angle": 0, + "content": "0" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.508, + 0.557, + 0.513 + ], + "angle": 0, + "content": "Q3:FLuency." + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.521, + 0.539, + 0.526 + ], + "angle": 0, + "content": "0" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.533, + 0.563, + 0.538 + ], + "angle": 0, + "content": "Q4: Coherence." + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.546, + 0.539, + 0.551 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.557, + 0.575, + 0.562 + ], + "angle": 0, + "content": "Q5: Entertainment." + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.571, + 0.539, + 0.576 + ], + "angle": 0, + "content": "。" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.601, + 0.529, + 0.607 + ], + "angle": 0, + "content": "Quit" + }, + { + "type": "text", + "bbox": [ + 0.824, + 0.601, + 0.849, + 0.607 + ], + "angle": 0, + "content": "Submit" + }, + { + "type": "image_caption", + "bbox": [ + 0.302, + 0.627, + 0.695, + 0.642 + ], + "angle": 0, + "content": "Figure 3: Human Evaluation Platform for Generated Outputs" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.667, + 0.489, + 0.844 + ], + "angle": 0, + "content": "are numerous existing works (Huang et al., 2022; Tang et al., 2022a,b) that provide approaches for injecting such knowledge into PLMs. However, the phonetic features differ from these text-format knowledge items, as phonemes are hard to encode with input text tokens when feeding into PLM encoders. Another promising approach is to explicitly model the phonetic features into text sequences (Tang et al., 2022c), though there is no observed method for transforming phonetic notation. We intend to perform further research based on these existing approaches." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.516, + 0.941 + ], + "angle": 0, + "content": "587" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.134, + 0.085, + 0.435, + 0.1 + ], + "angle": 0, + "content": "ACL 2023 Responsible NLP Checklist" + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.108, + 0.323, + 0.123 + ], + "angle": 0, + "content": "A For every submission:" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.127, + 0.534, + 0.144 + ], + "angle": 0, + "content": "A1. Did you describe the limitations of your work?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.145, + 0.85, + 0.16 + ], + "angle": 0, + "content": "Yes, in the required Limitations section as well as Section 4 (concerning our proposed metrics)" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.169, + 0.554, + 0.187 + ], + "angle": 0, + "content": "A2. Did you discuss any potential risks of your work?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.188, + 0.28, + 0.201 + ], + "angle": 0, + "content": "Ethics Statement" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.213, + 0.697, + 0.23 + ], + "angle": 0, + "content": "A3. Do the abstract and introduction summarize the paper's main claims?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.231, + 0.673, + 0.246 + ], + "angle": 0, + "content": "Abstract (all) and contribution summary at the end of the introduction." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.257, + 0.67, + 0.273 + ], + "angle": 0, + "content": "A4. Have you used AI writing assistants when working on this paper?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.274, + 0.233, + 0.288 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.3, + 0.489, + 0.317 + ], + "angle": 0, + "content": "B Did you use or create scientific artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.322, + 0.36, + 0.337 + ], + "angle": 0, + "content": "TwistList dataset (Section 3.2)" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.347, + 0.531, + 0.364 + ], + "angle": 0, + "content": "B1. Did you cite the creators of artifacts you used?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.365, + 0.728, + 0.38 + ], + "angle": 0, + "content": "Sources of all entries in the dataset are credited in the .json file for each entry." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.391, + 0.78, + 0.407 + ], + "angle": 0, + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.408, + 0.882, + 0.503 + ], + "angle": 0, + "content": "We did not discuss the licensing around our dataset. The dataset uses works that are freely available on the web and come from various sources such as websites, blogs, and eBooks. Many of these cases are Public Domain, and for those that are not, we believe we are in accordance with Fair Use, as the dataset does not encroach on the use case of the original works (no graphic design/other elements are maintained) and the dataset is for use as a research tool only. We will also reply promptly to any cases of copyright infringement that relevant copyright holders make us aware of." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.514, + 0.882, + 0.579 + ], + "angle": 0, + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.58, + 0.286, + 0.593 + ], + "angle": 0, + "content": "See answer to B2." + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.605, + 0.882, + 0.653 + ], + "angle": 0, + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.654, + 0.882, + 0.686 + ], + "angle": 0, + "content": "See the Ethics Statement regarding the potential for tongue twisters to be offensive. Additionally, all tongue twisters are believed to be about fictional characters, rather than individuals." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.697, + 0.882, + 0.729 + ], + "angle": 0, + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.73, + 0.882, + 0.777 + ], + "angle": 0, + "content": "Such details are not explicitly stated. However, it can be easily ascertained from the paper that the tongue twisters we focus on are entirely in English (and the range of domains the tongue twisters were taken from can be seen in the \"source\" entry for each example)." + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.787, + 0.882, + 0.868 + ], + "angle": 0, + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be." + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.869, + 0.395, + 0.884 + ], + "angle": 0, + "content": "See Table 1 for dataset statistics." + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.888, + 0.878, + 0.913 + ], + "angle": 0, + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.516, + 0.941 + ], + "angle": 0, + "content": "588" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.115, + 0.084, + 0.495, + 0.101 + ], + "angle": 0, + "content": "C Did you run computational experiments?" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.106, + 0.273, + 0.122 + ], + "angle": 0, + "content": "Section 4 (page 3)" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.131, + 0.88, + 0.163 + ], + "angle": 0, + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.165, + 0.255, + 0.181 + ], + "angle": 0, + "content": "Appendix A.3" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.19, + 0.882, + 0.223 + ], + "angle": 0, + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.224, + 0.254, + 0.24 + ], + "angle": 0, + "content": "Appendix A.3" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.249, + 0.882, + 0.297 + ], + "angle": 0, + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.299, + 0.503, + 0.314 + ], + "angle": 0, + "content": "Tables 3/5. Scores are the mean, as is standard." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.325, + 0.882, + 0.372 + ], + "angle": 0, + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.374, + 0.882, + 0.422 + ], + "angle": 0, + "content": "Exact details of evaluation implementations (except Phonetic Overlap) were not detailed. This is in part due to these metrics (BLEU/ROUGE/BERTScore) not being very reliable for creative language generation, and therefore the exact values from different implementations are not likely to be of use." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.431, + 0.876, + 0.449 + ], + "angle": 0, + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.454, + 0.538, + 0.469 + ], + "angle": 0, + "content": "Section 3.2 and Section 4. In addition to Appendix A.2" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.479, + 0.882, + 0.511 + ], + "angle": 0, + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.513, + 0.805, + 0.528 + ], + "angle": 0, + "content": "Screenshot of the annotation platforms can be found in Figures 2 and 3 in the Appendix" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.538, + 0.882, + 0.586 + ], + "angle": 0, + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.588, + 0.882, + 0.635 + ], + "angle": 0, + "content": "We declared that no monetary incentive was given to participants. We did not specify the recruitment process, but due to participants all holding or working towards university level qualifications, it can be inferred that they are colleagues." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.646, + 0.882, + 0.693 + ], + "angle": 0, + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.695, + 0.882, + 0.758 + ], + "angle": 0, + "content": "This information was not deemed necessary in the submitted paper (due to the limited risk of the data we were working with). However, it is stated in the Ethical Statement and Appendix A.2 that all shared information about human demographics was collected with the necessary permissions and approval." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.768, + 0.876, + 0.785 + ], + "angle": 0, + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.787, + 0.882, + 0.816 + ], + "angle": 0, + "content": "Ethical approval was gained for human evaluation of the dataset and generated outputs from the relevant institution" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.828, + 0.881, + 0.859 + ], + "angle": 0, + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.861, + 0.724, + 0.877 + ], + "angle": 0, + "content": "We provide demographic information for human participants in Appendix A.2" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.516, + 0.941 + ], + "angle": 0, + "content": "589" + } + ] +] \ No newline at end of file diff --git a/2023/TwistList_ Resources and Baselines for Tongue Twister Generation/450a8be3-d9ba-4e62-a952-97c7e24dcba6_origin.pdf b/2023/TwistList_ Resources and Baselines for Tongue Twister Generation/450a8be3-d9ba-4e62-a952-97c7e24dcba6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b1145dff6f433d6723f3573e3f8b06266a1fcce9 --- /dev/null +++ b/2023/TwistList_ Resources and Baselines for Tongue Twister Generation/450a8be3-d9ba-4e62-a952-97c7e24dcba6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0e38a479ef9be6bf5b912bab898946542cf9229c6cbe60627f3b7bef1f866df2 +size 515626 diff --git a/2023/TwistList_ Resources and Baselines for Tongue Twister Generation/full.md b/2023/TwistList_ Resources and Baselines for Tongue Twister Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a5ff871b381b1f9630f15c8d2afa28130c039796 --- /dev/null +++ b/2023/TwistList_ Resources and Baselines for Tongue Twister Generation/full.md @@ -0,0 +1,419 @@ +# TwistList: Resources and Baselines for Tongue Twister Generation + +Tyler Loakman $^{1*}$ , Chen Tang $^{2*}$ and Chenghua Lin $^{1\dagger}$ + +$^{1}$ Department of Computer Science, The University of Sheffield, UK + +$^{2}$ Department of Computer Science, The University of Surrey, UK + +{tcloakman1,c.lin}@sheffield.ac.uk + +chen.tang@surrey.ac.uk + +# Abstract + +Previous work in phonetically-grounded language generation has mainly focused on domains such as lyrics and poetry. In this paper, we present work on the generation of tongue twisters - a form of language that is required to be phonetically conditioned to maximise sound overlap, whilst maintaining semantic consistency with an input topic, and still being grammatically correct. We present TwistList, a large annotated dataset of tongue twisters, consisting of $2.1\mathrm{K}+$ human-authored examples. We additionally present several benchmark systems (referred to as TwisterMisters) for the proposed task of tongue twister generation, including models that both do and do not require training on in-domain data. We present the results of automatic and human evaluation to demonstrate the performance of existing mainstream pre-trained models in this task with limited (or no) task specific training and data, and no explicit phonetic knowledge. We find that the task of tongue twister generation is challenging for models under these conditions, yet some models are still capable of generating acceptable examples of this language type. + +# 1 Introduction + +Phonetically constrained language generation is a primary subarea of computational creativity in natural language generation (NLG), primarily encompassing lyric and poetry generation (Tian and Peng, 2022; Wöckener et al., 2021; Xue et al., 2021; Zhang et al., 2020a; Agarwal and Kann, 2020), as well as pun generation (Sun et al., 2022; He et al., 2019; Yu et al., 2018), and continues to prove challenging for myriad reasons. Primarily, such works require the inclusion of phonetic factors such as metre and rhyme, which involves careful consideration of candidate vocabulary on the syllable level, leading to a reduced pool of allowable vocabulary once these constraints are in place. + +![](images/df922e0bf87dd9a82794c843e2cecb8fc093ef3f8eecea4a89ad98ced56b52a3.jpg) +Figure 1: Tongue Twister Generation aims to generate an utterance with high levels of phonetic overlap, requiring understanding of semantics, grammar, and phonetics. + +In this paper, we present work on the generation of tongue twisters, a type of phonetically constrained language that is rarely explored in the NLG community. As a form of creative generation, tongue twisters can facilitate numerous useful applications, including: (1) being used as a pedagogical tool (Sugiharto et al., 2022; Somoff, 2014; Wilshire, 1999); (2) as a source of humorous entertainment stemming from unintentional mispronunciations; (3) as a stylistic device for engaging children in reading (e.g. Dr. Seuss stories (Geisel, 1965)); (4) as a method of designing memorable slogans and tag lines (Guerini et al., 2015); and (5) as stimuli in neuroscience/physiology research (Wong et al., 2019; O'Halloran, 2020; Kember et al., 2017). + +Tongue twister generation posits unique challenges compared to other generation tasks. One of the most pertinent features of tongue twisters is the presence of high levels of phonetic overlap across tokens (Wilshire, 1999). Consequently, whilst other types of creative generation may require only some output tokens to consider phonetics (such as rhyme or syllable counts), tongue twisters present an extreme version of this problem where the phonetics of almost all generated tokens must be considered. This leads to a very small vocabulary from which to choose + +semantically relevant words, and presents further challenges with maintaining grammatical validity. + +The only work that we are aware of on tongue twister generation at the time of conducting this research is by Keh et al. (2022), who present models that train on graphemes and phonemes, and take either a starting prompt to be continued, or keywords around which to theme an output. They release TT-Corp, a dataset of 644 tongue twisters with parallel non-twister equivalents. We differentiate our work through the release of a dataset that is over $3\mathrm{x}$ larger and which has undergone substantial human quality control. Furthermore, we assess the results of a wider range of popular pre-trained models on this task, including ChatGPT, without explicit injection of phonetic knowledge due to the difficulty in encoding phonetics and the expertise required to utilise phonetic characteristics appropriately. Our experimental results show that most popular pretrained language models (PLMs) rely on pure word repetition to generate tongue twisters, whilst some (i.e. BART) are able to generate more sophisticated examples. Additionally, very large zero-shot models (i.e. ChatGPT) are able to generate convincing tongue twisters almost on-par with human equivalents. $^{1}$ + +To summarise our contributions, we present: + +- TwistList, a large annotated dataset of human-authored tongue twisters, containing $2.1\mathrm{K}+$ examples with human evaluation of their quality. +- TwisterMisters, a series of baseline models for tongue twister generation using the most popular state-of-the-art PLMs. +- Extensive automatic and human evaluation to assess the ability of PLMs to implicitly model the complex phonetic phenomena in tongue twisters. + +# 2 Related Works + +Previous work in phonetically constrained generation has taken one of two approaches: 1) train a generation model on a collection of in-domain texts, or 2) train a generation model on prosaic out-of-domain text, with constraints imposed at decoding time. For example, Lau et al. (2018) collect 3,355 sonnets to produce novel poetry and train models to generate text in iambic pentameter, whilst Xue et al. (2021) train a rap generation model on 272,839 in-domain examples, infusing knowledge of rhythm afterwards. On the other hand, Van de Cruys (2020) train on a subset of CommonCrawl, imposing constraints on topic and + +
DatasetTrainValTestTotal
# Tongue Twisters19121061072128
Vocabulary Size955694688010358
# Total Phonemes55434656
# RAKE Keywords33333162883567
# BERTopic Keywords250132160250
Avg. # Input Keywords (RAKE)3.163.323.013.16
Avg. # Input Phonemes5.575.835.165.56
Avg. Tongue Twister Length (Words)15.0116.5913.5415.01
Avg. # Input Phonemes26.0628.2523.5026.04
+ +Table 1: The Statistics of TwistList. + +rhyme as a priori distributions, whilst Tian and Peng (2022) train a title-to-keyword module on narrative texts in addition to a sonnet generation model trained on news articles and short stories from Reddit. They imposed literary techniques (simile/metaphor) and metre/rhyme constraints at decoding time, owing to the lack of sufficient training data. $^2$ + +# 3 Tongue Twister Generation + +# 3.1 Task Definition + +We formulate the task of tongue twister generation as follows: for a given set of keywords, we aim to generate a tongue twister $T$ , whereby $T$ comprises a sequence of words $\{w_1, w_2, \dots, w_n\}$ . The generated output must satisfy the following constraints: (1) the output should be semantically related to the input keywords; (2) the output should show maximal levels of phonetic overlap across tokens; and (3) the output should be grammatically valid (Wilshire, 1999). Of these requirements, phonetic overlap is the most central to defining text as a "tongue twister". + +# 3.2 TwistList Dataset + +Dataset Construction. We present TwistList, an annotated dataset of $2.1\mathrm{K}+$ human-authored tongue twisters for use by the community. The examples contained therein come from a variety of sources available on the web. For each tongue twister, phonetic transcription is provided using the g2p-en package, in addition to keywords extracted with RAKE and BERTopic to represent the topic of the tongue twister. Following experimentation with both RAKE and BERTopic, only RAKE keywords are used in training due to human preference and issues regarding the use of BERTopic on short texts (where + +frequently no keywords are extracted). The main statistics of the dataset are presented in Table 1. + +
RAKE:sells thick socks
BERTopic:short shorts socks sock
Twister:Seth at Sainsbury's sells thick socks.
Phonetics:[S EH1 TH] [AE1 T] [S EY1 N S B ER0 IY0 Z] [S EH1 L Z] [TH IH1 K] [S AA1 K S]
+ +Table 2: Example from TwistList + +Quality Control. Quality control on our dataset was performed in multiple ways. Firstly, it was ensured that only sufficiently unique tongue twisters were kept in the dataset, as determined by removing examples with over $90\%$ word overlap (rather than keeping variants of the same tongue twister, such as "Peter Piper picked a pickled pepper" versus "Peter the Piper picked..."). Additionally, non-standard spellings were manually converted to standard US English5 to avoid G2P issues.6 Similarly, tongue-twisters containing obscure vocabulary (such as medicine and dinosaur names) were excluded to further minimise errors. An annotation platform was developed (see Appendix A.1), with which 3 human evaluators, who are native speakers of English, were provided with 100 sampled instances from the dataset to rate the quality of the resulting tongue twisters and the associated extracted keywords. The full dataset contains $2,500+$ tongue twisters, of which 2,128 are kept for training/development/testing after filtering examples with insufficient extracted keywords and excessive similarity to existing entries. + +To summarise, 3 annotators evaluated the quality of the dataset, where $88\%$ of assessed tongue twisters were considered high quality, and $6\%$ considered "suitable" (Kappa $= 0.321$ ). An example from TwistList is provided in Table 2. As Table 4 shows, the final dataset can be considered high quality, owing to fair/moderate levels of approval and agreement across evaluators. Demographic information of the evaluators can be found in Appendix A.2. + +# 3.3 Baseline Models + +We present the following baseline models (dubbed TwisterMisters) for the task of tongue twister generation on our TwistList dataset: + +Finetuned Baselines. For the finetuned baselines, we chose popular models for language generation, including GPT-2 (Radford et al., 2019), DialogGPT (Zhang et al., 2020c), T5 (Raffel et al., 2020), and BART (Lewis et al., 2020). These were finetuned with RAKE keywords extracted from human-authored tongue twisters as the input and the tongue twister text from TwistList as the target. This was in order to represent our baselines training on in-domain data. At inference time, the prompt "Generate tongue twisters about the keyword(s): X" is used, where X refers to the input consisting of one or more RAKE keywords extracted from tongue twisters. The full training details are given in Appendix A.3. We also conducted experiments on all aforementioned baselines without finetuning (i.e., a zero-shot setting), and the results were very poor. Therefore, we did not include these results in the paper. + +Training-Free Baseline We additionally provide a TwisterMister baseline that does not require any training. We utilise OpenAI's ChatGPT7 with the same prompt as a zero-shot setting for generation.8 Each request to ChatGPT was submitted as part of a separate session, to avoid the effects of extended dialogue influencing outputs. ChatGPT has been utilised in order to set a practical upper-bound of what may be expected from models without explicit phonetic knowledge, owing to its wealth of training data and 175B parameter architecture.9 It is assumed that ChatGPT's training data contains tongue twisters, and therefore it is able to abstract away the general patterns of such language in order to provide novel examples (though most likely based on graphemes rather than phonemes). + +# 4 Experiments + +Automatic Evaluation. We present the results of automatic evaluation on generated outputs and golden examples in Table 3 for the following metrics: Perplexity (PPL), BLEU (B-1/B-2) (Papineni et al., 2002), ROUGE (R-1/R-2/R-L) (Lin, 2004), and BERTScore Precision, Recall, and F-Measure (Zhang + +
ModelPPL↓B-1↑B-2↑R-1↑R-2↑R-L↑PO↓Init-PO↓BS-P↑BS-R↑BS-F↑
GPT-28.400.0070.0031.3010.1231.3150.0220.0200.6900.8100.744
DialoGPT3.830.0380.0257.7243.6107.6400.0690.0890.7540.8310.790
T510.160.0570.0389.7014.5739.5740.6890.7270.7950.8180.806
BART1.650.0730.05111.8836.10910.3530.0750.1200.7950.8450.819
ChatGPTN/A0.2000.13736.76520.65933.4370.0930.1570.8880.8940.883
+ +Table 3: Results of Automatic Evaluation. Golden PO = 0.385 and Golden Init-PO = 0.417. Due to the one-to-many issue in creative language generation, we acknowledge that the referenced metrics are imperfect. + +
Choices (%)Sample Quality
High.Suitable.Bad.Kappa
RAKE keywords82.018.00.00.321
BERTopic keywords15.085.00.00.445
Tongue Twisters88.06.04.00.321
+ +Table 4: Kappa refers to Fleiss' Kappa (Fleiss, 1971). All results achieve fair or moderate agreement. Good tongue twisters that are deemed a bit longer (3%) or shorter (3%) than expected are considered "suitable". + +et al., 2020b) (BS-P/BS-R/BS-F). PPL, BLEU and ROUGE are standard metrics in language generation to assess quality, whilst BERTScore assesses semantic similarity to a gold reference. Additionally, we propose two new metrics, Phonetic Overlap (PO) and Initial Phonetic Overlap (Init-PO). PO refers to the average overlap of all phonemes across tokens (# unique phonemes/#totalphonemes),whereas Init-PO is the ratio of unique word-initial phonemes to the number of words (# unique word-initial phonemes/#words). + +These phonetic metrics reward longer outputs. We argue that, all things equal, a longer tongue twister is better than a shorter one as it provides more entertainment and more opportunities for mispronunciation. Perfect scores on PO and Init-PO can be achieved by repetition of a single word. Whilst this does not lead to high quality outputs, these metrics are intended exclusively to be indicators of the phonetics, rather than an overall guide to quality. In both cases, higher levels of overlap results in lower ("better") scores, and the highest ("worst") achievable score is 1. + +The results in Table 3 show rather clear scaling, with the performance ranking on most metrics (except Perplexity and phoneme overlap) being identical. On the models explicitly finetuned for this task, GPT-2 is shown to be the worst, whilst BART performs the best. We hypothesise that GPT-2's poor performance is in part due to its simple causal language modelling objective alongside its decoder-only architecture (which is also in DialogGPT). Furthermore, whilst T5 performed well on the automatic metrics, manual + +inspection revealed that T5 often misinterpreted the task from the prompt, choosing to select its own keywords from the entire prompt, rather than using only the provided keyword list. On the other hand, the training-free zero-shot model, ChatGPT, was shown to perform best on all metrics. This is to be expected as ChatGPT has over 50x more parameters than any other tested PLM, with various pre-training objectives and reinforcement learning, leading to performant zero-shot capabilities. This further demonstrates that PLMs struggle to learn phonetic patterns implicitly from text, especially in English, which has high levels of irregular orthography. Furthermore, with limited data, PLMs struggle to learn the unusual probability distributions underlying tongue twisters, where word choices are intentionally "twisted", obscure, and anti-euphonious. Additionally, due to the wealth of training data seen by ChatGPT, it is likely that many examples have been seen during training. + +Human Evaluation. Due to tongue twisters being a creative domain where articulation abilities are tested, we also perform human evaluation. 3 evaluators were asked to rate 100 outputs from the best performing standard baseline (BART), in addition to ChatGPT outputs and gold examples from TwistList on the following criteria: Relevance (how relevant the tongue twister is given the keyword inputs), Fluency (how grammatically valid the output is), Difficulty of Articulation (how difficult a tongue twister is to say), Cohesion (how much sense the output makes), and Entertainment Value (how entertaining the output is, considering sounds and semantics). All ratings were on a 5-point Likert scale. Evaluator demographics and training materials are in Appendix A.2. + +The mean scores of human evaluation (Table 5) fall in line with expectations, with golden examples performing best on all metrics, and ChatGPT placing second on all but Difficulty of Articulation.[10] BART is able to produce outputs that are deemed to be the + +
Score (1 to 5)Human Evaluation
BARTChatGPTGolden
Relevance4.667*4.971†N/A
Difficulty of Articulation4.143*4.102*4.291*
Fluency3.028**4.915**4.938**
Coherence3.217*4.798*4.909*
Entertainment Value3.269*4.070*4.254*
+ +second most difficult to articulate, which we infer may be the result of slight morphological variants of input keywords being used repeatedly, making distinguishing between them during articulation quite challenging (whilst not being able to exploit deeper phonetic relations). The moderate score on Fluency (3.028) suggests instances of poor grammar may also hinder articulation abilities when expected grammatical structures are not found, leading to an interaction between grammatical validity and articulatory difficulty. Additionally, ChatGPT scoring the lowest for articulatory difficulty may be due to occasionally misunderstanding the requirements of a tongue twister, sometimes producing rhymes or standard prose (see Appendix A.4). However, ChatGPT scores well for Relevance and Fluency, highlighting its capability in producing high-quality coherent language. Perhaps most interestingly, none of the BART score averages on any human evaluation criteria fall below 3 ("neither agree nor disagree"). This performance is therefore quite good for a model finetuned on only 2128 examples, with no additional phonetic knowledge. + +Table 5: Results of Human Evaluation. The best scores are in bold, and the second best are underlined. We calculate Fleiss' Kappa for each metric, and mark the agreement fair*, moderate** and substantial†. + +
Inputassistant assistant assist
GPT-2assistant assistant assist assistant assist assistant
DialogGPTassistant assistant assistant assistant assistant assistant assistant assistant assistant assistant assistant assistant
T5assistant assistant assist assistant
BARTA assistant assist is an assistant assist, assistants assist to assist assistants.
ChatGPTAssistant ants assist ants in carrying leaves to the ant hill.
GoldenIf I assist a sister-assistant, will the sister's sister-assistant assist me?
+ +Table 6: Example outputs for the input "assistant assist". "Golden" refers to the human-authored tongue twisters. + +# 5 Case Study + +Within the example in Table 6, GPT-2 resorts to simply repeating the input, successfully achieving phonetic overlap, but failing to be grammatically valid or particularly sophisticated. This pattern is also demonstrated by DialogGPT and T5. Conversely, BART is able to introduce tokens unseen in the input to create an almost grammatically valid output (the primary mistake being indefinite article agreement, where in the first instance "an" would have been correct, rather than "a"). BART's output is also semantically and logically coherent, with "A assistant assist is an assistant assist" being valid (yet redundant), and "assistants assist to assist assistants" also being comprehensible. This example demonstrates why evaluators with high English proficiency and language/linguistics education were selected, as the same word may have different parts of speech, creating outputs that seem grammatically invalid, but do actually follow the rules of English.[11] + +Further investigation is needed to ascertain whether the models are intentionally exploiting this lexical ambiguity, or if human evaluators are demonstrating apophobia, where patterns are found in what is effectively noise (Brugger, 2001). Finally, ChatGPT utilises morphology to exploit the similarity of the plural noun "assistants" and the phrase "assist ants", and provides a continuation that is in line with the expected behaviour of ants. In comparison to the golden example, ChatGPT's output may be considered more interesting topic-wise, at the expense of not being as phonetically complex ("carrying leaves to the ant hill" contributes heavily to semantics, whilst not being recognisable as part of a tongue twister). For further analysis, please see Appendix A.4. + +# 6 Conclusion + +We present work on the topic of tongue twister generation, a form of phonetically-constrained language generation that aims to maximise phonetic overlap, whilst conveying meaningful semantics. We motivate the potential application domains for such generated language, and provide a large annotated dataset of tongue twisters, TwistList, to encourage further work. Finally, we present a series of benchmark models alongside automatic/human evaluation to assess generation quality. + +# Limitations + +Whilst the system presented within this paper is capable of allowing human-in-the-loop contributions (via selecting the input keywords on which to condition the output), it is not able to produce tongue-twisters that take advantage of particular features of speech sounds such as place and manner of articulation, in order to create more advanced outputs that exploit phonetic relatedness (rather than exact matches). The same can be said of our proposed metrics, PO and Init-PO, which do not account for phonetic similarity across sounds that share manner/place of articulation (e.g. "she sells sea shells"). Additionally, whilst commonly known tongue twisters may follow a particular format (e.g. rhyme schemes), such schemes and templates have not been enforced here. We also do not demonstrate the capabilities of these systems if they were trained on phonetic transcriptions explicitly, as we only aim to assess their performance when training on graphemes in standard orthography. + +# Ethics Statement + +All use of human participants in this study has been approved by the Ethics Board of the primary author's institution, including the disclosure of demographic information. Regarding the generation of tongue twisters, language generation is a necessarily creative domain that has the ability to reproduce content that some individuals may find offensive. Care was taken to check outputs in the human evaluation set for any such materials, and if they had been produced, they would have been removed from the evaluation set. Additionally, no egregiously offensive material has been provided in the TwistList dataset. However, the distinction between offensive and humorous content is a highly complex topic, and therefore some examples within the dataset may not be suitable for all individuals (e.g. suggestive content and swearing, such as "I'm not the pheasant plucker, I'm the pheasant plucker's son", and the clear relation to common expletives). + +# Acknowledgements + +Tyler Loakman is supported by the Centre for Doctoral Training in Speech and Language Technologies (SLT) and their Applications funded by UK Research and Innovation [grant number EP/S023062/1]. Chen Tang is supported by the China Scholarship Council (CSC) for his doctoral study (File No.202006120039). + +# References + +Rajat Agarwal and Katharina Kann. 2020. Acrostic poem generation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1230-1240, Online. Association for Computational Linguistics. +Peter Brugger. 2001. From haunted brain to haunted science: A cognitive neuroscience view of paranormal and pseudoscientific thought. In James Hournan and RenseEditors Lange, editors, *Hauntings and Poltergeists: Multidisciplinary Perspectives*, page 195-213. McFarland. +Joseph L Fleiss. 1971. Measuring nominal scale agreement among many raters. Psychological bulletin, 76(5):378. +Theodore Seuss Geisel. 1965. *Fox in socks: Dr. Seuss's book of tongue tanglers*. Random House. +Marco Guerini, Gözde Özbal, and Carlo Strapparava. 2015. Echoes of persuasion: The effect of euphony in persuasive communication. In Proceedings of the 2015 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1483-1493, Denver, Colorado. Association for Computational Linguistics. +He He, Nanyun Peng, and Percy Liang. 2019. Pun generation with surprise. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 1734-1744, Minneapolis, Minnesota. Association for Computational Linguistics. +Henglin Huang, Chen Tang, Tyler Loakman, Frank Guerin, and Chenghua Lin. 2022. Improving Chinese story generation via awareness of syntactic dependencies and semantics. In Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 2: Short Papers). +Sedrick Scott Keh, Steven Y. Feng, Varun Gangal, Malihe Alikhani, and Eduard Hovy. 2022. Pancetta: Phoneme aware neural completion to elicit tongue twisters automatically. +Heather Kember, Kathryn Connaghan, and Rupal Patel. 2017. Inducing speech errors in dysarthria using tongue twisters. International journal of language & communication disorders, 52(4):469-478. +Jey Han Lau, Trevor Cohn, Timothy Baldwin, Julian Brooke, and Adam Hammond. 2018. Deep-spare: A joint neural model of poetic language, meter and rhyme. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1948-1958, Melbourne, Australia. Association for Computational Linguistics. +Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, + +Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871-7880, Online. Association for Computational Linguistics. +Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81, Barcelona, Spain. Association for Computational Linguistics. +Ken D. O'Halloran. 2020. A tongue-twister to translation? increased complexity of genioglossus movement during wakefulness in persons with obstructive sleep apnoea. The Journal of Physiology, 598(3):435-436. +Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311-318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics. +Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9. +Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(140):1-67. +Victoria Somoff. 2014. Four is not fourteen: Tongue twister patterns and the unmastery of language. Western Folklore, 73(2/3):195-215. +Prasetyawan Sugiharto, Yan Santoso, and Maila Shofyana. 2022. Teaching english pronunciation using tongue twister. Acitya: Journal of Teaching and Education, 4(1):189-197. +Jiao Sun, Anjali Narayan-Chen, Shereen Oraby, Shuyang Gao, Tagyoung Chung, Jing Huang, Yang Liu, and Nanyun Peng. 2022. Context-situated pun generation. In EMNLP 2022. +Chen Tang, Chenghua Lin, Henglin Huang, Frank Guerin, and Zhihao Zhang. 2022a. EtrICA: Event-triggered context-aware story generation augmented by cross attention. In *Findings of the Association for Computational Linguistics: EMNLP* 2022. +Chen Tang, Hongbo Zhang, Tyler Loakman, Chenghua Lin, and Frank Guerin. 2022b. Terminology-aware medical dialogue generation. arXiv preprint arXiv:2210.15551. +Chen Tang, Zhihao Zhang, Tyler Loakman, Chenghua Lin, and Frank Guerin. 2022c. NGEP: A graph-based event planning framework for story generation. In Proceedings of AACL-IJCNLP, Online. + +Yufei Tian and Nanyun Peng. 2022. Zero-shot sonnet generation with discourse-level planning and aesthetics features. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3587-3597, Seattle, United States. Association for Computational Linguistics. +Tim Van de Cruys. 2020. Automatic poetry generation from prosaic text. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2471-2480, Online. Association for Computational Linguistics. +Carolyn E. Wilshire. 1999. The "tongue twister" paradigm as a technique for studying phonological encoding. Language and Speech, 42(1):57-82. +Jörg Wöckener, Thomas Haider, Tristan Miller, The-Khang Nguyen, Thanh Tung Linh Nguyen, Minh Vu Pham, Jonas Belouadi, and Steffen Eger. 2021. End-to-end style-conditioned poetry generation: What does it take to learn from examples alone? In Proceedings of the 5th Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature, pages 57-66, Punta Cana, Dominican Republic (online). Association for Computational Linguistics. +Min Ney Wong, Yanky Chan, Manwa L. Ng, and Frank F. Zhu. 2019. Effects of transcranial direct current stimulation over the broca's area on tongue twister production. International Journal of Speech-Language Pathology, 21(2):182-188. PMID: 29642741. +Lanqing Xue, Kaitao Song, Duocai Wu, Xu Tan, Nevin L. Zhang, Tao Qin, Wei-Qiang Zhang, and Tie-Yan Liu. 2021. DeepRapper: Neural rap generation with rhyme and rhythm modeling. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 69-81, Online. Association for Computational Linguistics. +Zhiwei Yu, Jiwei Tan, and Xiaojun Wan. 2018. A neural approach to pun generation. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1650-1660, Melbourne, Australia. Association for Computational Linguistics. +Rongsheng Zhang, Xiaoxi Mao, Le Li, Lin Jiang, Lin Chen, Zhiwei Hu, Yadong Xi, Changjie Fan, and Minlie Huang. 2020a. Youling: an AI-assisted lyrics creation system. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 85-91, Online. Association for Computational Linguistics. +Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi. 2020b. *Bertscore: Evaluating text generation with bert*. In International Conference on Learning Representations. + +Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, and Bill Dolan. 2020c. DIALOGPT: Large-scale generative pre-training for conversational response generation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 270-278, Online. Association for Computational Linguistics. + +# A Appendices + +# A.1 Dataset Quality Control + +An annotation platform was developed as shown in (Figure 2). + +# A.2 Human Participants + +Due to tongue twisters being highly reliant on articulation abilities, the demographics of the human participants used within this work are highly important. Additionally, tongue twisters are also a form of humour and entertainment, where individual perceptions of what may or may not be considered humorous or entertaining differ according to numerous factors. In an effort to remain as transparent as possible, and follow best practices for human evaluation, relevant demographic information of participants are outlined below (with the necessary requisite permission and ethical approval). + +Dataset Evaluation All evaluators involved in the quality control process of the TwistList dataset are native speakers of English, and either have or are working towards University level qualifications. Additionally, 2 of the 3 evaluators have extensive education in linguistics or modern languages. No monetary incentive was provided. + +Generation Evaluation All evaluators involved in the evaluation of the quality of generated tongue twisters are native speakers of English, and either hold or are working towards University level qualifications in Linguistics, Modern Languages or NLP. Additionally, all evaluators cited the United Kingdom as their country of socialisation, and no participants reported language processing difficulties that could affect results. No monetary incentive was provided. + +Materials Provided to Human Participants Additionally, all evaluators for both the dataset and generation outputs were presented with calibration examples to demonstrate the sort of outputs that would be presented, and the logic behind particular scores, in order to minimise individual interpretations of the scoring criteria. All evaluation was performed on a custom made online annotation platform (Figure 3). + +# A.3 Training Details + +All pre-trained models used (naturally excluding ChatGPT) are based on publicly available checkpoints from Hugging Face.12 Models are trained for up to 5 epochs on a Tesla A5000 machine with the best checkpoints selected based on the validation loss. The batch size is set to 32, and the learning rate is $8e^{-5}$ , with the Adam optimiser selected for training. To help the loss curve converge on our small few-shot dataset, we limit the generation length to 100 (covering all test tongue twisters). Meanwhile, the source length is limited to 150. The training and testing steps are set up with the implementation of the PyTorch Lightning13 framework to guarantee the reliability of the experiment. All language models are fairly trained and tested with the same steps. + +# A.4 Further Qualitative Comments + +Whilst the pattern of extreme word repetition is seen in many of the finetuned models (often with the exception of BART, which is demonstrated to be capable of producing slightly more sophisticated outputs), overall assessment of the tongue twisters produced at inference time reveals interesting patterns, particularly in regard to ChatGPT outputs. Firstly, the limits of ChatGPT are made apparent in a few examples such as the input "silver shiny ship sank" generating "How much wood would a woodchuck chuck if a woodchuck could chuck silver shiny ships?", a clear derivation of a famous woodchuck related tongue twister that it is rather safe to assume appears multiple times in ChatGPTs training material. Additionally, comments from evaluators also reveal that ChatGPT's output is often considered more of a rhyme or general literary text, rather than specifically a tongue twister. However, examples such as these are also found in the human-authored golden examples, demonstrating that there is no steadfast consistent opinion as to what constitutes a (good) tongue twister. Likewise, some examples may contain large amounts of sound repetition, but not in a way that necessarily presents articulatory difficulty. + +# A.5 Future Works + +In this paper, we mainly analyse the performance of large-scale pretrained language models (PLMs) on Tongue Twister Generation, and propose a corresponding dataset for further investigation. In further works, we aim to propose novel models which can better leverage phonetic symbols. There + +# Introduction + +1. Individually read the tongue twister, phonetics, and key words on the left side. +Select the options on the right side to evaluate the data quality from the following perspectives: +- The quality of the RAKE Keywords: Do these suitably represent the topic of the tongue twister? +- The quality of the BERTopic Keywords: Do these suitably represent the topic of the tongue twister? +- The quality of the Tongue Twister: Is it a good tongue twister, or too short/long, or generally bad quality? + +# Tongue Twisters + +Keywords for Tongue Twisters + +RAKE Keywords: + +pickled peppers peter piper picked + +[P I H1 K A H0 L D] [P E H1 P E R O Z] [P I Y1 T E R O] [P A Y1 P E R O] [P I H1 K T] + +BERTopic + +Keywords: + +pink peter piper peck + +[P I H1NGK][P I Y1TERO][P AY1PER0][P E H1K] + +Tongue Twister + +Tongue Twister + +Text: + +Peter Piper picked a peck of pickled peppers. A peck of pickled peppers Peter Piper picked. If Peter Piper picked a peck of pickled peppers, Where's the peck of pickled peppers Peter Piper picked? + +Tongue Twister + +Phonetics: + +[PI Y1 ER0] [PY A1 ER0] [PI H1 K T] [AOH] [PE H1 K] [AHV] [PI I H1 K AI O L D] [PE H1 ER0 Z] [JI AHO] [PE H1 K] [AHV] [PI I H1 K AI O L D] [PE H1 ER0 Z] [PI Y1 ER0] [PY A1 ER0] [PI I H1 K T] [JI (IH1 F) [PI Y1 T ER0] [PY A1 ER0] [PI I H1 K T] [AOH] [PE H1 K] [AHV] [PI I H1 K AI O L D] [PE H1 ER0 Z] [WEH1 K R2] [DH AHO] [PE H1 K] [AHV] [PI I H1 K AI O L D] [PE H1 ER0 Z] [PI Y1 T ER0] [PY A1 ER0] [PI I H1 K T] [Q]: + +# Sample Quality + +A1: Are the RAKE Keywords highly suitable for the Tongue Twister? + +Yes + +No + +A2: Are the BERTopic Keywords highly suitable for the Tongue Twister? + +Yes + +No + +A3: Is it a good tongue twister? + +Good +A Bit Too Short +A Bit Too Long +$\bigcirc$ Bad Quality + +A4: Please input up to 5 manual keywords for the tongue twister (these can either be extracted from the example, or any other words). + +Keywords + +Quit + +Submit + +Figure 2: TwistList Quality Control Annotation Platform + +# Tongue Twister Evaluation + +Home / Sample Annotation + +# Introduction + +1. Individually read the input keywords and the tongue twister on the left side. +2. Give a score for each metric on the right to evaluate the quality of generated tongue twisters: +- Relevance: The extent to which the tongue twister is remantically/topically related to the input keywords. +- Difficulty of Articulation: The extent to which the tongue twister is hard to say (aka. how much your tongue twists). +- Fluency: The extent to which the tongue twister can be considered grammatically acceptable. +- Coherence: The extent to which the tongue twister can be considered logically and semantically coherent. +Entertainment: The extent to which the tongue twister is considered enteratining (primarily relating to phonetics + +semantics) + +# Tongue Twisters + +Input Keywords + +RAKE Keywords: + +Generate tongue twisters about key words: nope + +Tongue Twister + +Tongue Twister Text: + +Nope, an antelope can't elope with a cantaloupe. + +# Evaluate the Quality of Tongue Twisters by giving a Score (1 to 5) + +Please select all options before submission. + +O1:Relevance. + +# + +O2: Difficulty of articulation. + +0 + +Q3:FLuency. + +0 + +Q4: Coherence. + +5 + +Q5: Entertainment. + +。 + +Quit + +Submit + +Figure 3: Human Evaluation Platform for Generated Outputs + +are numerous existing works (Huang et al., 2022; Tang et al., 2022a,b) that provide approaches for injecting such knowledge into PLMs. However, the phonetic features differ from these text-format knowledge items, as phonemes are hard to encode with input text tokens when feeding into PLM encoders. Another promising approach is to explicitly model the phonetic features into text sequences (Tang et al., 2022c), though there is no observed method for transforming phonetic notation. We intend to perform further research based on these existing approaches. + +# A For every submission: + +A1. Did you describe the limitations of your work? + +Yes, in the required Limitations section as well as Section 4 (concerning our proposed metrics) + +A2. Did you discuss any potential risks of your work? + +Ethics Statement + +A3. Do the abstract and introduction summarize the paper's main claims? + +Abstract (all) and contribution summary at the end of the introduction. + +A4. Have you used AI writing assistants when working on this paper? + +Left blank. + +# B Did you use or create scientific artifacts? + +TwistList dataset (Section 3.2) + +B1. Did you cite the creators of artifacts you used? + +Sources of all entries in the dataset are credited in the .json file for each entry. + +B2. Did you discuss the license or terms for use and / or distribution of any artifacts? + +We did not discuss the licensing around our dataset. The dataset uses works that are freely available on the web and come from various sources such as websites, blogs, and eBooks. Many of these cases are Public Domain, and for those that are not, we believe we are in accordance with Fair Use, as the dataset does not encroach on the use case of the original works (no graphic design/other elements are maintained) and the dataset is for use as a research tool only. We will also reply promptly to any cases of copyright infringement that relevant copyright holders make us aware of. + +B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? + +See answer to B2. + +B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? + +See the Ethics Statement regarding the potential for tongue twisters to be offensive. Additionally, all tongue twisters are believed to be about fictional characters, rather than individuals. + +B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? + +Such details are not explicitly stated. However, it can be easily ascertained from the paper that the tongue twisters we focus on are entirely in English (and the range of domains the tongue twisters were taken from can be seen in the "source" entry for each example). + +B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. + +See Table 1 for dataset statistics. + +# C Did you run computational experiments? + +Section 4 (page 3) + +C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? + +Appendix A.3 + +C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? + +Appendix A.3 + +C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? + +Tables 3/5. Scores are the mean, as is standard. + +C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? + +Exact details of evaluation implementations (except Phonetic Overlap) were not detailed. This is in part due to these metrics (BLEU/ROUGE/BERTScore) not being very reliable for creative language generation, and therefore the exact values from different implementations are not likely to be of use. + +# D Did you use human annotators (e.g., crowdworkers) or research with human participants? + +Section 3.2 and Section 4. In addition to Appendix A.2 + +D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? + +Screenshot of the annotation platforms can be found in Figures 2 and 3 in the Appendix + +D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? + +We declared that no monetary incentive was given to participants. We did not specify the recruitment process, but due to participants all holding or working towards university level qualifications, it can be inferred that they are colleagues. + +D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? + +This information was not deemed necessary in the submitted paper (due to the limited risk of the data we were working with). However, it is stated in the Ethical Statement and Appendix A.2 that all shared information about human demographics was collected with the necessary permissions and approval. + +D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? + +Ethical approval was gained for human evaluation of the dataset and generated outputs from the relevant institution + +D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? + +We provide demographic information for human participants in Appendix A.2 \ No newline at end of file diff --git a/2023/TwistList_ Resources and Baselines for Tongue Twister Generation/images.zip b/2023/TwistList_ Resources and Baselines for Tongue Twister Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7ec49179d3b6b673ac12da30efcb1831f32dd756 --- /dev/null +++ b/2023/TwistList_ Resources and Baselines for Tongue Twister Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eb3952bd321fe07bcb74eebe5779e96b9bc349dc1eaf6def88fe5ed3bbfdc55b +size 231507 diff --git a/2023/TwistList_ Resources and Baselines for Tongue Twister Generation/layout.json b/2023/TwistList_ Resources and Baselines for Tongue Twister Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..be52d3044fdae4833e178a1c98b0edb3b04f1bb1 --- /dev/null +++ b/2023/TwistList_ Resources and Baselines for Tongue Twister Generation/layout.json @@ -0,0 +1,9496 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 101, + 76, + 491, + 93 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 76, + 491, + 93 + ], + "spans": [ + { + "bbox": [ + 101, + 76, + 491, + 93 + ], + "type": "text", + "content": "TwistList: Resources and Baselines for Tongue Twister Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 174, + 110, + 424, + 124 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 110, + 424, + 124 + ], + "spans": [ + { + "bbox": [ + 174, + 110, + 424, + 124 + ], + "type": "text", + "content": "Tyler Loakman" + }, + { + "bbox": [ + 174, + 110, + 424, + 124 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 174, + 110, + 424, + 124 + ], + "type": "text", + "content": ", Chen Tang" + }, + { + "bbox": [ + 174, + 110, + 424, + 124 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 174, + 110, + 424, + 124 + ], + "type": "text", + "content": " and Chenghua Lin" + }, + { + "bbox": [ + 174, + 110, + 424, + 124 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 143, + 125, + 452, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 125, + 452, + 138 + ], + "spans": [ + { + "bbox": [ + 143, + 125, + 452, + 138 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 143, + 125, + 452, + 138 + ], + "type": "text", + "content": "Department of Computer Science, The University of Sheffield, UK" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 149, + 139, + 447, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 139, + 447, + 153 + ], + "spans": [ + { + "bbox": [ + 149, + 139, + 447, + 153 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 149, + 139, + 447, + 153 + ], + "type": "text", + "content": "Department of Computer Science, The University of Surrey, UK" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 201, + 153, + 396, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 153, + 396, + 166 + ], + "spans": [ + { + "bbox": [ + 201, + 153, + 396, + 166 + ], + "type": "text", + "content": "{tcloakman1,c.lin}@sheffield.ac.uk" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 235, + 168, + 362, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 168, + 362, + 179 + ], + "spans": [ + { + "bbox": [ + 235, + 168, + 362, + 179 + ], + "type": "text", + "content": "chen.tang@surrey.ac.uk" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 156, + 212, + 201, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 212, + 201, + 224 + ], + "spans": [ + { + "bbox": [ + 156, + 212, + 201, + 224 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 83, + 239, + 274, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 239, + 274, + 528 + ], + "spans": [ + { + "bbox": [ + 83, + 239, + 274, + 528 + ], + "type": "text", + "content": "Previous work in phonetically-grounded language generation has mainly focused on domains such as lyrics and poetry. In this paper, we present work on the generation of tongue twisters - a form of language that is required to be phonetically conditioned to maximise sound overlap, whilst maintaining semantic consistency with an input topic, and still being grammatically correct. We present TwistList, a large annotated dataset of tongue twisters, consisting of " + }, + { + "bbox": [ + 83, + 239, + 274, + 528 + ], + "type": "inline_equation", + "content": "2.1\\mathrm{K}+" + }, + { + "bbox": [ + 83, + 239, + 274, + 528 + ], + "type": "text", + "content": " human-authored examples. We additionally present several benchmark systems (referred to as TwisterMisters) for the proposed task of tongue twister generation, including models that both do and do not require training on in-domain data. We present the results of automatic and human evaluation to demonstrate the performance of existing mainstream pre-trained models in this task with limited (or no) task specific training and data, and no explicit phonetic knowledge. We find that the task of tongue twister generation is challenging for models under these conditions, yet some models are still capable of generating acceptable examples of this language type." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 68, + 541, + 151, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 541, + 151, + 553 + ], + "spans": [ + { + "bbox": [ + 68, + 541, + 151, + 553 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 565, + 291, + 741 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 565, + 291, + 741 + ], + "spans": [ + { + "bbox": [ + 67, + 565, + 291, + 741 + ], + "type": "text", + "content": "Phonetically constrained language generation is a primary subarea of computational creativity in natural language generation (NLG), primarily encompassing lyric and poetry generation (Tian and Peng, 2022; Wöckener et al., 2021; Xue et al., 2021; Zhang et al., 2020a; Agarwal and Kann, 2020), as well as pun generation (Sun et al., 2022; He et al., 2019; Yu et al., 2018), and continues to prove challenging for myriad reasons. Primarily, such works require the inclusion of phonetic factors such as metre and rhyme, which involves careful consideration of candidate vocabulary on the syllable level, leading to a reduced pool of allowable vocabulary once these constraints are in place." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 304, + 210, + 526, + 342 + ], + "blocks": [ + { + "bbox": [ + 304, + 210, + 526, + 342 + ], + "lines": [ + { + "bbox": [ + 304, + 210, + 526, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 210, + 526, + 342 + ], + "type": "image", + "image_path": "df922e0bf87dd9a82794c843e2cecb8fc093ef3f8eecea4a89ad98ced56b52a3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 349, + 525, + 386 + ], + "lines": [ + { + "bbox": [ + 302, + 349, + 525, + 386 + ], + "spans": [ + { + "bbox": [ + 302, + 349, + 525, + 386 + ], + "type": "text", + "content": "Figure 1: Tongue Twister Generation aims to generate an utterance with high levels of phonetic overlap, requiring understanding of semantics, grammar, and phonetics." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 417, + 525, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 417, + 525, + 618 + ], + "spans": [ + { + "bbox": [ + 302, + 417, + 525, + 618 + ], + "type": "text", + "content": "In this paper, we present work on the generation of tongue twisters, a type of phonetically constrained language that is rarely explored in the NLG community. As a form of creative generation, tongue twisters can facilitate numerous useful applications, including: (1) being used as a pedagogical tool (Sugiharto et al., 2022; Somoff, 2014; Wilshire, 1999); (2) as a source of humorous entertainment stemming from unintentional mispronunciations; (3) as a stylistic device for engaging children in reading (e.g. Dr. Seuss stories (Geisel, 1965)); (4) as a method of designing memorable slogans and tag lines (Guerini et al., 2015); and (5) as stimuli in neuroscience/physiology research (Wong et al., 2019; O'Halloran, 2020; Kember et al., 2017)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 624, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 624, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 624, + 526, + 772 + ], + "type": "text", + "content": "Tongue twister generation posits unique challenges compared to other generation tasks. One of the most pertinent features of tongue twisters is the presence of high levels of phonetic overlap across tokens (Wilshire, 1999). Consequently, whilst other types of creative generation may require only some output tokens to consider phonetics (such as rhyme or syllable counts), tongue twisters present an extreme version of this problem where the phonetics of almost all generated tokens must be considered. This leads to a very small vocabulary from which to choose" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 81, + 751, + 151, + 761 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 751, + 151, + 761 + ], + "spans": [ + { + "bbox": [ + 81, + 751, + 151, + 761 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 81, + 761, + 162, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 761, + 162, + 772 + ], + "spans": [ + { + "bbox": [ + 81, + 761, + 162, + 772 + ], + "type": "text", + "content": "†Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "text", + "content": "579" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "spans": [ + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "type": "text", + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 224, + 806, + 369, + 817 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 806, + 369, + 817 + ], + "spans": [ + { + "bbox": [ + 224, + 806, + 369, + 817 + ], + "type": "text", + "content": "Volume 2: Short Papers, pages 579-589" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 176, + 818, + 416, + 828 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 818, + 416, + 828 + ], + "spans": [ + { + "bbox": [ + 176, + 818, + 416, + 828 + ], + "type": "text", + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 290, + 98 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 290, + 98 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 290, + 98 + ], + "type": "text", + "content": "semantically relevant words, and presents further challenges with maintaining grammatical validity." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 100, + 291, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 100, + 291, + 409 + ], + "spans": [ + { + "bbox": [ + 69, + 100, + 291, + 409 + ], + "type": "text", + "content": "The only work that we are aware of on tongue twister generation at the time of conducting this research is by Keh et al. (2022), who present models that train on graphemes and phonemes, and take either a starting prompt to be continued, or keywords around which to theme an output. They release TT-Corp, a dataset of 644 tongue twisters with parallel non-twister equivalents. We differentiate our work through the release of a dataset that is over " + }, + { + "bbox": [ + 69, + 100, + 291, + 409 + ], + "type": "inline_equation", + "content": "3\\mathrm{x}" + }, + { + "bbox": [ + 69, + 100, + 291, + 409 + ], + "type": "text", + "content": " larger and which has undergone substantial human quality control. Furthermore, we assess the results of a wider range of popular pre-trained models on this task, including ChatGPT, without explicit injection of phonetic knowledge due to the difficulty in encoding phonetics and the expertise required to utilise phonetic characteristics appropriately. Our experimental results show that most popular pretrained language models (PLMs) rely on pure word repetition to generate tongue twisters, whilst some (i.e. BART) are able to generate more sophisticated examples. Additionally, very large zero-shot models (i.e. ChatGPT) are able to generate convincing tongue twisters almost on-par with human equivalents." + }, + { + "bbox": [ + 69, + 100, + 291, + 409 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 78, + 411, + 267, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 411, + 267, + 423 + ], + "spans": [ + { + "bbox": [ + 78, + 411, + 267, + 423 + ], + "type": "text", + "content": "To summarise our contributions, we present:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 425, + 290, + 546 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 68, + 425, + 290, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 425, + 290, + 465 + ], + "spans": [ + { + "bbox": [ + 68, + 425, + 290, + 465 + ], + "type": "text", + "content": "- TwistList, a large annotated dataset of human-authored tongue twisters, containing " + }, + { + "bbox": [ + 68, + 425, + 290, + 465 + ], + "type": "inline_equation", + "content": "2.1\\mathrm{K}+" + }, + { + "bbox": [ + 68, + 425, + 290, + 465 + ], + "type": "text", + "content": " examples with human evaluation of their quality." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 465, + 290, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 465, + 290, + 504 + ], + "spans": [ + { + "bbox": [ + 68, + 465, + 290, + 504 + ], + "type": "text", + "content": "- TwisterMisters, a series of baseline models for tongue twister generation using the most popular state-of-the-art PLMs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 506, + 290, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 506, + 290, + 546 + ], + "spans": [ + { + "bbox": [ + 68, + 506, + 290, + 546 + ], + "type": "text", + "content": "- Extensive automatic and human evaluation to assess the ability of PLMs to implicitly model the complex phonetic phenomena in tongue twisters." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 68, + 557, + 162, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 557, + 162, + 571 + ], + "spans": [ + { + "bbox": [ + 68, + 557, + 162, + 571 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 580, + 291, + 743 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 580, + 291, + 743 + ], + "spans": [ + { + "bbox": [ + 67, + 580, + 291, + 743 + ], + "type": "text", + "content": "Previous work in phonetically constrained generation has taken one of two approaches: 1) train a generation model on a collection of in-domain texts, or 2) train a generation model on prosaic out-of-domain text, with constraints imposed at decoding time. For example, Lau et al. (2018) collect 3,355 sonnets to produce novel poetry and train models to generate text in iambic pentameter, whilst Xue et al. (2021) train a rap generation model on 272,839 in-domain examples, infusing knowledge of rhythm afterwards. On the other hand, Van de Cruys (2020) train on a subset of CommonCrawl, imposing constraints on topic and" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 307, + 68, + 523, + 168 + ], + "blocks": [ + { + "bbox": [ + 307, + 68, + 523, + 168 + ], + "lines": [ + { + "bbox": [ + 307, + 68, + 523, + 168 + ], + "spans": [ + { + "bbox": [ + 307, + 68, + 523, + 168 + ], + "type": "table", + "html": "
DatasetTrainValTestTotal
# Tongue Twisters19121061072128
Vocabulary Size955694688010358
# Total Phonemes55434656
# RAKE Keywords33333162883567
# BERTopic Keywords250132160250
Avg. # Input Keywords (RAKE)3.163.323.013.16
Avg. # Input Phonemes5.575.835.165.56
Avg. Tongue Twister Length (Words)15.0116.5913.5415.01
Avg. # Input Phonemes26.0628.2523.5026.04
", + "image_path": "79ffdcc132cb6409a323d26bc731e27151df5d2b265c4e7f99818bda248f8a81.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 343, + 175, + 483, + 186 + ], + "lines": [ + { + "bbox": [ + 343, + 175, + 483, + 186 + ], + "spans": [ + { + "bbox": [ + 343, + 175, + 483, + 186 + ], + "type": "text", + "content": "Table 1: The Statistics of TwistList." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 302, + 211, + 525, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 211, + 525, + 306 + ], + "spans": [ + { + "bbox": [ + 302, + 211, + 525, + 306 + ], + "type": "text", + "content": "rhyme as a priori distributions, whilst Tian and Peng (2022) train a title-to-keyword module on narrative texts in addition to a sonnet generation model trained on news articles and short stories from Reddit. They imposed literary techniques (simile/metaphor) and metre/rhyme constraints at decoding time, owing to the lack of sufficient training data." + }, + { + "bbox": [ + 302, + 211, + 525, + 306 + ], + "type": "inline_equation", + "content": "^2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 320, + 456, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 320, + 456, + 333 + ], + "spans": [ + { + "bbox": [ + 302, + 320, + 456, + 333 + ], + "type": "text", + "content": "3 Tongue Twister Generation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 344, + 399, + 356 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 344, + 399, + 356 + ], + "spans": [ + { + "bbox": [ + 302, + 344, + 399, + 356 + ], + "type": "text", + "content": "3.1 Task Definition" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 364, + 525, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 364, + 525, + 513 + ], + "spans": [ + { + "bbox": [ + 302, + 364, + 525, + 513 + ], + "type": "text", + "content": "We formulate the task of tongue twister generation as follows: for a given set of keywords, we aim to generate a tongue twister " + }, + { + "bbox": [ + 302, + 364, + 525, + 513 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 302, + 364, + 525, + 513 + ], + "type": "text", + "content": ", whereby " + }, + { + "bbox": [ + 302, + 364, + 525, + 513 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 302, + 364, + 525, + 513 + ], + "type": "text", + "content": " comprises a sequence of words " + }, + { + "bbox": [ + 302, + 364, + 525, + 513 + ], + "type": "inline_equation", + "content": "\\{w_1, w_2, \\dots, w_n\\}" + }, + { + "bbox": [ + 302, + 364, + 525, + 513 + ], + "type": "text", + "content": ". The generated output must satisfy the following constraints: (1) the output should be semantically related to the input keywords; (2) the output should show maximal levels of phonetic overlap across tokens; and (3) the output should be grammatically valid (Wilshire, 1999). Of these requirements, phonetic overlap is the most central to defining text as a \"tongue twister\"." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 526, + 408, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 526, + 408, + 538 + ], + "spans": [ + { + "bbox": [ + 302, + 526, + 408, + 538 + ], + "type": "text", + "content": "3.2 TwistList Dataset" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 545, + 525, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 545, + 525, + 708 + ], + "spans": [ + { + "bbox": [ + 302, + 545, + 525, + 708 + ], + "type": "text", + "content": "Dataset Construction. We present TwistList, an annotated dataset of " + }, + { + "bbox": [ + 302, + 545, + 525, + 708 + ], + "type": "inline_equation", + "content": "2.1\\mathrm{K}+" + }, + { + "bbox": [ + 302, + 545, + 525, + 708 + ], + "type": "text", + "content": " human-authored tongue twisters for use by the community. The examples contained therein come from a variety of sources available on the web. For each tongue twister, phonetic transcription is provided using the g2p-en package, in addition to keywords extracted with RAKE and BERTopic to represent the topic of the tongue twister. Following experimentation with both RAKE and BERTopic, only RAKE keywords are used in training due to human preference and issues regarding the use of BERTopic on short texts (where" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 719, + 525, + 750 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 719, + 525, + 750 + ], + "spans": [ + { + "bbox": [ + 302, + 719, + 525, + 750 + ], + "type": "text", + "content": "2Additionally, there is often a reluctance in computational creativity to train on examples, under the assumption that the newly generated content will be overly derivative." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 750, + 517, + 761 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 750, + 517, + 761 + ], + "spans": [ + { + "bbox": [ + 317, + 750, + 517, + 761 + ], + "type": "text", + "content": "3The source of each tongue twister is stated for each entry." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 761, + 458, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 761, + 458, + 772 + ], + "spans": [ + { + "bbox": [ + 317, + 761, + 458, + 772 + ], + "type": "text", + "content": "4https://pypi.org/project/g2p-en/" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 67, + 750, + 290, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 750, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 750, + 290, + 772 + ], + "type": "text", + "content": "1Our code and resources can be accessed at https://github.com/tangg555/TwistList" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "580" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 71, + 290, + 98 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 71, + 290, + 98 + ], + "spans": [ + { + "bbox": [ + 68, + 71, + 290, + 98 + ], + "type": "text", + "content": "frequently no keywords are extracted). The main statistics of the dataset are presented in Table 1." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 71, + 117, + 289, + 207 + ], + "blocks": [ + { + "bbox": [ + 71, + 117, + 289, + 207 + ], + "lines": [ + { + "bbox": [ + 71, + 117, + 289, + 207 + ], + "spans": [ + { + "bbox": [ + 71, + 117, + 289, + 207 + ], + "type": "table", + "html": "
RAKE:sells thick socks
BERTopic:short shorts socks sock
Twister:Seth at Sainsbury's sells thick socks.
Phonetics:[S EH1 TH] [AE1 T] [S EY1 N S B ER0 IY0 Z] [S EH1 L Z] [TH IH1 K] [S AA1 K S]
", + "image_path": "80ed8d4600a98e49a2b4dfdb7896e4a66e2e302d1bc905de00d367d9e68711d2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 114, + 215, + 244, + 227 + ], + "lines": [ + { + "bbox": [ + 114, + 215, + 244, + 227 + ], + "spans": [ + { + "bbox": [ + 114, + 215, + 244, + 227 + ], + "type": "text", + "content": "Table 2: Example from TwistList" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 66, + 269, + 291, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 269, + 291, + 567 + ], + "spans": [ + { + "bbox": [ + 66, + 269, + 291, + 567 + ], + "type": "text", + "content": "Quality Control. Quality control on our dataset was performed in multiple ways. Firstly, it was ensured that only sufficiently unique tongue twisters were kept in the dataset, as determined by removing examples with over " + }, + { + "bbox": [ + 66, + 269, + 291, + 567 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 66, + 269, + 291, + 567 + ], + "type": "text", + "content": " word overlap (rather than keeping variants of the same tongue twister, such as \"Peter Piper picked a pickled pepper\" versus \"Peter the Piper picked...\"). Additionally, non-standard spellings were manually converted to standard US English5 to avoid G2P issues.6 Similarly, tongue-twisters containing obscure vocabulary (such as medicine and dinosaur names) were excluded to further minimise errors. An annotation platform was developed (see Appendix A.1), with which 3 human evaluators, who are native speakers of English, were provided with 100 sampled instances from the dataset to rate the quality of the resulting tongue twisters and the associated extracted keywords. The full dataset contains " + }, + { + "bbox": [ + 66, + 269, + 291, + 567 + ], + "type": "inline_equation", + "content": "2,500+" + }, + { + "bbox": [ + 66, + 269, + 291, + 567 + ], + "type": "text", + "content": " tongue twisters, of which 2,128 are kept for training/development/testing after filtering examples with insufficient extracted keywords and excessive similarity to existing entries." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 572, + 291, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 572, + 291, + 695 + ], + "spans": [ + { + "bbox": [ + 66, + 572, + 291, + 695 + ], + "type": "text", + "content": "To summarise, 3 annotators evaluated the quality of the dataset, where " + }, + { + "bbox": [ + 66, + 572, + 291, + 695 + ], + "type": "inline_equation", + "content": "88\\%" + }, + { + "bbox": [ + 66, + 572, + 291, + 695 + ], + "type": "text", + "content": " of assessed tongue twisters were considered high quality, and " + }, + { + "bbox": [ + 66, + 572, + 291, + 695 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 66, + 572, + 291, + 695 + ], + "type": "text", + "content": " considered \"suitable\" (Kappa " + }, + { + "bbox": [ + 66, + 572, + 291, + 695 + ], + "type": "inline_equation", + "content": "= 0.321" + }, + { + "bbox": [ + 66, + 572, + 291, + 695 + ], + "type": "text", + "content": "). An example from TwistList is provided in Table 2. As Table 4 shows, the final dataset can be considered high quality, owing to fair/moderate levels of approval and agreement across evaluators. Demographic information of the evaluators can be found in Appendix A.2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 302, + 71, + 402, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 402, + 83 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 402, + 83 + ], + "type": "text", + "content": "3.3 Baseline Models" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 89, + 525, + 130 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 89, + 525, + 130 + ], + "spans": [ + { + "bbox": [ + 302, + 89, + 525, + 130 + ], + "type": "text", + "content": "We present the following baseline models (dubbed TwisterMisters) for the task of tongue twister generation on our TwistList dataset:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 139, + 526, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 139, + 526, + 383 + ], + "spans": [ + { + "bbox": [ + 302, + 139, + 526, + 383 + ], + "type": "text", + "content": "Finetuned Baselines. For the finetuned baselines, we chose popular models for language generation, including GPT-2 (Radford et al., 2019), DialogGPT (Zhang et al., 2020c), T5 (Raffel et al., 2020), and BART (Lewis et al., 2020). These were finetuned with RAKE keywords extracted from human-authored tongue twisters as the input and the tongue twister text from TwistList as the target. This was in order to represent our baselines training on in-domain data. At inference time, the prompt \"Generate tongue twisters about the keyword(s): X\" is used, where X refers to the input consisting of one or more RAKE keywords extracted from tongue twisters. The full training details are given in Appendix A.3. We also conducted experiments on all aforementioned baselines without finetuning (i.e., a zero-shot setting), and the results were very poor. Therefore, we did not include these results in the paper." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 392, + 525, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 392, + 525, + 596 + ], + "spans": [ + { + "bbox": [ + 302, + 392, + 525, + 596 + ], + "type": "text", + "content": "Training-Free Baseline We additionally provide a TwisterMister baseline that does not require any training. We utilise OpenAI's ChatGPT7 with the same prompt as a zero-shot setting for generation.8 Each request to ChatGPT was submitted as part of a separate session, to avoid the effects of extended dialogue influencing outputs. ChatGPT has been utilised in order to set a practical upper-bound of what may be expected from models without explicit phonetic knowledge, owing to its wealth of training data and 175B parameter architecture.9 It is assumed that ChatGPT's training data contains tongue twisters, and therefore it is able to abstract away the general patterns of such language in order to provide novel examples (though most likely based on graphemes rather than phonemes)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 607, + 387, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 607, + 387, + 621 + ], + "spans": [ + { + "bbox": [ + 302, + 607, + 387, + 621 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 629, + 525, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 629, + 525, + 710 + ], + "spans": [ + { + "bbox": [ + 302, + 629, + 525, + 710 + ], + "type": "text", + "content": "Automatic Evaluation. We present the results of automatic evaluation on generated outputs and golden examples in Table 3 for the following metrics: Perplexity (PPL), BLEU (B-1/B-2) (Papineni et al., 2002), ROUGE (R-1/R-2/R-L) (Lin, 2004), and BERTScore Precision, Recall, and F-Measure (Zhang" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 719, + 290, + 750 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 719, + 290, + 750 + ], + "spans": [ + { + "bbox": [ + 67, + 719, + 290, + 750 + ], + "type": "text", + "content": "For example, where phonetic spellings or letter substitutions such as \"k\" for \"c\" were used for literary and visual effect, such as \"kwik\" for \"quick\"." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 67, + 751, + 289, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 751, + 289, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 751, + 289, + 772 + ], + "type": "inline_equation", + "content": "^6 g2p" + }, + { + "bbox": [ + 67, + 751, + 289, + 772 + ], + "type": "text", + "content": "-en uses the CMU Pronouncing Dictionary to retrieve transcriptions, which is an American English resource." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 719, + 442, + 729 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 719, + 442, + 729 + ], + "spans": [ + { + "bbox": [ + 317, + 719, + 442, + 729 + ], + "type": "text", + "content": "7https://chat.openai.com/chat" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 303, + 729, + 525, + 761 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 729, + 525, + 761 + ], + "spans": [ + { + "bbox": [ + 303, + 729, + 525, + 761 + ], + "type": "text", + "content": "8No direct comparison is made to PANCETTA (Keh et al., 2022) as no code has been publicly released at the time of writing, and essential implementation details are absent from the paper." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 761, + 485, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 761, + 485, + 772 + ], + "spans": [ + { + "bbox": [ + 317, + 761, + 485, + 772 + ], + "type": "text", + "content": "9ModelPPL↓B-1↑B-2↑R-1↑R-2↑R-L↑PO↓Init-PO↓BS-P↑BS-R↑BS-F↑GPT-28.400.0070.0031.3010.1231.3150.0220.0200.6900.8100.744DialoGPT3.830.0380.0257.7243.6107.6400.0690.0890.7540.8310.790T510.160.0570.0389.7014.5739.5740.6890.7270.7950.8180.806BART1.650.0730.05111.8836.10910.3530.0750.1200.7950.8450.819ChatGPTN/A0.2000.13736.76520.65933.4370.0930.1570.8880.8940.883", + "image_path": "63da84be1096e3824047a82151b4f89eb6e6f8e7227fd33b9948f8a728724a54.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 76, + 199, + 282, + 263 + ], + "blocks": [ + { + "bbox": [ + 67, + 155, + 524, + 179 + ], + "lines": [ + { + "bbox": [ + 67, + 155, + 524, + 179 + ], + "spans": [ + { + "bbox": [ + 67, + 155, + 524, + 179 + ], + "type": "text", + "content": "Table 3: Results of Automatic Evaluation. Golden PO = 0.385 and Golden Init-PO = 0.417. Due to the one-to-many issue in creative language generation, we acknowledge that the referenced metrics are imperfect." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 76, + 199, + 282, + 263 + ], + "lines": [ + { + "bbox": [ + 76, + 199, + 282, + 263 + ], + "spans": [ + { + "bbox": [ + 76, + 199, + 282, + 263 + ], + "type": "table", + "html": "
Choices (%)Sample Quality
High.Suitable.Bad.Kappa
RAKE keywords82.018.00.00.321
BERTopic keywords15.085.00.00.445
Tongue Twisters88.06.04.00.321
", + "image_path": "e6c3fa3ffb3747a02c714049eda8f7e2c7b40a1eb7868abf9c6028ce05620605.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 270, + 290, + 319 + ], + "lines": [ + { + "bbox": [ + 67, + 270, + 290, + 319 + ], + "spans": [ + { + "bbox": [ + 67, + 270, + 290, + 319 + ], + "type": "text", + "content": "Table 4: Kappa refers to Fleiss' Kappa (Fleiss, 1971). All results achieve fair or moderate agreement. Good tongue twisters that are deemed a bit longer (3%) or shorter (3%) than expected are considered \"suitable\"." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 346, + 290, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 346, + 290, + 481 + ], + "spans": [ + { + "bbox": [ + 67, + 346, + 290, + 481 + ], + "type": "text", + "content": "et al., 2020b) (BS-P/BS-R/BS-F). PPL, BLEU and ROUGE are standard metrics in language generation to assess quality, whilst BERTScore assesses semantic similarity to a gold reference. Additionally, we propose two new metrics, Phonetic Overlap (PO) and Initial Phonetic Overlap (Init-PO). PO refers to the average overlap of all phonemes across tokens (# unique phonemes/#totalphonemes),whereas Init-PO is the ratio of unique word-initial phonemes to the number of words (# unique word-initial phonemes/#words)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 486, + 290, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 486, + 290, + 634 + ], + "spans": [ + { + "bbox": [ + 67, + 486, + 290, + 634 + ], + "type": "text", + "content": "These phonetic metrics reward longer outputs. We argue that, all things equal, a longer tongue twister is better than a shorter one as it provides more entertainment and more opportunities for mispronunciation. Perfect scores on PO and Init-PO can be achieved by repetition of a single word. Whilst this does not lead to high quality outputs, these metrics are intended exclusively to be indicators of the phonetics, rather than an overall guide to quality. In both cases, higher levels of overlap results in lower (\"better\") scores, and the highest (\"worst\") achievable score is 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 638, + 290, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 638, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 638, + 290, + 772 + ], + "type": "text", + "content": "The results in Table 3 show rather clear scaling, with the performance ranking on most metrics (except Perplexity and phoneme overlap) being identical. On the models explicitly finetuned for this task, GPT-2 is shown to be the worst, whilst BART performs the best. We hypothesise that GPT-2's poor performance is in part due to its simple causal language modelling objective alongside its decoder-only architecture (which is also in DialogGPT). Furthermore, whilst T5 performed well on the automatic metrics, manual" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 301, + 200, + 525, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 200, + 525, + 459 + ], + "spans": [ + { + "bbox": [ + 301, + 200, + 525, + 459 + ], + "type": "text", + "content": "inspection revealed that T5 often misinterpreted the task from the prompt, choosing to select its own keywords from the entire prompt, rather than using only the provided keyword list. On the other hand, the training-free zero-shot model, ChatGPT, was shown to perform best on all metrics. This is to be expected as ChatGPT has over 50x more parameters than any other tested PLM, with various pre-training objectives and reinforcement learning, leading to performant zero-shot capabilities. This further demonstrates that PLMs struggle to learn phonetic patterns implicitly from text, especially in English, which has high levels of irregular orthography. Furthermore, with limited data, PLMs struggle to learn the unusual probability distributions underlying tongue twisters, where word choices are intentionally \"twisted\", obscure, and anti-euphonious. Additionally, due to the wealth of training data seen by ChatGPT, it is likely that many examples have been seen during training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 469, + 525, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 469, + 525, + 671 + ], + "spans": [ + { + "bbox": [ + 302, + 469, + 525, + 671 + ], + "type": "text", + "content": "Human Evaluation. Due to tongue twisters being a creative domain where articulation abilities are tested, we also perform human evaluation. 3 evaluators were asked to rate 100 outputs from the best performing standard baseline (BART), in addition to ChatGPT outputs and gold examples from TwistList on the following criteria: Relevance (how relevant the tongue twister is given the keyword inputs), Fluency (how grammatically valid the output is), Difficulty of Articulation (how difficult a tongue twister is to say), Cohesion (how much sense the output makes), and Entertainment Value (how entertaining the output is, considering sounds and semantics). All ratings were on a 5-point Likert scale. Evaluator demographics and training materials are in Appendix A.2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 672, + 525, + 740 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 672, + 525, + 740 + ], + "spans": [ + { + "bbox": [ + 302, + 672, + 525, + 740 + ], + "type": "text", + "content": "The mean scores of human evaluation (Table 5) fall in line with expectations, with golden examples performing best on all metrics, and ChatGPT placing second on all but Difficulty of Articulation.[10] BART is able to produce outputs that are deemed to be the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 751, + 524, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 751, + 524, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 524, + 772 + ], + "type": "text", + "content": "10We exclude relevance for the golden examples as these were collected from the web, not elicited with keyword prompts." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "text", + "content": "582" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 76, + 68, + 283, + 154 + ], + "blocks": [ + { + "bbox": [ + 76, + 68, + 283, + 154 + ], + "lines": [ + { + "bbox": [ + 76, + 68, + 283, + 154 + ], + "spans": [ + { + "bbox": [ + 76, + 68, + 283, + 154 + ], + "type": "table", + "html": "
Score (1 to 5)Human Evaluation
BARTChatGPTGolden
Relevance4.667*4.971†N/A
Difficulty of Articulation4.143*4.102*4.291*
Fluency3.028**4.915**4.938**
Coherence3.217*4.798*4.909*
Entertainment Value3.269*4.070*4.254*
", + "image_path": "65969a47912de5bc50d6f38177701cc9162d5660f9aad3bfc7d543af72c03022.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 238, + 291, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 238, + 291, + 535 + ], + "spans": [ + { + "bbox": [ + 67, + 238, + 291, + 535 + ], + "type": "text", + "content": "second most difficult to articulate, which we infer may be the result of slight morphological variants of input keywords being used repeatedly, making distinguishing between them during articulation quite challenging (whilst not being able to exploit deeper phonetic relations). The moderate score on Fluency (3.028) suggests instances of poor grammar may also hinder articulation abilities when expected grammatical structures are not found, leading to an interaction between grammatical validity and articulatory difficulty. Additionally, ChatGPT scoring the lowest for articulatory difficulty may be due to occasionally misunderstanding the requirements of a tongue twister, sometimes producing rhymes or standard prose (see Appendix A.4). However, ChatGPT scores well for Relevance and Fluency, highlighting its capability in producing high-quality coherent language. Perhaps most interestingly, none of the BART score averages on any human evaluation criteria fall below 3 (\"neither agree nor disagree\"). This performance is therefore quite good for a model finetuned on only 2128 examples, with no additional phonetic knowledge." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 71, + 552, + 289, + 722 + ], + "blocks": [ + { + "bbox": [ + 67, + 161, + 291, + 210 + ], + "lines": [ + { + "bbox": [ + 67, + 161, + 291, + 210 + ], + "spans": [ + { + "bbox": [ + 67, + 161, + 291, + 210 + ], + "type": "text", + "content": "Table 5: Results of Human Evaluation. The best scores are in bold, and the second best are underlined. We calculate Fleiss' Kappa for each metric, and mark the agreement fair*, moderate** and substantial†." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 71, + 552, + 289, + 722 + ], + "lines": [ + { + "bbox": [ + 71, + 552, + 289, + 722 + ], + "spans": [ + { + "bbox": [ + 71, + 552, + 289, + 722 + ], + "type": "table", + "html": "
Inputassistant assistant assist
GPT-2assistant assistant assist assistant assist assistant
DialogGPTassistant assistant assistant assistant assistant assistant assistant assistant assistant assistant assistant assistant
T5assistant assistant assist assistant
BARTA assistant assist is an assistant assist, assistants assist to assist assistants.
ChatGPTAssistant ants assist ants in carrying leaves to the ant hill.
GoldenIf I assist a sister-assistant, will the sister's sister-assistant assist me?
", + "image_path": "9c108b07e72f1da819ac8f0ed8a7050c7a2490c45b4f84cab83601b30b897d9f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 729, + 290, + 754 + ], + "lines": [ + { + "bbox": [ + 67, + 729, + 290, + 754 + ], + "spans": [ + { + "bbox": [ + 67, + 729, + 290, + 754 + ], + "type": "text", + "content": "Table 6: Example outputs for the input \"assistant assist\". \"Golden\" refers to the human-authored tongue twisters." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 303, + 70, + 379, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 70, + 379, + 84 + ], + "spans": [ + { + "bbox": [ + 303, + 70, + 379, + 84 + ], + "type": "text", + "content": "5 Case Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 301, + 100, + 527, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 100, + 527, + 356 + ], + "spans": [ + { + "bbox": [ + 301, + 100, + 527, + 356 + ], + "type": "text", + "content": "Within the example in Table 6, GPT-2 resorts to simply repeating the input, successfully achieving phonetic overlap, but failing to be grammatically valid or particularly sophisticated. This pattern is also demonstrated by DialogGPT and T5. Conversely, BART is able to introduce tokens unseen in the input to create an almost grammatically valid output (the primary mistake being indefinite article agreement, where in the first instance \"an\" would have been correct, rather than \"a\"). BART's output is also semantically and logically coherent, with \"A assistant assist is an assistant assist\" being valid (yet redundant), and \"assistants assist to assist assistants\" also being comprehensible. This example demonstrates why evaluators with high English proficiency and language/linguistics education were selected, as the same word may have different parts of speech, creating outputs that seem grammatically invalid, but do actually follow the rules of English.[11]" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 357, + 527, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 357, + 527, + 560 + ], + "spans": [ + { + "bbox": [ + 302, + 357, + 527, + 560 + ], + "type": "text", + "content": "Further investigation is needed to ascertain whether the models are intentionally exploiting this lexical ambiguity, or if human evaluators are demonstrating apophobia, where patterns are found in what is effectively noise (Brugger, 2001). Finally, ChatGPT utilises morphology to exploit the similarity of the plural noun \"assistants\" and the phrase \"assist ants\", and provides a continuation that is in line with the expected behaviour of ants. In comparison to the golden example, ChatGPT's output may be considered more interesting topic-wise, at the expense of not being as phonetically complex (\"carrying leaves to the ant hill\" contributes heavily to semantics, whilst not being recognisable as part of a tongue twister). For further analysis, please see Appendix A.4." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 580, + 379, + 593 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 580, + 379, + 593 + ], + "spans": [ + { + "bbox": [ + 302, + 580, + 379, + 593 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 609, + 527, + 731 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 609, + 527, + 731 + ], + "spans": [ + { + "bbox": [ + 302, + 609, + 527, + 731 + ], + "type": "text", + "content": "We present work on the topic of tongue twister generation, a form of phonetically-constrained language generation that aims to maximise phonetic overlap, whilst conveying meaningful semantics. We motivate the potential application domains for such generated language, and provide a large annotated dataset of tongue twisters, TwistList, to encourage further work. Finally, we present a series of benchmark models alongside automatic/human evaluation to assess generation quality." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 750, + 518, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 750, + 518, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 518, + 772 + ], + "type": "text", + "content": "11https://en.wikipedia.org/wiki/Buffalo_buffalo_Buffalo_buffalo_buffalo_buffalo_buffalo_buffalo" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "text", + "content": "583" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 70, + 127, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 70, + 127, + 83 + ], + "spans": [ + { + "bbox": [ + 68, + 70, + 127, + 83 + ], + "type": "text", + "content": "Limitations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 96, + 291, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 96, + 291, + 353 + ], + "spans": [ + { + "bbox": [ + 69, + 96, + 291, + 353 + ], + "type": "text", + "content": "Whilst the system presented within this paper is capable of allowing human-in-the-loop contributions (via selecting the input keywords on which to condition the output), it is not able to produce tongue-twisters that take advantage of particular features of speech sounds such as place and manner of articulation, in order to create more advanced outputs that exploit phonetic relatedness (rather than exact matches). The same can be said of our proposed metrics, PO and Init-PO, which do not account for phonetic similarity across sounds that share manner/place of articulation (e.g. \"she sells sea shells\"). Additionally, whilst commonly known tongue twisters may follow a particular format (e.g. rhyme schemes), such schemes and templates have not been enforced here. We also do not demonstrate the capabilities of these systems if they were trained on phonetic transcriptions explicitly, as we only aim to assess their performance when training on graphemes in standard orthography." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 68, + 369, + 153, + 381 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 369, + 153, + 381 + ], + "spans": [ + { + "bbox": [ + 68, + 369, + 153, + 381 + ], + "type": "text", + "content": "Ethics Statement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 394, + 290, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 394, + 290, + 637 + ], + "spans": [ + { + "bbox": [ + 69, + 394, + 290, + 637 + ], + "type": "text", + "content": "All use of human participants in this study has been approved by the Ethics Board of the primary author's institution, including the disclosure of demographic information. Regarding the generation of tongue twisters, language generation is a necessarily creative domain that has the ability to reproduce content that some individuals may find offensive. Care was taken to check outputs in the human evaluation set for any such materials, and if they had been produced, they would have been removed from the evaluation set. Additionally, no egregiously offensive material has been provided in the TwistList dataset. However, the distinction between offensive and humorous content is a highly complex topic, and therefore some examples within the dataset may not be suitable for all individuals (e.g. suggestive content and swearing, such as \"I'm not the pheasant plucker, I'm the pheasant plucker's son\", and the clear relation to common expletives)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 653, + 166, + 666 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 653, + 166, + 666 + ], + "spans": [ + { + "bbox": [ + 68, + 653, + 166, + 666 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 678, + 290, + 760 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 678, + 290, + 760 + ], + "spans": [ + { + "bbox": [ + 67, + 678, + 290, + 760 + ], + "type": "text", + "content": "Tyler Loakman is supported by the Centre for Doctoral Training in Speech and Language Technologies (SLT) and their Applications funded by UK Research and Innovation [grant number EP/S023062/1]. Chen Tang is supported by the China Scholarship Council (CSC) for his doctoral study (File No.202006120039)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 70, + 359, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 70, + 359, + 83 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 359, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 89, + 526, + 772 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 304, + 89, + 525, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 89, + 525, + 146 + ], + "spans": [ + { + "bbox": [ + 304, + 89, + 525, + 146 + ], + "type": "text", + "content": "Rajat Agarwal and Katharina Kann. 2020. Acrostic poem generation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1230-1240, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 153, + 526, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 153, + 526, + 220 + ], + "spans": [ + { + "bbox": [ + 304, + 153, + 526, + 220 + ], + "type": "text", + "content": "Peter Brugger. 2001. From haunted brain to haunted science: A cognitive neuroscience view of paranormal and pseudoscientific thought. In James Hournan and RenseEditors Lange, editors, *Hauntings and Poltergeists: Multidisciplinary Perspectives*, page 195-213. McFarland." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 228, + 525, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 525, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 525, + 251 + ], + "type": "text", + "content": "Joseph L Fleiss. 1971. Measuring nominal scale agreement among many raters. Psychological bulletin, 76(5):378." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 259, + 525, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 259, + 525, + 282 + ], + "spans": [ + { + "bbox": [ + 304, + 259, + 525, + 282 + ], + "type": "text", + "content": "Theodore Seuss Geisel. 1965. *Fox in socks: Dr. Seuss's book of tongue tanglers*. Random House." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 290, + 526, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 290, + 526, + 368 + ], + "spans": [ + { + "bbox": [ + 304, + 290, + 526, + 368 + ], + "type": "text", + "content": "Marco Guerini, Gözde Özbal, and Carlo Strapparava. 2015. Echoes of persuasion: The effect of euphony in persuasive communication. In Proceedings of the 2015 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1483-1493, Denver, Colorado. Association for Computational Linguistics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 375, + 525, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 375, + 525, + 454 + ], + "spans": [ + { + "bbox": [ + 304, + 375, + 525, + 454 + ], + "type": "text", + "content": "He He, Nanyun Peng, and Percy Liang. 2019. Pun generation with surprise. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 1734-1744, Minneapolis, Minnesota. Association for Computational Linguistics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 461, + 525, + 550 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 525, + 550 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 525, + 550 + ], + "type": "text", + "content": "Henglin Huang, Chen Tang, Tyler Loakman, Frank Guerin, and Chenghua Lin. 2022. Improving Chinese story generation via awareness of syntactic dependencies and semantics. In Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 2: Short Papers)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 558, + 525, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 525, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 525, + 602 + ], + "type": "text", + "content": "Sedrick Scott Keh, Steven Y. Feng, Varun Gangal, Malihe Alikhani, and Eduard Hovy. 2022. Pancetta: Phoneme aware neural completion to elicit tongue twisters automatically." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 611, + 525, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 611, + 525, + 655 + ], + "spans": [ + { + "bbox": [ + 304, + 611, + 525, + 655 + ], + "type": "text", + "content": "Heather Kember, Kathryn Connaghan, and Rupal Patel. 2017. Inducing speech errors in dysarthria using tongue twisters. International journal of language & communication disorders, 52(4):469-478." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 664, + 525, + 741 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 664, + 525, + 741 + ], + "spans": [ + { + "bbox": [ + 304, + 664, + 525, + 741 + ], + "type": "text", + "content": "Jey Han Lau, Trevor Cohn, Timothy Baldwin, Julian Brooke, and Adam Hammond. 2018. Deep-spare: A joint neural model of poetic language, meter and rhyme. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1948-1958, Melbourne, Australia. Association for Computational Linguistics." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 750, + 525, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 750, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 750, + 525, + 772 + ], + "type": "text", + "content": "Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy," + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "584" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 290, + 772 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 80, + 72, + 290, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 72, + 290, + 139 + ], + "spans": [ + { + "bbox": [ + 80, + 72, + 290, + 139 + ], + "type": "text", + "content": "Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871-7880, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 149, + 289, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 149, + 289, + 194 + ], + "spans": [ + { + "bbox": [ + 69, + 149, + 289, + 194 + ], + "type": "text", + "content": "Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81, Barcelona, Spain. Association for Computational Linguistics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 205, + 289, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 205, + 289, + 249 + ], + "spans": [ + { + "bbox": [ + 69, + 205, + 289, + 249 + ], + "type": "text", + "content": "Ken D. O'Halloran. 2020. A tongue-twister to translation? increased complexity of genioglossus movement during wakefulness in persons with obstructive sleep apnoea. The Journal of Physiology, 598(3):435-436." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 260, + 289, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 260, + 289, + 327 + ], + "spans": [ + { + "bbox": [ + 69, + 260, + 289, + 327 + ], + "type": "text", + "content": "Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311-318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 338, + 289, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 338, + 289, + 382 + ], + "spans": [ + { + "bbox": [ + 69, + 338, + 289, + 382 + ], + "type": "text", + "content": "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 394, + 289, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 394, + 289, + 449 + ], + "spans": [ + { + "bbox": [ + 69, + 394, + 289, + 449 + ], + "type": "text", + "content": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(140):1-67." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 460, + 289, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 460, + 289, + 493 + ], + "spans": [ + { + "bbox": [ + 69, + 460, + 289, + 493 + ], + "type": "text", + "content": "Victoria Somoff. 2014. Four is not fourteen: Tongue twister patterns and the unmastery of language. Western Folklore, 73(2/3):195-215." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 506, + 289, + 549 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 506, + 289, + 549 + ], + "spans": [ + { + "bbox": [ + 69, + 506, + 289, + 549 + ], + "type": "text", + "content": "Prasetyawan Sugiharto, Yan Santoso, and Maila Shofyana. 2022. Teaching english pronunciation using tongue twister. Acitya: Journal of Teaching and Education, 4(1):189-197." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 560, + 289, + 604 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 560, + 289, + 604 + ], + "spans": [ + { + "bbox": [ + 69, + 560, + 289, + 604 + ], + "type": "text", + "content": "Jiao Sun, Anjali Narayan-Chen, Shereen Oraby, Shuyang Gao, Tagyoung Chung, Jing Huang, Yang Liu, and Nanyun Peng. 2022. Context-situated pun generation. In EMNLP 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 616, + 289, + 671 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 616, + 289, + 671 + ], + "spans": [ + { + "bbox": [ + 69, + 616, + 289, + 671 + ], + "type": "text", + "content": "Chen Tang, Chenghua Lin, Henglin Huang, Frank Guerin, and Zhihao Zhang. 2022a. EtrICA: Event-triggered context-aware story generation augmented by cross attention. In *Findings of the Association for Computational Linguistics: EMNLP* 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 683, + 289, + 717 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 683, + 289, + 717 + ], + "spans": [ + { + "bbox": [ + 69, + 683, + 289, + 717 + ], + "type": "text", + "content": "Chen Tang, Hongbo Zhang, Tyler Loakman, Chenghua Lin, and Frank Guerin. 2022b. Terminology-aware medical dialogue generation. arXiv preprint arXiv:2210.15551." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 69, + 728, + 289, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 728, + 289, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 728, + 289, + 772 + ], + "type": "text", + "content": "Chen Tang, Zhihao Zhang, Tyler Loakman, Chenghua Lin, and Frank Guerin. 2022c. NGEP: A graph-based event planning framework for story generation. In Proceedings of AACL-IJCNLP, Online." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 524, + 772 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 305, + 72, + 524, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 524, + 149 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 524, + 149 + ], + "type": "text", + "content": "Yufei Tian and Nanyun Peng. 2022. Zero-shot sonnet generation with discourse-level planning and aesthetics features. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3587-3597, Seattle, United States. Association for Computational Linguistics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 161, + 524, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 161, + 524, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 161, + 524, + 216 + ], + "type": "text", + "content": "Tim Van de Cruys. 2020. Automatic poetry generation from prosaic text. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2471-2480, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 227, + 524, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 227, + 524, + 261 + ], + "spans": [ + { + "bbox": [ + 304, + 227, + 524, + 261 + ], + "type": "text", + "content": "Carolyn E. Wilshire. 1999. The \"tongue twister\" paradigm as a technique for studying phonological encoding. Language and Speech, 42(1):57-82." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 273, + 524, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 273, + 524, + 372 + ], + "spans": [ + { + "bbox": [ + 304, + 273, + 524, + 372 + ], + "type": "text", + "content": "Jörg Wöckener, Thomas Haider, Tristan Miller, The-Khang Nguyen, Thanh Tung Linh Nguyen, Minh Vu Pham, Jonas Belouadi, and Steffen Eger. 2021. End-to-end style-conditioned poetry generation: What does it take to learn from examples alone? In Proceedings of the 5th Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature, pages 57-66, Punta Cana, Dominican Republic (online). Association for Computational Linguistics." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 383, + 524, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 383, + 524, + 439 + ], + "spans": [ + { + "bbox": [ + 304, + 383, + 524, + 439 + ], + "type": "text", + "content": "Min Ney Wong, Yanky Chan, Manwa L. Ng, and Frank F. Zhu. 2019. Effects of transcranial direct current stimulation over the broca's area on tongue twister production. International Journal of Speech-Language Pathology, 21(2):182-188. PMID: 29642741." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 450, + 524, + 550 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 524, + 550 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 524, + 550 + ], + "type": "text", + "content": "Lanqing Xue, Kaitao Song, Duocai Wu, Xu Tan, Nevin L. Zhang, Tao Qin, Wei-Qiang Zhang, and Tie-Yan Liu. 2021. DeepRapper: Neural rap generation with rhyme and rhythm modeling. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 69-81, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 560, + 524, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 560, + 524, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 560, + 524, + 628 + ], + "type": "text", + "content": "Zhiwei Yu, Jiwei Tan, and Xiaojun Wan. 2018. A neural approach to pun generation. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1650-1660, Melbourne, Australia. Association for Computational Linguistics." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 639, + 524, + 717 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 639, + 524, + 717 + ], + "spans": [ + { + "bbox": [ + 304, + 639, + 524, + 717 + ], + "type": "text", + "content": "Rongsheng Zhang, Xiaoxi Mao, Le Li, Lin Jiang, Lin Chen, Zhiwei Hu, Yadong Xi, Changjie Fan, and Minlie Huang. 2020a. Youling: an AI-assisted lyrics creation system. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 85-91, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 728, + 524, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 728, + 524, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 728, + 524, + 772 + ], + "type": "text", + "content": "Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi. 2020b. *Bertscore: Evaluating text generation with bert*. In International Conference on Learning Representations." + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "585" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 291, + 161 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 291, + 161 + ], + "type": "text", + "content": "Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, and Bill Dolan. 2020c. DIALOGPT: Large-scale generative pre-training for conversational response generation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 270-278, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 171, + 148, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 171, + 148, + 185 + ], + "spans": [ + { + "bbox": [ + 68, + 171, + 148, + 185 + ], + "type": "text", + "content": "A Appendices" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 68, + 193, + 204, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 193, + 204, + 205 + ], + "spans": [ + { + "bbox": [ + 68, + 193, + 204, + 205 + ], + "type": "text", + "content": "A.1 Dataset Quality Control" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 211, + 290, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 211, + 290, + 238 + ], + "spans": [ + { + "bbox": [ + 67, + 211, + 290, + 238 + ], + "type": "text", + "content": "An annotation platform was developed as shown in (Figure 2)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 246, + 188, + 259 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 246, + 188, + 259 + ], + "spans": [ + { + "bbox": [ + 68, + 246, + 188, + 259 + ], + "type": "text", + "content": "A.2 Human Participants" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 264, + 289, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 264, + 289, + 425 + ], + "spans": [ + { + "bbox": [ + 67, + 264, + 289, + 425 + ], + "type": "text", + "content": "Due to tongue twisters being highly reliant on articulation abilities, the demographics of the human participants used within this work are highly important. Additionally, tongue twisters are also a form of humour and entertainment, where individual perceptions of what may or may not be considered humorous or entertaining differ according to numerous factors. In an effort to remain as transparent as possible, and follow best practices for human evaluation, relevant demographic information of participants are outlined below (with the necessary requisite permission and ethical approval)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 433, + 290, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 433, + 290, + 528 + ], + "spans": [ + { + "bbox": [ + 67, + 433, + 290, + 528 + ], + "type": "text", + "content": "Dataset Evaluation All evaluators involved in the quality control process of the TwistList dataset are native speakers of English, and either have or are working towards University level qualifications. Additionally, 2 of the 3 evaluators have extensive education in linguistics or modern languages. No monetary incentive was provided." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 535, + 291, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 535, + 291, + 657 + ], + "spans": [ + { + "bbox": [ + 67, + 535, + 291, + 657 + ], + "type": "text", + "content": "Generation Evaluation All evaluators involved in the evaluation of the quality of generated tongue twisters are native speakers of English, and either hold or are working towards University level qualifications in Linguistics, Modern Languages or NLP. Additionally, all evaluators cited the United Kingdom as their country of socialisation, and no participants reported language processing difficulties that could affect results. No monetary incentive was provided." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 665, + 290, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 665, + 290, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 665, + 290, + 773 + ], + "type": "text", + "content": "Materials Provided to Human Participants Additionally, all evaluators for both the dataset and generation outputs were presented with calibration examples to demonstrate the sort of outputs that would be presented, and the logic behind particular scores, in order to minimise individual interpretations of the scoring criteria. All evaluation was performed on a custom made online annotation platform (Figure 3)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 303, + 71, + 405, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 71, + 405, + 84 + ], + "spans": [ + { + "bbox": [ + 303, + 71, + 405, + 84 + ], + "type": "text", + "content": "A.3 Training Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 89, + 525, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 89, + 525, + 293 + ], + "spans": [ + { + "bbox": [ + 302, + 89, + 525, + 293 + ], + "type": "text", + "content": "All pre-trained models used (naturally excluding ChatGPT) are based on publicly available checkpoints from Hugging Face.12 Models are trained for up to 5 epochs on a Tesla A5000 machine with the best checkpoints selected based on the validation loss. The batch size is set to 32, and the learning rate is " + }, + { + "bbox": [ + 302, + 89, + 525, + 293 + ], + "type": "inline_equation", + "content": "8e^{-5}" + }, + { + "bbox": [ + 302, + 89, + 525, + 293 + ], + "type": "text", + "content": ", with the Adam optimiser selected for training. To help the loss curve converge on our small few-shot dataset, we limit the generation length to 100 (covering all test tongue twisters). Meanwhile, the source length is limited to 150. The training and testing steps are set up with the implementation of the PyTorch Lightning13 framework to guarantee the reliability of the experiment. All language models are fairly trained and tested with the same steps." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 303, + 470, + 315 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 303, + 470, + 315 + ], + "spans": [ + { + "bbox": [ + 302, + 303, + 470, + 315 + ], + "type": "text", + "content": "A.4 Further Qualitative Comments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 321, + 525, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 321, + 525, + 632 + ], + "spans": [ + { + "bbox": [ + 302, + 321, + 525, + 632 + ], + "type": "text", + "content": "Whilst the pattern of extreme word repetition is seen in many of the finetuned models (often with the exception of BART, which is demonstrated to be capable of producing slightly more sophisticated outputs), overall assessment of the tongue twisters produced at inference time reveals interesting patterns, particularly in regard to ChatGPT outputs. Firstly, the limits of ChatGPT are made apparent in a few examples such as the input \"silver shiny ship sank\" generating \"How much wood would a woodchuck chuck if a woodchuck could chuck silver shiny ships?\", a clear derivation of a famous woodchuck related tongue twister that it is rather safe to assume appears multiple times in ChatGPTs training material. Additionally, comments from evaluators also reveal that ChatGPT's output is often considered more of a rhyme or general literary text, rather than specifically a tongue twister. However, examples such as these are also found in the human-authored golden examples, demonstrating that there is no steadfast consistent opinion as to what constitutes a (good) tongue twister. Likewise, some examples may contain large amounts of sound repetition, but not in a way that necessarily presents articulatory difficulty." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 303, + 643, + 395, + 655 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 643, + 395, + 655 + ], + "spans": [ + { + "bbox": [ + 303, + 643, + 395, + 655 + ], + "type": "text", + "content": "A.5 Future Works" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 661, + 524, + 741 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 661, + 524, + 741 + ], + "spans": [ + { + "bbox": [ + 302, + 661, + 524, + 741 + ], + "type": "text", + "content": "In this paper, we mainly analyse the performance of large-scale pretrained language models (PLMs) on Tongue Twister Generation, and propose a corresponding dataset for further investigation. In further works, we aim to propose novel models which can better leverage phonetic symbols. There" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 749, + 446, + 761 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 749, + 446, + 761 + ], + "spans": [ + { + "bbox": [ + 313, + 749, + 446, + 761 + ], + "type": "text", + "content": "12https://huggingface.co/models" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 761, + 458, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 761, + 458, + 772 + ], + "spans": [ + { + "bbox": [ + 314, + 761, + 458, + 772 + ], + "type": "text", + "content": "13https://www.pytorchlightning.ai/" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "586" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 82, + 87, + 110, + 93 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 87, + 110, + 93 + ], + "spans": [ + { + "bbox": [ + 82, + 87, + 110, + 93 + ], + "type": "text", + "content": "Introduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 86, + 101, + 258, + 130 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 86, + 101, + 215, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 101, + 215, + 106 + ], + "spans": [ + { + "bbox": [ + 86, + 101, + 215, + 106 + ], + "type": "text", + "content": "1. Individually read the tongue twister, phonetics, and key words on the left side." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 89, + 108, + 243, + 113 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 108, + 243, + 113 + ], + "spans": [ + { + "bbox": [ + 89, + 108, + 243, + 113 + ], + "type": "text", + "content": "Select the options on the right side to evaluate the data quality from the following perspectives:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 90, + 114, + 242, + 118 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 114, + 242, + 118 + ], + "spans": [ + { + "bbox": [ + 90, + 114, + 242, + 118 + ], + "type": "text", + "content": "- The quality of the RAKE Keywords: Do these suitably represent the topic of the tongue twister?" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 90, + 119, + 249, + 123 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 119, + 249, + 123 + ], + "spans": [ + { + "bbox": [ + 90, + 119, + 249, + 123 + ], + "type": "text", + "content": "- The quality of the BERTopic Keywords: Do these suitably represent the topic of the tongue twister?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 90, + 124, + 258, + 130 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 124, + 258, + 130 + ], + "spans": [ + { + "bbox": [ + 90, + 124, + 258, + 130 + ], + "type": "text", + "content": "- The quality of the Tongue Twister: Is it a good tongue twister, or too short/long, or generally bad quality?" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 86, + 141, + 116, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 141, + 116, + 147 + ], + "spans": [ + { + "bbox": [ + 86, + 141, + 116, + 147 + ], + "type": "text", + "content": "Tongue Twisters" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 91, + 158, + 144, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 158, + 144, + 163 + ], + "spans": [ + { + "bbox": [ + 91, + 158, + 144, + 163 + ], + "type": "text", + "content": "Keywords for Tongue Twisters" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 94, + 174, + 123, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 174, + 123, + 179 + ], + "spans": [ + { + "bbox": [ + 94, + 174, + 123, + 179 + ], + "type": "text", + "content": "RAKE Keywords:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 174, + 189, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 174, + 189, + 179 + ], + "spans": [ + { + "bbox": [ + 132, + 174, + 189, + 179 + ], + "type": "text", + "content": "pickled peppers peter piper picked" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 179, + 245, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 179, + 245, + 185 + ], + "spans": [ + { + "bbox": [ + 132, + 179, + 245, + 185 + ], + "type": "text", + "content": "[P I H1 K A H0 L D] [P E H1 P E R O Z] [P I Y1 T E R O] [P A Y1 P E R O] [P I H1 K T]" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 94, + 191, + 111, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 191, + 111, + 196 + ], + "spans": [ + { + "bbox": [ + 94, + 191, + 111, + 196 + ], + "type": "text", + "content": "BERTopic" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 94, + 197, + 113, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 197, + 113, + 202 + ], + "spans": [ + { + "bbox": [ + 94, + 197, + 113, + 202 + ], + "type": "text", + "content": "Keywords:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 192, + 168, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 168, + 196 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 168, + 196 + ], + "type": "text", + "content": "pink peter piper peck" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 197, + 210, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 197, + 210, + 203 + ], + "spans": [ + { + "bbox": [ + 132, + 197, + 210, + 203 + ], + "type": "text", + "content": "[P I H1NGK][P I Y1TERO][P AY1PER0][P E H1K]" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 91, + 217, + 119, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 217, + 119, + 223 + ], + "spans": [ + { + "bbox": [ + 91, + 217, + 119, + 223 + ], + "type": "text", + "content": "Tongue Twister" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 94, + 235, + 121, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 235, + 121, + 241 + ], + "spans": [ + { + "bbox": [ + 94, + 235, + 121, + 241 + ], + "type": "text", + "content": "Tongue Twister" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 94, + 242, + 104, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 242, + 104, + 246 + ], + "spans": [ + { + "bbox": [ + 94, + 242, + 104, + 246 + ], + "type": "text", + "content": "Text:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 234, + 282, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 234, + 282, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 234, + 282, + 251 + ], + "type": "text", + "content": "Peter Piper picked a peck of pickled peppers. A peck of pickled peppers Peter Piper picked. If Peter Piper picked a peck of pickled peppers, Where's the peck of pickled peppers Peter Piper picked?" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 94, + 258, + 121, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 258, + 121, + 263 + ], + "spans": [ + { + "bbox": [ + 94, + 258, + 121, + 263 + ], + "type": "text", + "content": "Tongue Twister" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 94, + 264, + 113, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 264, + 113, + 268 + ], + "spans": [ + { + "bbox": [ + 94, + 264, + 113, + 268 + ], + "type": "text", + "content": "Phonetics:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 132, + 258, + 282, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 258, + 282, + 286 + ], + "spans": [ + { + "bbox": [ + 132, + 258, + 282, + 286 + ], + "type": "text", + "content": "[PI Y1 ER0] [PY A1 ER0] [PI H1 K T] [AOH] [PE H1 K] [AHV] [PI I H1 K AI O L D] [PE H1 ER0 Z] [JI AHO] [PE H1 K] [AHV] [PI I H1 K AI O L D] [PE H1 ER0 Z] [PI Y1 ER0] [PY A1 ER0] [PI I H1 K T] [JI (IH1 F) [PI Y1 T ER0] [PY A1 ER0] [PI I H1 K T] [AOH] [PE H1 K] [AHV] [PI I H1 K AI O L D] [PE H1 ER0 Z] [WEH1 K R2] [DH AHO] [PE H1 K] [AHV] [PI I H1 K AI O L D] [PE H1 ER0 Z] [PI Y1 T ER0] [PY A1 ER0] [PI I H1 K T] [Q]:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 303, + 88, + 331, + 94 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 88, + 331, + 94 + ], + "spans": [ + { + "bbox": [ + 303, + 88, + 331, + 94 + ], + "type": "text", + "content": "Sample Quality" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 104, + 424, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 104, + 424, + 111 + ], + "spans": [ + { + "bbox": [ + 307, + 104, + 424, + 111 + ], + "type": "text", + "content": "A1: Are the RAKE Keywords highly suitable for the Tongue Twister?" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 309, + 118, + 321, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 118, + 321, + 122 + ], + "spans": [ + { + "bbox": [ + 309, + 118, + 321, + 122 + ], + "type": "text", + "content": "Yes" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 309, + 123, + 321, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 123, + 321, + 128 + ], + "spans": [ + { + "bbox": [ + 309, + 123, + 321, + 128 + ], + "type": "text", + "content": "No" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 145, + 431, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 145, + 431, + 151 + ], + "spans": [ + { + "bbox": [ + 307, + 145, + 431, + 151 + ], + "type": "text", + "content": "A2: Are the BERTopic Keywords highly suitable for the Tongue Twister?" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 309, + 158, + 320, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 158, + 320, + 163 + ], + "spans": [ + { + "bbox": [ + 309, + 158, + 320, + 163 + ], + "type": "text", + "content": "Yes" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 309, + 164, + 320, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 164, + 320, + 169 + ], + "spans": [ + { + "bbox": [ + 309, + 164, + 320, + 169 + ], + "type": "text", + "content": "No" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 185, + 362, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 362, + 191 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 362, + 191 + ], + "type": "text", + "content": "A3: Is it a good tongue twister?" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 309, + 199, + 339, + 221 + ], + "type": "list", + "angle": 0, + "index": 37, + "blocks": [ + { + "bbox": [ + 309, + 199, + 324, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 199, + 324, + 203 + ], + "spans": [ + { + "bbox": [ + 309, + 199, + 324, + 203 + ], + "type": "text", + "content": "Good" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 309, + 205, + 339, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 205, + 339, + 209 + ], + "spans": [ + { + "bbox": [ + 309, + 205, + 339, + 209 + ], + "type": "text", + "content": "A Bit Too Short" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 309, + 210, + 337, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 210, + 337, + 214 + ], + "spans": [ + { + "bbox": [ + 309, + 210, + 337, + 214 + ], + "type": "text", + "content": "A Bit Too Long" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 309, + 215, + 334, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 215, + 334, + 221 + ], + "spans": [ + { + "bbox": [ + 309, + 215, + 334, + 221 + ], + "type": "inline_equation", + "content": "\\bigcirc" + }, + { + "bbox": [ + 309, + 215, + 334, + 221 + ], + "type": "text", + "content": " Bad Quality" + } + ] + } + ], + "index": 36 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 307, + 238, + 489, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 238, + 489, + 249 + ], + "spans": [ + { + "bbox": [ + 307, + 238, + 489, + 249 + ], + "type": "text", + "content": "A4: Please input up to 5 manual keywords for the tongue twister (these can either be extracted from the example, or any other words)." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 309, + 257, + 327, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 257, + 327, + 262 + ], + "spans": [ + { + "bbox": [ + 309, + 257, + 327, + 262 + ], + "type": "text", + "content": "Keywords" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 305, + 294, + 315, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 294, + 315, + 300 + ], + "spans": [ + { + "bbox": [ + 305, + 294, + 315, + 300 + ], + "type": "text", + "content": "Quit" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 491, + 294, + 505, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 491, + 294, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 491, + 294, + 505, + 300 + ], + "type": "text", + "content": "Submit" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 188, + 315, + 405, + 327 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 315, + 405, + 327 + ], + "spans": [ + { + "bbox": [ + 188, + 315, + 405, + 327 + ], + "type": "text", + "content": "Figure 2: TwistList Quality Control Annotation Platform" + } + ] + } + ], + "index": 42, + "type": "text" + }, + { + "bbox": [ + 82, + 339, + 158, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 339, + 158, + 348 + ], + "spans": [ + { + "bbox": [ + 82, + 339, + 158, + 348 + ], + "type": "text", + "content": "Tongue Twister Evaluation" + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 464, + 340, + 511, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 464, + 340, + 511, + 346 + ], + "spans": [ + { + "bbox": [ + 464, + 340, + 511, + 346 + ], + "type": "text", + "content": "Home / Sample Annotation" + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 85, + 355, + 108, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 355, + 108, + 361 + ], + "spans": [ + { + "bbox": [ + 85, + 355, + 108, + 361 + ], + "type": "text", + "content": "Introduction" + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 85, + 369, + 271, + 414 + ], + "type": "list", + "angle": 0, + "index": 54, + "blocks": [ + { + "bbox": [ + 85, + 369, + 208, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 369, + 208, + 374 + ], + "spans": [ + { + "bbox": [ + 85, + 369, + 208, + 374 + ], + "type": "text", + "content": "1. Individually read the input keywords and the tongue twister on the left side." + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 85, + 375, + 235, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 375, + 235, + 380 + ], + "spans": [ + { + "bbox": [ + 85, + 375, + 235, + 380 + ], + "type": "text", + "content": "2. Give a score for each metric on the right to evaluate the quality of generated tongue twisters:" + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 85, + 381, + 257, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 381, + 257, + 386 + ], + "spans": [ + { + "bbox": [ + 85, + 381, + 257, + 386 + ], + "type": "text", + "content": "- Relevance: The extent to which the tongue twister is remantically/topically related to the input keywords." + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 85, + 387, + 271, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 387, + 271, + 391 + ], + "spans": [ + { + "bbox": [ + 85, + 387, + 271, + 391 + ], + "type": "text", + "content": "- Difficulty of Articulation: The extent to which the tongue twister is hard to say (aka. how much your tongue twists)." + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 85, + 392, + 238, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 392, + 238, + 397 + ], + "spans": [ + { + "bbox": [ + 85, + 392, + 238, + 397 + ], + "type": "text", + "content": "- Fluency: The extent to which the tongue twister can be considered grammatically acceptable." + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 85, + 398, + 258, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 398, + 258, + 403 + ], + "spans": [ + { + "bbox": [ + 85, + 398, + 258, + 403 + ], + "type": "text", + "content": "- Coherence: The extent to which the tongue twister can be considered logically and semantically coherent." + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 85, + 404, + 271, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 404, + 271, + 409 + ], + "spans": [ + { + "bbox": [ + 85, + 404, + 271, + 409 + ], + "type": "text", + "content": "Entertainment: The extent to which the tongue twister is considered enteratining (primarily relating to phonetics +" + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 85, + 410, + 103, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 410, + 103, + 414 + ], + "spans": [ + { + "bbox": [ + 85, + 410, + 103, + 414 + ], + "type": "text", + "content": "semantics)" + } + ] + } + ], + "index": 53 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 85, + 425, + 114, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 425, + 114, + 431 + ], + "spans": [ + { + "bbox": [ + 85, + 425, + 114, + 431 + ], + "type": "text", + "content": "Tongue Twisters" + } + ] + } + ], + "index": 55 + }, + { + "bbox": [ + 90, + 443, + 119, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 443, + 119, + 448 + ], + "spans": [ + { + "bbox": [ + 90, + 443, + 119, + 448 + ], + "type": "text", + "content": "Input Keywords" + } + ] + } + ], + "index": 56 + }, + { + "bbox": [ + 93, + 459, + 121, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 459, + 121, + 464 + ], + "spans": [ + { + "bbox": [ + 93, + 459, + 121, + 464 + ], + "type": "text", + "content": "RAKE Keywords:" + } + ] + } + ], + "index": 57 + }, + { + "bbox": [ + 132, + 459, + 209, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 459, + 209, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 459, + 209, + 464 + ], + "type": "text", + "content": "Generate tongue twisters about key words: nope" + } + ] + } + ], + "index": 58 + }, + { + "bbox": [ + 90, + 479, + 117, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 479, + 117, + 484 + ], + "spans": [ + { + "bbox": [ + 90, + 479, + 117, + 484 + ], + "type": "text", + "content": "Tongue Twister" + } + ] + } + ], + "index": 59 + }, + { + "bbox": [ + 93, + 496, + 120, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 496, + 120, + 506 + ], + "spans": [ + { + "bbox": [ + 93, + 496, + 120, + 506 + ], + "type": "text", + "content": "Tongue Twister Text:" + } + ] + } + ], + "index": 60 + }, + { + "bbox": [ + 132, + 495, + 210, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 495, + 210, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 495, + 210, + 500 + ], + "type": "text", + "content": "Nope, an antelope can't elope with a cantaloupe." + } + ] + } + ], + "index": 61 + }, + { + "bbox": [ + 132, + 501, + 210, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 210, + 506 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 210, + 506 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 62 + }, + { + "bbox": [ + 301, + 354, + 415, + 362 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 354, + 415, + 362 + ], + "spans": [ + { + "bbox": [ + 301, + 354, + 415, + 362 + ], + "type": "text", + "content": "Evaluate the Quality of Tongue Twisters by giving a Score (1 to 5)" + } + ] + } + ], + "index": 63 + }, + { + "bbox": [ + 307, + 372, + 384, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 372, + 384, + 378 + ], + "spans": [ + { + "bbox": [ + 307, + 372, + 384, + 378 + ], + "type": "text", + "content": "Please select all options before submission." + } + ] + } + ], + "index": 64 + }, + { + "bbox": [ + 309, + 386, + 334, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 386, + 334, + 390 + ], + "spans": [ + { + "bbox": [ + 309, + 386, + 334, + 390 + ], + "type": "text", + "content": "O1:Relevance." + } + ] + } + ], + "index": 65 + }, + { + "bbox": [ + 309, + 391, + 320, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 391, + 320, + 396 + ], + "spans": [ + { + "bbox": [ + 309, + 391, + 320, + 396 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 66 + }, + { + "bbox": [ + 309, + 406, + 358, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 406, + 358, + 410 + ], + "spans": [ + { + "bbox": [ + 309, + 406, + 358, + 410 + ], + "type": "text", + "content": "O2: Difficulty of articulation." + } + ] + } + ], + "index": 67 + }, + { + "bbox": [ + 309, + 415, + 320, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 415, + 320, + 419 + ], + "spans": [ + { + "bbox": [ + 309, + 415, + 320, + 419 + ], + "type": "text", + "content": "0" + } + ] + } + ], + "index": 68 + }, + { + "bbox": [ + 309, + 427, + 331, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 427, + 331, + 431 + ], + "spans": [ + { + "bbox": [ + 309, + 427, + 331, + 431 + ], + "type": "text", + "content": "Q3:FLuency." + } + ] + } + ], + "index": 69 + }, + { + "bbox": [ + 309, + 438, + 320, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 438, + 320, + 442 + ], + "spans": [ + { + "bbox": [ + 309, + 438, + 320, + 442 + ], + "type": "text", + "content": "0" + } + ] + } + ], + "index": 70 + }, + { + "bbox": [ + 309, + 448, + 334, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 448, + 334, + 452 + ], + "spans": [ + { + "bbox": [ + 309, + 448, + 334, + 452 + ], + "type": "text", + "content": "Q4: Coherence." + } + ] + } + ], + "index": 71 + }, + { + "bbox": [ + 309, + 459, + 320, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 459, + 320, + 463 + ], + "spans": [ + { + "bbox": [ + 309, + 459, + 320, + 463 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 72 + }, + { + "bbox": [ + 309, + 468, + 342, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 468, + 342, + 472 + ], + "spans": [ + { + "bbox": [ + 309, + 468, + 342, + 472 + ], + "type": "text", + "content": "Q5: Entertainment." + } + ] + } + ], + "index": 73 + }, + { + "bbox": [ + 309, + 480, + 320, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 480, + 320, + 484 + ], + "spans": [ + { + "bbox": [ + 309, + 480, + 320, + 484 + ], + "type": "text", + "content": "。" + } + ] + } + ], + "index": 74 + }, + { + "bbox": [ + 305, + 505, + 314, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 505, + 314, + 510 + ], + "spans": [ + { + "bbox": [ + 305, + 505, + 314, + 510 + ], + "type": "text", + "content": "Quit" + } + ] + } + ], + "index": 75 + }, + { + "bbox": [ + 490, + 505, + 505, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 490, + 505, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 490, + 505, + 505, + 510 + ], + "type": "text", + "content": "Submit" + } + ] + } + ], + "index": 76 + }, + { + "bbox": [ + 179, + 527, + 413, + 539 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 527, + 413, + 539 + ], + "spans": [ + { + "bbox": [ + 179, + 527, + 413, + 539 + ], + "type": "text", + "content": "Figure 3: Human Evaluation Platform for Generated Outputs" + } + ] + } + ], + "index": 77, + "type": "text" + }, + { + "bbox": [ + 67, + 560, + 290, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 560, + 290, + 709 + ], + "spans": [ + { + "bbox": [ + 67, + 560, + 290, + 709 + ], + "type": "text", + "content": "are numerous existing works (Huang et al., 2022; Tang et al., 2022a,b) that provide approaches for injecting such knowledge into PLMs. However, the phonetic features differ from these text-format knowledge items, as phonemes are hard to encode with input text tokens when feeding into PLM encoders. Another promising approach is to explicitly model the phonetic features into text sequences (Tang et al., 2022c), though there is no observed method for transforming phonetic notation. We intend to perform further research based on these existing approaches." + } + ] + } + ], + "index": 78 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 82, + 71, + 183, + 79 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 71, + 183, + 79 + ], + "spans": [ + { + "bbox": [ + 82, + 71, + 183, + 79 + ], + "type": "text", + "content": "Tongue Twister Dataset Evaluation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 464, + 72, + 511, + 77 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 464, + 72, + 511, + 77 + ], + "spans": [ + { + "bbox": [ + 464, + 72, + 511, + 77 + ], + "type": "text", + "content": "Home / Sample Annotation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "587" + } + ] + } + ], + "index": 79 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 90, + 192, + 103 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 90, + 192, + 103 + ], + "spans": [ + { + "bbox": [ + 68, + 90, + 192, + 103 + ], + "type": "text", + "content": "A For every submission:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 106, + 317, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 106, + 317, + 121 + ], + "spans": [ + { + "bbox": [ + 77, + 106, + 317, + 121 + ], + "type": "text", + "content": "A1. Did you describe the limitations of your work?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 121, + 505, + 134 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 121, + 505, + 134 + ], + "spans": [ + { + "bbox": [ + 89, + 121, + 505, + 134 + ], + "type": "text", + "content": "Yes, in the required Limitations section as well as Section 4 (concerning our proposed metrics)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 142, + 329, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 142, + 329, + 157 + ], + "spans": [ + { + "bbox": [ + 77, + 142, + 329, + 157 + ], + "type": "text", + "content": "A2. Did you discuss any potential risks of your work?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 158, + 166, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 158, + 166, + 169 + ], + "spans": [ + { + "bbox": [ + 89, + 158, + 166, + 169 + ], + "type": "text", + "content": "Ethics Statement" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 77, + 179, + 414, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 179, + 414, + 193 + ], + "spans": [ + { + "bbox": [ + 77, + 179, + 414, + 193 + ], + "type": "text", + "content": "A3. Do the abstract and introduction summarize the paper's main claims?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 89, + 194, + 400, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 194, + 400, + 206 + ], + "spans": [ + { + "bbox": [ + 89, + 194, + 400, + 206 + ], + "type": "text", + "content": "Abstract (all) and contribution summary at the end of the introduction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 77, + 216, + 398, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 216, + 398, + 229 + ], + "spans": [ + { + "bbox": [ + 77, + 216, + 398, + 229 + ], + "type": "text", + "content": "A4. Have you used AI writing assistants when working on this paper?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 89, + 230, + 138, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 230, + 138, + 242 + ], + "spans": [ + { + "bbox": [ + 89, + 230, + 138, + 242 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 68, + 252, + 290, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 252, + 290, + 266 + ], + "spans": [ + { + "bbox": [ + 68, + 252, + 290, + 266 + ], + "type": "text", + "content": "B Did you use or create scientific artifacts?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 79, + 270, + 214, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 270, + 214, + 283 + ], + "spans": [ + { + "bbox": [ + 79, + 270, + 214, + 283 + ], + "type": "text", + "content": "TwistList dataset (Section 3.2)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 77, + 291, + 315, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 291, + 315, + 306 + ], + "spans": [ + { + "bbox": [ + 77, + 291, + 315, + 306 + ], + "type": "text", + "content": "B1. Did you cite the creators of artifacts you used?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 89, + 306, + 433, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 306, + 433, + 319 + ], + "spans": [ + { + "bbox": [ + 89, + 306, + 433, + 319 + ], + "type": "text", + "content": "Sources of all entries in the dataset are credited in the .json file for each entry." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 77, + 328, + 464, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 328, + 464, + 342 + ], + "spans": [ + { + "bbox": [ + 77, + 328, + 464, + 342 + ], + "type": "text", + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 89, + 343, + 524, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 343, + 524, + 423 + ], + "spans": [ + { + "bbox": [ + 89, + 343, + 524, + 423 + ], + "type": "text", + "content": "We did not discuss the licensing around our dataset. The dataset uses works that are freely available on the web and come from various sources such as websites, blogs, and eBooks. Many of these cases are Public Domain, and for those that are not, we believe we are in accordance with Fair Use, as the dataset does not encroach on the use case of the original works (no graphic design/other elements are maintained) and the dataset is for use as a research tool only. We will also reply promptly to any cases of copyright infringement that relevant copyright holders make us aware of." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 77, + 432, + 524, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 432, + 524, + 486 + ], + "spans": [ + { + "bbox": [ + 77, + 432, + 524, + 486 + ], + "type": "text", + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 89, + 487, + 170, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 487, + 170, + 498 + ], + "spans": [ + { + "bbox": [ + 89, + 487, + 170, + 498 + ], + "type": "text", + "content": "See answer to B2." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 77, + 508, + 524, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 508, + 524, + 549 + ], + "spans": [ + { + "bbox": [ + 77, + 508, + 524, + 549 + ], + "type": "text", + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 89, + 550, + 524, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 550, + 524, + 576 + ], + "spans": [ + { + "bbox": [ + 89, + 550, + 524, + 576 + ], + "type": "text", + "content": "See the Ethics Statement regarding the potential for tongue twisters to be offensive. Additionally, all tongue twisters are believed to be about fictional characters, rather than individuals." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 77, + 586, + 524, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 586, + 524, + 613 + ], + "spans": [ + { + "bbox": [ + 77, + 586, + 524, + 613 + ], + "type": "text", + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 89, + 613, + 524, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 613, + 524, + 653 + ], + "spans": [ + { + "bbox": [ + 89, + 613, + 524, + 653 + ], + "type": "text", + "content": "Such details are not explicitly stated. However, it can be easily ascertained from the paper that the tongue twisters we focus on are entirely in English (and the range of domains the tongue twisters were taken from can be seen in the \"source\" entry for each example)." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 77, + 661, + 524, + 729 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 661, + 524, + 729 + ], + "spans": [ + { + "bbox": [ + 77, + 661, + 524, + 729 + ], + "type": "text", + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 89, + 730, + 235, + 743 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 730, + 235, + 743 + ], + "spans": [ + { + "bbox": [ + 89, + 730, + 235, + 743 + ], + "type": "text", + "content": "See Table 1 for dataset statistics." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "spans": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "text", + "content": "ACL 2023 Responsible NLP Checklist" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 746, + 522, + 767 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 746, + 522, + 767 + ], + "spans": [ + { + "bbox": [ + 67, + 746, + 522, + 767 + ], + "type": "text", + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "588" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 70, + 294, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 70, + 294, + 84 + ], + "spans": [ + { + "bbox": [ + 68, + 70, + 294, + 84 + ], + "type": "text", + "content": "C Did you run computational experiments?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 79, + 89, + 162, + 102 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 89, + 162, + 102 + ], + "spans": [ + { + "bbox": [ + 79, + 89, + 162, + 102 + ], + "type": "text", + "content": "Section 4 (page 3)" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 110, + 523, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 110, + 523, + 137 + ], + "spans": [ + { + "bbox": [ + 77, + 110, + 523, + 137 + ], + "type": "text", + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 138, + 151, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 138, + 151, + 152 + ], + "spans": [ + { + "bbox": [ + 89, + 138, + 151, + 152 + ], + "type": "text", + "content": "Appendix A.3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 159, + 524, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 159, + 524, + 187 + ], + "spans": [ + { + "bbox": [ + 77, + 159, + 524, + 187 + ], + "type": "text", + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 188, + 151, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 188, + 151, + 201 + ], + "spans": [ + { + "bbox": [ + 89, + 188, + 151, + 201 + ], + "type": "text", + "content": "Appendix A.3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 77, + 209, + 524, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 209, + 524, + 249 + ], + "spans": [ + { + "bbox": [ + 77, + 209, + 524, + 249 + ], + "type": "text", + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 89, + 251, + 299, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 251, + 299, + 264 + ], + "spans": [ + { + "bbox": [ + 89, + 251, + 299, + 264 + ], + "type": "text", + "content": "Tables 3/5. Scores are the mean, as is standard." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 77, + 273, + 524, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 273, + 524, + 312 + ], + "spans": [ + { + "bbox": [ + 77, + 273, + 524, + 312 + ], + "type": "text", + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 88, + 314, + 524, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 314, + 524, + 354 + ], + "spans": [ + { + "bbox": [ + 88, + 314, + 524, + 354 + ], + "type": "text", + "content": "Exact details of evaluation implementations (except Phonetic Overlap) were not detailed. This is in part due to these metrics (BLEU/ROUGE/BERTScore) not being very reliable for creative language generation, and therefore the exact values from different implementations are not likely to be of use." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 68, + 362, + 521, + 377 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 362, + 521, + 377 + ], + "spans": [ + { + "bbox": [ + 68, + 362, + 521, + 377 + ], + "type": "text", + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 79, + 381, + 320, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 381, + 320, + 394 + ], + "spans": [ + { + "bbox": [ + 79, + 381, + 320, + 394 + ], + "type": "text", + "content": "Section 3.2 and Section 4. In addition to Appendix A.2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 77, + 402, + 524, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 402, + 524, + 429 + ], + "spans": [ + { + "bbox": [ + 77, + 402, + 524, + 429 + ], + "type": "text", + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 89, + 431, + 478, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 431, + 478, + 444 + ], + "spans": [ + { + "bbox": [ + 89, + 431, + 478, + 444 + ], + "type": "text", + "content": "Screenshot of the annotation platforms can be found in Figures 2 and 3 in the Appendix" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 77, + 452, + 524, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 452, + 524, + 492 + ], + "spans": [ + { + "bbox": [ + 77, + 452, + 524, + 492 + ], + "type": "text", + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 88, + 494, + 524, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 494, + 524, + 534 + ], + "spans": [ + { + "bbox": [ + 88, + 494, + 524, + 534 + ], + "type": "text", + "content": "We declared that no monetary incentive was given to participants. We did not specify the recruitment process, but due to participants all holding or working towards university level qualifications, it can be inferred that they are colleagues." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 77, + 543, + 524, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 543, + 524, + 582 + ], + "spans": [ + { + "bbox": [ + 77, + 543, + 524, + 582 + ], + "type": "text", + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 88, + 584, + 524, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 584, + 524, + 637 + ], + "spans": [ + { + "bbox": [ + 88, + 584, + 524, + 637 + ], + "type": "text", + "content": "This information was not deemed necessary in the submitted paper (due to the limited risk of the data we were working with). However, it is stated in the Ethical Statement and Appendix A.2 that all shared information about human demographics was collected with the necessary permissions and approval." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 77, + 645, + 521, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 645, + 521, + 660 + ], + "spans": [ + { + "bbox": [ + 77, + 645, + 521, + 660 + ], + "type": "text", + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 89, + 661, + 524, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 661, + 524, + 686 + ], + "spans": [ + { + "bbox": [ + 89, + 661, + 524, + 686 + ], + "type": "text", + "content": "Ethical approval was gained for human evaluation of the dataset and generated outputs from the relevant institution" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 77, + 696, + 524, + 722 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 696, + 524, + 722 + ], + "spans": [ + { + "bbox": [ + 77, + 696, + 524, + 722 + ], + "type": "text", + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 89, + 724, + 430, + 737 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 724, + 430, + 737 + ], + "spans": [ + { + "bbox": [ + 89, + 724, + 430, + 737 + ], + "type": "text", + "content": "We provide demographic information for human participants in Appendix A.2" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "589" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Typo-Robust Representation Learning for Dense Retrieval/8977932e-2af7-46c1-9920-a3e3abaf9a04_content_list.json b/2023/Typo-Robust Representation Learning for Dense Retrieval/8977932e-2af7-46c1-9920-a3e3abaf9a04_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a1bdd8334b7a44f0b2aaa9c032aaa9410656aa08 --- /dev/null +++ b/2023/Typo-Robust Representation Learning for Dense Retrieval/8977932e-2af7-46c1-9920-a3e3abaf9a04_content_list.json @@ -0,0 +1,1877 @@ +[ + { + "type": "text", + "text": "Typo-Robust Representation Learning for Dense Retrieval", + "text_level": 1, + "bbox": [ + 194, + 84, + 801, + 104 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Panuthep Tasawong†, Wuttikorn Ponwitayarat†, Peerat Limkonchotiwat†, Can Udomcharoenchaikit†, Ekapol Chuangsuwanich‡, Sarana Nutanong†", + "bbox": [ + 181, + 112, + 825, + 147 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "†School of Information Science and Technology, VISTEC, Thailand \n‡Department of Computer Engineering, Chulalongkorn University, Thailand {panuthep.t_s20, wuttikorn.p_s22, peerat.l_s19, canu_pro, snutanon} @vistec.ac.th, ekapolc@cp.eng.chula.ac.th", + "bbox": [ + 191, + 147, + 811, + 230 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 252, + 339, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dense retrieval is a basic building block of information retrieval applications. One of the main challenges of dense retrieval in real-world settings is the handling of queries containing misspelled words. A popular approach for handling misspelled queries is minimizing the representations discrepancy between misspelled queries and their pristine ones. Unlike the existing approaches, which only focus on the alignment between misspelled and pristine queries, our method also improves the contrast between each misspelled query and its surrounding queries. To assess the effectiveness of our proposed method, we compare it against the existing competitors using two benchmark datasets and two base encoders. Our method outperforms the competitors in all cases with misspelled queries. Our code and models are available at https://github.com/panuthept/DST-DenseRetrieval.", + "bbox": [ + 141, + 277, + 463, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 114, + 571, + 260, + 587 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dense retrieval is a fundamental component in many information retrieval applications, such as open-domain question answering and ad-hoc retrieval. The objective is to score and rank a large collection of candidate passages based on their similarity to a given query. The performance of dense retrieval relies on representation learning. A popular approach is to finetune a pre-trained language model to create an embedding space that puts each query closer to its corresponding passages (Zhan et al., 2020; Khattab and Zaharia, 2020; Xiong et al., 2021; Qu et al., 2021; Ren et al., 2021a,b).", + "bbox": [ + 112, + 596, + 489, + 789 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One of the major challenges of dense retrieval is the handling of misspelled queries which induces representations of the misspelled queries to be closer to irrelevant passages than their corresponding passages. Several studies have demonstrated that misspellings in search queries can substantially degrade retrieval performance (Zhuang and Zuccon, 2021; Penha et al., 2022), specifically", + "bbox": [ + 112, + 791, + 489, + 919 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "when informative terms, such as entity mentions, are misspelled (Sidiropoulos and Kanoulas, 2022).", + "bbox": [ + 507, + 253, + 884, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To create a retrieval model that is capable of handling misspelled queries, researchers have proposed different training methods to align representations of misspelled queries with their pristine ones. Zhuang and Zuccon (2021, 2022) devise augmentation methods to generate misspelled queries and propose training methods, Typos-aware Training and Self-Teaching (ST), to encourage consistency between outputs of misspelled queries and their non-misspelled counterparts. Alternatively, Sidiropoulos and Kanoulas (2022) apply contrastive loss to enforce representations of misspelled queries to be closer to their corresponding non-misspelled queries. Although these methods can improve the performance of retrieval models for misspelled queries, there is still a substantial performance drop for misspelled queries.", + "bbox": [ + 507, + 286, + 884, + 558 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose a training method to improve dense retrieval for handling misspelled queries based on the following desired properties:", + "bbox": [ + 507, + 562, + 882, + 609 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Alignment: the method should be able to align queries with their corresponding passages.", + "- Robustness: the method should be able to align misspelled queries with their pristine queries.", + "- Contrast: the method should be able to separate queries that refer to different passages and passages that correspond to different queries." + ], + "bbox": [ + 507, + 612, + 882, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In contrast to the existing methods for handling misspelled queries that only satisfy the Alignment and Robustness properties, our method also aims to satisfy the Contrast property. Increasing the distance between dissimilar queries should help distinguish misspelled queries from other distinct queries. We design the following components for our training method: (i) Dual Self-Teaching (DST) incorporates the ideas of Dual Learning (Xia et al., 2017; Li et al., 2021) and Self-Teaching (Zhuang and Zuccon, 2022) to train robust dense retrieval in a bidirectional manner: passage retrieval and", + "bbox": [ + 507, + 726, + 884, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1106", + "bbox": [ + 480, + 927, + 521, + 940 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics", + "bbox": [ + 226, + 945, + 771, + 957 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Volume 2: Short Papers, pages 1106-1115", + "bbox": [ + 368, + 958, + 628, + 971 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "July 9-14, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 295, + 972, + 700, + 985 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "query retrieval. (ii) Query Augmentation generates a numerous number of misspelling variations for each query to supply our training objective.", + "bbox": [ + 112, + 84, + 487, + 133 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experimental studies were conducted to assess the efficiency of the proposed method in comparison to existing approaches. We conduct experiments based on two different pre-trained language models. We evaluate using two passage retrieval benchmark datasets, a standard one and a specialized one for misspellings robustness evaluation. For each dataset, we measure performance on both misspelled and non-misspelled queries, where the misspelled queries are both generated and real-world queries. The experimental results show that the proposed method outperforms the best existing methods for enhancing the robustness of dense retrieval against misspellings without sacrificing performance for non-misspelled queries.", + "bbox": [ + 112, + 134, + 489, + 375 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize our contributions as follows:", + "bbox": [ + 131, + 376, + 460, + 391 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel training method to enhance the robustness of dense retrieval against misspellings by incorporating three desired properties: Alignment, Robustness, and Contrast.", + "- We introduce Dual Self-Teaching (DST) which adopts the idea of Dual Learning and Self-Teaching to learn robust representations. In addition, we propose Query Augmentation to generate multiple views of a particular query under different misspelling scenarios.", + "- We evaluate our method on misspelled and non-misspelled queries from two passage retrieval datasets. The results show that our method outperforms the previous state-of-the-art methods by a significant margin on misspelled queries." + ], + "bbox": [ + 114, + 393, + 489, + 634 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Methodology", + "text_level": 1, + "bbox": [ + 112, + 649, + 263, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose a training pipeline to enhance the dense retrieval capability for handling spelling variations and mistakes in queries. As shown in Figure 1, the training pipeline comprises three steps. (i) Query Augmentation: we augment each query in the training set into multiple misspelled queries using the typo generators provided by Zhuang and Zuccon (2021). (ii) Similarity Score Calculation: we compute similarity score distributions between queries and passages for passage retrieval and query retrieval tasks using in-batch negative queries and passages, with additional hard negative passages. (iii) Dual Self-Teaching Loss Calculation: we compute the DST loss using the similarity score distributions to achieve all three desired properties.", + "bbox": [ + 112, + 677, + 489, + 919 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 Query Augmentation", + "text_level": 1, + "bbox": [ + 507, + 84, + 721, + 99 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The purpose of this step is to guide the learning with a broad array of possible misspelling patterns. Let $\\mathbf{Q}$ denote a set $\\{q_{1}, q_{2}, \\ldots, q_{N}\\}$ of $N$ queries. From all queries in $\\mathbf{Q}$ , we generate a set of $K \\times N$ misspelled queries $\\mathcal{Q}' = \\{\\langle q_{1,k}', q_{2,k}', \\ldots, q_{N,k}'\\rangle\\}_{k=1}^{K}$ , where $K$ is the misspelling variations. We use five typo generators proposed by Zhuang and Zuccon (2021), including: RandInsert, RandDelete, RandSub, SwapNeighbor, and SwapAdjacent. Please refer to Appendix A.2 for examples of the misspelled queries.", + "bbox": [ + 507, + 105, + 884, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2 Similarity Score Calculation", + "text_level": 1, + "bbox": [ + 507, + 292, + 779, + 307 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Let $S(a, \\mathbf{B})$ denote a function that computes a similarity score distribution of any vector $a$ over any set of vectors $\\mathbf{B}$ :", + "bbox": [ + 507, + 313, + 884, + 360 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nS (a, \\mathbf {B}) = \\left\\{b _ {i} \\in \\mathbf {B} \\left| \\frac {\\exp (a \\cdot b _ {i})}{\\sum_ {b _ {j} \\in \\mathbf {B}} \\exp (a \\cdot b _ {j})} \\right. \\right\\} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 367, + 882, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given $\\mathbf{P} = \\{p_1, p_2, \\dots, p_M\\}$ to be a set of $M$ passages and $\\mathbf{Q}_k' = \\{q_{1,k}', q_{2,k}', \\dots, q_{N,k}'\\}$ to be the $k^{th}$ set of misspelled queries in $\\mathcal{Q}'$ , we compute two groups of score distributions as follows:", + "bbox": [ + 507, + 416, + 882, + 480 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "Passage retrieval: we calculate score distributions in a query-to-passages direction for each original query $s_p = S(q_n, \\mathbf{P})$ and misspelled query $s_p'^k = S(q_{n,k}', \\mathbf{P})$ .", + "- Query retrieval: we calculate score distributions in a passage-to-queries direction for original queries $s_q = S(p_m, \\mathbf{Q})$ and each set of misspellled queries $s_q^{\\prime k} = S(p_m, \\mathbf{Q}_k^{\\prime})$ ." + ], + "bbox": [ + 507, + 481, + 882, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This way, we produce four different score distributions $(s_p,s_p^{\\prime k},s_q,s_q^{\\prime k})$ for our training objective.", + "bbox": [ + 507, + 609, + 882, + 644 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3 Dual Self-Teaching Loss Calculation", + "text_level": 1, + "bbox": [ + 507, + 653, + 840, + 668 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We design the Dual Self-Teaching loss $(\\mathcal{L}_{\\mathrm{DST}})$ to capture the three desired properties: Alignment, Robustness, and Contrast.", + "bbox": [ + 507, + 673, + 882, + 720 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {D S T}} = \\underbrace {(1 - \\beta) \\mathcal {L} _ {\\mathrm {D C E}}} _ {\\text {D u a l C r o s s - E n t r o p y}} + \\underbrace {\\beta \\mathcal {L} _ {\\mathrm {D K L}}} _ {\\text {D u a l K L - D i v e r g e n c e}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 732, + 882, + 769 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dual Cross-Entropy loss $(\\mathcal{L}_{\\mathrm{DCE}})$ satisfies the Alignment and Contrast properties by utilizing cross-entropy losses to learn score distributions of the original queries for passage retrieval $(s_p)$ and query retrieval $(s_q)$ given labels $y_{p}$ and $y_{q}$ .", + "bbox": [ + 507, + 776, + 882, + 858 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {D C E}} = \\underbrace {(1 - \\gamma) \\mathcal {L} _ {\\mathrm {C E}} ^ {(P)} \\left(s _ {p} , y _ {p}\\right)} _ {\\text {P a s s a g e R e t i v e a l}} + \\underbrace {\\gamma \\mathcal {L} _ {\\mathrm {C E}} ^ {(Q)} \\left(s _ {q} , y _ {q}\\right)} _ {\\text {Q u e r y R e t i v e a l}} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 863, + 882, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1107", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9cb707006d2ab480650319b69a87a6bda5748b990dc59544a681e1c57616d9ac.jpg", + "image_caption": [ + "Figure 1: The proposed training pipeline consists of three steps: (a) Query Augmentation, (b) Similarity Score Calculation, and (c) Dual Self-Teaching Loss Calculation." + ], + "image_footnote": [], + "bbox": [ + 211, + 80, + 786, + 253 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Minimizing the $\\mathcal{L}_{\\mathrm{CE}}^{(P)}$ term will increase the similarity scores between queries and their relevant passages to be higher than other irrelevant passages by separating the relevant and irrelevant passages from one another. Minimizing the $\\mathcal{L}_{\\mathrm{CE}}^{(Q)}$ term will increase the similarity scores between passages and their relevant queries to be higher than other irrelevant queries by separating the relevant and irrelevant queries from one another. In this manner, minimizing one of the two terms will align queries with their corresponding passages, satisfying the Alignment property. Moreover, minimizing both terms will separate queries that refer to different passages and passages that belong to different queries, satisfying the Contrast property.", + "bbox": [ + 112, + 304, + 489, + 549 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dual KL-Divergence loss $(\\mathcal{L}_{\\mathrm{DKL}})$ aims to fulfill the Robustness property by using KL losses to match score distributions of misspelled queries $\\{s_p^{\\prime 1}, s_p^{\\prime 2}, \\ldots, s_p^{\\prime K}\\}$ and $\\{s_q^{\\prime 1}, s_q^{\\prime 2}, \\ldots, s_q^{\\prime K}\\}$ to the score distributions of the original query $s_p$ and $s_q$ .", + "bbox": [ + 112, + 551, + 489, + 633 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {D K L}} = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\underbrace {(1 - \\sigma) \\mathcal {L} _ {\\mathrm {K L}} ^ {(P)} \\left(s _ {p} ^ {\\prime k} , s _ {p}\\right)} _ {\\text {P a s s a g e R e t r i e v a l C o n s i s t e n c y}} \\tag {4} \\\\ + \\underbrace {\\sigma \\mathcal {L} _ {\\mathrm {K L}} ^ {(Q)} \\left(s _ {q} ^ {\\prime k} , s _ {q}\\right)} _ {\\text {Q u e r y R e t r i e v a l C o n s i s t e n c y}} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 644, + 489, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Minimizing $\\mathcal{L}_{\\mathrm{KL}}^{(P)}$ and $\\mathcal{L}_{\\mathrm{KL}}^{(Q)}$ will reduce the discrepancy between misspelled and non-misspelled queries for both query-to-passages and passage-to-queries score distributions. This way, we implicitly align representations of the misspelled queries to the original queries, satisfying the Robustness property. To stabilize training, we apply stop-gradient to the score distributions of the original queries $(s_p$ and $s_q)$ in the $\\mathcal{L}_{\\mathrm{DKL}}$ . The $\\beta$ , $\\gamma$ , and $\\sigma$ are the balancing coefficients selected by hyper-parameter tuning", + "bbox": [ + 112, + 753, + 490, + 919 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "on a development set. With this loss combination, we achieve all three desired properties.", + "bbox": [ + 507, + 300, + 884, + 332 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 Experimental Settings", + "text_level": 1, + "bbox": [ + 507, + 344, + 737, + 362 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1 Training Details", + "text_level": 1, + "bbox": [ + 507, + 372, + 684, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We experiment on two pre-trained language models, BERT (Devlin et al., 2019) and Character-BERT (El Boukkouri et al., 2020). We train models only on the training set of MS MARCO dataset (Nguyen et al., 2016). Moreover, the training data provided by the Tevatron toolkit (Gao et al., 2022) also contains hard negative passages. We include the training set details and hyper-parameter settings in Appendix A.1.", + "bbox": [ + 507, + 394, + 884, + 539 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2 Competitive Methods", + "text_level": 1, + "bbox": [ + 507, + 552, + 724, + 568 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To show the effectiveness of our method, we compare our work with the following baseline and competitive training methods.", + "bbox": [ + 507, + 574, + 882, + 621 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- DPR (Karpukhin et al., 2020) is a baseline training method that trains dense retrieval merely on non-misspelled queries using $\\mathcal{L}_{\\mathrm{CE}}^{(P)}$ loss.", + "- $DPR + Aug$ (Zhuang and Zuccon, 2021) is the Typos-aware Training method which trains dense retrieval on both misspelled and non-misspelled queries using $\\mathcal{L}_{\\mathrm{CE}}^{(P)}$ loss.", + "- $DPR + Aug + CL$ (Sidiropoulos and Kanoulas, 2022) employs additional contrastive loss to train the misspelled queries.", + "- $DPR + ST$ (Zhuang and Zuccon, 2022) is the Self-Teaching method that trains dense retrieval on both misspelled and non-misspelled queries using $\\mathcal{L}_{\\mathrm{CE}}^{(P)}$ and $\\mathcal{L}_{\\mathrm{KL}}^{(P)}$ losses." + ], + "bbox": [ + 507, + 623, + 882, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note that their query augmentation method is identical to the Query Augmentation with $K = 1$ . We retrain all models using the same setting described in the previous section.", + "bbox": [ + 507, + 854, + 882, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1108", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/d402ae89676c46cc2d674a2707b32f22a0776882a28eee8ea829c2885179e05f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsBERT-basedCharacterBERT-based
MS MARCODL-typoMS MARCODL-typo
MRR@10R@1000nDCG@10MRRMAPMRR@10R@1000nDCG@10MRRMAP
DPR.143 (.331).696 (.954).276 (.682).431 (.873).175 (.563).162 (.321).726 (.945).268 (.643).376 (.832).212 (.503)
+ Aug.227 (.334).857 (.950).398 (.682).530 (.806).286 (.565).258 (.326).883 (.946).414 (.631).578 (.783).318 (.512)
+ Aug + CL.234 (.335).867 (.951).387 (.668).536 (.864).267 (.544).263 (.330).894 (.947).466 (.677).635 (.819).360 (.544)
+ ST.237 (.333).874 (.950).392 (.677).525 (.852).283 (.557).274 (.332).900 (.947).469 (.650).619 (.810).359 (.517)
+ DST (our).260† (.336).894† (.954).432 (.673).558 (.833).343† (.568).288† (.332).918† (.949).529† (.673).742† (.854).403 (.537)
", + "bbox": [ + 121, + 80, + 884, + 200 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3 Dataset and Evaluation", + "text_level": 1, + "bbox": [ + 112, + 281, + 344, + 294 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Datasets. We evaluate the effectiveness of DST on two passage retrieval datasets, MS MARCO and DL-typo (Zhuang and Zuccon, 2022), each with misspelled and non-misspelled queries. There are 8.8 million candidate passages for both datasets. The development set of MS MARCO contains 6,980 non-misspelled queries. To obtain misspelled queries, we use the typos generator method proposed by Zhuang and Zuccon (2021) to generate 10 misspelled variations for each original query. The DL-typo provides 60 real-world misspelled queries and 60 corresponding non-misspelled queries that are corrected manually.", + "bbox": [ + 112, + 307, + 489, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Evaluation. We use the standard metrics originally used by each dataset's creators. For MS MARCO, each misspelled query performance is the average of 10 measurements. We employ Ranx evaluation library (Bassani, 2022) to measure performance and statistical significance. Specifically, we use a two-tailed paired t-test with Bonferroni correction to measure the statistical significance $(p < 0.05)$ .", + "bbox": [ + 112, + 519, + 489, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4 Experimental Results", + "text_level": 1, + "bbox": [ + 112, + 668, + 334, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1 Main Results", + "text_level": 1, + "bbox": [ + 112, + 700, + 265, + 714 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Table 1, the results indicate that DST outperforms competitive methods for misspelled queries in every case without sacrificing performance for non-misspelled queries in eight out of ten cases. We observe some performance trade-offs for the BERT-based model in the DL-typo dataset's non-misspelling scores (nDCG@10 and MRR). Aside from that, there is no performance trade-off for the CharacterBERT-based model. These outcomes conform with the observation in Figure 2 (Section 4.4) that DST improves the Robustness and Contrast of misspelled queries.", + "bbox": [ + 112, + 726, + 490, + 919 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2 Query Augmentation Size Study", + "text_level": 1, + "bbox": [ + 507, + 281, + 808, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To study the benefit of query augmentation and find the optimal augmentation size, we measure the performance of BERT-based dense retrieval models trained with DST using the query augmentation size $K$ of 1, 10, 20, 40, and 60. Note that the query augmentation method used in previous works is a special case of Query Augmentation when $K = 1$ . We report the results using MRR@10 for the development set of the MS MARCO dataset. We also report training time to show trade-offs between performance and computation.", + "bbox": [ + 505, + 302, + 884, + 479 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/65c27de57f78c38baea3116a028eb20cd2ef9fe80691c5d28c1a2f515b49baae.jpg", + "table_caption": [ + "Table 1: Results of different training methods on misspelled and non-misspelled queries. We report the results in the format of \"misspelled query performance (non-misspelled query performance)\". We emphasize the best score with bold text and the second-best score with underlined text. We use $\\dagger$ to denote DST results that significantly outperform the second-best result ( $p < 0.05$ )." + ], + "table_footnote": [], + "table_body": "
QueriesK
110204060
Original.334.334.335.336.332
Misspelled.251.258.260.260.260
Training time (hr)1820233139
", + "bbox": [ + 527, + 489, + 867, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 2: Results of query augmentation size study. We train all models in this experiment on a V100 32G GPU.", + "bbox": [ + 507, + 580, + 882, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Table 2, the results indicate that increasing $K$ improves the performance of both misspelled and non-misspelled queries, but only up to a certain point, after which the performance begins to decline. We observe that setting $K = 40$ produces the best results, and there is no further performance improvement after this point.", + "bbox": [ + 507, + 624, + 882, + 737 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3 Loss Ablation Study", + "text_level": 1, + "bbox": [ + 507, + 747, + 715, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this experiment, we study the benefit of each term in DST by training BERT-based dense retrieval models on variant loss combinations with $K = 40$ . The results in Table 3 reveal that $\\mathcal{L}_{\\mathrm{KL}}^{(P)}$ and $\\mathcal{L}_{\\mathrm{KL}}^{(Q)}$ terms positively contribute to the performance of misspelled and non-misspelled queries, with the $\\mathcal{L}_{\\mathrm{KL}}^{(P)}$ being more significant. The $\\mathcal{L}_{\\mathrm{CE}}^{(P)}$ term is crucial for retrieval performance, whereas the $\\mathcal{L}_{\\mathrm{CE}}^{(Q)}$ term indirectly improves the performance", + "bbox": [ + 507, + 768, + 884, + 920 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1109", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/a151b538d7850d6f9ac41fcb45f25b6288f560992290463285cfd1dd193618f6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
L(PCE)L(QCE)L(PKL)L(QKL)MRR@10
✓✓✓✓.260 (.336)
✓✓✓.257 (.335)
✓✓✓.228 (.326)
✓✓✓.251 (.337)
✓✓✓.087 (.114)
✓✓.249 (.336)
✓✓.120 (.158)
", + "bbox": [ + 159, + 80, + 421, + 200 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 3: Loss ablation study results on MS MARCO.", + "bbox": [ + 115, + 209, + 482, + 223 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of misspelled queries by separating their pristine queries from the surrounding queries. Disabling query retrieval terms $(\\mathcal{L}_{\\mathrm{CE}}^{(Q)})$ and $\\mathcal{L}_{\\mathrm{KL}}^{(Q)}$ greatly reduces performances for misspelled queries. The passage retrieval terms $(\\mathcal{L}_{\\mathrm{CE}}^{(P)})$ and $\\mathcal{L}_{\\mathrm{KL}}^{(P)}$ are indispensable and cannot be substituted.", + "bbox": [ + 112, + 242, + 487, + 342 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4 Query Distributions", + "text_level": 1, + "bbox": [ + 112, + 360, + 317, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The purpose of this section is to study the impact of our training method on the Robustness and Contrast of misspelled queries. We also compare our method against the baseline and competitive methods to show its effectiveness. The Robustness and Contrast of misspelled queries are illustrated using the following kernel density graphs:", + "bbox": [ + 112, + 384, + 487, + 495 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Original-to-Misspell: the cosine similarity distribution between original and misspelled queries.", + "- Original-to-Neighbor: the cosine similarity distribution between original and neighbor queries. The Robustness property is emphasized by the Original-to-Misspell distribution having high cosine similarity. On the other hand, the Contrast property is emphasized by the small overlapping between Original-to-Misspell and Original-to-Neighbor distributions. The results in Figure 2 show that our method (c) produces the best Robustness and Contrast properties for misspelled queries in comparison to other methods." + ], + "bbox": [ + 112, + 499, + 489, + 709 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 112, + 728, + 243, + 741 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This paper aims to address the misspelling problem in dense retrieval. We formulate three desired properties for making dense retrieval robust to misspellings: Alignment, Robustness, and Contrast. Unlike previous methods, which only focus on the Alignment and Robustness properties, our method considers all the desired properties. The empirical results show that our method performs best against misspelled queries, revealing the importance of the Contrast property for handling misspellings.", + "bbox": [ + 112, + 758, + 487, + 917 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d5b2d927d187e10c36b3de26eb7003e3a424cda229614d3b30ef44fc48c48a7d.jpg", + "image_caption": [ + "(a) DPR (Karpukhin et al., 2020)." + ], + "image_footnote": [], + "bbox": [ + 529, + 80, + 858, + 259 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4be0a7c423333aa8d7cbdccb802eb94c004878c415d2f6917ecdc767a5e037b9.jpg", + "image_caption": [ + "(b) Self-Teaching (Zhuang and Zuccon, 2022)." + ], + "image_footnote": [], + "bbox": [ + 527, + 287, + 858, + 467 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/497082c56023ac179644de161784f0c8592c628036e5c3c6e31863042aada675.jpg", + "image_caption": [ + "(c) Dual Self-Teaching (our).", + "Figure 2: Kernel density of Original-to-Neighbor (orange) and Original-to-Misspell (blue) of different training methods." + ], + "image_footnote": [], + "bbox": [ + 527, + 495, + 858, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6 Limitations", + "text_level": 1, + "bbox": [ + 509, + 760, + 643, + 774 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We list the limitations of our work as follows:", + "bbox": [ + 507, + 789, + 848, + 803 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The Query Augmentation is designed for the English alphabet; therefore, other languages with different alphabets will require further work.", + "- Since the training strategy relies on fine-tuning a pre-trained language model using a large passage retrieval dataset, it may not be suitable for languages with limited resources" + ], + "bbox": [ + 507, + 806, + 882, + 917 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1110", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 115, + 84, + 213, + 98 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Elias Bassani. 2022. ranx: A blazing-fast python library for ranking evaluation and comparison. In ECIR (2), volume 13186 of Lecture Notes in Computer Science, pages 259-264. Springer.", + "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics.", + "Hicham El Boukkouri, Olivier Ferret, Thomas Lavergne, Hiroshi Noji, Pierre Zweigenbaum, and Jun'ichi Tsujii. 2020. CharacterBERT: Reconciling ELMo and BERT for word-level open-vocabulary representations from characters. In Proceedings of the 28th International Conference on Computational Linguistics, pages 6903-6915, Barcelona, Spain (Online). International Committee on Computational Linguistics.", + "Luyu Gao, Xueguang Ma, Jimmy J. Lin, and Jamie Callan. 2022. Tevatron: An efficient and flexible toolkit for dense retrieval. ArXiv, abs/2203.05765.", + "Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6769-6781, Online. Association for Computational Linguistics.", + "Omar Khattab and Matei Zaharia. 2020. Colbert: Efficient and effective passage search via contextualized late interaction over bert. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '20, page 39-48, New York, NY, USA. Association for Computing Machinery.", + "Yizhi Li, Zhenghao Liu, Chenyan Xiong, and Zhiyuan Liu. 2021. More robust dense retrieval with contrastive dual learning. In Proceedings of the 2021 ACM SIGIR International Conference on Theory of Information Retrieval, ICTIR '21, page 287-296, New York, NY, USA. Association for Computing Machinery.", + "Tri Nguyen, Mir Rosenberg, Xia Song, Jianfeng Gao, Saurabh Tiwary, Rangan Majumder, and Li Deng. 2016. MS MARCO: A human generated machine reading comprehension dataset. In Proceedings of the Workshop on Cognitive Computation: Integrating neural and symbolic approaches 2016 co-located with the 30th Annual Conference on Neural Information Processing Systems (NIPS 2016), Barcelona, Spain, December 9, 2016, volume 1773 of CEUR Workshop Proceedings. CEUR-WS.org." + ], + "bbox": [ + 115, + 107, + 489, + 917 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Gustavo Penha, Arthur Camara, and Claudia Hauff. 2022. Evaluating the robustness of retrieval pipelines with query variation generators. In Advances in Information Retrieval, pages 397-412, Cham. Springer International Publishing.", + "Yingqi Qu, Yuchen Ding, Jing Liu, Kai Liu, Ruiyang Ren, Wayne Xin Zhao, Daxiang Dong, Hua Wu, and Haifeng Wang. 2021. RocketQA: An optimized training approach to dense passage retrieval for open-domain question answering. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5835-5847, Online. Association for Computational Linguistics.", + "Ruiyang Ren, Shangwen Lv, Yingqi Qu, Jing Liu, Wayne Xin Zhao, QiaoQiao She, Hua Wu, Haifeng Wang, and Ji-Rong Wen. 2021a. PAIR: Leveraging passage-centric similarity relation for improving dense passage retrieval. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 2173-2183, Online. Association for Computational Linguistics.", + "Ruiyang Ren, Yingqi Qu, Jing Liu, Wayne Xin Zhao, QiaoQiao She, Hua Wu, Haifeng Wang, and Ji-Rong Wen. 2021b. RocketQAv2: A joint training method for dense passage retrieval and passage re-ranking. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 2825-2835, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.", + "Georgios Sidiropoulos and Evangelos Kanoulas. 2022. Analysing the robustness of dual encoders for dense retrieval against misspellings. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval. ACM.", + "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander M. Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online. Association for Computational Linguistics.", + "Yingce Xia, Tao Qin, Wei Chen, Jiang Bian, Nenghai Yu, and Tie-Yan Liu. 2017. Dual supervised learning. In Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 3789-3798. PMLR.", + "Lee Xiong, Chenyan Xiong, Ye Li, Kwok-Fung Tang, Jialin Liu, Paul N. Bennett, Junaid Ahmed, and" + ], + "bbox": [ + 510, + 85, + 884, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1111", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Arnold Overwijk. 2021. Approximate nearest neighbor negative contrastive learning for dense text retrieval. In International Conference on Learning Representations.", + "Jingtao Zhan, Jiaxin Mao, Yiqun Liu, Min Zhang, and Shaoping Ma. 2020. Repbert: Contextualized text embeddings for first-stage retrieval. CoRR, abs/2006.15498.", + "Shengyao Zhuang and Guido Zuccon. 2021. Dealing with typos for BERT-based passage retrieval and ranking. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 2836-2842, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.", + "Shengyao Zhuang and Guido Zuccon. 2022. Character-bert and self-teaching for improving the robustness of dense retrievers on queries with typos. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '22, page 1444-1454, New York, NY, USA. Association for Computing Machinery." + ], + "bbox": [ + 115, + 85, + 489, + 405 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1112", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A Appendix", + "text_level": 1, + "bbox": [ + 114, + 84, + 236, + 99 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A.1 Training Setup and Hyperparameters", + "text_level": 1, + "bbox": [ + 114, + 117, + 460, + 133 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The MS MARCO is a large-scale English language dataset for machine reading comprehension (MRC). The dataset consists of anonymized queries sampled from Bing's search query logs, each with human generated answers. The training set we used contains 400,782 training samples, each consisting of a query, positive passage, and a set of hard negative passages, which we randomly select 7 hard negative passages for each training sample. We set a batch size to 16 and use in-batch negative sampling for each training sample. Therefore, we obtain $7 + 8 * 15 = 127$ negative passages for each training sample. We use the AdamW optimizer and learning rate of 1e-5 for 150,000 steps with a linear learning rate warm-up over the first 10,000 steps and a linear learning rate decay over the rest of the training steps. For our training method, we set the hyper-parameters $\\beta = 0.5$ , $\\gamma = 0.5$ , $\\sigma = 0.2$ , and the query augmentation size $K = 40$ . Using one V100 32G GPU, the BERT-based model training time is around 31 hours, while the CharacterBERT-based model training time is roughly 56 hours.", + "bbox": [ + 112, + 143, + 489, + 499 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A.2 Query Augmentation Examples", + "text_level": 1, + "bbox": [ + 114, + 520, + 410, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4 provides examples of misspelled queries generated by the Query Augmentation for each original query.", + "bbox": [ + 112, + 546, + 487, + 596 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Original query:", + "text_level": 1, + "bbox": [ + 144, + 617, + 240, + 629 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "what is the goddess of agriculture in greek mythology", + "bbox": [ + 144, + 632, + 438, + 645 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Misspelled queries:", + "text_level": 1, + "bbox": [ + 144, + 648, + 258, + 659 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "what is the goddoess of agriculture in greek mythology", + "bbox": [ + 144, + 663, + 455, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "what is the goddess of agriculture in greek mythology", + "bbox": [ + 144, + 678, + 442, + 692 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "what is the goddess of agriculture in greek mythologo", + "bbox": [ + 144, + 695, + 442, + 708 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "what is the goddesses of agriculture in greek mythology", + "bbox": [ + 144, + 712, + 448, + 725 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "what is the goddess of agriculture in greek mythology", + "bbox": [ + 144, + 728, + 448, + 741 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "what is the goddess of agriculture in greek mythology", + "bbox": [ + 144, + 744, + 453, + 757 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "what is the goddess of agriculture in greek mythology", + "bbox": [ + 144, + 760, + 440, + 774 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "what is the goddess of agriculture in grvek mythology", + "bbox": [ + 144, + 778, + 448, + 790 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "what is the goddess of agricultrue in greek mythology", + "bbox": [ + 144, + 794, + 448, + 807 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "what is the goddess of agriculture in greek mythology", + "bbox": [ + 144, + 810, + 448, + 822 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4: The outputs of Query Augmentation with $K = {10}$ . We use different colors to indicate different types of typo: RandInsert , RandDelete , RandSub ,", + "bbox": [ + 112, + 835, + 489, + 878 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SwapNeighbor, and SwapAdjacent.", + "bbox": [ + 117, + 879, + 378, + 896 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A.3 Licenses", + "text_level": 1, + "bbox": [ + 509, + 84, + 626, + 98 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Datasets: The MS MARCO dataset is available under the MIT license, and the DL-typo dataset is available under the Apache license 2.0. These licenses allow users to use the datasets under nonrestrictive agreements.", + "bbox": [ + 507, + 105, + 882, + 184 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Softwares: We employ Hugging Face (Wolf et al., 2020) and Tevatron (Gao et al., 2022) libraries to train dense retrieval models. We utilize Ranx library (Bassani, 2022) to evaluate retrieval performance. These libraries are available under the Apache license 2.0 which allows both academic and commercial usages. For this reason, we release our code under the Apache license 2.0 to make our code fully accessible and compatible with the other codes we use.", + "bbox": [ + 507, + 186, + 884, + 344 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1113", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A For every submission:", + "text_level": 1, + "bbox": [ + 114, + 107, + 322, + 122 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "A1. Did you describe the limitations of your work?", + "bbox": [ + 129, + 126, + 532, + 143 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Section 6", + "bbox": [ + 149, + 143, + 223, + 156 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "A2. Did you discuss any potential risks of your work?", + "bbox": [ + 127, + 170, + 552, + 186 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "There is no potential risk associated with increasing the robustness of information retrieval applications to question containing misspellings.", + "bbox": [ + 149, + 186, + 882, + 219 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "A3. Do the abstract and introduction summarize the paper's main claims?", + "bbox": [ + 129, + 228, + 695, + 244 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Section 1", + "bbox": [ + 149, + 244, + 223, + 259 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "A4. Have you used AI writing assistants when working on this paper?", + "bbox": [ + 129, + 271, + 670, + 288 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We use Grammarly to check grammatical errors and QuillBot to polish writing quality. These tools are applied to a certain number of sentences in each section, which are then reviewed by humans.", + "bbox": [ + 149, + 288, + 880, + 319 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "B Did you use or create scientific artifacts?", + "text_level": 1, + "bbox": [ + 114, + 331, + 487, + 347 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Section 3.1 for pre-trained language models, training dataset, and training toolkit. Section 3.2 for competitive methods. Section 3.3 for evaluation datasets and evaluation toolkit.", + "bbox": [ + 112, + 353, + 882, + 385 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "B1. Did you cite the creators of artifacts you used?", + "bbox": [ + 129, + 395, + 529, + 411 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Section 3.1 for pre-trained language models, training dataset, and training toolkit. Section 3.2 for competitive methods. Section 3.3 for evaluation datasets and evaluation toolkit.", + "bbox": [ + 149, + 412, + 880, + 444 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?", + "bbox": [ + 129, + 453, + 779, + 470 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Appendix A.3", + "bbox": [ + 149, + 470, + 253, + 486 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?", + "bbox": [ + 129, + 495, + 880, + 562 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Appendix A.3", + "bbox": [ + 149, + 563, + 253, + 577 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?", + "bbox": [ + 129, + 588, + 880, + 636 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We did not collect any data. The datasets we used are publicly available and widely used in information retrieval literature. The data is already anonymized by the creators of the datasets. Therefore we do not need to anonymize the data.", + "bbox": [ + 149, + 637, + 880, + 684 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?", + "bbox": [ + 129, + 695, + 880, + 728 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Appendix A.1", + "bbox": [ + 149, + 728, + 253, + 743 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be.", + "bbox": [ + 129, + 753, + 880, + 835 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Section 3.3 for the evaluation set Appendix A.1 for the training set", + "bbox": [ + 149, + 835, + 640, + 851 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.", + "bbox": [ + 112, + 854, + 877, + 879 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "ACL 2023 Responsible NLP Checklist", + "bbox": [ + 132, + 84, + 433, + 99 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1114", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "C Did you run computational experiments?", + "text_level": 1, + "bbox": [ + 114, + 83, + 494, + 99 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Section 4", + "bbox": [ + 132, + 105, + 206, + 118 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?", + "bbox": [ + 129, + 130, + 878, + 162 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Appendix A.1", + "bbox": [ + 149, + 164, + 253, + 180 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?", + "bbox": [ + 129, + 189, + 880, + 222 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Appendix A.1", + "bbox": [ + 149, + 223, + 252, + 239 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?", + "bbox": [ + 129, + 248, + 880, + 296 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Section 4.1 for Main Results Section 4.2 for Query Augmentation Size Study Section 4.3 for Loss Ablation Study Section 4.4 for Query Distributions", + "bbox": [ + 147, + 298, + 880, + 330 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?", + "bbox": [ + 129, + 340, + 880, + 387 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Our evaluation is parameter free, therefore there is no parameter settings.", + "bbox": [ + 149, + 388, + 695, + 405 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?", + "text_level": 1, + "bbox": [ + 112, + 416, + 877, + 432 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 132, + 436, + 213, + 451 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?", + "bbox": [ + 127, + 463, + 880, + 494 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 495, + 248, + 511 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?", + "bbox": [ + 127, + 521, + 880, + 569 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 571, + 248, + 586 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?", + "bbox": [ + 127, + 596, + 880, + 644 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 646, + 248, + 661 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?", + "bbox": [ + 127, + 671, + 873, + 687 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 689, + 248, + 703 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?", + "bbox": [ + 127, + 715, + 880, + 746 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 747, + 248, + 763 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1115", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Typo-Robust Representation Learning for Dense Retrieval/8977932e-2af7-46c1-9920-a3e3abaf9a04_model.json b/2023/Typo-Robust Representation Learning for Dense Retrieval/8977932e-2af7-46c1-9920-a3e3abaf9a04_model.json new file mode 100644 index 0000000000000000000000000000000000000000..294d6b12eb6c56d8fcd39a38a26c24d42a49eee8 --- /dev/null +++ b/2023/Typo-Robust Representation Learning for Dense Retrieval/8977932e-2af7-46c1-9920-a3e3abaf9a04_model.json @@ -0,0 +1,2244 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.195, + 0.085, + 0.803, + 0.105 + ], + "angle": 0, + "content": "Typo-Robust Representation Learning for Dense Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.183, + 0.114, + 0.826, + 0.148 + ], + "angle": 0, + "content": "Panuthep Tasawong†, Wuttikorn Ponwitayarat†, Peerat Limkonchotiwat†, Can Udomcharoenchaikit†, Ekapol Chuangsuwanich‡, Sarana Nutanong†" + }, + { + "type": "text", + "bbox": [ + 0.192, + 0.148, + 0.812, + 0.231 + ], + "angle": 0, + "content": "†School of Information Science and Technology, VISTEC, Thailand \n‡Department of Computer Engineering, Chulalongkorn University, Thailand {panuthep.t_s20, wuttikorn.p_s22, peerat.l_s19, canu_pro, snutanon} @vistec.ac.th, ekapolc@cp.eng.chula.ac.th" + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.253, + 0.341, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.278, + 0.464, + 0.562 + ], + "angle": 0, + "content": "Dense retrieval is a basic building block of information retrieval applications. One of the main challenges of dense retrieval in real-world settings is the handling of queries containing misspelled words. A popular approach for handling misspelled queries is minimizing the representations discrepancy between misspelled queries and their pristine ones. Unlike the existing approaches, which only focus on the alignment between misspelled and pristine queries, our method also improves the contrast between each misspelled query and its surrounding queries. To assess the effectiveness of our proposed method, we compare it against the existing competitors using two benchmark datasets and two base encoders. Our method outperforms the competitors in all cases with misspelled queries. Our code and models are available at https://github.com/panuthept/DST-DenseRetrieval." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.573, + 0.262, + 0.588 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.598, + 0.49, + 0.79 + ], + "angle": 0, + "content": "Dense retrieval is a fundamental component in many information retrieval applications, such as open-domain question answering and ad-hoc retrieval. The objective is to score and rank a large collection of candidate passages based on their similarity to a given query. The performance of dense retrieval relies on representation learning. A popular approach is to finetune a pre-trained language model to create an embedding space that puts each query closer to its corresponding passages (Zhan et al., 2020; Khattab and Zaharia, 2020; Xiong et al., 2021; Qu et al., 2021; Ren et al., 2021a,b)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.792, + 0.49, + 0.92 + ], + "angle": 0, + "content": "One of the major challenges of dense retrieval is the handling of misspelled queries which induces representations of the misspelled queries to be closer to irrelevant passages than their corresponding passages. Several studies have demonstrated that misspellings in search queries can substantially degrade retrieval performance (Zhuang and Zuccon, 2021; Penha et al., 2022), specifically" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.254, + 0.885, + 0.285 + ], + "angle": 0, + "content": "when informative terms, such as entity mentions, are misspelled (Sidiropoulos and Kanoulas, 2022)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.287, + 0.885, + 0.56 + ], + "angle": 0, + "content": "To create a retrieval model that is capable of handling misspelled queries, researchers have proposed different training methods to align representations of misspelled queries with their pristine ones. Zhuang and Zuccon (2021, 2022) devise augmentation methods to generate misspelled queries and propose training methods, Typos-aware Training and Self-Teaching (ST), to encourage consistency between outputs of misspelled queries and their non-misspelled counterparts. Alternatively, Sidiropoulos and Kanoulas (2022) apply contrastive loss to enforce representations of misspelled queries to be closer to their corresponding non-misspelled queries. Although these methods can improve the performance of retrieval models for misspelled queries, there is still a substantial performance drop for misspelled queries." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.563, + 0.884, + 0.611 + ], + "angle": 0, + "content": "In this paper, we propose a training method to improve dense retrieval for handling misspelled queries based on the following desired properties:" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.613, + 0.881, + 0.644 + ], + "angle": 0, + "content": "- Alignment: the method should be able to align queries with their corresponding passages." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.645, + 0.881, + 0.676 + ], + "angle": 0, + "content": "- Robustness: the method should be able to align misspelled queries with their pristine queries." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.677, + 0.884, + 0.725 + ], + "angle": 0, + "content": "- Contrast: the method should be able to separate queries that refer to different passages and passages that correspond to different queries." + }, + { + "type": "list", + "bbox": [ + 0.509, + 0.613, + 0.884, + 0.725 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.727, + 0.885, + 0.919 + ], + "angle": 0, + "content": "In contrast to the existing methods for handling misspelled queries that only satisfy the Alignment and Robustness properties, our method also aims to satisfy the Contrast property. Increasing the distance between dissimilar queries should help distinguish misspelled queries from other distinct queries. We design the following components for our training method: (i) Dual Self-Teaching (DST) incorporates the ideas of Dual Learning (Xia et al., 2017; Li et al., 2021) and Self-Teaching (Zhuang and Zuccon, 2022) to train robust dense retrieval in a bidirectional manner: passage retrieval and" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.928, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1106" + }, + { + "type": "footer", + "bbox": [ + 0.227, + 0.946, + 0.772, + 0.958 + ], + "angle": 0, + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + }, + { + "type": "footer", + "bbox": [ + 0.369, + 0.959, + 0.63, + 0.972 + ], + "angle": 0, + "content": "Volume 2: Short Papers, pages 1106-1115" + }, + { + "type": "footer", + "bbox": [ + 0.297, + 0.973, + 0.702, + 0.986 + ], + "angle": 0, + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.489, + 0.134 + ], + "angle": 0, + "content": "query retrieval. (ii) Query Augmentation generates a numerous number of misspelling variations for each query to supply our training objective." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.135, + 0.49, + 0.376 + ], + "angle": 0, + "content": "Experimental studies were conducted to assess the efficiency of the proposed method in comparison to existing approaches. We conduct experiments based on two different pre-trained language models. We evaluate using two passage retrieval benchmark datasets, a standard one and a specialized one for misspellings robustness evaluation. For each dataset, we measure performance on both misspelled and non-misspelled queries, where the misspelled queries are both generated and real-world queries. The experimental results show that the proposed method outperforms the best existing methods for enhancing the robustness of dense retrieval against misspellings without sacrificing performance for non-misspelled queries." + }, + { + "type": "text", + "bbox": [ + 0.132, + 0.377, + 0.462, + 0.392 + ], + "angle": 0, + "content": "We summarize our contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.394, + 0.49, + 0.458 + ], + "angle": 0, + "content": "- We propose a novel training method to enhance the robustness of dense retrieval against misspellings by incorporating three desired properties: Alignment, Robustness, and Contrast." + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.458, + 0.49, + 0.554 + ], + "angle": 0, + "content": "- We introduce Dual Self-Teaching (DST) which adopts the idea of Dual Learning and Self-Teaching to learn robust representations. In addition, we propose Query Augmentation to generate multiple views of a particular query under different misspelling scenarios." + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.555, + 0.49, + 0.636 + ], + "angle": 0, + "content": "- We evaluate our method on misspelled and non-misspelled queries from two passage retrieval datasets. The results show that our method outperforms the previous state-of-the-art methods by a significant margin on misspelled queries." + }, + { + "type": "list", + "bbox": [ + 0.115, + 0.394, + 0.49, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.65, + 0.264, + 0.668 + ], + "angle": 0, + "content": "2 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.678, + 0.49, + 0.92 + ], + "angle": 0, + "content": "We propose a training pipeline to enhance the dense retrieval capability for handling spelling variations and mistakes in queries. As shown in Figure 1, the training pipeline comprises three steps. (i) Query Augmentation: we augment each query in the training set into multiple misspelled queries using the typo generators provided by Zhuang and Zuccon (2021). (ii) Similarity Score Calculation: we compute similarity score distributions between queries and passages for passage retrieval and query retrieval tasks using in-batch negative queries and passages, with additional hard negative passages. (iii) Dual Self-Teaching Loss Calculation: we compute the DST loss using the similarity score distributions to achieve all three desired properties." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.085, + 0.722, + 0.101 + ], + "angle": 0, + "content": "2.1 Query Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.106, + 0.885, + 0.283 + ], + "angle": 0, + "content": "The purpose of this step is to guide the learning with a broad array of possible misspelling patterns. Let \\(\\mathbf{Q}\\) denote a set \\(\\{q_{1}, q_{2}, \\ldots, q_{N}\\}\\) of \\(N\\) queries. From all queries in \\(\\mathbf{Q}\\), we generate a set of \\(K \\times N\\) misspelled queries \\(\\mathcal{Q}' = \\{\\langle q_{1,k}', q_{2,k}', \\ldots, q_{N,k}'\\rangle\\}_{k=1}^{K}\\), where \\(K\\) is the misspelling variations. We use five typo generators proposed by Zhuang and Zuccon (2021), including: RandInsert, RandDelete, RandSub, SwapNeighbor, and SwapAdjacent. Please refer to Appendix A.2 for examples of the misspelled queries." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.293, + 0.78, + 0.308 + ], + "angle": 0, + "content": "2.2 Similarity Score Calculation" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.314, + 0.885, + 0.361 + ], + "angle": 0, + "content": "Let \\( S(a, \\mathbf{B}) \\) denote a function that computes a similarity score distribution of any vector \\( a \\) over any set of vectors \\( \\mathbf{B} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.368, + 0.884, + 0.412 + ], + "angle": 0, + "content": "\\[\nS (a, \\mathbf {B}) = \\left\\{b _ {i} \\in \\mathbf {B} \\left| \\frac {\\exp (a \\cdot b _ {i})}{\\sum_ {b _ {j} \\in \\mathbf {B}} \\exp (a \\cdot b _ {j})} \\right. \\right\\} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.417, + 0.884, + 0.481 + ], + "angle": 0, + "content": "Given \\(\\mathbf{P} = \\{p_1, p_2, \\dots, p_M\\}\\) to be a set of \\(M\\) passages and \\(\\mathbf{Q}_k' = \\{q_{1,k}', q_{2,k}', \\dots, q_{N,k}'\\}\\) to be the \\(k^{th}\\) set of misspelled queries in \\(\\mathcal{Q}'\\), we compute two groups of score distributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.482, + 0.884, + 0.547 + ], + "angle": 0, + "content": "Passage retrieval: we calculate score distributions in a query-to-passages direction for each original query \\( s_p = S(q_n, \\mathbf{P}) \\) and misspelled query \\( s_p'^k = S(q_{n,k}', \\mathbf{P}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.547, + 0.884, + 0.611 + ], + "angle": 0, + "content": "- Query retrieval: we calculate score distributions in a passage-to-queries direction for original queries \\( s_q = S(p_m, \\mathbf{Q}) \\) and each set of misspellled queries \\( s_q^{\\prime k} = S(p_m, \\mathbf{Q}_k^{\\prime}) \\)." + }, + { + "type": "list", + "bbox": [ + 0.509, + 0.482, + 0.884, + 0.611 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.611, + 0.884, + 0.645 + ], + "angle": 0, + "content": "This way, we produce four different score distributions \\((s_p,s_p^{\\prime k},s_q,s_q^{\\prime k})\\) for our training objective." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.654, + 0.842, + 0.669 + ], + "angle": 0, + "content": "2.3 Dual Self-Teaching Loss Calculation" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.674, + 0.884, + 0.721 + ], + "angle": 0, + "content": "We design the Dual Self-Teaching loss \\((\\mathcal{L}_{\\mathrm{DST}})\\) to capture the three desired properties: Alignment, Robustness, and Contrast." + }, + { + "type": "equation", + "bbox": [ + 0.531, + 0.733, + 0.884, + 0.77 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {D S T}} = \\underbrace {(1 - \\beta) \\mathcal {L} _ {\\mathrm {D C E}}} _ {\\text {D u a l C r o s s - E n t r o p y}} + \\underbrace {\\beta \\mathcal {L} _ {\\mathrm {D K L}}} _ {\\text {D u a l K L - D i v e r g e n c e}} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.777, + 0.884, + 0.859 + ], + "angle": 0, + "content": "Dual Cross-Entropy loss \\((\\mathcal{L}_{\\mathrm{DCE}})\\) satisfies the Alignment and Contrast properties by utilizing cross-entropy losses to learn score distributions of the original queries for passage retrieval \\((s_p)\\) and query retrieval \\((s_q)\\) given labels \\(y_{p}\\) and \\(y_{q}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.864, + 0.883, + 0.919 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {D C E}} = \\underbrace {(1 - \\gamma) \\mathcal {L} _ {\\mathrm {C E}} ^ {(P)} \\left(s _ {p} , y _ {p}\\right)} _ {\\text {P a s s a g e R e t i v e a l}} + \\underbrace {\\gamma \\mathcal {L} _ {\\mathrm {C E}} ^ {(Q)} \\left(s _ {q} , y _ {q}\\right)} _ {\\text {Q u e r y R e t i v e a l}} \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1107" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.212, + 0.082, + 0.788, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.263, + 0.884, + 0.293 + ], + "angle": 0, + "content": "Figure 1: The proposed training pipeline consists of three steps: (a) Query Augmentation, (b) Similarity Score Calculation, and (c) Dual Self-Teaching Loss Calculation." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.305, + 0.49, + 0.55 + ], + "angle": 0, + "content": "Minimizing the \\(\\mathcal{L}_{\\mathrm{CE}}^{(P)}\\) term will increase the similarity scores between queries and their relevant passages to be higher than other irrelevant passages by separating the relevant and irrelevant passages from one another. Minimizing the \\(\\mathcal{L}_{\\mathrm{CE}}^{(Q)}\\) term will increase the similarity scores between passages and their relevant queries to be higher than other irrelevant queries by separating the relevant and irrelevant queries from one another. In this manner, minimizing one of the two terms will align queries with their corresponding passages, satisfying the Alignment property. Moreover, minimizing both terms will separate queries that refer to different passages and passages that belong to different queries, satisfying the Contrast property." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.552, + 0.49, + 0.634 + ], + "angle": 0, + "content": "Dual KL-Divergence loss \\((\\mathcal{L}_{\\mathrm{DKL}})\\) aims to fulfill the Robustness property by using KL losses to match score distributions of misspelled queries \\(\\{s_p^{\\prime 1}, s_p^{\\prime 2}, \\ldots, s_p^{\\prime K}\\}\\) and \\(\\{s_q^{\\prime 1}, s_q^{\\prime 2}, \\ldots, s_q^{\\prime K}\\}\\) to the score distributions of the original query \\(s_p\\) and \\(s_q\\)." + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.645, + 0.49, + 0.74 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {D K L}} = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\underbrace {(1 - \\sigma) \\mathcal {L} _ {\\mathrm {K L}} ^ {(P)} \\left(s _ {p} ^ {\\prime k} , s _ {p}\\right)} _ {\\text {P a s s a g e R e t r i e v a l C o n s i s t e n c y}} \\tag {4} \\\\ + \\underbrace {\\sigma \\mathcal {L} _ {\\mathrm {K L}} ^ {(Q)} \\left(s _ {q} ^ {\\prime k} , s _ {q}\\right)} _ {\\text {Q u e r y R e t r i e v a l C o n s i s t e n c y}} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.755, + 0.491, + 0.92 + ], + "angle": 0, + "content": "Minimizing \\(\\mathcal{L}_{\\mathrm{KL}}^{(P)}\\) and \\(\\mathcal{L}_{\\mathrm{KL}}^{(Q)}\\) will reduce the discrepancy between misspelled and non-misspelled queries for both query-to-passages and passage-to-queries score distributions. This way, we implicitly align representations of the misspelled queries to the original queries, satisfying the Robustness property. To stabilize training, we apply stop-gradient to the score distributions of the original queries \\((s_p\\) and \\(s_q)\\) in the \\(\\mathcal{L}_{\\mathrm{DKL}}\\). The \\(\\beta\\), \\(\\gamma\\), and \\(\\sigma\\) are the balancing coefficients selected by hyper-parameter tuning" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.301, + 0.885, + 0.333 + ], + "angle": 0, + "content": "on a development set. With this loss combination, we achieve all three desired properties." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.346, + 0.738, + 0.363 + ], + "angle": 0, + "content": "3 Experimental Settings" + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.373, + 0.685, + 0.389 + ], + "angle": 0, + "content": "3.1 Training Details" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.395, + 0.885, + 0.541 + ], + "angle": 0, + "content": "We experiment on two pre-trained language models, BERT (Devlin et al., 2019) and Character-BERT (El Boukkouri et al., 2020). We train models only on the training set of MS MARCO dataset (Nguyen et al., 2016). Moreover, the training data provided by the Tevatron toolkit (Gao et al., 2022) also contains hard negative passages. We include the training set details and hyper-parameter settings in Appendix A.1." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.553, + 0.725, + 0.569 + ], + "angle": 0, + "content": "3.2 Competitive Methods" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.575, + 0.884, + 0.623 + ], + "angle": 0, + "content": "To show the effectiveness of our method, we compare our work with the following baseline and competitive training methods." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.624, + 0.883, + 0.674 + ], + "angle": 0, + "content": "- DPR (Karpukhin et al., 2020) is a baseline training method that trains dense retrieval merely on non-misspelled queries using \\(\\mathcal{L}_{\\mathrm{CE}}^{(P)}\\) loss." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.674, + 0.882, + 0.74 + ], + "angle": 0, + "content": "- \\(DPR + Aug\\) (Zhuang and Zuccon, 2021) is the Typos-aware Training method which trains dense retrieval on both misspelled and non-misspelled queries using \\(\\mathcal{L}_{\\mathrm{CE}}^{(P)}\\) loss." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.74, + 0.882, + 0.787 + ], + "angle": 0, + "content": "- \\(DPR + Aug + CL\\) (Sidiropoulos and Kanoulas, 2022) employs additional contrastive loss to train the misspelled queries." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.789, + 0.882, + 0.855 + ], + "angle": 0, + "content": "- \\(DPR + ST\\) (Zhuang and Zuccon, 2022) is the Self-Teaching method that trains dense retrieval on both misspelled and non-misspelled queries using \\(\\mathcal{L}_{\\mathrm{CE}}^{(P)}\\) and \\(\\mathcal{L}_{\\mathrm{KL}}^{(P)}\\) losses." + }, + { + "type": "list", + "bbox": [ + 0.509, + 0.624, + 0.883, + 0.855 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.856, + 0.884, + 0.919 + ], + "angle": 0, + "content": "Note that their query augmentation method is identical to the Query Augmentation with \\( K = 1 \\). We retrain all models using the same setting described in the previous section." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1108" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.122, + 0.082, + 0.885, + 0.201 + ], + "angle": 0, + "content": "
MethodsBERT-basedCharacterBERT-based
MS MARCODL-typoMS MARCODL-typo
MRR@10R@1000nDCG@10MRRMAPMRR@10R@1000nDCG@10MRRMAP
DPR.143 (.331).696 (.954).276 (.682).431 (.873).175 (.563).162 (.321).726 (.945).268 (.643).376 (.832).212 (.503)
+ Aug.227 (.334).857 (.950).398 (.682).530 (.806).286 (.565).258 (.326).883 (.946).414 (.631).578 (.783).318 (.512)
+ Aug + CL.234 (.335).867 (.951).387 (.668).536 (.864).267 (.544).263 (.330).894 (.947).466 (.677).635 (.819).360 (.544)
+ ST.237 (.333).874 (.950).392 (.677).525 (.852).283 (.557).274 (.332).900 (.947).469 (.650).619 (.810).359 (.517)
+ DST (our).260† (.336).894† (.954).432 (.673).558 (.833).343† (.568).288† (.332).918† (.949).529† (.673).742† (.854).403 (.537)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.209, + 0.885, + 0.268 + ], + "angle": 0, + "content": "Table 1: Results of different training methods on misspelled and non-misspelled queries. We report the results in the format of \"misspelled query performance (non-misspelled query performance)\". We emphasize the best score with bold text and the second-best score with underlined text. We use \\(\\dagger\\) to denote DST results that significantly outperform the second-best result (\\(p < 0.05\\))." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.282, + 0.345, + 0.296 + ], + "angle": 0, + "content": "3.3 Dataset and Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.308, + 0.49, + 0.518 + ], + "angle": 0, + "content": "Datasets. We evaluate the effectiveness of DST on two passage retrieval datasets, MS MARCO and DL-typo (Zhuang and Zuccon, 2022), each with misspelled and non-misspelled queries. There are 8.8 million candidate passages for both datasets. The development set of MS MARCO contains 6,980 non-misspelled queries. To obtain misspelled queries, we use the typos generator method proposed by Zhuang and Zuccon (2021) to generate 10 misspelled variations for each original query. The DL-typo provides 60 real-world misspelled queries and 60 corresponding non-misspelled queries that are corrected manually." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.52, + 0.49, + 0.651 + ], + "angle": 0, + "content": "Evaluation. We use the standard metrics originally used by each dataset's creators. For MS MARCO, each misspelled query performance is the average of 10 measurements. We employ Ranx evaluation library (Bassani, 2022) to measure performance and statistical significance. Specifically, we use a two-tailed paired t-test with Bonferroni correction to measure the statistical significance \\((p < 0.05)\\)." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.669, + 0.336, + 0.687 + ], + "angle": 0, + "content": "4 Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.701, + 0.266, + 0.715 + ], + "angle": 0, + "content": "4.1 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.727, + 0.491, + 0.92 + ], + "angle": 0, + "content": "As shown in Table 1, the results indicate that DST outperforms competitive methods for misspelled queries in every case without sacrificing performance for non-misspelled queries in eight out of ten cases. We observe some performance trade-offs for the BERT-based model in the DL-typo dataset's non-misspelling scores (nDCG@10 and MRR). Aside from that, there is no performance trade-off for the CharacterBERT-based model. These outcomes conform with the observation in Figure 2 (Section 4.4) that DST improves the Robustness and Contrast of misspelled queries." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.282, + 0.809, + 0.298 + ], + "angle": 0, + "content": "4.2 Query Augmentation Size Study" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.303, + 0.885, + 0.48 + ], + "angle": 0, + "content": "To study the benefit of query augmentation and find the optimal augmentation size, we measure the performance of BERT-based dense retrieval models trained with DST using the query augmentation size \\( K \\) of 1, 10, 20, 40, and 60. Note that the query augmentation method used in previous works is a special case of Query Augmentation when \\( K = 1 \\). We report the results using MRR@10 for the development set of the MS MARCO dataset. We also report training time to show trade-offs between performance and computation." + }, + { + "type": "table", + "bbox": [ + 0.528, + 0.49, + 0.868, + 0.572 + ], + "angle": 0, + "content": "
QueriesK
110204060
Original.334.334.335.336.332
Misspelled.251.258.260.260.260
Training time (hr)1820233139
" + }, + { + "type": "table_caption", + "bbox": [ + 0.508, + 0.581, + 0.883, + 0.611 + ], + "angle": 0, + "content": "Table 2: Results of query augmentation size study. We train all models in this experiment on a V100 32G GPU." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.625, + 0.884, + 0.738 + ], + "angle": 0, + "content": "As shown in Table 2, the results indicate that increasing \\( K \\) improves the performance of both misspelled and non-misspelled queries, but only up to a certain point, after which the performance begins to decline. We observe that setting \\( K = 40 \\) produces the best results, and there is no further performance improvement after this point." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.748, + 0.717, + 0.764 + ], + "angle": 0, + "content": "4.3 Loss Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.769, + 0.885, + 0.921 + ], + "angle": 0, + "content": "In this experiment, we study the benefit of each term in DST by training BERT-based dense retrieval models on variant loss combinations with \\( K = 40 \\). The results in Table 3 reveal that \\( \\mathcal{L}_{\\mathrm{KL}}^{(P)} \\) and \\( \\mathcal{L}_{\\mathrm{KL}}^{(Q)} \\) terms positively contribute to the performance of misspelled and non-misspelled queries, with the \\( \\mathcal{L}_{\\mathrm{KL}}^{(P)} \\) being more significant. The \\( \\mathcal{L}_{\\mathrm{CE}}^{(P)} \\) term is crucial for retrieval performance, whereas the \\( \\mathcal{L}_{\\mathrm{CE}}^{(Q)} \\) term indirectly improves the performance" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1109" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.161, + 0.081, + 0.422, + 0.201 + ], + "angle": 0, + "content": "
L(PCE)L(QCE)L(PKL)L(QKL)MRR@10
✓✓✓✓.260 (.336)
✓✓✓.257 (.335)
✓✓✓.228 (.326)
✓✓✓.251 (.337)
✓✓✓.087 (.114)
✓✓.249 (.336)
✓✓.120 (.158)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.117, + 0.21, + 0.483, + 0.224 + ], + "angle": 0, + "content": "Table 3: Loss ablation study results on MS MARCO." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.243, + 0.489, + 0.343 + ], + "angle": 0, + "content": "of misspelled queries by separating their pristine queries from the surrounding queries. Disabling query retrieval terms \\((\\mathcal{L}_{\\mathrm{CE}}^{(Q)})\\) and \\(\\mathcal{L}_{\\mathrm{KL}}^{(Q)}\\) greatly reduces performances for misspelled queries. The passage retrieval terms \\((\\mathcal{L}_{\\mathrm{CE}}^{(P)})\\) and \\(\\mathcal{L}_{\\mathrm{KL}}^{(P)}\\) are indispensable and cannot be substituted." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.361, + 0.319, + 0.376 + ], + "angle": 0, + "content": "4.4 Query Distributions" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.385, + 0.489, + 0.497 + ], + "angle": 0, + "content": "The purpose of this section is to study the impact of our training method on the Robustness and Contrast of misspelled queries. We also compare our method against the baseline and competitive methods to show its effectiveness. The Robustness and Contrast of misspelled queries are illustrated using the following kernel density graphs:" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.5, + 0.489, + 0.531 + ], + "angle": 0, + "content": "- Original-to-Misspell: the cosine similarity distribution between original and misspelled queries." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.532, + 0.49, + 0.71 + ], + "angle": 0, + "content": "- Original-to-Neighbor: the cosine similarity distribution between original and neighbor queries. The Robustness property is emphasized by the Original-to-Misspell distribution having high cosine similarity. On the other hand, the Contrast property is emphasized by the small overlapping between Original-to-Misspell and Original-to-Neighbor distributions. The results in Figure 2 show that our method (c) produces the best Robustness and Contrast properties for misspelled queries in comparison to other methods." + }, + { + "type": "list", + "bbox": [ + 0.114, + 0.5, + 0.49, + 0.71 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.729, + 0.245, + 0.743 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.759, + 0.489, + 0.919 + ], + "angle": 0, + "content": "This paper aims to address the misspelling problem in dense retrieval. We formulate three desired properties for making dense retrieval robust to misspellings: Alignment, Robustness, and Contrast. Unlike previous methods, which only focus on the Alignment and Robustness properties, our method considers all the desired properties. The empirical results show that our method performs best against misspelled queries, revealing the importance of the Contrast property for handling misspellings." + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.081, + 0.86, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.601, + 0.267, + 0.807, + 0.28 + ], + "angle": 0, + "content": "(a) DPR (Karpukhin et al., 2020)." + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.288, + 0.86, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.561, + 0.475, + 0.846, + 0.488 + ], + "angle": 0, + "content": "(b) Self-Teaching (Zhuang and Zuccon, 2022)." + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.497, + 0.86, + 0.675 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.614, + 0.683, + 0.792, + 0.695 + ], + "angle": 0, + "content": "(c) Dual Self-Teaching (our)." + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.707, + 0.884, + 0.75 + ], + "angle": 0, + "content": "Figure 2: Kernel density of Original-to-Neighbor (orange) and Original-to-Misspell (blue) of different training methods." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.761, + 0.645, + 0.775 + ], + "angle": 0, + "content": "6 Limitations" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.79, + 0.85, + 0.804 + ], + "angle": 0, + "content": "We list the limitations of our work as follows:" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.807, + 0.884, + 0.855 + ], + "angle": 0, + "content": "- The Query Augmentation is designed for the English alphabet; therefore, other languages with different alphabets will require further work." + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.856, + 0.884, + 0.919 + ], + "angle": 0, + "content": "- Since the training strategy relies on fine-tuning a pre-trained language model using a large passage retrieval dataset, it may not be suitable for languages with limited resources" + }, + { + "type": "list", + "bbox": [ + 0.509, + 0.807, + 0.884, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1110" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.116, + 0.085, + 0.214, + 0.099 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.108, + 0.49, + 0.162 + ], + "angle": 0, + "content": "Elias Bassani. 2022. ranx: A blazing-fast python library for ranking evaluation and comparison. In ECIR (2), volume 13186 of Lecture Notes in Computer Science, pages 259-264. Springer." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.172, + 0.49, + 0.29 + ], + "angle": 0, + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.3, + 0.49, + 0.417 + ], + "angle": 0, + "content": "Hicham El Boukkouri, Olivier Ferret, Thomas Lavergne, Hiroshi Noji, Pierre Zweigenbaum, and Jun'ichi Tsujii. 2020. CharacterBERT: Reconciling ELMo and BERT for word-level open-vocabulary representations from characters. In Proceedings of the 28th International Conference on Computational Linguistics, pages 6903-6915, Barcelona, Spain (Online). International Committee on Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.429, + 0.488, + 0.469 + ], + "angle": 0, + "content": "Luyu Gao, Xueguang Ma, Jimmy J. Lin, and Jamie Callan. 2022. Tevatron: An efficient and flexible toolkit for dense retrieval. ArXiv, abs/2203.05765." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.479, + 0.49, + 0.572 + ], + "angle": 0, + "content": "Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6769-6781, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.582, + 0.49, + 0.674 + ], + "angle": 0, + "content": "Omar Khattab and Matei Zaharia. 2020. Colbert: Efficient and effective passage search via contextualized late interaction over bert. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '20, page 39-48, New York, NY, USA. Association for Computing Machinery." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.685, + 0.49, + 0.777 + ], + "angle": 0, + "content": "Yizhi Li, Zhenghao Liu, Chenyan Xiong, and Zhiyuan Liu. 2021. More robust dense retrieval with contrastive dual learning. In Proceedings of the 2021 ACM SIGIR International Conference on Theory of Information Retrieval, ICTIR '21, page 287-296, New York, NY, USA. Association for Computing Machinery." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.787, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Tri Nguyen, Mir Rosenberg, Xia Song, Jianfeng Gao, Saurabh Tiwary, Rangan Majumder, and Li Deng. 2016. MS MARCO: A human generated machine reading comprehension dataset. In Proceedings of the Workshop on Cognitive Computation: Integrating neural and symbolic approaches 2016 co-located with the 30th Annual Conference on Neural Information Processing Systems (NIPS 2016), Barcelona, Spain, December 9, 2016, volume 1773 of CEUR Workshop Proceedings. CEUR-WS.org." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.108, + 0.49, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.086, + 0.885, + 0.152 + ], + "angle": 0, + "content": "Gustavo Penha, Arthur Camara, and Claudia Hauff. 2022. Evaluating the robustness of retrieval pipelines with query variation generators. In Advances in Information Retrieval, pages 397-412, Cham. Springer International Publishing." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.166, + 0.885, + 0.284 + ], + "angle": 0, + "content": "Yingqi Qu, Yuchen Ding, Jing Liu, Kai Liu, Ruiyang Ren, Wayne Xin Zhao, Daxiang Dong, Hua Wu, and Haifeng Wang. 2021. RocketQA: An optimized training approach to dense passage retrieval for open-domain question answering. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5835-5847, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.298, + 0.885, + 0.403 + ], + "angle": 0, + "content": "Ruiyang Ren, Shangwen Lv, Yingqi Qu, Jing Liu, Wayne Xin Zhao, QiaoQiao She, Hua Wu, Haifeng Wang, and Ji-Rong Wen. 2021a. PAIR: Leveraging passage-centric similarity relation for improving dense passage retrieval. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 2173-2183, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.416, + 0.885, + 0.522 + ], + "angle": 0, + "content": "Ruiyang Ren, Yingqi Qu, Jing Liu, Wayne Xin Zhao, QiaoQiao She, Hua Wu, Haifeng Wang, and Ji-Rong Wen. 2021b. RocketQAv2: A joint training method for dense passage retrieval and passage re-ranking. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 2825-2835, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.535, + 0.885, + 0.613 + ], + "angle": 0, + "content": "Georgios Sidiropoulos and Evangelos Kanoulas. 2022. Analysing the robustness of dual encoders for dense retrieval against misspellings. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval. ACM." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.628, + 0.885, + 0.785 + ], + "angle": 0, + "content": "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander M. Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.799, + 0.885, + 0.877 + ], + "angle": 0, + "content": "Yingce Xia, Tao Qin, Wei Chen, Jiang Bian, Nenghai Yu, and Tie-Yan Liu. 2017. Dual supervised learning. In Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 3789-3798. PMLR." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.891, + 0.885, + 0.919 + ], + "angle": 0, + "content": "Lee Xiong, Chenyan Xiong, Ye Li, Kwok-Fung Tang, Jialin Liu, Paul N. Bennett, Junaid Ahmed, and" + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.885, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.52, + 0.941 + ], + "angle": 0, + "content": "1111" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.135, + 0.086, + 0.49, + 0.139 + ], + "angle": 0, + "content": "Arnold Overwijk. 2021. Approximate nearest neighbor negative contrastive learning for dense text retrieval. In International Conference on Learning Representations." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.149, + 0.49, + 0.201 + ], + "angle": 0, + "content": "Jingtao Zhan, Jiaxin Mao, Yiqun Liu, Min Zhang, and Shaoping Ma. 2020. Repbert: Contextualized text embeddings for first-stage retrieval. CoRR, abs/2006.15498." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.212, + 0.49, + 0.304 + ], + "angle": 0, + "content": "Shengyao Zhuang and Guido Zuccon. 2021. Dealing with typos for BERT-based passage retrieval and ranking. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 2836-2842, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.313, + 0.49, + 0.406 + ], + "angle": 0, + "content": "Shengyao Zhuang and Guido Zuccon. 2022. Character-bert and self-teaching for improving the robustness of dense retrievers on queries with typos. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '22, page 1444-1454, New York, NY, USA. Association for Computing Machinery." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1112" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.115, + 0.085, + 0.237, + 0.1 + ], + "angle": 0, + "content": "A Appendix" + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.118, + 0.461, + 0.134 + ], + "angle": 0, + "content": "A.1 Training Setup and Hyperparameters" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.145, + 0.49, + 0.5 + ], + "angle": 0, + "content": "The MS MARCO is a large-scale English language dataset for machine reading comprehension (MRC). The dataset consists of anonymized queries sampled from Bing's search query logs, each with human generated answers. The training set we used contains 400,782 training samples, each consisting of a query, positive passage, and a set of hard negative passages, which we randomly select 7 hard negative passages for each training sample. We set a batch size to 16 and use in-batch negative sampling for each training sample. Therefore, we obtain \\(7 + 8 * 15 = 127\\) negative passages for each training sample. We use the AdamW optimizer and learning rate of 1e-5 for 150,000 steps with a linear learning rate warm-up over the first 10,000 steps and a linear learning rate decay over the rest of the training steps. For our training method, we set the hyper-parameters \\(\\beta = 0.5\\), \\(\\gamma = 0.5\\), \\(\\sigma = 0.2\\), and the query augmentation size \\(K = 40\\). Using one V100 32G GPU, the BERT-based model training time is around 31 hours, while the CharacterBERT-based model training time is roughly 56 hours." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.521, + 0.411, + 0.537 + ], + "angle": 0, + "content": "A.2 Query Augmentation Examples" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.548, + 0.489, + 0.597 + ], + "angle": 0, + "content": "Table 4 provides examples of misspelled queries generated by the Query Augmentation for each original query." + }, + { + "type": "title", + "bbox": [ + 0.145, + 0.618, + 0.242, + 0.63 + ], + "angle": 0, + "content": "Original query:" + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.633, + 0.44, + 0.646 + ], + "angle": 0, + "content": "what is the goddess of agriculture in greek mythology" + }, + { + "type": "title", + "bbox": [ + 0.146, + 0.649, + 0.26, + 0.66 + ], + "angle": 0, + "content": "Misspelled queries:" + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.664, + 0.456, + 0.677 + ], + "angle": 0, + "content": "what is the goddoess of agriculture in greek mythology" + }, + { + "type": "text", + "bbox": [ + 0.146, + 0.679, + 0.443, + 0.693 + ], + "angle": 0, + "content": "what is the goddess of agriculture in greek mythology" + }, + { + "type": "text", + "bbox": [ + 0.146, + 0.696, + 0.443, + 0.709 + ], + "angle": 0, + "content": "what is the goddess of agriculture in greek mythologo" + }, + { + "type": "text", + "bbox": [ + 0.146, + 0.713, + 0.449, + 0.726 + ], + "angle": 0, + "content": "what is the goddesses of agriculture in greek mythology" + }, + { + "type": "text", + "bbox": [ + 0.146, + 0.729, + 0.45, + 0.742 + ], + "angle": 0, + "content": "what is the goddess of agriculture in greek mythology" + }, + { + "type": "text", + "bbox": [ + 0.146, + 0.745, + 0.454, + 0.758 + ], + "angle": 0, + "content": "what is the goddess of agriculture in greek mythology" + }, + { + "type": "text", + "bbox": [ + 0.146, + 0.762, + 0.442, + 0.775 + ], + "angle": 0, + "content": "what is the goddess of agriculture in greek mythology" + }, + { + "type": "text", + "bbox": [ + 0.146, + 0.779, + 0.45, + 0.791 + ], + "angle": 0, + "content": "what is the goddess of agriculture in grvek mythology" + }, + { + "type": "text", + "bbox": [ + 0.146, + 0.795, + 0.45, + 0.808 + ], + "angle": 0, + "content": "what is the goddess of agricultrue in greek mythology" + }, + { + "type": "text", + "bbox": [ + 0.146, + 0.811, + 0.45, + 0.824 + ], + "angle": 0, + "content": "what is the goddess of agriculture in greek mythology" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.836, + 0.49, + 0.879 + ], + "angle": 0, + "content": "Table 4: The outputs of Query Augmentation with \\( K = {10} \\) . We use different colors to indicate different types of typo: RandInsert , RandDelete , RandSub ," + }, + { + "type": "text", + "bbox": [ + 0.119, + 0.881, + 0.379, + 0.897 + ], + "angle": 0, + "content": "SwapNeighbor, and SwapAdjacent." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.085, + 0.627, + 0.099 + ], + "angle": 0, + "content": "A.3 Licenses" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.106, + 0.884, + 0.185 + ], + "angle": 0, + "content": "Datasets: The MS MARCO dataset is available under the MIT license, and the DL-typo dataset is available under the Apache license 2.0. These licenses allow users to use the datasets under nonrestrictive agreements." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.187, + 0.885, + 0.346 + ], + "angle": 0, + "content": "Softwares: We employ Hugging Face (Wolf et al., 2020) and Tevatron (Gao et al., 2022) libraries to train dense retrieval models. We utilize Ranx library (Bassani, 2022) to evaluate retrieval performance. These libraries are available under the Apache license 2.0 which allows both academic and commercial usages. For this reason, we release our code under the Apache license 2.0 to make our code fully accessible and compatible with the other codes we use." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1113" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.134, + 0.085, + 0.435, + 0.1 + ], + "angle": 0, + "content": "ACL 2023 Responsible NLP Checklist" + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.108, + 0.323, + 0.123 + ], + "angle": 0, + "content": "A For every submission:" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.127, + 0.534, + 0.145 + ], + "angle": 0, + "content": "A1. Did you describe the limitations of your work?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.145, + 0.225, + 0.158 + ], + "angle": 0, + "content": "Section 6" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.171, + 0.554, + 0.187 + ], + "angle": 0, + "content": "A2. Did you discuss any potential risks of your work?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.187, + 0.884, + 0.22 + ], + "angle": 0, + "content": "There is no potential risk associated with increasing the robustness of information retrieval applications to question containing misspellings." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.229, + 0.697, + 0.246 + ], + "angle": 0, + "content": "A3. Do the abstract and introduction summarize the paper's main claims?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.246, + 0.225, + 0.26 + ], + "angle": 0, + "content": "Section 1" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.272, + 0.671, + 0.289 + ], + "angle": 0, + "content": "A4. Have you used AI writing assistants when working on this paper?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.29, + 0.882, + 0.321 + ], + "angle": 0, + "content": "We use Grammarly to check grammatical errors and QuillBot to polish writing quality. These tools are applied to a certain number of sentences in each section, which are then reviewed by humans." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.332, + 0.489, + 0.348 + ], + "angle": 0, + "content": "B Did you use or create scientific artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.354, + 0.884, + 0.386 + ], + "angle": 0, + "content": "Section 3.1 for pre-trained language models, training dataset, and training toolkit. Section 3.2 for competitive methods. Section 3.3 for evaluation datasets and evaluation toolkit." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.396, + 0.531, + 0.412 + ], + "angle": 0, + "content": "B1. Did you cite the creators of artifacts you used?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.413, + 0.882, + 0.445 + ], + "angle": 0, + "content": "Section 3.1 for pre-trained language models, training dataset, and training toolkit. Section 3.2 for competitive methods. Section 3.3 for evaluation datasets and evaluation toolkit." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.454, + 0.78, + 0.472 + ], + "angle": 0, + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.472, + 0.255, + 0.487 + ], + "angle": 0, + "content": "Appendix A.3" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.497, + 0.882, + 0.563 + ], + "angle": 0, + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.564, + 0.254, + 0.578 + ], + "angle": 0, + "content": "Appendix A.3" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.589, + 0.882, + 0.637 + ], + "angle": 0, + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.638, + 0.882, + 0.686 + ], + "angle": 0, + "content": "We did not collect any data. The datasets we used are publicly available and widely used in information retrieval literature. The data is already anonymized by the creators of the datasets. Therefore we do not need to anonymize the data." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.696, + 0.882, + 0.73 + ], + "angle": 0, + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.73, + 0.254, + 0.744 + ], + "angle": 0, + "content": "Appendix A.1" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.754, + 0.882, + 0.836 + ], + "angle": 0, + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be." + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.837, + 0.642, + 0.852 + ], + "angle": 0, + "content": "Section 3.3 for the evaluation set Appendix A.1 for the training set" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.856, + 0.878, + 0.881 + ], + "angle": 0, + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1114" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.115, + 0.084, + 0.495, + 0.101 + ], + "angle": 0, + "content": "C Did you run computational experiments?" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.106, + 0.207, + 0.12 + ], + "angle": 0, + "content": "Section 4" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.131, + 0.88, + 0.163 + ], + "angle": 0, + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.165, + 0.254, + 0.181 + ], + "angle": 0, + "content": "Appendix A.1" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.19, + 0.882, + 0.223 + ], + "angle": 0, + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.224, + 0.253, + 0.24 + ], + "angle": 0, + "content": "Appendix A.1" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.249, + 0.882, + 0.297 + ], + "angle": 0, + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?" + }, + { + "type": "text", + "bbox": [ + 0.148, + 0.299, + 0.882, + 0.331 + ], + "angle": 0, + "content": "Section 4.1 for Main Results Section 4.2 for Query Augmentation Size Study Section 4.3 for Loss Ablation Study Section 4.4 for Query Distributions" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.341, + 0.882, + 0.388 + ], + "angle": 0, + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.39, + 0.697, + 0.406 + ], + "angle": 0, + "content": "Our evaluation is parameter free, therefore there is no parameter settings." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.417, + 0.878, + 0.433 + ], + "angle": 0, + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.437, + 0.215, + 0.453 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.464, + 0.882, + 0.495 + ], + "angle": 0, + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.497, + 0.249, + 0.512 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.523, + 0.882, + 0.57 + ], + "angle": 0, + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.572, + 0.249, + 0.587 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.598, + 0.882, + 0.645 + ], + "angle": 0, + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.647, + 0.249, + 0.662 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.673, + 0.874, + 0.688 + ], + "angle": 0, + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.69, + 0.249, + 0.705 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.716, + 0.881, + 0.747 + ], + "angle": 0, + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.749, + 0.249, + 0.764 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1115" + } + ] +] \ No newline at end of file diff --git a/2023/Typo-Robust Representation Learning for Dense Retrieval/8977932e-2af7-46c1-9920-a3e3abaf9a04_origin.pdf b/2023/Typo-Robust Representation Learning for Dense Retrieval/8977932e-2af7-46c1-9920-a3e3abaf9a04_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..01fc47ae4909991e5ef1c01b6a53bd681998ba49 --- /dev/null +++ b/2023/Typo-Robust Representation Learning for Dense Retrieval/8977932e-2af7-46c1-9920-a3e3abaf9a04_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:97f949bb4629932f3cff142e5de91570c2bd71c77a48bcafba21da51d6ce5ddf +size 922755 diff --git a/2023/Typo-Robust Representation Learning for Dense Retrieval/full.md b/2023/Typo-Robust Representation Learning for Dense Retrieval/full.md new file mode 100644 index 0000000000000000000000000000000000000000..bf735cf7dfb1a1a1005ad655902e17d5098be924 --- /dev/null +++ b/2023/Typo-Robust Representation Learning for Dense Retrieval/full.md @@ -0,0 +1,334 @@ +# Typo-Robust Representation Learning for Dense Retrieval + +Panuthep Tasawong†, Wuttikorn Ponwitayarat†, Peerat Limkonchotiwat†, Can Udomcharoenchaikit†, Ekapol Chuangsuwanich‡, Sarana Nutanong† + +†School of Information Science and Technology, VISTEC, Thailand +‡Department of Computer Engineering, Chulalongkorn University, Thailand {panuthep.t_s20, wuttikorn.p_s22, peerat.l_s19, canu_pro, snutanon} @vistec.ac.th, ekapolc@cp.eng.chula.ac.th + +# Abstract + +Dense retrieval is a basic building block of information retrieval applications. One of the main challenges of dense retrieval in real-world settings is the handling of queries containing misspelled words. A popular approach for handling misspelled queries is minimizing the representations discrepancy between misspelled queries and their pristine ones. Unlike the existing approaches, which only focus on the alignment between misspelled and pristine queries, our method also improves the contrast between each misspelled query and its surrounding queries. To assess the effectiveness of our proposed method, we compare it against the existing competitors using two benchmark datasets and two base encoders. Our method outperforms the competitors in all cases with misspelled queries. Our code and models are available at https://github.com/panuthept/DST-DenseRetrieval. + +# 1 Introduction + +Dense retrieval is a fundamental component in many information retrieval applications, such as open-domain question answering and ad-hoc retrieval. The objective is to score and rank a large collection of candidate passages based on their similarity to a given query. The performance of dense retrieval relies on representation learning. A popular approach is to finetune a pre-trained language model to create an embedding space that puts each query closer to its corresponding passages (Zhan et al., 2020; Khattab and Zaharia, 2020; Xiong et al., 2021; Qu et al., 2021; Ren et al., 2021a,b). + +One of the major challenges of dense retrieval is the handling of misspelled queries which induces representations of the misspelled queries to be closer to irrelevant passages than their corresponding passages. Several studies have demonstrated that misspellings in search queries can substantially degrade retrieval performance (Zhuang and Zuccon, 2021; Penha et al., 2022), specifically + +when informative terms, such as entity mentions, are misspelled (Sidiropoulos and Kanoulas, 2022). + +To create a retrieval model that is capable of handling misspelled queries, researchers have proposed different training methods to align representations of misspelled queries with their pristine ones. Zhuang and Zuccon (2021, 2022) devise augmentation methods to generate misspelled queries and propose training methods, Typos-aware Training and Self-Teaching (ST), to encourage consistency between outputs of misspelled queries and their non-misspelled counterparts. Alternatively, Sidiropoulos and Kanoulas (2022) apply contrastive loss to enforce representations of misspelled queries to be closer to their corresponding non-misspelled queries. Although these methods can improve the performance of retrieval models for misspelled queries, there is still a substantial performance drop for misspelled queries. + +In this paper, we propose a training method to improve dense retrieval for handling misspelled queries based on the following desired properties: + +- Alignment: the method should be able to align queries with their corresponding passages. +- Robustness: the method should be able to align misspelled queries with their pristine queries. +- Contrast: the method should be able to separate queries that refer to different passages and passages that correspond to different queries. + +In contrast to the existing methods for handling misspelled queries that only satisfy the Alignment and Robustness properties, our method also aims to satisfy the Contrast property. Increasing the distance between dissimilar queries should help distinguish misspelled queries from other distinct queries. We design the following components for our training method: (i) Dual Self-Teaching (DST) incorporates the ideas of Dual Learning (Xia et al., 2017; Li et al., 2021) and Self-Teaching (Zhuang and Zuccon, 2022) to train robust dense retrieval in a bidirectional manner: passage retrieval and + +query retrieval. (ii) Query Augmentation generates a numerous number of misspelling variations for each query to supply our training objective. + +Experimental studies were conducted to assess the efficiency of the proposed method in comparison to existing approaches. We conduct experiments based on two different pre-trained language models. We evaluate using two passage retrieval benchmark datasets, a standard one and a specialized one for misspellings robustness evaluation. For each dataset, we measure performance on both misspelled and non-misspelled queries, where the misspelled queries are both generated and real-world queries. The experimental results show that the proposed method outperforms the best existing methods for enhancing the robustness of dense retrieval against misspellings without sacrificing performance for non-misspelled queries. + +We summarize our contributions as follows: + +- We propose a novel training method to enhance the robustness of dense retrieval against misspellings by incorporating three desired properties: Alignment, Robustness, and Contrast. +- We introduce Dual Self-Teaching (DST) which adopts the idea of Dual Learning and Self-Teaching to learn robust representations. In addition, we propose Query Augmentation to generate multiple views of a particular query under different misspelling scenarios. +- We evaluate our method on misspelled and non-misspelled queries from two passage retrieval datasets. The results show that our method outperforms the previous state-of-the-art methods by a significant margin on misspelled queries. + +# 2 Methodology + +We propose a training pipeline to enhance the dense retrieval capability for handling spelling variations and mistakes in queries. As shown in Figure 1, the training pipeline comprises three steps. (i) Query Augmentation: we augment each query in the training set into multiple misspelled queries using the typo generators provided by Zhuang and Zuccon (2021). (ii) Similarity Score Calculation: we compute similarity score distributions between queries and passages for passage retrieval and query retrieval tasks using in-batch negative queries and passages, with additional hard negative passages. (iii) Dual Self-Teaching Loss Calculation: we compute the DST loss using the similarity score distributions to achieve all three desired properties. + +# 2.1 Query Augmentation + +The purpose of this step is to guide the learning with a broad array of possible misspelling patterns. Let $\mathbf{Q}$ denote a set $\{q_{1}, q_{2}, \ldots, q_{N}\}$ of $N$ queries. From all queries in $\mathbf{Q}$ , we generate a set of $K \times N$ misspelled queries $\mathcal{Q}' = \{\langle q_{1,k}', q_{2,k}', \ldots, q_{N,k}'\rangle\}_{k=1}^{K}$ , where $K$ is the misspelling variations. We use five typo generators proposed by Zhuang and Zuccon (2021), including: RandInsert, RandDelete, RandSub, SwapNeighbor, and SwapAdjacent. Please refer to Appendix A.2 for examples of the misspelled queries. + +# 2.2 Similarity Score Calculation + +Let $S(a, \mathbf{B})$ denote a function that computes a similarity score distribution of any vector $a$ over any set of vectors $\mathbf{B}$ : + +$$ +S (a, \mathbf {B}) = \left\{b _ {i} \in \mathbf {B} \left| \frac {\exp (a \cdot b _ {i})}{\sum_ {b _ {j} \in \mathbf {B}} \exp (a \cdot b _ {j})} \right. \right\} \tag {1} +$$ + +Given $\mathbf{P} = \{p_1, p_2, \dots, p_M\}$ to be a set of $M$ passages and $\mathbf{Q}_k' = \{q_{1,k}', q_{2,k}', \dots, q_{N,k}'\}$ to be the $k^{th}$ set of misspelled queries in $\mathcal{Q}'$ , we compute two groups of score distributions as follows: + +Passage retrieval: we calculate score distributions in a query-to-passages direction for each original query $s_p = S(q_n, \mathbf{P})$ and misspelled query $s_p'^k = S(q_{n,k}', \mathbf{P})$ . +- Query retrieval: we calculate score distributions in a passage-to-queries direction for original queries $s_q = S(p_m, \mathbf{Q})$ and each set of misspellled queries $s_q^{\prime k} = S(p_m, \mathbf{Q}_k^{\prime})$ . + +This way, we produce four different score distributions $(s_p,s_p^{\prime k},s_q,s_q^{\prime k})$ for our training objective. + +# 2.3 Dual Self-Teaching Loss Calculation + +We design the Dual Self-Teaching loss $(\mathcal{L}_{\mathrm{DST}})$ to capture the three desired properties: Alignment, Robustness, and Contrast. + +$$ +\mathcal {L} _ {\mathrm {D S T}} = \underbrace {(1 - \beta) \mathcal {L} _ {\mathrm {D C E}}} _ {\text {D u a l C r o s s - E n t r o p y}} + \underbrace {\beta \mathcal {L} _ {\mathrm {D K L}}} _ {\text {D u a l K L - D i v e r g e n c e}} \tag {2} +$$ + +Dual Cross-Entropy loss $(\mathcal{L}_{\mathrm{DCE}})$ satisfies the Alignment and Contrast properties by utilizing cross-entropy losses to learn score distributions of the original queries for passage retrieval $(s_p)$ and query retrieval $(s_q)$ given labels $y_{p}$ and $y_{q}$ . + +$$ +\mathcal {L} _ {\mathrm {D C E}} = \underbrace {(1 - \gamma) \mathcal {L} _ {\mathrm {C E}} ^ {(P)} \left(s _ {p} , y _ {p}\right)} _ {\text {P a s s a g e R e t i v e a l}} + \underbrace {\gamma \mathcal {L} _ {\mathrm {C E}} ^ {(Q)} \left(s _ {q} , y _ {q}\right)} _ {\text {Q u e r y R e t i v e a l}} \tag {3} +$$ + +![](images/9cb707006d2ab480650319b69a87a6bda5748b990dc59544a681e1c57616d9ac.jpg) +Figure 1: The proposed training pipeline consists of three steps: (a) Query Augmentation, (b) Similarity Score Calculation, and (c) Dual Self-Teaching Loss Calculation. + +Minimizing the $\mathcal{L}_{\mathrm{CE}}^{(P)}$ term will increase the similarity scores between queries and their relevant passages to be higher than other irrelevant passages by separating the relevant and irrelevant passages from one another. Minimizing the $\mathcal{L}_{\mathrm{CE}}^{(Q)}$ term will increase the similarity scores between passages and their relevant queries to be higher than other irrelevant queries by separating the relevant and irrelevant queries from one another. In this manner, minimizing one of the two terms will align queries with their corresponding passages, satisfying the Alignment property. Moreover, minimizing both terms will separate queries that refer to different passages and passages that belong to different queries, satisfying the Contrast property. + +Dual KL-Divergence loss $(\mathcal{L}_{\mathrm{DKL}})$ aims to fulfill the Robustness property by using KL losses to match score distributions of misspelled queries $\{s_p^{\prime 1}, s_p^{\prime 2}, \ldots, s_p^{\prime K}\}$ and $\{s_q^{\prime 1}, s_q^{\prime 2}, \ldots, s_q^{\prime K}\}$ to the score distributions of the original query $s_p$ and $s_q$ . + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {D K L}} = \frac {1}{K} \sum_ {k = 1} ^ {K} \underbrace {(1 - \sigma) \mathcal {L} _ {\mathrm {K L}} ^ {(P)} \left(s _ {p} ^ {\prime k} , s _ {p}\right)} _ {\text {P a s s a g e R e t r i e v a l C o n s i s t e n c y}} \tag {4} \\ + \underbrace {\sigma \mathcal {L} _ {\mathrm {K L}} ^ {(Q)} \left(s _ {q} ^ {\prime k} , s _ {q}\right)} _ {\text {Q u e r y R e t r i e v a l C o n s i s t e n c y}} \\ \end{array} +$$ + +Minimizing $\mathcal{L}_{\mathrm{KL}}^{(P)}$ and $\mathcal{L}_{\mathrm{KL}}^{(Q)}$ will reduce the discrepancy between misspelled and non-misspelled queries for both query-to-passages and passage-to-queries score distributions. This way, we implicitly align representations of the misspelled queries to the original queries, satisfying the Robustness property. To stabilize training, we apply stop-gradient to the score distributions of the original queries $(s_p$ and $s_q)$ in the $\mathcal{L}_{\mathrm{DKL}}$ . The $\beta$ , $\gamma$ , and $\sigma$ are the balancing coefficients selected by hyper-parameter tuning + +on a development set. With this loss combination, we achieve all three desired properties. + +# 3 Experimental Settings + +# 3.1 Training Details + +We experiment on two pre-trained language models, BERT (Devlin et al., 2019) and Character-BERT (El Boukkouri et al., 2020). We train models only on the training set of MS MARCO dataset (Nguyen et al., 2016). Moreover, the training data provided by the Tevatron toolkit (Gao et al., 2022) also contains hard negative passages. We include the training set details and hyper-parameter settings in Appendix A.1. + +# 3.2 Competitive Methods + +To show the effectiveness of our method, we compare our work with the following baseline and competitive training methods. + +- DPR (Karpukhin et al., 2020) is a baseline training method that trains dense retrieval merely on non-misspelled queries using $\mathcal{L}_{\mathrm{CE}}^{(P)}$ loss. +- $DPR + Aug$ (Zhuang and Zuccon, 2021) is the Typos-aware Training method which trains dense retrieval on both misspelled and non-misspelled queries using $\mathcal{L}_{\mathrm{CE}}^{(P)}$ loss. +- $DPR + Aug + CL$ (Sidiropoulos and Kanoulas, 2022) employs additional contrastive loss to train the misspelled queries. +- $DPR + ST$ (Zhuang and Zuccon, 2022) is the Self-Teaching method that trains dense retrieval on both misspelled and non-misspelled queries using $\mathcal{L}_{\mathrm{CE}}^{(P)}$ and $\mathcal{L}_{\mathrm{KL}}^{(P)}$ losses. + +Note that their query augmentation method is identical to the Query Augmentation with $K = 1$ . We retrain all models using the same setting described in the previous section. + +
MethodsBERT-basedCharacterBERT-based
MS MARCODL-typoMS MARCODL-typo
MRR@10R@1000nDCG@10MRRMAPMRR@10R@1000nDCG@10MRRMAP
DPR.143 (.331).696 (.954).276 (.682).431 (.873).175 (.563).162 (.321).726 (.945).268 (.643).376 (.832).212 (.503)
+ Aug.227 (.334).857 (.950).398 (.682).530 (.806).286 (.565).258 (.326).883 (.946).414 (.631).578 (.783).318 (.512)
+ Aug + CL.234 (.335).867 (.951).387 (.668).536 (.864).267 (.544).263 (.330).894 (.947).466 (.677).635 (.819).360 (.544)
+ ST.237 (.333).874 (.950).392 (.677).525 (.852).283 (.557).274 (.332).900 (.947).469 (.650).619 (.810).359 (.517)
+ DST (our).260† (.336).894† (.954).432 (.673).558 (.833).343† (.568).288† (.332).918† (.949).529† (.673).742† (.854).403 (.537)
+ +# 3.3 Dataset and Evaluation + +Datasets. We evaluate the effectiveness of DST on two passage retrieval datasets, MS MARCO and DL-typo (Zhuang and Zuccon, 2022), each with misspelled and non-misspelled queries. There are 8.8 million candidate passages for both datasets. The development set of MS MARCO contains 6,980 non-misspelled queries. To obtain misspelled queries, we use the typos generator method proposed by Zhuang and Zuccon (2021) to generate 10 misspelled variations for each original query. The DL-typo provides 60 real-world misspelled queries and 60 corresponding non-misspelled queries that are corrected manually. + +Evaluation. We use the standard metrics originally used by each dataset's creators. For MS MARCO, each misspelled query performance is the average of 10 measurements. We employ Ranx evaluation library (Bassani, 2022) to measure performance and statistical significance. Specifically, we use a two-tailed paired t-test with Bonferroni correction to measure the statistical significance $(p < 0.05)$ . + +# 4 Experimental Results + +# 4.1 Main Results + +As shown in Table 1, the results indicate that DST outperforms competitive methods for misspelled queries in every case without sacrificing performance for non-misspelled queries in eight out of ten cases. We observe some performance trade-offs for the BERT-based model in the DL-typo dataset's non-misspelling scores (nDCG@10 and MRR). Aside from that, there is no performance trade-off for the CharacterBERT-based model. These outcomes conform with the observation in Figure 2 (Section 4.4) that DST improves the Robustness and Contrast of misspelled queries. + +# 4.2 Query Augmentation Size Study + +To study the benefit of query augmentation and find the optimal augmentation size, we measure the performance of BERT-based dense retrieval models trained with DST using the query augmentation size $K$ of 1, 10, 20, 40, and 60. Note that the query augmentation method used in previous works is a special case of Query Augmentation when $K = 1$ . We report the results using MRR@10 for the development set of the MS MARCO dataset. We also report training time to show trade-offs between performance and computation. + +Table 1: Results of different training methods on misspelled and non-misspelled queries. We report the results in the format of "misspelled query performance (non-misspelled query performance)". We emphasize the best score with bold text and the second-best score with underlined text. We use $\dagger$ to denote DST results that significantly outperform the second-best result ( $p < 0.05$ ). + +
QueriesK
110204060
Original.334.334.335.336.332
Misspelled.251.258.260.260.260
Training time (hr)1820233139
+ +Table 2: Results of query augmentation size study. We train all models in this experiment on a V100 32G GPU. + +As shown in Table 2, the results indicate that increasing $K$ improves the performance of both misspelled and non-misspelled queries, but only up to a certain point, after which the performance begins to decline. We observe that setting $K = 40$ produces the best results, and there is no further performance improvement after this point. + +# 4.3 Loss Ablation Study + +In this experiment, we study the benefit of each term in DST by training BERT-based dense retrieval models on variant loss combinations with $K = 40$ . The results in Table 3 reveal that $\mathcal{L}_{\mathrm{KL}}^{(P)}$ and $\mathcal{L}_{\mathrm{KL}}^{(Q)}$ terms positively contribute to the performance of misspelled and non-misspelled queries, with the $\mathcal{L}_{\mathrm{KL}}^{(P)}$ being more significant. The $\mathcal{L}_{\mathrm{CE}}^{(P)}$ term is crucial for retrieval performance, whereas the $\mathcal{L}_{\mathrm{CE}}^{(Q)}$ term indirectly improves the performance + +
L(PCE)L(QCE)L(PKL)L(QKL)MRR@10
✓✓✓✓.260 (.336)
✓✓✓.257 (.335)
✓✓✓.228 (.326)
✓✓✓.251 (.337)
✓✓✓.087 (.114)
✓✓.249 (.336)
✓✓.120 (.158)
+ +Table 3: Loss ablation study results on MS MARCO. + +of misspelled queries by separating their pristine queries from the surrounding queries. Disabling query retrieval terms $(\mathcal{L}_{\mathrm{CE}}^{(Q)})$ and $\mathcal{L}_{\mathrm{KL}}^{(Q)}$ greatly reduces performances for misspelled queries. The passage retrieval terms $(\mathcal{L}_{\mathrm{CE}}^{(P)})$ and $\mathcal{L}_{\mathrm{KL}}^{(P)}$ are indispensable and cannot be substituted. + +# 4.4 Query Distributions + +The purpose of this section is to study the impact of our training method on the Robustness and Contrast of misspelled queries. We also compare our method against the baseline and competitive methods to show its effectiveness. The Robustness and Contrast of misspelled queries are illustrated using the following kernel density graphs: + +- Original-to-Misspell: the cosine similarity distribution between original and misspelled queries. +- Original-to-Neighbor: the cosine similarity distribution between original and neighbor queries. The Robustness property is emphasized by the Original-to-Misspell distribution having high cosine similarity. On the other hand, the Contrast property is emphasized by the small overlapping between Original-to-Misspell and Original-to-Neighbor distributions. The results in Figure 2 show that our method (c) produces the best Robustness and Contrast properties for misspelled queries in comparison to other methods. + +# 5 Conclusion + +This paper aims to address the misspelling problem in dense retrieval. We formulate three desired properties for making dense retrieval robust to misspellings: Alignment, Robustness, and Contrast. Unlike previous methods, which only focus on the Alignment and Robustness properties, our method considers all the desired properties. The empirical results show that our method performs best against misspelled queries, revealing the importance of the Contrast property for handling misspellings. + +![](images/d5b2d927d187e10c36b3de26eb7003e3a424cda229614d3b30ef44fc48c48a7d.jpg) +(a) DPR (Karpukhin et al., 2020). + +![](images/4be0a7c423333aa8d7cbdccb802eb94c004878c415d2f6917ecdc767a5e037b9.jpg) +(b) Self-Teaching (Zhuang and Zuccon, 2022). + +![](images/497082c56023ac179644de161784f0c8592c628036e5c3c6e31863042aada675.jpg) +(c) Dual Self-Teaching (our). +Figure 2: Kernel density of Original-to-Neighbor (orange) and Original-to-Misspell (blue) of different training methods. + +# 6 Limitations + +We list the limitations of our work as follows: + +- The Query Augmentation is designed for the English alphabet; therefore, other languages with different alphabets will require further work. +- Since the training strategy relies on fine-tuning a pre-trained language model using a large passage retrieval dataset, it may not be suitable for languages with limited resources + +# References + +Elias Bassani. 2022. ranx: A blazing-fast python library for ranking evaluation and comparison. In ECIR (2), volume 13186 of Lecture Notes in Computer Science, pages 259-264. Springer. +Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics. +Hicham El Boukkouri, Olivier Ferret, Thomas Lavergne, Hiroshi Noji, Pierre Zweigenbaum, and Jun'ichi Tsujii. 2020. CharacterBERT: Reconciling ELMo and BERT for word-level open-vocabulary representations from characters. In Proceedings of the 28th International Conference on Computational Linguistics, pages 6903-6915, Barcelona, Spain (Online). International Committee on Computational Linguistics. +Luyu Gao, Xueguang Ma, Jimmy J. Lin, and Jamie Callan. 2022. Tevatron: An efficient and flexible toolkit for dense retrieval. ArXiv, abs/2203.05765. +Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6769-6781, Online. Association for Computational Linguistics. +Omar Khattab and Matei Zaharia. 2020. Colbert: Efficient and effective passage search via contextualized late interaction over bert. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '20, page 39-48, New York, NY, USA. Association for Computing Machinery. +Yizhi Li, Zhenghao Liu, Chenyan Xiong, and Zhiyuan Liu. 2021. More robust dense retrieval with contrastive dual learning. In Proceedings of the 2021 ACM SIGIR International Conference on Theory of Information Retrieval, ICTIR '21, page 287-296, New York, NY, USA. Association for Computing Machinery. +Tri Nguyen, Mir Rosenberg, Xia Song, Jianfeng Gao, Saurabh Tiwary, Rangan Majumder, and Li Deng. 2016. MS MARCO: A human generated machine reading comprehension dataset. In Proceedings of the Workshop on Cognitive Computation: Integrating neural and symbolic approaches 2016 co-located with the 30th Annual Conference on Neural Information Processing Systems (NIPS 2016), Barcelona, Spain, December 9, 2016, volume 1773 of CEUR Workshop Proceedings. CEUR-WS.org. + +Gustavo Penha, Arthur Camara, and Claudia Hauff. 2022. Evaluating the robustness of retrieval pipelines with query variation generators. In Advances in Information Retrieval, pages 397-412, Cham. Springer International Publishing. +Yingqi Qu, Yuchen Ding, Jing Liu, Kai Liu, Ruiyang Ren, Wayne Xin Zhao, Daxiang Dong, Hua Wu, and Haifeng Wang. 2021. RocketQA: An optimized training approach to dense passage retrieval for open-domain question answering. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5835-5847, Online. Association for Computational Linguistics. +Ruiyang Ren, Shangwen Lv, Yingqi Qu, Jing Liu, Wayne Xin Zhao, QiaoQiao She, Hua Wu, Haifeng Wang, and Ji-Rong Wen. 2021a. PAIR: Leveraging passage-centric similarity relation for improving dense passage retrieval. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 2173-2183, Online. Association for Computational Linguistics. +Ruiyang Ren, Yingqi Qu, Jing Liu, Wayne Xin Zhao, QiaoQiao She, Hua Wu, Haifeng Wang, and Ji-Rong Wen. 2021b. RocketQAv2: A joint training method for dense passage retrieval and passage re-ranking. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 2825-2835, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics. +Georgios Sidiropoulos and Evangelos Kanoulas. 2022. Analysing the robustness of dual encoders for dense retrieval against misspellings. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval. ACM. +Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander M. Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online. Association for Computational Linguistics. +Yingce Xia, Tao Qin, Wei Chen, Jiang Bian, Nenghai Yu, and Tie-Yan Liu. 2017. Dual supervised learning. In Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 3789-3798. PMLR. +Lee Xiong, Chenyan Xiong, Ye Li, Kwok-Fung Tang, Jialin Liu, Paul N. Bennett, Junaid Ahmed, and + +Arnold Overwijk. 2021. Approximate nearest neighbor negative contrastive learning for dense text retrieval. In International Conference on Learning Representations. +Jingtao Zhan, Jiaxin Mao, Yiqun Liu, Min Zhang, and Shaoping Ma. 2020. Repbert: Contextualized text embeddings for first-stage retrieval. CoRR, abs/2006.15498. +Shengyao Zhuang and Guido Zuccon. 2021. Dealing with typos for BERT-based passage retrieval and ranking. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 2836-2842, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics. +Shengyao Zhuang and Guido Zuccon. 2022. Character-bert and self-teaching for improving the robustness of dense retrievers on queries with typos. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '22, page 1444-1454, New York, NY, USA. Association for Computing Machinery. + +# A Appendix + +# A.1 Training Setup and Hyperparameters + +The MS MARCO is a large-scale English language dataset for machine reading comprehension (MRC). The dataset consists of anonymized queries sampled from Bing's search query logs, each with human generated answers. The training set we used contains 400,782 training samples, each consisting of a query, positive passage, and a set of hard negative passages, which we randomly select 7 hard negative passages for each training sample. We set a batch size to 16 and use in-batch negative sampling for each training sample. Therefore, we obtain $7 + 8 * 15 = 127$ negative passages for each training sample. We use the AdamW optimizer and learning rate of 1e-5 for 150,000 steps with a linear learning rate warm-up over the first 10,000 steps and a linear learning rate decay over the rest of the training steps. For our training method, we set the hyper-parameters $\beta = 0.5$ , $\gamma = 0.5$ , $\sigma = 0.2$ , and the query augmentation size $K = 40$ . Using one V100 32G GPU, the BERT-based model training time is around 31 hours, while the CharacterBERT-based model training time is roughly 56 hours. + +# A.2 Query Augmentation Examples + +Table 4 provides examples of misspelled queries generated by the Query Augmentation for each original query. + +# Original query: + +what is the goddess of agriculture in greek mythology + +# Misspelled queries: + +what is the goddoess of agriculture in greek mythology + +what is the goddess of agriculture in greek mythology + +what is the goddess of agriculture in greek mythologo + +what is the goddesses of agriculture in greek mythology + +what is the goddess of agriculture in greek mythology + +what is the goddess of agriculture in greek mythology + +what is the goddess of agriculture in greek mythology + +what is the goddess of agriculture in grvek mythology + +what is the goddess of agricultrue in greek mythology + +what is the goddess of agriculture in greek mythology + +Table 4: The outputs of Query Augmentation with $K = {10}$ . We use different colors to indicate different types of typo: RandInsert , RandDelete , RandSub , + +SwapNeighbor, and SwapAdjacent. + +# A.3 Licenses + +Datasets: The MS MARCO dataset is available under the MIT license, and the DL-typo dataset is available under the Apache license 2.0. These licenses allow users to use the datasets under nonrestrictive agreements. + +Softwares: We employ Hugging Face (Wolf et al., 2020) and Tevatron (Gao et al., 2022) libraries to train dense retrieval models. We utilize Ranx library (Bassani, 2022) to evaluate retrieval performance. These libraries are available under the Apache license 2.0 which allows both academic and commercial usages. For this reason, we release our code under the Apache license 2.0 to make our code fully accessible and compatible with the other codes we use. + +# A For every submission: + +A1. Did you describe the limitations of your work? + +Section 6 + +A2. Did you discuss any potential risks of your work? + +There is no potential risk associated with increasing the robustness of information retrieval applications to question containing misspellings. + +A3. Do the abstract and introduction summarize the paper's main claims? + +Section 1 + +A4. Have you used AI writing assistants when working on this paper? + +We use Grammarly to check grammatical errors and QuillBot to polish writing quality. These tools are applied to a certain number of sentences in each section, which are then reviewed by humans. + +# B Did you use or create scientific artifacts? + +Section 3.1 for pre-trained language models, training dataset, and training toolkit. Section 3.2 for competitive methods. Section 3.3 for evaluation datasets and evaluation toolkit. + +B1. Did you cite the creators of artifacts you used? + +Section 3.1 for pre-trained language models, training dataset, and training toolkit. Section 3.2 for competitive methods. Section 3.3 for evaluation datasets and evaluation toolkit. + +B2. Did you discuss the license or terms for use and / or distribution of any artifacts? + +Appendix A.3 + +B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? + +Appendix A.3 + +B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? + +We did not collect any data. The datasets we used are publicly available and widely used in information retrieval literature. The data is already anonymized by the creators of the datasets. Therefore we do not need to anonymize the data. + +B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? + +Appendix A.1 + +B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. + +Section 3.3 for the evaluation set Appendix A.1 for the training set + +The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance. + +# C Did you run computational experiments? + +Section 4 + +C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? + +Appendix A.1 + +C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? + +Appendix A.1 + +C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? + +Section 4.1 for Main Results Section 4.2 for Query Augmentation Size Study Section 4.3 for Loss Ablation Study Section 4.4 for Query Distributions + +C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? + +Our evaluation is parameter free, therefore there is no parameter settings. + +# D Did you use human annotators (e.g., crowdworkers) or research with human participants? + +Left blank. + +D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? + +No response. + +D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? + +No response. + +D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? + +No response. + +D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? + +No response. + +D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? + +No response. \ No newline at end of file diff --git a/2023/Typo-Robust Representation Learning for Dense Retrieval/images.zip b/2023/Typo-Robust Representation Learning for Dense Retrieval/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..72ada52f40cb4285c02296166ab1840ceabd5692 --- /dev/null +++ b/2023/Typo-Robust Representation Learning for Dense Retrieval/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1cb00796b0a15e50b4aa6ae5431602fd3d7bb4c069a8af493b039aeeee522433 +size 258770 diff --git a/2023/Typo-Robust Representation Learning for Dense Retrieval/layout.json b/2023/Typo-Robust Representation Learning for Dense Retrieval/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3889ef33e545b57ba9f813fc915281868952769a --- /dev/null +++ b/2023/Typo-Robust Representation Learning for Dense Retrieval/layout.json @@ -0,0 +1,8142 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 116, + 71, + 477, + 88 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 71, + 477, + 88 + ], + "spans": [ + { + "bbox": [ + 116, + 71, + 477, + 88 + ], + "type": "text", + "content": "Typo-Robust Representation Learning for Dense Retrieval" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 108, + 95, + 491, + 124 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 95, + 491, + 124 + ], + "spans": [ + { + "bbox": [ + 108, + 95, + 491, + 124 + ], + "type": "text", + "content": "Panuthep Tasawong†, Wuttikorn Ponwitayarat†, Peerat Limkonchotiwat†, Can Udomcharoenchaikit†, Ekapol Chuangsuwanich‡, Sarana Nutanong†" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 114, + 124, + 483, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 124, + 483, + 194 + ], + "spans": [ + { + "bbox": [ + 114, + 124, + 483, + 194 + ], + "type": "text", + "content": "†School of Information Science and Technology, VISTEC, Thailand \n‡Department of Computer Engineering, Chulalongkorn University, Thailand {panuthep.t_s20, wuttikorn.p_s22, peerat.l_s19, canu_pro, snutanon} @vistec.ac.th, ekapolc@cp.eng.chula.ac.th" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 84, + 233, + 276, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 233, + 276, + 472 + ], + "spans": [ + { + "bbox": [ + 84, + 233, + 276, + 472 + ], + "type": "text", + "content": "Dense retrieval is a basic building block of information retrieval applications. One of the main challenges of dense retrieval in real-world settings is the handling of queries containing misspelled words. A popular approach for handling misspelled queries is minimizing the representations discrepancy between misspelled queries and their pristine ones. Unlike the existing approaches, which only focus on the alignment between misspelled and pristine queries, our method also improves the contrast between each misspelled query and its surrounding queries. To assess the effectiveness of our proposed method, we compare it against the existing competitors using two benchmark datasets and two base encoders. Our method outperforms the competitors in all cases with misspelled queries. Our code and models are available at https://github.com/panuthept/DST-DenseRetrieval." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 481, + 155, + 494 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 481, + 155, + 494 + ], + "spans": [ + { + "bbox": [ + 68, + 481, + 155, + 494 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 502, + 291, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 502, + 291, + 664 + ], + "spans": [ + { + "bbox": [ + 67, + 502, + 291, + 664 + ], + "type": "text", + "content": "Dense retrieval is a fundamental component in many information retrieval applications, such as open-domain question answering and ad-hoc retrieval. The objective is to score and rank a large collection of candidate passages based on their similarity to a given query. The performance of dense retrieval relies on representation learning. A popular approach is to finetune a pre-trained language model to create an embedding space that puts each query closer to its corresponding passages (Zhan et al., 2020; Khattab and Zaharia, 2020; Xiong et al., 2021; Qu et al., 2021; Ren et al., 2021a,b)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 666, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 666, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 666, + 291, + 773 + ], + "type": "text", + "content": "One of the major challenges of dense retrieval is the handling of misspelled queries which induces representations of the misspelled queries to be closer to irrelevant passages than their corresponding passages. Several studies have demonstrated that misspellings in search queries can substantially degrade retrieval performance (Zhuang and Zuccon, 2021; Penha et al., 2022), specifically" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 213, + 526, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 213, + 526, + 239 + ], + "spans": [ + { + "bbox": [ + 302, + 213, + 526, + 239 + ], + "type": "text", + "content": "when informative terms, such as entity mentions, are misspelled (Sidiropoulos and Kanoulas, 2022)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 241, + 526, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 241, + 526, + 470 + ], + "spans": [ + { + "bbox": [ + 302, + 241, + 526, + 470 + ], + "type": "text", + "content": "To create a retrieval model that is capable of handling misspelled queries, researchers have proposed different training methods to align representations of misspelled queries with their pristine ones. Zhuang and Zuccon (2021, 2022) devise augmentation methods to generate misspelled queries and propose training methods, Typos-aware Training and Self-Teaching (ST), to encourage consistency between outputs of misspelled queries and their non-misspelled counterparts. Alternatively, Sidiropoulos and Kanoulas (2022) apply contrastive loss to enforce representations of misspelled queries to be closer to their corresponding non-misspelled queries. Although these methods can improve the performance of retrieval models for misspelled queries, there is still a substantial performance drop for misspelled queries." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 473, + 525, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 473, + 525, + 513 + ], + "spans": [ + { + "bbox": [ + 302, + 473, + 525, + 513 + ], + "type": "text", + "content": "In this paper, we propose a training method to improve dense retrieval for handling misspelled queries based on the following desired properties:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 515, + 525, + 609 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 302, + 515, + 524, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 515, + 524, + 541 + ], + "spans": [ + { + "bbox": [ + 302, + 515, + 524, + 541 + ], + "type": "text", + "content": "- Alignment: the method should be able to align queries with their corresponding passages." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 542, + 524, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 542, + 524, + 568 + ], + "spans": [ + { + "bbox": [ + 302, + 542, + 524, + 568 + ], + "type": "text", + "content": "- Robustness: the method should be able to align misspelled queries with their pristine queries." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 569, + 525, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 569, + 525, + 609 + ], + "spans": [ + { + "bbox": [ + 302, + 569, + 525, + 609 + ], + "type": "text", + "content": "- Contrast: the method should be able to separate queries that refer to different passages and passages that correspond to different queries." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 302, + 611, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 611, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 611, + 526, + 772 + ], + "type": "text", + "content": "In contrast to the existing methods for handling misspelled queries that only satisfy the Alignment and Robustness properties, our method also aims to satisfy the Contrast property. Increasing the distance between dissimilar queries should help distinguish misspelled queries from other distinct queries. We design the following components for our training method: (i) Dual Self-Teaching (DST) incorporates the ideas of Dual Learning (Xia et al., 2017; Li et al., 2021) and Self-Teaching (Zhuang and Zuccon, 2022) to train robust dense retrieval in a bidirectional manner: passage retrieval and" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 780, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 310, + 791 + ], + "type": "text", + "content": "1106" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 135, + 795, + 459, + 805 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 795, + 459, + 805 + ], + "spans": [ + { + "bbox": [ + 135, + 795, + 459, + 805 + ], + "type": "text", + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 219, + 806, + 374, + 817 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 806, + 374, + 817 + ], + "spans": [ + { + "bbox": [ + 219, + 806, + 374, + 817 + ], + "type": "text", + "content": "Volume 2: Short Papers, pages 1106-1115" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "spans": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "text", + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 290, + 112 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 290, + 112 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 290, + 112 + ], + "type": "text", + "content": "query retrieval. (ii) Query Augmentation generates a numerous number of misspelling variations for each query to supply our training objective." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 113, + 291, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 113, + 291, + 316 + ], + "spans": [ + { + "bbox": [ + 67, + 113, + 291, + 316 + ], + "type": "text", + "content": "Experimental studies were conducted to assess the efficiency of the proposed method in comparison to existing approaches. We conduct experiments based on two different pre-trained language models. We evaluate using two passage retrieval benchmark datasets, a standard one and a specialized one for misspellings robustness evaluation. For each dataset, we measure performance on both misspelled and non-misspelled queries, where the misspelled queries are both generated and real-world queries. The experimental results show that the proposed method outperforms the best existing methods for enhancing the robustness of dense retrieval against misspellings without sacrificing performance for non-misspelled queries." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 78, + 317, + 274, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 317, + 274, + 329 + ], + "spans": [ + { + "bbox": [ + 78, + 317, + 274, + 329 + ], + "type": "text", + "content": "We summarize our contributions as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 331, + 291, + 534 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 68, + 331, + 291, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 331, + 291, + 385 + ], + "spans": [ + { + "bbox": [ + 68, + 331, + 291, + 385 + ], + "type": "text", + "content": "- We propose a novel training method to enhance the robustness of dense retrieval against misspellings by incorporating three desired properties: Alignment, Robustness, and Contrast." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 385, + 291, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 385, + 291, + 465 + ], + "spans": [ + { + "bbox": [ + 68, + 385, + 291, + 465 + ], + "type": "text", + "content": "- We introduce Dual Self-Teaching (DST) which adopts the idea of Dual Learning and Self-Teaching to learn robust representations. In addition, we propose Query Augmentation to generate multiple views of a particular query under different misspelling scenarios." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 466, + 291, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 466, + 291, + 534 + ], + "spans": [ + { + "bbox": [ + 68, + 466, + 291, + 534 + ], + "type": "text", + "content": "- We evaluate our method on misspelled and non-misspelled queries from two passage retrieval datasets. The results show that our method outperforms the previous state-of-the-art methods by a significant margin on misspelled queries." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 67, + 546, + 157, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 546, + 157, + 561 + ], + "spans": [ + { + "bbox": [ + 67, + 546, + 157, + 561 + ], + "type": "text", + "content": "2 Methodology" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 570, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 570, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 570, + 291, + 773 + ], + "type": "text", + "content": "We propose a training pipeline to enhance the dense retrieval capability for handling spelling variations and mistakes in queries. As shown in Figure 1, the training pipeline comprises three steps. (i) Query Augmentation: we augment each query in the training set into multiple misspelled queries using the typo generators provided by Zhuang and Zuccon (2021). (ii) Similarity Score Calculation: we compute similarity score distributions between queries and passages for passage retrieval and query retrieval tasks using in-batch negative queries and passages, with additional hard negative passages. (iii) Dual Self-Teaching Loss Calculation: we compute the DST loss using the similarity score distributions to achieve all three desired properties." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 71, + 429, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 429, + 84 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 429, + 84 + ], + "type": "text", + "content": "2.1 Query Augmentation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 89, + 526, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 89, + 526, + 238 + ], + "spans": [ + { + "bbox": [ + 302, + 89, + 526, + 238 + ], + "type": "text", + "content": "The purpose of this step is to guide the learning with a broad array of possible misspelling patterns. Let " + }, + { + "bbox": [ + 302, + 89, + 526, + 238 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 302, + 89, + 526, + 238 + ], + "type": "text", + "content": " denote a set " + }, + { + "bbox": [ + 302, + 89, + 526, + 238 + ], + "type": "inline_equation", + "content": "\\{q_{1}, q_{2}, \\ldots, q_{N}\\}" + }, + { + "bbox": [ + 302, + 89, + 526, + 238 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 302, + 89, + 526, + 238 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 302, + 89, + 526, + 238 + ], + "type": "text", + "content": " queries. From all queries in " + }, + { + "bbox": [ + 302, + 89, + 526, + 238 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 302, + 89, + 526, + 238 + ], + "type": "text", + "content": ", we generate a set of " + }, + { + "bbox": [ + 302, + 89, + 526, + 238 + ], + "type": "inline_equation", + "content": "K \\times N" + }, + { + "bbox": [ + 302, + 89, + 526, + 238 + ], + "type": "text", + "content": " misspelled queries " + }, + { + "bbox": [ + 302, + 89, + 526, + 238 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}' = \\{\\langle q_{1,k}', q_{2,k}', \\ldots, q_{N,k}'\\rangle\\}_{k=1}^{K}" + }, + { + "bbox": [ + 302, + 89, + 526, + 238 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 302, + 89, + 526, + 238 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 302, + 89, + 526, + 238 + ], + "type": "text", + "content": " is the misspelling variations. We use five typo generators proposed by Zhuang and Zuccon (2021), including: RandInsert, RandDelete, RandSub, SwapNeighbor, and SwapAdjacent. Please refer to Appendix A.2 for examples of the misspelled queries." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 246, + 464, + 259 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 246, + 464, + 259 + ], + "spans": [ + { + "bbox": [ + 302, + 246, + 464, + 259 + ], + "type": "text", + "content": "2.2 Similarity Score Calculation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 264, + 526, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 264, + 526, + 303 + ], + "spans": [ + { + "bbox": [ + 302, + 264, + 526, + 303 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 302, + 264, + 526, + 303 + ], + "type": "inline_equation", + "content": "S(a, \\mathbf{B})" + }, + { + "bbox": [ + 302, + 264, + 526, + 303 + ], + "type": "text", + "content": " denote a function that computes a similarity score distribution of any vector " + }, + { + "bbox": [ + 302, + 264, + 526, + 303 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 302, + 264, + 526, + 303 + ], + "type": "text", + "content": " over any set of vectors " + }, + { + "bbox": [ + 302, + 264, + 526, + 303 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 302, + 264, + 526, + 303 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 311, + 309, + 525, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 309, + 525, + 346 + ], + "spans": [ + { + "bbox": [ + 311, + 309, + 525, + 346 + ], + "type": "interline_equation", + "content": "S (a, \\mathbf {B}) = \\left\\{b _ {i} \\in \\mathbf {B} \\left| \\frac {\\exp (a \\cdot b _ {i})}{\\sum_ {b _ {j} \\in \\mathbf {B}} \\exp (a \\cdot b _ {j})} \\right. \\right\\} \\tag {1}", + "image_path": "fd0e78b1fde269a8c87d2769af594d1426e260944cde0acc7a1262b5f7416b8a.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 350, + 525, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 350, + 525, + 404 + ], + "spans": [ + { + "bbox": [ + 302, + 350, + 525, + 404 + ], + "type": "text", + "content": "Given " + }, + { + "bbox": [ + 302, + 350, + 525, + 404 + ], + "type": "inline_equation", + "content": "\\mathbf{P} = \\{p_1, p_2, \\dots, p_M\\}" + }, + { + "bbox": [ + 302, + 350, + 525, + 404 + ], + "type": "text", + "content": " to be a set of " + }, + { + "bbox": [ + 302, + 350, + 525, + 404 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 302, + 350, + 525, + 404 + ], + "type": "text", + "content": " passages and " + }, + { + "bbox": [ + 302, + 350, + 525, + 404 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_k' = \\{q_{1,k}', q_{2,k}', \\dots, q_{N,k}'\\}" + }, + { + "bbox": [ + 302, + 350, + 525, + 404 + ], + "type": "text", + "content": " to be the " + }, + { + "bbox": [ + 302, + 350, + 525, + 404 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 302, + 350, + 525, + 404 + ], + "type": "text", + "content": " set of misspelled queries in " + }, + { + "bbox": [ + 302, + 350, + 525, + 404 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}'" + }, + { + "bbox": [ + 302, + 350, + 525, + 404 + ], + "type": "text", + "content": ", we compute two groups of score distributions as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 405, + 525, + 513 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 302, + 405, + 525, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 405, + 525, + 460 + ], + "spans": [ + { + "bbox": [ + 302, + 405, + 525, + 460 + ], + "type": "text", + "content": "Passage retrieval: we calculate score distributions in a query-to-passages direction for each original query " + }, + { + "bbox": [ + 302, + 405, + 525, + 460 + ], + "type": "inline_equation", + "content": "s_p = S(q_n, \\mathbf{P})" + }, + { + "bbox": [ + 302, + 405, + 525, + 460 + ], + "type": "text", + "content": " and misspelled query " + }, + { + "bbox": [ + 302, + 405, + 525, + 460 + ], + "type": "inline_equation", + "content": "s_p'^k = S(q_{n,k}', \\mathbf{P})" + }, + { + "bbox": [ + 302, + 405, + 525, + 460 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 460, + 525, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 460, + 525, + 513 + ], + "spans": [ + { + "bbox": [ + 302, + 460, + 525, + 513 + ], + "type": "text", + "content": "- Query retrieval: we calculate score distributions in a passage-to-queries direction for original queries " + }, + { + "bbox": [ + 302, + 460, + 525, + 513 + ], + "type": "inline_equation", + "content": "s_q = S(p_m, \\mathbf{Q})" + }, + { + "bbox": [ + 302, + 460, + 525, + 513 + ], + "type": "text", + "content": " and each set of misspellled queries " + }, + { + "bbox": [ + 302, + 460, + 525, + 513 + ], + "type": "inline_equation", + "content": "s_q^{\\prime k} = S(p_m, \\mathbf{Q}_k^{\\prime})" + }, + { + "bbox": [ + 302, + 460, + 525, + 513 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 302, + 513, + 525, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 513, + 525, + 542 + ], + "spans": [ + { + "bbox": [ + 302, + 513, + 525, + 542 + ], + "type": "text", + "content": "This way, we produce four different score distributions " + }, + { + "bbox": [ + 302, + 513, + 525, + 542 + ], + "type": "inline_equation", + "content": "(s_p,s_p^{\\prime k},s_q,s_q^{\\prime k})" + }, + { + "bbox": [ + 302, + 513, + 525, + 542 + ], + "type": "text", + "content": " for our training objective." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 302, + 550, + 500, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 550, + 500, + 562 + ], + "spans": [ + { + "bbox": [ + 302, + 550, + 500, + 562 + ], + "type": "text", + "content": "2.3 Dual Self-Teaching Loss Calculation" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 302, + 566, + 525, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 566, + 525, + 606 + ], + "spans": [ + { + "bbox": [ + 302, + 566, + 525, + 606 + ], + "type": "text", + "content": "We design the Dual Self-Teaching loss " + }, + { + "bbox": [ + 302, + 566, + 525, + 606 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{DST}})" + }, + { + "bbox": [ + 302, + 566, + 525, + 606 + ], + "type": "text", + "content": " to capture the three desired properties: Alignment, Robustness, and Contrast." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 616, + 525, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 616, + 525, + 647 + ], + "spans": [ + { + "bbox": [ + 315, + 616, + 525, + 647 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {D S T}} = \\underbrace {(1 - \\beta) \\mathcal {L} _ {\\mathrm {D C E}}} _ {\\text {D u a l C r o s s - E n t r o p y}} + \\underbrace {\\beta \\mathcal {L} _ {\\mathrm {D K L}}} _ {\\text {D u a l K L - D i v e r g e n c e}} \\tag {2}", + "image_path": "6794fe4976077851ecc798424dcab56399aaa79afec91278331bd51cf54d4a5a.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 302, + 653, + 525, + 722 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 653, + 525, + 722 + ], + "spans": [ + { + "bbox": [ + 302, + 653, + 525, + 722 + ], + "type": "text", + "content": "Dual Cross-Entropy loss " + }, + { + "bbox": [ + 302, + 653, + 525, + 722 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{DCE}})" + }, + { + "bbox": [ + 302, + 653, + 525, + 722 + ], + "type": "text", + "content": " satisfies the Alignment and Contrast properties by utilizing cross-entropy losses to learn score distributions of the original queries for passage retrieval " + }, + { + "bbox": [ + 302, + 653, + 525, + 722 + ], + "type": "inline_equation", + "content": "(s_p)" + }, + { + "bbox": [ + 302, + 653, + 525, + 722 + ], + "type": "text", + "content": " and query retrieval " + }, + { + "bbox": [ + 302, + 653, + 525, + 722 + ], + "type": "inline_equation", + "content": "(s_q)" + }, + { + "bbox": [ + 302, + 653, + 525, + 722 + ], + "type": "text", + "content": " given labels " + }, + { + "bbox": [ + 302, + 653, + 525, + 722 + ], + "type": "inline_equation", + "content": "y_{p}" + }, + { + "bbox": [ + 302, + 653, + 525, + 722 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 302, + 653, + 525, + 722 + ], + "type": "inline_equation", + "content": "y_{q}" + }, + { + "bbox": [ + 302, + 653, + 525, + 722 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 726, + 525, + 772 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 726, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 312, + 726, + 525, + 772 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {D C E}} = \\underbrace {(1 - \\gamma) \\mathcal {L} _ {\\mathrm {C E}} ^ {(P)} \\left(s _ {p} , y _ {p}\\right)} _ {\\text {P a s s a g e R e t i v e a l}} + \\underbrace {\\gamma \\mathcal {L} _ {\\mathrm {C E}} ^ {(Q)} \\left(s _ {q} , y _ {q}\\right)} _ {\\text {Q u e r y R e t i v e a l}} \\tag {3}", + "image_path": "0af589d7d95c993c3ef77edd0c2f07383aade9384cffeb43277bf412bf166613.jpg" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "1107" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 126, + 68, + 468, + 213 + ], + "blocks": [ + { + "bbox": [ + 126, + 68, + 468, + 213 + ], + "lines": [ + { + "bbox": [ + 126, + 68, + 468, + 213 + ], + "spans": [ + { + "bbox": [ + 126, + 68, + 468, + 213 + ], + "type": "image", + "image_path": "9cb707006d2ab480650319b69a87a6bda5748b990dc59544a681e1c57616d9ac.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 221, + 525, + 246 + ], + "lines": [ + { + "bbox": [ + 67, + 221, + 525, + 246 + ], + "spans": [ + { + "bbox": [ + 67, + 221, + 525, + 246 + ], + "type": "text", + "content": "Figure 1: The proposed training pipeline consists of three steps: (a) Query Augmentation, (b) Similarity Score Calculation, and (c) Dual Self-Teaching Loss Calculation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 256, + 291, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 256, + 291, + 462 + ], + "spans": [ + { + "bbox": [ + 67, + 256, + 291, + 462 + ], + "type": "text", + "content": "Minimizing the " + }, + { + "bbox": [ + 67, + 256, + 291, + 462 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CE}}^{(P)}" + }, + { + "bbox": [ + 67, + 256, + 291, + 462 + ], + "type": "text", + "content": " term will increase the similarity scores between queries and their relevant passages to be higher than other irrelevant passages by separating the relevant and irrelevant passages from one another. Minimizing the " + }, + { + "bbox": [ + 67, + 256, + 291, + 462 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CE}}^{(Q)}" + }, + { + "bbox": [ + 67, + 256, + 291, + 462 + ], + "type": "text", + "content": " term will increase the similarity scores between passages and their relevant queries to be higher than other irrelevant queries by separating the relevant and irrelevant queries from one another. In this manner, minimizing one of the two terms will align queries with their corresponding passages, satisfying the Alignment property. Moreover, minimizing both terms will separate queries that refer to different passages and passages that belong to different queries, satisfying the Contrast property." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 464, + 291, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 464, + 291, + 533 + ], + "spans": [ + { + "bbox": [ + 67, + 464, + 291, + 533 + ], + "type": "text", + "content": "Dual KL-Divergence loss " + }, + { + "bbox": [ + 67, + 464, + 291, + 533 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{DKL}})" + }, + { + "bbox": [ + 67, + 464, + 291, + 533 + ], + "type": "text", + "content": " aims to fulfill the Robustness property by using KL losses to match score distributions of misspelled queries " + }, + { + "bbox": [ + 67, + 464, + 291, + 533 + ], + "type": "inline_equation", + "content": "\\{s_p^{\\prime 1}, s_p^{\\prime 2}, \\ldots, s_p^{\\prime K}\\}" + }, + { + "bbox": [ + 67, + 464, + 291, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 464, + 291, + 533 + ], + "type": "inline_equation", + "content": "\\{s_q^{\\prime 1}, s_q^{\\prime 2}, \\ldots, s_q^{\\prime K}\\}" + }, + { + "bbox": [ + 67, + 464, + 291, + 533 + ], + "type": "text", + "content": " to the score distributions of the original query " + }, + { + "bbox": [ + 67, + 464, + 291, + 533 + ], + "type": "inline_equation", + "content": "s_p" + }, + { + "bbox": [ + 67, + 464, + 291, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 464, + 291, + 533 + ], + "type": "inline_equation", + "content": "s_q" + }, + { + "bbox": [ + 67, + 464, + 291, + 533 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 93, + 542, + 291, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 542, + 291, + 622 + ], + "spans": [ + { + "bbox": [ + 93, + 542, + 291, + 622 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {D K L}} = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\underbrace {(1 - \\sigma) \\mathcal {L} _ {\\mathrm {K L}} ^ {(P)} \\left(s _ {p} ^ {\\prime k} , s _ {p}\\right)} _ {\\text {P a s s a g e R e t r i e v a l C o n s i s t e n c y}} \\tag {4} \\\\ + \\underbrace {\\sigma \\mathcal {L} _ {\\mathrm {K L}} ^ {(Q)} \\left(s _ {q} ^ {\\prime k} , s _ {q}\\right)} _ {\\text {Q u e r y R e t r i e v a l C o n s i s t e n c y}} \\\\ \\end{array}", + "image_path": "534a704a12b1a17ef3adbc4465b7ce0081cb2f0a41e01113b1ead01d30a32c33.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 634, + 292, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 634, + 292, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 634, + 292, + 773 + ], + "type": "text", + "content": "Minimizing " + }, + { + "bbox": [ + 67, + 634, + 292, + 773 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{KL}}^{(P)}" + }, + { + "bbox": [ + 67, + 634, + 292, + 773 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 634, + 292, + 773 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{KL}}^{(Q)}" + }, + { + "bbox": [ + 67, + 634, + 292, + 773 + ], + "type": "text", + "content": " will reduce the discrepancy between misspelled and non-misspelled queries for both query-to-passages and passage-to-queries score distributions. This way, we implicitly align representations of the misspelled queries to the original queries, satisfying the Robustness property. To stabilize training, we apply stop-gradient to the score distributions of the original queries " + }, + { + "bbox": [ + 67, + 634, + 292, + 773 + ], + "type": "inline_equation", + "content": "(s_p" + }, + { + "bbox": [ + 67, + 634, + 292, + 773 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 634, + 292, + 773 + ], + "type": "inline_equation", + "content": "s_q)" + }, + { + "bbox": [ + 67, + 634, + 292, + 773 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 67, + 634, + 292, + 773 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{DKL}}" + }, + { + "bbox": [ + 67, + 634, + 292, + 773 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 67, + 634, + 292, + 773 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 67, + 634, + 292, + 773 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 67, + 634, + 292, + 773 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 67, + 634, + 292, + 773 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 67, + 634, + 292, + 773 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 67, + 634, + 292, + 773 + ], + "type": "text", + "content": " are the balancing coefficients selected by hyper-parameter tuning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 253, + 526, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 253, + 526, + 280 + ], + "spans": [ + { + "bbox": [ + 302, + 253, + 526, + 280 + ], + "type": "text", + "content": "on a development set. With this loss combination, we achieve all three desired properties." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 290, + 439, + 305 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 290, + 439, + 305 + ], + "spans": [ + { + "bbox": [ + 302, + 290, + 439, + 305 + ], + "type": "text", + "content": "3 Experimental Settings" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 313, + 407, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 313, + 407, + 327 + ], + "spans": [ + { + "bbox": [ + 302, + 313, + 407, + 327 + ], + "type": "text", + "content": "3.1 Training Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 332, + 526, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 332, + 526, + 454 + ], + "spans": [ + { + "bbox": [ + 302, + 332, + 526, + 454 + ], + "type": "text", + "content": "We experiment on two pre-trained language models, BERT (Devlin et al., 2019) and Character-BERT (El Boukkouri et al., 2020). We train models only on the training set of MS MARCO dataset (Nguyen et al., 2016). Moreover, the training data provided by the Tevatron toolkit (Gao et al., 2022) also contains hard negative passages. We include the training set details and hyper-parameter settings in Appendix A.1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 465, + 431, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 465, + 431, + 478 + ], + "spans": [ + { + "bbox": [ + 302, + 465, + 431, + 478 + ], + "type": "text", + "content": "3.2 Competitive Methods" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 483, + 525, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 483, + 525, + 523 + ], + "spans": [ + { + "bbox": [ + 302, + 483, + 525, + 523 + ], + "type": "text", + "content": "To show the effectiveness of our method, we compare our work with the following baseline and competitive training methods." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 524, + 525, + 719 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 302, + 524, + 525, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 524, + 525, + 566 + ], + "spans": [ + { + "bbox": [ + 302, + 524, + 525, + 566 + ], + "type": "text", + "content": "- DPR (Karpukhin et al., 2020) is a baseline training method that trains dense retrieval merely on non-misspelled queries using " + }, + { + "bbox": [ + 302, + 524, + 525, + 566 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CE}}^{(P)}" + }, + { + "bbox": [ + 302, + 524, + 525, + 566 + ], + "type": "text", + "content": " loss." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 566, + 524, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 566, + 524, + 622 + ], + "spans": [ + { + "bbox": [ + 302, + 566, + 524, + 622 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 302, + 566, + 524, + 622 + ], + "type": "inline_equation", + "content": "DPR + Aug" + }, + { + "bbox": [ + 302, + 566, + 524, + 622 + ], + "type": "text", + "content": " (Zhuang and Zuccon, 2021) is the Typos-aware Training method which trains dense retrieval on both misspelled and non-misspelled queries using " + }, + { + "bbox": [ + 302, + 566, + 524, + 622 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CE}}^{(P)}" + }, + { + "bbox": [ + 302, + 566, + 524, + 622 + ], + "type": "text", + "content": " loss." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 622, + 524, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 622, + 524, + 661 + ], + "spans": [ + { + "bbox": [ + 302, + 622, + 524, + 661 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 302, + 622, + 524, + 661 + ], + "type": "inline_equation", + "content": "DPR + Aug + CL" + }, + { + "bbox": [ + 302, + 622, + 524, + 661 + ], + "type": "text", + "content": " (Sidiropoulos and Kanoulas, 2022) employs additional contrastive loss to train the misspelled queries." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 663, + 524, + 719 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 663, + 524, + 719 + ], + "spans": [ + { + "bbox": [ + 302, + 663, + 524, + 719 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 302, + 663, + 524, + 719 + ], + "type": "inline_equation", + "content": "DPR + ST" + }, + { + "bbox": [ + 302, + 663, + 524, + 719 + ], + "type": "text", + "content": " (Zhuang and Zuccon, 2022) is the Self-Teaching method that trains dense retrieval on both misspelled and non-misspelled queries using " + }, + { + "bbox": [ + 302, + 663, + 524, + 719 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CE}}^{(P)}" + }, + { + "bbox": [ + 302, + 663, + 524, + 719 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 302, + 663, + 524, + 719 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{KL}}^{(P)}" + }, + { + "bbox": [ + 302, + 663, + 524, + 719 + ], + "type": "text", + "content": " losses." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 302, + 719, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 719, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 719, + 525, + 772 + ], + "type": "text", + "content": "Note that their query augmentation method is identical to the Query Augmentation with " + }, + { + "bbox": [ + 302, + 719, + 525, + 772 + ], + "type": "inline_equation", + "content": "K = 1" + }, + { + "bbox": [ + 302, + 719, + 525, + 772 + ], + "type": "text", + "content": ". We retrain all models using the same setting described in the previous section." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "1108" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 72, + 68, + 526, + 169 + ], + "blocks": [ + { + "bbox": [ + 72, + 68, + 526, + 169 + ], + "lines": [ + { + "bbox": [ + 72, + 68, + 526, + 169 + ], + "spans": [ + { + "bbox": [ + 72, + 68, + 526, + 169 + ], + "type": "table", + "html": "
MethodsBERT-basedCharacterBERT-based
MS MARCODL-typoMS MARCODL-typo
MRR@10R@1000nDCG@10MRRMAPMRR@10R@1000nDCG@10MRRMAP
DPR.143 (.331).696 (.954).276 (.682).431 (.873).175 (.563).162 (.321).726 (.945).268 (.643).376 (.832).212 (.503)
+ Aug.227 (.334).857 (.950).398 (.682).530 (.806).286 (.565).258 (.326).883 (.946).414 (.631).578 (.783).318 (.512)
+ Aug + CL.234 (.335).867 (.951).387 (.668).536 (.864).267 (.544).263 (.330).894 (.947).466 (.677).635 (.819).360 (.544)
+ ST.237 (.333).874 (.950).392 (.677).525 (.852).283 (.557).274 (.332).900 (.947).469 (.650).619 (.810).359 (.517)
+ DST (our).260† (.336).894† (.954).432 (.673).558 (.833).343† (.568).288† (.332).918† (.949).529† (.673).742† (.854).403 (.537)
", + "image_path": "d402ae89676c46cc2d674a2707b32f22a0776882a28eee8ea829c2885179e05f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 237, + 205, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 237, + 205, + 248 + ], + "spans": [ + { + "bbox": [ + 67, + 237, + 205, + 248 + ], + "type": "text", + "content": "3.3 Dataset and Evaluation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 259, + 291, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 259, + 291, + 435 + ], + "spans": [ + { + "bbox": [ + 67, + 259, + 291, + 435 + ], + "type": "text", + "content": "Datasets. We evaluate the effectiveness of DST on two passage retrieval datasets, MS MARCO and DL-typo (Zhuang and Zuccon, 2022), each with misspelled and non-misspelled queries. There are 8.8 million candidate passages for both datasets. The development set of MS MARCO contains 6,980 non-misspelled queries. To obtain misspelled queries, we use the typos generator method proposed by Zhuang and Zuccon (2021) to generate 10 misspelled variations for each original query. The DL-typo provides 60 real-world misspelled queries and 60 corresponding non-misspelled queries that are corrected manually." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 437, + 291, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 437, + 291, + 547 + ], + "spans": [ + { + "bbox": [ + 67, + 437, + 291, + 547 + ], + "type": "text", + "content": "Evaluation. We use the standard metrics originally used by each dataset's creators. For MS MARCO, each misspelled query performance is the average of 10 measurements. We employ Ranx evaluation library (Bassani, 2022) to measure performance and statistical significance. Specifically, we use a two-tailed paired t-test with Bonferroni correction to measure the statistical significance " + }, + { + "bbox": [ + 67, + 437, + 291, + 547 + ], + "type": "inline_equation", + "content": "(p < 0.05)" + }, + { + "bbox": [ + 67, + 437, + 291, + 547 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 562, + 199, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 562, + 199, + 577 + ], + "spans": [ + { + "bbox": [ + 67, + 562, + 199, + 577 + ], + "type": "text", + "content": "4 Experimental Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 589, + 158, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 589, + 158, + 601 + ], + "spans": [ + { + "bbox": [ + 67, + 589, + 158, + 601 + ], + "type": "text", + "content": "4.1 Main Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 611, + 292, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 611, + 292, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 611, + 292, + 773 + ], + "type": "text", + "content": "As shown in Table 1, the results indicate that DST outperforms competitive methods for misspelled queries in every case without sacrificing performance for non-misspelled queries in eight out of ten cases. We observe some performance trade-offs for the BERT-based model in the DL-typo dataset's non-misspelling scores (nDCG@10 and MRR). Aside from that, there is no performance trade-off for the CharacterBERT-based model. These outcomes conform with the observation in Figure 2 (Section 4.4) that DST improves the Robustness and Contrast of misspelled queries." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 237, + 481, + 250 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 237, + 481, + 250 + ], + "spans": [ + { + "bbox": [ + 302, + 237, + 481, + 250 + ], + "type": "text", + "content": "4.2 Query Augmentation Size Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 301, + 254, + 526, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 254, + 526, + 403 + ], + "spans": [ + { + "bbox": [ + 301, + 254, + 526, + 403 + ], + "type": "text", + "content": "To study the benefit of query augmentation and find the optimal augmentation size, we measure the performance of BERT-based dense retrieval models trained with DST using the query augmentation size " + }, + { + "bbox": [ + 301, + 254, + 526, + 403 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 301, + 254, + 526, + 403 + ], + "type": "text", + "content": " of 1, 10, 20, 40, and 60. Note that the query augmentation method used in previous works is a special case of Query Augmentation when " + }, + { + "bbox": [ + 301, + 254, + 526, + 403 + ], + "type": "inline_equation", + "content": "K = 1" + }, + { + "bbox": [ + 301, + 254, + 526, + 403 + ], + "type": "text", + "content": ". We report the results using MRR@10 for the development set of the MS MARCO dataset. We also report training time to show trade-offs between performance and computation." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 314, + 412, + 516, + 481 + ], + "blocks": [ + { + "bbox": [ + 67, + 175, + 526, + 225 + ], + "lines": [ + { + "bbox": [ + 67, + 175, + 526, + 225 + ], + "spans": [ + { + "bbox": [ + 67, + 175, + 526, + 225 + ], + "type": "text", + "content": "Table 1: Results of different training methods on misspelled and non-misspelled queries. We report the results in the format of \"misspelled query performance (non-misspelled query performance)\". We emphasize the best score with bold text and the second-best score with underlined text. We use " + }, + { + "bbox": [ + 67, + 175, + 526, + 225 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 67, + 175, + 526, + 225 + ], + "type": "text", + "content": " to denote DST results that significantly outperform the second-best result (" + }, + { + "bbox": [ + 67, + 175, + 526, + 225 + ], + "type": "inline_equation", + "content": "p < 0.05" + }, + { + "bbox": [ + 67, + 175, + 526, + 225 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 412, + 516, + 481 + ], + "lines": [ + { + "bbox": [ + 314, + 412, + 516, + 481 + ], + "spans": [ + { + "bbox": [ + 314, + 412, + 516, + 481 + ], + "type": "table", + "html": "
QueriesK
110204060
Original.334.334.335.336.332
Misspelled.251.258.260.260.260
Training time (hr)1820233139
", + "image_path": "65c27de57f78c38baea3116a028eb20cd2ef9fe80691c5d28c1a2f515b49baae.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 488, + 525, + 513 + ], + "lines": [ + { + "bbox": [ + 302, + 488, + 525, + 513 + ], + "spans": [ + { + "bbox": [ + 302, + 488, + 525, + 513 + ], + "type": "text", + "content": "Table 2: Results of query augmentation size study. We train all models in this experiment on a V100 32G GPU." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 302, + 525, + 525, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 525, + 525, + 620 + ], + "spans": [ + { + "bbox": [ + 302, + 525, + 525, + 620 + ], + "type": "text", + "content": "As shown in Table 2, the results indicate that increasing " + }, + { + "bbox": [ + 302, + 525, + 525, + 620 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 302, + 525, + 525, + 620 + ], + "type": "text", + "content": " improves the performance of both misspelled and non-misspelled queries, but only up to a certain point, after which the performance begins to decline. We observe that setting " + }, + { + "bbox": [ + 302, + 525, + 525, + 620 + ], + "type": "inline_equation", + "content": "K = 40" + }, + { + "bbox": [ + 302, + 525, + 525, + 620 + ], + "type": "text", + "content": " produces the best results, and there is no further performance improvement after this point." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 629, + 426, + 642 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 629, + 426, + 642 + ], + "spans": [ + { + "bbox": [ + 302, + 629, + 426, + 642 + ], + "type": "text", + "content": "4.3 Loss Ablation Study" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 646, + 526, + 774 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 646, + 526, + 774 + ], + "spans": [ + { + "bbox": [ + 302, + 646, + 526, + 774 + ], + "type": "text", + "content": "In this experiment, we study the benefit of each term in DST by training BERT-based dense retrieval models on variant loss combinations with " + }, + { + "bbox": [ + 302, + 646, + 526, + 774 + ], + "type": "inline_equation", + "content": "K = 40" + }, + { + "bbox": [ + 302, + 646, + 526, + 774 + ], + "type": "text", + "content": ". The results in Table 3 reveal that " + }, + { + "bbox": [ + 302, + 646, + 526, + 774 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{KL}}^{(P)}" + }, + { + "bbox": [ + 302, + 646, + 526, + 774 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 302, + 646, + 526, + 774 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{KL}}^{(Q)}" + }, + { + "bbox": [ + 302, + 646, + 526, + 774 + ], + "type": "text", + "content": " terms positively contribute to the performance of misspelled and non-misspelled queries, with the " + }, + { + "bbox": [ + 302, + 646, + 526, + 774 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{KL}}^{(P)}" + }, + { + "bbox": [ + 302, + 646, + 526, + 774 + ], + "type": "text", + "content": " being more significant. The " + }, + { + "bbox": [ + 302, + 646, + 526, + 774 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CE}}^{(P)}" + }, + { + "bbox": [ + 302, + 646, + 526, + 774 + ], + "type": "text", + "content": " term is crucial for retrieval performance, whereas the " + }, + { + "bbox": [ + 302, + 646, + 526, + 774 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CE}}^{(Q)}" + }, + { + "bbox": [ + 302, + 646, + 526, + 774 + ], + "type": "text", + "content": " term indirectly improves the performance" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "1109" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 95, + 68, + 251, + 169 + ], + "blocks": [ + { + "bbox": [ + 95, + 68, + 251, + 169 + ], + "lines": [ + { + "bbox": [ + 95, + 68, + 251, + 169 + ], + "spans": [ + { + "bbox": [ + 95, + 68, + 251, + 169 + ], + "type": "table", + "html": "
L(PCE)L(QCE)L(PKL)L(QKL)MRR@10
✓✓✓✓.260 (.336)
✓✓✓.257 (.335)
✓✓✓.228 (.326)
✓✓✓.251 (.337)
✓✓✓.087 (.114)
✓✓.249 (.336)
✓✓.120 (.158)
", + "image_path": "a151b538d7850d6f9ac41fcb45f25b6288f560992290463285cfd1dd193618f6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 176, + 287, + 188 + ], + "lines": [ + { + "bbox": [ + 69, + 176, + 287, + 188 + ], + "spans": [ + { + "bbox": [ + 69, + 176, + 287, + 188 + ], + "type": "text", + "content": "Table 3: Loss ablation study results on MS MARCO." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 204, + 290, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 204, + 290, + 288 + ], + "spans": [ + { + "bbox": [ + 67, + 204, + 290, + 288 + ], + "type": "text", + "content": "of misspelled queries by separating their pristine queries from the surrounding queries. Disabling query retrieval terms " + }, + { + "bbox": [ + 67, + 204, + 290, + 288 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{CE}}^{(Q)})" + }, + { + "bbox": [ + 67, + 204, + 290, + 288 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 204, + 290, + 288 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{KL}}^{(Q)}" + }, + { + "bbox": [ + 67, + 204, + 290, + 288 + ], + "type": "text", + "content": " greatly reduces performances for misspelled queries. The passage retrieval terms " + }, + { + "bbox": [ + 67, + 204, + 290, + 288 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{CE}}^{(P)})" + }, + { + "bbox": [ + 67, + 204, + 290, + 288 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 204, + 290, + 288 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{KL}}^{(P)}" + }, + { + "bbox": [ + 67, + 204, + 290, + 288 + ], + "type": "text", + "content": " are indispensable and cannot be substituted." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 303, + 189, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 303, + 189, + 316 + ], + "spans": [ + { + "bbox": [ + 67, + 303, + 189, + 316 + ], + "type": "text", + "content": "4.4 Query Distributions" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 323, + 290, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 323, + 290, + 417 + ], + "spans": [ + { + "bbox": [ + 67, + 323, + 290, + 417 + ], + "type": "text", + "content": "The purpose of this section is to study the impact of our training method on the Robustness and Contrast of misspelled queries. We also compare our method against the baseline and competitive methods to show its effectiveness. The Robustness and Contrast of misspelled queries are illustrated using the following kernel density graphs:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 420, + 291, + 597 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 67, + 420, + 290, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 420, + 290, + 446 + ], + "spans": [ + { + "bbox": [ + 67, + 420, + 290, + 446 + ], + "type": "text", + "content": "- Original-to-Misspell: the cosine similarity distribution between original and misspelled queries." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 447, + 291, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 447, + 291, + 597 + ], + "spans": [ + { + "bbox": [ + 67, + 447, + 291, + 597 + ], + "type": "text", + "content": "- Original-to-Neighbor: the cosine similarity distribution between original and neighbor queries. The Robustness property is emphasized by the Original-to-Misspell distribution having high cosine similarity. On the other hand, the Contrast property is emphasized by the small overlapping between Original-to-Misspell and Original-to-Neighbor distributions. The results in Figure 2 show that our method (c) produces the best Robustness and Contrast properties for misspelled queries in comparison to other methods." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 67, + 613, + 145, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 613, + 145, + 624 + ], + "spans": [ + { + "bbox": [ + 67, + 613, + 145, + 624 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 638, + 290, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 638, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 638, + 290, + 772 + ], + "type": "text", + "content": "This paper aims to address the misspelling problem in dense retrieval. We formulate three desired properties for making dense retrieval robust to misspellings: Alignment, Robustness, and Contrast. Unlike previous methods, which only focus on the Alignment and Robustness properties, our method considers all the desired properties. The empirical results show that our method performs best against misspelled queries, revealing the importance of the Contrast property for handling misspellings." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 315, + 68, + 511, + 218 + ], + "blocks": [ + { + "bbox": [ + 315, + 68, + 511, + 218 + ], + "lines": [ + { + "bbox": [ + 315, + 68, + 511, + 218 + ], + "spans": [ + { + "bbox": [ + 315, + 68, + 511, + 218 + ], + "type": "image", + "image_path": "d5b2d927d187e10c36b3de26eb7003e3a424cda229614d3b30ef44fc48c48a7d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 357, + 224, + 480, + 235 + ], + "lines": [ + { + "bbox": [ + 357, + 224, + 480, + 235 + ], + "spans": [ + { + "bbox": [ + 357, + 224, + 480, + 235 + ], + "type": "text", + "content": "(a) DPR (Karpukhin et al., 2020)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 314, + 242, + 511, + 393 + ], + "blocks": [ + { + "bbox": [ + 314, + 242, + 511, + 393 + ], + "lines": [ + { + "bbox": [ + 314, + 242, + 511, + 393 + ], + "spans": [ + { + "bbox": [ + 314, + 242, + 511, + 393 + ], + "type": "image", + "image_path": "4be0a7c423333aa8d7cbdccb802eb94c004878c415d2f6917ecdc767a5e037b9.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 399, + 503, + 410 + ], + "lines": [ + { + "bbox": [ + 333, + 399, + 503, + 410 + ], + "spans": [ + { + "bbox": [ + 333, + 399, + 503, + 410 + ], + "type": "text", + "content": "(b) Self-Teaching (Zhuang and Zuccon, 2022)." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 314, + 417, + 511, + 567 + ], + "blocks": [ + { + "bbox": [ + 314, + 417, + 511, + 567 + ], + "lines": [ + { + "bbox": [ + 314, + 417, + 511, + 567 + ], + "spans": [ + { + "bbox": [ + 314, + 417, + 511, + 567 + ], + "type": "image", + "image_path": "497082c56023ac179644de161784f0c8592c628036e5c3c6e31863042aada675.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 365, + 574, + 471, + 584 + ], + "lines": [ + { + "bbox": [ + 365, + 574, + 471, + 584 + ], + "spans": [ + { + "bbox": [ + 365, + 574, + 471, + 584 + ], + "type": "text", + "content": "(c) Dual Self-Teaching (our)." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 302, + 594, + 525, + 630 + ], + "lines": [ + { + "bbox": [ + 302, + 594, + 525, + 630 + ], + "spans": [ + { + "bbox": [ + 302, + 594, + 525, + 630 + ], + "type": "text", + "content": "Figure 2: Kernel density of Original-to-Neighbor (orange) and Original-to-Misspell (blue) of different training methods." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 303, + 640, + 383, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 640, + 383, + 651 + ], + "spans": [ + { + "bbox": [ + 303, + 640, + 383, + 651 + ], + "type": "text", + "content": "6 Limitations" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 302, + 664, + 505, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 664, + 505, + 676 + ], + "spans": [ + { + "bbox": [ + 302, + 664, + 505, + 676 + ], + "type": "text", + "content": "We list the limitations of our work as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 302, + 678, + 525, + 772 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 302, + 678, + 525, + 719 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 678, + 525, + 719 + ], + "spans": [ + { + "bbox": [ + 302, + 678, + 525, + 719 + ], + "type": "text", + "content": "- The Query Augmentation is designed for the English alphabet; therefore, other languages with different alphabets will require further work." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 303, + 719, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 719, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 303, + 719, + 525, + 772 + ], + "type": "text", + "content": "- Since the training strategy relies on fine-tuning a pre-trained language model using a large passage retrieval dataset, it may not be suitable for languages with limited resources" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "1110" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 71, + 127, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 71, + 127, + 83 + ], + "spans": [ + { + "bbox": [ + 69, + 71, + 127, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 90, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 69, + 90, + 291, + 136 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 90, + 291, + 136 + ], + "spans": [ + { + "bbox": [ + 69, + 90, + 291, + 136 + ], + "type": "text", + "content": "Elias Bassani. 2022. ranx: A blazing-fast python library for ranking evaluation and comparison. In ECIR (2), volume 13186 of Lecture Notes in Computer Science, pages 259-264. Springer." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 144, + 291, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 144, + 291, + 243 + ], + "spans": [ + { + "bbox": [ + 69, + 144, + 291, + 243 + ], + "type": "text", + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 252, + 291, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 252, + 291, + 350 + ], + "spans": [ + { + "bbox": [ + 69, + 252, + 291, + 350 + ], + "type": "text", + "content": "Hicham El Boukkouri, Olivier Ferret, Thomas Lavergne, Hiroshi Noji, Pierre Zweigenbaum, and Jun'ichi Tsujii. 2020. CharacterBERT: Reconciling ELMo and BERT for word-level open-vocabulary representations from characters. In Proceedings of the 28th International Conference on Computational Linguistics, pages 6903-6915, Barcelona, Spain (Online). International Committee on Computational Linguistics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 360, + 290, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 360, + 290, + 394 + ], + "spans": [ + { + "bbox": [ + 69, + 360, + 290, + 394 + ], + "type": "text", + "content": "Luyu Gao, Xueguang Ma, Jimmy J. Lin, and Jamie Callan. 2022. Tevatron: An efficient and flexible toolkit for dense retrieval. ArXiv, abs/2203.05765." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 402, + 291, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 402, + 291, + 481 + ], + "spans": [ + { + "bbox": [ + 69, + 402, + 291, + 481 + ], + "type": "text", + "content": "Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6769-6781, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 489, + 291, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 489, + 291, + 566 + ], + "spans": [ + { + "bbox": [ + 69, + 489, + 291, + 566 + ], + "type": "text", + "content": "Omar Khattab and Matei Zaharia. 2020. Colbert: Efficient and effective passage search via contextualized late interaction over bert. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '20, page 39-48, New York, NY, USA. Association for Computing Machinery." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 576, + 291, + 653 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 576, + 291, + 653 + ], + "spans": [ + { + "bbox": [ + 69, + 576, + 291, + 653 + ], + "type": "text", + "content": "Yizhi Li, Zhenghao Liu, Chenyan Xiong, and Zhiyuan Liu. 2021. More robust dense retrieval with contrastive dual learning. In Proceedings of the 2021 ACM SIGIR International Conference on Theory of Information Retrieval, ICTIR '21, page 287-296, New York, NY, USA. Association for Computing Machinery." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 661, + 291, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 661, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 661, + 291, + 772 + ], + "type": "text", + "content": "Tri Nguyen, Mir Rosenberg, Xia Song, Jianfeng Gao, Saurabh Tiwary, Rangan Majumder, and Li Deng. 2016. MS MARCO: A human generated machine reading comprehension dataset. In Proceedings of the Workshop on Cognitive Computation: Integrating neural and symbolic approaches 2016 co-located with the 30th Annual Conference on Neural Information Processing Systems (NIPS 2016), Barcelona, Spain, December 9, 2016, volume 1773 of CEUR Workshop Proceedings. CEUR-WS.org." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 526, + 772 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 304, + 72, + 526, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 526, + 127 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 526, + 127 + ], + "type": "text", + "content": "Gustavo Penha, Arthur Camara, and Claudia Hauff. 2022. Evaluating the robustness of retrieval pipelines with query variation generators. In Advances in Information Retrieval, pages 397-412, Cham. Springer International Publishing." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 139, + 526, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 139, + 526, + 238 + ], + "spans": [ + { + "bbox": [ + 304, + 139, + 526, + 238 + ], + "type": "text", + "content": "Yingqi Qu, Yuchen Ding, Jing Liu, Kai Liu, Ruiyang Ren, Wayne Xin Zhao, Daxiang Dong, Hua Wu, and Haifeng Wang. 2021. RocketQA: An optimized training approach to dense passage retrieval for open-domain question answering. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5835-5847, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 250, + 526, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 250, + 526, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 250, + 526, + 338 + ], + "type": "text", + "content": "Ruiyang Ren, Shangwen Lv, Yingqi Qu, Jing Liu, Wayne Xin Zhao, QiaoQiao She, Hua Wu, Haifeng Wang, and Ji-Rong Wen. 2021a. PAIR: Leveraging passage-centric similarity relation for improving dense passage retrieval. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 2173-2183, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 349, + 526, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 349, + 526, + 439 + ], + "spans": [ + { + "bbox": [ + 304, + 349, + 526, + 439 + ], + "type": "text", + "content": "Ruiyang Ren, Yingqi Qu, Jing Liu, Wayne Xin Zhao, QiaoQiao She, Hua Wu, Haifeng Wang, and Ji-Rong Wen. 2021b. RocketQAv2: A joint training method for dense passage retrieval and passage re-ranking. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 2825-2835, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 449, + 526, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 526, + 515 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 526, + 515 + ], + "type": "text", + "content": "Georgios Sidiropoulos and Evangelos Kanoulas. 2022. Analysing the robustness of dual encoders for dense retrieval against misspellings. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval. ACM." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 528, + 526, + 660 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 528, + 526, + 660 + ], + "spans": [ + { + "bbox": [ + 304, + 528, + 526, + 660 + ], + "type": "text", + "content": "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander M. Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 671, + 526, + 737 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 671, + 526, + 737 + ], + "spans": [ + { + "bbox": [ + 304, + 671, + 526, + 737 + ], + "type": "text", + "content": "Yingce Xia, Tao Qin, Wei Chen, Jiang Bian, Nenghai Yu, and Tie-Yan Liu. 2017. Dual supervised learning. In Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 3789-3798. PMLR." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 749, + 526, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 749, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 749, + 526, + 772 + ], + "type": "text", + "content": "Lee Xiong, Chenyan Xiong, Ye Li, Kwok-Fung Tang, Jialin Liu, Paul N. Bennett, Junaid Ahmed, and" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1111" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 341 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 80, + 72, + 291, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 72, + 291, + 116 + ], + "spans": [ + { + "bbox": [ + 80, + 72, + 291, + 116 + ], + "type": "text", + "content": "Arnold Overwijk. 2021. Approximate nearest neighbor negative contrastive learning for dense text retrieval. In International Conference on Learning Representations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 125, + 291, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 125, + 291, + 169 + ], + "spans": [ + { + "bbox": [ + 69, + 125, + 291, + 169 + ], + "type": "text", + "content": "Jingtao Zhan, Jiaxin Mao, Yiqun Liu, Min Zhang, and Shaoping Ma. 2020. Repbert: Contextualized text embeddings for first-stage retrieval. CoRR, abs/2006.15498." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 178, + 291, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 178, + 291, + 255 + ], + "spans": [ + { + "bbox": [ + 69, + 178, + 291, + 255 + ], + "type": "text", + "content": "Shengyao Zhuang and Guido Zuccon. 2021. Dealing with typos for BERT-based passage retrieval and ranking. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 2836-2842, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 263, + 291, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 263, + 291, + 341 + ], + "spans": [ + { + "bbox": [ + 69, + 263, + 291, + 341 + ], + "type": "text", + "content": "Shengyao Zhuang and Guido Zuccon. 2022. Character-bert and self-teaching for improving the robustness of dense retrievers on queries with typos. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '22, page 1444-1454, New York, NY, USA. Association for Computing Machinery." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1112" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 71, + 141, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 71, + 141, + 84 + ], + "spans": [ + { + "bbox": [ + 68, + 71, + 141, + 84 + ], + "type": "text", + "content": "A Appendix" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 99, + 274, + 112 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 99, + 274, + 112 + ], + "spans": [ + { + "bbox": [ + 68, + 99, + 274, + 112 + ], + "type": "text", + "content": "A.1 Training Setup and Hyperparameters" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 121, + 291, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 121, + 291, + 420 + ], + "spans": [ + { + "bbox": [ + 67, + 121, + 291, + 420 + ], + "type": "text", + "content": "The MS MARCO is a large-scale English language dataset for machine reading comprehension (MRC). The dataset consists of anonymized queries sampled from Bing's search query logs, each with human generated answers. The training set we used contains 400,782 training samples, each consisting of a query, positive passage, and a set of hard negative passages, which we randomly select 7 hard negative passages for each training sample. We set a batch size to 16 and use in-batch negative sampling for each training sample. Therefore, we obtain " + }, + { + "bbox": [ + 67, + 121, + 291, + 420 + ], + "type": "inline_equation", + "content": "7 + 8 * 15 = 127" + }, + { + "bbox": [ + 67, + 121, + 291, + 420 + ], + "type": "text", + "content": " negative passages for each training sample. We use the AdamW optimizer and learning rate of 1e-5 for 150,000 steps with a linear learning rate warm-up over the first 10,000 steps and a linear learning rate decay over the rest of the training steps. For our training method, we set the hyper-parameters " + }, + { + "bbox": [ + 67, + 121, + 291, + 420 + ], + "type": "inline_equation", + "content": "\\beta = 0.5" + }, + { + "bbox": [ + 67, + 121, + 291, + 420 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 67, + 121, + 291, + 420 + ], + "type": "inline_equation", + "content": "\\gamma = 0.5" + }, + { + "bbox": [ + 67, + 121, + 291, + 420 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 67, + 121, + 291, + 420 + ], + "type": "inline_equation", + "content": "\\sigma = 0.2" + }, + { + "bbox": [ + 67, + 121, + 291, + 420 + ], + "type": "text", + "content": ", and the query augmentation size " + }, + { + "bbox": [ + 67, + 121, + 291, + 420 + ], + "type": "inline_equation", + "content": "K = 40" + }, + { + "bbox": [ + 67, + 121, + 291, + 420 + ], + "type": "text", + "content": ". Using one V100 32G GPU, the BERT-based model training time is around 31 hours, while the CharacterBERT-based model training time is roughly 56 hours." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 438, + 244, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 438, + 244, + 451 + ], + "spans": [ + { + "bbox": [ + 68, + 438, + 244, + 451 + ], + "type": "text", + "content": "A.2 Query Augmentation Examples" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 460, + 290, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 460, + 290, + 502 + ], + "spans": [ + { + "bbox": [ + 67, + 460, + 290, + 502 + ], + "type": "text", + "content": "Table 4 provides examples of misspelled queries generated by the Query Augmentation for each original query." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 86, + 519, + 143, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 519, + 143, + 529 + ], + "spans": [ + { + "bbox": [ + 86, + 519, + 143, + 529 + ], + "type": "text", + "content": "Original query:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 86, + 532, + 261, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 532, + 261, + 543 + ], + "spans": [ + { + "bbox": [ + 86, + 532, + 261, + 543 + ], + "type": "text", + "content": "what is the goddess of agriculture in greek mythology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 86, + 545, + 154, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 545, + 154, + 555 + ], + "spans": [ + { + "bbox": [ + 86, + 545, + 154, + 555 + ], + "type": "text", + "content": "Misspelled queries:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 86, + 558, + 271, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 558, + 271, + 569 + ], + "spans": [ + { + "bbox": [ + 86, + 558, + 271, + 569 + ], + "type": "text", + "content": "what is the goddoess of agriculture in greek mythology" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 86, + 571, + 263, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 571, + 263, + 582 + ], + "spans": [ + { + "bbox": [ + 86, + 571, + 263, + 582 + ], + "type": "text", + "content": "what is the goddess of agriculture in greek mythology" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 86, + 585, + 263, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 585, + 263, + 596 + ], + "spans": [ + { + "bbox": [ + 86, + 585, + 263, + 596 + ], + "type": "text", + "content": "what is the goddess of agriculture in greek mythologo" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 86, + 599, + 267, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 599, + 267, + 610 + ], + "spans": [ + { + "bbox": [ + 86, + 599, + 267, + 610 + ], + "type": "text", + "content": "what is the goddesses of agriculture in greek mythology" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 86, + 613, + 267, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 613, + 267, + 624 + ], + "spans": [ + { + "bbox": [ + 86, + 613, + 267, + 624 + ], + "type": "text", + "content": "what is the goddess of agriculture in greek mythology" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 86, + 626, + 270, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 626, + 270, + 637 + ], + "spans": [ + { + "bbox": [ + 86, + 626, + 270, + 637 + ], + "type": "text", + "content": "what is the goddess of agriculture in greek mythology" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 86, + 640, + 262, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 640, + 262, + 651 + ], + "spans": [ + { + "bbox": [ + 86, + 640, + 262, + 651 + ], + "type": "text", + "content": "what is the goddess of agriculture in greek mythology" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 86, + 655, + 267, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 655, + 267, + 665 + ], + "spans": [ + { + "bbox": [ + 86, + 655, + 267, + 665 + ], + "type": "text", + "content": "what is the goddess of agriculture in grvek mythology" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 86, + 668, + 267, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 668, + 267, + 679 + ], + "spans": [ + { + "bbox": [ + 86, + 668, + 267, + 679 + ], + "type": "text", + "content": "what is the goddess of agricultrue in greek mythology" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 86, + 682, + 267, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 682, + 267, + 692 + ], + "spans": [ + { + "bbox": [ + 86, + 682, + 267, + 692 + ], + "type": "text", + "content": "what is the goddess of agriculture in greek mythology" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 67, + 703, + 291, + 739 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 291, + 739 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 291, + 739 + ], + "type": "text", + "content": "Table 4: The outputs of Query Augmentation with " + }, + { + "bbox": [ + 67, + 703, + 291, + 739 + ], + "type": "inline_equation", + "content": "K = {10}" + }, + { + "bbox": [ + 67, + 703, + 291, + 739 + ], + "type": "text", + "content": " . We use different colors to indicate different types of typo: RandInsert , RandDelete , RandSub ," + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 70, + 740, + 225, + 754 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 740, + 225, + 754 + ], + "spans": [ + { + "bbox": [ + 70, + 740, + 225, + 754 + ], + "type": "text", + "content": "SwapNeighbor, and SwapAdjacent." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 303, + 71, + 373, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 71, + 373, + 83 + ], + "spans": [ + { + "bbox": [ + 303, + 71, + 373, + 83 + ], + "type": "text", + "content": "A.3 Licenses" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 302, + 89, + 525, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 89, + 525, + 155 + ], + "spans": [ + { + "bbox": [ + 302, + 89, + 525, + 155 + ], + "type": "text", + "content": "Datasets: The MS MARCO dataset is available under the MIT license, and the DL-typo dataset is available under the Apache license 2.0. These licenses allow users to use the datasets under nonrestrictive agreements." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 302, + 157, + 526, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 157, + 526, + 290 + ], + "spans": [ + { + "bbox": [ + 302, + 157, + 526, + 290 + ], + "type": "text", + "content": "Softwares: We employ Hugging Face (Wolf et al., 2020) and Tevatron (Gao et al., 2022) libraries to train dense retrieval models. We utilize Ranx library (Bassani, 2022) to evaluate retrieval performance. These libraries are available under the Apache license 2.0 which allows both academic and commercial usages. For this reason, we release our code under the Apache license 2.0 to make our code fully accessible and compatible with the other codes we use." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "1113" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 90, + 192, + 103 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 90, + 192, + 103 + ], + "spans": [ + { + "bbox": [ + 68, + 90, + 192, + 103 + ], + "type": "text", + "content": "A For every submission:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 106, + 317, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 106, + 317, + 121 + ], + "spans": [ + { + "bbox": [ + 77, + 106, + 317, + 121 + ], + "type": "text", + "content": "A1. Did you describe the limitations of your work?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 121, + 133, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 121, + 133, + 132 + ], + "spans": [ + { + "bbox": [ + 89, + 121, + 133, + 132 + ], + "type": "text", + "content": "Section 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 76, + 143, + 329, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 143, + 329, + 157 + ], + "spans": [ + { + "bbox": [ + 76, + 143, + 329, + 157 + ], + "type": "text", + "content": "A2. Did you discuss any potential risks of your work?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 157, + 525, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 157, + 525, + 185 + ], + "spans": [ + { + "bbox": [ + 89, + 157, + 525, + 185 + ], + "type": "text", + "content": "There is no potential risk associated with increasing the robustness of information retrieval applications to question containing misspellings." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 77, + 192, + 414, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 192, + 414, + 206 + ], + "spans": [ + { + "bbox": [ + 77, + 192, + 414, + 206 + ], + "type": "text", + "content": "A3. Do the abstract and introduction summarize the paper's main claims?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 89, + 206, + 133, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 206, + 133, + 218 + ], + "spans": [ + { + "bbox": [ + 89, + 206, + 133, + 218 + ], + "type": "text", + "content": "Section 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 77, + 228, + 399, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 228, + 399, + 243 + ], + "spans": [ + { + "bbox": [ + 77, + 228, + 399, + 243 + ], + "type": "text", + "content": "A4. Have you used AI writing assistants when working on this paper?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 89, + 243, + 524, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 243, + 524, + 269 + ], + "spans": [ + { + "bbox": [ + 89, + 243, + 524, + 269 + ], + "type": "text", + "content": "We use Grammarly to check grammatical errors and QuillBot to polish writing quality. These tools are applied to a certain number of sentences in each section, which are then reviewed by humans." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 68, + 279, + 290, + 292 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 279, + 290, + 292 + ], + "spans": [ + { + "bbox": [ + 68, + 279, + 290, + 292 + ], + "type": "text", + "content": "B Did you use or create scientific artifacts?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 297, + 525, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 297, + 525, + 324 + ], + "spans": [ + { + "bbox": [ + 67, + 297, + 525, + 324 + ], + "type": "text", + "content": "Section 3.1 for pre-trained language models, training dataset, and training toolkit. Section 3.2 for competitive methods. Section 3.3 for evaluation datasets and evaluation toolkit." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 77, + 333, + 315, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 333, + 315, + 346 + ], + "spans": [ + { + "bbox": [ + 77, + 333, + 315, + 346 + ], + "type": "text", + "content": "B1. Did you cite the creators of artifacts you used?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 89, + 347, + 524, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 347, + 524, + 374 + ], + "spans": [ + { + "bbox": [ + 89, + 347, + 524, + 374 + ], + "type": "text", + "content": "Section 3.1 for pre-trained language models, training dataset, and training toolkit. Section 3.2 for competitive methods. Section 3.3 for evaluation datasets and evaluation toolkit." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 77, + 381, + 464, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 381, + 464, + 396 + ], + "spans": [ + { + "bbox": [ + 77, + 381, + 464, + 396 + ], + "type": "text", + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 89, + 396, + 151, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 396, + 151, + 409 + ], + "spans": [ + { + "bbox": [ + 89, + 396, + 151, + 409 + ], + "type": "text", + "content": "Appendix A.3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 77, + 417, + 524, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 417, + 524, + 473 + ], + "spans": [ + { + "bbox": [ + 77, + 417, + 524, + 473 + ], + "type": "text", + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 89, + 474, + 151, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 474, + 151, + 486 + ], + "spans": [ + { + "bbox": [ + 89, + 474, + 151, + 486 + ], + "type": "text", + "content": "Appendix A.3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 77, + 495, + 524, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 495, + 524, + 535 + ], + "spans": [ + { + "bbox": [ + 77, + 495, + 524, + 535 + ], + "type": "text", + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 89, + 536, + 524, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 536, + 524, + 576 + ], + "spans": [ + { + "bbox": [ + 89, + 536, + 524, + 576 + ], + "type": "text", + "content": "We did not collect any data. The datasets we used are publicly available and widely used in information retrieval literature. The data is already anonymized by the creators of the datasets. Therefore we do not need to anonymize the data." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 77, + 585, + 524, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 585, + 524, + 613 + ], + "spans": [ + { + "bbox": [ + 77, + 585, + 524, + 613 + ], + "type": "text", + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 89, + 613, + 151, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 613, + 151, + 625 + ], + "spans": [ + { + "bbox": [ + 89, + 613, + 151, + 625 + ], + "type": "text", + "content": "Appendix A.1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 77, + 634, + 524, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 634, + 524, + 703 + ], + "spans": [ + { + "bbox": [ + 77, + 634, + 524, + 703 + ], + "type": "text", + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 89, + 703, + 381, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 703, + 381, + 716 + ], + "spans": [ + { + "bbox": [ + 89, + 703, + 381, + 716 + ], + "type": "text", + "content": "Section 3.3 for the evaluation set Appendix A.1 for the training set" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 67, + 719, + 522, + 740 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 719, + 522, + 740 + ], + "spans": [ + { + "bbox": [ + 67, + 719, + 522, + 740 + ], + "type": "text", + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "spans": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "text", + "content": "ACL 2023 Responsible NLP Checklist" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "1114" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 70, + 294, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 70, + 294, + 84 + ], + "spans": [ + { + "bbox": [ + 68, + 70, + 294, + 84 + ], + "type": "text", + "content": "C Did you run computational experiments?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 79, + 89, + 123, + 100 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 89, + 123, + 100 + ], + "spans": [ + { + "bbox": [ + 79, + 89, + 123, + 100 + ], + "type": "text", + "content": "Section 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 110, + 523, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 110, + 523, + 137 + ], + "spans": [ + { + "bbox": [ + 77, + 110, + 523, + 137 + ], + "type": "text", + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 138, + 151, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 138, + 151, + 152 + ], + "spans": [ + { + "bbox": [ + 89, + 138, + 151, + 152 + ], + "type": "text", + "content": "Appendix A.1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 159, + 524, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 159, + 524, + 187 + ], + "spans": [ + { + "bbox": [ + 77, + 159, + 524, + 187 + ], + "type": "text", + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 188, + 150, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 188, + 150, + 201 + ], + "spans": [ + { + "bbox": [ + 89, + 188, + 150, + 201 + ], + "type": "text", + "content": "Appendix A.1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 77, + 209, + 524, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 209, + 524, + 249 + ], + "spans": [ + { + "bbox": [ + 77, + 209, + 524, + 249 + ], + "type": "text", + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 88, + 251, + 524, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 251, + 524, + 278 + ], + "spans": [ + { + "bbox": [ + 88, + 251, + 524, + 278 + ], + "type": "text", + "content": "Section 4.1 for Main Results Section 4.2 for Query Augmentation Size Study Section 4.3 for Loss Ablation Study Section 4.4 for Query Distributions" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 77, + 286, + 524, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 286, + 524, + 326 + ], + "spans": [ + { + "bbox": [ + 77, + 286, + 524, + 326 + ], + "type": "text", + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 89, + 327, + 414, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 327, + 414, + 341 + ], + "spans": [ + { + "bbox": [ + 89, + 327, + 414, + 341 + ], + "type": "text", + "content": "Our evaluation is parameter free, therefore there is no parameter settings." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 350, + 522, + 364 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 350, + 522, + 364 + ], + "spans": [ + { + "bbox": [ + 67, + 350, + 522, + 364 + ], + "type": "text", + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 79, + 367, + 127, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 367, + 127, + 380 + ], + "spans": [ + { + "bbox": [ + 79, + 367, + 127, + 380 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 76, + 390, + 524, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 390, + 524, + 416 + ], + "spans": [ + { + "bbox": [ + 76, + 390, + 524, + 416 + ], + "type": "text", + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 89, + 417, + 148, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 417, + 148, + 430 + ], + "spans": [ + { + "bbox": [ + 89, + 417, + 148, + 430 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 76, + 439, + 524, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 439, + 524, + 479 + ], + "spans": [ + { + "bbox": [ + 76, + 439, + 524, + 479 + ], + "type": "text", + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 89, + 481, + 148, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 481, + 148, + 493 + ], + "spans": [ + { + "bbox": [ + 89, + 481, + 148, + 493 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 76, + 502, + 524, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 502, + 524, + 542 + ], + "spans": [ + { + "bbox": [ + 76, + 502, + 524, + 542 + ], + "type": "text", + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 89, + 544, + 148, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 544, + 148, + 556 + ], + "spans": [ + { + "bbox": [ + 89, + 544, + 148, + 556 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 76, + 565, + 520, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 565, + 520, + 578 + ], + "spans": [ + { + "bbox": [ + 76, + 565, + 520, + 578 + ], + "type": "text", + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 89, + 580, + 148, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 580, + 148, + 592 + ], + "spans": [ + { + "bbox": [ + 89, + 580, + 148, + 592 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 76, + 602, + 524, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 602, + 524, + 628 + ], + "spans": [ + { + "bbox": [ + 76, + 602, + 524, + 628 + ], + "type": "text", + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 89, + 629, + 148, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 629, + 148, + 642 + ], + "spans": [ + { + "bbox": [ + 89, + 629, + 148, + 642 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "1115" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Uncertainty-Aware Bootstrap Learning for Joint Extraction on Distantly-Supervised Data/a0f755f8-19b9-4154-a835-8d33c06b518f_content_list.json b/2023/Uncertainty-Aware Bootstrap Learning for Joint Extraction on Distantly-Supervised Data/a0f755f8-19b9-4154-a835-8d33c06b518f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0206c9a7ef504489e77dbb292525d3b608867c20 --- /dev/null +++ b/2023/Uncertainty-Aware Bootstrap Learning for Joint Extraction on Distantly-Supervised Data/a0f755f8-19b9-4154-a835-8d33c06b518f_content_list.json @@ -0,0 +1,1385 @@ +[ + { + "type": "text", + "text": "Uncertainty-Aware Bootstrap Learning for Joint Extraction on Distantly-Supervised Data", + "text_level": 1, + "bbox": [ + 174, + 89, + 823, + 130 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yufei Li $^{1}$ , Xiao Yu $^{2}$ , Yanchi Liu $^{3}$ , Haifeng Chen $^{3}$ , Cong Liu $^{1}$", + "bbox": [ + 233, + 148, + 769, + 165 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of California, Riverside 2Stellar Cyber 3NEC Labs America", + "bbox": [ + 206, + 166, + 794, + 181 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}\\{yli927,congl\\} @ucr.edu,$ $^{2}\\mathrm{xyu@stellarcyber.ai},$", + "bbox": [ + 278, + 183, + 724, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3{yanchi,haifeng}@nec-labs.com", + "bbox": [ + 349, + 199, + 652, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 252, + 339, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jointly extracting entity pairs and their relations is challenging when working on distantly-supervised data with ambiguous or noisy labels. To mitigate such impact, we propose uncertainty-aware bootstrap learning, which is motivated by the intuition that the higher uncertainty of an instance, the more likely the model confidence is inconsistent with the ground truths. Specifically, we first explore instance-level data uncertainty to create an initial high-confident examples. Such subset serves as filtering noisy instances and facilitating the model to converge fast at the early stage. During bootstrap learning, we propose self-ensembling as a regularizer to alleviate inter-model uncertainty produced by noisy labels. We further define probability variance of joint tagging probabilities to estimate inner-model parametric uncertainty, which is used to select and build up new reliable training instances for the next iteration. Experimental results on two large datasets reveal that our approach outperforms existing strong baselines and related methods.", + "bbox": [ + 144, + 274, + 460, + 613 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 114, + 621, + 258, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Joint extraction involves extracting multiple types of entities and relations between them using a single model, which is necessary in automatic knowledge base construction (Yu et al., 2020). One way to cheaply acquire a large amount of labeled data for training joint extraction models is through distant supervision (DS) (Mintz et al., 2009). DS involves aligning a knowledge base (KB) with an unlabeled corpus using hand-crafted rules or logic constraints. Due to the lack of human annotators, DS brings a large proportion of noisy labels, e.g., over $30\\%$ noisy instances in some cases (Mintz et al., 2009), making it impossible to learn useful features. The noise can be either false relations due to the aforementioned rule-based matching assumption or wrong entity tags due to limited coverage over entities in open-domain KBs.", + "bbox": [ + 115, + 645, + 489, + 917 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing distantly-supervised approaches model noise relying either on heuristics such as reinforcement learning (RL) (Nooralahzadeh et al., 2019; Hu et al., 2021) and adversarial learning (Chen et al., 2021), or pattern-based methods (Jia et al., 2019; Shang et al., 2022) to select trustable instances. Nevertheless, these methods require designing heuristics or hand-crafted patterns which may encourage a model to leverage spurious features without considering the confidence or uncertainty of its predictions.", + "bbox": [ + 507, + 253, + 885, + 429 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In response to these problems, we propose UnBED—Uncertainty-aware Bootstrap learning for joint Extraction on Distantly-supervised data. UnBED assumes that 1) low data uncertainty indicates reliable instances using a pre-trained language model (PLM) in the initial stage, 2) model should be aware of trustable entity and relation labels regarding its uncertainty after training. Our bootstrap serves uncertainty as a principle to mitigate the impact of noise labels on model learning and validate input sequences to control the number of training examples in each step. Particularly, we quantify data uncertainty of an instance according to its winning score (Hendrycks and Gimpel, 2017) and entropy (Shannon, 1948). We define averaged maximum probability that is estimated by a joint PLM over each token in a sequence to adapt previous techniques in joint extraction scheme. Instances with low data uncertainty are collected to form an initial subset, which is used to tune the joint PLM tagger and facilitate fast convergence. Then, we define parametric uncertainty in two perspectives—inter-model and inner-model uncertainty. The former is quantified by self-ensembling (Wang and Wang, 2022) and serves as a regularizer to improve model robustness against noisy labels during training. The latter is captured by probability variance in MC Dropout (Gal and Ghahramani, 2016) for selecting new confident instances for the next training iteration. Such two", + "bbox": [ + 507, + 437, + 884, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1349", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics", + "bbox": [ + 226, + 945, + 771, + 957 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Volume 2: Short Papers, pages 1349-1358", + "bbox": [ + 368, + 958, + 630, + 971 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "July 9-14, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 295, + 972, + 700, + 985 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "fold model uncertainties reinforce with each other to guide the model to iteratively improve its robustness and learn from reliable knowledge.", + "bbox": [ + 112, + 84, + 489, + 134 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 114, + 148, + 270, + 162 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Joint Extraction Methods Joint extraction detects entities and their relations using a single model, which effectively integrates the information from both sources and therefore achieves better results in both subtasks compared to pipelined methods (Zheng et al., 2017). For example, unified methods tag entities and relation simultaneously, e.g., (Zheng et al., 2017) proposes a novel tagging scheme which converts joint extraction to a sequence labeling problem; (Dai et al., 2019) introduces query position and sequential tagging to extract overlapping relations. Such methods avoid producing redundant information compared to parameter-sharing neural models (Miwa and Bansal, 2016; Gupta et al., 2016), and require no hand-crafted features that are used in structured systems (Yu et al., 2020).", + "bbox": [ + 112, + 175, + 489, + 449 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the challenge of learning from DS, pre-trained transformers (e.g., BERT, GPT-2) have gain much attention. They model strong expressive context-aware representations for text sequence through multiple attention layers, and achieve state-of-the-art performance on various NLP tasks (Radford et al., 2019; Devlin et al., 2019; Li et al., 2022). They can be cheaply fine-tuned to solve different downstream tasks including NER and RC. Specifically, BERT is trained on large English corpus using masked language modeling. The multi-head attention weights indicate interactions between each pair of words and its hidden states integrate semantic information of the whole sentence, which are used to decode different tagging results.", + "bbox": [ + 112, + 451, + 489, + 692 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Uncertainty Methods Uncertainty generally comes from two sources—aleatoric uncertainty and epistemic uncertainty. The former is also referred to as data uncertainty, describing noise inherent in the data generation. Methods mitigating such uncertainty include data interpolation (Dong et al., 2018), winning score, and temperature scale (Guo et al., 2017). The latter is also called model uncertainty, describing whether the structure choice and model parameters best describe the data distribution. One main solution to mitigate model uncertainty is Bayesian Neural Network (BNN) (Klein et al., 2017) that puts a prior distribution on its weights. To save computational cost, Monte Carlo", + "bbox": [ + 112, + 694, + 489, + 919 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c1f2c545fd2d830bab363ce8adb2f0142a1c0f955d0de4e723956682688a7890.jpg", + "image_caption": [ + "Figure 1: Joint extraction as a token classification task." + ], + "image_footnote": [], + "bbox": [ + 589, + 80, + 806, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "dropout is proposed as an approximation of variational Bayesian inference (Gal and Ghahramani, 2016), realized by training models with dropout layers and testing with stochastic inference to quantify probability variance. Besides BNN, self-ensembling (Wang and Wang, 2022) which measures the outputs variance between models with the same architecture has been shown effective to reduce parametric uncertainty across models.", + "bbox": [ + 507, + 326, + 884, + 472 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3 Joint Extraction Model", + "text_level": 1, + "bbox": [ + 507, + 487, + 744, + 502 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Tagging Scheme For an input sequence $\\mathcal{X} = \\{x_1, \\dots, x_n\\}$ , we tag $n$ sequences according to different query position $p$ following (Dai et al., 2019). If $p$ is the start of an entity (query entity $e_1$ ), the sequence is an instance. The entity type is labeled at $p$ and other entities $e_2$ which have relationship with the query entity are labeled with relation types $re$ . The rest of tokens are labeled \"O\" (Outside), meaning they do not correspond to the query entity. Accordingly, we convert joint extraction into a token classification task and extract relation triplets $\\{e_1, re, e_2\\}$ in each instance, as shown in Figure 1.", + "bbox": [ + 507, + 514, + 884, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Position-Attentive Encoder we use BERT (Devlin et al., 2019) to encode a sentence $\\mathcal{X}$ into token-level representations $h = \\{h_1,..,h_n\\}$ , where $h_i\\in \\mathbb{R}^d$ is a $d$ -dimensional vector corresponding to the $i$ -th token in $\\mathcal{X}$ . For each query $p$ , self-matching is applied to calculate the position-attention $\\boldsymbol{a}_t\\in \\mathbb{R}^T$ between token at $p$ and each token at target position $t$ , which compares the sentence representations against itself to collect context information (Tan et al., 2018). The produced position-aware representation $\\boldsymbol{c}_t\\in \\mathbb{R}^{T\\times d}$ is an attention-weighted sentence vector $\\boldsymbol{c}_t = \\boldsymbol{a}_t^\\top \\boldsymbol{h}$ . Finally, we concatenate $\\boldsymbol{h}_t$ and $\\boldsymbol{c}_t$ to generate position-aware and context", + "bbox": [ + 507, + 709, + 885, + 919 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1350", + "bbox": [ + 480, + 927, + 521, + 940 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "aware representations $\\pmb{u}_t = [\\pmb{h}_t|\\pmb{c}_t]$ .", + "bbox": [ + 112, + 84, + 379, + 99 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CRF Decoder (Lafferty et al., 2001) For each position-aware representation $\\boldsymbol{u}_t$ , we first learn a linear transformation $\\boldsymbol{z}_t = \\boldsymbol{W}\\boldsymbol{u}_t \\in \\mathbb{R}^C$ to represent tag scores for the $t$ -th token. Here $C$ is the number of distinct tags. For an instance with labels $\\boldsymbol{y} = \\{y_1, \\dots, y_n\\}$ , the decoding score $s(\\boldsymbol{z}, \\boldsymbol{y})$ is the sum of transition scores from tag $y_t$ to tag $y_{t+1}$ plus the input score $z_t^{yt}$ . The conditional probability $p(\\boldsymbol{y}|\\boldsymbol{z})$ is the softmax over $s(\\boldsymbol{z}, \\boldsymbol{y})$ for all possible label sequences $\\boldsymbol{y}'$ . We maximize the log-likelihood of correct tag sequences during training $\\mathcal{L}_c = \\sum \\log p(\\boldsymbol{y}|\\boldsymbol{z})$ .", + "bbox": [ + 112, + 99, + 489, + 294 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4 Uncertainty-Aware Bootstrap Learning", + "text_level": 1, + "bbox": [ + 112, + 305, + 487, + 322 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motivation One of the main challenges in bootstrap learning is to evaluate the \"correctness\" of a labeled instance. We consider this problem from an uncertainty perspective and assume instances with lower uncertainty are more likely to be correctly labeled. In this section, we first propose instance-level data uncertainty which is used to filter noisy examples and build an initial subset. Then, we introduce our two-fold model uncertainties which helps iteratively mitigate DS effect and build up trustable examples during bootstrap learning.", + "bbox": [ + 112, + 330, + 489, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1 Data Uncertainty", + "text_level": 1, + "bbox": [ + 112, + 518, + 297, + 533 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Presenting examples in an easy-to-hard order at different training stages can benefit models (Platanios et al., 2019; Zhou et al., 2020), we propose data uncertainty as a way to quantify the \"hardness\" of an instance. To better estimate the data uncertainty, we use pre-trained language models (PLMs) to generate tag probability for each token in a sequence. Our intuition is that higher uncertain inputs are \"harder\" to be generated by a PLM, as it already has rationales of language. Accordingly, we propose two data uncertainties, which can be used individually or combined together:", + "bbox": [ + 112, + 539, + 489, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Winning Score (WS) The maximum softmax probability reflects data uncertainty of an input (Hendrycks and Gimpel, 2017). Given an input instance $\\mathcal{I} = \\{x_1,\\dots,x_n\\}$ , we define data uncertainty $u^{d}(\\mathcal{I})$ as the minus averaged token classification winning score:", + "bbox": [ + 112, + 732, + 489, + 829 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nu ^ {d} (\\mathcal {I}) = - \\frac {1}{n} \\sum_ {t = 1} ^ {n} \\max _ {c \\in [ 1, C ]} P (y _ {t} = c | x _ {t}) \\quad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 131, + 856, + 487, + 876 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Entropy Shannon entropy (Shannon, 1948) is widely used to reflect information uncertainty. We", + "bbox": [ + 112, + 887, + 487, + 917 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "propose data uncertainty $u^{d}(\\mathcal{I})$ as the averaged token classification entropy:", + "bbox": [ + 507, + 84, + 882, + 116 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nu ^ {d} (\\mathcal {I}) = \\frac {1}{n} \\sum_ {t = 1} ^ {n} \\sum_ {c = 1} ^ {C} P \\left(y _ {t} = c \\mid x _ {t}\\right) \\log P \\left(y _ {t} = c \\mid x _ {t}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 510, + 147, + 882, + 180 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We filter out examples with high uncertainty scores and build an initial subset with \"simple\" examples. At the early training stage, a model is not aware of what a decent distribution $P(y|x)$ should be, thus data uncertainty facilitates it to converge fast by tuning on a fairly \"simple\" subset.", + "bbox": [ + 507, + 181, + 884, + 279 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.2 Model Uncertainty", + "text_level": 1, + "bbox": [ + 507, + 292, + 705, + 307 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our bootstrap learning, we define model uncertainty, i.e., epistemic uncertainty (Kendall and Gal, 2017), to measure whether model parameters can best describe the data distribution following (Zhou et al., 2020). A small model uncertainty indicates the model is confident that the current training data has been well learned (Wang et al., 2019). We adopt Monte Carlo Dropout (Gal and Ghahramani, 2016) to approximate Bayesian inference which captures inner-model parametric uncertainty. Specifically, we perform $K$ forward passes through our joint model. In each pass, part of network neurons $\\theta$ are randomly deactivated. Finally, we yield $K$ samples on model parameters $\\{\\hat{\\theta}_1,\\dots,\\hat{\\theta}_K\\}$ . We use the averaged token classification Probability Variance (PV) (Shelmanov et al., 2021) over all tags for instance $\\mathcal{I}$ :", + "bbox": [ + 507, + 313, + 884, + 588 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nu ^ {m} (\\theta) = \\frac {1}{n} \\sum_ {t = 1} ^ {n} \\sum_ {c = 1} ^ {C} \\operatorname {V a r} \\left[ P \\left(y _ {t} = c \\mid x _ {t}, \\hat {\\theta} _ {k}\\right) \\right] _ {k = 1} ^ {K} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 510, + 615, + 882, + 656 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\operatorname{Var}[\\cdot]$ is the variance of distribution over the $K$ passes following the common settings in (Dong et al., 2018; Xiao and Wang, 2019). Accordingly, model is aware of its confidence over each instance and how likely the label is noisy.", + "bbox": [ + 507, + 658, + 882, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.3 Training Strategy", + "text_level": 1, + "bbox": [ + 507, + 752, + 695, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Uncertainty-Aware Loss Besides MC Dropout which measures parametric uncertainty within a model, we also consider mitigating parametric uncertainty between models to stabilize the weights during training. Specifically, we use self-ensembling (He et al., 2020; Wang and Wang, 2022) to calculate the loss between the same models to improve model robustness and reduce the label noise effect on model performance.", + "bbox": [ + 507, + 774, + 884, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1351", + "bbox": [ + 482, + 927, + 517, + 940 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Algorithm 1 Bootstrap Learning", + "text_level": 1, + "bbox": [ + 115, + 83, + 361, + 99 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Input: Original dataset $\\mathcal{D} = \\{(\\mathcal{I}^n,y^n)\\}_{n = 1}^N$ two joint models $f_{1},f_{2}$ with parameters $\\theta_{1},\\theta_{2}$", + "bbox": [ + 115, + 104, + 485, + 137 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Compute data uncertainty $u^{d}(\\mathcal{I})$ for each instance $\\mathcal{I}$ in $\\mathcal{D}$ ;", + "2: Initial dataset $\\mathcal{C} \\gets$ Select data pairs $(\\mathcal{I}^n, y^n)$ such that $u^d(\\mathcal{I}) < \\tau^d$ from $\\mathcal{D}$ ;", + "3: for epoch $e = 1, \\ldots$ do", + "4: Train $f_{1}, f_{2}$ on $\\mathcal{C}$ using Eq. (5);", + "5: Calculate model uncertainty $u^{m}(\\theta_{1})$ on $\\mathcal{D}$ ;", + "6: $\\mathcal{C} \\gets$ Select data pairs $(\\mathcal{I}^n, y^n)$ such that $u^m(\\mathcal{I}; \\theta_1) < \\tau^m$ from $\\mathcal{D}$ ;" + ], + "bbox": [ + 124, + 139, + 487, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We create another joint model with identical framework, e.g., architecture, loss functions, hyperparameters, and compute a self-ensemble loss $\\mathcal{L}_e$ to minimize the difference between two outputs from the two models regarding the same inputs:", + "bbox": [ + 112, + 318, + 487, + 399 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {e} = \\sum K L (f (\\mathcal {I}; \\theta_ {1}), f (\\mathcal {I}; \\theta_ {2})) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 424, + 487, + 445 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $KL(.)$ is the Kullback-Leibler divergence between two probabilistic distributions, $\\theta_{1},\\theta_{2}$ denote the parameters of first and second models. We formulate our final uncertainty-aware objective $\\mathcal{L}$ as the sum of CRF and self-ensemble loss:", + "bbox": [ + 112, + 464, + 487, + 544 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {c} + \\alpha \\mathcal {L} _ {e} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 242, + 574, + 487, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha$ denotes the weight of self-ensembling, and $\\mathcal{L}_c$ means the token classification loss.", + "bbox": [ + 112, + 609, + 487, + 640 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Bootstrap Learning Procedure To mitigate the DS effect on model performance, we propose a twofold bootstrap learning strategy (see Algorithm 1). Specifically, we first apply data uncertainty to filter \"harder\" examples and redistribute a reliable initial training data $\\mathcal{M}$ . Then, we iteratively feed examples following an easy-to-hard order to the model. In each training iteration, we regularize the joint model with self-ensembling loss to reduce the impact of noisy labels on model parameters. Then we use probability variance to select new confident training instances $\\mathcal{D}'$ that can be explained by the model as the next training inputs. The more certain examples are selected, the more likely the model will learn beneficial information and will converge faster. We repeat the above procedure until the F1 score on the validation set converges.", + "bbox": [ + 112, + 645, + 487, + 917 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 509, + 83, + 653, + 99 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.1 Setup", + "text_level": 1, + "bbox": [ + 509, + 109, + 601, + 124 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We evaluate the performance of UnBED on two datasets, NYT and Wiki-KBP. The NYT (Riedel et al., 2010) dataset collects news from New York Times and its training data is automatically labeled by DS. We use the revised test dataset (Jia et al., 2019) that is manually annotated to ensure quality. The Wiki-KBP (Ling and Weld, 2012) dataset collects articles from Wikipedia. Its training data is labeled by DS (Liu et al., 2017), and the test set is manually annotated (Ellis et al., 2013).", + "bbox": [ + 507, + 130, + 882, + 290 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We compare UnBED with the following baselines: ARNOR (Jia et al., 2019), a pattern-based method to reduce noise for distantly-supervised triplet extraction. PURE (Zhong and Chen, 2021), a pipeline approach that uses pre-trained BERT entity model to first recognize entities and then employs a relation model to detect underlying relations. FAN (Hao et al., 2021), an adversarial method including a transformers encoder to reduce noise for distantly-supervised triplet extraction.", + "bbox": [ + 507, + 291, + 882, + 451 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Evaluation We evaluate the extracted triplets for each sentence based on Precision (Prec.), Recall (Rec.), and F1. A triplet $\\{e_1, re, e_2\\}$ is marked correct if the relation type $re$ , two entities $e_1, e_2$ are all correct. We build a validation set by randomly sampling $10\\%$ sentences from the test set.", + "bbox": [ + 507, + 453, + 882, + 548 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Implementation Details We use Hugging Face bert-large-uncased (Devlin et al., 2019) pre-trained model as backbone. For ARNOR, the hidden vector size is set to 300. In regularization training, we find optimal parameters $\\alpha$ as 1 for both datasets. We implement UnBED and all baselines in PyTorch, with Adam optimizer, initial learning rate $10^{-5}$ , dropout rate 0.1, and batch size 8. For initial subset configuration, we choose data uncertainty threshold 0.5. For bootstrap learning, an empirical model uncertainty threshold is set to 0.6 with the best validation F1.", + "bbox": [ + 507, + 549, + 882, + 741 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.2 Overall Results", + "text_level": 1, + "bbox": [ + 507, + 753, + 675, + 766 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Table 1, UnBED significantly outperforms all baselines in precision and F1 metric. Specifically, UnBED achieves $8\\%$ F1 improvement on NYT (3% on Wiki-KBP) over denoising approaches—ARNOR and FAN. Our approach also outperforms baselines using pretrained transformers (PURE and FAN), showing that uncertainty-aware bootstrap learning effectively reduces the impact of noisy labels.", + "bbox": [ + 507, + 774, + 882, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1352", + "bbox": [ + 482, + 927, + 521, + 940 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/233e3b0b2d247d6c323133f9234c64f0d04deb61933c9c4eae036c8234df3694.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodNYTWiki-KBP
Prec.Rec.F1Prec.Rec.F1
ARNOR (Jia et al., 2019)0.5880.6140.6000.4020.4710.434
PURE (Zhong and Chen, 2021)0.5360.6640.5930.3950.4330.413
FAN (Hao et al., 2021)0.5790.6460.6110.3910.4670.426
UnBED-WS0.6620.7300.6940.4290.5010.462
UnBED-Entropy0.6510.7410.6930.4220.5090.461
", + "bbox": [ + 186, + 80, + 811, + 199 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1: Evaluation results on NYT and Wiki-KBP datasets. Bold numbers denote the best metrics. UnBED-WS and UnBED-Entropy denote UnBED with winning score and entropy as the data uncertainty, respectively.", + "bbox": [ + 112, + 208, + 882, + 237 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0b1cd8eb90ceb16ce072eac5bd6e552b67596131705df3883ebf1cfe9ff3f8eb.jpg", + "image_caption": [ + "Figure 2: F1 score vs. Epochs under different settings. Vanilla-PV-enssembled denotes UnBED-WS, and entropy-PV-enssembled denotes UnBED-Entropy." + ], + "image_footnote": [], + "bbox": [ + 132, + 262, + 470, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.3 Further Analysis", + "text_level": 1, + "bbox": [ + 112, + 491, + 294, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We analyze the functionality of different components in Figure 2. We observe that both the entropy-PV and vanilla-PV outperform the baseline (joint model directly trained on the original DS dataset) in terms of F1 $(5\\sim 7\\%)$ increase), demonstrating the effect of filtering noisy labels and selecting trustable instance using probability variance. Besides, self-ensembling further enhances the performance in later training stage $(2\\sim 4$ F1 increase), proving that mitigating the inter-model uncertainty benefits model robustness against noisy labels.", + "bbox": [ + 112, + 511, + 489, + 690 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6 Conclusions", + "text_level": 1, + "bbox": [ + 112, + 701, + 253, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this paper, we propose a novel uncertainty-aware bootstrap learning framework for distantly-supervised joint extraction. Specifically, we define data uncertainty in generally token classification to filter out highly-error-prone instances and build an initial high-confident subset, which is used to tune the joint extraction model for fast convergence. We then propose a two-fold bootstrap learning procedure which iteratively mitigates the DS impact on model robustness and selects new trustable training instances. Experimental results on two benchmark datasets show that UnBED significantly out", + "bbox": [ + 112, + 726, + 490, + 917 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "performs other denoising techniques.", + "bbox": [ + 507, + 263, + 784, + 279 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Limitations", + "text_level": 1, + "bbox": [ + 509, + 290, + 613, + 305 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this work we propose an uncertainty-aware bootstrap learning framework for joint extraction. Though it achieves state-of-the-art performance compared to other denoising techniques, UnBED requires large training resources considering the ensemble loss calculated between two large PLMs and the probability variance calculated on the PLM joint extraction model. In our future work, we hope to incorporate pruning techniques during training to improve the efficiency. We will also consider more complex relations between entities, e.g., relations beyond the sentence boundary, to fit in real-world information extraction scenarios.", + "bbox": [ + 507, + 315, + 884, + 524 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 509, + 536, + 680, + 552 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This work was supported by NSF CNS 2135625, CPS 2038727, CNS Career 1750263, and a Darpa Shell grant.", + "bbox": [ + 507, + 561, + 882, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 510, + 636, + 608, + 652 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Tao Chen, Haochen Shi, Liyuan Liu, Siliang Tang, Jian Shao, Zhigang Chen, and Yueting Zhuang. 2021. Empower distantly supervised relation extraction with collaborative adversarial training. In Thirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2021, Thirty-Third Conference on Innovative Applications of Artificial Intelligence, IAAI 2021, The Eleventh Symposium on Educational Advances in Artificial Intelligence, EAAI 2021, Virtual Event, February 2-9, 2021, pages 12675-12682. AAAI Press.", + "Dai Dai, Xinyan Xiao, Yajuan Lyu, Shan Dou, Qiaoqiao She, and Haifeng Wang. 2019. Joint extraction of entities and overlapping relations using position-attentive sequence labeling. In The Thirty-Third AAAI Conference on Artificial Intelligence, AAAI 2019, The Thirty-First Innovative Applications of Artificial Intelligence Conference, IAAI 2019, The Ninth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2019, Honolulu, Hawaii," + ], + "bbox": [ + 510, + 658, + 885, + 917 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1353", + "bbox": [ + 482, + 927, + 519, + 940 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "USA, January 27 - February 1, 2019, pages 6300-6308. AAAI Press.", + "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics.", + "Li Dong, Chris Quirk, and Mirella Lapata. 2018. Confidence modeling for neural semantic parsing. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 743-753, Melbourne, Australia. Association for Computational Linguistics.", + "Joe Ellis, Jeremy Getman, Justin Mott, Xuansong Li, Kira Griffith, Stephanie M. Strassel, and Jonathan Wright. 2013. Linguistic resources for 2013 knowledge base population evaluations. In Proceedings of the Sixth Text Analysis Conference, TAC 2013, Gaithersburg, Maryland, USA, November 18-19, 2013. NIST.", + "Yarin Gal and Zoubin Ghahramani. 2016. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In Proceedings of the 33nd International Conference on Machine Learning, ICML 2016, New York City, NY, USA, June 19-24, 2016, volume 48 of JMLR Workshop and Conference Proceedings, pages 1050-1059. JMLR.org.", + "Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. 2017. On calibration of modern neural networks. In Proceedings of the 34th International Conference on Machine Learning, ICML 2017, Sydney, NSW, Australia, 6-11 August 2017, volume 70 of Proceedings of Machine Learning Research, pages 1321-1330. PMLR.", + "Pankaj Gupta, Hinrich Schütze, and Bernt Andrassy. 2016. Table filling multi-task recurrent neural network for joint entity and relation extraction. In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 2537-2547, Osaka, Japan. The COLING 2016 Organizing Committee.", + "Kailong Hao, Botao Yu, and Wei Hu. 2021. Knowing false negatives: An adversarial training method for distantly supervised relation extraction. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 9661-9672, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.", + "Jianfeng He, Xuchao Zhang, Shuo Lei, Zhiqian Chen, Fanglan Chen, Abdulaziz Alhamadani, Bei Xiao, and ChangTien Lu. 2020. Towards more accurate uncertainty estimation in text classification. In Proceedings of the 2020 Conference on Empirical Methods" + ], + "bbox": [ + 115, + 85, + 489, + 917 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "in Natural Language Processing (EMNLP), pages 8362-8372, Online. Association for Computational Linguistics.", + "Dan Hendrycks and Kevin Gimpel. 2017. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net.", + "Xuming Hu, Chenwei Zhang, Yawen Yang, Xiaohe Li, Li Lin, Lijie Wen, and Philip S. Yu. 2021. Gradient imitation reinforcement learning for low resource relation extraction. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, EMNLP 2021, Virtual Event / Punta Cana, Dominican Republic, 7-11 November; 2021, pages 2737-2746. Association for Computational Linguistics.", + "Wei Jia, Dai Dai, Xinyan Xiao, and Hua Wu. 2019. ARNOR: Attention regularization based noise reduction for distant supervision relation classification. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1399-1408, Florence, Italy. Association for Computational Linguistics.", + "Alex Kendall and Yarin Gal. 2017. What uncertainties do we need in bayesian deep learning for computer vision? In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 5574-5584.", + "Aaron Klein, Stefan Falkner, Jost Tobias Springenberg, and Frank Hutter. 2017. Learning curve prediction with bayesian neural networks. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net.", + "John D. Lafferty, Andrew McCallum, and Fernando C. N. Pereira. 2001. Conditional random fields: Probabilistic models for segmenting and labeling sequence data. In Proceedings of the Eighteenth International Conference on Machine Learning, ICML '01, page 282-289, San Francisco, CA, USA. Morgan Kaufmann Publishers Inc.", + "Shuyang Li, Yufei Li, Jianmo Ni, and Julian McAuley. 2022. SHARE: a system for hierarchical assistive recipe editing. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 11077-11090, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.", + "Xiao Ling and Daniel S. Weld. 2012. Fine-grained entity recognition. In Proceedings of the Twenty-Sixth AAAI Conference on Artificial Intelligence, July 22-26, 2012, Toronto, Ontario, Canada. AAAI Press." + ], + "bbox": [ + 510, + 85, + 884, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1354", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Liyuan Liu, Xiang Ren, Qi Zhu, Shi Zhi, Huan Gui, Heng Ji, and Jiawei Han. 2017. Heterogeneous supervision for relation extraction: A representation learning approach. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 46-56, Copenhagen, Denmark. Association for Computational Linguistics.", + "Mike Mintz, Steven Bills, Rion Snow, and Daniel Jurafsky. 2009. Distant supervision for relation extraction without labeled data. In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP, pages 1003-1011, Suntec, Singapore. Association for Computational Linguistics.", + "Makoto Miwa and Mohit Bansal. 2016. End-to-end relation extraction using lstms on sequences and tree structures. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, ACL 2016, August 7-12, 2016, Berlin, Germany, Volume 1: Long Papers. The Association for Computer Linguistics.", + "Farhad Nooralahzadeh, Jan Tore Lönning, and Lilja Øvrelid. 2019. Reinforcement-based denoising of distantly supervised NER with partial annotation. In Proceedings of the 2nd Workshop on Deep Learning Approaches for Low-Resource NLP, DeepLo@EMNLP-IJCNLP 2019, Hong Kong, China, November 3, 2019, pages 225–233. Association for Computational Linguistics.", + "Emmanouil Antonios Platanios, Otilia Stretcu, Graham Neubig, Barnabás Póczos, and Tom M. Mitchell. 2019. Competence-based curriculum learning for neural machine translation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 1162-1172. Association for Computational Linguistics.", + "Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners.", + "Sebastian Riedel, Limin Yao, and Andrew McCallum. 2010. Modeling relations and their mentions without labeled text. In Machine Learning and Knowledge Discovery in Databases, European Conference, ECML PKDD 2010, Barcelona, Spain, September 20-24, 2010, Proceedings, Part III, volume 6323 of Lecture Notes in Computer Science, pages 148-163. Springer.", + "Yuming Shang, Heyan Huang, Xin Sun, Wei Wei, and Xian-Ling Mao. 2022. A pattern-aware self-attention network for distant supervised relation extraction. Inf. Sci., 584:269-279.", + "Claude E. Shannon. 1948. A mathematical theory of communication. Bell Syst. Tech. J., 27(3):379-423." + ], + "bbox": [ + 115, + 85, + 489, + 917 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Artem Shelmanov, Evgenii Tsymbalov, Dmitri Puzyrev, Kirill Fedyanin, Alexander Panchenko, and Maxim Panov. 2021. How certain is your Transformer? In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 1833-1840, Online. Association for Computational Linguistics.", + "Zhixing Tan, Mingxuan Wang, Jun Xie, Yidong Chen, and Xiaodong Shi. 2018. Deep semantic role labeling with self-attention. In Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence and Thirtieth Innovative Applications of Artificial Intelligence Conference and Eighth AAAI Symposium on Educational Advances in Artificial Intelligence, AAAI'18/IAAI'18/EAAI'18. AAAI Press.", + "Hongjun Wang and Yisen Wang. 2022. Self-ensemble adversarial training for improved robustness. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net.", + "Shuo Wang, Yang Liu, Chao Wang, Huanbo Luan, and Maosong Sun. 2019. Improving back-translation with uncertainty-based confidence estimation. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLP-IJCNLP 2019, Hong Kong, China, November 3-7, 2019, pages 791-802. Association for Computational Linguistics.", + "Yijun Xiao and William Yang Wang. 2019. Quantifying uncertainties in natural language processing tasks. In The Thirty-Third AAAI Conference on Artificial Intelligence, AAAI 2019, The Thirty-First Innovative Applications of Artificial Intelligence Conference, IAAI 2019, The Ninth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2019, Honolulu, Hawaii, USA, January 27 - February 1, 2019, pages 7322-7329. AAAI Press.", + "Bowen Yu, Zhenyu Zhang, Xiaobo Shu, Tingwen Liu, Yubin Wang, Bin Wang, and Sujian Li. 2020. Joint extraction of entities and relations based on a novel decomposition strategy. In ECAI 2020 - 24th European Conference on Artificial Intelligence, 29 August-8 September 2020, Santiago de Compostela, Spain, August 29 - September 8, 2020 - Including 10th Conference on Prestigious Applications of Artificial Intelligence (PAIS 2020), volume 325 of Frontiers in Artificial Intelligence and Applications, pages 2282-2289. IOS Press.", + "Suncong Zheng, Feng Wang, Hongyun Bao, Yuexing Hao, Peng Zhou, and Bo Xu. 2017. Joint extraction of entities and relations based on a novel tagging scheme. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1227-1236, Vancouver, Canada. Association for Computational Linguistics.", + "Zexuan Zhong and Danqi Chen. 2021. A frustratingly easy approach for entity and relation extraction. In" + ], + "bbox": [ + 510, + 85, + 884, + 917 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1355", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 50-61, Online. Association for Computational Linguistics.", + "Yikai Zhou, Baosong Yang, Derek F. Wong, Yu Wan, and Lidia S. Chao. 2020. Uncertainty-aware curriculum learning for neural machine translation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6934-6944, Online. Association for Computational Linguistics." + ], + "bbox": [ + 115, + 85, + 489, + 256 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1356", + "bbox": [ + 482, + 928, + 521, + 940 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A For every submission:", + "bbox": [ + 115, + 107, + 322, + 122 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "A1. Did you describe the limitations of your work? Section 7", + "A2. Did you discuss any potential risks of your work? We study open-domain information extraction for researches in this area", + "A3. Do the abstract and introduction summarize the paper's main claims? Abstract and Section 1", + "A4. Have you used AI writing assistants when working on this paper? Left blank." + ], + "bbox": [ + 129, + 126, + 695, + 287 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "B Did you use or create scientific artifacts?", + "bbox": [ + 114, + 300, + 489, + 316 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 132, + 321, + 215, + 336 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "B1. Did you cite the creators of artifacts you used? No response.", + "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? No response.", + "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? No response.", + "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? No response.", + "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? No response.", + "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. No response." + ], + "bbox": [ + 127, + 347, + 880, + 753 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "C Did you run computational experiments?", + "bbox": [ + 114, + 764, + 492, + 781 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Section 5", + "bbox": [ + 132, + 785, + 205, + 800 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? No response.", + "bbox": [ + 127, + 813, + 880, + 860 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "ACL 2023 Responsible NLP Checklist", + "bbox": [ + 132, + 84, + 433, + 99 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.", + "bbox": [ + 112, + 868, + 877, + 892 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1357", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?", + "bbox": [ + 129, + 83, + 878, + 115 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Section 5", + "bbox": [ + 149, + 117, + 221, + 130 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?", + "bbox": [ + 127, + 143, + 878, + 190 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 192, + 248, + 206 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?", + "bbox": [ + 129, + 217, + 878, + 263 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Section 5", + "bbox": [ + 149, + 267, + 221, + 280 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?", + "bbox": [ + 112, + 293, + 877, + 309 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 132, + 313, + 213, + 329 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?", + "bbox": [ + 127, + 340, + 878, + 370 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 374, + 248, + 388 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?", + "bbox": [ + 127, + 399, + 878, + 445 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 449, + 248, + 463 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?", + "bbox": [ + 127, + 475, + 878, + 521 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 524, + 248, + 538 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?", + "bbox": [ + 127, + 549, + 872, + 564 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 567, + 248, + 581 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?", + "bbox": [ + 127, + 592, + 878, + 621 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 626, + 248, + 640 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1358", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Uncertainty-Aware Bootstrap Learning for Joint Extraction on Distantly-Supervised Data/a0f755f8-19b9-4154-a835-8d33c06b518f_model.json b/2023/Uncertainty-Aware Bootstrap Learning for Joint Extraction on Distantly-Supervised Data/a0f755f8-19b9-4154-a835-8d33c06b518f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e4e1e459415187c3046eb9886e3a6d92d642d1a8 --- /dev/null +++ b/2023/Uncertainty-Aware Bootstrap Learning for Joint Extraction on Distantly-Supervised Data/a0f755f8-19b9-4154-a835-8d33c06b518f_model.json @@ -0,0 +1,1925 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.175, + 0.09, + 0.825, + 0.131 + ], + "angle": 0, + "content": "Uncertainty-Aware Bootstrap Learning for Joint Extraction on Distantly-Supervised Data" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.149, + 0.771, + 0.166 + ], + "angle": 0, + "content": "Yufei Li\\(^{1}\\), Xiao Yu\\(^{2}\\), Yanchi Liu\\(^{3}\\), Haifeng Chen\\(^{3}\\), Cong Liu\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.207, + 0.167, + 0.796, + 0.183 + ], + "angle": 0, + "content": "1University of California, Riverside 2Stellar Cyber 3NEC Labs America" + }, + { + "type": "text", + "bbox": [ + 0.28, + 0.184, + 0.725, + 0.2 + ], + "angle": 0, + "content": "\\(^{1}\\{yli927,congl\\} @ucr.edu,\\) \\(^{2}\\mathrm{xyu@stellarcyber.ai},\\)" + }, + { + "type": "text", + "bbox": [ + 0.351, + 0.2, + 0.653, + 0.216 + ], + "angle": 0, + "content": "3{yanchi,haifeng}@nec-labs.com" + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.253, + 0.341, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.275, + 0.461, + 0.614 + ], + "angle": 0, + "content": "Jointly extracting entity pairs and their relations is challenging when working on distantly-supervised data with ambiguous or noisy labels. To mitigate such impact, we propose uncertainty-aware bootstrap learning, which is motivated by the intuition that the higher uncertainty of an instance, the more likely the model confidence is inconsistent with the ground truths. Specifically, we first explore instance-level data uncertainty to create an initial high-confident examples. Such subset serves as filtering noisy instances and facilitating the model to converge fast at the early stage. During bootstrap learning, we propose self-ensembling as a regularizer to alleviate inter-model uncertainty produced by noisy labels. We further define probability variance of joint tagging probabilities to estimate inner-model parametric uncertainty, which is used to select and build up new reliable training instances for the next iteration. Experimental results on two large datasets reveal that our approach outperforms existing strong baselines and related methods." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.622, + 0.26, + 0.637 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.646, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Joint extraction involves extracting multiple types of entities and relations between them using a single model, which is necessary in automatic knowledge base construction (Yu et al., 2020). One way to cheaply acquire a large amount of labeled data for training joint extraction models is through distant supervision (DS) (Mintz et al., 2009). DS involves aligning a knowledge base (KB) with an unlabeled corpus using hand-crafted rules or logic constraints. Due to the lack of human annotators, DS brings a large proportion of noisy labels, e.g., over \\(30\\%\\) noisy instances in some cases (Mintz et al., 2009), making it impossible to learn useful features. The noise can be either false relations due to the aforementioned rule-based matching assumption or wrong entity tags due to limited coverage over entities in open-domain KBs." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.254, + 0.886, + 0.43 + ], + "angle": 0, + "content": "Existing distantly-supervised approaches model noise relying either on heuristics such as reinforcement learning (RL) (Nooralahzadeh et al., 2019; Hu et al., 2021) and adversarial learning (Chen et al., 2021), or pattern-based methods (Jia et al., 2019; Shang et al., 2022) to select trustable instances. Nevertheless, these methods require designing heuristics or hand-crafted patterns which may encourage a model to leverage spurious features without considering the confidence or uncertainty of its predictions." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.438, + 0.885, + 0.919 + ], + "angle": 0, + "content": "In response to these problems, we propose UnBED—Uncertainty-aware Bootstrap learning for joint Extraction on Distantly-supervised data. UnBED assumes that 1) low data uncertainty indicates reliable instances using a pre-trained language model (PLM) in the initial stage, 2) model should be aware of trustable entity and relation labels regarding its uncertainty after training. Our bootstrap serves uncertainty as a principle to mitigate the impact of noise labels on model learning and validate input sequences to control the number of training examples in each step. Particularly, we quantify data uncertainty of an instance according to its winning score (Hendrycks and Gimpel, 2017) and entropy (Shannon, 1948). We define averaged maximum probability that is estimated by a joint PLM over each token in a sequence to adapt previous techniques in joint extraction scheme. Instances with low data uncertainty are collected to form an initial subset, which is used to tune the joint PLM tagger and facilitate fast convergence. Then, we define parametric uncertainty in two perspectives—inter-model and inner-model uncertainty. The former is quantified by self-ensembling (Wang and Wang, 2022) and serves as a regularizer to improve model robustness against noisy labels during training. The latter is captured by probability variance in MC Dropout (Gal and Ghahramani, 2016) for selecting new confident instances for the next training iteration. Such two" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1349" + }, + { + "type": "footer", + "bbox": [ + 0.227, + 0.946, + 0.772, + 0.958 + ], + "angle": 0, + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + }, + { + "type": "footer", + "bbox": [ + 0.369, + 0.959, + 0.631, + 0.972 + ], + "angle": 0, + "content": "Volume 2: Short Papers, pages 1349-1358" + }, + { + "type": "footer", + "bbox": [ + 0.297, + 0.973, + 0.702, + 0.986 + ], + "angle": 0, + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.114, + 0.085, + 0.49, + 0.135 + ], + "angle": 0, + "content": "fold model uncertainties reinforce with each other to guide the model to iteratively improve its robustness and learn from reliable knowledge." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.149, + 0.271, + 0.163 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.177, + 0.49, + 0.45 + ], + "angle": 0, + "content": "Joint Extraction Methods Joint extraction detects entities and their relations using a single model, which effectively integrates the information from both sources and therefore achieves better results in both subtasks compared to pipelined methods (Zheng et al., 2017). For example, unified methods tag entities and relation simultaneously, e.g., (Zheng et al., 2017) proposes a novel tagging scheme which converts joint extraction to a sequence labeling problem; (Dai et al., 2019) introduces query position and sequential tagging to extract overlapping relations. Such methods avoid producing redundant information compared to parameter-sharing neural models (Miwa and Bansal, 2016; Gupta et al., 2016), and require no hand-crafted features that are used in structured systems (Yu et al., 2020)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.452, + 0.49, + 0.693 + ], + "angle": 0, + "content": "To address the challenge of learning from DS, pre-trained transformers (e.g., BERT, GPT-2) have gain much attention. They model strong expressive context-aware representations for text sequence through multiple attention layers, and achieve state-of-the-art performance on various NLP tasks (Radford et al., 2019; Devlin et al., 2019; Li et al., 2022). They can be cheaply fine-tuned to solve different downstream tasks including NER and RC. Specifically, BERT is trained on large English corpus using masked language modeling. The multi-head attention weights indicate interactions between each pair of words and its hidden states integrate semantic information of the whole sentence, which are used to decode different tagging results." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.695, + 0.49, + 0.92 + ], + "angle": 0, + "content": "Uncertainty Methods Uncertainty generally comes from two sources—aleatoric uncertainty and epistemic uncertainty. The former is also referred to as data uncertainty, describing noise inherent in the data generation. Methods mitigating such uncertainty include data interpolation (Dong et al., 2018), winning score, and temperature scale (Guo et al., 2017). The latter is also called model uncertainty, describing whether the structure choice and model parameters best describe the data distribution. One main solution to mitigate model uncertainty is Bayesian Neural Network (BNN) (Klein et al., 2017) that puts a prior distribution on its weights. To save computational cost, Monte Carlo" + }, + { + "type": "image", + "bbox": [ + 0.59, + 0.082, + 0.807, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.296, + 0.882, + 0.311 + ], + "angle": 0, + "content": "Figure 1: Joint extraction as a token classification task." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.328, + 0.885, + 0.473 + ], + "angle": 0, + "content": "dropout is proposed as an approximation of variational Bayesian inference (Gal and Ghahramani, 2016), realized by training models with dropout layers and testing with stochastic inference to quantify probability variance. Besides BNN, self-ensembling (Wang and Wang, 2022) which measures the outputs variance between models with the same architecture has been shown effective to reduce parametric uncertainty across models." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.488, + 0.746, + 0.503 + ], + "angle": 0, + "content": "3 Joint Extraction Model" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.516, + 0.885, + 0.71 + ], + "angle": 0, + "content": "Tagging Scheme For an input sequence \\(\\mathcal{X} = \\{x_1, \\dots, x_n\\}\\), we tag \\(n\\) sequences according to different query position \\(p\\) following (Dai et al., 2019). If \\(p\\) is the start of an entity (query entity \\(e_1\\)), the sequence is an instance. The entity type is labeled at \\(p\\) and other entities \\(e_2\\) which have relationship with the query entity are labeled with relation types \\(re\\). The rest of tokens are labeled \"O\" (Outside), meaning they do not correspond to the query entity. Accordingly, we convert joint extraction into a token classification task and extract relation triplets \\(\\{e_1, re, e_2\\}\\) in each instance, as shown in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.711, + 0.887, + 0.92 + ], + "angle": 0, + "content": "Position-Attentive Encoder we use BERT (Devlin et al., 2019) to encode a sentence \\(\\mathcal{X}\\) into token-level representations \\(h = \\{h_1,..,h_n\\}\\), where \\(h_i\\in \\mathbb{R}^d\\) is a \\(d\\)-dimensional vector corresponding to the \\(i\\)-th token in \\(\\mathcal{X}\\). For each query \\(p\\), self-matching is applied to calculate the position-attention \\(\\boldsymbol{a}_t\\in \\mathbb{R}^T\\) between token at \\(p\\) and each token at target position \\(t\\), which compares the sentence representations against itself to collect context information (Tan et al., 2018). The produced position-aware representation \\(\\boldsymbol{c}_t\\in \\mathbb{R}^{T\\times d}\\) is an attention-weighted sentence vector \\(\\boldsymbol{c}_t = \\boldsymbol{a}_t^\\top \\boldsymbol{h}\\). Finally, we concatenate \\(\\boldsymbol{h}_t\\) and \\(\\boldsymbol{c}_t\\) to generate position-aware and context" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.928, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1350" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.114, + 0.085, + 0.381, + 0.101 + ], + "angle": 0, + "content": "aware representations \\(\\pmb{u}_t = [\\pmb{h}_t|\\pmb{c}_t]\\)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.101, + 0.49, + 0.295 + ], + "angle": 0, + "content": "CRF Decoder (Lafferty et al., 2001) For each position-aware representation \\( \\boldsymbol{u}_t \\), we first learn a linear transformation \\( \\boldsymbol{z}_t = \\boldsymbol{W}\\boldsymbol{u}_t \\in \\mathbb{R}^C \\) to represent tag scores for the \\( t \\)-th token. Here \\( C \\) is the number of distinct tags. For an instance with labels \\( \\boldsymbol{y} = \\{y_1, \\dots, y_n\\} \\), the decoding score \\( s(\\boldsymbol{z}, \\boldsymbol{y}) \\) is the sum of transition scores from tag \\( y_t \\) to tag \\( y_{t+1} \\) plus the input score \\( z_t^{yt} \\). The conditional probability \\( p(\\boldsymbol{y}|\\boldsymbol{z}) \\) is the softmax over \\( s(\\boldsymbol{z}, \\boldsymbol{y}) \\) for all possible label sequences \\( \\boldsymbol{y}' \\). We maximize the log-likelihood of correct tag sequences during training \\( \\mathcal{L}_c = \\sum \\log p(\\boldsymbol{y}|\\boldsymbol{z}) \\)." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.306, + 0.489, + 0.323 + ], + "angle": 0, + "content": "4 Uncertainty-Aware Bootstrap Learning" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.331, + 0.49, + 0.509 + ], + "angle": 0, + "content": "Motivation One of the main challenges in bootstrap learning is to evaluate the \"correctness\" of a labeled instance. We consider this problem from an uncertainty perspective and assume instances with lower uncertainty are more likely to be correctly labeled. In this section, we first propose instance-level data uncertainty which is used to filter noisy examples and build an initial subset. Then, we introduce our two-fold model uncertainties which helps iteratively mitigate DS effect and build up trustable examples during bootstrap learning." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.519, + 0.298, + 0.535 + ], + "angle": 0, + "content": "4.1 Data Uncertainty" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.54, + 0.49, + 0.732 + ], + "angle": 0, + "content": "Presenting examples in an easy-to-hard order at different training stages can benefit models (Platanios et al., 2019; Zhou et al., 2020), we propose data uncertainty as a way to quantify the \"hardness\" of an instance. To better estimate the data uncertainty, we use pre-trained language models (PLMs) to generate tag probability for each token in a sequence. Our intuition is that higher uncertain inputs are \"harder\" to be generated by a PLM, as it already has rationales of language. Accordingly, we propose two data uncertainties, which can be used individually or combined together:" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.733, + 0.49, + 0.83 + ], + "angle": 0, + "content": "Winning Score (WS) The maximum softmax probability reflects data uncertainty of an input (Hendrycks and Gimpel, 2017). Given an input instance \\(\\mathcal{I} = \\{x_1,\\dots,x_n\\}\\), we define data uncertainty \\(u^{d}(\\mathcal{I})\\) as the minus averaged token classification winning score:" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.857, + 0.489, + 0.877 + ], + "angle": 0, + "content": "\\[\nu ^ {d} (\\mathcal {I}) = - \\frac {1}{n} \\sum_ {t = 1} ^ {n} \\max _ {c \\in [ 1, C ]} P (y _ {t} = c | x _ {t}) \\quad (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.888, + 0.489, + 0.919 + ], + "angle": 0, + "content": "Entropy Shannon entropy (Shannon, 1948) is widely used to reflect information uncertainty. We" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.883, + 0.117 + ], + "angle": 0, + "content": "propose data uncertainty \\( u^{d}(\\mathcal{I}) \\) as the averaged token classification entropy:" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.148, + 0.883, + 0.181 + ], + "angle": 0, + "content": "\\[\nu ^ {d} (\\mathcal {I}) = \\frac {1}{n} \\sum_ {t = 1} ^ {n} \\sum_ {c = 1} ^ {C} P \\left(y _ {t} = c \\mid x _ {t}\\right) \\log P \\left(y _ {t} = c \\mid x _ {t}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.183, + 0.885, + 0.28 + ], + "angle": 0, + "content": "We filter out examples with high uncertainty scores and build an initial subset with \"simple\" examples. At the early training stage, a model is not aware of what a decent distribution \\( P(y|x) \\) should be, thus data uncertainty facilitates it to converge fast by tuning on a fairly \"simple\" subset." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.293, + 0.706, + 0.308 + ], + "angle": 0, + "content": "4.2 Model Uncertainty" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.315, + 0.885, + 0.589 + ], + "angle": 0, + "content": "In our bootstrap learning, we define model uncertainty, i.e., epistemic uncertainty (Kendall and Gal, 2017), to measure whether model parameters can best describe the data distribution following (Zhou et al., 2020). A small model uncertainty indicates the model is confident that the current training data has been well learned (Wang et al., 2019). We adopt Monte Carlo Dropout (Gal and Ghahramani, 2016) to approximate Bayesian inference which captures inner-model parametric uncertainty. Specifically, we perform \\( K \\) forward passes through our joint model. In each pass, part of network neurons \\( \\theta \\) are randomly deactivated. Finally, we yield \\( K \\) samples on model parameters \\( \\{\\hat{\\theta}_1,\\dots,\\hat{\\theta}_K\\} \\). We use the averaged token classification Probability Variance (PV) (Shelmanov et al., 2021) over all tags for instance \\( \\mathcal{I} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.616, + 0.883, + 0.657 + ], + "angle": 0, + "content": "\\[\nu ^ {m} (\\theta) = \\frac {1}{n} \\sum_ {t = 1} ^ {n} \\sum_ {c = 1} ^ {C} \\operatorname {V a r} \\left[ P \\left(y _ {t} = c \\mid x _ {t}, \\hat {\\theta} _ {k}\\right) \\right] _ {k = 1} ^ {K} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.659, + 0.884, + 0.74 + ], + "angle": 0, + "content": "where \\(\\operatorname{Var}[\\cdot]\\) is the variance of distribution over the \\(K\\) passes following the common settings in (Dong et al., 2018; Xiao and Wang, 2019). Accordingly, model is aware of its confidence over each instance and how likely the label is noisy." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.753, + 0.696, + 0.769 + ], + "angle": 0, + "content": "4.3 Training Strategy" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.775, + 0.885, + 0.919 + ], + "angle": 0, + "content": "Uncertainty-Aware Loss Besides MC Dropout which measures parametric uncertainty within a model, we also consider mitigating parametric uncertainty between models to stabilize the weights during training. Specifically, we use self-ensembling (He et al., 2020; Wang and Wang, 2022) to calculate the loss between the same models to improve model robustness and reduce the label noise effect on model performance." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.928, + 0.519, + 0.941 + ], + "angle": 0, + "content": "1351" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.116, + 0.084, + 0.362, + 0.1 + ], + "angle": 0, + "content": "Algorithm 1 Bootstrap Learning" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.105, + 0.487, + 0.139 + ], + "angle": 0, + "content": "Input: Original dataset \\(\\mathcal{D} = \\{(\\mathcal{I}^n,y^n)\\}_{n = 1}^N\\) two joint models \\(f_{1},f_{2}\\) with parameters \\(\\theta_{1},\\theta_{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.14, + 0.489, + 0.169 + ], + "angle": 0, + "content": "1: Compute data uncertainty \\( u^{d}(\\mathcal{I}) \\) for each instance \\( \\mathcal{I} \\) in \\( \\mathcal{D} \\);" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.172, + 0.488, + 0.203 + ], + "angle": 0, + "content": "2: Initial dataset \\(\\mathcal{C} \\gets\\) Select data pairs \\((\\mathcal{I}^n, y^n)\\) such that \\(u^d(\\mathcal{I}) < \\tau^d\\) from \\(\\mathcal{D}\\);" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.205, + 0.317, + 0.219 + ], + "angle": 0, + "content": "3: for epoch \\( e = 1, \\ldots \\) do" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.221, + 0.408, + 0.235 + ], + "angle": 0, + "content": "4: Train \\( f_{1}, f_{2} \\) on \\( \\mathcal{C} \\) using Eq. (5);" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.237, + 0.488, + 0.251 + ], + "angle": 0, + "content": "5: Calculate model uncertainty \\( u^{m}(\\theta_{1}) \\) on \\( \\mathcal{D} \\);" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.253, + 0.488, + 0.283 + ], + "angle": 0, + "content": "6: \\(\\mathcal{C} \\gets\\) Select data pairs \\((\\mathcal{I}^n, y^n)\\) such that \\(u^m(\\mathcal{I}; \\theta_1) < \\tau^m\\) from \\(\\mathcal{D}\\);" + }, + { + "type": "list", + "bbox": [ + 0.126, + 0.14, + 0.489, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.319, + 0.489, + 0.4 + ], + "angle": 0, + "content": "We create another joint model with identical framework, e.g., architecture, loss functions, hyperparameters, and compute a self-ensemble loss \\(\\mathcal{L}_e\\) to minimize the difference between two outputs from the two models regarding the same inputs:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.425, + 0.488, + 0.447 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {e} = \\sum K L (f (\\mathcal {I}; \\theta_ {1}), f (\\mathcal {I}; \\theta_ {2})) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.466, + 0.489, + 0.545 + ], + "angle": 0, + "content": "where \\(KL(.)\\) is the Kullback-Leibler divergence between two probabilistic distributions, \\(\\theta_{1},\\theta_{2}\\) denote the parameters of first and second models. We formulate our final uncertainty-aware objective \\(\\mathcal{L}\\) as the sum of CRF and self-ensemble loss:" + }, + { + "type": "equation", + "bbox": [ + 0.243, + 0.575, + 0.488, + 0.59 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {c} + \\alpha \\mathcal {L} _ {e} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.611, + 0.489, + 0.642 + ], + "angle": 0, + "content": "where \\(\\alpha\\) denotes the weight of self-ensembling, and \\(\\mathcal{L}_c\\) means the token classification loss." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.646, + 0.489, + 0.919 + ], + "angle": 0, + "content": "Bootstrap Learning Procedure To mitigate the DS effect on model performance, we propose a twofold bootstrap learning strategy (see Algorithm 1). Specifically, we first apply data uncertainty to filter \"harder\" examples and redistribute a reliable initial training data \\(\\mathcal{M}\\). Then, we iteratively feed examples following an easy-to-hard order to the model. In each training iteration, we regularize the joint model with self-ensembling loss to reduce the impact of noisy labels on model parameters. Then we use probability variance to select new confident training instances \\(\\mathcal{D}'\\) that can be explained by the model as the next training inputs. The more certain examples are selected, the more likely the model will learn beneficial information and will converge faster. We repeat the above procedure until the F1 score on the validation set converges." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.084, + 0.655, + 0.101 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.11, + 0.602, + 0.126 + ], + "angle": 0, + "content": "5.1 Setup" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.131, + 0.884, + 0.291 + ], + "angle": 0, + "content": "We evaluate the performance of UnBED on two datasets, NYT and Wiki-KBP. The NYT (Riedel et al., 2010) dataset collects news from New York Times and its training data is automatically labeled by DS. We use the revised test dataset (Jia et al., 2019) that is manually annotated to ensure quality. The Wiki-KBP (Ling and Weld, 2012) dataset collects articles from Wikipedia. Its training data is labeled by DS (Liu et al., 2017), and the test set is manually annotated (Ellis et al., 2013)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.292, + 0.884, + 0.452 + ], + "angle": 0, + "content": "We compare UnBED with the following baselines: ARNOR (Jia et al., 2019), a pattern-based method to reduce noise for distantly-supervised triplet extraction. PURE (Zhong and Chen, 2021), a pipeline approach that uses pre-trained BERT entity model to first recognize entities and then employs a relation model to detect underlying relations. FAN (Hao et al., 2021), an adversarial method including a transformers encoder to reduce noise for distantly-supervised triplet extraction." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.454, + 0.883, + 0.549 + ], + "angle": 0, + "content": "Evaluation We evaluate the extracted triplets for each sentence based on Precision (Prec.), Recall (Rec.), and F1. A triplet \\(\\{e_1, re, e_2\\}\\) is marked correct if the relation type \\(re\\), two entities \\(e_1, e_2\\) are all correct. We build a validation set by randomly sampling \\(10\\%\\) sentences from the test set." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.55, + 0.884, + 0.742 + ], + "angle": 0, + "content": "Implementation Details We use Hugging Face bert-large-uncased (Devlin et al., 2019) pre-trained model as backbone. For ARNOR, the hidden vector size is set to 300. In regularization training, we find optimal parameters \\(\\alpha\\) as 1 for both datasets. We implement UnBED and all baselines in PyTorch, with Adam optimizer, initial learning rate \\(10^{-5}\\), dropout rate 0.1, and batch size 8. For initial subset configuration, we choose data uncertainty threshold 0.5. For bootstrap learning, an empirical model uncertainty threshold is set to 0.6 with the best validation F1." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.754, + 0.677, + 0.768 + ], + "angle": 0, + "content": "5.2 Overall Results" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.775, + 0.884, + 0.919 + ], + "angle": 0, + "content": "As shown in Table 1, UnBED significantly outperforms all baselines in precision and F1 metric. Specifically, UnBED achieves \\(8\\%\\) F1 improvement on NYT (3% on Wiki-KBP) over denoising approaches—ARNOR and FAN. Our approach also outperforms baselines using pretrained transformers (PURE and FAN), showing that uncertainty-aware bootstrap learning effectively reduces the impact of noisy labels." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.928, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1352" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.187, + 0.082, + 0.813, + 0.2 + ], + "angle": 0, + "content": "
MethodNYTWiki-KBP
Prec.Rec.F1Prec.Rec.F1
ARNOR (Jia et al., 2019)0.5880.6140.6000.4020.4710.434
PURE (Zhong and Chen, 2021)0.5360.6640.5930.3950.4330.413
FAN (Hao et al., 2021)0.5790.6460.6110.3910.4670.426
UnBED-WS0.6620.7300.6940.4290.5010.462
UnBED-Entropy0.6510.7410.6930.4220.5090.461
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.209, + 0.883, + 0.239 + ], + "angle": 0, + "content": "Table 1: Evaluation results on NYT and Wiki-KBP datasets. Bold numbers denote the best metrics. UnBED-WS and UnBED-Entropy denote UnBED with winning score and entropy as the data uncertainty, respectively." + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.263, + 0.472, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.423, + 0.49, + 0.468 + ], + "angle": 0, + "content": "Figure 2: F1 score vs. Epochs under different settings. Vanilla-PV-enssembled denotes UnBED-WS, and entropy-PV-enssembled denotes UnBED-Entropy." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.492, + 0.295, + 0.508 + ], + "angle": 0, + "content": "5.3 Further Analysis" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.512, + 0.49, + 0.691 + ], + "angle": 0, + "content": "We analyze the functionality of different components in Figure 2. We observe that both the entropy-PV and vanilla-PV outperform the baseline (joint model directly trained on the original DS dataset) in terms of F1 \\((5\\sim 7\\%)\\) increase), demonstrating the effect of filtering noisy labels and selecting trustable instance using probability variance. Besides, self-ensembling further enhances the performance in later training stage \\((2\\sim 4\\) F1 increase), proving that mitigating the inter-model uncertainty benefits model robustness against noisy labels." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.702, + 0.254, + 0.716 + ], + "angle": 0, + "content": "6 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.727, + 0.491, + 0.919 + ], + "angle": 0, + "content": "In this paper, we propose a novel uncertainty-aware bootstrap learning framework for distantly-supervised joint extraction. Specifically, we define data uncertainty in generally token classification to filter out highly-error-prone instances and build an initial high-confident subset, which is used to tune the joint extraction model for fast convergence. We then propose a two-fold bootstrap learning procedure which iteratively mitigates the DS impact on model robustness and selects new trustable training instances. Experimental results on two benchmark datasets show that UnBED significantly out" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.264, + 0.785, + 0.28 + ], + "angle": 0, + "content": "performs other denoising techniques." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.291, + 0.615, + 0.306 + ], + "angle": 0, + "content": "Limitations" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.316, + 0.885, + 0.525 + ], + "angle": 0, + "content": "In this work we propose an uncertainty-aware bootstrap learning framework for joint extraction. Though it achieves state-of-the-art performance compared to other denoising techniques, UnBED requires large training resources considering the ensemble loss calculated between two large PLMs and the probability variance calculated on the PLM joint extraction model. In our future work, we hope to incorporate pruning techniques during training to improve the efficiency. We will also consider more complex relations between entities, e.g., relations beyond the sentence boundary, to fit in real-world information extraction scenarios." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.537, + 0.682, + 0.554 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.562, + 0.884, + 0.61 + ], + "angle": 0, + "content": "This work was supported by NSF CNS 2135625, CPS 2038727, CNS Career 1750263, and a Darpa Shell grant." + }, + { + "type": "title", + "bbox": [ + 0.511, + 0.637, + 0.61, + 0.653 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.659, + 0.885, + 0.792 + ], + "angle": 0, + "content": "Tao Chen, Haochen Shi, Liyuan Liu, Siliang Tang, Jian Shao, Zhigang Chen, and Yueting Zhuang. 2021. Empower distantly supervised relation extraction with collaborative adversarial training. In Thirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2021, Thirty-Third Conference on Innovative Applications of Artificial Intelligence, IAAI 2021, The Eleventh Symposium on Educational Advances in Artificial Intelligence, EAAI 2021, Virtual Event, February 2-9, 2021, pages 12675-12682. AAAI Press." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.8, + 0.887, + 0.919 + ], + "angle": 0, + "content": "Dai Dai, Xinyan Xiao, Yajuan Lyu, Shan Dou, Qiaoqiao She, and Haifeng Wang. 2019. Joint extraction of entities and overlapping relations using position-attentive sequence labeling. In The Thirty-Third AAAI Conference on Artificial Intelligence, AAAI 2019, The Thirty-First Innovative Applications of Artificial Intelligence Conference, IAAI 2019, The Ninth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2019, Honolulu, Hawaii," + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.659, + 0.887, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1353" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.135, + 0.086, + 0.49, + 0.113 + ], + "angle": 0, + "content": "USA, January 27 - February 1, 2019, pages 6300-6308. AAAI Press." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.123, + 0.49, + 0.241 + ], + "angle": 0, + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.252, + 0.49, + 0.331 + ], + "angle": 0, + "content": "Li Dong, Chris Quirk, and Mirella Lapata. 2018. Confidence modeling for neural semantic parsing. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 743-753, Melbourne, Australia. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.341, + 0.49, + 0.432 + ], + "angle": 0, + "content": "Joe Ellis, Jeremy Getman, Justin Mott, Xuansong Li, Kira Griffith, Stephanie M. Strassel, and Jonathan Wright. 2013. Linguistic resources for 2013 knowledge base population evaluations. In Proceedings of the Sixth Text Analysis Conference, TAC 2013, Gaithersburg, Maryland, USA, November 18-19, 2013. NIST." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.443, + 0.49, + 0.536 + ], + "angle": 0, + "content": "Yarin Gal and Zoubin Ghahramani. 2016. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In Proceedings of the 33nd International Conference on Machine Learning, ICML 2016, New York City, NY, USA, June 19-24, 2016, volume 48 of JMLR Workshop and Conference Proceedings, pages 1050-1059. JMLR.org." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.545, + 0.49, + 0.637 + ], + "angle": 0, + "content": "Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. 2017. On calibration of modern neural networks. In Proceedings of the 34th International Conference on Machine Learning, ICML 2017, Sydney, NSW, Australia, 6-11 August 2017, volume 70 of Proceedings of Machine Learning Research, pages 1321-1330. PMLR." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.648, + 0.49, + 0.74 + ], + "angle": 0, + "content": "Pankaj Gupta, Hinrich Schütze, and Bernt Andrassy. 2016. Table filling multi-task recurrent neural network for joint entity and relation extraction. In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 2537-2547, Osaka, Japan. The COLING 2016 Organizing Committee." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.75, + 0.49, + 0.843 + ], + "angle": 0, + "content": "Kailong Hao, Botao Yu, and Wei Hu. 2021. Knowing false negatives: An adversarial training method for distantly supervised relation extraction. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 9661-9672, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.853, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Jianfeng He, Xuchao Zhang, Shuo Lei, Zhiqian Chen, Fanglan Chen, Abdulaziz Alhamadani, Bei Xiao, and ChangTien Lu. 2020. Towards more accurate uncertainty estimation in text classification. In Proceedings of the 2020 Conference on Empirical Methods" + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.529, + 0.086, + 0.882, + 0.126 + ], + "angle": 0, + "content": "in Natural Language Processing (EMNLP), pages 8362-8372, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.139, + 0.885, + 0.219 + ], + "angle": 0, + "content": "Dan Hendrycks and Kevin Gimpel. 2017. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.232, + 0.885, + 0.349 + ], + "angle": 0, + "content": "Xuming Hu, Chenwei Zhang, Yawen Yang, Xiaohe Li, Li Lin, Lijie Wen, and Philip S. Yu. 2021. Gradient imitation reinforcement learning for low resource relation extraction. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, EMNLP 2021, Virtual Event / Punta Cana, Dominican Republic, 7-11 November; 2021, pages 2737-2746. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.363, + 0.885, + 0.456 + ], + "angle": 0, + "content": "Wei Jia, Dai Dai, Xinyan Xiao, and Hua Wu. 2019. ARNOR: Attention regularization based noise reduction for distant supervision relation classification. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1399-1408, Florence, Italy. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.469, + 0.885, + 0.549 + ], + "angle": 0, + "content": "Alex Kendall and Yarin Gal. 2017. What uncertainties do we need in bayesian deep learning for computer vision? In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 5574-5584." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.562, + 0.885, + 0.641 + ], + "angle": 0, + "content": "Aaron Klein, Stefan Falkner, Jost Tobias Springenberg, and Frank Hutter. 2017. Learning curve prediction with bayesian neural networks. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.655, + 0.885, + 0.746 + ], + "angle": 0, + "content": "John D. Lafferty, Andrew McCallum, and Fernando C. N. Pereira. 2001. Conditional random fields: Probabilistic models for segmenting and labeling sequence data. In Proceedings of the Eighteenth International Conference on Machine Learning, ICML '01, page 282-289, San Francisco, CA, USA. Morgan Kaufmann Publishers Inc." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.76, + 0.885, + 0.853 + ], + "angle": 0, + "content": "Shuyang Li, Yufei Li, Jianmo Ni, and Julian McAuley. 2022. SHARE: a system for hierarchical assistive recipe editing. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 11077-11090, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.866, + 0.885, + 0.918 + ], + "angle": 0, + "content": "Xiao Ling and Daniel S. Weld. 2012. Fine-grained entity recognition. In Proceedings of the Twenty-Sixth AAAI Conference on Artificial Intelligence, July 22-26, 2012, Toronto, Ontario, Canada. AAAI Press." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.885, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1354" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.179 + ], + "angle": 0, + "content": "Liyuan Liu, Xiang Ren, Qi Zhu, Shi Zhi, Huan Gui, Heng Ji, and Jiawei Han. 2017. Heterogeneous supervision for relation extraction: A representation learning approach. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 46-56, Copenhagen, Denmark. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.189, + 0.49, + 0.294 + ], + "angle": 0, + "content": "Mike Mintz, Steven Bills, Rion Snow, and Daniel Jurafsky. 2009. Distant supervision for relation extraction without labeled data. In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP, pages 1003-1011, Suntec, Singapore. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.304, + 0.49, + 0.396 + ], + "angle": 0, + "content": "Makoto Miwa and Mohit Bansal. 2016. End-to-end relation extraction using lstms on sequences and tree structures. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, ACL 2016, August 7-12, 2016, Berlin, Germany, Volume 1: Long Papers. The Association for Computer Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.406, + 0.49, + 0.512 + ], + "angle": 0, + "content": "Farhad Nooralahzadeh, Jan Tore Lönning, and Lilja Øvrelid. 2019. Reinforcement-based denoising of distantly supervised NER with partial annotation. In Proceedings of the 2nd Workshop on Deep Learning Approaches for Low-Resource NLP, DeepLo@EMNLP-IJCNLP 2019, Hong Kong, China, November 3, 2019, pages 225–233. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.522, + 0.49, + 0.653 + ], + "angle": 0, + "content": "Emmanouil Antonios Platanios, Otilia Stretcu, Graham Neubig, Barnabás Póczos, and Tom M. Mitchell. 2019. Competence-based curriculum learning for neural machine translation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 1162-1172. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.663, + 0.489, + 0.703 + ], + "angle": 0, + "content": "Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.713, + 0.49, + 0.819 + ], + "angle": 0, + "content": "Sebastian Riedel, Limin Yao, and Andrew McCallum. 2010. Modeling relations and their mentions without labeled text. In Machine Learning and Knowledge Discovery in Databases, European Conference, ECML PKDD 2010, Barcelona, Spain, September 20-24, 2010, Proceedings, Part III, volume 6323 of Lecture Notes in Computer Science, pages 148-163. Springer." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.829, + 0.49, + 0.881 + ], + "angle": 0, + "content": "Yuming Shang, Heyan Huang, Xin Sun, Wei Wei, and Xian-Ling Mao. 2022. A pattern-aware self-attention network for distant supervised relation extraction. Inf. Sci., 584:269-279." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.892, + 0.489, + 0.918 + ], + "angle": 0, + "content": "Claude E. Shannon. 1948. A mathematical theory of communication. Bell Syst. Tech. J., 27(3):379-423." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.086, + 0.885, + 0.179 + ], + "angle": 0, + "content": "Artem Shelmanov, Evgenii Tsymbalov, Dmitri Puzyrev, Kirill Fedyanin, Alexander Panchenko, and Maxim Panov. 2021. How certain is your Transformer? In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 1833-1840, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.188, + 0.885, + 0.293 + ], + "angle": 0, + "content": "Zhixing Tan, Mingxuan Wang, Jun Xie, Yidong Chen, and Xiaodong Shi. 2018. Deep semantic role labeling with self-attention. In Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence and Thirtieth Innovative Applications of Artificial Intelligence Conference and Eighth AAAI Symposium on Educational Advances in Artificial Intelligence, AAAI'18/IAAI'18/EAAI'18. AAAI Press." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.303, + 0.885, + 0.37 + ], + "angle": 0, + "content": "Hongjun Wang and Yisen Wang. 2022. Self-ensemble adversarial training for improved robustness. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.379, + 0.885, + 0.498 + ], + "angle": 0, + "content": "Shuo Wang, Yang Liu, Chao Wang, Huanbo Luan, and Maosong Sun. 2019. Improving back-translation with uncertainty-based confidence estimation. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLP-IJCNLP 2019, Hong Kong, China, November 3-7, 2019, pages 791-802. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.507, + 0.885, + 0.626 + ], + "angle": 0, + "content": "Yijun Xiao and William Yang Wang. 2019. Quantifying uncertainties in natural language processing tasks. In The Thirty-Third AAAI Conference on Artificial Intelligence, AAAI 2019, The Thirty-First Innovative Applications of Artificial Intelligence Conference, IAAI 2019, The Ninth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2019, Honolulu, Hawaii, USA, January 27 - February 1, 2019, pages 7322-7329. AAAI Press." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.635, + 0.885, + 0.78 + ], + "angle": 0, + "content": "Bowen Yu, Zhenyu Zhang, Xiaobo Shu, Tingwen Liu, Yubin Wang, Bin Wang, and Sujian Li. 2020. Joint extraction of entities and relations based on a novel decomposition strategy. In ECAI 2020 - 24th European Conference on Artificial Intelligence, 29 August-8 September 2020, Santiago de Compostela, Spain, August 29 - September 8, 2020 - Including 10th Conference on Prestigious Applications of Artificial Intelligence (PAIS 2020), volume 325 of Frontiers in Artificial Intelligence and Applications, pages 2282-2289. IOS Press." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.79, + 0.885, + 0.882 + ], + "angle": 0, + "content": "Suncong Zheng, Feng Wang, Hongyun Bao, Yuexing Hao, Peng Zhou, and Bo Xu. 2017. Joint extraction of entities and relations based on a novel tagging scheme. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1227-1236, Vancouver, Canada. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.892, + 0.885, + 0.919 + ], + "angle": 0, + "content": "Zexuan Zhong and Danqi Chen. 2021. A frustratingly easy approach for entity and relation extraction. In" + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.885, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1355" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.086, + 0.49, + 0.153 + ], + "angle": 0, + "content": "Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 50-61, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.162, + 0.49, + 0.257 + ], + "angle": 0, + "content": "Yikai Zhou, Baosong Yang, Derek F. Wong, Yu Wan, and Lidia S. Chao. 2020. Uncertainty-aware curriculum learning for neural machine translation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6934-6944, Online. Association for Computational Linguistics." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1356" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.133, + 0.085, + 0.435, + 0.1 + ], + "angle": 0, + "content": "ACL 2023 Responsible NLP Checklist" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.108, + 0.323, + 0.123 + ], + "angle": 0, + "content": "A For every submission:" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.127, + 0.533, + 0.158 + ], + "angle": 0, + "content": "A1. Did you describe the limitations of your work? Section 7" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.171, + 0.687, + 0.202 + ], + "angle": 0, + "content": "A2. Did you discuss any potential risks of your work? We study open-domain information extraction for researches in this area" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.213, + 0.697, + 0.244 + ], + "angle": 0, + "content": "A3. Do the abstract and introduction summarize the paper's main claims? Abstract and Section 1" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.256, + 0.67, + 0.288 + ], + "angle": 0, + "content": "A4. Have you used AI writing assistants when working on this paper? Left blank." + }, + { + "type": "list", + "bbox": [ + 0.13, + 0.127, + 0.697, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.301, + 0.49, + 0.317 + ], + "angle": 0, + "content": "B Did you use or create scientific artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.322, + 0.216, + 0.337 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.348, + 0.531, + 0.379 + ], + "angle": 0, + "content": "B1. Did you cite the creators of artifacts you used? No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.391, + 0.779, + 0.423 + ], + "angle": 0, + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.434, + 0.881, + 0.513 + ], + "angle": 0, + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.525, + 0.881, + 0.588 + ], + "angle": 0, + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.6, + 0.881, + 0.648 + ], + "angle": 0, + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.659, + 0.881, + 0.755 + ], + "angle": 0, + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. No response." + }, + { + "type": "list", + "bbox": [ + 0.129, + 0.348, + 0.881, + 0.755 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.765, + 0.494, + 0.782 + ], + "angle": 0, + "content": "C Did you run computational experiments?" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.787, + 0.206, + 0.801 + ], + "angle": 0, + "content": "Section 5" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.814, + 0.881, + 0.861 + ], + "angle": 0, + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? No response." + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.869, + 0.878, + 0.893 + ], + "angle": 0, + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1357" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.13, + 0.084, + 0.88, + 0.116 + ], + "angle": 0, + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.118, + 0.223, + 0.131 + ], + "angle": 0, + "content": "Section 5" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.144, + 0.88, + 0.191 + ], + "angle": 0, + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.193, + 0.249, + 0.208 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.218, + 0.88, + 0.265 + ], + "angle": 0, + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.268, + 0.223, + 0.281 + ], + "angle": 0, + "content": "Section 5" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.294, + 0.878, + 0.31 + ], + "angle": 0, + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.315, + 0.215, + 0.33 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.341, + 0.88, + 0.372 + ], + "angle": 0, + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.375, + 0.249, + 0.389 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.4, + 0.88, + 0.447 + ], + "angle": 0, + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.45, + 0.249, + 0.464 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.476, + 0.88, + 0.522 + ], + "angle": 0, + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.525, + 0.249, + 0.539 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.55, + 0.873, + 0.565 + ], + "angle": 0, + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.568, + 0.249, + 0.582 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.593, + 0.88, + 0.623 + ], + "angle": 0, + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.627, + 0.249, + 0.641 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1358" + } + ] +] \ No newline at end of file diff --git a/2023/Uncertainty-Aware Bootstrap Learning for Joint Extraction on Distantly-Supervised Data/a0f755f8-19b9-4154-a835-8d33c06b518f_origin.pdf b/2023/Uncertainty-Aware Bootstrap Learning for Joint Extraction on Distantly-Supervised Data/a0f755f8-19b9-4154-a835-8d33c06b518f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0f0ed5472adc8151226dd07645d8b3f0270c983f --- /dev/null +++ b/2023/Uncertainty-Aware Bootstrap Learning for Joint Extraction on Distantly-Supervised Data/a0f755f8-19b9-4154-a835-8d33c06b518f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dd547702f63d3161eb1eb8366f29d787da51bcf52efe67712888d623c53cfb85 +size 496709 diff --git a/2023/Uncertainty-Aware Bootstrap Learning for Joint Extraction on Distantly-Supervised Data/full.md b/2023/Uncertainty-Aware Bootstrap Learning for Joint Extraction on Distantly-Supervised Data/full.md new file mode 100644 index 0000000000000000000000000000000000000000..13898a057021c3aabe4d5842b8588cf9ed5bae41 --- /dev/null +++ b/2023/Uncertainty-Aware Bootstrap Learning for Joint Extraction on Distantly-Supervised Data/full.md @@ -0,0 +1,259 @@ +# Uncertainty-Aware Bootstrap Learning for Joint Extraction on Distantly-Supervised Data + +Yufei Li $^{1}$ , Xiao Yu $^{2}$ , Yanchi Liu $^{3}$ , Haifeng Chen $^{3}$ , Cong Liu $^{1}$ + +1University of California, Riverside 2Stellar Cyber 3NEC Labs America + +$^{1}\{yli927,congl\} @ucr.edu,$ $^{2}\mathrm{xyu@stellarcyber.ai},$ + +3{yanchi,haifeng}@nec-labs.com + +# Abstract + +Jointly extracting entity pairs and their relations is challenging when working on distantly-supervised data with ambiguous or noisy labels. To mitigate such impact, we propose uncertainty-aware bootstrap learning, which is motivated by the intuition that the higher uncertainty of an instance, the more likely the model confidence is inconsistent with the ground truths. Specifically, we first explore instance-level data uncertainty to create an initial high-confident examples. Such subset serves as filtering noisy instances and facilitating the model to converge fast at the early stage. During bootstrap learning, we propose self-ensembling as a regularizer to alleviate inter-model uncertainty produced by noisy labels. We further define probability variance of joint tagging probabilities to estimate inner-model parametric uncertainty, which is used to select and build up new reliable training instances for the next iteration. Experimental results on two large datasets reveal that our approach outperforms existing strong baselines and related methods. + +# 1 Introduction + +Joint extraction involves extracting multiple types of entities and relations between them using a single model, which is necessary in automatic knowledge base construction (Yu et al., 2020). One way to cheaply acquire a large amount of labeled data for training joint extraction models is through distant supervision (DS) (Mintz et al., 2009). DS involves aligning a knowledge base (KB) with an unlabeled corpus using hand-crafted rules or logic constraints. Due to the lack of human annotators, DS brings a large proportion of noisy labels, e.g., over $30\%$ noisy instances in some cases (Mintz et al., 2009), making it impossible to learn useful features. The noise can be either false relations due to the aforementioned rule-based matching assumption or wrong entity tags due to limited coverage over entities in open-domain KBs. + +Existing distantly-supervised approaches model noise relying either on heuristics such as reinforcement learning (RL) (Nooralahzadeh et al., 2019; Hu et al., 2021) and adversarial learning (Chen et al., 2021), or pattern-based methods (Jia et al., 2019; Shang et al., 2022) to select trustable instances. Nevertheless, these methods require designing heuristics or hand-crafted patterns which may encourage a model to leverage spurious features without considering the confidence or uncertainty of its predictions. + +In response to these problems, we propose UnBED—Uncertainty-aware Bootstrap learning for joint Extraction on Distantly-supervised data. UnBED assumes that 1) low data uncertainty indicates reliable instances using a pre-trained language model (PLM) in the initial stage, 2) model should be aware of trustable entity and relation labels regarding its uncertainty after training. Our bootstrap serves uncertainty as a principle to mitigate the impact of noise labels on model learning and validate input sequences to control the number of training examples in each step. Particularly, we quantify data uncertainty of an instance according to its winning score (Hendrycks and Gimpel, 2017) and entropy (Shannon, 1948). We define averaged maximum probability that is estimated by a joint PLM over each token in a sequence to adapt previous techniques in joint extraction scheme. Instances with low data uncertainty are collected to form an initial subset, which is used to tune the joint PLM tagger and facilitate fast convergence. Then, we define parametric uncertainty in two perspectives—inter-model and inner-model uncertainty. The former is quantified by self-ensembling (Wang and Wang, 2022) and serves as a regularizer to improve model robustness against noisy labels during training. The latter is captured by probability variance in MC Dropout (Gal and Ghahramani, 2016) for selecting new confident instances for the next training iteration. Such two + +fold model uncertainties reinforce with each other to guide the model to iteratively improve its robustness and learn from reliable knowledge. + +# 2 Related Work + +Joint Extraction Methods Joint extraction detects entities and their relations using a single model, which effectively integrates the information from both sources and therefore achieves better results in both subtasks compared to pipelined methods (Zheng et al., 2017). For example, unified methods tag entities and relation simultaneously, e.g., (Zheng et al., 2017) proposes a novel tagging scheme which converts joint extraction to a sequence labeling problem; (Dai et al., 2019) introduces query position and sequential tagging to extract overlapping relations. Such methods avoid producing redundant information compared to parameter-sharing neural models (Miwa and Bansal, 2016; Gupta et al., 2016), and require no hand-crafted features that are used in structured systems (Yu et al., 2020). + +To address the challenge of learning from DS, pre-trained transformers (e.g., BERT, GPT-2) have gain much attention. They model strong expressive context-aware representations for text sequence through multiple attention layers, and achieve state-of-the-art performance on various NLP tasks (Radford et al., 2019; Devlin et al., 2019; Li et al., 2022). They can be cheaply fine-tuned to solve different downstream tasks including NER and RC. Specifically, BERT is trained on large English corpus using masked language modeling. The multi-head attention weights indicate interactions between each pair of words and its hidden states integrate semantic information of the whole sentence, which are used to decode different tagging results. + +Uncertainty Methods Uncertainty generally comes from two sources—aleatoric uncertainty and epistemic uncertainty. The former is also referred to as data uncertainty, describing noise inherent in the data generation. Methods mitigating such uncertainty include data interpolation (Dong et al., 2018), winning score, and temperature scale (Guo et al., 2017). The latter is also called model uncertainty, describing whether the structure choice and model parameters best describe the data distribution. One main solution to mitigate model uncertainty is Bayesian Neural Network (BNN) (Klein et al., 2017) that puts a prior distribution on its weights. To save computational cost, Monte Carlo + +![](images/c1f2c545fd2d830bab363ce8adb2f0142a1c0f955d0de4e723956682688a7890.jpg) +Figure 1: Joint extraction as a token classification task. + +dropout is proposed as an approximation of variational Bayesian inference (Gal and Ghahramani, 2016), realized by training models with dropout layers and testing with stochastic inference to quantify probability variance. Besides BNN, self-ensembling (Wang and Wang, 2022) which measures the outputs variance between models with the same architecture has been shown effective to reduce parametric uncertainty across models. + +# 3 Joint Extraction Model + +Tagging Scheme For an input sequence $\mathcal{X} = \{x_1, \dots, x_n\}$ , we tag $n$ sequences according to different query position $p$ following (Dai et al., 2019). If $p$ is the start of an entity (query entity $e_1$ ), the sequence is an instance. The entity type is labeled at $p$ and other entities $e_2$ which have relationship with the query entity are labeled with relation types $re$ . The rest of tokens are labeled "O" (Outside), meaning they do not correspond to the query entity. Accordingly, we convert joint extraction into a token classification task and extract relation triplets $\{e_1, re, e_2\}$ in each instance, as shown in Figure 1. + +Position-Attentive Encoder we use BERT (Devlin et al., 2019) to encode a sentence $\mathcal{X}$ into token-level representations $h = \{h_1,..,h_n\}$ , where $h_i\in \mathbb{R}^d$ is a $d$ -dimensional vector corresponding to the $i$ -th token in $\mathcal{X}$ . For each query $p$ , self-matching is applied to calculate the position-attention $\boldsymbol{a}_t\in \mathbb{R}^T$ between token at $p$ and each token at target position $t$ , which compares the sentence representations against itself to collect context information (Tan et al., 2018). The produced position-aware representation $\boldsymbol{c}_t\in \mathbb{R}^{T\times d}$ is an attention-weighted sentence vector $\boldsymbol{c}_t = \boldsymbol{a}_t^\top \boldsymbol{h}$ . Finally, we concatenate $\boldsymbol{h}_t$ and $\boldsymbol{c}_t$ to generate position-aware and context + +aware representations $\pmb{u}_t = [\pmb{h}_t|\pmb{c}_t]$ . + +CRF Decoder (Lafferty et al., 2001) For each position-aware representation $\boldsymbol{u}_t$ , we first learn a linear transformation $\boldsymbol{z}_t = \boldsymbol{W}\boldsymbol{u}_t \in \mathbb{R}^C$ to represent tag scores for the $t$ -th token. Here $C$ is the number of distinct tags. For an instance with labels $\boldsymbol{y} = \{y_1, \dots, y_n\}$ , the decoding score $s(\boldsymbol{z}, \boldsymbol{y})$ is the sum of transition scores from tag $y_t$ to tag $y_{t+1}$ plus the input score $z_t^{yt}$ . The conditional probability $p(\boldsymbol{y}|\boldsymbol{z})$ is the softmax over $s(\boldsymbol{z}, \boldsymbol{y})$ for all possible label sequences $\boldsymbol{y}'$ . We maximize the log-likelihood of correct tag sequences during training $\mathcal{L}_c = \sum \log p(\boldsymbol{y}|\boldsymbol{z})$ . + +# 4 Uncertainty-Aware Bootstrap Learning + +Motivation One of the main challenges in bootstrap learning is to evaluate the "correctness" of a labeled instance. We consider this problem from an uncertainty perspective and assume instances with lower uncertainty are more likely to be correctly labeled. In this section, we first propose instance-level data uncertainty which is used to filter noisy examples and build an initial subset. Then, we introduce our two-fold model uncertainties which helps iteratively mitigate DS effect and build up trustable examples during bootstrap learning. + +# 4.1 Data Uncertainty + +Presenting examples in an easy-to-hard order at different training stages can benefit models (Platanios et al., 2019; Zhou et al., 2020), we propose data uncertainty as a way to quantify the "hardness" of an instance. To better estimate the data uncertainty, we use pre-trained language models (PLMs) to generate tag probability for each token in a sequence. Our intuition is that higher uncertain inputs are "harder" to be generated by a PLM, as it already has rationales of language. Accordingly, we propose two data uncertainties, which can be used individually or combined together: + +Winning Score (WS) The maximum softmax probability reflects data uncertainty of an input (Hendrycks and Gimpel, 2017). Given an input instance $\mathcal{I} = \{x_1,\dots,x_n\}$ , we define data uncertainty $u^{d}(\mathcal{I})$ as the minus averaged token classification winning score: + +$$ +u ^ {d} (\mathcal {I}) = - \frac {1}{n} \sum_ {t = 1} ^ {n} \max _ {c \in [ 1, C ]} P (y _ {t} = c | x _ {t}) \quad (1) +$$ + +Entropy Shannon entropy (Shannon, 1948) is widely used to reflect information uncertainty. We + +propose data uncertainty $u^{d}(\mathcal{I})$ as the averaged token classification entropy: + +$$ +u ^ {d} (\mathcal {I}) = \frac {1}{n} \sum_ {t = 1} ^ {n} \sum_ {c = 1} ^ {C} P \left(y _ {t} = c \mid x _ {t}\right) \log P \left(y _ {t} = c \mid x _ {t}\right) \tag {2} +$$ + +We filter out examples with high uncertainty scores and build an initial subset with "simple" examples. At the early training stage, a model is not aware of what a decent distribution $P(y|x)$ should be, thus data uncertainty facilitates it to converge fast by tuning on a fairly "simple" subset. + +# 4.2 Model Uncertainty + +In our bootstrap learning, we define model uncertainty, i.e., epistemic uncertainty (Kendall and Gal, 2017), to measure whether model parameters can best describe the data distribution following (Zhou et al., 2020). A small model uncertainty indicates the model is confident that the current training data has been well learned (Wang et al., 2019). We adopt Monte Carlo Dropout (Gal and Ghahramani, 2016) to approximate Bayesian inference which captures inner-model parametric uncertainty. Specifically, we perform $K$ forward passes through our joint model. In each pass, part of network neurons $\theta$ are randomly deactivated. Finally, we yield $K$ samples on model parameters $\{\hat{\theta}_1,\dots,\hat{\theta}_K\}$ . We use the averaged token classification Probability Variance (PV) (Shelmanov et al., 2021) over all tags for instance $\mathcal{I}$ : + +$$ +u ^ {m} (\theta) = \frac {1}{n} \sum_ {t = 1} ^ {n} \sum_ {c = 1} ^ {C} \operatorname {V a r} \left[ P \left(y _ {t} = c \mid x _ {t}, \hat {\theta} _ {k}\right) \right] _ {k = 1} ^ {K} \tag {3} +$$ + +where $\operatorname{Var}[\cdot]$ is the variance of distribution over the $K$ passes following the common settings in (Dong et al., 2018; Xiao and Wang, 2019). Accordingly, model is aware of its confidence over each instance and how likely the label is noisy. + +# 4.3 Training Strategy + +Uncertainty-Aware Loss Besides MC Dropout which measures parametric uncertainty within a model, we also consider mitigating parametric uncertainty between models to stabilize the weights during training. Specifically, we use self-ensembling (He et al., 2020; Wang and Wang, 2022) to calculate the loss between the same models to improve model robustness and reduce the label noise effect on model performance. + +# Algorithm 1 Bootstrap Learning + +Input: Original dataset $\mathcal{D} = \{(\mathcal{I}^n,y^n)\}_{n = 1}^N$ two joint models $f_{1},f_{2}$ with parameters $\theta_{1},\theta_{2}$ + +1: Compute data uncertainty $u^{d}(\mathcal{I})$ for each instance $\mathcal{I}$ in $\mathcal{D}$ ; +2: Initial dataset $\mathcal{C} \gets$ Select data pairs $(\mathcal{I}^n, y^n)$ such that $u^d(\mathcal{I}) < \tau^d$ from $\mathcal{D}$ ; +3: for epoch $e = 1, \ldots$ do +4: Train $f_{1}, f_{2}$ on $\mathcal{C}$ using Eq. (5); +5: Calculate model uncertainty $u^{m}(\theta_{1})$ on $\mathcal{D}$ ; +6: $\mathcal{C} \gets$ Select data pairs $(\mathcal{I}^n, y^n)$ such that $u^m(\mathcal{I}; \theta_1) < \tau^m$ from $\mathcal{D}$ ; + +We create another joint model with identical framework, e.g., architecture, loss functions, hyperparameters, and compute a self-ensemble loss $\mathcal{L}_e$ to minimize the difference between two outputs from the two models regarding the same inputs: + +$$ +\mathcal {L} _ {e} = \sum K L (f (\mathcal {I}; \theta_ {1}), f (\mathcal {I}; \theta_ {2})) \tag {4} +$$ + +where $KL(.)$ is the Kullback-Leibler divergence between two probabilistic distributions, $\theta_{1},\theta_{2}$ denote the parameters of first and second models. We formulate our final uncertainty-aware objective $\mathcal{L}$ as the sum of CRF and self-ensemble loss: + +$$ +\mathcal {L} = \mathcal {L} _ {c} + \alpha \mathcal {L} _ {e} \tag {5} +$$ + +where $\alpha$ denotes the weight of self-ensembling, and $\mathcal{L}_c$ means the token classification loss. + +Bootstrap Learning Procedure To mitigate the DS effect on model performance, we propose a twofold bootstrap learning strategy (see Algorithm 1). Specifically, we first apply data uncertainty to filter "harder" examples and redistribute a reliable initial training data $\mathcal{M}$ . Then, we iteratively feed examples following an easy-to-hard order to the model. In each training iteration, we regularize the joint model with self-ensembling loss to reduce the impact of noisy labels on model parameters. Then we use probability variance to select new confident training instances $\mathcal{D}'$ that can be explained by the model as the next training inputs. The more certain examples are selected, the more likely the model will learn beneficial information and will converge faster. We repeat the above procedure until the F1 score on the validation set converges. + +# 5 Experiments + +# 5.1 Setup + +We evaluate the performance of UnBED on two datasets, NYT and Wiki-KBP. The NYT (Riedel et al., 2010) dataset collects news from New York Times and its training data is automatically labeled by DS. We use the revised test dataset (Jia et al., 2019) that is manually annotated to ensure quality. The Wiki-KBP (Ling and Weld, 2012) dataset collects articles from Wikipedia. Its training data is labeled by DS (Liu et al., 2017), and the test set is manually annotated (Ellis et al., 2013). + +We compare UnBED with the following baselines: ARNOR (Jia et al., 2019), a pattern-based method to reduce noise for distantly-supervised triplet extraction. PURE (Zhong and Chen, 2021), a pipeline approach that uses pre-trained BERT entity model to first recognize entities and then employs a relation model to detect underlying relations. FAN (Hao et al., 2021), an adversarial method including a transformers encoder to reduce noise for distantly-supervised triplet extraction. + +Evaluation We evaluate the extracted triplets for each sentence based on Precision (Prec.), Recall (Rec.), and F1. A triplet $\{e_1, re, e_2\}$ is marked correct if the relation type $re$ , two entities $e_1, e_2$ are all correct. We build a validation set by randomly sampling $10\%$ sentences from the test set. + +Implementation Details We use Hugging Face bert-large-uncased (Devlin et al., 2019) pre-trained model as backbone. For ARNOR, the hidden vector size is set to 300. In regularization training, we find optimal parameters $\alpha$ as 1 for both datasets. We implement UnBED and all baselines in PyTorch, with Adam optimizer, initial learning rate $10^{-5}$ , dropout rate 0.1, and batch size 8. For initial subset configuration, we choose data uncertainty threshold 0.5. For bootstrap learning, an empirical model uncertainty threshold is set to 0.6 with the best validation F1. + +# 5.2 Overall Results + +As shown in Table 1, UnBED significantly outperforms all baselines in precision and F1 metric. Specifically, UnBED achieves $8\%$ F1 improvement on NYT (3% on Wiki-KBP) over denoising approaches—ARNOR and FAN. Our approach also outperforms baselines using pretrained transformers (PURE and FAN), showing that uncertainty-aware bootstrap learning effectively reduces the impact of noisy labels. + +
MethodNYTWiki-KBP
Prec.Rec.F1Prec.Rec.F1
ARNOR (Jia et al., 2019)0.5880.6140.6000.4020.4710.434
PURE (Zhong and Chen, 2021)0.5360.6640.5930.3950.4330.413
FAN (Hao et al., 2021)0.5790.6460.6110.3910.4670.426
UnBED-WS0.6620.7300.6940.4290.5010.462
UnBED-Entropy0.6510.7410.6930.4220.5090.461
+ +Table 1: Evaluation results on NYT and Wiki-KBP datasets. Bold numbers denote the best metrics. UnBED-WS and UnBED-Entropy denote UnBED with winning score and entropy as the data uncertainty, respectively. + +![](images/0b1cd8eb90ceb16ce072eac5bd6e552b67596131705df3883ebf1cfe9ff3f8eb.jpg) +Figure 2: F1 score vs. Epochs under different settings. Vanilla-PV-enssembled denotes UnBED-WS, and entropy-PV-enssembled denotes UnBED-Entropy. + +# 5.3 Further Analysis + +We analyze the functionality of different components in Figure 2. We observe that both the entropy-PV and vanilla-PV outperform the baseline (joint model directly trained on the original DS dataset) in terms of F1 $(5\sim 7\%)$ increase), demonstrating the effect of filtering noisy labels and selecting trustable instance using probability variance. Besides, self-ensembling further enhances the performance in later training stage $(2\sim 4$ F1 increase), proving that mitigating the inter-model uncertainty benefits model robustness against noisy labels. + +# 6 Conclusions + +In this paper, we propose a novel uncertainty-aware bootstrap learning framework for distantly-supervised joint extraction. Specifically, we define data uncertainty in generally token classification to filter out highly-error-prone instances and build an initial high-confident subset, which is used to tune the joint extraction model for fast convergence. We then propose a two-fold bootstrap learning procedure which iteratively mitigates the DS impact on model robustness and selects new trustable training instances. Experimental results on two benchmark datasets show that UnBED significantly out + +performs other denoising techniques. + +# Limitations + +In this work we propose an uncertainty-aware bootstrap learning framework for joint extraction. Though it achieves state-of-the-art performance compared to other denoising techniques, UnBED requires large training resources considering the ensemble loss calculated between two large PLMs and the probability variance calculated on the PLM joint extraction model. In our future work, we hope to incorporate pruning techniques during training to improve the efficiency. We will also consider more complex relations between entities, e.g., relations beyond the sentence boundary, to fit in real-world information extraction scenarios. + +# Acknowledgements + +This work was supported by NSF CNS 2135625, CPS 2038727, CNS Career 1750263, and a Darpa Shell grant. + +# References + +Tao Chen, Haochen Shi, Liyuan Liu, Siliang Tang, Jian Shao, Zhigang Chen, and Yueting Zhuang. 2021. Empower distantly supervised relation extraction with collaborative adversarial training. In Thirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2021, Thirty-Third Conference on Innovative Applications of Artificial Intelligence, IAAI 2021, The Eleventh Symposium on Educational Advances in Artificial Intelligence, EAAI 2021, Virtual Event, February 2-9, 2021, pages 12675-12682. AAAI Press. +Dai Dai, Xinyan Xiao, Yajuan Lyu, Shan Dou, Qiaoqiao She, and Haifeng Wang. 2019. Joint extraction of entities and overlapping relations using position-attentive sequence labeling. In The Thirty-Third AAAI Conference on Artificial Intelligence, AAAI 2019, The Thirty-First Innovative Applications of Artificial Intelligence Conference, IAAI 2019, The Ninth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2019, Honolulu, Hawaii, + +USA, January 27 - February 1, 2019, pages 6300-6308. AAAI Press. +Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics. +Li Dong, Chris Quirk, and Mirella Lapata. 2018. Confidence modeling for neural semantic parsing. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 743-753, Melbourne, Australia. Association for Computational Linguistics. +Joe Ellis, Jeremy Getman, Justin Mott, Xuansong Li, Kira Griffith, Stephanie M. Strassel, and Jonathan Wright. 2013. Linguistic resources for 2013 knowledge base population evaluations. In Proceedings of the Sixth Text Analysis Conference, TAC 2013, Gaithersburg, Maryland, USA, November 18-19, 2013. NIST. +Yarin Gal and Zoubin Ghahramani. 2016. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In Proceedings of the 33nd International Conference on Machine Learning, ICML 2016, New York City, NY, USA, June 19-24, 2016, volume 48 of JMLR Workshop and Conference Proceedings, pages 1050-1059. JMLR.org. +Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. 2017. On calibration of modern neural networks. In Proceedings of the 34th International Conference on Machine Learning, ICML 2017, Sydney, NSW, Australia, 6-11 August 2017, volume 70 of Proceedings of Machine Learning Research, pages 1321-1330. PMLR. +Pankaj Gupta, Hinrich Schütze, and Bernt Andrassy. 2016. Table filling multi-task recurrent neural network for joint entity and relation extraction. In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 2537-2547, Osaka, Japan. The COLING 2016 Organizing Committee. +Kailong Hao, Botao Yu, and Wei Hu. 2021. Knowing false negatives: An adversarial training method for distantly supervised relation extraction. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 9661-9672, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics. +Jianfeng He, Xuchao Zhang, Shuo Lei, Zhiqian Chen, Fanglan Chen, Abdulaziz Alhamadani, Bei Xiao, and ChangTien Lu. 2020. Towards more accurate uncertainty estimation in text classification. In Proceedings of the 2020 Conference on Empirical Methods + +in Natural Language Processing (EMNLP), pages 8362-8372, Online. Association for Computational Linguistics. +Dan Hendrycks and Kevin Gimpel. 2017. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net. +Xuming Hu, Chenwei Zhang, Yawen Yang, Xiaohe Li, Li Lin, Lijie Wen, and Philip S. Yu. 2021. Gradient imitation reinforcement learning for low resource relation extraction. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, EMNLP 2021, Virtual Event / Punta Cana, Dominican Republic, 7-11 November; 2021, pages 2737-2746. Association for Computational Linguistics. +Wei Jia, Dai Dai, Xinyan Xiao, and Hua Wu. 2019. ARNOR: Attention regularization based noise reduction for distant supervision relation classification. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1399-1408, Florence, Italy. Association for Computational Linguistics. +Alex Kendall and Yarin Gal. 2017. What uncertainties do we need in bayesian deep learning for computer vision? In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 5574-5584. +Aaron Klein, Stefan Falkner, Jost Tobias Springenberg, and Frank Hutter. 2017. Learning curve prediction with bayesian neural networks. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net. +John D. Lafferty, Andrew McCallum, and Fernando C. N. Pereira. 2001. Conditional random fields: Probabilistic models for segmenting and labeling sequence data. In Proceedings of the Eighteenth International Conference on Machine Learning, ICML '01, page 282-289, San Francisco, CA, USA. Morgan Kaufmann Publishers Inc. +Shuyang Li, Yufei Li, Jianmo Ni, and Julian McAuley. 2022. SHARE: a system for hierarchical assistive recipe editing. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 11077-11090, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics. +Xiao Ling and Daniel S. Weld. 2012. Fine-grained entity recognition. In Proceedings of the Twenty-Sixth AAAI Conference on Artificial Intelligence, July 22-26, 2012, Toronto, Ontario, Canada. AAAI Press. + +Liyuan Liu, Xiang Ren, Qi Zhu, Shi Zhi, Huan Gui, Heng Ji, and Jiawei Han. 2017. Heterogeneous supervision for relation extraction: A representation learning approach. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 46-56, Copenhagen, Denmark. Association for Computational Linguistics. +Mike Mintz, Steven Bills, Rion Snow, and Daniel Jurafsky. 2009. Distant supervision for relation extraction without labeled data. In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP, pages 1003-1011, Suntec, Singapore. Association for Computational Linguistics. +Makoto Miwa and Mohit Bansal. 2016. End-to-end relation extraction using lstms on sequences and tree structures. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, ACL 2016, August 7-12, 2016, Berlin, Germany, Volume 1: Long Papers. The Association for Computer Linguistics. +Farhad Nooralahzadeh, Jan Tore Lönning, and Lilja Øvrelid. 2019. Reinforcement-based denoising of distantly supervised NER with partial annotation. In Proceedings of the 2nd Workshop on Deep Learning Approaches for Low-Resource NLP, DeepLo@EMNLP-IJCNLP 2019, Hong Kong, China, November 3, 2019, pages 225–233. Association for Computational Linguistics. +Emmanouil Antonios Platanios, Otilia Stretcu, Graham Neubig, Barnabás Póczos, and Tom M. Mitchell. 2019. Competence-based curriculum learning for neural machine translation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 1162-1172. Association for Computational Linguistics. +Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners. +Sebastian Riedel, Limin Yao, and Andrew McCallum. 2010. Modeling relations and their mentions without labeled text. In Machine Learning and Knowledge Discovery in Databases, European Conference, ECML PKDD 2010, Barcelona, Spain, September 20-24, 2010, Proceedings, Part III, volume 6323 of Lecture Notes in Computer Science, pages 148-163. Springer. +Yuming Shang, Heyan Huang, Xin Sun, Wei Wei, and Xian-Ling Mao. 2022. A pattern-aware self-attention network for distant supervised relation extraction. Inf. Sci., 584:269-279. +Claude E. Shannon. 1948. A mathematical theory of communication. Bell Syst. Tech. J., 27(3):379-423. + +Artem Shelmanov, Evgenii Tsymbalov, Dmitri Puzyrev, Kirill Fedyanin, Alexander Panchenko, and Maxim Panov. 2021. How certain is your Transformer? In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 1833-1840, Online. Association for Computational Linguistics. +Zhixing Tan, Mingxuan Wang, Jun Xie, Yidong Chen, and Xiaodong Shi. 2018. Deep semantic role labeling with self-attention. In Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence and Thirtieth Innovative Applications of Artificial Intelligence Conference and Eighth AAAI Symposium on Educational Advances in Artificial Intelligence, AAAI'18/IAAI'18/EAAI'18. AAAI Press. +Hongjun Wang and Yisen Wang. 2022. Self-ensemble adversarial training for improved robustness. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net. +Shuo Wang, Yang Liu, Chao Wang, Huanbo Luan, and Maosong Sun. 2019. Improving back-translation with uncertainty-based confidence estimation. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLP-IJCNLP 2019, Hong Kong, China, November 3-7, 2019, pages 791-802. Association for Computational Linguistics. +Yijun Xiao and William Yang Wang. 2019. Quantifying uncertainties in natural language processing tasks. In The Thirty-Third AAAI Conference on Artificial Intelligence, AAAI 2019, The Thirty-First Innovative Applications of Artificial Intelligence Conference, IAAI 2019, The Ninth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2019, Honolulu, Hawaii, USA, January 27 - February 1, 2019, pages 7322-7329. AAAI Press. +Bowen Yu, Zhenyu Zhang, Xiaobo Shu, Tingwen Liu, Yubin Wang, Bin Wang, and Sujian Li. 2020. Joint extraction of entities and relations based on a novel decomposition strategy. In ECAI 2020 - 24th European Conference on Artificial Intelligence, 29 August-8 September 2020, Santiago de Compostela, Spain, August 29 - September 8, 2020 - Including 10th Conference on Prestigious Applications of Artificial Intelligence (PAIS 2020), volume 325 of Frontiers in Artificial Intelligence and Applications, pages 2282-2289. IOS Press. +Suncong Zheng, Feng Wang, Hongyun Bao, Yuexing Hao, Peng Zhou, and Bo Xu. 2017. Joint extraction of entities and relations based on a novel tagging scheme. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1227-1236, Vancouver, Canada. Association for Computational Linguistics. +Zexuan Zhong and Danqi Chen. 2021. A frustratingly easy approach for entity and relation extraction. In + +Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 50-61, Online. Association for Computational Linguistics. +Yikai Zhou, Baosong Yang, Derek F. Wong, Yu Wan, and Lidia S. Chao. 2020. Uncertainty-aware curriculum learning for neural machine translation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6934-6944, Online. Association for Computational Linguistics. + +A For every submission: + +A1. Did you describe the limitations of your work? Section 7 +A2. Did you discuss any potential risks of your work? We study open-domain information extraction for researches in this area +A3. Do the abstract and introduction summarize the paper's main claims? Abstract and Section 1 +A4. Have you used AI writing assistants when working on this paper? Left blank. + +B Did you use or create scientific artifacts? + +Left blank. + +B1. Did you cite the creators of artifacts you used? No response. +B2. Did you discuss the license or terms for use and / or distribution of any artifacts? No response. +B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? No response. +B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? No response. +B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? No response. +B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. No response. + +C Did you run computational experiments? + +Section 5 + +C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? No response. + +C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? + +Section 5 + +C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? + +No response. + +C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? + +Section 5 + +D Did you use human annotators (e.g., crowdworkers) or research with human participants? + +Left blank. + +D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? + +No response. + +D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? + +No response. + +D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? + +No response. + +D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? + +No response. + +D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? + +No response. \ No newline at end of file diff --git a/2023/Uncertainty-Aware Bootstrap Learning for Joint Extraction on Distantly-Supervised Data/images.zip b/2023/Uncertainty-Aware Bootstrap Learning for Joint Extraction on Distantly-Supervised Data/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6475875192b1dd61969ac167d4e7bb00529c6231 --- /dev/null +++ b/2023/Uncertainty-Aware Bootstrap Learning for Joint Extraction on Distantly-Supervised Data/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2c9770abd590810f7ce7aefdab4522a650de313458d43862ffa31c6b1dcb10b2 +size 128762 diff --git a/2023/Uncertainty-Aware Bootstrap Learning for Joint Extraction on Distantly-Supervised Data/layout.json b/2023/Uncertainty-Aware Bootstrap Learning for Joint Extraction on Distantly-Supervised Data/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d07a5fee441e1c7f9495697ec89022fda3c93d09 --- /dev/null +++ b/2023/Uncertainty-Aware Bootstrap Learning for Joint Extraction on Distantly-Supervised Data/layout.json @@ -0,0 +1,7486 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 104, + 75, + 490, + 110 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 75, + 490, + 110 + ], + "spans": [ + { + "bbox": [ + 104, + 75, + 490, + 110 + ], + "type": "text", + "content": "Uncertainty-Aware Bootstrap Learning for Joint Extraction on Distantly-Supervised Data" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 139, + 125, + 458, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 125, + 458, + 139 + ], + "spans": [ + { + "bbox": [ + 139, + 125, + 458, + 139 + ], + "type": "text", + "content": "Yufei Li" + }, + { + "bbox": [ + 139, + 125, + 458, + 139 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 139, + 125, + 458, + 139 + ], + "type": "text", + "content": ", Xiao Yu" + }, + { + "bbox": [ + 139, + 125, + 458, + 139 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 139, + 125, + 458, + 139 + ], + "type": "text", + "content": ", Yanchi Liu" + }, + { + "bbox": [ + 139, + 125, + 458, + 139 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 139, + 125, + 458, + 139 + ], + "type": "text", + "content": ", Haifeng Chen" + }, + { + "bbox": [ + 139, + 125, + 458, + 139 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 139, + 125, + 458, + 139 + ], + "type": "text", + "content": ", Cong Liu" + }, + { + "bbox": [ + 139, + 125, + 458, + 139 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 123, + 140, + 473, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 140, + 473, + 153 + ], + "spans": [ + { + "bbox": [ + 123, + 140, + 473, + 153 + ], + "type": "text", + "content": "1University of California, Riverside 2Stellar Cyber 3NEC Labs America" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 166, + 154, + 431, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 154, + 431, + 168 + ], + "spans": [ + { + "bbox": [ + 166, + 154, + 431, + 168 + ], + "type": "inline_equation", + "content": "^{1}\\{yli927,congl\\} @ucr.edu," + }, + { + "bbox": [ + 166, + 154, + 431, + 168 + ], + "type": "inline_equation", + "content": "^{2}\\mathrm{xyu@stellarcyber.ai}," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 208, + 168, + 388, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 168, + 388, + 181 + ], + "spans": [ + { + "bbox": [ + 208, + 168, + 388, + 181 + ], + "type": "text", + "content": "3{yanchi,haifeng}@nec-labs.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 86, + 231, + 274, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 231, + 274, + 516 + ], + "spans": [ + { + "bbox": [ + 86, + 231, + 274, + 516 + ], + "type": "text", + "content": "Jointly extracting entity pairs and their relations is challenging when working on distantly-supervised data with ambiguous or noisy labels. To mitigate such impact, we propose uncertainty-aware bootstrap learning, which is motivated by the intuition that the higher uncertainty of an instance, the more likely the model confidence is inconsistent with the ground truths. Specifically, we first explore instance-level data uncertainty to create an initial high-confident examples. Such subset serves as filtering noisy instances and facilitating the model to converge fast at the early stage. During bootstrap learning, we propose self-ensembling as a regularizer to alleviate inter-model uncertainty produced by noisy labels. We further define probability variance of joint tagging probabilities to estimate inner-model parametric uncertainty, which is used to select and build up new reliable training instances for the next iteration. Experimental results on two large datasets reveal that our approach outperforms existing strong baselines and related methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 68, + 523, + 154, + 535 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 523, + 154, + 535 + ], + "spans": [ + { + "bbox": [ + 68, + 523, + 154, + 535 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 543, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 543, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 543, + 291, + 772 + ], + "type": "text", + "content": "Joint extraction involves extracting multiple types of entities and relations between them using a single model, which is necessary in automatic knowledge base construction (Yu et al., 2020). One way to cheaply acquire a large amount of labeled data for training joint extraction models is through distant supervision (DS) (Mintz et al., 2009). DS involves aligning a knowledge base (KB) with an unlabeled corpus using hand-crafted rules or logic constraints. Due to the lack of human annotators, DS brings a large proportion of noisy labels, e.g., over " + }, + { + "bbox": [ + 69, + 543, + 291, + 772 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 69, + 543, + 291, + 772 + ], + "type": "text", + "content": " noisy instances in some cases (Mintz et al., 2009), making it impossible to learn useful features. The noise can be either false relations due to the aforementioned rule-based matching assumption or wrong entity tags due to limited coverage over entities in open-domain KBs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 213, + 527, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 213, + 527, + 361 + ], + "spans": [ + { + "bbox": [ + 302, + 213, + 527, + 361 + ], + "type": "text", + "content": "Existing distantly-supervised approaches model noise relying either on heuristics such as reinforcement learning (RL) (Nooralahzadeh et al., 2019; Hu et al., 2021) and adversarial learning (Chen et al., 2021), or pattern-based methods (Jia et al., 2019; Shang et al., 2022) to select trustable instances. Nevertheless, these methods require designing heuristics or hand-crafted patterns which may encourage a model to leverage spurious features without considering the confidence or uncertainty of its predictions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 368, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 368, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 368, + 526, + 772 + ], + "type": "text", + "content": "In response to these problems, we propose UnBED—Uncertainty-aware Bootstrap learning for joint Extraction on Distantly-supervised data. UnBED assumes that 1) low data uncertainty indicates reliable instances using a pre-trained language model (PLM) in the initial stage, 2) model should be aware of trustable entity and relation labels regarding its uncertainty after training. Our bootstrap serves uncertainty as a principle to mitigate the impact of noise labels on model learning and validate input sequences to control the number of training examples in each step. Particularly, we quantify data uncertainty of an instance according to its winning score (Hendrycks and Gimpel, 2017) and entropy (Shannon, 1948). We define averaged maximum probability that is estimated by a joint PLM over each token in a sequence to adapt previous techniques in joint extraction scheme. Instances with low data uncertainty are collected to form an initial subset, which is used to tune the joint PLM tagger and facilitate fast convergence. Then, we define parametric uncertainty in two perspectives—inter-model and inner-model uncertainty. The former is quantified by self-ensembling (Wang and Wang, 2022) and serves as a regularizer to improve model robustness against noisy labels during training. The latter is captured by probability variance in MC Dropout (Gal and Ghahramani, 2016) for selecting new confident instances for the next training iteration. Such two" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "1349" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 135, + 795, + 459, + 805 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 795, + 459, + 805 + ], + "spans": [ + { + "bbox": [ + 135, + 795, + 459, + 805 + ], + "type": "text", + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 219, + 806, + 375, + 817 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 806, + 375, + 817 + ], + "spans": [ + { + "bbox": [ + 219, + 806, + 375, + 817 + ], + "type": "text", + "content": "Volume 2: Short Papers, pages 1349-1358" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "spans": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "text", + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 291, + 113 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 291, + 113 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 291, + 113 + ], + "type": "text", + "content": "fold model uncertainties reinforce with each other to guide the model to iteratively improve its robustness and learn from reliable knowledge." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 125, + 161, + 137 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 125, + 161, + 137 + ], + "spans": [ + { + "bbox": [ + 68, + 125, + 161, + 137 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 148, + 291, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 148, + 291, + 378 + ], + "spans": [ + { + "bbox": [ + 67, + 148, + 291, + 378 + ], + "type": "text", + "content": "Joint Extraction Methods Joint extraction detects entities and their relations using a single model, which effectively integrates the information from both sources and therefore achieves better results in both subtasks compared to pipelined methods (Zheng et al., 2017). For example, unified methods tag entities and relation simultaneously, e.g., (Zheng et al., 2017) proposes a novel tagging scheme which converts joint extraction to a sequence labeling problem; (Dai et al., 2019) introduces query position and sequential tagging to extract overlapping relations. Such methods avoid producing redundant information compared to parameter-sharing neural models (Miwa and Bansal, 2016; Gupta et al., 2016), and require no hand-crafted features that are used in structured systems (Yu et al., 2020)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 380, + 291, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 380, + 291, + 582 + ], + "spans": [ + { + "bbox": [ + 67, + 380, + 291, + 582 + ], + "type": "text", + "content": "To address the challenge of learning from DS, pre-trained transformers (e.g., BERT, GPT-2) have gain much attention. They model strong expressive context-aware representations for text sequence through multiple attention layers, and achieve state-of-the-art performance on various NLP tasks (Radford et al., 2019; Devlin et al., 2019; Li et al., 2022). They can be cheaply fine-tuned to solve different downstream tasks including NER and RC. Specifically, BERT is trained on large English corpus using masked language modeling. The multi-head attention weights indicate interactions between each pair of words and its hidden states integrate semantic information of the whole sentence, which are used to decode different tagging results." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 584, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 584, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 584, + 291, + 773 + ], + "type": "text", + "content": "Uncertainty Methods Uncertainty generally comes from two sources—aleatoric uncertainty and epistemic uncertainty. The former is also referred to as data uncertainty, describing noise inherent in the data generation. Methods mitigating such uncertainty include data interpolation (Dong et al., 2018), winning score, and temperature scale (Guo et al., 2017). The latter is also called model uncertainty, describing whether the structure choice and model parameters best describe the data distribution. One main solution to mitigate model uncertainty is Bayesian Neural Network (BNN) (Klein et al., 2017) that puts a prior distribution on its weights. To save computational cost, Monte Carlo" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 351, + 68, + 480, + 242 + ], + "blocks": [ + { + "bbox": [ + 351, + 68, + 480, + 242 + ], + "lines": [ + { + "bbox": [ + 351, + 68, + 480, + 242 + ], + "spans": [ + { + "bbox": [ + 351, + 68, + 480, + 242 + ], + "type": "image", + "image_path": "c1f2c545fd2d830bab363ce8adb2f0142a1c0f955d0de4e723956682688a7890.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 248, + 524, + 261 + ], + "lines": [ + { + "bbox": [ + 302, + 248, + 524, + 261 + ], + "spans": [ + { + "bbox": [ + 302, + 248, + 524, + 261 + ], + "type": "text", + "content": "Figure 1: Joint extraction as a token classification task." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 275, + 526, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 275, + 526, + 397 + ], + "spans": [ + { + "bbox": [ + 302, + 275, + 526, + 397 + ], + "type": "text", + "content": "dropout is proposed as an approximation of variational Bayesian inference (Gal and Ghahramani, 2016), realized by training models with dropout layers and testing with stochastic inference to quantify probability variance. Besides BNN, self-ensembling (Wang and Wang, 2022) which measures the outputs variance between models with the same architecture has been shown effective to reduce parametric uncertainty across models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 410, + 443, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 410, + 443, + 423 + ], + "spans": [ + { + "bbox": [ + 302, + 410, + 443, + 423 + ], + "type": "text", + "content": "3 Joint Extraction Model" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 433, + 526, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 433, + 526, + 597 + ], + "spans": [ + { + "bbox": [ + 302, + 433, + 526, + 597 + ], + "type": "text", + "content": "Tagging Scheme For an input sequence " + }, + { + "bbox": [ + 302, + 433, + 526, + 597 + ], + "type": "inline_equation", + "content": "\\mathcal{X} = \\{x_1, \\dots, x_n\\}" + }, + { + "bbox": [ + 302, + 433, + 526, + 597 + ], + "type": "text", + "content": ", we tag " + }, + { + "bbox": [ + 302, + 433, + 526, + 597 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 302, + 433, + 526, + 597 + ], + "type": "text", + "content": " sequences according to different query position " + }, + { + "bbox": [ + 302, + 433, + 526, + 597 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 302, + 433, + 526, + 597 + ], + "type": "text", + "content": " following (Dai et al., 2019). If " + }, + { + "bbox": [ + 302, + 433, + 526, + 597 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 302, + 433, + 526, + 597 + ], + "type": "text", + "content": " is the start of an entity (query entity " + }, + { + "bbox": [ + 302, + 433, + 526, + 597 + ], + "type": "inline_equation", + "content": "e_1" + }, + { + "bbox": [ + 302, + 433, + 526, + 597 + ], + "type": "text", + "content": "), the sequence is an instance. The entity type is labeled at " + }, + { + "bbox": [ + 302, + 433, + 526, + 597 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 302, + 433, + 526, + 597 + ], + "type": "text", + "content": " and other entities " + }, + { + "bbox": [ + 302, + 433, + 526, + 597 + ], + "type": "inline_equation", + "content": "e_2" + }, + { + "bbox": [ + 302, + 433, + 526, + 597 + ], + "type": "text", + "content": " which have relationship with the query entity are labeled with relation types " + }, + { + "bbox": [ + 302, + 433, + 526, + 597 + ], + "type": "inline_equation", + "content": "re" + }, + { + "bbox": [ + 302, + 433, + 526, + 597 + ], + "type": "text", + "content": ". The rest of tokens are labeled \"O\" (Outside), meaning they do not correspond to the query entity. Accordingly, we convert joint extraction into a token classification task and extract relation triplets " + }, + { + "bbox": [ + 302, + 433, + 526, + 597 + ], + "type": "inline_equation", + "content": "\\{e_1, re, e_2\\}" + }, + { + "bbox": [ + 302, + 433, + 526, + 597 + ], + "type": "text", + "content": " in each instance, as shown in Figure 1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 597, + 527, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 597, + 527, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 597, + 527, + 773 + ], + "type": "text", + "content": "Position-Attentive Encoder we use BERT (Devlin et al., 2019) to encode a sentence " + }, + { + "bbox": [ + 302, + 597, + 527, + 773 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 302, + 597, + 527, + 773 + ], + "type": "text", + "content": " into token-level representations " + }, + { + "bbox": [ + 302, + 597, + 527, + 773 + ], + "type": "inline_equation", + "content": "h = \\{h_1,..,h_n\\}" + }, + { + "bbox": [ + 302, + 597, + 527, + 773 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 302, + 597, + 527, + 773 + ], + "type": "inline_equation", + "content": "h_i\\in \\mathbb{R}^d" + }, + { + "bbox": [ + 302, + 597, + 527, + 773 + ], + "type": "text", + "content": " is a " + }, + { + "bbox": [ + 302, + 597, + 527, + 773 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 302, + 597, + 527, + 773 + ], + "type": "text", + "content": "-dimensional vector corresponding to the " + }, + { + "bbox": [ + 302, + 597, + 527, + 773 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 302, + 597, + 527, + 773 + ], + "type": "text", + "content": "-th token in " + }, + { + "bbox": [ + 302, + 597, + 527, + 773 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 302, + 597, + 527, + 773 + ], + "type": "text", + "content": ". For each query " + }, + { + "bbox": [ + 302, + 597, + 527, + 773 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 302, + 597, + 527, + 773 + ], + "type": "text", + "content": ", self-matching is applied to calculate the position-attention " + }, + { + "bbox": [ + 302, + 597, + 527, + 773 + ], + "type": "inline_equation", + "content": "\\boldsymbol{a}_t\\in \\mathbb{R}^T" + }, + { + "bbox": [ + 302, + 597, + 527, + 773 + ], + "type": "text", + "content": " between token at " + }, + { + "bbox": [ + 302, + 597, + 527, + 773 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 302, + 597, + 527, + 773 + ], + "type": "text", + "content": " and each token at target position " + }, + { + "bbox": [ + 302, + 597, + 527, + 773 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 302, + 597, + 527, + 773 + ], + "type": "text", + "content": ", which compares the sentence representations against itself to collect context information (Tan et al., 2018). The produced position-aware representation " + }, + { + "bbox": [ + 302, + 597, + 527, + 773 + ], + "type": "inline_equation", + "content": "\\boldsymbol{c}_t\\in \\mathbb{R}^{T\\times d}" + }, + { + "bbox": [ + 302, + 597, + 527, + 773 + ], + "type": "text", + "content": " is an attention-weighted sentence vector " + }, + { + "bbox": [ + 302, + 597, + 527, + 773 + ], + "type": "inline_equation", + "content": "\\boldsymbol{c}_t = \\boldsymbol{a}_t^\\top \\boldsymbol{h}" + }, + { + "bbox": [ + 302, + 597, + 527, + 773 + ], + "type": "text", + "content": ". Finally, we concatenate " + }, + { + "bbox": [ + 302, + 597, + 527, + 773 + ], + "type": "inline_equation", + "content": "\\boldsymbol{h}_t" + }, + { + "bbox": [ + 302, + 597, + 527, + 773 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 302, + 597, + 527, + 773 + ], + "type": "inline_equation", + "content": "\\boldsymbol{c}_t" + }, + { + "bbox": [ + 302, + 597, + 527, + 773 + ], + "type": "text", + "content": " to generate position-aware and context" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 780, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 310, + 791 + ], + "type": "text", + "content": "1350" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 226, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 226, + 84 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 226, + 84 + ], + "type": "text", + "content": "aware representations " + }, + { + "bbox": [ + 67, + 71, + 226, + 84 + ], + "type": "inline_equation", + "content": "\\pmb{u}_t = [\\pmb{h}_t|\\pmb{c}_t]" + }, + { + "bbox": [ + 67, + 71, + 226, + 84 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 84, + 291, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 84, + 291, + 248 + ], + "spans": [ + { + "bbox": [ + 67, + 84, + 291, + 248 + ], + "type": "text", + "content": "CRF Decoder (Lafferty et al., 2001) For each position-aware representation " + }, + { + "bbox": [ + 67, + 84, + 291, + 248 + ], + "type": "inline_equation", + "content": "\\boldsymbol{u}_t" + }, + { + "bbox": [ + 67, + 84, + 291, + 248 + ], + "type": "text", + "content": ", we first learn a linear transformation " + }, + { + "bbox": [ + 67, + 84, + 291, + 248 + ], + "type": "inline_equation", + "content": "\\boldsymbol{z}_t = \\boldsymbol{W}\\boldsymbol{u}_t \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 67, + 84, + 291, + 248 + ], + "type": "text", + "content": " to represent tag scores for the " + }, + { + "bbox": [ + 67, + 84, + 291, + 248 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 67, + 84, + 291, + 248 + ], + "type": "text", + "content": "-th token. Here " + }, + { + "bbox": [ + 67, + 84, + 291, + 248 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 67, + 84, + 291, + 248 + ], + "type": "text", + "content": " is the number of distinct tags. For an instance with labels " + }, + { + "bbox": [ + 67, + 84, + 291, + 248 + ], + "type": "inline_equation", + "content": "\\boldsymbol{y} = \\{y_1, \\dots, y_n\\}" + }, + { + "bbox": [ + 67, + 84, + 291, + 248 + ], + "type": "text", + "content": ", the decoding score " + }, + { + "bbox": [ + 67, + 84, + 291, + 248 + ], + "type": "inline_equation", + "content": "s(\\boldsymbol{z}, \\boldsymbol{y})" + }, + { + "bbox": [ + 67, + 84, + 291, + 248 + ], + "type": "text", + "content": " is the sum of transition scores from tag " + }, + { + "bbox": [ + 67, + 84, + 291, + 248 + ], + "type": "inline_equation", + "content": "y_t" + }, + { + "bbox": [ + 67, + 84, + 291, + 248 + ], + "type": "text", + "content": " to tag " + }, + { + "bbox": [ + 67, + 84, + 291, + 248 + ], + "type": "inline_equation", + "content": "y_{t+1}" + }, + { + "bbox": [ + 67, + 84, + 291, + 248 + ], + "type": "text", + "content": " plus the input score " + }, + { + "bbox": [ + 67, + 84, + 291, + 248 + ], + "type": "inline_equation", + "content": "z_t^{yt}" + }, + { + "bbox": [ + 67, + 84, + 291, + 248 + ], + "type": "text", + "content": ". The conditional probability " + }, + { + "bbox": [ + 67, + 84, + 291, + 248 + ], + "type": "inline_equation", + "content": "p(\\boldsymbol{y}|\\boldsymbol{z})" + }, + { + "bbox": [ + 67, + 84, + 291, + 248 + ], + "type": "text", + "content": " is the softmax over " + }, + { + "bbox": [ + 67, + 84, + 291, + 248 + ], + "type": "inline_equation", + "content": "s(\\boldsymbol{z}, \\boldsymbol{y})" + }, + { + "bbox": [ + 67, + 84, + 291, + 248 + ], + "type": "text", + "content": " for all possible label sequences " + }, + { + "bbox": [ + 67, + 84, + 291, + 248 + ], + "type": "inline_equation", + "content": "\\boldsymbol{y}'" + }, + { + "bbox": [ + 67, + 84, + 291, + 248 + ], + "type": "text", + "content": ". We maximize the log-likelihood of correct tag sequences during training " + }, + { + "bbox": [ + 67, + 84, + 291, + 248 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c = \\sum \\log p(\\boldsymbol{y}|\\boldsymbol{z})" + }, + { + "bbox": [ + 67, + 84, + 291, + 248 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 257, + 290, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 257, + 290, + 271 + ], + "spans": [ + { + "bbox": [ + 67, + 257, + 290, + 271 + ], + "type": "text", + "content": "4 Uncertainty-Aware Bootstrap Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 278, + 291, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 278, + 291, + 428 + ], + "spans": [ + { + "bbox": [ + 67, + 278, + 291, + 428 + ], + "type": "text", + "content": "Motivation One of the main challenges in bootstrap learning is to evaluate the \"correctness\" of a labeled instance. We consider this problem from an uncertainty perspective and assume instances with lower uncertainty are more likely to be correctly labeled. In this section, we first propose instance-level data uncertainty which is used to filter noisy examples and build an initial subset. Then, we introduce our two-fold model uncertainties which helps iteratively mitigate DS effect and build up trustable examples during bootstrap learning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 436, + 177, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 436, + 177, + 449 + ], + "spans": [ + { + "bbox": [ + 67, + 436, + 177, + 449 + ], + "type": "text", + "content": "4.1 Data Uncertainty" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 454, + 291, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 454, + 291, + 615 + ], + "spans": [ + { + "bbox": [ + 67, + 454, + 291, + 615 + ], + "type": "text", + "content": "Presenting examples in an easy-to-hard order at different training stages can benefit models (Platanios et al., 2019; Zhou et al., 2020), we propose data uncertainty as a way to quantify the \"hardness\" of an instance. To better estimate the data uncertainty, we use pre-trained language models (PLMs) to generate tag probability for each token in a sequence. Our intuition is that higher uncertain inputs are \"harder\" to be generated by a PLM, as it already has rationales of language. Accordingly, we propose two data uncertainties, which can be used individually or combined together:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 616, + 291, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 616, + 291, + 698 + ], + "spans": [ + { + "bbox": [ + 67, + 616, + 291, + 698 + ], + "type": "text", + "content": "Winning Score (WS) The maximum softmax probability reflects data uncertainty of an input (Hendrycks and Gimpel, 2017). Given an input instance " + }, + { + "bbox": [ + 67, + 616, + 291, + 698 + ], + "type": "inline_equation", + "content": "\\mathcal{I} = \\{x_1,\\dots,x_n\\}" + }, + { + "bbox": [ + 67, + 616, + 291, + 698 + ], + "type": "text", + "content": ", we define data uncertainty " + }, + { + "bbox": [ + 67, + 616, + 291, + 698 + ], + "type": "inline_equation", + "content": "u^{d}(\\mathcal{I})" + }, + { + "bbox": [ + 67, + 616, + 291, + 698 + ], + "type": "text", + "content": " as the minus averaged token classification winning score:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 78, + 720, + 290, + 737 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 720, + 290, + 737 + ], + "spans": [ + { + "bbox": [ + 78, + 720, + 290, + 737 + ], + "type": "interline_equation", + "content": "u ^ {d} (\\mathcal {I}) = - \\frac {1}{n} \\sum_ {t = 1} ^ {n} \\max _ {c \\in [ 1, C ]} P (y _ {t} = c | x _ {t}) \\quad (1)", + "image_path": "4373222d8a9a704a601802936de516587ec81e301917bcdda4d29657d27881b1.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "type": "text", + "content": "Entropy Shannon entropy (Shannon, 1948) is widely used to reflect information uncertainty. We" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 71, + 525, + 98 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 525, + 98 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 525, + 98 + ], + "type": "text", + "content": "propose data uncertainty " + }, + { + "bbox": [ + 302, + 71, + 525, + 98 + ], + "type": "inline_equation", + "content": "u^{d}(\\mathcal{I})" + }, + { + "bbox": [ + 302, + 71, + 525, + 98 + ], + "type": "text", + "content": " as the averaged token classification entropy:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 124, + 525, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 124, + 525, + 152 + ], + "spans": [ + { + "bbox": [ + 304, + 124, + 525, + 152 + ], + "type": "interline_equation", + "content": "u ^ {d} (\\mathcal {I}) = \\frac {1}{n} \\sum_ {t = 1} ^ {n} \\sum_ {c = 1} ^ {C} P \\left(y _ {t} = c \\mid x _ {t}\\right) \\log P \\left(y _ {t} = c \\mid x _ {t}\\right) \\tag {2}", + "image_path": "e6afb5b75f4a25b7d3e83d30bb82d5070086dc6ebc51dd780c204510022416c0.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 153, + 526, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 153, + 526, + 235 + ], + "spans": [ + { + "bbox": [ + 302, + 153, + 526, + 235 + ], + "type": "text", + "content": "We filter out examples with high uncertainty scores and build an initial subset with \"simple\" examples. At the early training stage, a model is not aware of what a decent distribution " + }, + { + "bbox": [ + 302, + 153, + 526, + 235 + ], + "type": "inline_equation", + "content": "P(y|x)" + }, + { + "bbox": [ + 302, + 153, + 526, + 235 + ], + "type": "text", + "content": " should be, thus data uncertainty facilitates it to converge fast by tuning on a fairly \"simple\" subset." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 246, + 420, + 259 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 246, + 420, + 259 + ], + "spans": [ + { + "bbox": [ + 302, + 246, + 420, + 259 + ], + "type": "text", + "content": "4.2 Model Uncertainty" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 264, + 526, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 264, + 526, + 495 + ], + "spans": [ + { + "bbox": [ + 302, + 264, + 526, + 495 + ], + "type": "text", + "content": "In our bootstrap learning, we define model uncertainty, i.e., epistemic uncertainty (Kendall and Gal, 2017), to measure whether model parameters can best describe the data distribution following (Zhou et al., 2020). A small model uncertainty indicates the model is confident that the current training data has been well learned (Wang et al., 2019). We adopt Monte Carlo Dropout (Gal and Ghahramani, 2016) to approximate Bayesian inference which captures inner-model parametric uncertainty. Specifically, we perform " + }, + { + "bbox": [ + 302, + 264, + 526, + 495 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 302, + 264, + 526, + 495 + ], + "type": "text", + "content": " forward passes through our joint model. In each pass, part of network neurons " + }, + { + "bbox": [ + 302, + 264, + 526, + 495 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 302, + 264, + 526, + 495 + ], + "type": "text", + "content": " are randomly deactivated. Finally, we yield " + }, + { + "bbox": [ + 302, + 264, + 526, + 495 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 302, + 264, + 526, + 495 + ], + "type": "text", + "content": " samples on model parameters " + }, + { + "bbox": [ + 302, + 264, + 526, + 495 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\theta}_1,\\dots,\\hat{\\theta}_K\\}" + }, + { + "bbox": [ + 302, + 264, + 526, + 495 + ], + "type": "text", + "content": ". We use the averaged token classification Probability Variance (PV) (Shelmanov et al., 2021) over all tags for instance " + }, + { + "bbox": [ + 302, + 264, + 526, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 302, + 264, + 526, + 495 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 518, + 525, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 525, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 525, + 552 + ], + "type": "interline_equation", + "content": "u ^ {m} (\\theta) = \\frac {1}{n} \\sum_ {t = 1} ^ {n} \\sum_ {c = 1} ^ {C} \\operatorname {V a r} \\left[ P \\left(y _ {t} = c \\mid x _ {t}, \\hat {\\theta} _ {k}\\right) \\right] _ {k = 1} ^ {K} \\tag {3}", + "image_path": "82f1a6cbd39dba95fe805a15a1c2bb0fd2804165bf7a7feccfb97a9af2fa8ec3.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 554, + 525, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 554, + 525, + 622 + ], + "spans": [ + { + "bbox": [ + 302, + 554, + 525, + 622 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 302, + 554, + 525, + 622 + ], + "type": "inline_equation", + "content": "\\operatorname{Var}[\\cdot]" + }, + { + "bbox": [ + 302, + 554, + 525, + 622 + ], + "type": "text", + "content": " is the variance of distribution over the " + }, + { + "bbox": [ + 302, + 554, + 525, + 622 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 302, + 554, + 525, + 622 + ], + "type": "text", + "content": " passes following the common settings in (Dong et al., 2018; Xiao and Wang, 2019). Accordingly, model is aware of its confidence over each instance and how likely the label is noisy." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 633, + 414, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 633, + 414, + 646 + ], + "spans": [ + { + "bbox": [ + 302, + 633, + 414, + 646 + ], + "type": "text", + "content": "4.3 Training Strategy" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 302, + 651, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 651, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 651, + 526, + 772 + ], + "type": "text", + "content": "Uncertainty-Aware Loss Besides MC Dropout which measures parametric uncertainty within a model, we also consider mitigating parametric uncertainty between models to stabilize the weights during training. Specifically, we use self-ensembling (He et al., 2020; Wang and Wang, 2022) to calculate the loss between the same models to improve model robustness and reduce the label noise effect on model performance." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 780, + 308, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 780, + 308, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 780, + 308, + 791 + ], + "type": "text", + "content": "1351" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 70, + 215, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 70, + 215, + 84 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 215, + 84 + ], + "type": "text", + "content": "Algorithm 1 Bootstrap Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 88, + 289, + 116 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 88, + 289, + 116 + ], + "spans": [ + { + "bbox": [ + 69, + 88, + 289, + 116 + ], + "type": "text", + "content": "Input: Original dataset " + }, + { + "bbox": [ + 69, + 88, + 289, + 116 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(\\mathcal{I}^n,y^n)\\}_{n = 1}^N" + }, + { + "bbox": [ + 69, + 88, + 289, + 116 + ], + "type": "text", + "content": " two joint models " + }, + { + "bbox": [ + 69, + 88, + 289, + 116 + ], + "type": "inline_equation", + "content": "f_{1},f_{2}" + }, + { + "bbox": [ + 69, + 88, + 289, + 116 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 69, + 88, + 289, + 116 + ], + "type": "inline_equation", + "content": "\\theta_{1},\\theta_{2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 74, + 117, + 290, + 238 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 74, + 117, + 290, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 117, + 290, + 142 + ], + "spans": [ + { + "bbox": [ + 74, + 117, + 290, + 142 + ], + "type": "text", + "content": "1: Compute data uncertainty " + }, + { + "bbox": [ + 74, + 117, + 290, + 142 + ], + "type": "inline_equation", + "content": "u^{d}(\\mathcal{I})" + }, + { + "bbox": [ + 74, + 117, + 290, + 142 + ], + "type": "text", + "content": " for each instance " + }, + { + "bbox": [ + 74, + 117, + 290, + 142 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 74, + 117, + 290, + 142 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 74, + 117, + 290, + 142 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 74, + 117, + 290, + 142 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 74, + 144, + 290, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 144, + 290, + 170 + ], + "spans": [ + { + "bbox": [ + 74, + 144, + 290, + 170 + ], + "type": "text", + "content": "2: Initial dataset " + }, + { + "bbox": [ + 74, + 144, + 290, + 170 + ], + "type": "inline_equation", + "content": "\\mathcal{C} \\gets" + }, + { + "bbox": [ + 74, + 144, + 290, + 170 + ], + "type": "text", + "content": " Select data pairs " + }, + { + "bbox": [ + 74, + 144, + 290, + 170 + ], + "type": "inline_equation", + "content": "(\\mathcal{I}^n, y^n)" + }, + { + "bbox": [ + 74, + 144, + 290, + 170 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 74, + 144, + 290, + 170 + ], + "type": "inline_equation", + "content": "u^d(\\mathcal{I}) < \\tau^d" + }, + { + "bbox": [ + 74, + 144, + 290, + 170 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 74, + 144, + 290, + 170 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 74, + 144, + 290, + 170 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 74, + 172, + 188, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 172, + 188, + 184 + ], + "spans": [ + { + "bbox": [ + 74, + 172, + 188, + 184 + ], + "type": "text", + "content": "3: for epoch " + }, + { + "bbox": [ + 74, + 172, + 188, + 184 + ], + "type": "inline_equation", + "content": "e = 1, \\ldots" + }, + { + "bbox": [ + 74, + 172, + 188, + 184 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 74, + 185, + 242, + 197 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 185, + 242, + 197 + ], + "spans": [ + { + "bbox": [ + 74, + 185, + 242, + 197 + ], + "type": "text", + "content": "4: Train " + }, + { + "bbox": [ + 74, + 185, + 242, + 197 + ], + "type": "inline_equation", + "content": "f_{1}, f_{2}" + }, + { + "bbox": [ + 74, + 185, + 242, + 197 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 74, + 185, + 242, + 197 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 74, + 185, + 242, + 197 + ], + "type": "text", + "content": " using Eq. (5);" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 74, + 199, + 290, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 199, + 290, + 211 + ], + "spans": [ + { + "bbox": [ + 74, + 199, + 290, + 211 + ], + "type": "text", + "content": "5: Calculate model uncertainty " + }, + { + "bbox": [ + 74, + 199, + 290, + 211 + ], + "type": "inline_equation", + "content": "u^{m}(\\theta_{1})" + }, + { + "bbox": [ + 74, + 199, + 290, + 211 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 74, + 199, + 290, + 211 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 74, + 199, + 290, + 211 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 74, + 212, + 290, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 212, + 290, + 238 + ], + "spans": [ + { + "bbox": [ + 74, + 212, + 290, + 238 + ], + "type": "text", + "content": "6: " + }, + { + "bbox": [ + 74, + 212, + 290, + 238 + ], + "type": "inline_equation", + "content": "\\mathcal{C} \\gets" + }, + { + "bbox": [ + 74, + 212, + 290, + 238 + ], + "type": "text", + "content": " Select data pairs " + }, + { + "bbox": [ + 74, + 212, + 290, + 238 + ], + "type": "inline_equation", + "content": "(\\mathcal{I}^n, y^n)" + }, + { + "bbox": [ + 74, + 212, + 290, + 238 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 74, + 212, + 290, + 238 + ], + "type": "inline_equation", + "content": "u^m(\\mathcal{I}; \\theta_1) < \\tau^m" + }, + { + "bbox": [ + 74, + 212, + 290, + 238 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 74, + 212, + 290, + 238 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 74, + 212, + 290, + 238 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 67, + 268, + 290, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 268, + 290, + 336 + ], + "spans": [ + { + "bbox": [ + 67, + 268, + 290, + 336 + ], + "type": "text", + "content": "We create another joint model with identical framework, e.g., architecture, loss functions, hyperparameters, and compute a self-ensemble loss " + }, + { + "bbox": [ + 67, + 268, + 290, + 336 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_e" + }, + { + "bbox": [ + 67, + 268, + 290, + 336 + ], + "type": "text", + "content": " to minimize the difference between two outputs from the two models regarding the same inputs:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 103, + 357, + 290, + 375 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 357, + 290, + 375 + ], + "spans": [ + { + "bbox": [ + 103, + 357, + 290, + 375 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {e} = \\sum K L (f (\\mathcal {I}; \\theta_ {1}), f (\\mathcal {I}; \\theta_ {2})) \\tag {4}", + "image_path": "4dab930f79bbad6186c0f78949ef2cdfb14d41941d1e5f2a94ee465f3e045fa6.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 391, + 290, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 391, + 290, + 458 + ], + "spans": [ + { + "bbox": [ + 67, + 391, + 290, + 458 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 67, + 391, + 290, + 458 + ], + "type": "inline_equation", + "content": "KL(.)" + }, + { + "bbox": [ + 67, + 391, + 290, + 458 + ], + "type": "text", + "content": " is the Kullback-Leibler divergence between two probabilistic distributions, " + }, + { + "bbox": [ + 67, + 391, + 290, + 458 + ], + "type": "inline_equation", + "content": "\\theta_{1},\\theta_{2}" + }, + { + "bbox": [ + 67, + 391, + 290, + 458 + ], + "type": "text", + "content": " denote the parameters of first and second models. We formulate our final uncertainty-aware objective " + }, + { + "bbox": [ + 67, + 391, + 290, + 458 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 67, + 391, + 290, + 458 + ], + "type": "text", + "content": " as the sum of CRF and self-ensemble loss:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 144, + 483, + 290, + 496 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 483, + 290, + 496 + ], + "spans": [ + { + "bbox": [ + 144, + 483, + 290, + 496 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {c} + \\alpha \\mathcal {L} _ {e} \\tag {5}", + "image_path": "e3d24fab2dc0d4d1cb3a588414708e0b4b67e827cc7098cf3686138f35bbf611.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 67, + 513, + 290, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 513, + 290, + 539 + ], + "spans": [ + { + "bbox": [ + 67, + 513, + 290, + 539 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 67, + 513, + 290, + 539 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 67, + 513, + 290, + 539 + ], + "type": "text", + "content": " denotes the weight of self-ensembling, and " + }, + { + "bbox": [ + 67, + 513, + 290, + 539 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 67, + 513, + 290, + 539 + ], + "type": "text", + "content": " means the token classification loss." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 67, + 543, + 290, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 543, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 543, + 290, + 772 + ], + "type": "text", + "content": "Bootstrap Learning Procedure To mitigate the DS effect on model performance, we propose a twofold bootstrap learning strategy (see Algorithm 1). Specifically, we first apply data uncertainty to filter \"harder\" examples and redistribute a reliable initial training data " + }, + { + "bbox": [ + 67, + 543, + 290, + 772 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 67, + 543, + 290, + 772 + ], + "type": "text", + "content": ". Then, we iteratively feed examples following an easy-to-hard order to the model. In each training iteration, we regularize the joint model with self-ensembling loss to reduce the impact of noisy labels on model parameters. Then we use probability variance to select new confident training instances " + }, + { + "bbox": [ + 67, + 543, + 290, + 772 + ], + "type": "inline_equation", + "content": "\\mathcal{D}'" + }, + { + "bbox": [ + 67, + 543, + 290, + 772 + ], + "type": "text", + "content": " that can be explained by the model as the next training inputs. The more certain examples are selected, the more likely the model will learn beneficial information and will converge faster. We repeat the above procedure until the F1 score on the validation set converges." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 303, + 70, + 389, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 70, + 389, + 84 + ], + "spans": [ + { + "bbox": [ + 303, + 70, + 389, + 84 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 303, + 92, + 358, + 105 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 92, + 358, + 105 + ], + "spans": [ + { + "bbox": [ + 303, + 92, + 358, + 105 + ], + "type": "text", + "content": "5.1 Setup" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 302, + 110, + 525, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 110, + 525, + 244 + ], + "spans": [ + { + "bbox": [ + 302, + 110, + 525, + 244 + ], + "type": "text", + "content": "We evaluate the performance of UnBED on two datasets, NYT and Wiki-KBP. The NYT (Riedel et al., 2010) dataset collects news from New York Times and its training data is automatically labeled by DS. We use the revised test dataset (Jia et al., 2019) that is manually annotated to ensure quality. The Wiki-KBP (Ling and Weld, 2012) dataset collects articles from Wikipedia. Its training data is labeled by DS (Liu et al., 2017), and the test set is manually annotated (Ellis et al., 2013)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 302, + 245, + 525, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 245, + 525, + 380 + ], + "spans": [ + { + "bbox": [ + 302, + 245, + 525, + 380 + ], + "type": "text", + "content": "We compare UnBED with the following baselines: ARNOR (Jia et al., 2019), a pattern-based method to reduce noise for distantly-supervised triplet extraction. PURE (Zhong and Chen, 2021), a pipeline approach that uses pre-trained BERT entity model to first recognize entities and then employs a relation model to detect underlying relations. FAN (Hao et al., 2021), an adversarial method including a transformers encoder to reduce noise for distantly-supervised triplet extraction." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 302, + 381, + 525, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 381, + 525, + 461 + ], + "spans": [ + { + "bbox": [ + 302, + 381, + 525, + 461 + ], + "type": "text", + "content": "Evaluation We evaluate the extracted triplets for each sentence based on Precision (Prec.), Recall (Rec.), and F1. A triplet " + }, + { + "bbox": [ + 302, + 381, + 525, + 461 + ], + "type": "inline_equation", + "content": "\\{e_1, re, e_2\\}" + }, + { + "bbox": [ + 302, + 381, + 525, + 461 + ], + "type": "text", + "content": " is marked correct if the relation type " + }, + { + "bbox": [ + 302, + 381, + 525, + 461 + ], + "type": "inline_equation", + "content": "re" + }, + { + "bbox": [ + 302, + 381, + 525, + 461 + ], + "type": "text", + "content": ", two entities " + }, + { + "bbox": [ + 302, + 381, + 525, + 461 + ], + "type": "inline_equation", + "content": "e_1, e_2" + }, + { + "bbox": [ + 302, + 381, + 525, + 461 + ], + "type": "text", + "content": " are all correct. We build a validation set by randomly sampling " + }, + { + "bbox": [ + 302, + 381, + 525, + 461 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 302, + 381, + 525, + 461 + ], + "type": "text", + "content": " sentences from the test set." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 302, + 462, + 525, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 462, + 525, + 624 + ], + "spans": [ + { + "bbox": [ + 302, + 462, + 525, + 624 + ], + "type": "text", + "content": "Implementation Details We use Hugging Face bert-large-uncased (Devlin et al., 2019) pre-trained model as backbone. For ARNOR, the hidden vector size is set to 300. In regularization training, we find optimal parameters " + }, + { + "bbox": [ + 302, + 462, + 525, + 624 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 302, + 462, + 525, + 624 + ], + "type": "text", + "content": " as 1 for both datasets. We implement UnBED and all baselines in PyTorch, with Adam optimizer, initial learning rate " + }, + { + "bbox": [ + 302, + 462, + 525, + 624 + ], + "type": "inline_equation", + "content": "10^{-5}" + }, + { + "bbox": [ + 302, + 462, + 525, + 624 + ], + "type": "text", + "content": ", dropout rate 0.1, and batch size 8. For initial subset configuration, we choose data uncertainty threshold 0.5. For bootstrap learning, an empirical model uncertainty threshold is set to 0.6 with the best validation F1." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 302, + 634, + 402, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 634, + 402, + 645 + ], + "spans": [ + { + "bbox": [ + 302, + 634, + 402, + 645 + ], + "type": "text", + "content": "5.2 Overall Results" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 302, + 651, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 651, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 651, + 525, + 772 + ], + "type": "text", + "content": "As shown in Table 1, UnBED significantly outperforms all baselines in precision and F1 metric. Specifically, UnBED achieves " + }, + { + "bbox": [ + 302, + 651, + 525, + 772 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 302, + 651, + 525, + 772 + ], + "type": "text", + "content": " F1 improvement on NYT (3% on Wiki-KBP) over denoising approaches—ARNOR and FAN. Our approach also outperforms baselines using pretrained transformers (PURE and FAN), showing that uncertainty-aware bootstrap learning effectively reduces the impact of noisy labels." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 780, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 780, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 780, + 310, + 791 + ], + "type": "text", + "content": "1352" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 111, + 68, + 483, + 168 + ], + "blocks": [ + { + "bbox": [ + 111, + 68, + 483, + 168 + ], + "lines": [ + { + "bbox": [ + 111, + 68, + 483, + 168 + ], + "spans": [ + { + "bbox": [ + 111, + 68, + 483, + 168 + ], + "type": "table", + "html": "
MethodNYTWiki-KBP
Prec.Rec.F1Prec.Rec.F1
ARNOR (Jia et al., 2019)0.5880.6140.6000.4020.4710.434
PURE (Zhong and Chen, 2021)0.5360.6640.5930.3950.4330.413
FAN (Hao et al., 2021)0.5790.6460.6110.3910.4670.426
UnBED-WS0.6620.7300.6940.4290.5010.462
UnBED-Entropy0.6510.7410.6930.4220.5090.461
", + "image_path": "233e3b0b2d247d6c323133f9234c64f0d04deb61933c9c4eae036c8234df3694.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 175, + 525, + 200 + ], + "lines": [ + { + "bbox": [ + 67, + 175, + 525, + 200 + ], + "spans": [ + { + "bbox": [ + 67, + 175, + 525, + 200 + ], + "type": "text", + "content": "Table 1: Evaluation results on NYT and Wiki-KBP datasets. Bold numbers denote the best metrics. UnBED-WS and UnBED-Entropy denote UnBED with winning score and entropy as the data uncertainty, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 79, + 221, + 280, + 344 + ], + "blocks": [ + { + "bbox": [ + 79, + 221, + 280, + 344 + ], + "lines": [ + { + "bbox": [ + 79, + 221, + 280, + 344 + ], + "spans": [ + { + "bbox": [ + 79, + 221, + 280, + 344 + ], + "type": "image", + "image_path": "0b1cd8eb90ceb16ce072eac5bd6e552b67596131705df3883ebf1cfe9ff3f8eb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 355, + 291, + 393 + ], + "lines": [ + { + "bbox": [ + 67, + 355, + 291, + 393 + ], + "spans": [ + { + "bbox": [ + 67, + 355, + 291, + 393 + ], + "type": "text", + "content": "Figure 2: F1 score vs. Epochs under different settings. Vanilla-PV-enssembled denotes UnBED-WS, and entropy-PV-enssembled denotes UnBED-Entropy." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 413, + 175, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 413, + 175, + 427 + ], + "spans": [ + { + "bbox": [ + 67, + 413, + 175, + 427 + ], + "type": "text", + "content": "5.3 Further Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 430, + 291, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 430, + 291, + 581 + ], + "spans": [ + { + "bbox": [ + 67, + 430, + 291, + 581 + ], + "type": "text", + "content": "We analyze the functionality of different components in Figure 2. We observe that both the entropy-PV and vanilla-PV outperform the baseline (joint model directly trained on the original DS dataset) in terms of F1 " + }, + { + "bbox": [ + 67, + 430, + 291, + 581 + ], + "type": "inline_equation", + "content": "(5\\sim 7\\%)" + }, + { + "bbox": [ + 67, + 430, + 291, + 581 + ], + "type": "text", + "content": " increase), demonstrating the effect of filtering noisy labels and selecting trustable instance using probability variance. Besides, self-ensembling further enhances the performance in later training stage " + }, + { + "bbox": [ + 67, + 430, + 291, + 581 + ], + "type": "inline_equation", + "content": "(2\\sim 4" + }, + { + "bbox": [ + 67, + 430, + 291, + 581 + ], + "type": "text", + "content": " F1 increase), proving that mitigating the inter-model uncertainty benefits model robustness against noisy labels." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 590, + 151, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 590, + 151, + 602 + ], + "spans": [ + { + "bbox": [ + 67, + 590, + 151, + 602 + ], + "type": "text", + "content": "6 Conclusions" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 611, + 292, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 611, + 292, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 611, + 292, + 772 + ], + "type": "text", + "content": "In this paper, we propose a novel uncertainty-aware bootstrap learning framework for distantly-supervised joint extraction. Specifically, we define data uncertainty in generally token classification to filter out highly-error-prone instances and build an initial high-confident subset, which is used to tune the joint extraction model for fast convergence. We then propose a two-fold bootstrap learning procedure which iteratively mitigates the DS impact on model robustness and selects new trustable training instances. Experimental results on two benchmark datasets show that UnBED significantly out" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 222, + 467, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 222, + 467, + 235 + ], + "spans": [ + { + "bbox": [ + 302, + 222, + 467, + 235 + ], + "type": "text", + "content": "performs other denoising techniques." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 303, + 244, + 365, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 244, + 365, + 257 + ], + "spans": [ + { + "bbox": [ + 303, + 244, + 365, + 257 + ], + "type": "text", + "content": "Limitations" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 265, + 526, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 265, + 526, + 441 + ], + "spans": [ + { + "bbox": [ + 302, + 265, + 526, + 441 + ], + "type": "text", + "content": "In this work we propose an uncertainty-aware bootstrap learning framework for joint extraction. Though it achieves state-of-the-art performance compared to other denoising techniques, UnBED requires large training resources considering the ensemble loss calculated between two large PLMs and the probability variance calculated on the PLM joint extraction model. In our future work, we hope to incorporate pruning techniques during training to improve the efficiency. We will also consider more complex relations between entities, e.g., relations beyond the sentence boundary, to fit in real-world information extraction scenarios." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 303, + 451, + 405, + 465 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 451, + 405, + 465 + ], + "spans": [ + { + "bbox": [ + 303, + 451, + 405, + 465 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 472, + 525, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 472, + 525, + 513 + ], + "spans": [ + { + "bbox": [ + 302, + 472, + 525, + 513 + ], + "type": "text", + "content": "This work was supported by NSF CNS 2135625, CPS 2038727, CNS Career 1750263, and a Darpa Shell grant." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 535, + 362, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 535, + 362, + 549 + ], + "spans": [ + { + "bbox": [ + 304, + 535, + 362, + 549 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 554, + 527, + 772 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 304, + 554, + 526, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 554, + 526, + 666 + ], + "spans": [ + { + "bbox": [ + 304, + 554, + 526, + 666 + ], + "type": "text", + "content": "Tao Chen, Haochen Shi, Liyuan Liu, Siliang Tang, Jian Shao, Zhigang Chen, and Yueting Zhuang. 2021. Empower distantly supervised relation extraction with collaborative adversarial training. In Thirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2021, Thirty-Third Conference on Innovative Applications of Artificial Intelligence, IAAI 2021, The Eleventh Symposium on Educational Advances in Artificial Intelligence, EAAI 2021, Virtual Event, February 2-9, 2021, pages 12675-12682. AAAI Press." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 672, + 527, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 672, + 527, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 672, + 527, + 772 + ], + "type": "text", + "content": "Dai Dai, Xinyan Xiao, Yajuan Lyu, Shan Dou, Qiaoqiao She, and Haifeng Wang. 2019. Joint extraction of entities and overlapping relations using position-attentive sequence labeling. In The Thirty-Third AAAI Conference on Artificial Intelligence, AAAI 2019, The Thirty-First Innovative Applications of Artificial Intelligence Conference, IAAI 2019, The Ninth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2019, Honolulu, Hawaii," + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 780, + 309, + 791 + ], + "type": "text", + "content": "1353" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 80, + 72, + 291, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 72, + 291, + 95 + ], + "spans": [ + { + "bbox": [ + 80, + 72, + 291, + 95 + ], + "type": "text", + "content": "USA, January 27 - February 1, 2019, pages 6300-6308. AAAI Press." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 103, + 291, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 103, + 291, + 202 + ], + "spans": [ + { + "bbox": [ + 69, + 103, + 291, + 202 + ], + "type": "text", + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 211, + 291, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 211, + 291, + 278 + ], + "spans": [ + { + "bbox": [ + 69, + 211, + 291, + 278 + ], + "type": "text", + "content": "Li Dong, Chris Quirk, and Mirella Lapata. 2018. Confidence modeling for neural semantic parsing. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 743-753, Melbourne, Australia. Association for Computational Linguistics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 286, + 291, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 286, + 291, + 363 + ], + "spans": [ + { + "bbox": [ + 69, + 286, + 291, + 363 + ], + "type": "text", + "content": "Joe Ellis, Jeremy Getman, Justin Mott, Xuansong Li, Kira Griffith, Stephanie M. Strassel, and Jonathan Wright. 2013. Linguistic resources for 2013 knowledge base population evaluations. In Proceedings of the Sixth Text Analysis Conference, TAC 2013, Gaithersburg, Maryland, USA, November 18-19, 2013. NIST." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 372, + 291, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 372, + 291, + 450 + ], + "spans": [ + { + "bbox": [ + 69, + 372, + 291, + 450 + ], + "type": "text", + "content": "Yarin Gal and Zoubin Ghahramani. 2016. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In Proceedings of the 33nd International Conference on Machine Learning, ICML 2016, New York City, NY, USA, June 19-24, 2016, volume 48 of JMLR Workshop and Conference Proceedings, pages 1050-1059. JMLR.org." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 458, + 291, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 458, + 291, + 535 + ], + "spans": [ + { + "bbox": [ + 69, + 458, + 291, + 535 + ], + "type": "text", + "content": "Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. 2017. On calibration of modern neural networks. In Proceedings of the 34th International Conference on Machine Learning, ICML 2017, Sydney, NSW, Australia, 6-11 August 2017, volume 70 of Proceedings of Machine Learning Research, pages 1321-1330. PMLR." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 544, + 291, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 544, + 291, + 622 + ], + "spans": [ + { + "bbox": [ + 69, + 544, + 291, + 622 + ], + "type": "text", + "content": "Pankaj Gupta, Hinrich Schütze, and Bernt Andrassy. 2016. Table filling multi-task recurrent neural network for joint entity and relation extraction. In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 2537-2547, Osaka, Japan. The COLING 2016 Organizing Committee." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 630, + 291, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 630, + 291, + 708 + ], + "spans": [ + { + "bbox": [ + 69, + 630, + 291, + 708 + ], + "type": "text", + "content": "Kailong Hao, Botao Yu, and Wei Hu. 2021. Knowing false negatives: An adversarial training method for distantly supervised relation extraction. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 9661-9672, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 717, + 291, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 717, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 717, + 291, + 772 + ], + "type": "text", + "content": "Jianfeng He, Xuchao Zhang, Shuo Lei, Zhiqian Chen, Fanglan Chen, Abdulaziz Alhamadani, Bei Xiao, and ChangTien Lu. 2020. Towards more accurate uncertainty estimation in text classification. In Proceedings of the 2020 Conference on Empirical Methods" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 526, + 772 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 314, + 72, + 524, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 524, + 105 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 524, + 105 + ], + "type": "text", + "content": "in Natural Language Processing (EMNLP), pages 8362-8372, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 116, + 526, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 116, + 526, + 184 + ], + "spans": [ + { + "bbox": [ + 304, + 116, + 526, + 184 + ], + "type": "text", + "content": "Dan Hendrycks and Kevin Gimpel. 2017. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 195, + 526, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 195, + 526, + 293 + ], + "spans": [ + { + "bbox": [ + 304, + 195, + 526, + 293 + ], + "type": "text", + "content": "Xuming Hu, Chenwei Zhang, Yawen Yang, Xiaohe Li, Li Lin, Lijie Wen, and Philip S. Yu. 2021. Gradient imitation reinforcement learning for low resource relation extraction. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, EMNLP 2021, Virtual Event / Punta Cana, Dominican Republic, 7-11 November; 2021, pages 2737-2746. Association for Computational Linguistics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 305, + 526, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 305, + 526, + 383 + ], + "spans": [ + { + "bbox": [ + 304, + 305, + 526, + 383 + ], + "type": "text", + "content": "Wei Jia, Dai Dai, Xinyan Xiao, and Hua Wu. 2019. ARNOR: Attention regularization based noise reduction for distant supervision relation classification. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1399-1408, Florence, Italy. Association for Computational Linguistics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 394, + 526, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 394, + 526, + 461 + ], + "spans": [ + { + "bbox": [ + 304, + 394, + 526, + 461 + ], + "type": "text", + "content": "Alex Kendall and Yarin Gal. 2017. What uncertainties do we need in bayesian deep learning for computer vision? In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 5574-5584." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 472, + 526, + 539 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 526, + 539 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 526, + 539 + ], + "type": "text", + "content": "Aaron Klein, Stefan Falkner, Jost Tobias Springenberg, and Frank Hutter. 2017. Learning curve prediction with bayesian neural networks. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 550, + 526, + 627 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 550, + 526, + 627 + ], + "spans": [ + { + "bbox": [ + 304, + 550, + 526, + 627 + ], + "type": "text", + "content": "John D. Lafferty, Andrew McCallum, and Fernando C. N. Pereira. 2001. Conditional random fields: Probabilistic models for segmenting and labeling sequence data. In Proceedings of the Eighteenth International Conference on Machine Learning, ICML '01, page 282-289, San Francisco, CA, USA. Morgan Kaufmann Publishers Inc." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 639, + 526, + 717 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 639, + 526, + 717 + ], + "spans": [ + { + "bbox": [ + 304, + 639, + 526, + 717 + ], + "type": "text", + "content": "Shuyang Li, Yufei Li, Jianmo Ni, and Julian McAuley. 2022. SHARE: a system for hierarchical assistive recipe editing. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 11077-11090, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 728, + 526, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 728, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 728, + 526, + 772 + ], + "type": "text", + "content": "Xiao Ling and Daniel S. Weld. 2012. Fine-grained entity recognition. In Proceedings of the Twenty-Sixth AAAI Conference on Artificial Intelligence, July 22-26, 2012, Toronto, Ontario, Canada. AAAI Press." + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1354" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 291, + 150 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 291, + 150 + ], + "type": "text", + "content": "Liyuan Liu, Xiang Ren, Qi Zhu, Shi Zhi, Huan Gui, Heng Ji, and Jiawei Han. 2017. Heterogeneous supervision for relation extraction: A representation learning approach. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 46-56, Copenhagen, Denmark. Association for Computational Linguistics." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 158, + 291, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 158, + 291, + 247 + ], + "spans": [ + { + "bbox": [ + 69, + 158, + 291, + 247 + ], + "type": "text", + "content": "Mike Mintz, Steven Bills, Rion Snow, and Daniel Jurafsky. 2009. Distant supervision for relation extraction without labeled data. In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP, pages 1003-1011, Suntec, Singapore. Association for Computational Linguistics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 255, + 291, + 333 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 255, + 291, + 333 + ], + "spans": [ + { + "bbox": [ + 69, + 255, + 291, + 333 + ], + "type": "text", + "content": "Makoto Miwa and Mohit Bansal. 2016. End-to-end relation extraction using lstms on sequences and tree structures. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, ACL 2016, August 7-12, 2016, Berlin, Germany, Volume 1: Long Papers. The Association for Computer Linguistics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 341, + 291, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 341, + 291, + 430 + ], + "spans": [ + { + "bbox": [ + 69, + 341, + 291, + 430 + ], + "type": "text", + "content": "Farhad Nooralahzadeh, Jan Tore Lönning, and Lilja Øvrelid. 2019. Reinforcement-based denoising of distantly supervised NER with partial annotation. In Proceedings of the 2nd Workshop on Deep Learning Approaches for Low-Resource NLP, DeepLo@EMNLP-IJCNLP 2019, Hong Kong, China, November 3, 2019, pages 225–233. Association for Computational Linguistics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 439, + 291, + 549 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 439, + 291, + 549 + ], + "spans": [ + { + "bbox": [ + 69, + 439, + 291, + 549 + ], + "type": "text", + "content": "Emmanouil Antonios Platanios, Otilia Stretcu, Graham Neubig, Barnabás Póczos, and Tom M. Mitchell. 2019. Competence-based curriculum learning for neural machine translation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 1162-1172. Association for Computational Linguistics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 557, + 290, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 557, + 290, + 591 + ], + "spans": [ + { + "bbox": [ + 69, + 557, + 290, + 591 + ], + "type": "text", + "content": "Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 599, + 291, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 599, + 291, + 688 + ], + "spans": [ + { + "bbox": [ + 69, + 599, + 291, + 688 + ], + "type": "text", + "content": "Sebastian Riedel, Limin Yao, and Andrew McCallum. 2010. Modeling relations and their mentions without labeled text. In Machine Learning and Knowledge Discovery in Databases, European Conference, ECML PKDD 2010, Barcelona, Spain, September 20-24, 2010, Proceedings, Part III, volume 6323 of Lecture Notes in Computer Science, pages 148-163. Springer." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 697, + 291, + 740 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 697, + 291, + 740 + ], + "spans": [ + { + "bbox": [ + 69, + 697, + 291, + 740 + ], + "type": "text", + "content": "Yuming Shang, Heyan Huang, Xin Sun, Wei Wei, and Xian-Ling Mao. 2022. A pattern-aware self-attention network for distant supervised relation extraction. Inf. Sci., 584:269-279." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 750, + 290, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 750, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 750, + 290, + 772 + ], + "type": "text", + "content": "Claude E. Shannon. 1948. A mathematical theory of communication. Bell Syst. Tech. J., 27(3):379-423." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 526, + 772 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 304, + 72, + 526, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 526, + 150 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 526, + 150 + ], + "type": "text", + "content": "Artem Shelmanov, Evgenii Tsymbalov, Dmitri Puzyrev, Kirill Fedyanin, Alexander Panchenko, and Maxim Panov. 2021. How certain is your Transformer? In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 1833-1840, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 158, + 526, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 158, + 526, + 246 + ], + "spans": [ + { + "bbox": [ + 304, + 158, + 526, + 246 + ], + "type": "text", + "content": "Zhixing Tan, Mingxuan Wang, Jun Xie, Yidong Chen, and Xiaodong Shi. 2018. Deep semantic role labeling with self-attention. In Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence and Thirtieth Innovative Applications of Artificial Intelligence Conference and Eighth AAAI Symposium on Educational Advances in Artificial Intelligence, AAAI'18/IAAI'18/EAAI'18. AAAI Press." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 254, + 526, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 254, + 526, + 311 + ], + "spans": [ + { + "bbox": [ + 304, + 254, + 526, + 311 + ], + "type": "text", + "content": "Hongjun Wang and Yisen Wang. 2022. Self-ensemble adversarial training for improved robustness. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 318, + 526, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 318, + 526, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 318, + 526, + 418 + ], + "type": "text", + "content": "Shuo Wang, Yang Liu, Chao Wang, Huanbo Luan, and Maosong Sun. 2019. Improving back-translation with uncertainty-based confidence estimation. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLP-IJCNLP 2019, Hong Kong, China, November 3-7, 2019, pages 791-802. Association for Computational Linguistics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 426, + 526, + 526 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 526, + 526 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 526, + 526 + ], + "type": "text", + "content": "Yijun Xiao and William Yang Wang. 2019. Quantifying uncertainties in natural language processing tasks. In The Thirty-Third AAAI Conference on Artificial Intelligence, AAAI 2019, The Thirty-First Innovative Applications of Artificial Intelligence Conference, IAAI 2019, The Ninth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2019, Honolulu, Hawaii, USA, January 27 - February 1, 2019, pages 7322-7329. AAAI Press." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 534, + 526, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 526, + 655 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 526, + 655 + ], + "type": "text", + "content": "Bowen Yu, Zhenyu Zhang, Xiaobo Shu, Tingwen Liu, Yubin Wang, Bin Wang, and Sujian Li. 2020. Joint extraction of entities and relations based on a novel decomposition strategy. In ECAI 2020 - 24th European Conference on Artificial Intelligence, 29 August-8 September 2020, Santiago de Compostela, Spain, August 29 - September 8, 2020 - Including 10th Conference on Prestigious Applications of Artificial Intelligence (PAIS 2020), volume 325 of Frontiers in Artificial Intelligence and Applications, pages 2282-2289. IOS Press." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 664, + 526, + 741 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 664, + 526, + 741 + ], + "spans": [ + { + "bbox": [ + 304, + 664, + 526, + 741 + ], + "type": "text", + "content": "Suncong Zheng, Feng Wang, Hongyun Bao, Yuexing Hao, Peng Zhou, and Bo Xu. 2017. Joint extraction of entities and relations based on a novel tagging scheme. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1227-1236, Vancouver, Canada. Association for Computational Linguistics." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 750, + 526, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 750, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 750, + 526, + 772 + ], + "type": "text", + "content": "Zexuan Zhong and Danqi Chen. 2021. A frustratingly easy approach for entity and relation extraction. In" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1355" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 216 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 79, + 72, + 291, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 72, + 291, + 128 + ], + "spans": [ + { + "bbox": [ + 79, + 72, + 291, + 128 + ], + "type": "text", + "content": "Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 50-61, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 136, + 291, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 136, + 291, + 216 + ], + "spans": [ + { + "bbox": [ + 69, + 136, + 291, + 216 + ], + "type": "text", + "content": "Yikai Zhou, Baosong Yang, Derek F. Wong, Yu Wan, and Lidia S. Chao. 2020. Uncertainty-aware curriculum learning for neural machine translation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6934-6944, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "text", + "content": "1356" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "spans": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "type": "text", + "content": "A For every submission:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 106, + 414, + 242 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "spans": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "type": "text", + "content": "A1. Did you describe the limitations of your work? Section 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 143, + 408, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 143, + 408, + 169 + ], + "spans": [ + { + "bbox": [ + 77, + 143, + 408, + 169 + ], + "type": "text", + "content": "A2. Did you discuss any potential risks of your work? We study open-domain information extraction for researches in this area" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 179, + 414, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 179, + 414, + 205 + ], + "spans": [ + { + "bbox": [ + 77, + 179, + 414, + 205 + ], + "type": "text", + "content": "A3. Do the abstract and introduction summarize the paper's main claims? Abstract and Section 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 77, + 215, + 398, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 215, + 398, + 242 + ], + "spans": [ + { + "bbox": [ + 77, + 215, + 398, + 242 + ], + "type": "text", + "content": "A4. Have you used AI writing assistants when working on this paper? Left blank." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 68, + 253, + 291, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 253, + 291, + 266 + ], + "spans": [ + { + "bbox": [ + 68, + 253, + 291, + 266 + ], + "type": "text", + "content": "B Did you use or create scientific artifacts?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 79, + 270, + 128, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 270, + 128, + 283 + ], + "spans": [ + { + "bbox": [ + 79, + 270, + 128, + 283 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 76, + 292, + 524, + 634 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 292, + 315, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 292, + 315, + 318 + ], + "spans": [ + { + "bbox": [ + 76, + 292, + 315, + 318 + ], + "type": "text", + "content": "B1. Did you cite the creators of artifacts you used? No response." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 76, + 328, + 463, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 328, + 463, + 355 + ], + "spans": [ + { + "bbox": [ + 76, + 328, + 463, + 355 + ], + "type": "text", + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? No response." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 76, + 364, + 524, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 364, + 524, + 431 + ], + "spans": [ + { + "bbox": [ + 76, + 364, + 524, + 431 + ], + "type": "text", + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? No response." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 76, + 441, + 524, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 441, + 524, + 494 + ], + "spans": [ + { + "bbox": [ + 76, + 441, + 524, + 494 + ], + "type": "text", + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? No response." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 76, + 504, + 524, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 504, + 524, + 544 + ], + "spans": [ + { + "bbox": [ + 76, + 504, + 524, + 544 + ], + "type": "text", + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? No response." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 76, + 554, + 524, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 554, + 524, + 634 + ], + "spans": [ + { + "bbox": [ + 76, + 554, + 524, + 634 + ], + "type": "text", + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. No response." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "spans": [ + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "type": "text", + "content": "C Did you run computational experiments?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 79, + 661, + 122, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 661, + 122, + 673 + ], + "spans": [ + { + "bbox": [ + 79, + 661, + 122, + 673 + ], + "type": "text", + "content": "Section 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 76, + 684, + 524, + 724 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 684, + 524, + 724 + ], + "spans": [ + { + "bbox": [ + 76, + 684, + 524, + 724 + ], + "type": "text", + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? No response." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "spans": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "text", + "content": "ACL 2023 Responsible NLP Checklist" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "spans": [ + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "type": "text", + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1357" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 77, + 70, + 523, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 70, + 523, + 97 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 523, + 97 + ], + "type": "text", + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 89, + 99, + 132, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 99, + 132, + 110 + ], + "spans": [ + { + "bbox": [ + 89, + 99, + 132, + 110 + ], + "type": "text", + "content": "Section 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 76, + 121, + 523, + 160 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 121, + 523, + 160 + ], + "spans": [ + { + "bbox": [ + 76, + 121, + 523, + 160 + ], + "type": "text", + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 162, + 148, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 162, + 148, + 174 + ], + "spans": [ + { + "bbox": [ + 89, + 162, + 148, + 174 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 183, + 523, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 183, + 523, + 222 + ], + "spans": [ + { + "bbox": [ + 77, + 183, + 523, + 222 + ], + "type": "text", + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 225, + 132, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 225, + 132, + 236 + ], + "spans": [ + { + "bbox": [ + 89, + 225, + 132, + 236 + ], + "type": "text", + "content": "Section 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 247, + 522, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 247, + 522, + 260 + ], + "spans": [ + { + "bbox": [ + 67, + 247, + 522, + 260 + ], + "type": "text", + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 79, + 264, + 127, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 264, + 127, + 277 + ], + "spans": [ + { + "bbox": [ + 79, + 264, + 127, + 277 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 76, + 286, + 523, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 286, + 523, + 312 + ], + "spans": [ + { + "bbox": [ + 76, + 286, + 523, + 312 + ], + "type": "text", + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 89, + 315, + 148, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 315, + 148, + 327 + ], + "spans": [ + { + "bbox": [ + 89, + 315, + 148, + 327 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 76, + 336, + 523, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 336, + 523, + 375 + ], + "spans": [ + { + "bbox": [ + 76, + 336, + 523, + 375 + ], + "type": "text", + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 89, + 378, + 148, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 378, + 148, + 390 + ], + "spans": [ + { + "bbox": [ + 89, + 378, + 148, + 390 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 76, + 400, + 523, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 400, + 523, + 439 + ], + "spans": [ + { + "bbox": [ + 76, + 400, + 523, + 439 + ], + "type": "text", + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 89, + 441, + 148, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 441, + 148, + 453 + ], + "spans": [ + { + "bbox": [ + 89, + 441, + 148, + 453 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 76, + 462, + 519, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 462, + 519, + 475 + ], + "spans": [ + { + "bbox": [ + 76, + 462, + 519, + 475 + ], + "type": "text", + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 89, + 477, + 148, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 477, + 148, + 489 + ], + "spans": [ + { + "bbox": [ + 89, + 477, + 148, + 489 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 76, + 498, + 523, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 498, + 523, + 523 + ], + "spans": [ + { + "bbox": [ + 76, + 498, + 523, + 523 + ], + "type": "text", + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 89, + 527, + 148, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 527, + 148, + 539 + ], + "spans": [ + { + "bbox": [ + 89, + 527, + 148, + 539 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1358" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Understanding Demonstration-based Learning from a Causal Perspective/58464278-fa19-45b3-a09f-9c2dd2d2dc18_content_list.json b/2023/Understanding Demonstration-based Learning from a Causal Perspective/58464278-fa19-45b3-a09f-9c2dd2d2dc18_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6a4bfae468e6e5b340deb5a209698fc6a2edfb87 --- /dev/null +++ b/2023/Understanding Demonstration-based Learning from a Causal Perspective/58464278-fa19-45b3-a09f-9c2dd2d2dc18_content_list.json @@ -0,0 +1,1273 @@ +[ + { + "type": "text", + "text": "Understanding Demonstration-based Learning from a Causal Perspective", + "text_level": 1, + "bbox": [ + 119, + 90, + 878, + 110 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ruiyi Zhang", + "bbox": [ + 275, + 143, + 389, + 158 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Adobe Research ruizhang@adobe.com", + "bbox": [ + 240, + 160, + 426, + 192 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tong Yu", + "bbox": [ + 626, + 143, + 705, + 158 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Adobe Research tyu@adobe.com", + "bbox": [ + 598, + 160, + 734, + 192 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 252, + 339, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Demonstration-based learning has shown impressive performance in exploiting pretrained language models under few-shot learning settings. It is interesting to see that demonstrations, even those composed of random tokens, can still improve performance. In this paper, we build a Structural Causal Model (SCM) to understand demonstration-based learning from causal perspectives and interpret random demonstrations as interventions on the demonstration variable within the causal model. We investigate the causal effects and find that the concurrence of specific words in the demonstration will induce bias, while randomly sampled tokens in the demonstration do not. Based on this finding, we further propose simple ways to construct random demonstrations, which even outperform hand-crafted, meaningful demonstrations on public sequence labeling benchmarks1.", + "bbox": [ + 141, + 275, + 460, + 558 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 114, + 567, + 258, + 583 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large pretrained language models (PLMs) have recently shown great progress (Devlin et al., 2019; Liu et al., 2019a; Lewis et al., 2020; Xie et al., 2020; Huang et al., 2021). These models, such as GPT-4 (Peng et al., 2023), PALM (Anil et al., 2023), and Llama (Touvron et al., 2023), have shown human-level capability with only a few illustrative examples (Lake et al., 2015). Specifically, demonstration-based learning has been introduced to augment the input with demonstrations, i.e., the input and expected output pairs. Brown et al. (2020) simply picked up to a small number of sampled instances and directly concatenated them with the input to perform in-context learning. Lee et al. (2022) concatenated the input with task demonstrations to create augmented input and fed them into PLMs to obtain improved token representations to do sequence labeling in a classifier-based fine-tuning way.", + "bbox": [ + 112, + 592, + 489, + 898 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, how and why such demonstrations help still remains unclear, and there has been a growing amount of work investigating the mechanisms of demonstration-based learning. Min et al. (2022) investigated in-context learning with demonstrations under zero-shot settings and found that input with random labels can still produce performance comparable to that of correct labels. Zhang et al. (2022a) replaced every token in the demonstration with random ones and still surprisingly observed good few-shot learners even when the demonstration is meaningless. These observations conflict with some existing hypotheses (Gao et al., 2021; Lee et al., 2022) that models are learning meaningful knowledge from demonstrations.", + "bbox": [ + 507, + 253, + 884, + 492 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To better understand demonstration-based learning, we take a deeper dive into the random construction of demonstrations. Specifically, we first build a Structural Causal Model (SCM) to understand demonstration-based learning from a Causal Perspective. A causal view is developed to explore the spurious correlations between demonstrations and few-shot training samples. Based on the intervention on the demonstration variable in the SCM, we design multiple simple and effective ways to construct random demonstrations. These methods are evaluated on structured prediction tasks with carefully designed experiment setups. Empirical results show that carefully designed random demonstrations can outperform meaningful demonstrations under the few-shot learning setting. This finding suggests that meaningless demonstrations can still provide valid information for PLMs. Moreover, random demonstrations allow the learning algorithm to identify important features and patterns in the data more effectively than homogeneous handcrafted demonstrations.", + "bbox": [ + 507, + 495, + 884, + 847 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Background", + "text_level": 1, + "bbox": [ + 507, + 860, + 650, + 877 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this section, we introduce the background of sequence labeling and demonstration-based learning.", + "bbox": [ + 507, + 887, + 884, + 919 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Code available at: github.com/zhangry868/RandDemo", + "bbox": [ + 136, + 903, + 475, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1465", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics", + "bbox": [ + 226, + 945, + 769, + 957 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Volume 2: Short Papers, pages 1465-1475", + "bbox": [ + 368, + 958, + 628, + 971 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "July 9-14, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 295, + 972, + 700, + 985 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/b9a0e5e6260873fdda05b1796d73eb62ddec68965915154d28b93b9b380aeac8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Sentence:The Algerian War of Independence marked the end of French colonial rule in North Africa .
Labels:O B-MISC I-MISC I-MISC O O O O B-ORG O O O B-LOC I-LOC O
Biased: French -> [ORG] Desired: French -> [MISC]
Standard:[SEP] The unnamed suspect left the British colony after being detained and then freed by the Independent Commission Against Corruption ( ICAC ) , the radio said . Independent Commission Against Corruption is ORG . [SEP] [...]
Random:[SEP] Lebanon First Ed ##up CBOE suspect CB Chicago K Chicago Board Options Exchange ##ty Paul Gascoigne CBOE Monday Les into vintage I ##tion Ferdinand ##ca Op [SEP] [...]
", + "bbox": [ + 115, + 82, + 878, + 175 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1: An example from the CoNLL03 dataset with different demonstrations. The NER model takes both the sentence and a demonstration as its inputs. The top two rows show examples of the NER model inputs and outputs with standard demonstrations. A biased prediction for 'French' is caused by the demonstration bias. The bottom three lines show three different demonstrations: Standard and Random demonstrations. The notation '[SEP][...]' indicates that there are demonstrations for other classes, which have been omitted due to limited space.", + "bbox": [ + 110, + 181, + 884, + 254 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Sequence Labeling Given an input sentence $\\mathbf{x} = [x_{1}, x_{2}, \\dots, x_{n}]$ composed of $n$ tokens, the sequence labeling task is to predict a tag $y_{i} \\in Y \\cup \\{O\\}$ for each token $x_{i}$ , where $Y$ is a predefined set of tags, and $O$ denotes outside a tagged span. In the few-shot setting, we only have $K$ -shot support set $S$ for training which contains $K$ examples for each tag type. This setting usually refers to $K$ -shot learning. Modern sequence labeling models are usually composed of an encoder and a classification head. The encoders are PLMs such as BERT (Devlin et al., 2018) and RoBERTa (Liu et al., 2019b), which provides contextualized representations for each token $\\mathbf{h} = [h_{1}, h_{2}, \\dots, h_{n}]$ given the natural language sequence $\\mathbf{x} = [x_{1}, x_{2}, \\dots, x_{n}]$ . The classification head takes these contextualized representations and predicts the label $l_{i}$ for each token $x_{i}$ . The model is optimized with the standard cross-entropy loss.", + "bbox": [ + 112, + 262, + 489, + 568 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Demonstration-based Learning Given some demonstration $\\tilde{\\mathbf{x}}$ , we concatenate the original input $\\mathbf{x}$ with its demonstration $\\tilde{\\mathbf{x}}$ as $[\\mathbf{x};\\tilde{\\mathbf{x}}]$ . We then feed the demonstration-augmented input $[\\mathbf{x};\\tilde{\\mathbf{x}}]$ into the encoder, and get the contextualized representation $[\\mathbf{h};\\tilde{\\mathbf{h}}]$ . The classification head takes $\\mathbf{h}$ as the input and estimate the corresponding token's label $l_{i}$ in the original natural-language sequence. Please note that we use identical demonstrations during training and testing (Lee et al., 2022).", + "bbox": [ + 112, + 574, + 489, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Demonstration Construction To construct demonstrations, we first sample an entity $e^{(c)}$ for each label type $t^{(c)}$ , and its context $s^{(c)}$ from support set $S$ . Then we convert them into a natural language sequence $d^{(c)} = T(s^{(c)}, e^{(c)}, t^{(c)})$ , where $T$ is the template operator and previous works (Lee et al., 2022) focus on finding more effective templates. With these sequences $[d^{(c_i)}]_{i=1}^{|Y|}$ with different tags $c_i$ , a demonstration $\\tilde{\\mathbf{x}}$ is built by concatenating them together: $\\tilde{\\mathbf{x}} = d^{(c_1)} \\oplus d^{(c_2)} \\oplus \\dots \\oplus d^{(c_{|Y|})}$ , where $\\oplus$ is the concatenation operator. An effective tem", + "bbox": [ + 112, + 740, + 489, + 917 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "plate, such as the one used in Lee et al. (2022), is \" $s^{(c)}$ . $e^{(c)}$ is $t^{(c)}$ .\" Here, we refer the \" $e^{(c)}$ is $t^{(c)}$ \" part in the template as labeling part of the demonstration.", + "bbox": [ + 507, + 262, + 884, + 325 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3 Demonstration-based Learning from a Causal Perspective", + "text_level": 1, + "bbox": [ + 507, + 340, + 875, + 373 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we give a specific example to show the potential bias and understand demonstration-based learning from a causal perspective. Specifically, we first introduce a Structural Causal Model (SCM) (Pearl et al., 2000) to describe the mechanism and identify the induced bias. Then, we perform demonstration variable intervention and propose multiple simple and effective random demonstration templates inspired by our causal model.", + "bbox": [ + 507, + 384, + 884, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We observe that the frequent co-occurrence of tokens in the classical demonstrations generate harmful superficial patterns which is misleading to the model and leads to biased predictions (Zhang et al., 2022a; Min et al., 2022). A specific example with different demonstrations is provided in Table 1, where the entity to predict is French. Following previous work (Zhang et al., 2022a), the observed demonstrations (i.e., standard demonstration) provides some biased information: the concurrency of British and ICAC, which is an organization (ORG), may lead to biased predictions: French is labeled as an Organization while its desired prediction is other classes (MISC). Intuitively, the co-occurrence of two specific words in the demonstration may induce bias, while randomly sampled tokens in the demonstration do not. This specific example suggests why random demonstrations may sometimes perform better than standard ones.", + "bbox": [ + 507, + 530, + 884, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1 Causal Model", + "text_level": 1, + "bbox": [ + 507, + 848, + 665, + 862 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To study the causal relationship between the NER model and its training data, and explain the role of the demonstration, we introduce a SCM to describe", + "bbox": [ + 507, + 871, + 882, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1466", + "bbox": [ + 482, + 928, + 521, + 940 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c02e510f97f02f010e978031e4d8f0b374c424d313250b464c45bd264c6682c3.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 115, + 85, + 292, + 179 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/987321a8336f314009714664bfe25b3a1faaccc13230764ececde8df55dbed24.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 300, + 86, + 475, + 178 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b90ea756a254755f329ba8c508bd923ee037cbaa91498920bb1a1dd606c0ce96.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 482, + 86, + 660, + 178 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d43a2f5e031397c86a55695cbf9634643370921a08ffb03291bd3f22ee9f8765.jpg", + "image_caption": [ + "(d)", + "Figure 1: Causal views of NER. (a) shows a traditional NER model (Zeng et al., 2020), (b) shows the demonstration-based NER model under the causal view. With demonstration $D$ , the backdoor path $G \\rightarrow D \\rightarrow X$ exists, which further introduces the bias. (c) shows the demonstration-based NER model with debiasing techniques, and the red cross means intervention. (d) is model architecture overview between classical and demonstration-based learning." + ], + "image_footnote": [], + "bbox": [ + 670, + 82, + 880, + 190 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the inference step in NER models. Figure 1 shows the SCM of NER models. There are mainly 6 variables in NER models: 1) Demonstration Tokens $D$ , the tokens which form the demonstration; 2) Context Tokens $C$ , the tokens that are related to the context; 3) Entity Tokens $E$ , the tokens which are entities; 4) Input Example $X$ , which is composed of $C$ and $E$ in the traditional model and composed of $C$ , $E$ and $D$ in the demonstration-based models; 5) Unobserved confounders $G$ , a confounding variable (not a concrete token) that influences the generation of $C$ , $E$ and $D$ ; 6) Evaluation result $Y$ , the evaluation result (the F1 score) of the NER models. Under the causal view, the key difference between the traditional NER model and the demonstration-based NER model is that, the demonstration-based NER model has an additional node $D$ . With the introduction of the demonstration $D$ , a backdoor path $G \\rightarrow D \\rightarrow X$ exists, which further introduces the bias.", + "bbox": [ + 112, + 285, + 489, + 605 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by our SCM model (Figure 1b), we develop sampling techniques to generate new counterfactual examples by the interventions on the existing observational examples to alleviate this bias. The benefits of interventions on $E$ and $C$ have been studied in (Zeng et al., 2020). In this paper, we focus on understanding the role of demonstrations in NER models under the causal view. We understand the co-occurrence of tokens and harmful superficial patterns from the causal perspective and focus on using interventions on the demonstration variable to create new counterfactual demonstrations.", + "bbox": [ + 112, + 608, + 489, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2 Controllable Random Demonstrations", + "text_level": 1, + "bbox": [ + 112, + 816, + 460, + 829 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first provide a running example to better understand the induced bias from human-crafted demonstrations and then present different ways of intervention on the demonstration tokens. The intervention is implemented via control", + "bbox": [ + 112, + 839, + 490, + 917 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "lable random demonstrations to create new counterfactual examples, as replacing standard demonstrations with random tokens can remove induce bias and still make the model a good few-shot learner (Zhang et al., 2022a).", + "bbox": [ + 507, + 285, + 884, + 365 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In Lee et al. (2022), an effective template $T$ is \" $s^{(c)}$ . $e^{(c)}$ is $t^{(c)}$ , and an example demonstration $d^{(c)}$ can be \"[SEP] Obama returns to White House. Obama is PER.\" Intuitively, the model understands the demonstrations and then better performs inference. However, random demonstrations can still bring performance improvement (Zhang et al., 2022a). The random template is as simple as $[s_i]_{i=1}^L$ , where $s_i \\in p$ , and $p$ is a token distribution. Random demonstrations are composed of $L$ tokens randomly sampled from $p$ .", + "bbox": [ + 507, + 367, + 884, + 545 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Demonstration Intervention We use the intervention on the demonstration tokens to create new counterfactual examples, to alleviate the biases. If we do not carefully design D, the backdoor path will exist and the model performance is degraded. Our causal framework enables us to think about the problem from a causal perspective and guides us how to properly design D. We denote uniform distribution composed of vocabulary words of the PLMs as $p_{\\mathcal{V}}$ . Given the token distribution $p_{\\mathcal{V}}$ , for any word $w_i \\in p_{\\mathcal{V}}$ , we have $p_{\\mathcal{V}}(w_i) = \\frac{1}{|\\mathcal{V}|}$ . Then we have a plain way to construct random demonstrations.", + "bbox": [ + 507, + 546, + 882, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An important observation is that not all counterfactual examples are correct or useful. Hence, the intervention can be better implemented by replacing the uniform distribution with a non-uniform distribution, i.e., by adding or removing words and changing specific words' probabilities. Some mechanism is needed to identify good counterfactual demonstrations, to avoid introducing noise. An intuitive solution is that we consider tokens from the support set are more helpful as PLMs are fine", + "bbox": [ + 507, + 758, + 884, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1467", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/3e10791f930af0474d8a42b65d96ba922faabadbafba002a9eacef57294219d7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModeNERChunking
CoNLL03OntoNotes 5.0CoNLL00
F1PrecisionRecallF1PrecisionRecallF1PrecisionRecall
No Demo.28.71±10.3139.96±11.2522.68±9.0937.37±7.5833.80±6.7941.92±8.8563.17±4.2259.28±5.0567.72±3.51
Standard45.86±6.0847.38±5.9344.75±7.0740.21±7.6532.51±6.8752.82±8.2870.55±3.0866.53±4.4075.21±2.11
Random41.33±7.3645.41±7.3738.22±7.6539.71±7.5632.28±6.5651.63±8.7569.28±2.7864.75±3.8574.57±1.66
Rand-S45.55±8.0246.84±7.7144.60±8.6241.60±7.0533.96±6.2953.75±7.8070.63±3.0166.24±4.2975.75±1.70
Rand-W45.93±7.5747.79±7.4244.50±8.1345.49±3.7737.82±3.6457.18±4.1772.15±3.1668.00±4.4276.94±1.67
Rand-E47.32±7.4248.96±7.0246.02±8.1146.06±3.8438.32±3.6557.81±4.3174.02±2.9370.37±4.2378.18±1.75
", + "bbox": [ + 119, + 80, + 873, + 225 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 2: Main results for traditional token classification method (No Demo.) and demonstration-based learning with different modes of demonstrations under 5-shot scenario.", + "bbox": [ + 112, + 231, + 882, + 260 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tuned on the support set. We expect to see a better downstream predictor when the demonstrations are constructed randomly from a intervened token distribution.", + "bbox": [ + 112, + 269, + 487, + 332 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The difference between random demonstrations lies in the vocabulary and its associated probability distributions. We perform the interventions by controlling the vocabulary and changing the probability of random tokens. We encourage entity words (e.g., ICAC, British) to appear more frequently compared to the others (e.g., is). Based on the previous theoretical justification, we consider the following variants of constructing random demonstrations2 construction methods as counterfactual alternatives of the standard demonstrations3:", + "bbox": [ + 112, + 334, + 487, + 508 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Random: random context with tokens uniformly sampled from PLMs vocabulary $\\nu$ .", + "- Rand-S: random context with tokens uniformly sampled from unique words (i.e., vocabulary) of support set, denoted as $S$ .", + "- Rand-W4: random context with tokens sampled from $S$ , and entity tokens in support set, denoted as $W$ ; tokens from $W$ have four times higher probability compared with those from $S$ .", + "- Rand-E: similar to Rand-W, but replace entity tokens with entities composed of coherent tokens in support set, denoted as $\\mathcal{U}$ ." + ], + "bbox": [ + 112, + 511, + 485, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4 Experimental Results", + "text_level": 1, + "bbox": [ + 112, + 715, + 334, + 732 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1 Experiment Setup", + "text_level": 1, + "bbox": [ + 112, + 741, + 302, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Datasets We conduct experiments on two sequence labeling tasks: (i) named entity recognition (NER) on dataset CoNLL03 (Tjong Kim Sang and De Meulder, 2003), and OntoNotes 5.0 (Weischedel et al., 2013); and (ii) chunking on dataset CoNLL00 (Tjong Kim Sang and Buchholz,", + "bbox": [ + 112, + 762, + 487, + 859 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2000). Following previous works Ma et al. (2021); Zhang et al. (2022a), we omit the 7 value types in OntoNotes and only consider the 6 most frequent types in CoNLL00. For few-shot data sampling, we follow the greedy sampling strategy proposed by Yang and Katiyar (2020) to sample $K$ shots for each type in an increasing order with respect to their frequencies, the detailed algorithm can be found. For each dataset, we sample 5 different $K$ -shot support sets and report mean and standard deviation of metrics. For each $K$ -shot support set, we run the experiments with 3 random seeds.", + "bbox": [ + 507, + 269, + 884, + 462 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Main Results We show the results for demonstration-based learning with different modes of demonstrations as well as classical sequence labeling with no demonstration in Table 2. The results show that demonstration-based method can consistently improve model performance. In demonstration-based methods, the Random approach shows the worst performance and Rand-S shows comparable results with the standard demonstrations, and the conclusion is consistent with previous works (Zhang et al., 2022a). Interestingly, if we modify the token sampling distributions and sample more entity or entity-related words as Rand-W and Rand-E, our model shows even better performance than standard meaningful demonstrations. The difference between Rand-W and Rand-E lies in whether there are complete entities, and the results show that adding complete entities instead of random entity words can lead to better performance. At the same time, it shows adding random tokens related to the support set can reduce the fine-tuned bias, which verifies our hypothesis in Section 3.1. Intuitively, the benefits of demonstration-based methods come from tokens of support sets $S$ instead of meaningful demonstrations, as the standard demonstration sampled from the support set also shows good performance.", + "bbox": [ + 507, + 468, + 882, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "2Random: [SEP] {random context}", + "bbox": [ + 134, + 866, + 373, + 879 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "3Standard: [SEP] {context} {entity} is {tag}.", + "bbox": [ + 134, + 879, + 460, + 892 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "4 Empirical results show sampling only from $\\mathcal{W}$ leads to poor performance.", + "bbox": [ + 115, + 892, + 485, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1468", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c081e6404c2d14411e0af537694d5665345a88764e5bbd0c6639d760b30d3098.jpg", + "image_caption": [ + "Figure 2: Results with different support set size on CoNLL03, NRB and WTS datasets." + ], + "image_footnote": [], + "bbox": [ + 117, + 80, + 378, + 219 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/47e737b77f0e5157053023e79ebe490557513eb838b5aff71480e50c7ecc9650.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 386, + 80, + 640, + 221 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b5bc4c4268bb8d2f235c89ba405172e60206d99cb1707f253ba822159311a774.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 80, + 880, + 221 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d3e28a47a040cbb0835e15b9fce8e7ef3f14b440fc930ded59b0ce350481d560.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModeCoNLL03OntoNotes5.0CoNLL00
No Demo.45.70±8.1351.62±2.7672.80±3.53
Standard45.73±7.2954.76±2.3675.90±1.95
Rand-S46.86±6.5054.35±2.6772.23±3.42
Rand-W52.11±6.1554.48±2.3573.84±2.19
Rand-E52.87±7.6455.94±2.3875.30±3.06
", + "bbox": [ + 122, + 260, + 473, + 361 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 3: Main results (F1 scores) of RoBERTa-Large for traditional token classification with different modes of demonstrations under 5-shot scenario.", + "bbox": [ + 112, + 367, + 487, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2 Analysis", + "text_level": 1, + "bbox": [ + 112, + 425, + 226, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Ablation Studies We further investigate whether the performance gain of demonstration-based learning changes over the size of support set. We present results of different modes of demonstrations under $K = 5, 10, 20$ shots in Figure 2. With more training examples in the support set, the relative performance gap between Rand-E and Standard remains, but it becomes smaller. This indicates that carefully designed random demonstrations show a consistent performance improvement upon standard demonstration. We also observe that the variance within each group becomes smaller as more data becomes available. Among random demonstrations, Rand-E consistently shows better performance than Rand-W and Rand-S, which verifies our hypothesis based on the SCM.", + "bbox": [ + 112, + 449, + 489, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Additionally, we investigate the effect of using different base models and replace BERT with RoBERTa. The observed results for RoBERTa in Table 3 are consistent with those of BERT, demonstrating that Rand-E exhibits superior performance across different model architectures.", + "bbox": [ + 112, + 708, + 489, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Name Regularity Bias Name Regularity Bias (Ghaddar et al., 2021; Lin et al., 2020) in NER occurs when a model relies on a signal from the entity name to make predictions and disregards evidence from the local context. Ghaddar et al. (2021) carefully designed a testbed utilizing Wikipedia disambiguation pages to diagnose the Name Regu", + "bbox": [ + 112, + 806, + 489, + 917 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "larity Bias of NER models. Details about the NRB dataset are provided in the appendix.", + "bbox": [ + 507, + 256, + 880, + 288 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use both the NRB and WTS (as control sets) datasets to evaluate the model trained with different modes of demonstrations on CoNLL03. The results show a smaller gap for random demonstrations, suggesting that random demonstration-based learning can better leverage context information instead of the name regularity patterns.", + "bbox": [ + 507, + 290, + 882, + 403 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5 Conclusions", + "text_level": 1, + "bbox": [ + 509, + 416, + 648, + 432 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this paper, we present a casual view to understand demonstration-based learning. Based on the structural causal model we constructed, we investigate the causal effects and discover that the concurrence of specific words in the demonstration can induce bias. To address this issue, we perform interventions by constructing random demonstrations. Our empirical results indicate that carefully designed random demonstrations consistently outperform meaningful demonstrations on public sequence labeling benchmarks.", + "bbox": [ + 507, + 444, + 884, + 620 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6 Limitations", + "text_level": 1, + "bbox": [ + 507, + 634, + 645, + 649 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "All our experiments are done on the sequence labeling task, and they can be further evaluated on sentence classification tasks with classifier-based fine-tuning since the [CLS] token used for classification represents the whole sentence. We provide a causal opinion on demonstration-based learning and a simple but not systematic method to alleviate the induced bias. Our demonstration-based learning builds upon previous works (Lee et al., 2022; Zhang et al., 2022a), where BERT or RoBERTa are used instead of Large Language Models, such as InstructGPT (Ouyang et al., 2022), PaLM (Chowdhery et al., 2022), and OPT (Zhang et al., 2022b). Furthermore, our conclusions are drawn from few-shot learning settings and cannot be directly applied to zero-shot inference.", + "bbox": [ + 507, + 661, + 884, + 917 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1469", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 115, + 84, + 213, + 98 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Rohan Anil, Andrew M Dai, Orhan Firat, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, et al. 2023. Palm 2 technical report. arXiv preprint arXiv:2305.10403.", + "Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual.", + "Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. 2022. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311.", + "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805.", + "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics.", + "Tianyu Gao, Adam Fisch, and Danqi Chen. 2021. Making pre-trained language models better few-shot learners. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 3816-3830, Online. Association for Computational Linguistics.", + "Abbas Ghaddar, Philippe Langlais, Ahmad Rashid, and Mehdi Rezagholizadeh. 2021. Context-aware Adversarial Training for Name Regularity Bias in Named Entity Recognition. Transactions of the Association for Computational Linguistics, 9:586-604.", + "Jiaxin Huang, Chunyuan Li, Krishan Subudhi, Damien Jose, Shobana Balakrishnan, Weizhu Chen, Baolin Peng, Jianfeng Gao, and Jiawei Han. 2021. Few-shot named entity recognition: An empirical baseline" + ], + "bbox": [ + 115, + 107, + 485, + 917 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "study. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 10408-10423, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.", + "Brenden Lake, Ruslan Salakhutdinov, and Joshua Tenenbaum. 2015. Human-level concept learning through probabilistic program induction. Science, 350:1332-1338.", + "Dong-Ho Lee, Akshen Kadakia, Kangmin Tan, Mahak Agarwal, Xinyu Feng, Takashi Shibuya, Ryosuke Mitani, Toshiyuki Sekiya, Jay Pujara, and Xiang Ren. 2022. Good examples make a faster learner: Simple demonstration-based learning for low-resource NER. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2687-2700, Dublin, Ireland. Association for Computational Linguistics.", + "Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871-7880, Online. Association for Computational Linguistics.", + "Hongyu Lin, Yaojie Lu, Jialong Tang, Xianpei Han, Le Sun, Zhicheng Wei, and Nicholas Jing Yuan. 2020. A rigorous study on named entity recognition: Can fine-tuning pretrained model lead to the promised land? In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7291-7300, Online. Association for Computational Linguistics.", + "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019a. Roberta: A robustly optimized BERT pretraining approach. CoRR, abs/1907.11692.", + "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019b. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692.", + "Ruotian Ma, Xin Zhou, Tao Gui, Yiding Tan, Qi Zhang, and Xuanjing Huang. 2021. Template-free prompt tuning for few-shot NER. CoRR, abs/2109.13532.", + "Sewon Min, Xinxi Lyu, Ari Holtzman, Mikel Artetxe, Mike Lewis, Hannaneh Hajishirzi, and Luke Zettlemoyer. 2022. Rethinking the role of demonstrations: What makes in-context learning work? arXiv preprint arXiv:2202.12837.", + "Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al." + ], + "bbox": [ + 510, + 85, + 880, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1470", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "2022. Training language models to follow instructions with human feedback. NeurIPS.", + "Judea Pearl et al. 2000. Models, reasoning and inference. Cambridge, UK: Cambridge University Press, 19(2).", + "Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. 2023. Instruction tuning with gpt-4. arXiv preprint arXiv:2304.03277.", + "Erik F. Tjong Kim Sang and Sabine Buchholz. 2000. Introduction to the CoNLL-2000 shared task chunking. In Fourth Conference on Computational Natural Language Learning and the Second Learning Language in Logic Workshop.", + "Erik F. Tjong Kim Sang and Fien De Meulder. 2003. Introduction to the CoNLL-2003 shared task: Language-independent named entity recognition. In Proceedings of the Seventh Conference on Natural Language Learning at HLT-NAACL 2003, pages 142-147.", + "Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971.", + "Ralph Weischedel, Martha Palmer, Mitchell Marcus, Edward Hovy, Sameer Pradhan, Lance Ramshaw, Nianwen Xue, Ann Taylor, Jeff Kaufman, Michelle Franchini, et al. 2013. Ontonotes release 5.0 ldc2013t19. Linguistic Data Consortium, Philadelphia, PA, 23.", + "Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. 2020. Unsupervised data augmentation for consistency training. In Advances in Neural Information Processing Systems, volume 33, pages 6256-6268. Curran Associates, Inc.", + "Yi Yang and Arzoo Katiyar. 2020. Simple and effective few-shot named entity recognition with structured nearest neighbor learning. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6365-6375, Online. Association for Computational Linguistics.", + "Xiangji Zeng, Yunliang Li, Yuchen Zhai, and Yin Zhang. 2020. Counterfactual generator: A weakly-supervised method for named entity recognition. In EMNLP.", + "Hongxin Zhang, Yanzhe Zhang, Ruiyi Zhang, and Diyi Yang. 2022a. Robustness of demonstration-based learning under limited data scenario. In EMNLP.", + "Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. 2022b. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068." + ], + "bbox": [ + 115, + 85, + 489, + 892 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1471", + "bbox": [ + 482, + 928, + 517, + 940 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/40efaad6896ce44d18c6741c18b875940e58d4da5d602aa409436da4c5c42ff7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Dataset|Y|L|Dtest|
CoNLL034183453
OntoNotes 5.0112112217
CoNLL006362012
", + "bbox": [ + 168, + 82, + 433, + 162 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4: Data Statistics. $|Y|$ : # of entity types. L: average # of tokens in input sentence. $|D_{support}|$ : average # of sentences in 5-shot support set over 5 different sub-samples. $|D_{test}|$ : # of sentences in test set.", + "bbox": [ + 112, + 173, + 489, + 231 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A Appendix", + "text_level": 1, + "bbox": [ + 114, + 255, + 238, + 272 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "NRB Dataset Details The NRB dataset contains examples whose labels can be easily inferred from the local context, but they are difficult to be tagged by a popular NER system. The WTS dataset is a domain control set that includes the same query terms covered by NRB, but these can be correctly labeled by both the popular NER tagger and the local context-only tagger. Therefore, the gap between the NRB and WTS sets measures how effectively the model captures context information to predict token labels.", + "bbox": [ + 112, + 280, + 487, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effects of Sampling Probability We present two variants, Random-E[X] and Random-W[X], where X refers to how many times the probability of preferred tokens is higher. In this ablation study, we consistently observe that Random-E4 performs better than Random-E2, and Random-W4 outperforms Random-E4. However, if we increase the X value to a very large number, the performance deteriorates.", + "bbox": [ + 112, + 458, + 489, + 602 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1472", + "bbox": [ + 482, + 928, + 521, + 940 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4d287e32f49c4496c1ede1687bce3ce1b2488a7c6804de1f0e6fee6e7e19d13b.jpg", + "image_caption": [ + "Figure 3: F1, Precision and Recall with more variants of Random Demonstrations on CoNLL03." + ], + "image_footnote": [], + "bbox": [ + 115, + 405, + 366, + 579 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/947aa770ce7060abb984559ce3a3fbf4b3483e8097b906e888edfd68305777ea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 376, + 405, + 623, + 580 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/c5d6fc26cb4ac6b2d6ad7911b284eefcd60b7610b7d3d6a9ff1046eaa2fff073.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 407, + 890, + 580 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1473", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "A For every submission:", + "bbox": [ + 115, + 107, + 322, + 122 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "A1. Did you describe the limitations of your work? Section 6", + "A2. Did you discuss any potential risks of your work? Section 6", + "A3. Do the abstract and introduction summarize the paper's main claims? Left blank.", + "A4. Have you used AI writing assistants when working on this paper? Left blank." + ], + "bbox": [ + 129, + 126, + 695, + 288 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "B Did you use or create scientific artifacts?", + "bbox": [ + 114, + 300, + 489, + 316 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 132, + 321, + 215, + 336 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "B1. Did you cite the creators of artifacts you used? Not applicable. Left blank.", + "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Not applicable. Left blank.", + "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Not applicable. Left blank.", + "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Not applicable. Left blank.", + "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Not applicable. Left blank.", + "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Not applicable. Left blank." + ], + "bbox": [ + 127, + 347, + 880, + 753 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "C Did you run computational experiments?", + "bbox": [ + 114, + 764, + 492, + 781 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Section 4", + "bbox": [ + 132, + 785, + 206, + 800 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Section 4", + "bbox": [ + 129, + 812, + 880, + 859 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.", + "bbox": [ + 112, + 866, + 877, + 889 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "ACL 2023 Responsible NLP Checklist", + "bbox": [ + 132, + 84, + 433, + 99 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1474", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Section 4", + "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Section 4", + "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Section 4" + ], + "bbox": [ + 129, + 83, + 880, + 280 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank.", + "bbox": [ + 112, + 293, + 877, + 330 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Not applicable. Left blank.", + "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Not applicable. Left blank.", + "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Not applicable. Left blank.", + "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Not applicable. Left blank.", + "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Not applicable. Left blank." + ], + "bbox": [ + 127, + 340, + 880, + 640 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1475", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Understanding Demonstration-based Learning from a Causal Perspective/58464278-fa19-45b3-a09f-9c2dd2d2dc18_model.json b/2023/Understanding Demonstration-based Learning from a Causal Perspective/58464278-fa19-45b3-a09f-9c2dd2d2dc18_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7d598996f97d59d8d2e31b782d00e258d5d58f7b --- /dev/null +++ b/2023/Understanding Demonstration-based Learning from a Causal Perspective/58464278-fa19-45b3-a09f-9c2dd2d2dc18_model.json @@ -0,0 +1,1828 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.12, + 0.091, + 0.88, + 0.111 + ], + "angle": 0, + "content": "Understanding Demonstration-based Learning from a Causal Perspective" + }, + { + "type": "text", + "bbox": [ + 0.277, + 0.144, + 0.391, + 0.159 + ], + "angle": 0, + "content": "Ruiyi Zhang" + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.161, + 0.428, + 0.193 + ], + "angle": 0, + "content": "Adobe Research ruizhang@adobe.com" + }, + { + "type": "text", + "bbox": [ + 0.628, + 0.144, + 0.706, + 0.159 + ], + "angle": 0, + "content": "Tong Yu" + }, + { + "type": "text", + "bbox": [ + 0.599, + 0.161, + 0.735, + 0.193 + ], + "angle": 0, + "content": "Adobe Research tyu@adobe.com" + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.253, + 0.341, + 0.268 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.276, + 0.461, + 0.56 + ], + "angle": 0, + "content": "Demonstration-based learning has shown impressive performance in exploiting pretrained language models under few-shot learning settings. It is interesting to see that demonstrations, even those composed of random tokens, can still improve performance. In this paper, we build a Structural Causal Model (SCM) to understand demonstration-based learning from causal perspectives and interpret random demonstrations as interventions on the demonstration variable within the causal model. We investigate the causal effects and find that the concurrence of specific words in the demonstration will induce bias, while randomly sampled tokens in the demonstration do not. Based on this finding, we further propose simple ways to construct random demonstrations, which even outperform hand-crafted, meaningful demonstrations on public sequence labeling benchmarks1." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.568, + 0.26, + 0.584 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.593, + 0.49, + 0.899 + ], + "angle": 0, + "content": "Large pretrained language models (PLMs) have recently shown great progress (Devlin et al., 2019; Liu et al., 2019a; Lewis et al., 2020; Xie et al., 2020; Huang et al., 2021). These models, such as GPT-4 (Peng et al., 2023), PALM (Anil et al., 2023), and Llama (Touvron et al., 2023), have shown human-level capability with only a few illustrative examples (Lake et al., 2015). Specifically, demonstration-based learning has been introduced to augment the input with demonstrations, i.e., the input and expected output pairs. Brown et al. (2020) simply picked up to a small number of sampled instances and directly concatenated them with the input to perform in-context learning. Lee et al. (2022) concatenated the input with task demonstrations to create augmented input and fed them into PLMs to obtain improved token representations to do sequence labeling in a classifier-based fine-tuning way." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.254, + 0.885, + 0.493 + ], + "angle": 0, + "content": "However, how and why such demonstrations help still remains unclear, and there has been a growing amount of work investigating the mechanisms of demonstration-based learning. Min et al. (2022) investigated in-context learning with demonstrations under zero-shot settings and found that input with random labels can still produce performance comparable to that of correct labels. Zhang et al. (2022a) replaced every token in the demonstration with random ones and still surprisingly observed good few-shot learners even when the demonstration is meaningless. These observations conflict with some existing hypotheses (Gao et al., 2021; Lee et al., 2022) that models are learning meaningful knowledge from demonstrations." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.496, + 0.885, + 0.848 + ], + "angle": 0, + "content": "To better understand demonstration-based learning, we take a deeper dive into the random construction of demonstrations. Specifically, we first build a Structural Causal Model (SCM) to understand demonstration-based learning from a Causal Perspective. A causal view is developed to explore the spurious correlations between demonstrations and few-shot training samples. Based on the intervention on the demonstration variable in the SCM, we design multiple simple and effective ways to construct random demonstrations. These methods are evaluated on structured prediction tasks with carefully designed experiment setups. Empirical results show that carefully designed random demonstrations can outperform meaningful demonstrations under the few-shot learning setting. This finding suggests that meaningless demonstrations can still provide valid information for PLMs. Moreover, random demonstrations allow the learning algorithm to identify important features and patterns in the data more effectively than homogeneous handcrafted demonstrations." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.861, + 0.652, + 0.878 + ], + "angle": 0, + "content": "2 Background" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.888, + 0.885, + 0.92 + ], + "angle": 0, + "content": "In this section, we introduce the background of sequence labeling and demonstration-based learning." + }, + { + "type": "page_footnote", + "bbox": [ + 0.137, + 0.904, + 0.476, + 0.919 + ], + "angle": 0, + "content": "\\(^{1}\\)Code available at: github.com/zhangry868/RandDemo" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1465" + }, + { + "type": "footer", + "bbox": [ + 0.227, + 0.946, + 0.771, + 0.958 + ], + "angle": 0, + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + }, + { + "type": "footer", + "bbox": [ + 0.369, + 0.959, + 0.63, + 0.972 + ], + "angle": 0, + "content": "Volume 2: Short Papers, pages 1465-1475" + }, + { + "type": "footer", + "bbox": [ + 0.297, + 0.973, + 0.702, + 0.986 + ], + "angle": 0, + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.117, + 0.083, + 0.88, + 0.177 + ], + "angle": 0, + "content": "
Sentence:The Algerian War of Independence marked the end of French colonial rule in North Africa .
Labels:O B-MISC I-MISC I-MISC O O O O B-ORG O O O B-LOC I-LOC O
Biased: French -> [ORG] Desired: French -> [MISC]
Standard:[SEP] The unnamed suspect left the British colony after being detained and then freed by the Independent Commission Against Corruption ( ICAC ) , the radio said . Independent Commission Against Corruption is ORG . [SEP] [...]
Random:[SEP] Lebanon First Ed ##up CBOE suspect CB Chicago K Chicago Board Options Exchange ##ty Paul Gascoigne CBOE Monday Les into vintage I ##tion Ferdinand ##ca Op [SEP] [...]
" + }, + { + "type": "table_caption", + "bbox": [ + 0.112, + 0.183, + 0.885, + 0.255 + ], + "angle": 0, + "content": "Table 1: An example from the CoNLL03 dataset with different demonstrations. The NER model takes both the sentence and a demonstration as its inputs. The top two rows show examples of the NER model inputs and outputs with standard demonstrations. A biased prediction for 'French' is caused by the demonstration bias. The bottom three lines show three different demonstrations: Standard and Random demonstrations. The notation '[SEP][...]' indicates that there are demonstrations for other classes, which have been omitted due to limited space." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.263, + 0.49, + 0.569 + ], + "angle": 0, + "content": "Sequence Labeling Given an input sentence \\(\\mathbf{x} = [x_{1}, x_{2}, \\dots, x_{n}]\\) composed of \\(n\\) tokens, the sequence labeling task is to predict a tag \\(y_{i} \\in Y \\cup \\{O\\}\\) for each token \\(x_{i}\\), where \\(Y\\) is a predefined set of tags, and \\(O\\) denotes outside a tagged span. In the few-shot setting, we only have \\(K\\)-shot support set \\(S\\) for training which contains \\(K\\) examples for each tag type. This setting usually refers to \\(K\\)-shot learning. Modern sequence labeling models are usually composed of an encoder and a classification head. The encoders are PLMs such as BERT (Devlin et al., 2018) and RoBERTa (Liu et al., 2019b), which provides contextualized representations for each token \\(\\mathbf{h} = [h_{1}, h_{2}, \\dots, h_{n}]\\) given the natural language sequence \\(\\mathbf{x} = [x_{1}, x_{2}, \\dots, x_{n}]\\). The classification head takes these contextualized representations and predicts the label \\(l_{i}\\) for each token \\(x_{i}\\). The model is optimized with the standard cross-entropy loss." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.575, + 0.49, + 0.736 + ], + "angle": 0, + "content": "Demonstration-based Learning Given some demonstration \\(\\tilde{\\mathbf{x}}\\), we concatenate the original input \\(\\mathbf{x}\\) with its demonstration \\(\\tilde{\\mathbf{x}}\\) as \\([\\mathbf{x};\\tilde{\\mathbf{x}}]\\). We then feed the demonstration-augmented input \\([\\mathbf{x};\\tilde{\\mathbf{x}}]\\) into the encoder, and get the contextualized representation \\([\\mathbf{h};\\tilde{\\mathbf{h}}]\\). The classification head takes \\(\\mathbf{h}\\) as the input and estimate the corresponding token's label \\(l_{i}\\) in the original natural-language sequence. Please note that we use identical demonstrations during training and testing (Lee et al., 2022)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.741, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Demonstration Construction To construct demonstrations, we first sample an entity \\( e^{(c)} \\) for each label type \\( t^{(c)} \\), and its context \\( s^{(c)} \\) from support set \\( S \\). Then we convert them into a natural language sequence \\( d^{(c)} = T(s^{(c)}, e^{(c)}, t^{(c)}) \\), where \\( T \\) is the template operator and previous works (Lee et al., 2022) focus on finding more effective templates. With these sequences \\( [d^{(c_i)}]_{i=1}^{|Y|} \\) with different tags \\( c_i \\), a demonstration \\( \\tilde{\\mathbf{x}} \\) is built by concatenating them together: \\( \\tilde{\\mathbf{x}} = d^{(c_1)} \\oplus d^{(c_2)} \\oplus \\dots \\oplus d^{(c_{|Y|})} \\), where \\( \\oplus \\) is the concatenation operator. An effective tem" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.263, + 0.885, + 0.326 + ], + "angle": 0, + "content": "plate, such as the one used in Lee et al. (2022), is \"\\(s^{(c)}\\). \\(e^{(c)}\\) is \\(t^{(c)}\\).\" Here, we refer the \"\\(e^{(c)}\\) is \\(t^{(c)}\\)\" part in the template as labeling part of the demonstration." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.341, + 0.877, + 0.374 + ], + "angle": 0, + "content": "3 Demonstration-based Learning from a Causal Perspective" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.385, + 0.885, + 0.529 + ], + "angle": 0, + "content": "In this section, we give a specific example to show the potential bias and understand demonstration-based learning from a causal perspective. Specifically, we first introduce a Structural Causal Model (SCM) (Pearl et al., 2000) to describe the mechanism and identify the induced bias. Then, we perform demonstration variable intervention and propose multiple simple and effective random demonstration templates inspired by our causal model." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.531, + 0.885, + 0.837 + ], + "angle": 0, + "content": "We observe that the frequent co-occurrence of tokens in the classical demonstrations generate harmful superficial patterns which is misleading to the model and leads to biased predictions (Zhang et al., 2022a; Min et al., 2022). A specific example with different demonstrations is provided in Table 1, where the entity to predict is French. Following previous work (Zhang et al., 2022a), the observed demonstrations (i.e., standard demonstration) provides some biased information: the concurrency of British and ICAC, which is an organization (ORG), may lead to biased predictions: French is labeled as an Organization while its desired prediction is other classes (MISC). Intuitively, the co-occurrence of two specific words in the demonstration may induce bias, while randomly sampled tokens in the demonstration do not. This specific example suggests why random demonstrations may sometimes perform better than standard ones." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.85, + 0.667, + 0.863 + ], + "angle": 0, + "content": "3.1 Causal Model" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.872, + 0.884, + 0.919 + ], + "angle": 0, + "content": "To study the causal relationship between the NER model and its training data, and explain the role of the demonstration, we introduce a SCM to describe" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1466" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.086, + 0.293, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.168, + 0.192, + 0.191, + 0.206 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.087, + 0.477, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.355, + 0.192, + 0.378, + 0.207 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.087, + 0.661, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.192, + 0.56, + 0.207 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.083, + 0.882, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.747, + 0.193, + 0.768, + 0.207 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.216, + 0.885, + 0.275 + ], + "angle": 0, + "content": "Figure 1: Causal views of NER. (a) shows a traditional NER model (Zeng et al., 2020), (b) shows the demonstration-based NER model under the causal view. With demonstration \\( D \\), the backdoor path \\( G \\rightarrow D \\rightarrow X \\) exists, which further introduces the bias. (c) shows the demonstration-based NER model with debiasing techniques, and the red cross means intervention. (d) is model architecture overview between classical and demonstration-based learning." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.286, + 0.49, + 0.606 + ], + "angle": 0, + "content": "the inference step in NER models. Figure 1 shows the SCM of NER models. There are mainly 6 variables in NER models: 1) Demonstration Tokens \\( D \\), the tokens which form the demonstration; 2) Context Tokens \\( C \\), the tokens that are related to the context; 3) Entity Tokens \\( E \\), the tokens which are entities; 4) Input Example \\( X \\), which is composed of \\( C \\) and \\( E \\) in the traditional model and composed of \\( C \\), \\( E \\) and \\( D \\) in the demonstration-based models; 5) Unobserved confounders \\( G \\), a confounding variable (not a concrete token) that influences the generation of \\( C \\), \\( E \\) and \\( D \\); 6) Evaluation result \\( Y \\), the evaluation result (the F1 score) of the NER models. Under the causal view, the key difference between the traditional NER model and the demonstration-based NER model is that, the demonstration-based NER model has an additional node \\( D \\). With the introduction of the demonstration \\( D \\), a backdoor path \\( G \\rightarrow D \\rightarrow X \\) exists, which further introduces the bias." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.609, + 0.49, + 0.801 + ], + "angle": 0, + "content": "Inspired by our SCM model (Figure 1b), we develop sampling techniques to generate new counterfactual examples by the interventions on the existing observational examples to alleviate this bias. The benefits of interventions on \\( E \\) and \\( C \\) have been studied in (Zeng et al., 2020). In this paper, we focus on understanding the role of demonstrations in NER models under the causal view. We understand the co-occurrence of tokens and harmful superficial patterns from the causal perspective and focus on using interventions on the demonstration variable to create new counterfactual demonstrations." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.817, + 0.461, + 0.831 + ], + "angle": 0, + "content": "3.2 Controllable Random Demonstrations" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.84, + 0.491, + 0.919 + ], + "angle": 0, + "content": "In this section, we first provide a running example to better understand the induced bias from human-crafted demonstrations and then present different ways of intervention on the demonstration tokens. The intervention is implemented via control" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.286, + 0.885, + 0.366 + ], + "angle": 0, + "content": "lable random demonstrations to create new counterfactual examples, as replacing standard demonstrations with random tokens can remove induce bias and still make the model a good few-shot learner (Zhang et al., 2022a)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.368, + 0.885, + 0.546 + ], + "angle": 0, + "content": "In Lee et al. (2022), an effective template \\( T \\) is \"\\( s^{(c)} \\). \\( e^{(c)} \\) is \\( t^{(c)} \\), and an example demonstration \\( d^{(c)} \\) can be \"[SEP] Obama returns to White House. Obama is PER.\" Intuitively, the model understands the demonstrations and then better performs inference. However, random demonstrations can still bring performance improvement (Zhang et al., 2022a). The random template is as simple as \\( [s_i]_{i=1}^L \\), where \\( s_i \\in p \\), and \\( p \\) is a token distribution. Random demonstrations are composed of \\( L \\) tokens randomly sampled from \\( p \\)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.548, + 0.884, + 0.756 + ], + "angle": 0, + "content": "Demonstration Intervention We use the intervention on the demonstration tokens to create new counterfactual examples, to alleviate the biases. If we do not carefully design D, the backdoor path will exist and the model performance is degraded. Our causal framework enables us to think about the problem from a causal perspective and guides us how to properly design D. We denote uniform distribution composed of vocabulary words of the PLMs as \\( p_{\\mathcal{V}} \\). Given the token distribution \\( p_{\\mathcal{V}} \\), for any word \\( w_i \\in p_{\\mathcal{V}} \\), we have \\( p_{\\mathcal{V}}(w_i) = \\frac{1}{|\\mathcal{V}|} \\). Then we have a plain way to construct random demonstrations." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.759, + 0.885, + 0.919 + ], + "angle": 0, + "content": "An important observation is that not all counterfactual examples are correct or useful. Hence, the intervention can be better implemented by replacing the uniform distribution with a non-uniform distribution, i.e., by adding or removing words and changing specific words' probabilities. Some mechanism is needed to identify good counterfactual demonstrations, to avoid introducing noise. An intuitive solution is that we consider tokens from the support set are more helpful as PLMs are fine" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1467" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.12, + 0.082, + 0.875, + 0.226 + ], + "angle": 0, + "content": "
ModeNERChunking
CoNLL03OntoNotes 5.0CoNLL00
F1PrecisionRecallF1PrecisionRecallF1PrecisionRecall
No Demo.28.71±10.3139.96±11.2522.68±9.0937.37±7.5833.80±6.7941.92±8.8563.17±4.2259.28±5.0567.72±3.51
Standard45.86±6.0847.38±5.9344.75±7.0740.21±7.6532.51±6.8752.82±8.2870.55±3.0866.53±4.4075.21±2.11
Random41.33±7.3645.41±7.3738.22±7.6539.71±7.5632.28±6.5651.63±8.7569.28±2.7864.75±3.8574.57±1.66
Rand-S45.55±8.0246.84±7.7144.60±8.6241.60±7.0533.96±6.2953.75±7.8070.63±3.0166.24±4.2975.75±1.70
Rand-W45.93±7.5747.79±7.4244.50±8.1345.49±3.7737.82±3.6457.18±4.1772.15±3.1668.00±4.4276.94±1.67
Rand-E47.32±7.4248.96±7.0246.02±8.1146.06±3.8438.32±3.6557.81±4.3174.02±2.9370.37±4.2378.18±1.75
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.233, + 0.883, + 0.261 + ], + "angle": 0, + "content": "Table 2: Main results for traditional token classification method (No Demo.) and demonstration-based learning with different modes of demonstrations under 5-shot scenario." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.271, + 0.489, + 0.333 + ], + "angle": 0, + "content": "tuned on the support set. We expect to see a better downstream predictor when the demonstrations are constructed randomly from a intervened token distribution." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.335, + 0.489, + 0.51 + ], + "angle": 0, + "content": "The difference between random demonstrations lies in the vocabulary and its associated probability distributions. We perform the interventions by controlling the vocabulary and changing the probability of random tokens. We encourage entity words (e.g., ICAC, British) to appear more frequently compared to the others (e.g., is). Based on the previous theoretical justification, we consider the following variants of constructing random demonstrations2 construction methods as counterfactual alternatives of the standard demonstrations3:" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.512, + 0.487, + 0.543 + ], + "angle": 0, + "content": "- Random: random context with tokens uniformly sampled from PLMs vocabulary \\(\\nu\\)." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.544, + 0.487, + 0.59 + ], + "angle": 0, + "content": "- Rand-S: random context with tokens uniformly sampled from unique words (i.e., vocabulary) of support set, denoted as \\(S\\)." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.592, + 0.486, + 0.655 + ], + "angle": 0, + "content": "- Rand-W4: random context with tokens sampled from \\( S \\), and entity tokens in support set, denoted as \\( W \\); tokens from \\( W \\) have four times higher probability compared with those from \\( S \\)." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.657, + 0.486, + 0.704 + ], + "angle": 0, + "content": "- Rand-E: similar to Rand-W, but replace entity tokens with entities composed of coherent tokens in support set, denoted as \\(\\mathcal{U}\\)." + }, + { + "type": "list", + "bbox": [ + 0.114, + 0.512, + 0.487, + 0.704 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.717, + 0.336, + 0.733 + ], + "angle": 0, + "content": "4 Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.742, + 0.304, + 0.759 + ], + "angle": 0, + "content": "4.1 Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.763, + 0.489, + 0.86 + ], + "angle": 0, + "content": "Datasets We conduct experiments on two sequence labeling tasks: (i) named entity recognition (NER) on dataset CoNLL03 (Tjong Kim Sang and De Meulder, 2003), and OntoNotes 5.0 (Weischedel et al., 2013); and (ii) chunking on dataset CoNLL00 (Tjong Kim Sang and Buchholz," + }, + { + "type": "page_footnote", + "bbox": [ + 0.136, + 0.867, + 0.374, + 0.88 + ], + "angle": 0, + "content": "2Random: [SEP] {random context}" + }, + { + "type": "page_footnote", + "bbox": [ + 0.136, + 0.88, + 0.462, + 0.894 + ], + "angle": 0, + "content": "3Standard: [SEP] {context} {entity} is {tag}." + }, + { + "type": "page_footnote", + "bbox": [ + 0.116, + 0.894, + 0.486, + 0.919 + ], + "angle": 0, + "content": "4 Empirical results show sampling only from \\(\\mathcal{W}\\) leads to poor performance." + }, + { + "type": "list", + "bbox": [ + 0.116, + 0.867, + 0.486, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.27, + 0.885, + 0.463 + ], + "angle": 0, + "content": "2000). Following previous works Ma et al. (2021); Zhang et al. (2022a), we omit the 7 value types in OntoNotes and only consider the 6 most frequent types in CoNLL00. For few-shot data sampling, we follow the greedy sampling strategy proposed by Yang and Katiyar (2020) to sample \\(K\\) shots for each type in an increasing order with respect to their frequencies, the detailed algorithm can be found. For each dataset, we sample 5 different \\(K\\)-shot support sets and report mean and standard deviation of metrics. For each \\(K\\)-shot support set, we run the experiments with 3 random seeds." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.469, + 0.884, + 0.919 + ], + "angle": 0, + "content": "Main Results We show the results for demonstration-based learning with different modes of demonstrations as well as classical sequence labeling with no demonstration in Table 2. The results show that demonstration-based method can consistently improve model performance. In demonstration-based methods, the Random approach shows the worst performance and Rand-S shows comparable results with the standard demonstrations, and the conclusion is consistent with previous works (Zhang et al., 2022a). Interestingly, if we modify the token sampling distributions and sample more entity or entity-related words as Rand-W and Rand-E, our model shows even better performance than standard meaningful demonstrations. The difference between Rand-W and Rand-E lies in whether there are complete entities, and the results show that adding complete entities instead of random entity words can lead to better performance. At the same time, it shows adding random tokens related to the support set can reduce the fine-tuned bias, which verifies our hypothesis in Section 3.1. Intuitively, the benefits of demonstration-based methods come from tokens of support sets \\( S \\) instead of meaningful demonstrations, as the standard demonstration sampled from the support set also shows good performance." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1468" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.082, + 0.379, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.082, + 0.641, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.082, + 0.882, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.203, + 0.231, + 0.794, + 0.246 + ], + "angle": 0, + "content": "Figure 2: Results with different support set size on CoNLL03, NRB and WTS datasets." + }, + { + "type": "table", + "bbox": [ + 0.124, + 0.261, + 0.475, + 0.362 + ], + "angle": 0, + "content": "
ModeCoNLL03OntoNotes5.0CoNLL00
No Demo.45.70±8.1351.62±2.7672.80±3.53
Standard45.73±7.2954.76±2.3675.90±1.95
Rand-S46.86±6.5054.35±2.6772.23±3.42
Rand-W52.11±6.1554.48±2.3573.84±2.19
Rand-E52.87±7.6455.94±2.3875.30±3.06
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.368, + 0.488, + 0.41 + ], + "angle": 0, + "content": "Table 3: Main results (F1 scores) of RoBERTa-Large for traditional token classification with different modes of demonstrations under 5-shot scenario." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.426, + 0.228, + 0.441 + ], + "angle": 0, + "content": "4.2 Analysis" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.45, + 0.49, + 0.706 + ], + "angle": 0, + "content": "Ablation Studies We further investigate whether the performance gain of demonstration-based learning changes over the size of support set. We present results of different modes of demonstrations under \\( K = 5, 10, 20 \\) shots in Figure 2. With more training examples in the support set, the relative performance gap between Rand-E and Standard remains, but it becomes smaller. This indicates that carefully designed random demonstrations show a consistent performance improvement upon standard demonstration. We also observe that the variance within each group becomes smaller as more data becomes available. Among random demonstrations, Rand-E consistently shows better performance than Rand-W and Rand-S, which verifies our hypothesis based on the SCM." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.709, + 0.49, + 0.804 + ], + "angle": 0, + "content": "Additionally, we investigate the effect of using different base models and replace BERT with RoBERTa. The observed results for RoBERTa in Table 3 are consistent with those of BERT, demonstrating that Rand-E exhibits superior performance across different model architectures." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.807, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Name Regularity Bias Name Regularity Bias (Ghaddar et al., 2021; Lin et al., 2020) in NER occurs when a model relies on a signal from the entity name to make predictions and disregards evidence from the local context. Ghaddar et al. (2021) carefully designed a testbed utilizing Wikipedia disambiguation pages to diagnose the Name Regu" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.258, + 0.882, + 0.29 + ], + "angle": 0, + "content": "larity Bias of NER models. Details about the NRB dataset are provided in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.291, + 0.884, + 0.404 + ], + "angle": 0, + "content": "We use both the NRB and WTS (as control sets) datasets to evaluate the model trained with different modes of demonstrations on CoNLL03. The results show a smaller gap for random demonstrations, suggesting that random demonstration-based learning can better leverage context information instead of the name regularity patterns." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.417, + 0.65, + 0.433 + ], + "angle": 0, + "content": "5 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.445, + 0.885, + 0.621 + ], + "angle": 0, + "content": "In this paper, we present a casual view to understand demonstration-based learning. Based on the structural causal model we constructed, we investigate the causal effects and discover that the concurrence of specific words in the demonstration can induce bias. To address this issue, we perform interventions by constructing random demonstrations. Our empirical results indicate that carefully designed random demonstrations consistently outperform meaningful demonstrations on public sequence labeling benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.635, + 0.646, + 0.65 + ], + "angle": 0, + "content": "6 Limitations" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.662, + 0.885, + 0.919 + ], + "angle": 0, + "content": "All our experiments are done on the sequence labeling task, and they can be further evaluated on sentence classification tasks with classifier-based fine-tuning since the [CLS] token used for classification represents the whole sentence. We provide a causal opinion on demonstration-based learning and a simple but not systematic method to alleviate the induced bias. Our demonstration-based learning builds upon previous works (Lee et al., 2022; Zhang et al., 2022a), where BERT or RoBERTa are used instead of Large Language Models, such as InstructGPT (Ouyang et al., 2022), PaLM (Chowdhery et al., 2022), and OPT (Zhang et al., 2022b). Furthermore, our conclusions are drawn from few-shot learning settings and cannot be directly applied to zero-shot inference." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1469" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.116, + 0.085, + 0.214, + 0.099 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.108, + 0.487, + 0.173 + ], + "angle": 0, + "content": "Rohan Anil, Andrew M Dai, Orhan Firat, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, et al. 2023. Palm 2 technical report. arXiv preprint arXiv:2305.10403." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.185, + 0.487, + 0.379 + ], + "angle": 0, + "content": "Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.392, + 0.487, + 0.469 + ], + "angle": 0, + "content": "Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. 2022. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.481, + 0.487, + 0.533 + ], + "angle": 0, + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.545, + 0.487, + 0.662 + ], + "angle": 0, + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.674, + 0.487, + 0.778 + ], + "angle": 0, + "content": "Tianyu Gao, Adam Fisch, and Danqi Chen. 2021. Making pre-trained language models better few-shot learners. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 3816-3830, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.789, + 0.487, + 0.854 + ], + "angle": 0, + "content": "Abbas Ghaddar, Philippe Langlais, Ahmad Rashid, and Mehdi Rezagholizadeh. 2021. Context-aware Adversarial Training for Name Regularity Bias in Named Entity Recognition. Transactions of the Association for Computational Linguistics, 9:586-604." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.866, + 0.487, + 0.918 + ], + "angle": 0, + "content": "Jiaxin Huang, Chunyuan Li, Krishan Subudhi, Damien Jose, Shobana Balakrishnan, Weizhu Chen, Baolin Peng, Jianfeng Gao, and Jiawei Han. 2021. Few-shot named entity recognition: An empirical baseline" + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.108, + 0.487, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.529, + 0.086, + 0.882, + 0.151 + ], + "angle": 0, + "content": "study. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 10408-10423, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.163, + 0.882, + 0.214 + ], + "angle": 0, + "content": "Brenden Lake, Ruslan Salakhutdinov, and Joshua Tenenbaum. 2015. Human-level concept learning through probabilistic program induction. Science, 350:1332-1338." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.226, + 0.882, + 0.344 + ], + "angle": 0, + "content": "Dong-Ho Lee, Akshen Kadakia, Kangmin Tan, Mahak Agarwal, Xinyu Feng, Takashi Shibuya, Ryosuke Mitani, Toshiyuki Sekiya, Jay Pujara, and Xiang Ren. 2022. Good examples make a faster learner: Simple demonstration-based learning for low-resource NER. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2687-2700, Dublin, Ireland. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.355, + 0.882, + 0.472 + ], + "angle": 0, + "content": "Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871-7880, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.483, + 0.882, + 0.588 + ], + "angle": 0, + "content": "Hongyu Lin, Yaojie Lu, Jialong Tang, Xianpei Han, Le Sun, Zhicheng Wei, and Nicholas Jing Yuan. 2020. A rigorous study on named entity recognition: Can fine-tuning pretrained model lead to the promised land? In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7291-7300, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.599, + 0.882, + 0.664 + ], + "angle": 0, + "content": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019a. Roberta: A robustly optimized BERT pretraining approach. CoRR, abs/1907.11692." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.675, + 0.882, + 0.741 + ], + "angle": 0, + "content": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019b. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.752, + 0.882, + 0.791 + ], + "angle": 0, + "content": "Ruotian Ma, Xin Zhou, Tao Gui, Yiding Tan, Qi Zhang, and Xuanjing Huang. 2021. Template-free prompt tuning for few-shot NER. CoRR, abs/2109.13532." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.802, + 0.882, + 0.868 + ], + "angle": 0, + "content": "Sewon Min, Xinxi Lyu, Ari Holtzman, Mikel Artetxe, Mike Lewis, Hannaneh Hajishirzi, and Luke Zettlemoyer. 2022. Rethinking the role of demonstrations: What makes in-context learning work? arXiv preprint arXiv:2202.12837." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.879, + 0.882, + 0.918 + ], + "angle": 0, + "content": "Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.882, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1470" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.135, + 0.086, + 0.49, + 0.113 + ], + "angle": 0, + "content": "2022. Training language models to follow instructions with human feedback. NeurIPS." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.123, + 0.49, + 0.162 + ], + "angle": 0, + "content": "Judea Pearl et al. 2000. Models, reasoning and inference. Cambridge, UK: Cambridge University Press, 19(2)." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.173, + 0.489, + 0.213 + ], + "angle": 0, + "content": "Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. 2023. Instruction tuning with gpt-4. arXiv preprint arXiv:2304.03277." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.222, + 0.489, + 0.288 + ], + "angle": 0, + "content": "Erik F. Tjong Kim Sang and Sabine Buchholz. 2000. Introduction to the CoNLL-2000 shared task chunking. In Fourth Conference on Computational Natural Language Learning and the Second Learning Language in Logic Workshop." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.298, + 0.489, + 0.376 + ], + "angle": 0, + "content": "Erik F. Tjong Kim Sang and Fien De Meulder. 2003. Introduction to the CoNLL-2003 shared task: Language-independent named entity recognition. In Proceedings of the Seventh Conference on Natural Language Learning at HLT-NAACL 2003, pages 142-147." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.387, + 0.489, + 0.464 + ], + "angle": 0, + "content": "Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.476, + 0.489, + 0.541 + ], + "angle": 0, + "content": "Ralph Weischedel, Martha Palmer, Mitchell Marcus, Edward Hovy, Sameer Pradhan, Lance Ramshaw, Nianwen Xue, Ann Taylor, Jeff Kaufman, Michelle Franchini, et al. 2013. Ontonotes release 5.0 ldc2013t19. Linguistic Data Consortium, Philadelphia, PA, 23." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.551, + 0.489, + 0.616 + ], + "angle": 0, + "content": "Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. 2020. Unsupervised data augmentation for consistency training. In Advances in Neural Information Processing Systems, volume 33, pages 6256-6268. Curran Associates, Inc." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.627, + 0.489, + 0.706 + ], + "angle": 0, + "content": "Yi Yang and Arzoo Katiyar. 2020. Simple and effective few-shot named entity recognition with structured nearest neighbor learning. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6365-6375, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.716, + 0.489, + 0.768 + ], + "angle": 0, + "content": "Xiangji Zeng, Yunliang Li, Yuchen Zhai, and Yin Zhang. 2020. Counterfactual generator: A weakly-supervised method for named entity recognition. In EMNLP." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.779, + 0.489, + 0.818 + ], + "angle": 0, + "content": "Hongxin Zhang, Yanzhe Zhang, Ruiyi Zhang, and Diyi Yang. 2022a. Robustness of demonstration-based learning under limited data scenario. In EMNLP." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.828, + 0.489, + 0.894 + ], + "angle": 0, + "content": "Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. 2022b. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.894 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.519, + 0.941 + ], + "angle": 0, + "content": "1471" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.169, + 0.083, + 0.434, + 0.163 + ], + "angle": 0, + "content": "
Dataset|Y|L|Dtest|
CoNLL034183453
OntoNotes 5.0112112217
CoNLL006362012
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.174, + 0.49, + 0.232 + ], + "angle": 0, + "content": "Table 4: Data Statistics. \\( |Y| \\): # of entity types. L: average # of tokens in input sentence. \\( |D_{support}| \\): average # of sentences in 5-shot support set over 5 different sub-samples. \\( |D_{test}| \\): # of sentences in test set." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.256, + 0.239, + 0.273 + ], + "angle": 0, + "content": "A Appendix" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.281, + 0.489, + 0.456 + ], + "angle": 0, + "content": "NRB Dataset Details The NRB dataset contains examples whose labels can be easily inferred from the local context, but they are difficult to be tagged by a popular NER system. The WTS dataset is a domain control set that includes the same query terms covered by NRB, but these can be correctly labeled by both the popular NER tagger and the local context-only tagger. Therefore, the gap between the NRB and WTS sets measures how effectively the model captures context information to predict token labels." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.459, + 0.49, + 0.603 + ], + "angle": 0, + "content": "Effects of Sampling Probability We present two variants, Random-E[X] and Random-W[X], where X refers to how many times the probability of preferred tokens is higher. In this ablation study, we consistently observe that Random-E4 performs better than Random-E2, and Random-W4 outperforms Random-E4. However, if we increase the X value to a very large number, the performance deteriorates." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1472" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.406, + 0.368, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.406, + 0.625, + 0.581 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.408, + 0.891, + 0.581 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.172, + 0.588, + 0.825, + 0.602 + ], + "angle": 0, + "content": "Figure 3: F1, Precision and Recall with more variants of Random Demonstrations on CoNLL03." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1473" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.133, + 0.085, + 0.435, + 0.1 + ], + "angle": 0, + "content": "ACL 2023 Responsible NLP Checklist" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.108, + 0.323, + 0.123 + ], + "angle": 0, + "content": "A For every submission:" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.127, + 0.533, + 0.158 + ], + "angle": 0, + "content": "A1. Did you describe the limitations of your work? Section 6" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.169, + 0.553, + 0.201 + ], + "angle": 0, + "content": "A2. Did you discuss any potential risks of your work? Section 6" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.213, + 0.696, + 0.246 + ], + "angle": 0, + "content": "A3. Do the abstract and introduction summarize the paper's main claims? Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.256, + 0.669, + 0.289 + ], + "angle": 0, + "content": "A4. Have you used AI writing assistants when working on this paper? Left blank." + }, + { + "type": "list", + "bbox": [ + 0.13, + 0.127, + 0.696, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.301, + 0.49, + 0.317 + ], + "angle": 0, + "content": "B Did you use or create scientific artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.322, + 0.216, + 0.337 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.348, + 0.531, + 0.38 + ], + "angle": 0, + "content": "B1. Did you cite the creators of artifacts you used? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.391, + 0.779, + 0.423 + ], + "angle": 0, + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.434, + 0.881, + 0.514 + ], + "angle": 0, + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.525, + 0.881, + 0.59 + ], + "angle": 0, + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.6, + 0.881, + 0.648 + ], + "angle": 0, + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.659, + 0.881, + 0.755 + ], + "angle": 0, + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Not applicable. Left blank." + }, + { + "type": "list", + "bbox": [ + 0.129, + 0.348, + 0.881, + 0.755 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.765, + 0.494, + 0.782 + ], + "angle": 0, + "content": "C Did you run computational experiments?" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.787, + 0.207, + 0.801 + ], + "angle": 0, + "content": "Section 4" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.813, + 0.881, + 0.86 + ], + "angle": 0, + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Section 4" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.867, + 0.878, + 0.89 + ], + "angle": 0, + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1474" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.13, + 0.084, + 0.88, + 0.131 + ], + "angle": 0, + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Section 4" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.143, + 0.881, + 0.207 + ], + "angle": 0, + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Section 4" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.218, + 0.881, + 0.281 + ], + "angle": 0, + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Section 4" + }, + { + "type": "list", + "bbox": [ + 0.13, + 0.084, + 0.881, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.294, + 0.878, + 0.331 + ], + "angle": 0, + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.341, + 0.881, + 0.388 + ], + "angle": 0, + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.4, + 0.881, + 0.464 + ], + "angle": 0, + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.476, + 0.881, + 0.54 + ], + "angle": 0, + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.55, + 0.873, + 0.582 + ], + "angle": 0, + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.593, + 0.881, + 0.641 + ], + "angle": 0, + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Not applicable. Left blank." + }, + { + "type": "list", + "bbox": [ + 0.128, + 0.341, + 0.881, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1475" + } + ] +] \ No newline at end of file diff --git a/2023/Understanding Demonstration-based Learning from a Causal Perspective/58464278-fa19-45b3-a09f-9c2dd2d2dc18_origin.pdf b/2023/Understanding Demonstration-based Learning from a Causal Perspective/58464278-fa19-45b3-a09f-9c2dd2d2dc18_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6e9abcfb0f5f8a8cb1fbe3f63682b0280acd54f7 --- /dev/null +++ b/2023/Understanding Demonstration-based Learning from a Causal Perspective/58464278-fa19-45b3-a09f-9c2dd2d2dc18_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cfbb88189094e25d5f55d22e28e8b7971dc90f76a09ed843cee4a7b51caaff6f +size 584191 diff --git a/2023/Understanding Demonstration-based Learning from a Causal Perspective/full.md b/2023/Understanding Demonstration-based Learning from a Causal Perspective/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d0d3e5a40cd9322b561593c9975eb97970691021 --- /dev/null +++ b/2023/Understanding Demonstration-based Learning from a Causal Perspective/full.md @@ -0,0 +1,220 @@ +# Understanding Demonstration-based Learning from a Causal Perspective + +Ruiyi Zhang + +Adobe Research ruizhang@adobe.com + +Tong Yu + +Adobe Research tyu@adobe.com + +# Abstract + +Demonstration-based learning has shown impressive performance in exploiting pretrained language models under few-shot learning settings. It is interesting to see that demonstrations, even those composed of random tokens, can still improve performance. In this paper, we build a Structural Causal Model (SCM) to understand demonstration-based learning from causal perspectives and interpret random demonstrations as interventions on the demonstration variable within the causal model. We investigate the causal effects and find that the concurrence of specific words in the demonstration will induce bias, while randomly sampled tokens in the demonstration do not. Based on this finding, we further propose simple ways to construct random demonstrations, which even outperform hand-crafted, meaningful demonstrations on public sequence labeling benchmarks1. + +# 1 Introduction + +Large pretrained language models (PLMs) have recently shown great progress (Devlin et al., 2019; Liu et al., 2019a; Lewis et al., 2020; Xie et al., 2020; Huang et al., 2021). These models, such as GPT-4 (Peng et al., 2023), PALM (Anil et al., 2023), and Llama (Touvron et al., 2023), have shown human-level capability with only a few illustrative examples (Lake et al., 2015). Specifically, demonstration-based learning has been introduced to augment the input with demonstrations, i.e., the input and expected output pairs. Brown et al. (2020) simply picked up to a small number of sampled instances and directly concatenated them with the input to perform in-context learning. Lee et al. (2022) concatenated the input with task demonstrations to create augmented input and fed them into PLMs to obtain improved token representations to do sequence labeling in a classifier-based fine-tuning way. + +However, how and why such demonstrations help still remains unclear, and there has been a growing amount of work investigating the mechanisms of demonstration-based learning. Min et al. (2022) investigated in-context learning with demonstrations under zero-shot settings and found that input with random labels can still produce performance comparable to that of correct labels. Zhang et al. (2022a) replaced every token in the demonstration with random ones and still surprisingly observed good few-shot learners even when the demonstration is meaningless. These observations conflict with some existing hypotheses (Gao et al., 2021; Lee et al., 2022) that models are learning meaningful knowledge from demonstrations. + +To better understand demonstration-based learning, we take a deeper dive into the random construction of demonstrations. Specifically, we first build a Structural Causal Model (SCM) to understand demonstration-based learning from a Causal Perspective. A causal view is developed to explore the spurious correlations between demonstrations and few-shot training samples. Based on the intervention on the demonstration variable in the SCM, we design multiple simple and effective ways to construct random demonstrations. These methods are evaluated on structured prediction tasks with carefully designed experiment setups. Empirical results show that carefully designed random demonstrations can outperform meaningful demonstrations under the few-shot learning setting. This finding suggests that meaningless demonstrations can still provide valid information for PLMs. Moreover, random demonstrations allow the learning algorithm to identify important features and patterns in the data more effectively than homogeneous handcrafted demonstrations. + +# 2 Background + +In this section, we introduce the background of sequence labeling and demonstration-based learning. + +
Sentence:The Algerian War of Independence marked the end of French colonial rule in North Africa .
Labels:O B-MISC I-MISC I-MISC O O O O B-ORG O O O B-LOC I-LOC O
Biased: French -> [ORG] Desired: French -> [MISC]
Standard:[SEP] The unnamed suspect left the British colony after being detained and then freed by the Independent Commission Against Corruption ( ICAC ) , the radio said . Independent Commission Against Corruption is ORG . [SEP] [...]
Random:[SEP] Lebanon First Ed ##up CBOE suspect CB Chicago K Chicago Board Options Exchange ##ty Paul Gascoigne CBOE Monday Les into vintage I ##tion Ferdinand ##ca Op [SEP] [...]
+ +Table 1: An example from the CoNLL03 dataset with different demonstrations. The NER model takes both the sentence and a demonstration as its inputs. The top two rows show examples of the NER model inputs and outputs with standard demonstrations. A biased prediction for 'French' is caused by the demonstration bias. The bottom three lines show three different demonstrations: Standard and Random demonstrations. The notation '[SEP][...]' indicates that there are demonstrations for other classes, which have been omitted due to limited space. + +Sequence Labeling Given an input sentence $\mathbf{x} = [x_{1}, x_{2}, \dots, x_{n}]$ composed of $n$ tokens, the sequence labeling task is to predict a tag $y_{i} \in Y \cup \{O\}$ for each token $x_{i}$ , where $Y$ is a predefined set of tags, and $O$ denotes outside a tagged span. In the few-shot setting, we only have $K$ -shot support set $S$ for training which contains $K$ examples for each tag type. This setting usually refers to $K$ -shot learning. Modern sequence labeling models are usually composed of an encoder and a classification head. The encoders are PLMs such as BERT (Devlin et al., 2018) and RoBERTa (Liu et al., 2019b), which provides contextualized representations for each token $\mathbf{h} = [h_{1}, h_{2}, \dots, h_{n}]$ given the natural language sequence $\mathbf{x} = [x_{1}, x_{2}, \dots, x_{n}]$ . The classification head takes these contextualized representations and predicts the label $l_{i}$ for each token $x_{i}$ . The model is optimized with the standard cross-entropy loss. + +Demonstration-based Learning Given some demonstration $\tilde{\mathbf{x}}$ , we concatenate the original input $\mathbf{x}$ with its demonstration $\tilde{\mathbf{x}}$ as $[\mathbf{x};\tilde{\mathbf{x}}]$ . We then feed the demonstration-augmented input $[\mathbf{x};\tilde{\mathbf{x}}]$ into the encoder, and get the contextualized representation $[\mathbf{h};\tilde{\mathbf{h}}]$ . The classification head takes $\mathbf{h}$ as the input and estimate the corresponding token's label $l_{i}$ in the original natural-language sequence. Please note that we use identical demonstrations during training and testing (Lee et al., 2022). + +Demonstration Construction To construct demonstrations, we first sample an entity $e^{(c)}$ for each label type $t^{(c)}$ , and its context $s^{(c)}$ from support set $S$ . Then we convert them into a natural language sequence $d^{(c)} = T(s^{(c)}, e^{(c)}, t^{(c)})$ , where $T$ is the template operator and previous works (Lee et al., 2022) focus on finding more effective templates. With these sequences $[d^{(c_i)}]_{i=1}^{|Y|}$ with different tags $c_i$ , a demonstration $\tilde{\mathbf{x}}$ is built by concatenating them together: $\tilde{\mathbf{x}} = d^{(c_1)} \oplus d^{(c_2)} \oplus \dots \oplus d^{(c_{|Y|})}$ , where $\oplus$ is the concatenation operator. An effective tem + +plate, such as the one used in Lee et al. (2022), is " $s^{(c)}$ . $e^{(c)}$ is $t^{(c)}$ ." Here, we refer the " $e^{(c)}$ is $t^{(c)}$ " part in the template as labeling part of the demonstration. + +# 3 Demonstration-based Learning from a Causal Perspective + +In this section, we give a specific example to show the potential bias and understand demonstration-based learning from a causal perspective. Specifically, we first introduce a Structural Causal Model (SCM) (Pearl et al., 2000) to describe the mechanism and identify the induced bias. Then, we perform demonstration variable intervention and propose multiple simple and effective random demonstration templates inspired by our causal model. + +We observe that the frequent co-occurrence of tokens in the classical demonstrations generate harmful superficial patterns which is misleading to the model and leads to biased predictions (Zhang et al., 2022a; Min et al., 2022). A specific example with different demonstrations is provided in Table 1, where the entity to predict is French. Following previous work (Zhang et al., 2022a), the observed demonstrations (i.e., standard demonstration) provides some biased information: the concurrency of British and ICAC, which is an organization (ORG), may lead to biased predictions: French is labeled as an Organization while its desired prediction is other classes (MISC). Intuitively, the co-occurrence of two specific words in the demonstration may induce bias, while randomly sampled tokens in the demonstration do not. This specific example suggests why random demonstrations may sometimes perform better than standard ones. + +# 3.1 Causal Model + +To study the causal relationship between the NER model and its training data, and explain the role of the demonstration, we introduce a SCM to describe + +![](images/c02e510f97f02f010e978031e4d8f0b374c424d313250b464c45bd264c6682c3.jpg) +(a) + +![](images/987321a8336f314009714664bfe25b3a1faaccc13230764ececde8df55dbed24.jpg) +(b) + +![](images/b90ea756a254755f329ba8c508bd923ee037cbaa91498920bb1a1dd606c0ce96.jpg) +(c) + +![](images/d43a2f5e031397c86a55695cbf9634643370921a08ffb03291bd3f22ee9f8765.jpg) +(d) +Figure 1: Causal views of NER. (a) shows a traditional NER model (Zeng et al., 2020), (b) shows the demonstration-based NER model under the causal view. With demonstration $D$ , the backdoor path $G \rightarrow D \rightarrow X$ exists, which further introduces the bias. (c) shows the demonstration-based NER model with debiasing techniques, and the red cross means intervention. (d) is model architecture overview between classical and demonstration-based learning. + +the inference step in NER models. Figure 1 shows the SCM of NER models. There are mainly 6 variables in NER models: 1) Demonstration Tokens $D$ , the tokens which form the demonstration; 2) Context Tokens $C$ , the tokens that are related to the context; 3) Entity Tokens $E$ , the tokens which are entities; 4) Input Example $X$ , which is composed of $C$ and $E$ in the traditional model and composed of $C$ , $E$ and $D$ in the demonstration-based models; 5) Unobserved confounders $G$ , a confounding variable (not a concrete token) that influences the generation of $C$ , $E$ and $D$ ; 6) Evaluation result $Y$ , the evaluation result (the F1 score) of the NER models. Under the causal view, the key difference between the traditional NER model and the demonstration-based NER model is that, the demonstration-based NER model has an additional node $D$ . With the introduction of the demonstration $D$ , a backdoor path $G \rightarrow D \rightarrow X$ exists, which further introduces the bias. + +Inspired by our SCM model (Figure 1b), we develop sampling techniques to generate new counterfactual examples by the interventions on the existing observational examples to alleviate this bias. The benefits of interventions on $E$ and $C$ have been studied in (Zeng et al., 2020). In this paper, we focus on understanding the role of demonstrations in NER models under the causal view. We understand the co-occurrence of tokens and harmful superficial patterns from the causal perspective and focus on using interventions on the demonstration variable to create new counterfactual demonstrations. + +# 3.2 Controllable Random Demonstrations + +In this section, we first provide a running example to better understand the induced bias from human-crafted demonstrations and then present different ways of intervention on the demonstration tokens. The intervention is implemented via control + +lable random demonstrations to create new counterfactual examples, as replacing standard demonstrations with random tokens can remove induce bias and still make the model a good few-shot learner (Zhang et al., 2022a). + +In Lee et al. (2022), an effective template $T$ is " $s^{(c)}$ . $e^{(c)}$ is $t^{(c)}$ , and an example demonstration $d^{(c)}$ can be "[SEP] Obama returns to White House. Obama is PER." Intuitively, the model understands the demonstrations and then better performs inference. However, random demonstrations can still bring performance improvement (Zhang et al., 2022a). The random template is as simple as $[s_i]_{i=1}^L$ , where $s_i \in p$ , and $p$ is a token distribution. Random demonstrations are composed of $L$ tokens randomly sampled from $p$ . + +Demonstration Intervention We use the intervention on the demonstration tokens to create new counterfactual examples, to alleviate the biases. If we do not carefully design D, the backdoor path will exist and the model performance is degraded. Our causal framework enables us to think about the problem from a causal perspective and guides us how to properly design D. We denote uniform distribution composed of vocabulary words of the PLMs as $p_{\mathcal{V}}$ . Given the token distribution $p_{\mathcal{V}}$ , for any word $w_i \in p_{\mathcal{V}}$ , we have $p_{\mathcal{V}}(w_i) = \frac{1}{|\mathcal{V}|}$ . Then we have a plain way to construct random demonstrations. + +An important observation is that not all counterfactual examples are correct or useful. Hence, the intervention can be better implemented by replacing the uniform distribution with a non-uniform distribution, i.e., by adding or removing words and changing specific words' probabilities. Some mechanism is needed to identify good counterfactual demonstrations, to avoid introducing noise. An intuitive solution is that we consider tokens from the support set are more helpful as PLMs are fine + +
ModeNERChunking
CoNLL03OntoNotes 5.0CoNLL00
F1PrecisionRecallF1PrecisionRecallF1PrecisionRecall
No Demo.28.71±10.3139.96±11.2522.68±9.0937.37±7.5833.80±6.7941.92±8.8563.17±4.2259.28±5.0567.72±3.51
Standard45.86±6.0847.38±5.9344.75±7.0740.21±7.6532.51±6.8752.82±8.2870.55±3.0866.53±4.4075.21±2.11
Random41.33±7.3645.41±7.3738.22±7.6539.71±7.5632.28±6.5651.63±8.7569.28±2.7864.75±3.8574.57±1.66
Rand-S45.55±8.0246.84±7.7144.60±8.6241.60±7.0533.96±6.2953.75±7.8070.63±3.0166.24±4.2975.75±1.70
Rand-W45.93±7.5747.79±7.4244.50±8.1345.49±3.7737.82±3.6457.18±4.1772.15±3.1668.00±4.4276.94±1.67
Rand-E47.32±7.4248.96±7.0246.02±8.1146.06±3.8438.32±3.6557.81±4.3174.02±2.9370.37±4.2378.18±1.75
+ +Table 2: Main results for traditional token classification method (No Demo.) and demonstration-based learning with different modes of demonstrations under 5-shot scenario. + +tuned on the support set. We expect to see a better downstream predictor when the demonstrations are constructed randomly from a intervened token distribution. + +The difference between random demonstrations lies in the vocabulary and its associated probability distributions. We perform the interventions by controlling the vocabulary and changing the probability of random tokens. We encourage entity words (e.g., ICAC, British) to appear more frequently compared to the others (e.g., is). Based on the previous theoretical justification, we consider the following variants of constructing random demonstrations2 construction methods as counterfactual alternatives of the standard demonstrations3: + +- Random: random context with tokens uniformly sampled from PLMs vocabulary $\nu$ . +- Rand-S: random context with tokens uniformly sampled from unique words (i.e., vocabulary) of support set, denoted as $S$ . +- Rand-W4: random context with tokens sampled from $S$ , and entity tokens in support set, denoted as $W$ ; tokens from $W$ have four times higher probability compared with those from $S$ . +- Rand-E: similar to Rand-W, but replace entity tokens with entities composed of coherent tokens in support set, denoted as $\mathcal{U}$ . + +# 4 Experimental Results + +# 4.1 Experiment Setup + +Datasets We conduct experiments on two sequence labeling tasks: (i) named entity recognition (NER) on dataset CoNLL03 (Tjong Kim Sang and De Meulder, 2003), and OntoNotes 5.0 (Weischedel et al., 2013); and (ii) chunking on dataset CoNLL00 (Tjong Kim Sang and Buchholz, + +2000). Following previous works Ma et al. (2021); Zhang et al. (2022a), we omit the 7 value types in OntoNotes and only consider the 6 most frequent types in CoNLL00. For few-shot data sampling, we follow the greedy sampling strategy proposed by Yang and Katiyar (2020) to sample $K$ shots for each type in an increasing order with respect to their frequencies, the detailed algorithm can be found. For each dataset, we sample 5 different $K$ -shot support sets and report mean and standard deviation of metrics. For each $K$ -shot support set, we run the experiments with 3 random seeds. + +Main Results We show the results for demonstration-based learning with different modes of demonstrations as well as classical sequence labeling with no demonstration in Table 2. The results show that demonstration-based method can consistently improve model performance. In demonstration-based methods, the Random approach shows the worst performance and Rand-S shows comparable results with the standard demonstrations, and the conclusion is consistent with previous works (Zhang et al., 2022a). Interestingly, if we modify the token sampling distributions and sample more entity or entity-related words as Rand-W and Rand-E, our model shows even better performance than standard meaningful demonstrations. The difference between Rand-W and Rand-E lies in whether there are complete entities, and the results show that adding complete entities instead of random entity words can lead to better performance. At the same time, it shows adding random tokens related to the support set can reduce the fine-tuned bias, which verifies our hypothesis in Section 3.1. Intuitively, the benefits of demonstration-based methods come from tokens of support sets $S$ instead of meaningful demonstrations, as the standard demonstration sampled from the support set also shows good performance. + +![](images/c081e6404c2d14411e0af537694d5665345a88764e5bbd0c6639d760b30d3098.jpg) +Figure 2: Results with different support set size on CoNLL03, NRB and WTS datasets. + +![](images/47e737b77f0e5157053023e79ebe490557513eb838b5aff71480e50c7ecc9650.jpg) + +![](images/b5bc4c4268bb8d2f235c89ba405172e60206d99cb1707f253ba822159311a774.jpg) + +
ModeCoNLL03OntoNotes5.0CoNLL00
No Demo.45.70±8.1351.62±2.7672.80±3.53
Standard45.73±7.2954.76±2.3675.90±1.95
Rand-S46.86±6.5054.35±2.6772.23±3.42
Rand-W52.11±6.1554.48±2.3573.84±2.19
Rand-E52.87±7.6455.94±2.3875.30±3.06
+ +Table 3: Main results (F1 scores) of RoBERTa-Large for traditional token classification with different modes of demonstrations under 5-shot scenario. + +# 4.2 Analysis + +Ablation Studies We further investigate whether the performance gain of demonstration-based learning changes over the size of support set. We present results of different modes of demonstrations under $K = 5, 10, 20$ shots in Figure 2. With more training examples in the support set, the relative performance gap between Rand-E and Standard remains, but it becomes smaller. This indicates that carefully designed random demonstrations show a consistent performance improvement upon standard demonstration. We also observe that the variance within each group becomes smaller as more data becomes available. Among random demonstrations, Rand-E consistently shows better performance than Rand-W and Rand-S, which verifies our hypothesis based on the SCM. + +Additionally, we investigate the effect of using different base models and replace BERT with RoBERTa. The observed results for RoBERTa in Table 3 are consistent with those of BERT, demonstrating that Rand-E exhibits superior performance across different model architectures. + +Name Regularity Bias Name Regularity Bias (Ghaddar et al., 2021; Lin et al., 2020) in NER occurs when a model relies on a signal from the entity name to make predictions and disregards evidence from the local context. Ghaddar et al. (2021) carefully designed a testbed utilizing Wikipedia disambiguation pages to diagnose the Name Regu + +larity Bias of NER models. Details about the NRB dataset are provided in the appendix. + +We use both the NRB and WTS (as control sets) datasets to evaluate the model trained with different modes of demonstrations on CoNLL03. The results show a smaller gap for random demonstrations, suggesting that random demonstration-based learning can better leverage context information instead of the name regularity patterns. + +# 5 Conclusions + +In this paper, we present a casual view to understand demonstration-based learning. Based on the structural causal model we constructed, we investigate the causal effects and discover that the concurrence of specific words in the demonstration can induce bias. To address this issue, we perform interventions by constructing random demonstrations. Our empirical results indicate that carefully designed random demonstrations consistently outperform meaningful demonstrations on public sequence labeling benchmarks. + +# 6 Limitations + +All our experiments are done on the sequence labeling task, and they can be further evaluated on sentence classification tasks with classifier-based fine-tuning since the [CLS] token used for classification represents the whole sentence. We provide a causal opinion on demonstration-based learning and a simple but not systematic method to alleviate the induced bias. Our demonstration-based learning builds upon previous works (Lee et al., 2022; Zhang et al., 2022a), where BERT or RoBERTa are used instead of Large Language Models, such as InstructGPT (Ouyang et al., 2022), PaLM (Chowdhery et al., 2022), and OPT (Zhang et al., 2022b). Furthermore, our conclusions are drawn from few-shot learning settings and cannot be directly applied to zero-shot inference. + +# References + +Rohan Anil, Andrew M Dai, Orhan Firat, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, et al. 2023. Palm 2 technical report. arXiv preprint arXiv:2305.10403. +Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual. +Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. 2022. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311. +Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805. +Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics. +Tianyu Gao, Adam Fisch, and Danqi Chen. 2021. Making pre-trained language models better few-shot learners. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 3816-3830, Online. Association for Computational Linguistics. +Abbas Ghaddar, Philippe Langlais, Ahmad Rashid, and Mehdi Rezagholizadeh. 2021. Context-aware Adversarial Training for Name Regularity Bias in Named Entity Recognition. Transactions of the Association for Computational Linguistics, 9:586-604. +Jiaxin Huang, Chunyuan Li, Krishan Subudhi, Damien Jose, Shobana Balakrishnan, Weizhu Chen, Baolin Peng, Jianfeng Gao, and Jiawei Han. 2021. Few-shot named entity recognition: An empirical baseline + +study. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 10408-10423, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics. +Brenden Lake, Ruslan Salakhutdinov, and Joshua Tenenbaum. 2015. Human-level concept learning through probabilistic program induction. Science, 350:1332-1338. +Dong-Ho Lee, Akshen Kadakia, Kangmin Tan, Mahak Agarwal, Xinyu Feng, Takashi Shibuya, Ryosuke Mitani, Toshiyuki Sekiya, Jay Pujara, and Xiang Ren. 2022. Good examples make a faster learner: Simple demonstration-based learning for low-resource NER. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2687-2700, Dublin, Ireland. Association for Computational Linguistics. +Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871-7880, Online. Association for Computational Linguistics. +Hongyu Lin, Yaojie Lu, Jialong Tang, Xianpei Han, Le Sun, Zhicheng Wei, and Nicholas Jing Yuan. 2020. A rigorous study on named entity recognition: Can fine-tuning pretrained model lead to the promised land? In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7291-7300, Online. Association for Computational Linguistics. +Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019a. Roberta: A robustly optimized BERT pretraining approach. CoRR, abs/1907.11692. +Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019b. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692. +Ruotian Ma, Xin Zhou, Tao Gui, Yiding Tan, Qi Zhang, and Xuanjing Huang. 2021. Template-free prompt tuning for few-shot NER. CoRR, abs/2109.13532. +Sewon Min, Xinxi Lyu, Ari Holtzman, Mikel Artetxe, Mike Lewis, Hannaneh Hajishirzi, and Luke Zettlemoyer. 2022. Rethinking the role of demonstrations: What makes in-context learning work? arXiv preprint arXiv:2202.12837. +Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. + +2022. Training language models to follow instructions with human feedback. NeurIPS. +Judea Pearl et al. 2000. Models, reasoning and inference. Cambridge, UK: Cambridge University Press, 19(2). +Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. 2023. Instruction tuning with gpt-4. arXiv preprint arXiv:2304.03277. +Erik F. Tjong Kim Sang and Sabine Buchholz. 2000. Introduction to the CoNLL-2000 shared task chunking. In Fourth Conference on Computational Natural Language Learning and the Second Learning Language in Logic Workshop. +Erik F. Tjong Kim Sang and Fien De Meulder. 2003. Introduction to the CoNLL-2003 shared task: Language-independent named entity recognition. In Proceedings of the Seventh Conference on Natural Language Learning at HLT-NAACL 2003, pages 142-147. +Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971. +Ralph Weischedel, Martha Palmer, Mitchell Marcus, Edward Hovy, Sameer Pradhan, Lance Ramshaw, Nianwen Xue, Ann Taylor, Jeff Kaufman, Michelle Franchini, et al. 2013. Ontonotes release 5.0 ldc2013t19. Linguistic Data Consortium, Philadelphia, PA, 23. +Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. 2020. Unsupervised data augmentation for consistency training. In Advances in Neural Information Processing Systems, volume 33, pages 6256-6268. Curran Associates, Inc. +Yi Yang and Arzoo Katiyar. 2020. Simple and effective few-shot named entity recognition with structured nearest neighbor learning. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6365-6375, Online. Association for Computational Linguistics. +Xiangji Zeng, Yunliang Li, Yuchen Zhai, and Yin Zhang. 2020. Counterfactual generator: A weakly-supervised method for named entity recognition. In EMNLP. +Hongxin Zhang, Yanzhe Zhang, Ruiyi Zhang, and Diyi Yang. 2022a. Robustness of demonstration-based learning under limited data scenario. In EMNLP. +Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. 2022b. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068. + +
Dataset|Y|L|Dtest|
CoNLL034183453
OntoNotes 5.0112112217
CoNLL006362012
+ +Table 4: Data Statistics. $|Y|$ : # of entity types. L: average # of tokens in input sentence. $|D_{support}|$ : average # of sentences in 5-shot support set over 5 different sub-samples. $|D_{test}|$ : # of sentences in test set. + +# A Appendix + +NRB Dataset Details The NRB dataset contains examples whose labels can be easily inferred from the local context, but they are difficult to be tagged by a popular NER system. The WTS dataset is a domain control set that includes the same query terms covered by NRB, but these can be correctly labeled by both the popular NER tagger and the local context-only tagger. Therefore, the gap between the NRB and WTS sets measures how effectively the model captures context information to predict token labels. + +Effects of Sampling Probability We present two variants, Random-E[X] and Random-W[X], where X refers to how many times the probability of preferred tokens is higher. In this ablation study, we consistently observe that Random-E4 performs better than Random-E2, and Random-W4 outperforms Random-E4. However, if we increase the X value to a very large number, the performance deteriorates. + +![](images/4d287e32f49c4496c1ede1687bce3ce1b2488a7c6804de1f0e6fee6e7e19d13b.jpg) +Figure 3: F1, Precision and Recall with more variants of Random Demonstrations on CoNLL03. + +![](images/947aa770ce7060abb984559ce3a3fbf4b3483e8097b906e888edfd68305777ea.jpg) + +![](images/c5d6fc26cb4ac6b2d6ad7911b284eefcd60b7610b7d3d6a9ff1046eaa2fff073.jpg) + +A For every submission: + +A1. Did you describe the limitations of your work? Section 6 +A2. Did you discuss any potential risks of your work? Section 6 +A3. Do the abstract and introduction summarize the paper's main claims? Left blank. +A4. Have you used AI writing assistants when working on this paper? Left blank. + +B Did you use or create scientific artifacts? + +Left blank. + +B1. Did you cite the creators of artifacts you used? Not applicable. Left blank. +B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Not applicable. Left blank. +B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Not applicable. Left blank. +B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Not applicable. Left blank. +B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Not applicable. Left blank. +B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Not applicable. Left blank. + +C Did you run computational experiments? + +Section 4 + +C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Section 4 + +The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance. + +C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Section 4 +C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Section 4 +C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Section 4 + +D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank. + +D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Not applicable. Left blank. +D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Not applicable. Left blank. +D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Not applicable. Left blank. +D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Not applicable. Left blank. +D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Not applicable. Left blank. \ No newline at end of file diff --git a/2023/Understanding Demonstration-based Learning from a Causal Perspective/images.zip b/2023/Understanding Demonstration-based Learning from a Causal Perspective/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3bc199d1628738a54815d41d6a27b5ee6c7e4e1d --- /dev/null +++ b/2023/Understanding Demonstration-based Learning from a Causal Perspective/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7a635010914d368f8fd9a3252b59efceb61d6fc8666f60bdca46184a9d9d461c +size 265597 diff --git a/2023/Understanding Demonstration-based Learning from a Causal Perspective/layout.json b/2023/Understanding Demonstration-based Learning from a Causal Perspective/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f5bcdaf4bce880b602ca937cb32939aff6bc11dd --- /dev/null +++ b/2023/Understanding Demonstration-based Learning from a Causal Perspective/layout.json @@ -0,0 +1,7316 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 71, + 76, + 523, + 93 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 76, + 523, + 93 + ], + "spans": [ + { + "bbox": [ + 71, + 76, + 523, + 93 + ], + "type": "text", + "content": "Understanding Demonstration-based Learning from a Causal Perspective" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 164, + 121, + 232, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 121, + 232, + 133 + ], + "spans": [ + { + "bbox": [ + 164, + 121, + 232, + 133 + ], + "type": "text", + "content": "Ruiyi Zhang" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 143, + 135, + 254, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 135, + 254, + 162 + ], + "spans": [ + { + "bbox": [ + 143, + 135, + 254, + 162 + ], + "type": "text", + "content": "Adobe Research ruizhang@adobe.com" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 373, + 121, + 420, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 121, + 420, + 133 + ], + "spans": [ + { + "bbox": [ + 373, + 121, + 420, + 133 + ], + "type": "text", + "content": "Tong Yu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 356, + 135, + 437, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 135, + 437, + 162 + ], + "spans": [ + { + "bbox": [ + 356, + 135, + 437, + 162 + ], + "type": "text", + "content": "Adobe Research tyu@adobe.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 155, + 212, + 202, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 212, + 202, + 225 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 202, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 84, + 232, + 274, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 232, + 274, + 470 + ], + "spans": [ + { + "bbox": [ + 84, + 232, + 274, + 470 + ], + "type": "text", + "content": "Demonstration-based learning has shown impressive performance in exploiting pretrained language models under few-shot learning settings. It is interesting to see that demonstrations, even those composed of random tokens, can still improve performance. In this paper, we build a Structural Causal Model (SCM) to understand demonstration-based learning from causal perspectives and interpret random demonstrations as interventions on the demonstration variable within the causal model. We investigate the causal effects and find that the concurrence of specific words in the demonstration will induce bias, while randomly sampled tokens in the demonstration do not. Based on this finding, we further propose simple ways to construct random demonstrations, which even outperform hand-crafted, meaningful demonstrations on public sequence labeling benchmarks1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 68, + 477, + 154, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 477, + 154, + 491 + ], + "spans": [ + { + "bbox": [ + 68, + 477, + 154, + 491 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 498, + 291, + 756 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 498, + 291, + 756 + ], + "spans": [ + { + "bbox": [ + 67, + 498, + 291, + 756 + ], + "type": "text", + "content": "Large pretrained language models (PLMs) have recently shown great progress (Devlin et al., 2019; Liu et al., 2019a; Lewis et al., 2020; Xie et al., 2020; Huang et al., 2021). These models, such as GPT-4 (Peng et al., 2023), PALM (Anil et al., 2023), and Llama (Touvron et al., 2023), have shown human-level capability with only a few illustrative examples (Lake et al., 2015). Specifically, demonstration-based learning has been introduced to augment the input with demonstrations, i.e., the input and expected output pairs. Brown et al. (2020) simply picked up to a small number of sampled instances and directly concatenated them with the input to perform in-context learning. Lee et al. (2022) concatenated the input with task demonstrations to create augmented input and fed them into PLMs to obtain improved token representations to do sequence labeling in a classifier-based fine-tuning way." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 213, + 526, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 213, + 526, + 414 + ], + "spans": [ + { + "bbox": [ + 302, + 213, + 526, + 414 + ], + "type": "text", + "content": "However, how and why such demonstrations help still remains unclear, and there has been a growing amount of work investigating the mechanisms of demonstration-based learning. Min et al. (2022) investigated in-context learning with demonstrations under zero-shot settings and found that input with random labels can still produce performance comparable to that of correct labels. Zhang et al. (2022a) replaced every token in the demonstration with random ones and still surprisingly observed good few-shot learners even when the demonstration is meaningless. These observations conflict with some existing hypotheses (Gao et al., 2021; Lee et al., 2022) that models are learning meaningful knowledge from demonstrations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 417, + 526, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 417, + 526, + 713 + ], + "spans": [ + { + "bbox": [ + 302, + 417, + 526, + 713 + ], + "type": "text", + "content": "To better understand demonstration-based learning, we take a deeper dive into the random construction of demonstrations. Specifically, we first build a Structural Causal Model (SCM) to understand demonstration-based learning from a Causal Perspective. A causal view is developed to explore the spurious correlations between demonstrations and few-shot training samples. Based on the intervention on the demonstration variable in the SCM, we design multiple simple and effective ways to construct random demonstrations. These methods are evaluated on structured prediction tasks with carefully designed experiment setups. Empirical results show that carefully designed random demonstrations can outperform meaningful demonstrations under the few-shot learning setting. This finding suggests that meaningless demonstrations can still provide valid information for PLMs. Moreover, random demonstrations allow the learning algorithm to identify important features and patterns in the data more effectively than homogeneous handcrafted demonstrations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 724, + 387, + 738 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 724, + 387, + 738 + ], + "spans": [ + { + "bbox": [ + 302, + 724, + 387, + 738 + ], + "type": "text", + "content": "2 Background" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 746, + 526, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 746, + 526, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 746, + 526, + 773 + ], + "type": "text", + "content": "In this section, we introduce the background of sequence labeling and demonstration-based learning." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 81, + 760, + 283, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 760, + 283, + 772 + ], + "spans": [ + { + "bbox": [ + 81, + 760, + 283, + 772 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 81, + 760, + 283, + 772 + ], + "type": "text", + "content": "Code available at: github.com/zhangry868/RandDemo" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "1465" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "spans": [ + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "type": "text", + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 219, + 806, + 374, + 817 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 806, + 374, + 817 + ], + "spans": [ + { + "bbox": [ + 219, + 806, + 374, + 817 + ], + "type": "text", + "content": "Volume 2: Short Papers, pages 1465-1475" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "spans": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "text", + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 69, + 523, + 148 + ], + "blocks": [ + { + "bbox": [ + 69, + 69, + 523, + 148 + ], + "lines": [ + { + "bbox": [ + 69, + 69, + 523, + 148 + ], + "spans": [ + { + "bbox": [ + 69, + 69, + 523, + 148 + ], + "type": "table", + "html": "
Sentence:The Algerian War of Independence marked the end of French colonial rule in North Africa .
Labels:O B-MISC I-MISC I-MISC O O O O B-ORG O O O B-LOC I-LOC O
Biased: French -> [ORG] Desired: French -> [MISC]
Standard:[SEP] The unnamed suspect left the British colony after being detained and then freed by the Independent Commission Against Corruption ( ICAC ) , the radio said . Independent Commission Against Corruption is ORG . [SEP] [...]
Random:[SEP] Lebanon First Ed ##up CBOE suspect CB Chicago K Chicago Board Options Exchange ##ty Paul Gascoigne CBOE Monday Les into vintage I ##tion Ferdinand ##ca Op [SEP] [...]
", + "image_path": "b9a0e5e6260873fdda05b1796d73eb62ddec68965915154d28b93b9b380aeac8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 153, + 526, + 214 + ], + "lines": [ + { + "bbox": [ + 66, + 153, + 526, + 214 + ], + "spans": [ + { + "bbox": [ + 66, + 153, + 526, + 214 + ], + "type": "text", + "content": "Table 1: An example from the CoNLL03 dataset with different demonstrations. The NER model takes both the sentence and a demonstration as its inputs. The top two rows show examples of the NER model inputs and outputs with standard demonstrations. A biased prediction for 'French' is caused by the demonstration bias. The bottom three lines show three different demonstrations: Standard and Random demonstrations. The notation '[SEP][...]' indicates that there are demonstrations for other classes, which have been omitted due to limited space." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 221, + 291, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 221, + 291, + 478 + ], + "spans": [ + { + "bbox": [ + 67, + 221, + 291, + 478 + ], + "type": "text", + "content": "Sequence Labeling Given an input sentence " + }, + { + "bbox": [ + 67, + 221, + 291, + 478 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = [x_{1}, x_{2}, \\dots, x_{n}]" + }, + { + "bbox": [ + 67, + 221, + 291, + 478 + ], + "type": "text", + "content": " composed of " + }, + { + "bbox": [ + 67, + 221, + 291, + 478 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 67, + 221, + 291, + 478 + ], + "type": "text", + "content": " tokens, the sequence labeling task is to predict a tag " + }, + { + "bbox": [ + 67, + 221, + 291, + 478 + ], + "type": "inline_equation", + "content": "y_{i} \\in Y \\cup \\{O\\}" + }, + { + "bbox": [ + 67, + 221, + 291, + 478 + ], + "type": "text", + "content": " for each token " + }, + { + "bbox": [ + 67, + 221, + 291, + 478 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 67, + 221, + 291, + 478 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 67, + 221, + 291, + 478 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 67, + 221, + 291, + 478 + ], + "type": "text", + "content": " is a predefined set of tags, and " + }, + { + "bbox": [ + 67, + 221, + 291, + 478 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 67, + 221, + 291, + 478 + ], + "type": "text", + "content": " denotes outside a tagged span. In the few-shot setting, we only have " + }, + { + "bbox": [ + 67, + 221, + 291, + 478 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 67, + 221, + 291, + 478 + ], + "type": "text", + "content": "-shot support set " + }, + { + "bbox": [ + 67, + 221, + 291, + 478 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 67, + 221, + 291, + 478 + ], + "type": "text", + "content": " for training which contains " + }, + { + "bbox": [ + 67, + 221, + 291, + 478 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 67, + 221, + 291, + 478 + ], + "type": "text", + "content": " examples for each tag type. This setting usually refers to " + }, + { + "bbox": [ + 67, + 221, + 291, + 478 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 67, + 221, + 291, + 478 + ], + "type": "text", + "content": "-shot learning. Modern sequence labeling models are usually composed of an encoder and a classification head. The encoders are PLMs such as BERT (Devlin et al., 2018) and RoBERTa (Liu et al., 2019b), which provides contextualized representations for each token " + }, + { + "bbox": [ + 67, + 221, + 291, + 478 + ], + "type": "inline_equation", + "content": "\\mathbf{h} = [h_{1}, h_{2}, \\dots, h_{n}]" + }, + { + "bbox": [ + 67, + 221, + 291, + 478 + ], + "type": "text", + "content": " given the natural language sequence " + }, + { + "bbox": [ + 67, + 221, + 291, + 478 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = [x_{1}, x_{2}, \\dots, x_{n}]" + }, + { + "bbox": [ + 67, + 221, + 291, + 478 + ], + "type": "text", + "content": ". The classification head takes these contextualized representations and predicts the label " + }, + { + "bbox": [ + 67, + 221, + 291, + 478 + ], + "type": "inline_equation", + "content": "l_{i}" + }, + { + "bbox": [ + 67, + 221, + 291, + 478 + ], + "type": "text", + "content": " for each token " + }, + { + "bbox": [ + 67, + 221, + 291, + 478 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 67, + 221, + 291, + 478 + ], + "type": "text", + "content": ". The model is optimized with the standard cross-entropy loss." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 483, + 291, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 483, + 291, + 618 + ], + "spans": [ + { + "bbox": [ + 67, + 483, + 291, + 618 + ], + "type": "text", + "content": "Demonstration-based Learning Given some demonstration " + }, + { + "bbox": [ + 67, + 483, + 291, + 618 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}" + }, + { + "bbox": [ + 67, + 483, + 291, + 618 + ], + "type": "text", + "content": ", we concatenate the original input " + }, + { + "bbox": [ + 67, + 483, + 291, + 618 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 67, + 483, + 291, + 618 + ], + "type": "text", + "content": " with its demonstration " + }, + { + "bbox": [ + 67, + 483, + 291, + 618 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}" + }, + { + "bbox": [ + 67, + 483, + 291, + 618 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 67, + 483, + 291, + 618 + ], + "type": "inline_equation", + "content": "[\\mathbf{x};\\tilde{\\mathbf{x}}]" + }, + { + "bbox": [ + 67, + 483, + 291, + 618 + ], + "type": "text", + "content": ". We then feed the demonstration-augmented input " + }, + { + "bbox": [ + 67, + 483, + 291, + 618 + ], + "type": "inline_equation", + "content": "[\\mathbf{x};\\tilde{\\mathbf{x}}]" + }, + { + "bbox": [ + 67, + 483, + 291, + 618 + ], + "type": "text", + "content": " into the encoder, and get the contextualized representation " + }, + { + "bbox": [ + 67, + 483, + 291, + 618 + ], + "type": "inline_equation", + "content": "[\\mathbf{h};\\tilde{\\mathbf{h}}]" + }, + { + "bbox": [ + 67, + 483, + 291, + 618 + ], + "type": "text", + "content": ". The classification head takes " + }, + { + "bbox": [ + 67, + 483, + 291, + 618 + ], + "type": "inline_equation", + "content": "\\mathbf{h}" + }, + { + "bbox": [ + 67, + 483, + 291, + 618 + ], + "type": "text", + "content": " as the input and estimate the corresponding token's label " + }, + { + "bbox": [ + 67, + 483, + 291, + 618 + ], + "type": "inline_equation", + "content": "l_{i}" + }, + { + "bbox": [ + 67, + 483, + 291, + 618 + ], + "type": "text", + "content": " in the original natural-language sequence. Please note that we use identical demonstrations during training and testing (Lee et al., 2022)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 623, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 623, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 623, + 291, + 772 + ], + "type": "text", + "content": "Demonstration Construction To construct demonstrations, we first sample an entity " + }, + { + "bbox": [ + 67, + 623, + 291, + 772 + ], + "type": "inline_equation", + "content": "e^{(c)}" + }, + { + "bbox": [ + 67, + 623, + 291, + 772 + ], + "type": "text", + "content": " for each label type " + }, + { + "bbox": [ + 67, + 623, + 291, + 772 + ], + "type": "inline_equation", + "content": "t^{(c)}" + }, + { + "bbox": [ + 67, + 623, + 291, + 772 + ], + "type": "text", + "content": ", and its context " + }, + { + "bbox": [ + 67, + 623, + 291, + 772 + ], + "type": "inline_equation", + "content": "s^{(c)}" + }, + { + "bbox": [ + 67, + 623, + 291, + 772 + ], + "type": "text", + "content": " from support set " + }, + { + "bbox": [ + 67, + 623, + 291, + 772 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 67, + 623, + 291, + 772 + ], + "type": "text", + "content": ". Then we convert them into a natural language sequence " + }, + { + "bbox": [ + 67, + 623, + 291, + 772 + ], + "type": "inline_equation", + "content": "d^{(c)} = T(s^{(c)}, e^{(c)}, t^{(c)})" + }, + { + "bbox": [ + 67, + 623, + 291, + 772 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 67, + 623, + 291, + 772 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 67, + 623, + 291, + 772 + ], + "type": "text", + "content": " is the template operator and previous works (Lee et al., 2022) focus on finding more effective templates. With these sequences " + }, + { + "bbox": [ + 67, + 623, + 291, + 772 + ], + "type": "inline_equation", + "content": "[d^{(c_i)}]_{i=1}^{|Y|}" + }, + { + "bbox": [ + 67, + 623, + 291, + 772 + ], + "type": "text", + "content": " with different tags " + }, + { + "bbox": [ + 67, + 623, + 291, + 772 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 67, + 623, + 291, + 772 + ], + "type": "text", + "content": ", a demonstration " + }, + { + "bbox": [ + 67, + 623, + 291, + 772 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}" + }, + { + "bbox": [ + 67, + 623, + 291, + 772 + ], + "type": "text", + "content": " is built by concatenating them together: " + }, + { + "bbox": [ + 67, + 623, + 291, + 772 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}} = d^{(c_1)} \\oplus d^{(c_2)} \\oplus \\dots \\oplus d^{(c_{|Y|})}" + }, + { + "bbox": [ + 67, + 623, + 291, + 772 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 67, + 623, + 291, + 772 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 67, + 623, + 291, + 772 + ], + "type": "text", + "content": " is the concatenation operator. An effective tem" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 302, + 221, + 526, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 221, + 526, + 274 + ], + "spans": [ + { + "bbox": [ + 302, + 221, + 526, + 274 + ], + "type": "text", + "content": "plate, such as the one used in Lee et al. (2022), is \"" + }, + { + "bbox": [ + 302, + 221, + 526, + 274 + ], + "type": "inline_equation", + "content": "s^{(c)}" + }, + { + "bbox": [ + 302, + 221, + 526, + 274 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 302, + 221, + 526, + 274 + ], + "type": "inline_equation", + "content": "e^{(c)}" + }, + { + "bbox": [ + 302, + 221, + 526, + 274 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 302, + 221, + 526, + 274 + ], + "type": "inline_equation", + "content": "t^{(c)}" + }, + { + "bbox": [ + 302, + 221, + 526, + 274 + ], + "type": "text", + "content": ".\" Here, we refer the \"" + }, + { + "bbox": [ + 302, + 221, + 526, + 274 + ], + "type": "inline_equation", + "content": "e^{(c)}" + }, + { + "bbox": [ + 302, + 221, + 526, + 274 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 302, + 221, + 526, + 274 + ], + "type": "inline_equation", + "content": "t^{(c)}" + }, + { + "bbox": [ + 302, + 221, + 526, + 274 + ], + "type": "text", + "content": "\" part in the template as labeling part of the demonstration." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 286, + 521, + 314 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 286, + 521, + 314 + ], + "spans": [ + { + "bbox": [ + 302, + 286, + 521, + 314 + ], + "type": "text", + "content": "3 Demonstration-based Learning from a Causal Perspective" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 323, + 526, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 323, + 526, + 444 + ], + "spans": [ + { + "bbox": [ + 302, + 323, + 526, + 444 + ], + "type": "text", + "content": "In this section, we give a specific example to show the potential bias and understand demonstration-based learning from a causal perspective. Specifically, we first introduce a Structural Causal Model (SCM) (Pearl et al., 2000) to describe the mechanism and identify the induced bias. Then, we perform demonstration variable intervention and propose multiple simple and effective random demonstration templates inspired by our causal model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 446, + 526, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 446, + 526, + 703 + ], + "spans": [ + { + "bbox": [ + 302, + 446, + 526, + 703 + ], + "type": "text", + "content": "We observe that the frequent co-occurrence of tokens in the classical demonstrations generate harmful superficial patterns which is misleading to the model and leads to biased predictions (Zhang et al., 2022a; Min et al., 2022). A specific example with different demonstrations is provided in Table 1, where the entity to predict is French. Following previous work (Zhang et al., 2022a), the observed demonstrations (i.e., standard demonstration) provides some biased information: the concurrency of British and ICAC, which is an organization (ORG), may lead to biased predictions: French is labeled as an Organization while its desired prediction is other classes (MISC). Intuitively, the co-occurrence of two specific words in the demonstration may induce bias, while randomly sampled tokens in the demonstration do not. This specific example suggests why random demonstrations may sometimes perform better than standard ones." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 714, + 396, + 725 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 714, + 396, + 725 + ], + "spans": [ + { + "bbox": [ + 302, + 714, + 396, + 725 + ], + "type": "text", + "content": "3.1 Causal Model" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 733, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 733, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 733, + 525, + 772 + ], + "type": "text", + "content": "To study the causal relationship between the NER model and its training data, and explain the role of the demonstration, we introduce a SCM to describe" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "text", + "content": "1466" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 72, + 174, + 151 + ], + "blocks": [ + { + "bbox": [ + 69, + 72, + 174, + 151 + ], + "lines": [ + { + "bbox": [ + 69, + 72, + 174, + 151 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 174, + 151 + ], + "type": "image", + "image_path": "c02e510f97f02f010e978031e4d8f0b374c424d313250b464c45bd264c6682c3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 99, + 161, + 113, + 173 + ], + "lines": [ + { + "bbox": [ + 99, + 161, + 113, + 173 + ], + "spans": [ + { + "bbox": [ + 99, + 161, + 113, + 173 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 179, + 73, + 283, + 150 + ], + "blocks": [ + { + "bbox": [ + 179, + 73, + 283, + 150 + ], + "lines": [ + { + "bbox": [ + 179, + 73, + 283, + 150 + ], + "spans": [ + { + "bbox": [ + 179, + 73, + 283, + 150 + ], + "type": "image", + "image_path": "987321a8336f314009714664bfe25b3a1faaccc13230764ececde8df55dbed24.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 161, + 224, + 174 + ], + "lines": [ + { + "bbox": [ + 211, + 161, + 224, + 174 + ], + "spans": [ + { + "bbox": [ + 211, + 161, + 224, + 174 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 287, + 73, + 393, + 150 + ], + "blocks": [ + { + "bbox": [ + 287, + 73, + 393, + 150 + ], + "lines": [ + { + "bbox": [ + 287, + 73, + 393, + 150 + ], + "spans": [ + { + "bbox": [ + 287, + 73, + 393, + 150 + ], + "type": "image", + "image_path": "b90ea756a254755f329ba8c508bd923ee037cbaa91498920bb1a1dd606c0ce96.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 161, + 333, + 174 + ], + "lines": [ + { + "bbox": [ + 320, + 161, + 333, + 174 + ], + "spans": [ + { + "bbox": [ + 320, + 161, + 333, + 174 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 399, + 69, + 524, + 160 + ], + "blocks": [ + { + "bbox": [ + 399, + 69, + 524, + 160 + ], + "lines": [ + { + "bbox": [ + 399, + 69, + 524, + 160 + ], + "spans": [ + { + "bbox": [ + 399, + 69, + 524, + 160 + ], + "type": "image", + "image_path": "d43a2f5e031397c86a55695cbf9634643370921a08ffb03291bd3f22ee9f8765.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 444, + 162, + 456, + 174 + ], + "lines": [ + { + "bbox": [ + 444, + 162, + 456, + 174 + ], + "spans": [ + { + "bbox": [ + 444, + 162, + 456, + 174 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 181, + 526, + 231 + ], + "lines": [ + { + "bbox": [ + 67, + 181, + 526, + 231 + ], + "spans": [ + { + "bbox": [ + 67, + 181, + 526, + 231 + ], + "type": "text", + "content": "Figure 1: Causal views of NER. (a) shows a traditional NER model (Zeng et al., 2020), (b) shows the demonstration-based NER model under the causal view. With demonstration " + }, + { + "bbox": [ + 67, + 181, + 526, + 231 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 67, + 181, + 526, + 231 + ], + "type": "text", + "content": ", the backdoor path " + }, + { + "bbox": [ + 67, + 181, + 526, + 231 + ], + "type": "inline_equation", + "content": "G \\rightarrow D \\rightarrow X" + }, + { + "bbox": [ + 67, + 181, + 526, + 231 + ], + "type": "text", + "content": " exists, which further introduces the bias. (c) shows the demonstration-based NER model with debiasing techniques, and the red cross means intervention. (d) is model architecture overview between classical and demonstration-based learning." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 240, + 291, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 240, + 291, + 509 + ], + "spans": [ + { + "bbox": [ + 67, + 240, + 291, + 509 + ], + "type": "text", + "content": "the inference step in NER models. Figure 1 shows the SCM of NER models. There are mainly 6 variables in NER models: 1) Demonstration Tokens " + }, + { + "bbox": [ + 67, + 240, + 291, + 509 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 67, + 240, + 291, + 509 + ], + "type": "text", + "content": ", the tokens which form the demonstration; 2) Context Tokens " + }, + { + "bbox": [ + 67, + 240, + 291, + 509 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 67, + 240, + 291, + 509 + ], + "type": "text", + "content": ", the tokens that are related to the context; 3) Entity Tokens " + }, + { + "bbox": [ + 67, + 240, + 291, + 509 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 67, + 240, + 291, + 509 + ], + "type": "text", + "content": ", the tokens which are entities; 4) Input Example " + }, + { + "bbox": [ + 67, + 240, + 291, + 509 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 67, + 240, + 291, + 509 + ], + "type": "text", + "content": ", which is composed of " + }, + { + "bbox": [ + 67, + 240, + 291, + 509 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 67, + 240, + 291, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 240, + 291, + 509 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 67, + 240, + 291, + 509 + ], + "type": "text", + "content": " in the traditional model and composed of " + }, + { + "bbox": [ + 67, + 240, + 291, + 509 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 67, + 240, + 291, + 509 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 67, + 240, + 291, + 509 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 67, + 240, + 291, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 240, + 291, + 509 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 67, + 240, + 291, + 509 + ], + "type": "text", + "content": " in the demonstration-based models; 5) Unobserved confounders " + }, + { + "bbox": [ + 67, + 240, + 291, + 509 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 67, + 240, + 291, + 509 + ], + "type": "text", + "content": ", a confounding variable (not a concrete token) that influences the generation of " + }, + { + "bbox": [ + 67, + 240, + 291, + 509 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 67, + 240, + 291, + 509 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 67, + 240, + 291, + 509 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 67, + 240, + 291, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 240, + 291, + 509 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 67, + 240, + 291, + 509 + ], + "type": "text", + "content": "; 6) Evaluation result " + }, + { + "bbox": [ + 67, + 240, + 291, + 509 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 67, + 240, + 291, + 509 + ], + "type": "text", + "content": ", the evaluation result (the F1 score) of the NER models. Under the causal view, the key difference between the traditional NER model and the demonstration-based NER model is that, the demonstration-based NER model has an additional node " + }, + { + "bbox": [ + 67, + 240, + 291, + 509 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 67, + 240, + 291, + 509 + ], + "type": "text", + "content": ". With the introduction of the demonstration " + }, + { + "bbox": [ + 67, + 240, + 291, + 509 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 67, + 240, + 291, + 509 + ], + "type": "text", + "content": ", a backdoor path " + }, + { + "bbox": [ + 67, + 240, + 291, + 509 + ], + "type": "inline_equation", + "content": "G \\rightarrow D \\rightarrow X" + }, + { + "bbox": [ + 67, + 240, + 291, + 509 + ], + "type": "text", + "content": " exists, which further introduces the bias." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 512, + 291, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 512, + 291, + 673 + ], + "spans": [ + { + "bbox": [ + 67, + 512, + 291, + 673 + ], + "type": "text", + "content": "Inspired by our SCM model (Figure 1b), we develop sampling techniques to generate new counterfactual examples by the interventions on the existing observational examples to alleviate this bias. The benefits of interventions on " + }, + { + "bbox": [ + 67, + 512, + 291, + 673 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 67, + 512, + 291, + 673 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 512, + 291, + 673 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 67, + 512, + 291, + 673 + ], + "type": "text", + "content": " have been studied in (Zeng et al., 2020). In this paper, we focus on understanding the role of demonstrations in NER models under the causal view. We understand the co-occurrence of tokens and harmful superficial patterns from the causal perspective and focus on using interventions on the demonstration variable to create new counterfactual demonstrations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 687, + 274, + 698 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 687, + 274, + 698 + ], + "spans": [ + { + "bbox": [ + 67, + 687, + 274, + 698 + ], + "type": "text", + "content": "3.2 Controllable Random Demonstrations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 67, + 706, + 292, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 706, + 292, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 706, + 292, + 772 + ], + "type": "text", + "content": "In this section, we first provide a running example to better understand the induced bias from human-crafted demonstrations and then present different ways of intervention on the demonstration tokens. The intervention is implemented via control" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 240, + 526, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 240, + 526, + 307 + ], + "spans": [ + { + "bbox": [ + 302, + 240, + 526, + 307 + ], + "type": "text", + "content": "lable random demonstrations to create new counterfactual examples, as replacing standard demonstrations with random tokens can remove induce bias and still make the model a good few-shot learner (Zhang et al., 2022a)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 309, + 526, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 309, + 526, + 459 + ], + "spans": [ + { + "bbox": [ + 302, + 309, + 526, + 459 + ], + "type": "text", + "content": "In Lee et al. (2022), an effective template " + }, + { + "bbox": [ + 302, + 309, + 526, + 459 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 302, + 309, + 526, + 459 + ], + "type": "text", + "content": " is \"" + }, + { + "bbox": [ + 302, + 309, + 526, + 459 + ], + "type": "inline_equation", + "content": "s^{(c)}" + }, + { + "bbox": [ + 302, + 309, + 526, + 459 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 302, + 309, + 526, + 459 + ], + "type": "inline_equation", + "content": "e^{(c)}" + }, + { + "bbox": [ + 302, + 309, + 526, + 459 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 302, + 309, + 526, + 459 + ], + "type": "inline_equation", + "content": "t^{(c)}" + }, + { + "bbox": [ + 302, + 309, + 526, + 459 + ], + "type": "text", + "content": ", and an example demonstration " + }, + { + "bbox": [ + 302, + 309, + 526, + 459 + ], + "type": "inline_equation", + "content": "d^{(c)}" + }, + { + "bbox": [ + 302, + 309, + 526, + 459 + ], + "type": "text", + "content": " can be \"[SEP] Obama returns to White House. Obama is PER.\" Intuitively, the model understands the demonstrations and then better performs inference. However, random demonstrations can still bring performance improvement (Zhang et al., 2022a). The random template is as simple as " + }, + { + "bbox": [ + 302, + 309, + 526, + 459 + ], + "type": "inline_equation", + "content": "[s_i]_{i=1}^L" + }, + { + "bbox": [ + 302, + 309, + 526, + 459 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 302, + 309, + 526, + 459 + ], + "type": "inline_equation", + "content": "s_i \\in p" + }, + { + "bbox": [ + 302, + 309, + 526, + 459 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 302, + 309, + 526, + 459 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 302, + 309, + 526, + 459 + ], + "type": "text", + "content": " is a token distribution. Random demonstrations are composed of " + }, + { + "bbox": [ + 302, + 309, + 526, + 459 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 302, + 309, + 526, + 459 + ], + "type": "text", + "content": " tokens randomly sampled from " + }, + { + "bbox": [ + 302, + 309, + 526, + 459 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 302, + 309, + 526, + 459 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 460, + 525, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 460, + 525, + 635 + ], + "spans": [ + { + "bbox": [ + 302, + 460, + 525, + 635 + ], + "type": "text", + "content": "Demonstration Intervention We use the intervention on the demonstration tokens to create new counterfactual examples, to alleviate the biases. If we do not carefully design D, the backdoor path will exist and the model performance is degraded. Our causal framework enables us to think about the problem from a causal perspective and guides us how to properly design D. We denote uniform distribution composed of vocabulary words of the PLMs as " + }, + { + "bbox": [ + 302, + 460, + 525, + 635 + ], + "type": "inline_equation", + "content": "p_{\\mathcal{V}}" + }, + { + "bbox": [ + 302, + 460, + 525, + 635 + ], + "type": "text", + "content": ". Given the token distribution " + }, + { + "bbox": [ + 302, + 460, + 525, + 635 + ], + "type": "inline_equation", + "content": "p_{\\mathcal{V}}" + }, + { + "bbox": [ + 302, + 460, + 525, + 635 + ], + "type": "text", + "content": ", for any word " + }, + { + "bbox": [ + 302, + 460, + 525, + 635 + ], + "type": "inline_equation", + "content": "w_i \\in p_{\\mathcal{V}}" + }, + { + "bbox": [ + 302, + 460, + 525, + 635 + ], + "type": "text", + "content": ", we have " + }, + { + "bbox": [ + 302, + 460, + 525, + 635 + ], + "type": "inline_equation", + "content": "p_{\\mathcal{V}}(w_i) = \\frac{1}{|\\mathcal{V}|}" + }, + { + "bbox": [ + 302, + 460, + 525, + 635 + ], + "type": "text", + "content": ". Then we have a plain way to construct random demonstrations." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 638, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 638, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 638, + 526, + 772 + ], + "type": "text", + "content": "An important observation is that not all counterfactual examples are correct or useful. Hence, the intervention can be better implemented by replacing the uniform distribution with a non-uniform distribution, i.e., by adding or removing words and changing specific words' probabilities. Some mechanism is needed to identify good counterfactual demonstrations, to avoid introducing noise. An intuitive solution is that we consider tokens from the support set are more helpful as PLMs are fine" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1467" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 71, + 68, + 520, + 190 + ], + "blocks": [ + { + "bbox": [ + 71, + 68, + 520, + 190 + ], + "lines": [ + { + "bbox": [ + 71, + 68, + 520, + 190 + ], + "spans": [ + { + "bbox": [ + 71, + 68, + 520, + 190 + ], + "type": "table", + "html": "
ModeNERChunking
CoNLL03OntoNotes 5.0CoNLL00
F1PrecisionRecallF1PrecisionRecallF1PrecisionRecall
No Demo.28.71±10.3139.96±11.2522.68±9.0937.37±7.5833.80±6.7941.92±8.8563.17±4.2259.28±5.0567.72±3.51
Standard45.86±6.0847.38±5.9344.75±7.0740.21±7.6532.51±6.8752.82±8.2870.55±3.0866.53±4.4075.21±2.11
Random41.33±7.3645.41±7.3738.22±7.6539.71±7.5632.28±6.5651.63±8.7569.28±2.7864.75±3.8574.57±1.66
Rand-S45.55±8.0246.84±7.7144.60±8.6241.60±7.0533.96±6.2953.75±7.8070.63±3.0166.24±4.2975.75±1.70
Rand-W45.93±7.5747.79±7.4244.50±8.1345.49±3.7737.82±3.6457.18±4.1772.15±3.1668.00±4.4276.94±1.67
Rand-E47.32±7.4248.96±7.0246.02±8.1146.06±3.8438.32±3.6557.81±4.3174.02±2.9370.37±4.2378.18±1.75
", + "image_path": "3e10791f930af0474d8a42b65d96ba922faabadbafba002a9eacef57294219d7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 195, + 525, + 219 + ], + "lines": [ + { + "bbox": [ + 67, + 195, + 525, + 219 + ], + "spans": [ + { + "bbox": [ + 67, + 195, + 525, + 219 + ], + "type": "text", + "content": "Table 2: Main results for traditional token classification method (No Demo.) and demonstration-based learning with different modes of demonstrations under 5-shot scenario." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 227, + 290, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 227, + 290, + 280 + ], + "spans": [ + { + "bbox": [ + 67, + 227, + 290, + 280 + ], + "type": "text", + "content": "tuned on the support set. We expect to see a better downstream predictor when the demonstrations are constructed randomly from a intervened token distribution." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 281, + 290, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 281, + 290, + 428 + ], + "spans": [ + { + "bbox": [ + 67, + 281, + 290, + 428 + ], + "type": "text", + "content": "The difference between random demonstrations lies in the vocabulary and its associated probability distributions. We perform the interventions by controlling the vocabulary and changing the probability of random tokens. We encourage entity words (e.g., ICAC, British) to appear more frequently compared to the others (e.g., is). Based on the previous theoretical justification, we consider the following variants of constructing random demonstrations2 construction methods as counterfactual alternatives of the standard demonstrations3:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 430, + 289, + 592 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 67, + 430, + 289, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 430, + 289, + 456 + ], + "spans": [ + { + "bbox": [ + 67, + 430, + 289, + 456 + ], + "type": "text", + "content": "- Random: random context with tokens uniformly sampled from PLMs vocabulary " + }, + { + "bbox": [ + 67, + 430, + 289, + 456 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 67, + 430, + 289, + 456 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 457, + 289, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 457, + 289, + 496 + ], + "spans": [ + { + "bbox": [ + 67, + 457, + 289, + 496 + ], + "type": "text", + "content": "- Rand-S: random context with tokens uniformly sampled from unique words (i.e., vocabulary) of support set, denoted as " + }, + { + "bbox": [ + 67, + 457, + 289, + 496 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 67, + 457, + 289, + 496 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 497, + 289, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 497, + 289, + 550 + ], + "spans": [ + { + "bbox": [ + 67, + 497, + 289, + 550 + ], + "type": "text", + "content": "- Rand-W4: random context with tokens sampled from " + }, + { + "bbox": [ + 67, + 497, + 289, + 550 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 67, + 497, + 289, + 550 + ], + "type": "text", + "content": ", and entity tokens in support set, denoted as " + }, + { + "bbox": [ + 67, + 497, + 289, + 550 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 67, + 497, + 289, + 550 + ], + "type": "text", + "content": "; tokens from " + }, + { + "bbox": [ + 67, + 497, + 289, + 550 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 67, + 497, + 289, + 550 + ], + "type": "text", + "content": " have four times higher probability compared with those from " + }, + { + "bbox": [ + 67, + 497, + 289, + 550 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 67, + 497, + 289, + 550 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 552, + 289, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 552, + 289, + 592 + ], + "spans": [ + { + "bbox": [ + 67, + 552, + 289, + 592 + ], + "type": "text", + "content": "- Rand-E: similar to Rand-W, but replace entity tokens with entities composed of coherent tokens in support set, denoted as " + }, + { + "bbox": [ + 67, + 552, + 289, + 592 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 67, + 552, + 289, + 592 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 67, + 602, + 199, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 602, + 199, + 616 + ], + "spans": [ + { + "bbox": [ + 67, + 602, + 199, + 616 + ], + "type": "text", + "content": "4 Experimental Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 624, + 180, + 638 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 624, + 180, + 638 + ], + "spans": [ + { + "bbox": [ + 67, + 624, + 180, + 638 + ], + "type": "text", + "content": "4.1 Experiment Setup" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 641, + 290, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 641, + 290, + 723 + ], + "spans": [ + { + "bbox": [ + 67, + 641, + 290, + 723 + ], + "type": "text", + "content": "Datasets We conduct experiments on two sequence labeling tasks: (i) named entity recognition (NER) on dataset CoNLL03 (Tjong Kim Sang and De Meulder, 2003), and OntoNotes 5.0 (Weischedel et al., 2013); and (ii) chunking on dataset CoNLL00 (Tjong Kim Sang and Buchholz," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 227, + 526, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 227, + 526, + 389 + ], + "spans": [ + { + "bbox": [ + 302, + 227, + 526, + 389 + ], + "type": "text", + "content": "2000). Following previous works Ma et al. (2021); Zhang et al. (2022a), we omit the 7 value types in OntoNotes and only consider the 6 most frequent types in CoNLL00. For few-shot data sampling, we follow the greedy sampling strategy proposed by Yang and Katiyar (2020) to sample " + }, + { + "bbox": [ + 302, + 227, + 526, + 389 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 302, + 227, + 526, + 389 + ], + "type": "text", + "content": " shots for each type in an increasing order with respect to their frequencies, the detailed algorithm can be found. For each dataset, we sample 5 different " + }, + { + "bbox": [ + 302, + 227, + 526, + 389 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 302, + 227, + 526, + 389 + ], + "type": "text", + "content": "-shot support sets and report mean and standard deviation of metrics. For each " + }, + { + "bbox": [ + 302, + 227, + 526, + 389 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 302, + 227, + 526, + 389 + ], + "type": "text", + "content": "-shot support set, we run the experiments with 3 random seeds." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 302, + 394, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 394, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 394, + 525, + 772 + ], + "type": "text", + "content": "Main Results We show the results for demonstration-based learning with different modes of demonstrations as well as classical sequence labeling with no demonstration in Table 2. The results show that demonstration-based method can consistently improve model performance. In demonstration-based methods, the Random approach shows the worst performance and Rand-S shows comparable results with the standard demonstrations, and the conclusion is consistent with previous works (Zhang et al., 2022a). Interestingly, if we modify the token sampling distributions and sample more entity or entity-related words as Rand-W and Rand-E, our model shows even better performance than standard meaningful demonstrations. The difference between Rand-W and Rand-E lies in whether there are complete entities, and the results show that adding complete entities instead of random entity words can lead to better performance. At the same time, it shows adding random tokens related to the support set can reduce the fine-tuned bias, which verifies our hypothesis in Section 3.1. Intuitively, the benefits of demonstration-based methods come from tokens of support sets " + }, + { + "bbox": [ + 302, + 394, + 525, + 772 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 302, + 394, + 525, + 772 + ], + "type": "text", + "content": " instead of meaningful demonstrations, as the standard demonstration sampled from the support set also shows good performance." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 80, + 729, + 222, + 740 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 729, + 222, + 740 + ], + "spans": [ + { + "bbox": [ + 80, + 729, + 222, + 740 + ], + "type": "text", + "content": "2Random: [SEP] {random context}" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 80, + 740, + 274, + 751 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 740, + 274, + 751 + ], + "spans": [ + { + "bbox": [ + 80, + 740, + 274, + 751 + ], + "type": "text", + "content": "3Standard: [SEP] {context} {entity} is {tag}." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 69, + 751, + 289, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 751, + 289, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 751, + 289, + 772 + ], + "type": "text", + "content": "4 Empirical results show sampling only from " + }, + { + "bbox": [ + 69, + 751, + 289, + 772 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 69, + 751, + 289, + 772 + ], + "type": "text", + "content": " leads to poor performance." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1468" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 68, + 225, + 185 + ], + "blocks": [ + { + "bbox": [ + 70, + 68, + 225, + 185 + ], + "lines": [ + { + "bbox": [ + 70, + 68, + 225, + 185 + ], + "spans": [ + { + "bbox": [ + 70, + 68, + 225, + 185 + ], + "type": "image", + "image_path": "c081e6404c2d14411e0af537694d5665345a88764e5bbd0c6639d760b30d3098.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 120, + 194, + 472, + 206 + ], + "lines": [ + { + "bbox": [ + 120, + 194, + 472, + 206 + ], + "spans": [ + { + "bbox": [ + 120, + 194, + 472, + 206 + ], + "type": "text", + "content": "Figure 2: Results with different support set size on CoNLL03, NRB and WTS datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 230, + 68, + 381, + 186 + ], + "blocks": [ + { + "bbox": [ + 230, + 68, + 381, + 186 + ], + "lines": [ + { + "bbox": [ + 230, + 68, + 381, + 186 + ], + "spans": [ + { + "bbox": [ + 230, + 68, + 381, + 186 + ], + "type": "image", + "image_path": "47e737b77f0e5157053023e79ebe490557513eb838b5aff71480e50c7ecc9650.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 381, + 68, + 524, + 186 + ], + "blocks": [ + { + "bbox": [ + 381, + 68, + 524, + 186 + ], + "lines": [ + { + "bbox": [ + 381, + 68, + 524, + 186 + ], + "spans": [ + { + "bbox": [ + 381, + 68, + 524, + 186 + ], + "type": "image", + "image_path": "b5bc4c4268bb8d2f235c89ba405172e60206d99cb1707f253ba822159311a774.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 73, + 219, + 282, + 304 + ], + "blocks": [ + { + "bbox": [ + 73, + 219, + 282, + 304 + ], + "lines": [ + { + "bbox": [ + 73, + 219, + 282, + 304 + ], + "spans": [ + { + "bbox": [ + 73, + 219, + 282, + 304 + ], + "type": "table", + "html": "
ModeCoNLL03OntoNotes5.0CoNLL00
No Demo.45.70±8.1351.62±2.7672.80±3.53
Standard45.73±7.2954.76±2.3675.90±1.95
Rand-S46.86±6.5054.35±2.6772.23±3.42
Rand-W52.11±6.1554.48±2.3573.84±2.19
Rand-E52.87±7.6455.94±2.3875.30±3.06
", + "image_path": "d3e28a47a040cbb0835e15b9fce8e7ef3f14b440fc930ded59b0ce350481d560.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 309, + 290, + 344 + ], + "lines": [ + { + "bbox": [ + 67, + 309, + 290, + 344 + ], + "spans": [ + { + "bbox": [ + 67, + 309, + 290, + 344 + ], + "type": "text", + "content": "Table 3: Main results (F1 scores) of RoBERTa-Large for traditional token classification with different modes of demonstrations under 5-shot scenario." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 358, + 135, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 358, + 135, + 370 + ], + "spans": [ + { + "bbox": [ + 67, + 358, + 135, + 370 + ], + "type": "text", + "content": "4.2 Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 378, + 291, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 378, + 291, + 593 + ], + "spans": [ + { + "bbox": [ + 67, + 378, + 291, + 593 + ], + "type": "text", + "content": "Ablation Studies We further investigate whether the performance gain of demonstration-based learning changes over the size of support set. We present results of different modes of demonstrations under " + }, + { + "bbox": [ + 67, + 378, + 291, + 593 + ], + "type": "inline_equation", + "content": "K = 5, 10, 20" + }, + { + "bbox": [ + 67, + 378, + 291, + 593 + ], + "type": "text", + "content": " shots in Figure 2. With more training examples in the support set, the relative performance gap between Rand-E and Standard remains, but it becomes smaller. This indicates that carefully designed random demonstrations show a consistent performance improvement upon standard demonstration. We also observe that the variance within each group becomes smaller as more data becomes available. Among random demonstrations, Rand-E consistently shows better performance than Rand-W and Rand-S, which verifies our hypothesis based on the SCM." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 596, + 291, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 596, + 291, + 676 + ], + "spans": [ + { + "bbox": [ + 67, + 596, + 291, + 676 + ], + "type": "text", + "content": "Additionally, we investigate the effect of using different base models and replace BERT with RoBERTa. The observed results for RoBERTa in Table 3 are consistent with those of BERT, demonstrating that Rand-E exhibits superior performance across different model architectures." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 678, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 678, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 678, + 291, + 772 + ], + "type": "text", + "content": "Name Regularity Bias Name Regularity Bias (Ghaddar et al., 2021; Lin et al., 2020) in NER occurs when a model relies on a signal from the entity name to make predictions and disregards evidence from the local context. Ghaddar et al. (2021) carefully designed a testbed utilizing Wikipedia disambiguation pages to diagnose the Name Regu" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 216, + 524, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 216, + 524, + 243 + ], + "spans": [ + { + "bbox": [ + 302, + 216, + 524, + 243 + ], + "type": "text", + "content": "larity Bias of NER models. Details about the NRB dataset are provided in the appendix." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 244, + 525, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 244, + 525, + 339 + ], + "spans": [ + { + "bbox": [ + 302, + 244, + 525, + 339 + ], + "type": "text", + "content": "We use both the NRB and WTS (as control sets) datasets to evaluate the model trained with different modes of demonstrations on CoNLL03. The results show a smaller gap for random demonstrations, suggesting that random demonstration-based learning can better leverage context information instead of the name regularity patterns." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 303, + 350, + 386, + 364 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 350, + 386, + 364 + ], + "spans": [ + { + "bbox": [ + 303, + 350, + 386, + 364 + ], + "type": "text", + "content": "5 Conclusions" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 374, + 526, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 374, + 526, + 522 + ], + "spans": [ + { + "bbox": [ + 302, + 374, + 526, + 522 + ], + "type": "text", + "content": "In this paper, we present a casual view to understand demonstration-based learning. Based on the structural causal model we constructed, we investigate the causal effects and discover that the concurrence of specific words in the demonstration can induce bias. To address this issue, we perform interventions by constructing random demonstrations. Our empirical results indicate that carefully designed random demonstrations consistently outperform meaningful demonstrations on public sequence labeling benchmarks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 534, + 384, + 546 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 534, + 384, + 546 + ], + "spans": [ + { + "bbox": [ + 302, + 534, + 384, + 546 + ], + "type": "text", + "content": "6 Limitations" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 556, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 556, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 556, + 526, + 772 + ], + "type": "text", + "content": "All our experiments are done on the sequence labeling task, and they can be further evaluated on sentence classification tasks with classifier-based fine-tuning since the [CLS] token used for classification represents the whole sentence. We provide a causal opinion on demonstration-based learning and a simple but not systematic method to alleviate the induced bias. Our demonstration-based learning builds upon previous works (Lee et al., 2022; Zhang et al., 2022a), where BERT or RoBERTa are used instead of Large Language Models, such as InstructGPT (Ouyang et al., 2022), PaLM (Chowdhery et al., 2022), and OPT (Zhang et al., 2022b). Furthermore, our conclusions are drawn from few-shot learning settings and cannot be directly applied to zero-shot inference." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1469" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 71, + 127, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 71, + 127, + 83 + ], + "spans": [ + { + "bbox": [ + 69, + 71, + 127, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 90, + 289, + 772 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 69, + 90, + 289, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 90, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 69, + 90, + 289, + 145 + ], + "type": "text", + "content": "Rohan Anil, Andrew M Dai, Orhan Firat, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, et al. 2023. Palm 2 technical report. arXiv preprint arXiv:2305.10403." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 155, + 289, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 155, + 289, + 318 + ], + "spans": [ + { + "bbox": [ + 69, + 155, + 289, + 318 + ], + "type": "text", + "content": "Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 329, + 289, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 329, + 289, + 394 + ], + "spans": [ + { + "bbox": [ + 69, + 329, + 289, + 394 + ], + "type": "text", + "content": "Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. 2022. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 404, + 289, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 404, + 289, + 448 + ], + "spans": [ + { + "bbox": [ + 69, + 404, + 289, + 448 + ], + "type": "text", + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 458, + 289, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 458, + 289, + 556 + ], + "spans": [ + { + "bbox": [ + 69, + 458, + 289, + 556 + ], + "type": "text", + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 566, + 289, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 566, + 289, + 654 + ], + "spans": [ + { + "bbox": [ + 69, + 566, + 289, + 654 + ], + "type": "text", + "content": "Tianyu Gao, Adam Fisch, and Danqi Chen. 2021. Making pre-trained language models better few-shot learners. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 3816-3830, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 663, + 289, + 718 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 663, + 289, + 718 + ], + "spans": [ + { + "bbox": [ + 69, + 663, + 289, + 718 + ], + "type": "text", + "content": "Abbas Ghaddar, Philippe Langlais, Ahmad Rashid, and Mehdi Rezagholizadeh. 2021. Context-aware Adversarial Training for Name Regularity Bias in Named Entity Recognition. Transactions of the Association for Computational Linguistics, 9:586-604." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 728, + 289, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 728, + 289, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 728, + 289, + 772 + ], + "type": "text", + "content": "Jiaxin Huang, Chunyuan Li, Krishan Subudhi, Damien Jose, Shobana Balakrishnan, Weizhu Chen, Baolin Peng, Jianfeng Gao, and Jiawei Han. 2021. Few-shot named entity recognition: An empirical baseline" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 524, + 772 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 314, + 72, + 524, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 524, + 126 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 524, + 126 + ], + "type": "text", + "content": "study. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 10408-10423, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 137, + 524, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 137, + 524, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 137, + 524, + 179 + ], + "type": "text", + "content": "Brenden Lake, Ruslan Salakhutdinov, and Joshua Tenenbaum. 2015. Human-level concept learning through probabilistic program induction. Science, 350:1332-1338." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 190, + 524, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 190, + 524, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 190, + 524, + 289 + ], + "type": "text", + "content": "Dong-Ho Lee, Akshen Kadakia, Kangmin Tan, Mahak Agarwal, Xinyu Feng, Takashi Shibuya, Ryosuke Mitani, Toshiyuki Sekiya, Jay Pujara, and Xiang Ren. 2022. Good examples make a faster learner: Simple demonstration-based learning for low-resource NER. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2687-2700, Dublin, Ireland. Association for Computational Linguistics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 298, + 524, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 298, + 524, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 298, + 524, + 396 + ], + "type": "text", + "content": "Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871-7880, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 406, + 524, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 406, + 524, + 494 + ], + "spans": [ + { + "bbox": [ + 304, + 406, + 524, + 494 + ], + "type": "text", + "content": "Hongyu Lin, Yaojie Lu, Jialong Tang, Xianpei Han, Le Sun, Zhicheng Wei, and Nicholas Jing Yuan. 2020. A rigorous study on named entity recognition: Can fine-tuning pretrained model lead to the promised land? In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7291-7300, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 503, + 524, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 503, + 524, + 558 + ], + "spans": [ + { + "bbox": [ + 304, + 503, + 524, + 558 + ], + "type": "text", + "content": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019a. Roberta: A robustly optimized BERT pretraining approach. CoRR, abs/1907.11692." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 567, + 524, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 524, + 623 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 524, + 623 + ], + "type": "text", + "content": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019b. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 632, + 524, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 632, + 524, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 632, + 524, + 665 + ], + "type": "text", + "content": "Ruotian Ma, Xin Zhou, Tao Gui, Yiding Tan, Qi Zhang, and Xuanjing Huang. 2021. Template-free prompt tuning for few-shot NER. CoRR, abs/2109.13532." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 674, + 524, + 729 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 674, + 524, + 729 + ], + "spans": [ + { + "bbox": [ + 304, + 674, + 524, + 729 + ], + "type": "text", + "content": "Sewon Min, Xinxi Lyu, Ari Holtzman, Mikel Artetxe, Mike Lewis, Hannaneh Hajishirzi, and Luke Zettlemoyer. 2022. Rethinking the role of demonstrations: What makes in-context learning work? arXiv preprint arXiv:2202.12837." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 739, + 524, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 739, + 524, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 739, + 524, + 772 + ], + "type": "text", + "content": "Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al." + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1470" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 751 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 80, + 72, + 291, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 72, + 291, + 95 + ], + "spans": [ + { + "bbox": [ + 80, + 72, + 291, + 95 + ], + "type": "text", + "content": "2022. Training language models to follow instructions with human feedback. NeurIPS." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 103, + 291, + 136 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 103, + 291, + 136 + ], + "spans": [ + { + "bbox": [ + 69, + 103, + 291, + 136 + ], + "type": "text", + "content": "Judea Pearl et al. 2000. Models, reasoning and inference. Cambridge, UK: Cambridge University Press, 19(2)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 145, + 290, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 145, + 290, + 179 + ], + "spans": [ + { + "bbox": [ + 69, + 145, + 290, + 179 + ], + "type": "text", + "content": "Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. 2023. Instruction tuning with gpt-4. arXiv preprint arXiv:2304.03277." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 186, + 290, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 186, + 290, + 242 + ], + "spans": [ + { + "bbox": [ + 69, + 186, + 290, + 242 + ], + "type": "text", + "content": "Erik F. Tjong Kim Sang and Sabine Buchholz. 2000. Introduction to the CoNLL-2000 shared task chunking. In Fourth Conference on Computational Natural Language Learning and the Second Learning Language in Logic Workshop." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 250, + 290, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 250, + 290, + 316 + ], + "spans": [ + { + "bbox": [ + 69, + 250, + 290, + 316 + ], + "type": "text", + "content": "Erik F. Tjong Kim Sang and Fien De Meulder. 2003. Introduction to the CoNLL-2003 shared task: Language-independent named entity recognition. In Proceedings of the Seventh Conference on Natural Language Learning at HLT-NAACL 2003, pages 142-147." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 325, + 290, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 325, + 290, + 390 + ], + "spans": [ + { + "bbox": [ + 69, + 325, + 290, + 390 + ], + "type": "text", + "content": "Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 400, + 290, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 400, + 290, + 454 + ], + "spans": [ + { + "bbox": [ + 69, + 400, + 290, + 454 + ], + "type": "text", + "content": "Ralph Weischedel, Martha Palmer, Mitchell Marcus, Edward Hovy, Sameer Pradhan, Lance Ramshaw, Nianwen Xue, Ann Taylor, Jeff Kaufman, Michelle Franchini, et al. 2013. Ontonotes release 5.0 ldc2013t19. Linguistic Data Consortium, Philadelphia, PA, 23." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 463, + 290, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 463, + 290, + 518 + ], + "spans": [ + { + "bbox": [ + 69, + 463, + 290, + 518 + ], + "type": "text", + "content": "Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. 2020. Unsupervised data augmentation for consistency training. In Advances in Neural Information Processing Systems, volume 33, pages 6256-6268. Curran Associates, Inc." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 527, + 290, + 593 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 527, + 290, + 593 + ], + "spans": [ + { + "bbox": [ + 69, + 527, + 290, + 593 + ], + "type": "text", + "content": "Yi Yang and Arzoo Katiyar. 2020. Simple and effective few-shot named entity recognition with structured nearest neighbor learning. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6365-6375, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 602, + 290, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 602, + 290, + 645 + ], + "spans": [ + { + "bbox": [ + 69, + 602, + 290, + 645 + ], + "type": "text", + "content": "Xiangji Zeng, Yunliang Li, Yuchen Zhai, and Yin Zhang. 2020. Counterfactual generator: A weakly-supervised method for named entity recognition. In EMNLP." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 655, + 290, + 687 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 655, + 290, + 687 + ], + "spans": [ + { + "bbox": [ + 69, + 655, + 290, + 687 + ], + "type": "text", + "content": "Hongxin Zhang, Yanzhe Zhang, Ruiyi Zhang, and Diyi Yang. 2022a. Robustness of demonstration-based learning under limited data scenario. In EMNLP." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 69, + 696, + 290, + 751 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 696, + 290, + 751 + ], + "spans": [ + { + "bbox": [ + 69, + 696, + 290, + 751 + ], + "type": "text", + "content": "Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. 2022b. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 308, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 308, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 308, + 791 + ], + "type": "text", + "content": "1471" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 100, + 69, + 258, + 137 + ], + "blocks": [ + { + "bbox": [ + 100, + 69, + 258, + 137 + ], + "lines": [ + { + "bbox": [ + 100, + 69, + 258, + 137 + ], + "spans": [ + { + "bbox": [ + 100, + 69, + 258, + 137 + ], + "type": "table", + "html": "
Dataset|Y|L|Dtest|
CoNLL034183453
OntoNotes 5.0112112217
CoNLL006362012
", + "image_path": "40efaad6896ce44d18c6741c18b875940e58d4da5d602aa409436da4c5c42ff7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 146, + 291, + 195 + ], + "lines": [ + { + "bbox": [ + 67, + 146, + 291, + 195 + ], + "spans": [ + { + "bbox": [ + 67, + 146, + 291, + 195 + ], + "type": "text", + "content": "Table 4: Data Statistics. " + }, + { + "bbox": [ + 67, + 146, + 291, + 195 + ], + "type": "inline_equation", + "content": "|Y|" + }, + { + "bbox": [ + 67, + 146, + 291, + 195 + ], + "type": "text", + "content": ": # of entity types. L: average # of tokens in input sentence. " + }, + { + "bbox": [ + 67, + 146, + 291, + 195 + ], + "type": "inline_equation", + "content": "|D_{support}|" + }, + { + "bbox": [ + 67, + 146, + 291, + 195 + ], + "type": "text", + "content": ": average # of sentences in 5-shot support set over 5 different sub-samples. " + }, + { + "bbox": [ + 67, + 146, + 291, + 195 + ], + "type": "inline_equation", + "content": "|D_{test}|" + }, + { + "bbox": [ + 67, + 146, + 291, + 195 + ], + "type": "text", + "content": ": # of sentences in test set." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 68, + 215, + 142, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 215, + 142, + 229 + ], + "spans": [ + { + "bbox": [ + 68, + 215, + 142, + 229 + ], + "type": "text", + "content": "A Appendix" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 236, + 290, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 236, + 290, + 383 + ], + "spans": [ + { + "bbox": [ + 67, + 236, + 290, + 383 + ], + "type": "text", + "content": "NRB Dataset Details The NRB dataset contains examples whose labels can be easily inferred from the local context, but they are difficult to be tagged by a popular NER system. The WTS dataset is a domain control set that includes the same query terms covered by NRB, but these can be correctly labeled by both the popular NER tagger and the local context-only tagger. Therefore, the gap between the NRB and WTS sets measures how effectively the model captures context information to predict token labels." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 386, + 291, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 386, + 291, + 507 + ], + "spans": [ + { + "bbox": [ + 67, + 386, + 291, + 507 + ], + "type": "text", + "content": "Effects of Sampling Probability We present two variants, Random-E[X] and Random-W[X], where X refers to how many times the probability of preferred tokens is higher. In this ablation study, we consistently observe that Random-E4 performs better than Random-E2, and Random-W4 outperforms Random-E4. However, if we increase the X value to a very large number, the performance deteriorates." + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "text", + "content": "1472" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 341, + 218, + 487 + ], + "blocks": [ + { + "bbox": [ + 69, + 341, + 218, + 487 + ], + "lines": [ + { + "bbox": [ + 69, + 341, + 218, + 487 + ], + "spans": [ + { + "bbox": [ + 69, + 341, + 218, + 487 + ], + "type": "image", + "image_path": "4d287e32f49c4496c1ede1687bce3ce1b2488a7c6804de1f0e6fee6e7e19d13b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 494, + 490, + 506 + ], + "lines": [ + { + "bbox": [ + 102, + 494, + 490, + 506 + ], + "spans": [ + { + "bbox": [ + 102, + 494, + 490, + 506 + ], + "type": "text", + "content": "Figure 3: F1, Precision and Recall with more variants of Random Demonstrations on CoNLL03." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 224, + 341, + 371, + 488 + ], + "blocks": [ + { + "bbox": [ + 224, + 341, + 371, + 488 + ], + "lines": [ + { + "bbox": [ + 224, + 341, + 371, + 488 + ], + "spans": [ + { + "bbox": [ + 224, + 341, + 371, + 488 + ], + "type": "image", + "image_path": "947aa770ce7060abb984559ce3a3fbf4b3483e8097b906e888edfd68305777ea.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 380, + 343, + 530, + 488 + ], + "blocks": [ + { + "bbox": [ + 380, + 343, + 530, + 488 + ], + "lines": [ + { + "bbox": [ + 380, + 343, + 530, + 488 + ], + "spans": [ + { + "bbox": [ + 380, + 343, + 530, + 488 + ], + "type": "image", + "image_path": "c5d6fc26cb4ac6b2d6ad7911b284eefcd60b7610b7d3d6a9ff1046eaa2fff073.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1473" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "spans": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "type": "text", + "content": "A For every submission:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 106, + 414, + 243 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "spans": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "type": "text", + "content": "A1. Did you describe the limitations of your work? Section 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 142, + 329, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 142, + 329, + 169 + ], + "spans": [ + { + "bbox": [ + 77, + 142, + 329, + 169 + ], + "type": "text", + "content": "A2. Did you discuss any potential risks of your work? Section 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 179, + 414, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 179, + 414, + 206 + ], + "spans": [ + { + "bbox": [ + 77, + 179, + 414, + 206 + ], + "type": "text", + "content": "A3. Do the abstract and introduction summarize the paper's main claims? Left blank." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 77, + 215, + 398, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 215, + 398, + 243 + ], + "spans": [ + { + "bbox": [ + 77, + 215, + 398, + 243 + ], + "type": "text", + "content": "A4. Have you used AI writing assistants when working on this paper? Left blank." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 68, + 253, + 291, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 253, + 291, + 266 + ], + "spans": [ + { + "bbox": [ + 68, + 253, + 291, + 266 + ], + "type": "text", + "content": "B Did you use or create scientific artifacts?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 79, + 270, + 128, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 270, + 128, + 283 + ], + "spans": [ + { + "bbox": [ + 79, + 270, + 128, + 283 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 76, + 292, + 524, + 634 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 292, + 315, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 292, + 315, + 319 + ], + "spans": [ + { + "bbox": [ + 76, + 292, + 315, + 319 + ], + "type": "text", + "content": "B1. Did you cite the creators of artifacts you used? Not applicable. Left blank." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 76, + 328, + 463, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 328, + 463, + 355 + ], + "spans": [ + { + "bbox": [ + 76, + 328, + 463, + 355 + ], + "type": "text", + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Not applicable. Left blank." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 76, + 364, + 524, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 364, + 524, + 432 + ], + "spans": [ + { + "bbox": [ + 76, + 364, + 524, + 432 + ], + "type": "text", + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Not applicable. Left blank." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 76, + 441, + 524, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 441, + 524, + 496 + ], + "spans": [ + { + "bbox": [ + 76, + 441, + 524, + 496 + ], + "type": "text", + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Not applicable. Left blank." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 76, + 504, + 524, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 504, + 524, + 544 + ], + "spans": [ + { + "bbox": [ + 76, + 504, + 524, + 544 + ], + "type": "text", + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Not applicable. Left blank." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 76, + 554, + 524, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 554, + 524, + 634 + ], + "spans": [ + { + "bbox": [ + 76, + 554, + 524, + 634 + ], + "type": "text", + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Not applicable. Left blank." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "spans": [ + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "type": "text", + "content": "C Did you run computational experiments?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 79, + 661, + 123, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 661, + 123, + 673 + ], + "spans": [ + { + "bbox": [ + 79, + 661, + 123, + 673 + ], + "type": "text", + "content": "Section 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 77, + 683, + 524, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 683, + 524, + 723 + ], + "spans": [ + { + "bbox": [ + 77, + 683, + 524, + 723 + ], + "type": "text", + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Section 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 67, + 729, + 522, + 748 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 729, + 522, + 748 + ], + "spans": [ + { + "bbox": [ + 67, + 729, + 522, + 748 + ], + "type": "text", + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "spans": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "text", + "content": "ACL 2023 Responsible NLP Checklist" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1474" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 77, + 70, + 524, + 236 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 77, + 70, + 523, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 70, + 523, + 110 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 523, + 110 + ], + "type": "text", + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Section 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 77, + 120, + 524, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 120, + 524, + 174 + ], + "spans": [ + { + "bbox": [ + 77, + 120, + 524, + 174 + ], + "type": "text", + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Section 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 183, + 524, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 183, + 524, + 236 + ], + "spans": [ + { + "bbox": [ + 77, + 183, + 524, + 236 + ], + "type": "text", + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Section 4" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 67, + 247, + 522, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 247, + 522, + 278 + ], + "spans": [ + { + "bbox": [ + 67, + 247, + 522, + 278 + ], + "type": "text", + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 76, + 286, + 524, + 539 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 76, + 286, + 524, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 286, + 524, + 326 + ], + "spans": [ + { + "bbox": [ + 76, + 286, + 524, + 326 + ], + "type": "text", + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Not applicable. Left blank." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 76, + 336, + 524, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 336, + 524, + 390 + ], + "spans": [ + { + "bbox": [ + 76, + 336, + 524, + 390 + ], + "type": "text", + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Not applicable. Left blank." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 76, + 400, + 524, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 400, + 524, + 454 + ], + "spans": [ + { + "bbox": [ + 76, + 400, + 524, + 454 + ], + "type": "text", + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Not applicable. Left blank." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 76, + 462, + 519, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 462, + 519, + 489 + ], + "spans": [ + { + "bbox": [ + 76, + 462, + 519, + 489 + ], + "type": "text", + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Not applicable. Left blank." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 76, + 498, + 524, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 498, + 524, + 539 + ], + "spans": [ + { + "bbox": [ + 76, + 498, + 524, + 539 + ], + "type": "text", + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Not applicable. Left blank." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1475" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/UniTRec_ A Unified Text-to-Text Transformer and Joint Contrastive Learning Framework for Text-based Recommendation/31a35e47-82fb-45ce-aa23-0bc4dd29fd03_content_list.json b/2023/UniTRec_ A Unified Text-to-Text Transformer and Joint Contrastive Learning Framework for Text-based Recommendation/31a35e47-82fb-45ce-aa23-0bc4dd29fd03_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..961002ac1d6d8818519e9c57ac68aa3bf481c96e --- /dev/null +++ b/2023/UniTRec_ A Unified Text-to-Text Transformer and Joint Contrastive Learning Framework for Text-based Recommendation/31a35e47-82fb-45ce-aa23-0bc4dd29fd03_content_list.json @@ -0,0 +1,1529 @@ +[ + { + "type": "text", + "text": "UniTRec: A Unified Text-to-Text Transformer and Joint Contrastive Learning Framework for Text-based Recommendation", + "text_level": 1, + "bbox": [ + 144, + 80, + 852, + 120 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhiming Mao $^{1,2}$ , Huimin Wang $^{1,3}$ , Yiming Du $^{1,2}$ , Kam-Fai Wong $^{1,2}$", + "bbox": [ + 216, + 129, + 789, + 147 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1The Chinese University of Hong Kong, Hong Kong, China", + "bbox": [ + 258, + 148, + 742, + 164 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ MoE Key Laboratory of High Confidence Software Technologies, China", + "bbox": [ + 206, + 164, + 796, + 181 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Jarvis Lab, Tencent, Shenzhen, China", + "bbox": [ + 344, + 181, + 657, + 197 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zmmao,ydu,kfwong}@se.cuhk.edu.hk", + "bbox": [ + 304, + 199, + 700, + 214 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "hmmmwang@tencent.com", + "bbox": [ + 379, + 216, + 623, + 230 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 252, + 339, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Prior study has shown that pretrained language models (PLM) can boost the performance of text-based recommendation. In contrast to previous works that either use PLM to encode user history as a whole input text, or impose an additional aggregation network to fuse multi-turn history representations, we propose a unified local- and global-attention Transformer encoder to better model two-level contexts of user history. Moreover, conditioned on user history encoded by Transformer encoders, our framework leverages Transformer decoders to estimate the language perplexity of candidate text items, which can serve as a straightforward yet significant contrastive signal for user-item text matching. Based on this, our framework, UniTRec, unifies the contrastive objectives of discriminative matching scores and candidate text perplexity to jointly enhance text-based recommendation. Extensive evaluation shows that UniTRec delivers SOTA performance on three text-based recommendation tasks. $^{1}$", + "bbox": [ + 141, + 278, + 460, + 590 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 114, + 602, + 258, + 617 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-based recommendation (Li et al., 2010; Gu et al., 2016; Okura et al., 2017; Malkiel et al., 2020) aims to recommend relevant textual content (e.g., news articles, Twitter posts) to people based on their behaviors as represented in historical log texts. For instance, engagement recommendation (Cheng et al., 2022) on social media (e.g., Twitter and Reddit) helps users discover and engage with interested threads by modeling their browsing history.", + "bbox": [ + 112, + 627, + 489, + 771 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pretrained language models (Devlin et al., 2019; Brown et al., 2020) have made waves in recent text-based recommendation research (Zhang et al., 2021; Qi et al., 2022; Geng et al., 2022). The most common practice is using PLM encoders (BERT family) to learn representations of user history and candidate item texts. Recommendation", + "bbox": [ + 112, + 772, + 489, + 885 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "matching scores are computed over the user and item representations and finally optimized by noise contrastive estimation (NCE) loss (Gutmann and Hyvarinen, 2010) for ranking multiple candidates.", + "bbox": [ + 507, + 253, + 884, + 316 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unlike encoding single text, using PLM to encode multi-turn texts of user history is nontrivial. Existing works (Malkiel et al., 2020; Qi et al., 2022; Geng et al., 2022) concatenate multi-turn history texts as a whole input text, then use one PLM encoder to learn the holistic user representation. This is a standard PLM encoding manner but ignores the relation among history turns, as all word tokens from different history turns are equally attended2. In contrast, previous studies point out that learning the relation among user history turns is also beneficial (Zeng et al., 2020; Qi et al., 2021). Another approach is using PLM encoders to learn representations from multi-turn history texts, followed by an additional aggregation network to fuse the multi-turn representations (Wu et al., 2021; Li et al., 2022). However, the imposed aggregation networks (with newly initialized parameters) weaken the representation power of PLM encoders which are already pretrained on large-scale corpora.", + "bbox": [ + 507, + 317, + 884, + 640 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This work introduces UniTRec, a Unified text-to-text Transformer framework for text-based Recommendation. In the encoder component of UniTRec, we design local- and global-attention to learn user history representations through tailored attention masking, which aims to jointly model word-level and turn-level relations of user history. UniTRec can utilize the full power of PLM encoders because it preserves the intact structure of PLM encoders without newly imposed parameters.", + "bbox": [ + 507, + 640, + 882, + 802 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Different from most previous works that predict user-candidate matching scores solely based on the representations learned by Transformer encoders, we argue that conditioned on user representations", + "bbox": [ + 507, + 803, + 882, + 868 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "2There is no inductive bias of turn-level and history-level relations introduced to Transformer self-attention computation, where each token plays an equal role.", + "bbox": [ + 507, + 879, + 882, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1Our code is available at https://github.com/Veason-silverbullet/UniTRec.", + "bbox": [ + 112, + 892, + 489, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1160", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics", + "bbox": [ + 226, + 945, + 769, + 957 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Volume 2: Short Papers, pages 1160-1170", + "bbox": [ + 368, + 958, + 628, + 971 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "July 9-14, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 295, + 972, + 700, + 985 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/40a6b845906692592a6a73d5a2c235d837c405da9535503c9288c81a7c8d3118.jpg", + "image_caption": [ + "Figure 1: An example of perplexity-based ranking for candidate item texts, conditioned on user history. The illustrated task is text-based news recommendation." + ], + "image_footnote": [], + "bbox": [ + 115, + 83, + 499, + 240 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "learned by Transformer encoders, candidate text perplexity (PPL) estimated by pretrained Transformer decoders is also a straightforward yet significant signal for text-based recommendation. As shown in Figure 1, we hypothesize that the candidate text perplexity estimated by pretrained LM decoders can directly measure the text matching degree between user history and candidate texts. It is because the perplexity estimates the likelihood of candidate texts based on encoder outputs, which naturally indicates the probabilities of candidate texts given the user history. Besides, UniTRec can use the last hidden states of Transformer decoders to directly predict matching scores. Hence, this work unifies the contrastive objectives of discriminative matching scores and candidate text perplexity to jointly enhance text-based recommendation.", + "bbox": [ + 112, + 319, + 489, + 593 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of this work are: (1) We propose local- and global-attention to model two-level relation of user history without additional parameters, which enjoys the full power of PLM encoders. (2) We introduce PLM perplexity to measure user-candidate text matching and unify the objectives of discriminative matching scores and candidate text perplexity to enhance text-based recommendation. (3) Experiments on three text-based recommendation datasets validate the effectiveness of UniTRec.", + "bbox": [ + 112, + 595, + 489, + 755 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Approach", + "text_level": 1, + "bbox": [ + 112, + 770, + 236, + 785 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 Unified User-history Modeling", + "text_level": 1, + "bbox": [ + 112, + 797, + 400, + 813 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Formally, multi-turn history of a user is represented as $H = [t_1, t_2, \\dots, t_N]$ , and each turn text $t_i$ contains $|t_i|$ words as $t_i = [x_i^1, x_i^2, \\dots, x_i^{|t_i|}]$ . UniTRec aims to unify learning word- and turn-level context representations in one Transformer encoder.", + "bbox": [ + 112, + 819, + 487, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Local attention on word-level context. We first", + "bbox": [ + 131, + 903, + 487, + 917 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "concatenate the multi-turn history texts as the input tokens $X = [x_{1}^{1}, x_{1}^{2}, \\dots, x_{1}^{|t_{1}|}, \\dots, x_{N}^{1}, x_{N}^{2}, \\dots, x_{N}^{|t_{N}|}]$ . Inspired by Dong et al. (2019), we tailor the attention masking in Transformer self-attention to learn the word-level context of each turn. Specifically, we allow word tokens from the same turn to attend to each other, while tokens from different turns are excluded from self-attention computation:", + "bbox": [ + 507, + 84, + 884, + 214 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "$\\mathbf{M}_{i,j} = \\left\\{ \\begin{array}{ll}0, & \\mathrm{token} x_i\\mathrm{and}x_j\\mathrm{in the same turn}\\\\ -\\infty , & \\mathrm{otherwise} \\end{array} \\right.$", + "bbox": [ + 507, + 218, + 875, + 260 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}} + \\mathbf {M}\\right) V \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 261, + 882, + 310 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": ", where $Q, K, V$ are self-attention query, key, and value in Vaswani et al. (2017), $\\mathbf{M}$ is the mask matrix to achieve local-attention inside each turn text. The local self-attention blocks consist of $L_{1}$ layers, by which original PLM encoders can be adapted to learn word-level context representations of turns.", + "bbox": [ + 507, + 313, + 882, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Global attention on turn-level context. Over the local self-attention layers, we leverage global self-attention to model the relation among history turns. Specifically, tokens from all turns attend to each other in self-attention computation (by setting the mask matrix $\\mathbf{M} = \\mathbf{0}$ ). In this way, Transformer encoders can perform global interaction among each token (and turn) to learn turn-level context representations of user history. There are $L_{2}$ layers in the global self-attention blocks, which can also be inherited from PLM encoders directly.", + "bbox": [ + 507, + 410, + 882, + 586 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2 Joint Contrastive Ranking Objectives", + "text_level": 1, + "bbox": [ + 507, + 596, + 850, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Conditioned on the history representation, we input the candidate text to Transformer decoders to predict how likely it should be recommended. It is worth noting that Transformer decoders can naturally perform effective cross-attention interaction between history and candidate hidden states.", + "bbox": [ + 507, + 618, + 882, + 713 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2.1 Objective on Discriminative Scores", + "text_level": 1, + "bbox": [ + 507, + 722, + 843, + 738 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by Lewis et al. (2020), we feed the last hidden state of decoder output $h_{T}$ to an MLP scorehead which predicts the user-candidate matching score $S^{d} = \\mathrm{ScoreHead}(h_{T})$ . The matching score is discriminative, as higher scores indicate higher user-candidate matching probabilities.", + "bbox": [ + 507, + 741, + 882, + 837 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Following previous works (Li et al., 2022; Qi et al., 2022), we adopt negative sampling with NCE loss to optimize matching score prediction. Given the user history and its ground truth matched candidate $C_i$ , UniTRec predicts the matching score", + "bbox": [ + 507, + 839, + 882, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1161", + "bbox": [ + 480, + 927, + 517, + 940 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e95c2a2c75f3e5b5a1f797a305cfcdafd75832efb99e732dd3ad387e4f424bbf.jpg", + "image_caption": [ + "Figure 2: Overview of UniTRec. In training, matching scores $S^d$ and $S^p$ are optimized by the NCE loss, respectively. In inference, $S^d$ and $S^p$ are normalized and combined to derive the final output ranking." + ], + "image_footnote": [], + "bbox": [ + 198, + 80, + 801, + 256 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "as $S_{i}^{d + }$ . In addition, $K$ unmatched negative candidates $\\{C_j\\}_{j = 1}^K$ are sampled from the candidate set, and their matching scores are $\\{S_j^{d - }\\}_{j = 1}^K$ . The NCE loss is represented in a contrastive form:", + "bbox": [ + 112, + 313, + 487, + 382 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i} ^ {d} = - \\log \\frac {\\exp (S _ {i} ^ {d +})}{\\exp (S _ {i} ^ {d +}) + \\sum_ {j = 1} ^ {K} \\exp (S _ {j} ^ {d -})} \\quad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 131, + 385, + 487, + 428 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2.2 Objective on Candidate Text Perplexity", + "text_level": 1, + "bbox": [ + 112, + 439, + 482, + 456 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As aforementioned, UniTRec leverages perplexity to rank candidate texts. Since lower perplexity indicates higher user-candidate matching probability, regarding the candidate text $Y = [y_{1}, y_{2}, \\dots, y_{T}]$ , we define the perplexity-based matching score $S^{p}$ as its negative perplexity:", + "bbox": [ + 112, + 464, + 487, + 561 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nS ^ {p} = - \\operatorname {P P L} (Y) = \\frac {1}{T} \\sum_ {i = 1} ^ {T} \\log p _ {\\theta} \\left(y _ {i} \\mid y _ {< i}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 124, + 568, + 487, + 596 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": ", where $p_{\\theta}(\\cdot)$ denotes the target probability output from the UniTRec Transformer decoder. Similar to Eq. (2), we optimize the perplexity-based matching score $S^p$ in the NCE loss form. As perplexity empirically varies in a wide range, we introduce a temperature parameter $\\tau$ to balance the joint NCE loss gradients following Radford et al. (2021).", + "bbox": [ + 112, + 605, + 487, + 718 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i} ^ {p} = - \\log \\frac {\\exp \\left(\\tau \\cdot S _ {i} ^ {p +}\\right)}{\\exp \\left(\\tau \\cdot S _ {i} ^ {p +}\\right) + \\sum_ {j = 1} ^ {K} \\exp \\left(\\tau \\cdot S _ {j} ^ {p -}\\right)} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 725, + 485, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": ", where $\\tau$ is learnable and initialized to 1. On the training dataset $\\mathcal{D}$ , the joint contrastive learning objective is formulated as:", + "bbox": [ + 112, + 778, + 487, + 826 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\sum_ {i = 1} ^ {| \\mathcal {D} |} \\left(\\mathcal {L} _ {i} ^ {d} + \\mathcal {L} _ {i} ^ {p}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 211, + 829, + 487, + 858 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3 Model Initialization and Inference", + "text_level": 1, + "bbox": [ + 507, + 316, + 823, + 331 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As UniTRec is a standard text-to-text Transformer, we initialize the parameters from pretrained BART (Lewis et al., 2020). In inference, UniTRec predicts the discriminative and perplexity-based scores for each candidate item, respectively. The two separate scores $S^d$ and $S^p$ are normalized, averaged, and finally ranked as the output. Detailed ranking process is provided in Appendix B.", + "bbox": [ + 507, + 336, + 882, + 464 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 Experiments", + "text_level": 1, + "bbox": [ + 507, + 476, + 655, + 492 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We evaluate UniTRec on three text-based recommendation tasks: 1) NewsRec, to recommend news articles to users based on their browsing history. We use the MIND-small dataset (Wu et al., 2020) for experiments. 2) QuoteRec, to recommend quotations to users based on their conversation history. We use the Reddit-quotation dataset (Wang et al., 2021) for experiments. 3) EngageRec, to recommend social media posts for users to engage with based on their comment history. We use the dataset released by Zeng et al. (2020) for experiments. Detailed dataset statistics is provided in Appendix A.", + "bbox": [ + 507, + 500, + 882, + 693 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Implementation Details. The UniTRec encoder and decoder both consist of 6 Transformer layers with 768-dimensional hidden states and 12 attention heads. We set $L_{1} = 3$ and $L_{2} = 3$ . We use AdamW optimizer (Loshchilov and Hutter, 2019) to train UniTRec with cosine learning rate decay.", + "bbox": [ + 507, + 694, + 882, + 789 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Baselines. We compare UniTRec with competitive baselines: 1) GRU4Rec (Balázs et al., 2016) utilizes a GRU network to learn multi-turn history. 2) SASRec (Kang and McAuley, 2018) encodes user history with a self-attention based sequential model. 3) BERT4Rec (Sun et al., 2019) employs bidirectional self-attention to model user history. 4) RoBERTa-Sim, a simple yet strong baseline men", + "bbox": [ + 507, + 790, + 882, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "3Note https://huggingface.co/docs/transformers/perplexity for LM perplexity calculation. We empirically discard the outer exponential term in the PPL formula, because it already exists in NCE loss Eq. (4) and does not affect the final ranking.", + "bbox": [ + 112, + 868, + 487, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1162", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/d5616c46fab06466614308db2cbf65cc97437d9fc394f7a6c790be94c198ffa6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelNewsRecQuoteRecEngageRec
MRRNDCG@5/10HR@5/10MRRNDCG@5/10HR@5/10MRRNDCG@5/10HR@5/10
GRU4Rec32.9136.20/42.5350.33/68.3534.0834.65/37.9344.45/54.632.121.04/1.511.27/2.65
SASRec32.6036.03/42.3750.63/68.6433.6334.30/37.4944.32/54.202.401.49/1.952.16/3.47
BERT4Rec32.8736.18/42.4050.21/67.9733.5934.26/37.2743.76/53.053.041.98/3.232.81/6.67
RoBERTa-Sim32.9636.47/42.8151.06/69.0837.1337.96/41.1848.14/58.063.742.66/3.754.42/7.70
UNBERT33.0936.53/42.8450.87/68.8239.7540.74/43.6950.90/60.042.831.96/2.673.11/5.24
UniTRec33.7637.63/43.7452.61/69.8941.2442.38/45.3152.87/61.884.063.23/4.294.58/7.68
", + "bbox": [ + 119, + 79, + 877, + 175 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/7ce67a8851ac89198c0864f344a456b91b633e6a541a95b1778ca7924257b456.jpg", + "table_caption": [ + "Table 1: Experiment results on three text-based recommendation tasks. MRR denotes mean reciprocal rank, NDCG denotes normalized discounted cumulative gain, and HR denotes hit ratio (presented in percentage). The overall performance of UniTRec is better than other baseline models with $p$ -value $< 0.05$ , validated by unpaired t-test." + ], + "table_footnote": [], + "table_body": "
ModelNewsRecQuoteRecEngageRec
MRRNDCG@5/10HR@5/10MRRNDCG@5/10HR@5/10MRRNDCG@5/10HR@5/10
UniTRec33.7637.63/43.7452.61/69.8941.2442.38/45.3152.87/61.884.063.23/4.294.58/7.68
w/o BART Init30.3133.32/39.6947.55/65.7819.0217.66/20.8022.45/32.162.240.86/1.611.27/3.62
w/o Local-Att33.3437.22/43.3252.28/69.5440.4441.63/44.5652.09/61.153.923.19/4.154.38/7.36
w/o Global-Att33.2237.06/43.1752.14/69.4740.2541.47/44.2652.07/60.763.642.78/3.593.89/6.35
Disc-Score only33.0736.76/43.0351.68/69.4640.5941.81/44.6552.39/61.143.822.99/3.604.49/6.85
PPL-Score only32.8336.39/42.5951.05/68.6740.3141.43/44.4752.13/61.203.292.39/3.033.86/5.66
", + "bbox": [ + 119, + 242, + 877, + 338 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 2: Recommendation performance of ablation model variants.", + "bbox": [ + 270, + 349, + 724, + 363 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tioned in Qi et al. (2022), uses the hidden states of [CLS] tokens to measure user-candidate similarity. 5) UNBERT, implemented as Zhang et al. (2021), concatenates history and candidate texts as the input to BERT and predicts matching scores from the final hidden states of [CLS] tokens.", + "bbox": [ + 112, + 388, + 487, + 483 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that we do not consider other methods that use non-text inputs (e.g., user profile, text topic labels). For fair comparison, all baseline models use pretrained 12-layer RoBERTa-base (Liu et al., 2019) as text encoders to learn embeddings of texts.", + "bbox": [ + 112, + 486, + 489, + 565 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Main Results", + "text_level": 1, + "bbox": [ + 112, + 577, + 263, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1 shows the performance of experiment models. From the results of NewsRec and QuoteRec, we can see that UniTRec outperforms all baseline models by a clear margin. Also, RoBERTa-Sim and UNBERT that directly use the [CLS] hidden states to represent user history, surpass other baselines that build additional aggregation networks upon the whole RoBERTa outputs. As displayed in the results, EngageRec is the most difficult task. We inspect the dataset and find that the texts on social media contain too much noise (e.g., URL and emoji), and the user history contains less number of turns. Nevertheless, UniTRec achieves better overall performance than other baseline models, validating its robustness on noisy text inputs and limited user history.", + "bbox": [ + 112, + 598, + 489, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 Ablation Studies and Analyses", + "text_level": 1, + "bbox": [ + 112, + 866, + 400, + 881 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We further conduct ablation studies on UniTRec. The experiment results are reported in Table 2.", + "bbox": [ + 112, + 887, + 487, + 917 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Initialization of UniTRec. We train UniTRec from scratch without initialization from pretrained BART (refer to w/o BART Init). The recommendation performance significantly drops in all three tasks, which indicates that acquiring effective text understanding ability from PLM is a necessary key to UniTRec performance.", + "bbox": [ + 507, + 388, + 882, + 501 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Local and global attention. We investigate the function of two-level attention modules of the UniTRec history encoder. Concretely, we set $L_{1} = 0$ in w/o Local-Att and $L_{2} = 0$ in w/o Global-Att, where $L_{1} + L_{2} = 6$ . We can observe that removing local and global attention from the original UniTRec history encoder both lead to suboptimal performance, while the performance drop is more significant in w/o Global-Att. The results justify the effectiveness of jointly modeling two-level history contexts through adapted Transformer attention masking without additional parameters.", + "bbox": [ + 507, + 508, + 884, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Discriminative and perplexity-based objectives. We probe into training UniTRec with standalone discriminative (Disc-Score only) and perplexity-based (PPL-Score only) contrastive objectives, respectively. We can see that the discriminative objective yields better performance than the perplexity-based objective. Besides, the model performance on both standalone objectives declines compared to the original joint objective. The results indicate that the discriminative and perplexity-based matching scores are complementary and can jointly provide more accurate signals of user history and candidate text matching for text-based recommendation.", + "bbox": [ + 507, + 709, + 882, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1163", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4 Conclusion", + "text_level": 1, + "bbox": [ + 112, + 84, + 247, + 98 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We present a unified Transformer UniTRec for text-based recommendation. UniTRec learns two-level contexts of multi-turn user history and jointly exploits discriminative matching scores and candidate text perplexity as matching objectives. Empirical experiments on three text-based recommendation datasets corroborate the effectiveness of UniTRec.", + "bbox": [ + 112, + 110, + 489, + 221 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5 Limitations", + "text_level": 1, + "bbox": [ + 112, + 234, + 250, + 249 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our model only focuses on utilizing text information for recommendation, which is a key limitation of this work. In real-world settings, recommender systems are usually required to handle heterogeneous information inputs. UniTRec is a pure text-based recommender modeling user history and candidate texts as inputs. However, incorporating additional side information (e.g., user profile, text topic labels, and dwell time of user behaviors) could further improve the recommendation performance and alleviate the cold start problem. Furthermore, UniTRec only models two-level relations of user behavior history. Nonetheless, incorporating more user behavior information, such as implicit and negative feedback, could further enhance the recommendation performance.", + "bbox": [ + 112, + 260, + 490, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 114, + 531, + 285, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We appreciate constructive comments from anonymous reviewers. The research described in this paper is partially supported by CUHK under Project No. 3230366.", + "bbox": [ + 112, + 556, + 489, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 114, + 648, + 213, + 662 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Hidasi Balázs, Karatzoglou Alexandros, Baltrunas Linas, and Tikk Domonkos. 2016. Session-based recommendations with recurrent neural networks. In 4th International Conference on Learning Representations ICLR 2016, San Juan, Puerto Rico, May 2-4, 2016, Conference Track Proceedings.", + "Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems," + ], + "bbox": [ + 115, + 671, + 490, + 917 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "volume 33, pages 1877-1901. Curran Associates, Inc.", + "Daniel Cheng, Kyle Yan, Phillip Keung, and Noah A. Smith. 2022. The engage corpus: A social media dataset for text-based recommender systems. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 1885-1889, Marseille, France. European Language Resources Association.", + "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 4171-4186. Association for Computational Linguistics.", + "Li Dong, Nan Yang, Wenhui Wang, Furu Wei, Xiaodong Liu, Yu Wang, Jianfeng Gao, Ming Zhou, and Hsiao-Wuen Hon. 2019. Unified language model pre-training for natural language understanding and generation. In 33rd Conference on Neural Information Processing Systems (NeurIPS 2019).", + "Shijie Geng, Shuchang Liu, Zuohui Fu, Yingqiang Ge, and Yongfeng Zhang. 2022. Recommendation as language processing (rlp): A unified pretrain, personalized prompt & predict paradigm (p5). In Proceedings of the 16th ACM Conference on Recommender Systems, RecSys '22, page 299-315, New York, NY, USA. Association for Computing Machinery.", + "Youyang Gu, Tao Lei, Regina Barzilay, and Tommi Jaakkola. 2016. Learning to refine text based recommendations. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2103-2108, Austin, Texas. Association for Computational Linguistics.", + "Michael Gutmann and Aapo Hyvarinen. 2010. Noise-contrastive estimation: A new estimation principle for unnormalized statistical models. In Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics, volume 9 of Proceedings of Machine Learning Research, pages 297-304, Chia Laguna Resort, Sardinia, Italy. PMLR.", + "Wang-Cheng Kang and Julian McAuley. 2018. Self-attentive sequential recommendation. In 2018 IEEE International Conference on Data Mining (ICDM), pages 197-206.", + "Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871-7880, Online. Association for Computational Linguistics." + ], + "bbox": [ + 509, + 85, + 885, + 917 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1164", + "bbox": [ + 482, + 928, + 521, + 940 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Jian Li, Jieming Zhu, Qiwei Bi, Guohao Cai, Lifeng Shang, Zhenhua Dong, Xin Jiang, and Qun Liu. 2022. MINER: Multi-interest matching network for news recommendation. In Findings of the Association for Computational Linguistics: ACL 2022, pages 343-352, Dublin, Ireland. Association for Computational Linguistics.", + "Yize Li, Jiazhong Nie, Yi Zhang, Bingqing Wang, Baoshi Yan, and Fuliang Weng. 2010. Contextual recommendation based on text mining. In *Coling* 2010: Posters, pages 692-700, Beijing, China. Coling 2010 Organizing Committee.", + "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. In arXiv preprint arXiv: 1907.11692. arXiv.", + "Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net.", + "Itzik Malkiel, Oren Barkan, Avi Caciularu, Noam Razin, Ori Katz, and Noam Koenigstein. 2020. *RecoBERT: A catalog language model for text-based recommendations*. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, Online. Association for Computational Linguistics.", + "Shumpei Okura, Yukihiro Tagami, Shingo Ono, and Akira Tajima. 2017. Embedding-based news recommendation for millions of users. In Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, page 1933-1942, New York, NY, USA. Association for Computing Machinery.", + "Fanchao Qi, Yanhui Yang, Jing Yi, Zhili Cheng, Zhiyuan Liu, and Maosong Sun. 2022. QuoteR: A benchmark of quote recommendation for writing. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 336-348, Dublin, Ireland. Association for Computational Linguistics.", + "Tao Qi, Fangzhao Wu, Chuhan Wu, Peiru Yang, Yang Yu, Xing Xie, and Yongfeng Huang. 2021. HieRec: Hierarchical user interest modeling for personalized news recommendation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 5446-5456, Online. Association for Computational Linguistics.", + "Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International" + ], + "bbox": [ + 115, + 85, + 489, + 917 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 8748-8763. PMLR.", + "Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang. 2019. Bert4rec: Sequential recommendation with bidirectional encoder representations from transformer. In Proceedings of the 28th ACM International Conference on Information and Knowledge Management, CIKM '19, page 1441-1450, New York, NY, USA. Association for Computing Machinery.", + "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30, pages 5998-6008. Curran Associates, Inc.", + "Lingzhi Wang, Xingshan Zeng, and Kam-Fai Wong. 2021. Quotation recommendation and interpretation based on transformation from queries to quotations. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 754-758, Online. Association for Computational Linguistics.", + "Chuhan Wu, Fangzhao Wu, Tao Qi, and Yongfeng Huang. 2021. Empowering news recommendation with pre-trained language models. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '21, page 1652-1656, New York, NY, USA. Association for Computing Machinery.", + "Fangzhao Wu, Ying Qiao, Jiun-Hung Chen, Chuhan Wu, Tao Qi, Jianxun Lian, Danyang Liu, Xing Xie, Jianfeng Gao, Winnie Wu, and Ming Zhou. 2020. MIND: A large-scale dataset for news recommendation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3597-3606, Online. Association for Computational Linguistics.", + "Xingshan Zeng, Jing Li, Lu Wang, Zhiming Mao, and Kam-Fai Wong. 2020. Dynamic online conversation recommendation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3331-3341, Online. Association for Computational Linguistics.", + "Qi Zhang, Jingjie Li, Qinglin Jia, Chuyuan Wang, Jieming Zhu, Zhaowei Wang, and Xiuqiang He. 2021. Umbert: User-news matching bert for news recommendation. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-21, pages 3356-3362. International Joint Conferences on Artificial Intelligence Organization. Main Track." + ], + "bbox": [ + 510, + 85, + 884, + 851 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1165", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/cee588b2bc04a0511e0aedc46adfb0b28ffdec05c6da196ac45df5a80b5bc0f0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetNewsRecQuoteRecEngageRec
Avg. history turns26.094.243.29
Avg. history tokens414.40279.82286.82
Avg. candidates37.2311117163
Avg. candidate tokens16.1519.11102.42
", + "bbox": [ + 117, + 80, + 497, + 147 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3: Statistics of three text-based recommendation training datasets. History and candidate tokens denote the number of BPE-tokenized tokens. The test set distribution is closed to the training sets (except candidates of EngageRec) and hence omitted. Note that the max length of each history log is truncated to 1024 tokens.", + "bbox": [ + 112, + 156, + 489, + 243 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A Dataset Statistics", + "text_level": 1, + "bbox": [ + 112, + 263, + 302, + 279 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The detailed statistics of the three text-based recommendation datasets are displayed in Table 3. Note that we use news titles as the text inputs for NewsRec following Qi et al. (2021). NewsRec regards the user clicked and non-clicked news as candidate texts, while QuoteRec and EngageRec regard all potential quotation texts and post texts as candidates. Different from Zeng et al. (2020) that formulates the task as recommending candidate users to given posts based on post content, we formulate the task as recommending candidate posts to given users based on user history.", + "bbox": [ + 112, + 288, + 489, + 481 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Algorithm 1 Candidate Ranking Processes", + "text_level": 1, + "bbox": [ + 115, + 494, + 426, + 510 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Input: discriminative scores $S^d = \\{S_1^d,S_2^d,\\dots,S_M^d\\}$ perplexity-based scores $S^{p} = \\{S_{1}^{p},S_{2}^{p},\\dots,S_{M}^{p}\\}$", + "bbox": [ + 115, + 514, + 463, + 541 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Output: final averaged ranking $R$ .", + "bbox": [ + 115, + 541, + 334, + 552 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Derive the normalized discriminative scores $S_{norm}^{d} =$ softmax(Sd).", + "2: Derive the normalized perplexity-based scores $S_{norm}^{p} =$ softmax( $S^p$ ).", + "3: Derive the geometric average scores $\\bar{S} = \\log (S_{norm}^d) + \\log (S_{norm}^p)$ .", + "4: Sort the averaged scores $\\bar{S}$ by descending order to derive the final ranking: $\\bar{R} \\gets \\mathrm{Rank}_{\\mathrm{des}}(\\bar{S})$", + "5: return $R$" + ], + "bbox": [ + 127, + 552, + 487, + 659 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "B Inference Ranking", + "text_level": 1, + "bbox": [ + 114, + 688, + 312, + 703 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given the user history and $M$ candidate texts, UniTRec first predicts the discriminative ranking scores $S^d = \\{S_1^d,S_2^d,\\dots,S_M^d\\}$ and perplexity-based ranking scores $S^{p} = \\{S_{1}^{p},S_{2}^{p},\\dots,S_{M}^{p}\\}$ of the candidates. Algorithm 1 outlines an approach to aggregate the final ranking based on $S^d$ and $S^p$ . Note that the function $\\mathrm{Rank}(S)^4$ denotes outputting the sorted order of elements in a score list $S$ . There exist other ways to average the ranking of $S^d$ and $S^p$ , which we leave for future work to explore.", + "bbox": [ + 112, + 713, + 489, + 873 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "C Qualitative Analysis", + "text_level": 1, + "bbox": [ + 509, + 83, + 722, + 99 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We show randomly sampled outputs of UniTRec, for instance, demonstrated on the news recommendation and quote recommendation tasks. Table 4 and 5 showcase the qualitative samples.", + "bbox": [ + 507, + 109, + 884, + 174 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "$^4$ Rank(S) works similarly to scipy.stats.rankdata(). For example in ascending order, $\\mathrm{Ran_{asc}}(\\{0.2, 0.6, 0.7, 0.4\\}) = \\mathrm{scipy.stats.rankdata}([0.2, 0.6, 0.7, 0.4]) = [1, 3, 4, 2]$", + "bbox": [ + 112, + 879, + 487, + 919 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1166", + "bbox": [ + 482, + 928, + 521, + 940 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7e5a33c24396675552e585b8b72a78527dafeb15cdd5137b89b50a70b3050531.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TurnHistory News Texts
#1Mac Engel: As long as these results are acceptable, Dallas Cowboys will continue to be losers
#2NFL world reacts to officials handing Packers win over Lions
#3Maryland Congressman Elijah Cummings, a Democrat and Chair of House Oversight and Reform Committee, has died: CNN
#4Unprecedented movement detected on California earthquake fault capable of 8.0 temblor
#5Bag Explodes While Being Loaded On Volaris Flight At Midway Airport
#6Orlando Scandrick rips Eagles: They have "accountability issues"
#7Meghan King Edmonds, Jim Edmonds' Nanny Denies Cheating Allegations
#8Nearly $400M worth of cocaine and marijuana intercepted by US Coast Guard
#9Former NBA first-round pick arrested in sex sting operation
#10China's trade with US shrinks in October despite optimism
", + "bbox": [ + 115, + 177, + 878, + 306 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/78d34d5e2448d59210826ce9ff2e37c62420605836ead25f73126d6614991189.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Candidate News TextsSdSpRClicked
Taylor Swift Rep Hits Back at Big Machine, Claims She's Actually Owed $7.9 Million in Unpaid Royalties0.0950.0694X
Former North Carolina State, NBA player Anthony Grundy dies in stabbing, police say0.1720.1553X
13 Reasons Why's Christian Navarro Slams Disney for Casting "the White Guy" in The Little Mermaid0.0480.0657X
Opinion: Colin Kaepernick is about to get what he deserves: a chance0.3030.2501✓
3 Indiana judges suspended after a night of drinking turned into a White Castle brawl0.0760.0595X
66 Cool Tech Gifts Anyone Would Be Thrilled to Receive0.0090.0059X
Police find 26 children behind false wall at Colorado day care0.0340.1166X
I've been writing about tiny homes for a year and spent 2 nights in a 300-foot home to see what it is all about0.0290.0198X
Report: Police investigating woman's death after Redskins' player Montae Nicholson took her to hospital0.2350.2612✓
", + "bbox": [ + 117, + 315, + 878, + 432 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/623da3869d86d65f0daf66f9f99d3cc4428c59f94f4599544aee48256d8e2177.jpg", + "table_caption": [ + "(i) Qualitative Example-A from news recommendation." + ], + "table_footnote": [], + "table_body": "
TurnHistory News Texts
#1Toddler dancing to celebrate 11 months cancer-free goes viral
#2NFL Week 8 Power Rankings: Old-school football rules the day
#3The 25 US cities where it's easiest to get a mortgage
#4Burning questions for Cowboys vs Giants on "Monday Night Football"
#5Who's the favorite to win 2019 NFL rushing title?
#6Grading all 32 NFL teams heading into the last eight weeks of the 2019 season
#7Jennifer Aniston looks amazing in a makeup-free selfie, plus more news
#8This $12 million "mansion yacht" is made entirely of stainless steel and it's a first for the industry. Take a peek inside
", + "bbox": [ + 115, + 475, + 878, + 581 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3e351c6e27ce61a217bda77205221667d02759d07489a0de62e543df624e0ecc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Candidate News TextsSdSpRClicked
Opinion: Colin Kaepernick is about to get what he deserves: a chance0.3300.4001✓
U.S. Troops Will Die If They Remain in Syria, Bashar Al-Assad Warns0.0240.01110✗
Pete Davidson, Kaia Gerber Are Dating, Trying to Stay "Low Profile"0.0640.0336✗
The Hottest Tech Gifts This Holiday Season0.0500.0278✗
Taylor Swift Rep Hits Back at Big Machine, Claims She's Actually Owed $7.9 Million in Unpaid Royalties0.0460.0387✗
13 Reasons Why's Christian Navarro Slams Disney for Casting "the White Guy" in The Little Mermaid0.0600.0964✓
Some believe Mason Rudolph, hit in head with his own helmet, isn't getting enough blame0.1540.1792✓
South Carolina teen gets life in prison for deadly elementary school shooting0.0660.0465✗
The Unlikely Star of My Family's Thanksgiving Table0.0470.0219✗
Police investigating woman's death after Redskins' player Montae Nicholson took her to hospital0.1580.1493✗
", + "bbox": [ + 119, + 590, + 875, + 721 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(ii) Qualitative Example-B from news recommendation.", + "bbox": [ + 307, + 726, + 685, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4: Case analyses of news recommendation. History News Texts are sorted by user-clicked timestamps. $S^d$ , $S^p$ , and $\\bar{R}$ are normalized discriminative, perplexity-based scores, and average ranking as described in Appendix B. Clicked denotes the ground truth user-click labels. Note that the experiment history logs are anonymized and delinked, which is always the first priority of the recommendation study.", + "bbox": [ + 112, + 760, + 882, + 819 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1167", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b1f9abb62882335295bc058772916268cad21fe7a1d0a5746ddce5588e3aa348.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TurnConversation Threading History
#1I own an FJ. It's a great car and even on stockies. It's great offroad.
#2I feel bad for you that you run the risk of being associated with the typical FJ owner.
#3What is a typical FJ owner? I've not heard anything bad about FJ owners.
#4It's like someone who drives a jeep wrangler in NYC. There's no need. Tons of FJ owners do that have it and not use it for what it's made for.
#5God forbid someone likes the design of a car and doesn't use it offroad.
#6Then buy a much more economic environmentalist friendly version. If you buy something and always use it for much less than it's purpose, why buy it?
#7Or people can buy whatever the hell they want because it's their money and not yours.
#8You're entirely right. Just like people can be rude just because you can do it, because you have the ability but why should you ass.
#9I wasn't aware that somebody buying a vehicle that they like and you don't was morally wrong.
#10I love FJs. It's perfectly fine to buy whatever you think looks nice.
", + "bbox": [ + 115, + 80, + 878, + 221 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/196da27e5d28c749cbc667bcd84304052778a81fecdd12867e526adf4d514a1a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Candidate Quote Texts\\( S^d \\)\\( S^P \\)\\( \\bar{R} \\)Ground truth
Beauty is in the eye of the beholder.0.4800.4711✓
A fool and his money are soon parted.0.1760.1402
Form follows function.0.0510.0463
Everything is worth what its purchaser will pay for it.0.0400.0584
Because it's there.0.0380.0295
You can't fix stupid.0.0210.0346
The lady doth protest too much, methinks.0.0220.0137
It's all about the money.0.0200.0138
Anybody driving slower than you is an idiot, and anyone going faster than you is a maniac?0.0120.0189
Opportunity is missed by most people.0.0180.00810
", + "bbox": [ + 146, + 231, + 848, + 360 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/50d963228235f11772ef629de780327ca9a770d9d496f5bd0aa2fcdd8470d501.jpg", + "table_caption": [ + "(iii) Qualitative Example-C from quote recommendation." + ], + "table_footnote": [], + "table_body": "
TurnConversation Threading History
#1Society is becoming more efficient, which is a good thing. People should realize there's no point in holding back this technology just for the sake of keeping people employed. If this were beneficial, then calculators and computers shouldn't exist either.
#2One small problem is that people need to pay rent and eat.
#3So we should ditch computers and go back to the typing pool? Should we get rid of heavy earth moving equipment and just use hundreds of guys with hand tools to build everything? It would employ a hell of a lot more people.
#4No one's saying that. I don't think anyone is really against automation, but as it increases, there are soon going to be more people that there are jobs that actually need doing. I actually believe we've already passed this point. So what do we do with the people, who can't get jobs simply because there are none? It's an issue that need assessed immediately.
#5Tons and tons and tons of American jobs have been replaced by new jobs created by technology or in support of technology years ago. An office might have needed people to handle filing paperwork, keeping it in order, and retrieving, where now a document management system has made them completely redundant. The upshot is that to access that DMS, people are out there selling computers, installing computers, servicing computers, and supporting end users building the servers installing, supporting monitoring backing them up, and all that jobs that come in support of those progress is progress. And it advances human efficiency and knowledge. These are just one or two examples, but the answer is not to kill progress. Other countries simply won't. The answer is to push education to the forefront, so people are prepared for these jobs and whatever other challenges the future may bring.
#6This is true. But it's unfortunate technological advances tend to reduce low skill jobs and replace them with high skill jobs. It would feel more fair if the low skilled workers could all do training programs and become high skilled workers. But this isn't really the case. Those jobs end up being taken by someone who had better educational opportunities or someone younger who still has time to take advantage of education.
#7The reality is the reality. Unfortunately or not educating people will create more educated people to handle high skill jobs, and I'll tell you being a desktop support technician isn't high skill. As that's where we push in the future, any amount of hand wringing won't change the facts. We must educate our people if we want to be a global leader in more than homelessness poverty.
#8Education won't matter. We are at the end of the job age at some point in the near future. We are going to have to deal with the fact that getting a job isn't a reality for a significant percentage of the population. Society will have to radically change as it did during the industrial revolution.
#9Much cheaper to heavily discourage having more children free abortions. Then in years there won't be so many useless people who can apparently be replaced by a simple robot.
#10Virtually every job will be replaced by automation name skilled trades that can't be automated. I imagine you'd be surprised at how hard this is. Are pharmacists useless, surgeons, accountants? I'd bet that your job is just as replaceable as these.
", + "bbox": [ + 115, + 403, + 878, + 734 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/895d3eb52d984a32d554e29ffc6f140ebfd2e1ce8023b632c7a3e11ed7c0bd71.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Candidate Quote TextsSdSpRGround truth
There's no such thing as a free lunch.0.3650.4171
I can't predict the future.0.1850.2102✓
I have never let my schooling interfere with my education.0.1040.0593
Prevention is better than cure.0.0440.0834
Knowledge is power.0.0590.0525
Don't let schooling interfere with your education.0.0440.0436
Nature abhors a vacuum.0.0360.0247
There is no substitute for hard work.0.0240.0178
There are three kinds of lies: lies, damned lies, and statistics.0.0220.0139
You can't fix stupid.0.0190.01010
", + "bbox": [ + 223, + 743, + 773, + 873 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "(iv) Qualitative Example-D from quote recommendation.", + "bbox": [ + 304, + 878, + 690, + 891 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 5: Case analyses of quote recommendation. We demonstrate the candidate quotes of the top 10 rankings out of all candidates. Note that there is only one ground truth quote for each conversation history.", + "bbox": [ + 112, + 914, + 880, + 942 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "A For every submission:", + "bbox": [ + 115, + 107, + 322, + 122 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "A1. Did you describe the limitations of your work? Section 5", + "A2. Did you discuss any potential risks of your work? We see no concern about potential risks.", + "A3. Do the abstract and introduction summarize the paper's main claims? Section 1", + "A4. Have you used AI writing assistants when working on this paper? Left blank." + ], + "bbox": [ + 129, + 126, + 695, + 287 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "B Did you use or create scientific artifacts?", + "bbox": [ + 114, + 299, + 489, + 316 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The Abstract provides the link to our code.", + "bbox": [ + 132, + 321, + 448, + 336 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "B1. Did you cite the creators of artifacts you used? Not applicable. Left blank.", + "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Not applicable. Left blank.", + "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Not applicable. Left blank.", + "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Not applicable. Left blank.", + "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? In the Abstract, a Github repository with documentation is released.", + "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Appendix A" + ], + "bbox": [ + 127, + 347, + 880, + 753 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "C Did you run computational experiments?", + "bbox": [ + 114, + 764, + 492, + 781 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Section 3", + "bbox": [ + 132, + 787, + 205, + 800 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Section 3", + "bbox": [ + 129, + 812, + 880, + 859 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.", + "bbox": [ + 112, + 866, + 877, + 889 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "ACL 2023 Responsible NLP Checklist", + "bbox": [ + 132, + 84, + 433, + 99 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1169", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Section 3", + "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Section 3", + "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Not applicable. Left blank." + ], + "bbox": [ + 127, + 83, + 880, + 282 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank.", + "bbox": [ + 112, + 293, + 877, + 330 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Not applicable. Left blank.", + "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Not applicable. Left blank.", + "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Not applicable. Left blank.", + "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Not applicable. Left blank.", + "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Not applicable. Left blank." + ], + "bbox": [ + 127, + 340, + 880, + 640 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1170", + "bbox": [ + 482, + 927, + 521, + 940 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/UniTRec_ A Unified Text-to-Text Transformer and Joint Contrastive Learning Framework for Text-based Recommendation/31a35e47-82fb-45ce-aa23-0bc4dd29fd03_model.json b/2023/UniTRec_ A Unified Text-to-Text Transformer and Joint Contrastive Learning Framework for Text-based Recommendation/31a35e47-82fb-45ce-aa23-0bc4dd29fd03_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8fe9c2622e4003c8e440f29eeebc34655869801a --- /dev/null +++ b/2023/UniTRec_ A Unified Text-to-Text Transformer and Joint Contrastive Learning Framework for Text-based Recommendation/31a35e47-82fb-45ce-aa23-0bc4dd29fd03_model.json @@ -0,0 +1,2026 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.146, + 0.082, + 0.853, + 0.121 + ], + "angle": 0, + "content": "UniTRec: A Unified Text-to-Text Transformer and Joint Contrastive Learning Framework for Text-based Recommendation" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.13, + 0.791, + 0.148 + ], + "angle": 0, + "content": "Zhiming Mao\\(^{1,2}\\), Huimin Wang\\(^{1,3}\\), Yiming Du\\(^{1,2}\\), Kam-Fai Wong\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.26, + 0.149, + 0.743, + 0.165 + ], + "angle": 0, + "content": "1The Chinese University of Hong Kong, Hong Kong, China" + }, + { + "type": "text", + "bbox": [ + 0.207, + 0.165, + 0.798, + 0.182 + ], + "angle": 0, + "content": "\\(^{2}\\)MoE Key Laboratory of High Confidence Software Technologies, China" + }, + { + "type": "text", + "bbox": [ + 0.346, + 0.183, + 0.658, + 0.198 + ], + "angle": 0, + "content": "\\(^{3}\\)Jarvis Lab, Tencent, Shenzhen, China" + }, + { + "type": "text", + "bbox": [ + 0.305, + 0.2, + 0.701, + 0.215 + ], + "angle": 0, + "content": "{zmmao,ydu,kfwong}@se.cuhk.edu.hk" + }, + { + "type": "text", + "bbox": [ + 0.381, + 0.217, + 0.624, + 0.231 + ], + "angle": 0, + "content": "hmmmwang@tencent.com" + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.253, + 0.341, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.279, + 0.461, + 0.591 + ], + "angle": 0, + "content": "Prior study has shown that pretrained language models (PLM) can boost the performance of text-based recommendation. In contrast to previous works that either use PLM to encode user history as a whole input text, or impose an additional aggregation network to fuse multi-turn history representations, we propose a unified local- and global-attention Transformer encoder to better model two-level contexts of user history. Moreover, conditioned on user history encoded by Transformer encoders, our framework leverages Transformer decoders to estimate the language perplexity of candidate text items, which can serve as a straightforward yet significant contrastive signal for user-item text matching. Based on this, our framework, UniTRec, unifies the contrastive objectives of discriminative matching scores and candidate text perplexity to jointly enhance text-based recommendation. Extensive evaluation shows that UniTRec delivers SOTA performance on three text-based recommendation tasks.\\(^{1}\\)" + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.603, + 0.26, + 0.618 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.628, + 0.49, + 0.772 + ], + "angle": 0, + "content": "Text-based recommendation (Li et al., 2010; Gu et al., 2016; Okura et al., 2017; Malkiel et al., 2020) aims to recommend relevant textual content (e.g., news articles, Twitter posts) to people based on their behaviors as represented in historical log texts. For instance, engagement recommendation (Cheng et al., 2022) on social media (e.g., Twitter and Reddit) helps users discover and engage with interested threads by modeling their browsing history." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.773, + 0.49, + 0.886 + ], + "angle": 0, + "content": "Pretrained language models (Devlin et al., 2019; Brown et al., 2020) have made waves in recent text-based recommendation research (Zhang et al., 2021; Qi et al., 2022; Geng et al., 2022). The most common practice is using PLM encoders (BERT family) to learn representations of user history and candidate item texts. Recommendation" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.254, + 0.885, + 0.317 + ], + "angle": 0, + "content": "matching scores are computed over the user and item representations and finally optimized by noise contrastive estimation (NCE) loss (Gutmann and Hyvarinen, 2010) for ranking multiple candidates." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.318, + 0.885, + 0.641 + ], + "angle": 0, + "content": "Unlike encoding single text, using PLM to encode multi-turn texts of user history is nontrivial. Existing works (Malkiel et al., 2020; Qi et al., 2022; Geng et al., 2022) concatenate multi-turn history texts as a whole input text, then use one PLM encoder to learn the holistic user representation. This is a standard PLM encoding manner but ignores the relation among history turns, as all word tokens from different history turns are equally attended2. In contrast, previous studies point out that learning the relation among user history turns is also beneficial (Zeng et al., 2020; Qi et al., 2021). Another approach is using PLM encoders to learn representations from multi-turn history texts, followed by an additional aggregation network to fuse the multi-turn representations (Wu et al., 2021; Li et al., 2022). However, the imposed aggregation networks (with newly initialized parameters) weaken the representation power of PLM encoders which are already pretrained on large-scale corpora." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.642, + 0.884, + 0.803 + ], + "angle": 0, + "content": "This work introduces UniTRec, a Unified text-to-text Transformer framework for text-based Recommendation. In the encoder component of UniTRec, we design local- and global-attention to learn user history representations through tailored attention masking, which aims to jointly model word-level and turn-level relations of user history. UniTRec can utilize the full power of PLM encoders because it preserves the intact structure of PLM encoders without newly imposed parameters." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.804, + 0.884, + 0.869 + ], + "angle": 0, + "content": "Different from most previous works that predict user-candidate matching scores solely based on the representations learned by Transformer encoders, we argue that conditioned on user representations" + }, + { + "type": "page_footnote", + "bbox": [ + 0.508, + 0.881, + 0.883, + 0.919 + ], + "angle": 0, + "content": "2There is no inductive bias of turn-level and history-level relations introduced to Transformer self-attention computation, where each token plays an equal role." + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.893, + 0.49, + 0.918 + ], + "angle": 0, + "content": "1Our code is available at https://github.com/Veason-silverbullet/UniTRec." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1160" + }, + { + "type": "footer", + "bbox": [ + 0.227, + 0.946, + 0.77, + 0.958 + ], + "angle": 0, + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + }, + { + "type": "footer", + "bbox": [ + 0.369, + 0.959, + 0.63, + 0.972 + ], + "angle": 0, + "content": "Volume 2: Short Papers, pages 1160-1170" + }, + { + "type": "footer", + "bbox": [ + 0.297, + 0.973, + 0.702, + 0.986 + ], + "angle": 0, + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.116, + 0.084, + 0.5, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.251, + 0.489, + 0.293 + ], + "angle": 0, + "content": "Figure 1: An example of perplexity-based ranking for candidate item texts, conditioned on user history. The illustrated task is text-based news recommendation." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.321, + 0.49, + 0.594 + ], + "angle": 0, + "content": "learned by Transformer encoders, candidate text perplexity (PPL) estimated by pretrained Transformer decoders is also a straightforward yet significant signal for text-based recommendation. As shown in Figure 1, we hypothesize that the candidate text perplexity estimated by pretrained LM decoders can directly measure the text matching degree between user history and candidate texts. It is because the perplexity estimates the likelihood of candidate texts based on encoder outputs, which naturally indicates the probabilities of candidate texts given the user history. Besides, UniTRec can use the last hidden states of Transformer decoders to directly predict matching scores. Hence, this work unifies the contrastive objectives of discriminative matching scores and candidate text perplexity to jointly enhance text-based recommendation." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.596, + 0.49, + 0.756 + ], + "angle": 0, + "content": "The contributions of this work are: (1) We propose local- and global-attention to model two-level relation of user history without additional parameters, which enjoys the full power of PLM encoders. (2) We introduce PLM perplexity to measure user-candidate text matching and unify the objectives of discriminative matching scores and candidate text perplexity to enhance text-based recommendation. (3) Experiments on three text-based recommendation datasets validate the effectiveness of UniTRec." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.771, + 0.237, + 0.787 + ], + "angle": 0, + "content": "2 Approach" + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.798, + 0.4, + 0.814 + ], + "angle": 0, + "content": "2.1 Unified User-history Modeling" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.82, + 0.489, + 0.902 + ], + "angle": 0, + "content": "Formally, multi-turn history of a user is represented as \\( H = [t_1, t_2, \\dots, t_N] \\), and each turn text \\( t_i \\) contains \\( |t_i| \\) words as \\( t_i = [x_i^1, x_i^2, \\dots, x_i^{|t_i|}] \\). UniTRec aims to unify learning word- and turn-level context representations in one Transformer encoder." + }, + { + "type": "text", + "bbox": [ + 0.132, + 0.904, + 0.488, + 0.918 + ], + "angle": 0, + "content": "Local attention on word-level context. We first" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.215 + ], + "angle": 0, + "content": "concatenate the multi-turn history texts as the input tokens \\( X = [x_{1}^{1}, x_{1}^{2}, \\dots, x_{1}^{|t_{1}|}, \\dots, x_{N}^{1}, x_{N}^{2}, \\dots, x_{N}^{|t_{N}|}] \\). Inspired by Dong et al. (2019), we tailor the attention masking in Transformer self-attention to learn the word-level context of each turn. Specifically, we allow word tokens from the same turn to attend to each other, while tokens from different turns are excluded from self-attention computation:" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.219, + 0.877, + 0.261 + ], + "angle": 0, + "content": "\\(\\mathbf{M}_{i,j} = \\left\\{ \\begin{array}{ll}0, & \\mathrm{token} x_i\\mathrm{and}x_j\\mathrm{in the same turn}\\\\ -\\infty , & \\mathrm{otherwise} \\end{array} \\right.\\)" + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.262, + 0.883, + 0.311 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}} + \\mathbf {M}\\right) V \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.314, + 0.884, + 0.409 + ], + "angle": 0, + "content": ", where \\( Q, K, V \\) are self-attention query, key, and value in Vaswani et al. (2017), \\( \\mathbf{M} \\) is the mask matrix to achieve local-attention inside each turn text. The local self-attention blocks consist of \\( L_{1} \\) layers, by which original PLM encoders can be adapted to learn word-level context representations of turns." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.411, + 0.884, + 0.587 + ], + "angle": 0, + "content": "Global attention on turn-level context. Over the local self-attention layers, we leverage global self-attention to model the relation among history turns. Specifically, tokens from all turns attend to each other in self-attention computation (by setting the mask matrix \\(\\mathbf{M} = \\mathbf{0}\\)). In this way, Transformer encoders can perform global interaction among each token (and turn) to learn turn-level context representations of user history. There are \\(L_{2}\\) layers in the global self-attention blocks, which can also be inherited from PLM encoders directly." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.598, + 0.851, + 0.614 + ], + "angle": 0, + "content": "2.2 Joint Contrastive Ranking Objectives" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.619, + 0.884, + 0.714 + ], + "angle": 0, + "content": "Conditioned on the history representation, we input the candidate text to Transformer decoders to predict how likely it should be recommended. It is worth noting that Transformer decoders can naturally perform effective cross-attention interaction between history and candidate hidden states." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.724, + 0.845, + 0.739 + ], + "angle": 0, + "content": "2.2.1 Objective on Discriminative Scores" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.743, + 0.883, + 0.838 + ], + "angle": 0, + "content": "Motivated by Lewis et al. (2020), we feed the last hidden state of decoder output \\( h_{T} \\) to an MLP scorehead which predicts the user-candidate matching score \\( S^{d} = \\mathrm{ScoreHead}(h_{T}) \\). The matching score is discriminative, as higher scores indicate higher user-candidate matching probabilities." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.84, + 0.884, + 0.919 + ], + "angle": 0, + "content": "Following previous works (Li et al., 2022; Qi et al., 2022), we adopt negative sampling with NCE loss to optimize matching score prediction. Given the user history and its ground truth matched candidate \\( C_i \\), UniTRec predicts the matching score" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.928, + 0.519, + 0.941 + ], + "angle": 0, + "content": "1161" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.199, + 0.081, + 0.803, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.263, + 0.885, + 0.294 + ], + "angle": 0, + "content": "Figure 2: Overview of UniTRec. In training, matching scores \\( S^d \\) and \\( S^p \\) are optimized by the NCE loss, respectively. In inference, \\( S^d \\) and \\( S^p \\) are normalized and combined to derive the final output ranking." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.315, + 0.489, + 0.384 + ], + "angle": 0, + "content": "as \\(S_{i}^{d + }\\) . In addition, \\(K\\) unmatched negative candidates \\(\\{C_j\\}_{j = 1}^K\\) are sampled from the candidate set, and their matching scores are \\(\\{S_j^{d - }\\}_{j = 1}^K\\) . The NCE loss is represented in a contrastive form:" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.386, + 0.488, + 0.429 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i} ^ {d} = - \\log \\frac {\\exp (S _ {i} ^ {d +})}{\\exp (S _ {i} ^ {d +}) + \\sum_ {j = 1} ^ {K} \\exp (S _ {j} ^ {d -})} \\quad (2)\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.441, + 0.483, + 0.457 + ], + "angle": 0, + "content": "2.2.2 Objective on Candidate Text Perplexity" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.465, + 0.489, + 0.562 + ], + "angle": 0, + "content": "As aforementioned, UniTRec leverages perplexity to rank candidate texts. Since lower perplexity indicates higher user-candidate matching probability, regarding the candidate text \\( Y = [y_{1}, y_{2}, \\dots, y_{T}] \\), we define the perplexity-based matching score \\( S^{p} \\) as its negative perplexity:" + }, + { + "type": "equation", + "bbox": [ + 0.125, + 0.569, + 0.488, + 0.598 + ], + "angle": 0, + "content": "\\[\nS ^ {p} = - \\operatorname {P P L} (Y) = \\frac {1}{T} \\sum_ {i = 1} ^ {T} \\log p _ {\\theta} \\left(y _ {i} \\mid y _ {< i}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.606, + 0.489, + 0.719 + ], + "angle": 0, + "content": ", where \\( p_{\\theta}(\\cdot) \\) denotes the target probability output from the UniTRec Transformer decoder. Similar to Eq. (2), we optimize the perplexity-based matching score \\( S^p \\) in the NCE loss form. As perplexity empirically varies in a wide range, we introduce a temperature parameter \\( \\tau \\) to balance the joint NCE loss gradients following Radford et al. (2021)." + }, + { + "type": "equation", + "bbox": [ + 0.119, + 0.726, + 0.486, + 0.778 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i} ^ {p} = - \\log \\frac {\\exp \\left(\\tau \\cdot S _ {i} ^ {p +}\\right)}{\\exp \\left(\\tau \\cdot S _ {i} ^ {p +}\\right) + \\sum_ {j = 1} ^ {K} \\exp \\left(\\tau \\cdot S _ {j} ^ {p -}\\right)} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.78, + 0.488, + 0.827 + ], + "angle": 0, + "content": ", where \\(\\tau\\) is learnable and initialized to 1. On the training dataset \\(\\mathcal{D}\\), the joint contrastive learning objective is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.83, + 0.488, + 0.859 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\sum_ {i = 1} ^ {| \\mathcal {D} |} \\left(\\mathcal {L} _ {i} ^ {d} + \\mathcal {L} _ {i} ^ {p}\\right) \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.317, + 0.825, + 0.332 + ], + "angle": 0, + "content": "2.3 Model Initialization and Inference" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.337, + 0.884, + 0.466 + ], + "angle": 0, + "content": "As UniTRec is a standard text-to-text Transformer, we initialize the parameters from pretrained BART (Lewis et al., 2020). In inference, UniTRec predicts the discriminative and perplexity-based scores for each candidate item, respectively. The two separate scores \\( S^d \\) and \\( S^p \\) are normalized, averaged, and finally ranked as the output. Detailed ranking process is provided in Appendix B." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.477, + 0.656, + 0.493 + ], + "angle": 0, + "content": "3 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.501, + 0.884, + 0.694 + ], + "angle": 0, + "content": "We evaluate UniTRec on three text-based recommendation tasks: 1) NewsRec, to recommend news articles to users based on their browsing history. We use the MIND-small dataset (Wu et al., 2020) for experiments. 2) QuoteRec, to recommend quotations to users based on their conversation history. We use the Reddit-quotation dataset (Wang et al., 2021) for experiments. 3) EngageRec, to recommend social media posts for users to engage with based on their comment history. We use the dataset released by Zeng et al. (2020) for experiments. Detailed dataset statistics is provided in Appendix A." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.695, + 0.883, + 0.79 + ], + "angle": 0, + "content": "Implementation Details. The UniTRec encoder and decoder both consist of 6 Transformer layers with 768-dimensional hidden states and 12 attention heads. We set \\( L_{1} = 3 \\) and \\( L_{2} = 3 \\). We use AdamW optimizer (Loshchilov and Hutter, 2019) to train UniTRec with cosine learning rate decay." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.791, + 0.884, + 0.919 + ], + "angle": 0, + "content": "Baselines. We compare UniTRec with competitive baselines: 1) GRU4Rec (Balázs et al., 2016) utilizes a GRU network to learn multi-turn history. 2) SASRec (Kang and McAuley, 2018) encodes user history with a self-attention based sequential model. 3) BERT4Rec (Sun et al., 2019) employs bidirectional self-attention to model user history. 4) RoBERTa-Sim, a simple yet strong baseline men" + }, + { + "type": "page_footnote", + "bbox": [ + 0.113, + 0.869, + 0.488, + 0.919 + ], + "angle": 0, + "content": "3Note https://huggingface.co/docs/transformers/perplexity for LM perplexity calculation. We empirically discard the outer exponential term in the PPL formula, because it already exists in NCE loss Eq. (4) and does not affect the final ranking." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1162" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.12, + 0.08, + 0.878, + 0.177 + ], + "angle": 0, + "content": "
ModelNewsRecQuoteRecEngageRec
MRRNDCG@5/10HR@5/10MRRNDCG@5/10HR@5/10MRRNDCG@5/10HR@5/10
GRU4Rec32.9136.20/42.5350.33/68.3534.0834.65/37.9344.45/54.632.121.04/1.511.27/2.65
SASRec32.6036.03/42.3750.63/68.6433.6334.30/37.4944.32/54.202.401.49/1.952.16/3.47
BERT4Rec32.8736.18/42.4050.21/67.9733.5934.26/37.2743.76/53.053.041.98/3.232.81/6.67
RoBERTa-Sim32.9636.47/42.8151.06/69.0837.1337.96/41.1848.14/58.063.742.66/3.754.42/7.70
UNBERT33.0936.53/42.8450.87/68.8239.7540.74/43.6950.90/60.042.831.96/2.673.11/5.24
UniTRec33.7637.63/43.7452.61/69.8941.2442.38/45.3152.87/61.884.063.23/4.294.58/7.68
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.187, + 0.882, + 0.232 + ], + "angle": 0, + "content": "Table 1: Experiment results on three text-based recommendation tasks. MRR denotes mean reciprocal rank, NDCG denotes normalized discounted cumulative gain, and HR denotes hit ratio (presented in percentage). The overall performance of UniTRec is better than other baseline models with \\( p \\)-value \\( < 0.05 \\), validated by unpaired t-test." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.243, + 0.878, + 0.34 + ], + "angle": 0, + "content": "
ModelNewsRecQuoteRecEngageRec
MRRNDCG@5/10HR@5/10MRRNDCG@5/10HR@5/10MRRNDCG@5/10HR@5/10
UniTRec33.7637.63/43.7452.61/69.8941.2442.38/45.3152.87/61.884.063.23/4.294.58/7.68
w/o BART Init30.3133.32/39.6947.55/65.7819.0217.66/20.8022.45/32.162.240.86/1.611.27/3.62
w/o Local-Att33.3437.22/43.3252.28/69.5440.4441.63/44.5652.09/61.153.923.19/4.154.38/7.36
w/o Global-Att33.2237.06/43.1752.14/69.4740.2541.47/44.2652.07/60.763.642.78/3.593.89/6.35
Disc-Score only33.0736.76/43.0351.68/69.4640.5941.81/44.6552.39/61.143.822.99/3.604.49/6.85
PPL-Score only32.8336.39/42.5951.05/68.6740.3141.43/44.4752.13/61.203.292.39/3.033.86/5.66
" + }, + { + "type": "table_caption", + "bbox": [ + 0.271, + 0.35, + 0.725, + 0.364 + ], + "angle": 0, + "content": "Table 2: Recommendation performance of ablation model variants." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.39, + 0.489, + 0.485 + ], + "angle": 0, + "content": "tioned in Qi et al. (2022), uses the hidden states of [CLS] tokens to measure user-candidate similarity. 5) UNBERT, implemented as Zhang et al. (2021), concatenates history and candidate texts as the input to BERT and predicts matching scores from the final hidden states of [CLS] tokens." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.487, + 0.49, + 0.567 + ], + "angle": 0, + "content": "Note that we do not consider other methods that use non-text inputs (e.g., user profile, text topic labels). For fair comparison, all baseline models use pretrained 12-layer RoBERTa-base (Liu et al., 2019) as text encoders to learn embeddings of texts." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.578, + 0.265, + 0.592 + ], + "angle": 0, + "content": "3.1 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.599, + 0.49, + 0.856 + ], + "angle": 0, + "content": "Table 1 shows the performance of experiment models. From the results of NewsRec and QuoteRec, we can see that UniTRec outperforms all baseline models by a clear margin. Also, RoBERTa-Sim and UNBERT that directly use the [CLS] hidden states to represent user history, surpass other baselines that build additional aggregation networks upon the whole RoBERTa outputs. As displayed in the results, EngageRec is the most difficult task. We inspect the dataset and find that the texts on social media contain too much noise (e.g., URL and emoji), and the user history contains less number of turns. Nevertheless, UniTRec achieves better overall performance than other baseline models, validating its robustness on noisy text inputs and limited user history." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.867, + 0.4, + 0.882 + ], + "angle": 0, + "content": "3.2 Ablation Studies and Analyses" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.888, + 0.489, + 0.919 + ], + "angle": 0, + "content": "We further conduct ablation studies on UniTRec. The experiment results are reported in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.39, + 0.884, + 0.502 + ], + "angle": 0, + "content": "Initialization of UniTRec. We train UniTRec from scratch without initialization from pretrained BART (refer to w/o BART Init). The recommendation performance significantly drops in all three tasks, which indicates that acquiring effective text understanding ability from PLM is a necessary key to UniTRec performance." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.51, + 0.885, + 0.703 + ], + "angle": 0, + "content": "Local and global attention. We investigate the function of two-level attention modules of the UniTRec history encoder. Concretely, we set \\( L_{1} = 0 \\) in w/o Local-Att and \\( L_{2} = 0 \\) in w/o Global-Att, where \\( L_{1} + L_{2} = 6 \\). We can observe that removing local and global attention from the original UniTRec history encoder both lead to suboptimal performance, while the performance drop is more significant in w/o Global-Att. The results justify the effectiveness of jointly modeling two-level history contexts through adapted Transformer attention masking without additional parameters." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.71, + 0.884, + 0.919 + ], + "angle": 0, + "content": "Discriminative and perplexity-based objectives. We probe into training UniTRec with standalone discriminative (Disc-Score only) and perplexity-based (PPL-Score only) contrastive objectives, respectively. We can see that the discriminative objective yields better performance than the perplexity-based objective. Besides, the model performance on both standalone objectives declines compared to the original joint objective. The results indicate that the discriminative and perplexity-based matching scores are complementary and can jointly provide more accurate signals of user history and candidate text matching for text-based recommendation." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1163" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.114, + 0.085, + 0.248, + 0.099 + ], + "angle": 0, + "content": "4 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.111, + 0.49, + 0.222 + ], + "angle": 0, + "content": "We present a unified Transformer UniTRec for text-based recommendation. UniTRec learns two-level contexts of multi-turn user history and jointly exploits discriminative matching scores and candidate text perplexity as matching objectives. Empirical experiments on three text-based recommendation datasets corroborate the effectiveness of UniTRec." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.235, + 0.251, + 0.25 + ], + "angle": 0, + "content": "5 Limitations" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.261, + 0.491, + 0.518 + ], + "angle": 0, + "content": "Our model only focuses on utilizing text information for recommendation, which is a key limitation of this work. In real-world settings, recommender systems are usually required to handle heterogeneous information inputs. UniTRec is a pure text-based recommender modeling user history and candidate texts as inputs. However, incorporating additional side information (e.g., user profile, text topic labels, and dwell time of user behaviors) could further improve the recommendation performance and alleviate the cold start problem. Furthermore, UniTRec only models two-level relations of user behavior history. Nonetheless, incorporating more user behavior information, such as implicit and negative feedback, could further enhance the recommendation performance." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.532, + 0.287, + 0.547 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.557, + 0.49, + 0.62 + ], + "angle": 0, + "content": "We appreciate constructive comments from anonymous reviewers. The research described in this paper is partially supported by CUHK under Project No. 3230366." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.649, + 0.214, + 0.663 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.672, + 0.49, + 0.752 + ], + "angle": 0, + "content": "Hidasi Balázs, Karatzoglou Alexandros, Baltrunas Linas, and Tikk Domonkos. 2016. Session-based recommendations with recurrent neural networks. In 4th International Conference on Learning Representations ICLR 2016, San Juan, Puerto Rico, May 2-4, 2016, Conference Track Proceedings." + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.761, + 0.491, + 0.919 + ], + "angle": 0, + "content": "Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems," + }, + { + "type": "list", + "bbox": [ + 0.116, + 0.672, + 0.491, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.528, + 0.086, + 0.884, + 0.112 + ], + "angle": 0, + "content": "volume 33, pages 1877-1901. Curran Associates, Inc." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.123, + 0.886, + 0.203 + ], + "angle": 0, + "content": "Daniel Cheng, Kyle Yan, Phillip Keung, and Noah A. Smith. 2022. The engage corpus: A social media dataset for text-based recommender systems. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 1885-1889, Marseille, France. European Language Resources Association." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.213, + 0.885, + 0.344 + ], + "angle": 0, + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 4171-4186. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.354, + 0.885, + 0.433 + ], + "angle": 0, + "content": "Li Dong, Nan Yang, Wenhui Wang, Furu Wei, Xiaodong Liu, Yu Wang, Jianfeng Gao, Ming Zhou, and Hsiao-Wuen Hon. 2019. Unified language model pre-training for natural language understanding and generation. In 33rd Conference on Neural Information Processing Systems (NeurIPS 2019)." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.444, + 0.886, + 0.537 + ], + "angle": 0, + "content": "Shijie Geng, Shuchang Liu, Zuohui Fu, Yingqiang Ge, and Yongfeng Zhang. 2022. Recommendation as language processing (rlp): A unified pretrain, personalized prompt & predict paradigm (p5). In Proceedings of the 16th ACM Conference on Recommender Systems, RecSys '22, page 299-315, New York, NY, USA. Association for Computing Machinery." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.545, + 0.884, + 0.626 + ], + "angle": 0, + "content": "Youyang Gu, Tao Lei, Regina Barzilay, and Tommi Jaakkola. 2016. Learning to refine text based recommendations. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2103-2108, Austin, Texas. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.635, + 0.885, + 0.728 + ], + "angle": 0, + "content": "Michael Gutmann and Aapo Hyvarinen. 2010. Noise-contrastive estimation: A new estimation principle for unnormalized statistical models. In Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics, volume 9 of Proceedings of Machine Learning Research, pages 297-304, Chia Laguna Resort, Sardinia, Italy. PMLR." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.737, + 0.885, + 0.791 + ], + "angle": 0, + "content": "Wang-Cheng Kang and Julian McAuley. 2018. Self-attentive sequential recommendation. In 2018 IEEE International Conference on Data Mining (ICDM), pages 197-206." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.801, + 0.886, + 0.919 + ], + "angle": 0, + "content": "Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871-7880, Online. Association for Computational Linguistics." + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.086, + 0.886, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1164" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.179 + ], + "angle": 0, + "content": "Jian Li, Jieming Zhu, Qiwei Bi, Guohao Cai, Lifeng Shang, Zhenhua Dong, Xin Jiang, and Qun Liu. 2022. MINER: Multi-interest matching network for news recommendation. In Findings of the Association for Computational Linguistics: ACL 2022, pages 343-352, Dublin, Ireland. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.189, + 0.49, + 0.255 + ], + "angle": 0, + "content": "Yize Li, Jiazhong Nie, Yi Zhang, Bingqing Wang, Baoshi Yan, and Fuliang Weng. 2010. Contextual recommendation based on text mining. In *Coling* 2010: Posters, pages 692-700, Beijing, China. Coling 2010 Organizing Committee." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.265, + 0.49, + 0.332 + ], + "angle": 0, + "content": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. In arXiv preprint arXiv: 1907.11692. arXiv." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.341, + 0.49, + 0.406 + ], + "angle": 0, + "content": "Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.417, + 0.49, + 0.497 + ], + "angle": 0, + "content": "Itzik Malkiel, Oren Barkan, Avi Caciularu, Noam Razin, Ori Katz, and Noam Koenigstein. 2020. *RecoBERT: A catalog language model for text-based recommendations*. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.507, + 0.49, + 0.6 + ], + "angle": 0, + "content": "Shumpei Okura, Yukihiro Tagami, Shingo Ono, and Akira Tajima. 2017. Embedding-based news recommendation for millions of users. In Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, page 1933-1942, New York, NY, USA. Association for Computing Machinery." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.608, + 0.49, + 0.702 + ], + "angle": 0, + "content": "Fanchao Qi, Yanhui Yang, Jing Yi, Zhili Cheng, Zhiyuan Liu, and Maosong Sun. 2022. QuoteR: A benchmark of quote recommendation for writing. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 336-348, Dublin, Ireland. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.711, + 0.49, + 0.83 + ], + "angle": 0, + "content": "Tao Qi, Fangzhao Wu, Chuhan Wu, Peiru Yang, Yang Yu, Xing Xie, and Yongfeng Huang. 2021. HieRec: Hierarchical user interest modeling for personalized news recommendation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 5446-5456, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.84, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International" + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.529, + 0.086, + 0.882, + 0.125 + ], + "angle": 0, + "content": "Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 8748-8763. PMLR." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.136, + 0.885, + 0.241 + ], + "angle": 0, + "content": "Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang. 2019. Bert4rec: Sequential recommendation with bidirectional encoder representations from transformer. In Proceedings of the 28th ACM International Conference on Information and Knowledge Management, CIKM '19, page 1441-1450, New York, NY, USA. Association for Computing Machinery." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.25, + 0.885, + 0.33 + ], + "angle": 0, + "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30, pages 5998-6008. Curran Associates, Inc." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.339, + 0.885, + 0.445 + ], + "angle": 0, + "content": "Lingzhi Wang, Xingshan Zeng, and Kam-Fai Wong. 2021. Quotation recommendation and interpretation based on transformation from queries to quotations. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 754-758, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.454, + 0.885, + 0.547 + ], + "angle": 0, + "content": "Chuhan Wu, Fangzhao Wu, Tao Qi, and Yongfeng Huang. 2021. Empowering news recommendation with pre-trained language models. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '21, page 1652-1656, New York, NY, USA. Association for Computing Machinery." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.556, + 0.885, + 0.661 + ], + "angle": 0, + "content": "Fangzhao Wu, Ying Qiao, Jiun-Hung Chen, Chuhan Wu, Tao Qi, Jianxun Lian, Danyang Liu, Xing Xie, Jianfeng Gao, Winnie Wu, and Ming Zhou. 2020. MIND: A large-scale dataset for news recommendation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3597-3606, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.671, + 0.885, + 0.75 + ], + "angle": 0, + "content": "Xingshan Zeng, Jing Li, Lu Wang, Zhiming Mao, and Kam-Fai Wong. 2020. Dynamic online conversation recommendation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3331-3341, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.759, + 0.885, + 0.852 + ], + "angle": 0, + "content": "Qi Zhang, Jingjie Li, Qinglin Jia, Chuyuan Wang, Jieming Zhu, Zhaowei Wang, and Xiuqiang He. 2021. Umbert: User-news matching bert for news recommendation. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-21, pages 3356-3362. International Joint Conferences on Artificial Intelligence Organization. Main Track." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.885, + 0.852 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1165" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.118, + 0.082, + 0.498, + 0.148 + ], + "angle": 0, + "content": "
DatasetNewsRecQuoteRecEngageRec
Avg. history turns26.094.243.29
Avg. history tokens414.40279.82286.82
Avg. candidates37.2311117163
Avg. candidate tokens16.1519.11102.42
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.157, + 0.49, + 0.244 + ], + "angle": 0, + "content": "Table 3: Statistics of three text-based recommendation training datasets. History and candidate tokens denote the number of BPE-tokenized tokens. The test set distribution is closed to the training sets (except candidates of EngageRec) and hence omitted. Note that the max length of each history log is truncated to 1024 tokens." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.265, + 0.303, + 0.28 + ], + "angle": 0, + "content": "A Dataset Statistics" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.29, + 0.49, + 0.482 + ], + "angle": 0, + "content": "The detailed statistics of the three text-based recommendation datasets are displayed in Table 3. Note that we use news titles as the text inputs for NewsRec following Qi et al. (2021). NewsRec regards the user clicked and non-clicked news as candidate texts, while QuoteRec and EngageRec regard all potential quotation texts and post texts as candidates. Different from Zeng et al. (2020) that formulates the task as recommending candidate users to given posts based on post content, we formulate the task as recommending candidate posts to given users based on user history." + }, + { + "type": "title", + "bbox": [ + 0.116, + 0.495, + 0.428, + 0.511 + ], + "angle": 0, + "content": "Algorithm 1 Candidate Ranking Processes" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.515, + 0.464, + 0.542 + ], + "angle": 0, + "content": "Input: discriminative scores \\(S^d = \\{S_1^d,S_2^d,\\dots,S_M^d\\}\\) perplexity-based scores \\(S^{p} = \\{S_{1}^{p},S_{2}^{p},\\dots,S_{M}^{p}\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.542, + 0.335, + 0.553 + ], + "angle": 0, + "content": "Output: final averaged ranking \\( R \\)." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.553, + 0.486, + 0.576 + ], + "angle": 0, + "content": "1: Derive the normalized discriminative scores \\( S_{norm}^{d} = \\) softmax(Sd)." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.577, + 0.486, + 0.6 + ], + "angle": 0, + "content": "2: Derive the normalized perplexity-based scores \\( S_{norm}^{p} = \\) softmax( \\( S^p \\) )." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.601, + 0.488, + 0.625 + ], + "angle": 0, + "content": "3: Derive the geometric average scores \\(\\bar{S} = \\log (S_{norm}^d) + \\log (S_{norm}^p)\\)." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.626, + 0.486, + 0.649 + ], + "angle": 0, + "content": "4: Sort the averaged scores \\(\\bar{S}\\) by descending order to derive the final ranking: \\(\\bar{R} \\gets \\mathrm{Rank}_{\\mathrm{des}}(\\bar{S})\\)" + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.65, + 0.212, + 0.66 + ], + "angle": 0, + "content": "5: return \\( R \\)" + }, + { + "type": "list", + "bbox": [ + 0.128, + 0.553, + 0.488, + 0.66 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.689, + 0.314, + 0.705 + ], + "angle": 0, + "content": "B Inference Ranking" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.714, + 0.49, + 0.875 + ], + "angle": 0, + "content": "Given the user history and \\(M\\) candidate texts, UniTRec first predicts the discriminative ranking scores \\(S^d = \\{S_1^d,S_2^d,\\dots,S_M^d\\}\\) and perplexity-based ranking scores \\(S^{p} = \\{S_{1}^{p},S_{2}^{p},\\dots,S_{M}^{p}\\}\\) of the candidates. Algorithm 1 outlines an approach to aggregate the final ranking based on \\(S^d\\) and \\(S^p\\). Note that the function \\(\\mathrm{Rank}(S)^4\\) denotes outputting the sorted order of elements in a score list \\(S\\). There exist other ways to average the ranking of \\(S^d\\) and \\(S^p\\), which we leave for future work to explore." + }, + { + "type": "page_footnote", + "bbox": [ + 0.113, + 0.881, + 0.489, + 0.92 + ], + "angle": 0, + "content": "\\(^4\\)Rank(S) works similarly to scipy.stats.rankdata(). For example in ascending order, \\(\\mathrm{Ran_{asc}}(\\{0.2, 0.6, 0.7, 0.4\\}) = \\mathrm{scipy.stats.rankdata}([0.2, 0.6, 0.7, 0.4]) = [1, 3, 4, 2]\\)" + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.084, + 0.724, + 0.101 + ], + "angle": 0, + "content": "C Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.11, + 0.885, + 0.175 + ], + "angle": 0, + "content": "We show randomly sampled outputs of UniTRec, for instance, demonstrated on the news recommendation and quote recommendation tasks. Table 4 and 5 showcase the qualitative samples." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1166" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.117, + 0.178, + 0.88, + 0.307 + ], + "angle": 0, + "content": "
TurnHistory News Texts
#1Mac Engel: As long as these results are acceptable, Dallas Cowboys will continue to be losers
#2NFL world reacts to officials handing Packers win over Lions
#3Maryland Congressman Elijah Cummings, a Democrat and Chair of House Oversight and Reform Committee, has died: CNN
#4Unprecedented movement detected on California earthquake fault capable of 8.0 temblor
#5Bag Explodes While Being Loaded On Volaris Flight At Midway Airport
#6Orlando Scandrick rips Eagles: They have "accountability issues"
#7Meghan King Edmonds, Jim Edmonds' Nanny Denies Cheating Allegations
#8Nearly $400M worth of cocaine and marijuana intercepted by US Coast Guard
#9Former NBA first-round pick arrested in sex sting operation
#10China's trade with US shrinks in October despite optimism
" + }, + { + "type": "table", + "bbox": [ + 0.118, + 0.316, + 0.88, + 0.434 + ], + "angle": 0, + "content": "
Candidate News TextsSdSpRClicked
Taylor Swift Rep Hits Back at Big Machine, Claims She's Actually Owed $7.9 Million in Unpaid Royalties0.0950.0694X
Former North Carolina State, NBA player Anthony Grundy dies in stabbing, police say0.1720.1553X
13 Reasons Why's Christian Navarro Slams Disney for Casting "the White Guy" in The Little Mermaid0.0480.0657X
Opinion: Colin Kaepernick is about to get what he deserves: a chance0.3030.2501✓
3 Indiana judges suspended after a night of drinking turned into a White Castle brawl0.0760.0595X
66 Cool Tech Gifts Anyone Would Be Thrilled to Receive0.0090.0059X
Police find 26 children behind false wall at Colorado day care0.0340.1166X
I've been writing about tiny homes for a year and spent 2 nights in a 300-foot home to see what it is all about0.0290.0198X
Report: Police investigating woman's death after Redskins' player Montae Nicholson took her to hospital0.2350.2612✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.311, + 0.44, + 0.685, + 0.453 + ], + "angle": 0, + "content": "(i) Qualitative Example-A from news recommendation." + }, + { + "type": "table", + "bbox": [ + 0.117, + 0.476, + 0.88, + 0.582 + ], + "angle": 0, + "content": "
TurnHistory News Texts
#1Toddler dancing to celebrate 11 months cancer-free goes viral
#2NFL Week 8 Power Rankings: Old-school football rules the day
#3The 25 US cities where it's easiest to get a mortgage
#4Burning questions for Cowboys vs Giants on "Monday Night Football"
#5Who's the favorite to win 2019 NFL rushing title?
#6Grading all 32 NFL teams heading into the last eight weeks of the 2019 season
#7Jennifer Aniston looks amazing in a makeup-free selfie, plus more news
#8This $12 million "mansion yacht" is made entirely of stainless steel and it's a first for the industry. Take a peek inside
" + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.592, + 0.876, + 0.722 + ], + "angle": 0, + "content": "
Candidate News TextsSdSpRClicked
Opinion: Colin Kaepernick is about to get what he deserves: a chance0.3300.4001✓
U.S. Troops Will Die If They Remain in Syria, Bashar Al-Assad Warns0.0240.01110✗
Pete Davidson, Kaia Gerber Are Dating, Trying to Stay "Low Profile"0.0640.0336✗
The Hottest Tech Gifts This Holiday Season0.0500.0278✗
Taylor Swift Rep Hits Back at Big Machine, Claims She's Actually Owed $7.9 Million in Unpaid Royalties0.0460.0387✗
13 Reasons Why's Christian Navarro Slams Disney for Casting "the White Guy" in The Little Mermaid0.0600.0964✓
Some believe Mason Rudolph, hit in head with his own helmet, isn't getting enough blame0.1540.1792✓
South Carolina teen gets life in prison for deadly elementary school shooting0.0660.0465✗
The Unlikely Star of My Family's Thanksgiving Table0.0470.0219✗
Police investigating woman's death after Redskins' player Montae Nicholson took her to hospital0.1580.1493✗
" + }, + { + "type": "table_caption", + "bbox": [ + 0.308, + 0.727, + 0.687, + 0.74 + ], + "angle": 0, + "content": "(ii) Qualitative Example-B from news recommendation." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.762, + 0.884, + 0.82 + ], + "angle": 0, + "content": "Table 4: Case analyses of news recommendation. History News Texts are sorted by user-clicked timestamps. \\( S^d \\), \\( S^p \\), and \\( \\bar{R} \\) are normalized discriminative, perplexity-based scores, and average ranking as described in Appendix B. Clicked denotes the ground truth user-click labels. Note that the experiment history logs are anonymized and delinked, which is always the first priority of the recommendation study." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1167" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.117, + 0.082, + 0.88, + 0.222 + ], + "angle": 0, + "content": "
TurnConversation Threading History
#1I own an FJ. It's a great car and even on stockies. It's great offroad.
#2I feel bad for you that you run the risk of being associated with the typical FJ owner.
#3What is a typical FJ owner? I've not heard anything bad about FJ owners.
#4It's like someone who drives a jeep wrangler in NYC. There's no need. Tons of FJ owners do that have it and not use it for what it's made for.
#5God forbid someone likes the design of a car and doesn't use it offroad.
#6Then buy a much more economic environmentalist friendly version. If you buy something and always use it for much less than it's purpose, why buy it?
#7Or people can buy whatever the hell they want because it's their money and not yours.
#8You're entirely right. Just like people can be rude just because you can do it, because you have the ability but why should you ass.
#9I wasn't aware that somebody buying a vehicle that they like and you don't was morally wrong.
#10I love FJs. It's perfectly fine to buy whatever you think looks nice.
" + }, + { + "type": "table", + "bbox": [ + 0.147, + 0.232, + 0.85, + 0.361 + ], + "angle": 0, + "content": "
Candidate Quote Texts\\( S^d \\)\\( S^P \\)\\( \\bar{R} \\)Ground truth
Beauty is in the eye of the beholder.0.4800.4711✓
A fool and his money are soon parted.0.1760.1402
Form follows function.0.0510.0463
Everything is worth what its purchaser will pay for it.0.0400.0584
Because it's there.0.0380.0295
You can't fix stupid.0.0210.0346
The lady doth protest too much, methinks.0.0220.0137
It's all about the money.0.0200.0138
Anybody driving slower than you is an idiot, and anyone going faster than you is a maniac?0.0120.0189
Opportunity is missed by most people.0.0180.00810
" + }, + { + "type": "table_caption", + "bbox": [ + 0.305, + 0.367, + 0.692, + 0.38 + ], + "angle": 0, + "content": "(iii) Qualitative Example-C from quote recommendation." + }, + { + "type": "table", + "bbox": [ + 0.117, + 0.404, + 0.879, + 0.735 + ], + "angle": 0, + "content": "
TurnConversation Threading History
#1Society is becoming more efficient, which is a good thing. People should realize there's no point in holding back this technology just for the sake of keeping people employed. If this were beneficial, then calculators and computers shouldn't exist either.
#2One small problem is that people need to pay rent and eat.
#3So we should ditch computers and go back to the typing pool? Should we get rid of heavy earth moving equipment and just use hundreds of guys with hand tools to build everything? It would employ a hell of a lot more people.
#4No one's saying that. I don't think anyone is really against automation, but as it increases, there are soon going to be more people that there are jobs that actually need doing. I actually believe we've already passed this point. So what do we do with the people, who can't get jobs simply because there are none? It's an issue that need assessed immediately.
#5Tons and tons and tons of American jobs have been replaced by new jobs created by technology or in support of technology years ago. An office might have needed people to handle filing paperwork, keeping it in order, and retrieving, where now a document management system has made them completely redundant. The upshot is that to access that DMS, people are out there selling computers, installing computers, servicing computers, and supporting end users building the servers installing, supporting monitoring backing them up, and all that jobs that come in support of those progress is progress. And it advances human efficiency and knowledge. These are just one or two examples, but the answer is not to kill progress. Other countries simply won't. The answer is to push education to the forefront, so people are prepared for these jobs and whatever other challenges the future may bring.
#6This is true. But it's unfortunate technological advances tend to reduce low skill jobs and replace them with high skill jobs. It would feel more fair if the low skilled workers could all do training programs and become high skilled workers. But this isn't really the case. Those jobs end up being taken by someone who had better educational opportunities or someone younger who still has time to take advantage of education.
#7The reality is the reality. Unfortunately or not educating people will create more educated people to handle high skill jobs, and I'll tell you being a desktop support technician isn't high skill. As that's where we push in the future, any amount of hand wringing won't change the facts. We must educate our people if we want to be a global leader in more than homelessness poverty.
#8Education won't matter. We are at the end of the job age at some point in the near future. We are going to have to deal with the fact that getting a job isn't a reality for a significant percentage of the population. Society will have to radically change as it did during the industrial revolution.
#9Much cheaper to heavily discourage having more children free abortions. Then in years there won't be so many useless people who can apparently be replaced by a simple robot.
#10Virtually every job will be replaced by automation name skilled trades that can't be automated. I imagine you'd be surprised at how hard this is. Are pharmacists useless, surgeons, accountants? I'd bet that your job is just as replaceable as these.
" + }, + { + "type": "table", + "bbox": [ + 0.224, + 0.744, + 0.774, + 0.874 + ], + "angle": 0, + "content": "
Candidate Quote TextsSdSpRGround truth
There's no such thing as a free lunch.0.3650.4171
I can't predict the future.0.1850.2102✓
I have never let my schooling interfere with my education.0.1040.0593
Prevention is better than cure.0.0440.0834
Knowledge is power.0.0590.0525
Don't let schooling interfere with your education.0.0440.0436
Nature abhors a vacuum.0.0360.0247
There is no substitute for hard work.0.0240.0178
There are three kinds of lies: lies, damned lies, and statistics.0.0220.0139
You can't fix stupid.0.0190.01010
" + }, + { + "type": "table_caption", + "bbox": [ + 0.305, + 0.879, + 0.692, + 0.892 + ], + "angle": 0, + "content": "(iv) Qualitative Example-D from quote recommendation." + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.915, + 0.882, + 0.944 + ], + "angle": 0, + "content": "Table 5: Case analyses of quote recommendation. We demonstrate the candidate quotes of the top 10 rankings out of all candidates. Note that there is only one ground truth quote for each conversation history." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.133, + 0.085, + 0.435, + 0.1 + ], + "angle": 0, + "content": "ACL 2023 Responsible NLP Checklist" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.108, + 0.323, + 0.123 + ], + "angle": 0, + "content": "A For every submission:" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.127, + 0.533, + 0.158 + ], + "angle": 0, + "content": "A1. Did you describe the limitations of your work? Section 5" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.171, + 0.553, + 0.202 + ], + "angle": 0, + "content": "A2. Did you discuss any potential risks of your work? We see no concern about potential risks." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.213, + 0.696, + 0.244 + ], + "angle": 0, + "content": "A3. Do the abstract and introduction summarize the paper's main claims? Section 1" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.256, + 0.669, + 0.288 + ], + "angle": 0, + "content": "A4. Have you used AI writing assistants when working on this paper? Left blank." + }, + { + "type": "list", + "bbox": [ + 0.13, + 0.127, + 0.696, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.3, + 0.49, + 0.317 + ], + "angle": 0, + "content": "B Did you use or create scientific artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.322, + 0.449, + 0.337 + ], + "angle": 0, + "content": "The Abstract provides the link to our code." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.348, + 0.531, + 0.38 + ], + "angle": 0, + "content": "B1. Did you cite the creators of artifacts you used? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.391, + 0.779, + 0.423 + ], + "angle": 0, + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.434, + 0.881, + 0.514 + ], + "angle": 0, + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.525, + 0.881, + 0.588 + ], + "angle": 0, + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.599, + 0.881, + 0.648 + ], + "angle": 0, + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? In the Abstract, a Github repository with documentation is released." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.658, + 0.882, + 0.755 + ], + "angle": 0, + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Appendix A" + }, + { + "type": "list", + "bbox": [ + 0.129, + 0.348, + 0.882, + 0.755 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.765, + 0.494, + 0.782 + ], + "angle": 0, + "content": "C Did you run computational experiments?" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.788, + 0.206, + 0.801 + ], + "angle": 0, + "content": "Section 3" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.813, + 0.881, + 0.86 + ], + "angle": 0, + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Section 3" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.867, + 0.878, + 0.89 + ], + "angle": 0, + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1169" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.13, + 0.084, + 0.88, + 0.131 + ], + "angle": 0, + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Section 3" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.143, + 0.881, + 0.207 + ], + "angle": 0, + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Section 3" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.219, + 0.881, + 0.283 + ], + "angle": 0, + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Not applicable. Left blank." + }, + { + "type": "list", + "bbox": [ + 0.129, + 0.084, + 0.881, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.294, + 0.878, + 0.331 + ], + "angle": 0, + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.341, + 0.881, + 0.388 + ], + "angle": 0, + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.4, + 0.881, + 0.464 + ], + "angle": 0, + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.476, + 0.881, + 0.54 + ], + "angle": 0, + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.55, + 0.873, + 0.582 + ], + "angle": 0, + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.593, + 0.881, + 0.641 + ], + "angle": 0, + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Not applicable. Left blank." + }, + { + "type": "list", + "bbox": [ + 0.128, + 0.341, + 0.881, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.928, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1170" + } + ] +] \ No newline at end of file diff --git a/2023/UniTRec_ A Unified Text-to-Text Transformer and Joint Contrastive Learning Framework for Text-based Recommendation/31a35e47-82fb-45ce-aa23-0bc4dd29fd03_origin.pdf b/2023/UniTRec_ A Unified Text-to-Text Transformer and Joint Contrastive Learning Framework for Text-based Recommendation/31a35e47-82fb-45ce-aa23-0bc4dd29fd03_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4a264fd6a249f64f8e4aa71b3f2b6c85a5dd5ce6 --- /dev/null +++ b/2023/UniTRec_ A Unified Text-to-Text Transformer and Joint Contrastive Learning Framework for Text-based Recommendation/31a35e47-82fb-45ce-aa23-0bc4dd29fd03_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:15766ea02fedf748dca8a60d6f2f1c91f48803c42119a269d758e930b6140a08 +size 739440 diff --git a/2023/UniTRec_ A Unified Text-to-Text Transformer and Joint Contrastive Learning Framework for Text-based Recommendation/full.md b/2023/UniTRec_ A Unified Text-to-Text Transformer and Joint Contrastive Learning Framework for Text-based Recommendation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3d6c85cce80ed69f3b20672b542debd0af766abb --- /dev/null +++ b/2023/UniTRec_ A Unified Text-to-Text Transformer and Joint Contrastive Learning Framework for Text-based Recommendation/full.md @@ -0,0 +1,275 @@ +# UniTRec: A Unified Text-to-Text Transformer and Joint Contrastive Learning Framework for Text-based Recommendation + +Zhiming Mao $^{1,2}$ , Huimin Wang $^{1,3}$ , Yiming Du $^{1,2}$ , Kam-Fai Wong $^{1,2}$ + +1The Chinese University of Hong Kong, Hong Kong, China + +$^{2}$ MoE Key Laboratory of High Confidence Software Technologies, China + +$^{3}$ Jarvis Lab, Tencent, Shenzhen, China + +{zmmao,ydu,kfwong}@se.cuhk.edu.hk + +hmmmwang@tencent.com + +# Abstract + +Prior study has shown that pretrained language models (PLM) can boost the performance of text-based recommendation. In contrast to previous works that either use PLM to encode user history as a whole input text, or impose an additional aggregation network to fuse multi-turn history representations, we propose a unified local- and global-attention Transformer encoder to better model two-level contexts of user history. Moreover, conditioned on user history encoded by Transformer encoders, our framework leverages Transformer decoders to estimate the language perplexity of candidate text items, which can serve as a straightforward yet significant contrastive signal for user-item text matching. Based on this, our framework, UniTRec, unifies the contrastive objectives of discriminative matching scores and candidate text perplexity to jointly enhance text-based recommendation. Extensive evaluation shows that UniTRec delivers SOTA performance on three text-based recommendation tasks. $^{1}$ + +# 1 Introduction + +Text-based recommendation (Li et al., 2010; Gu et al., 2016; Okura et al., 2017; Malkiel et al., 2020) aims to recommend relevant textual content (e.g., news articles, Twitter posts) to people based on their behaviors as represented in historical log texts. For instance, engagement recommendation (Cheng et al., 2022) on social media (e.g., Twitter and Reddit) helps users discover and engage with interested threads by modeling their browsing history. + +Pretrained language models (Devlin et al., 2019; Brown et al., 2020) have made waves in recent text-based recommendation research (Zhang et al., 2021; Qi et al., 2022; Geng et al., 2022). The most common practice is using PLM encoders (BERT family) to learn representations of user history and candidate item texts. Recommendation + +matching scores are computed over the user and item representations and finally optimized by noise contrastive estimation (NCE) loss (Gutmann and Hyvarinen, 2010) for ranking multiple candidates. + +Unlike encoding single text, using PLM to encode multi-turn texts of user history is nontrivial. Existing works (Malkiel et al., 2020; Qi et al., 2022; Geng et al., 2022) concatenate multi-turn history texts as a whole input text, then use one PLM encoder to learn the holistic user representation. This is a standard PLM encoding manner but ignores the relation among history turns, as all word tokens from different history turns are equally attended2. In contrast, previous studies point out that learning the relation among user history turns is also beneficial (Zeng et al., 2020; Qi et al., 2021). Another approach is using PLM encoders to learn representations from multi-turn history texts, followed by an additional aggregation network to fuse the multi-turn representations (Wu et al., 2021; Li et al., 2022). However, the imposed aggregation networks (with newly initialized parameters) weaken the representation power of PLM encoders which are already pretrained on large-scale corpora. + +This work introduces UniTRec, a Unified text-to-text Transformer framework for text-based Recommendation. In the encoder component of UniTRec, we design local- and global-attention to learn user history representations through tailored attention masking, which aims to jointly model word-level and turn-level relations of user history. UniTRec can utilize the full power of PLM encoders because it preserves the intact structure of PLM encoders without newly imposed parameters. + +Different from most previous works that predict user-candidate matching scores solely based on the representations learned by Transformer encoders, we argue that conditioned on user representations + +![](images/40a6b845906692592a6a73d5a2c235d837c405da9535503c9288c81a7c8d3118.jpg) +Figure 1: An example of perplexity-based ranking for candidate item texts, conditioned on user history. The illustrated task is text-based news recommendation. + +learned by Transformer encoders, candidate text perplexity (PPL) estimated by pretrained Transformer decoders is also a straightforward yet significant signal for text-based recommendation. As shown in Figure 1, we hypothesize that the candidate text perplexity estimated by pretrained LM decoders can directly measure the text matching degree between user history and candidate texts. It is because the perplexity estimates the likelihood of candidate texts based on encoder outputs, which naturally indicates the probabilities of candidate texts given the user history. Besides, UniTRec can use the last hidden states of Transformer decoders to directly predict matching scores. Hence, this work unifies the contrastive objectives of discriminative matching scores and candidate text perplexity to jointly enhance text-based recommendation. + +The contributions of this work are: (1) We propose local- and global-attention to model two-level relation of user history without additional parameters, which enjoys the full power of PLM encoders. (2) We introduce PLM perplexity to measure user-candidate text matching and unify the objectives of discriminative matching scores and candidate text perplexity to enhance text-based recommendation. (3) Experiments on three text-based recommendation datasets validate the effectiveness of UniTRec. + +# 2 Approach + +# 2.1 Unified User-history Modeling + +Formally, multi-turn history of a user is represented as $H = [t_1, t_2, \dots, t_N]$ , and each turn text $t_i$ contains $|t_i|$ words as $t_i = [x_i^1, x_i^2, \dots, x_i^{|t_i|}]$ . UniTRec aims to unify learning word- and turn-level context representations in one Transformer encoder. + +Local attention on word-level context. We first + +concatenate the multi-turn history texts as the input tokens $X = [x_{1}^{1}, x_{1}^{2}, \dots, x_{1}^{|t_{1}|}, \dots, x_{N}^{1}, x_{N}^{2}, \dots, x_{N}^{|t_{N}|}]$ . Inspired by Dong et al. (2019), we tailor the attention masking in Transformer self-attention to learn the word-level context of each turn. Specifically, we allow word tokens from the same turn to attend to each other, while tokens from different turns are excluded from self-attention computation: + +$\mathbf{M}_{i,j} = \left\{ \begin{array}{ll}0, & \mathrm{token} x_i\mathrm{and}x_j\mathrm{in the same turn}\\ -\infty , & \mathrm{otherwise} \end{array} \right.$ + +$$ +\operatorname {A t t e n t i o n} (Q, K, V) = \operatorname {s o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {d _ {k}}} + \mathbf {M}\right) V \tag {1} +$$ + +, where $Q, K, V$ are self-attention query, key, and value in Vaswani et al. (2017), $\mathbf{M}$ is the mask matrix to achieve local-attention inside each turn text. The local self-attention blocks consist of $L_{1}$ layers, by which original PLM encoders can be adapted to learn word-level context representations of turns. + +Global attention on turn-level context. Over the local self-attention layers, we leverage global self-attention to model the relation among history turns. Specifically, tokens from all turns attend to each other in self-attention computation (by setting the mask matrix $\mathbf{M} = \mathbf{0}$ ). In this way, Transformer encoders can perform global interaction among each token (and turn) to learn turn-level context representations of user history. There are $L_{2}$ layers in the global self-attention blocks, which can also be inherited from PLM encoders directly. + +# 2.2 Joint Contrastive Ranking Objectives + +Conditioned on the history representation, we input the candidate text to Transformer decoders to predict how likely it should be recommended. It is worth noting that Transformer decoders can naturally perform effective cross-attention interaction between history and candidate hidden states. + +# 2.2.1 Objective on Discriminative Scores + +Motivated by Lewis et al. (2020), we feed the last hidden state of decoder output $h_{T}$ to an MLP scorehead which predicts the user-candidate matching score $S^{d} = \mathrm{ScoreHead}(h_{T})$ . The matching score is discriminative, as higher scores indicate higher user-candidate matching probabilities. + +Following previous works (Li et al., 2022; Qi et al., 2022), we adopt negative sampling with NCE loss to optimize matching score prediction. Given the user history and its ground truth matched candidate $C_i$ , UniTRec predicts the matching score + +![](images/e95c2a2c75f3e5b5a1f797a305cfcdafd75832efb99e732dd3ad387e4f424bbf.jpg) +Figure 2: Overview of UniTRec. In training, matching scores $S^d$ and $S^p$ are optimized by the NCE loss, respectively. In inference, $S^d$ and $S^p$ are normalized and combined to derive the final output ranking. + +as $S_{i}^{d + }$ . In addition, $K$ unmatched negative candidates $\{C_j\}_{j = 1}^K$ are sampled from the candidate set, and their matching scores are $\{S_j^{d - }\}_{j = 1}^K$ . The NCE loss is represented in a contrastive form: + +$$ +\mathcal {L} _ {i} ^ {d} = - \log \frac {\exp (S _ {i} ^ {d +})}{\exp (S _ {i} ^ {d +}) + \sum_ {j = 1} ^ {K} \exp (S _ {j} ^ {d -})} \quad (2) +$$ + +# 2.2.2 Objective on Candidate Text Perplexity + +As aforementioned, UniTRec leverages perplexity to rank candidate texts. Since lower perplexity indicates higher user-candidate matching probability, regarding the candidate text $Y = [y_{1}, y_{2}, \dots, y_{T}]$ , we define the perplexity-based matching score $S^{p}$ as its negative perplexity: + +$$ +S ^ {p} = - \operatorname {P P L} (Y) = \frac {1}{T} \sum_ {i = 1} ^ {T} \log p _ {\theta} \left(y _ {i} \mid y _ {< i}\right) \tag {3} +$$ + +, where $p_{\theta}(\cdot)$ denotes the target probability output from the UniTRec Transformer decoder. Similar to Eq. (2), we optimize the perplexity-based matching score $S^p$ in the NCE loss form. As perplexity empirically varies in a wide range, we introduce a temperature parameter $\tau$ to balance the joint NCE loss gradients following Radford et al. (2021). + +$$ +\mathcal {L} _ {i} ^ {p} = - \log \frac {\exp \left(\tau \cdot S _ {i} ^ {p +}\right)}{\exp \left(\tau \cdot S _ {i} ^ {p +}\right) + \sum_ {j = 1} ^ {K} \exp \left(\tau \cdot S _ {j} ^ {p -}\right)} \tag {4} +$$ + +, where $\tau$ is learnable and initialized to 1. On the training dataset $\mathcal{D}$ , the joint contrastive learning objective is formulated as: + +$$ +\mathcal {L} = \sum_ {i = 1} ^ {| \mathcal {D} |} \left(\mathcal {L} _ {i} ^ {d} + \mathcal {L} _ {i} ^ {p}\right) \tag {5} +$$ + +# 2.3 Model Initialization and Inference + +As UniTRec is a standard text-to-text Transformer, we initialize the parameters from pretrained BART (Lewis et al., 2020). In inference, UniTRec predicts the discriminative and perplexity-based scores for each candidate item, respectively. The two separate scores $S^d$ and $S^p$ are normalized, averaged, and finally ranked as the output. Detailed ranking process is provided in Appendix B. + +# 3 Experiments + +We evaluate UniTRec on three text-based recommendation tasks: 1) NewsRec, to recommend news articles to users based on their browsing history. We use the MIND-small dataset (Wu et al., 2020) for experiments. 2) QuoteRec, to recommend quotations to users based on their conversation history. We use the Reddit-quotation dataset (Wang et al., 2021) for experiments. 3) EngageRec, to recommend social media posts for users to engage with based on their comment history. We use the dataset released by Zeng et al. (2020) for experiments. Detailed dataset statistics is provided in Appendix A. + +Implementation Details. The UniTRec encoder and decoder both consist of 6 Transformer layers with 768-dimensional hidden states and 12 attention heads. We set $L_{1} = 3$ and $L_{2} = 3$ . We use AdamW optimizer (Loshchilov and Hutter, 2019) to train UniTRec with cosine learning rate decay. + +Baselines. We compare UniTRec with competitive baselines: 1) GRU4Rec (Balázs et al., 2016) utilizes a GRU network to learn multi-turn history. 2) SASRec (Kang and McAuley, 2018) encodes user history with a self-attention based sequential model. 3) BERT4Rec (Sun et al., 2019) employs bidirectional self-attention to model user history. 4) RoBERTa-Sim, a simple yet strong baseline men + +
ModelNewsRecQuoteRecEngageRec
MRRNDCG@5/10HR@5/10MRRNDCG@5/10HR@5/10MRRNDCG@5/10HR@5/10
GRU4Rec32.9136.20/42.5350.33/68.3534.0834.65/37.9344.45/54.632.121.04/1.511.27/2.65
SASRec32.6036.03/42.3750.63/68.6433.6334.30/37.4944.32/54.202.401.49/1.952.16/3.47
BERT4Rec32.8736.18/42.4050.21/67.9733.5934.26/37.2743.76/53.053.041.98/3.232.81/6.67
RoBERTa-Sim32.9636.47/42.8151.06/69.0837.1337.96/41.1848.14/58.063.742.66/3.754.42/7.70
UNBERT33.0936.53/42.8450.87/68.8239.7540.74/43.6950.90/60.042.831.96/2.673.11/5.24
UniTRec33.7637.63/43.7452.61/69.8941.2442.38/45.3152.87/61.884.063.23/4.294.58/7.68
+ +Table 1: Experiment results on three text-based recommendation tasks. MRR denotes mean reciprocal rank, NDCG denotes normalized discounted cumulative gain, and HR denotes hit ratio (presented in percentage). The overall performance of UniTRec is better than other baseline models with $p$ -value $< 0.05$ , validated by unpaired t-test. + +
ModelNewsRecQuoteRecEngageRec
MRRNDCG@5/10HR@5/10MRRNDCG@5/10HR@5/10MRRNDCG@5/10HR@5/10
UniTRec33.7637.63/43.7452.61/69.8941.2442.38/45.3152.87/61.884.063.23/4.294.58/7.68
w/o BART Init30.3133.32/39.6947.55/65.7819.0217.66/20.8022.45/32.162.240.86/1.611.27/3.62
w/o Local-Att33.3437.22/43.3252.28/69.5440.4441.63/44.5652.09/61.153.923.19/4.154.38/7.36
w/o Global-Att33.2237.06/43.1752.14/69.4740.2541.47/44.2652.07/60.763.642.78/3.593.89/6.35
Disc-Score only33.0736.76/43.0351.68/69.4640.5941.81/44.6552.39/61.143.822.99/3.604.49/6.85
PPL-Score only32.8336.39/42.5951.05/68.6740.3141.43/44.4752.13/61.203.292.39/3.033.86/5.66
+ +Table 2: Recommendation performance of ablation model variants. + +tioned in Qi et al. (2022), uses the hidden states of [CLS] tokens to measure user-candidate similarity. 5) UNBERT, implemented as Zhang et al. (2021), concatenates history and candidate texts as the input to BERT and predicts matching scores from the final hidden states of [CLS] tokens. + +Note that we do not consider other methods that use non-text inputs (e.g., user profile, text topic labels). For fair comparison, all baseline models use pretrained 12-layer RoBERTa-base (Liu et al., 2019) as text encoders to learn embeddings of texts. + +# 3.1 Main Results + +Table 1 shows the performance of experiment models. From the results of NewsRec and QuoteRec, we can see that UniTRec outperforms all baseline models by a clear margin. Also, RoBERTa-Sim and UNBERT that directly use the [CLS] hidden states to represent user history, surpass other baselines that build additional aggregation networks upon the whole RoBERTa outputs. As displayed in the results, EngageRec is the most difficult task. We inspect the dataset and find that the texts on social media contain too much noise (e.g., URL and emoji), and the user history contains less number of turns. Nevertheless, UniTRec achieves better overall performance than other baseline models, validating its robustness on noisy text inputs and limited user history. + +# 3.2 Ablation Studies and Analyses + +We further conduct ablation studies on UniTRec. The experiment results are reported in Table 2. + +Initialization of UniTRec. We train UniTRec from scratch without initialization from pretrained BART (refer to w/o BART Init). The recommendation performance significantly drops in all three tasks, which indicates that acquiring effective text understanding ability from PLM is a necessary key to UniTRec performance. + +Local and global attention. We investigate the function of two-level attention modules of the UniTRec history encoder. Concretely, we set $L_{1} = 0$ in w/o Local-Att and $L_{2} = 0$ in w/o Global-Att, where $L_{1} + L_{2} = 6$ . We can observe that removing local and global attention from the original UniTRec history encoder both lead to suboptimal performance, while the performance drop is more significant in w/o Global-Att. The results justify the effectiveness of jointly modeling two-level history contexts through adapted Transformer attention masking without additional parameters. + +Discriminative and perplexity-based objectives. We probe into training UniTRec with standalone discriminative (Disc-Score only) and perplexity-based (PPL-Score only) contrastive objectives, respectively. We can see that the discriminative objective yields better performance than the perplexity-based objective. Besides, the model performance on both standalone objectives declines compared to the original joint objective. The results indicate that the discriminative and perplexity-based matching scores are complementary and can jointly provide more accurate signals of user history and candidate text matching for text-based recommendation. + +# 4 Conclusion + +We present a unified Transformer UniTRec for text-based recommendation. UniTRec learns two-level contexts of multi-turn user history and jointly exploits discriminative matching scores and candidate text perplexity as matching objectives. Empirical experiments on three text-based recommendation datasets corroborate the effectiveness of UniTRec. + +# 5 Limitations + +Our model only focuses on utilizing text information for recommendation, which is a key limitation of this work. In real-world settings, recommender systems are usually required to handle heterogeneous information inputs. UniTRec is a pure text-based recommender modeling user history and candidate texts as inputs. However, incorporating additional side information (e.g., user profile, text topic labels, and dwell time of user behaviors) could further improve the recommendation performance and alleviate the cold start problem. Furthermore, UniTRec only models two-level relations of user behavior history. Nonetheless, incorporating more user behavior information, such as implicit and negative feedback, could further enhance the recommendation performance. + +# Acknowledgements + +We appreciate constructive comments from anonymous reviewers. The research described in this paper is partially supported by CUHK under Project No. 3230366. + +# References + +Hidasi Balázs, Karatzoglou Alexandros, Baltrunas Linas, and Tikk Domonkos. 2016. Session-based recommendations with recurrent neural networks. In 4th International Conference on Learning Representations ICLR 2016, San Juan, Puerto Rico, May 2-4, 2016, Conference Track Proceedings. +Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems, + +volume 33, pages 1877-1901. Curran Associates, Inc. +Daniel Cheng, Kyle Yan, Phillip Keung, and Noah A. Smith. 2022. The engage corpus: A social media dataset for text-based recommender systems. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 1885-1889, Marseille, France. European Language Resources Association. +Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 4171-4186. Association for Computational Linguistics. +Li Dong, Nan Yang, Wenhui Wang, Furu Wei, Xiaodong Liu, Yu Wang, Jianfeng Gao, Ming Zhou, and Hsiao-Wuen Hon. 2019. Unified language model pre-training for natural language understanding and generation. In 33rd Conference on Neural Information Processing Systems (NeurIPS 2019). +Shijie Geng, Shuchang Liu, Zuohui Fu, Yingqiang Ge, and Yongfeng Zhang. 2022. Recommendation as language processing (rlp): A unified pretrain, personalized prompt & predict paradigm (p5). In Proceedings of the 16th ACM Conference on Recommender Systems, RecSys '22, page 299-315, New York, NY, USA. Association for Computing Machinery. +Youyang Gu, Tao Lei, Regina Barzilay, and Tommi Jaakkola. 2016. Learning to refine text based recommendations. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2103-2108, Austin, Texas. Association for Computational Linguistics. +Michael Gutmann and Aapo Hyvarinen. 2010. Noise-contrastive estimation: A new estimation principle for unnormalized statistical models. In Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics, volume 9 of Proceedings of Machine Learning Research, pages 297-304, Chia Laguna Resort, Sardinia, Italy. PMLR. +Wang-Cheng Kang and Julian McAuley. 2018. Self-attentive sequential recommendation. In 2018 IEEE International Conference on Data Mining (ICDM), pages 197-206. +Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871-7880, Online. Association for Computational Linguistics. + +Jian Li, Jieming Zhu, Qiwei Bi, Guohao Cai, Lifeng Shang, Zhenhua Dong, Xin Jiang, and Qun Liu. 2022. MINER: Multi-interest matching network for news recommendation. In Findings of the Association for Computational Linguistics: ACL 2022, pages 343-352, Dublin, Ireland. Association for Computational Linguistics. +Yize Li, Jiazhong Nie, Yi Zhang, Bingqing Wang, Baoshi Yan, and Fuliang Weng. 2010. Contextual recommendation based on text mining. In *Coling* 2010: Posters, pages 692-700, Beijing, China. Coling 2010 Organizing Committee. +Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. In arXiv preprint arXiv: 1907.11692. arXiv. +Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net. +Itzik Malkiel, Oren Barkan, Avi Caciularu, Noam Razin, Ori Katz, and Noam Koenigstein. 2020. *RecoBERT: A catalog language model for text-based recommendations*. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, Online. Association for Computational Linguistics. +Shumpei Okura, Yukihiro Tagami, Shingo Ono, and Akira Tajima. 2017. Embedding-based news recommendation for millions of users. In Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, page 1933-1942, New York, NY, USA. Association for Computing Machinery. +Fanchao Qi, Yanhui Yang, Jing Yi, Zhili Cheng, Zhiyuan Liu, and Maosong Sun. 2022. QuoteR: A benchmark of quote recommendation for writing. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 336-348, Dublin, Ireland. Association for Computational Linguistics. +Tao Qi, Fangzhao Wu, Chuhan Wu, Peiru Yang, Yang Yu, Xing Xie, and Yongfeng Huang. 2021. HieRec: Hierarchical user interest modeling for personalized news recommendation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 5446-5456, Online. Association for Computational Linguistics. +Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International + +Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 8748-8763. PMLR. +Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang. 2019. Bert4rec: Sequential recommendation with bidirectional encoder representations from transformer. In Proceedings of the 28th ACM International Conference on Information and Knowledge Management, CIKM '19, page 1441-1450, New York, NY, USA. Association for Computing Machinery. +Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30, pages 5998-6008. Curran Associates, Inc. +Lingzhi Wang, Xingshan Zeng, and Kam-Fai Wong. 2021. Quotation recommendation and interpretation based on transformation from queries to quotations. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 754-758, Online. Association for Computational Linguistics. +Chuhan Wu, Fangzhao Wu, Tao Qi, and Yongfeng Huang. 2021. Empowering news recommendation with pre-trained language models. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '21, page 1652-1656, New York, NY, USA. Association for Computing Machinery. +Fangzhao Wu, Ying Qiao, Jiun-Hung Chen, Chuhan Wu, Tao Qi, Jianxun Lian, Danyang Liu, Xing Xie, Jianfeng Gao, Winnie Wu, and Ming Zhou. 2020. MIND: A large-scale dataset for news recommendation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3597-3606, Online. Association for Computational Linguistics. +Xingshan Zeng, Jing Li, Lu Wang, Zhiming Mao, and Kam-Fai Wong. 2020. Dynamic online conversation recommendation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3331-3341, Online. Association for Computational Linguistics. +Qi Zhang, Jingjie Li, Qinglin Jia, Chuyuan Wang, Jieming Zhu, Zhaowei Wang, and Xiuqiang He. 2021. Umbert: User-news matching bert for news recommendation. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-21, pages 3356-3362. International Joint Conferences on Artificial Intelligence Organization. Main Track. + +
DatasetNewsRecQuoteRecEngageRec
Avg. history turns26.094.243.29
Avg. history tokens414.40279.82286.82
Avg. candidates37.2311117163
Avg. candidate tokens16.1519.11102.42
+ +Table 3: Statistics of three text-based recommendation training datasets. History and candidate tokens denote the number of BPE-tokenized tokens. The test set distribution is closed to the training sets (except candidates of EngageRec) and hence omitted. Note that the max length of each history log is truncated to 1024 tokens. + +# A Dataset Statistics + +The detailed statistics of the three text-based recommendation datasets are displayed in Table 3. Note that we use news titles as the text inputs for NewsRec following Qi et al. (2021). NewsRec regards the user clicked and non-clicked news as candidate texts, while QuoteRec and EngageRec regard all potential quotation texts and post texts as candidates. Different from Zeng et al. (2020) that formulates the task as recommending candidate users to given posts based on post content, we formulate the task as recommending candidate posts to given users based on user history. + +# Algorithm 1 Candidate Ranking Processes + +Input: discriminative scores $S^d = \{S_1^d,S_2^d,\dots,S_M^d\}$ perplexity-based scores $S^{p} = \{S_{1}^{p},S_{2}^{p},\dots,S_{M}^{p}\}$ + +Output: final averaged ranking $R$ . + +1: Derive the normalized discriminative scores $S_{norm}^{d} =$ softmax(Sd). +2: Derive the normalized perplexity-based scores $S_{norm}^{p} =$ softmax( $S^p$ ). +3: Derive the geometric average scores $\bar{S} = \log (S_{norm}^d) + \log (S_{norm}^p)$ . +4: Sort the averaged scores $\bar{S}$ by descending order to derive the final ranking: $\bar{R} \gets \mathrm{Rank}_{\mathrm{des}}(\bar{S})$ +5: return $R$ + +# B Inference Ranking + +Given the user history and $M$ candidate texts, UniTRec first predicts the discriminative ranking scores $S^d = \{S_1^d,S_2^d,\dots,S_M^d\}$ and perplexity-based ranking scores $S^{p} = \{S_{1}^{p},S_{2}^{p},\dots,S_{M}^{p}\}$ of the candidates. Algorithm 1 outlines an approach to aggregate the final ranking based on $S^d$ and $S^p$ . Note that the function $\mathrm{Rank}(S)^4$ denotes outputting the sorted order of elements in a score list $S$ . There exist other ways to average the ranking of $S^d$ and $S^p$ , which we leave for future work to explore. + +# C Qualitative Analysis + +We show randomly sampled outputs of UniTRec, for instance, demonstrated on the news recommendation and quote recommendation tasks. Table 4 and 5 showcase the qualitative samples. + +
TurnHistory News Texts
#1Mac Engel: As long as these results are acceptable, Dallas Cowboys will continue to be losers
#2NFL world reacts to officials handing Packers win over Lions
#3Maryland Congressman Elijah Cummings, a Democrat and Chair of House Oversight and Reform Committee, has died: CNN
#4Unprecedented movement detected on California earthquake fault capable of 8.0 temblor
#5Bag Explodes While Being Loaded On Volaris Flight At Midway Airport
#6Orlando Scandrick rips Eagles: They have "accountability issues"
#7Meghan King Edmonds, Jim Edmonds' Nanny Denies Cheating Allegations
#8Nearly $400M worth of cocaine and marijuana intercepted by US Coast Guard
#9Former NBA first-round pick arrested in sex sting operation
#10China's trade with US shrinks in October despite optimism
+ +
Candidate News TextsSdSpRClicked
Taylor Swift Rep Hits Back at Big Machine, Claims She's Actually Owed $7.9 Million in Unpaid Royalties0.0950.0694X
Former North Carolina State, NBA player Anthony Grundy dies in stabbing, police say0.1720.1553X
13 Reasons Why's Christian Navarro Slams Disney for Casting "the White Guy" in The Little Mermaid0.0480.0657X
Opinion: Colin Kaepernick is about to get what he deserves: a chance0.3030.2501✓
3 Indiana judges suspended after a night of drinking turned into a White Castle brawl0.0760.0595X
66 Cool Tech Gifts Anyone Would Be Thrilled to Receive0.0090.0059X
Police find 26 children behind false wall at Colorado day care0.0340.1166X
I've been writing about tiny homes for a year and spent 2 nights in a 300-foot home to see what it is all about0.0290.0198X
Report: Police investigating woman's death after Redskins' player Montae Nicholson took her to hospital0.2350.2612✓
+ +(i) Qualitative Example-A from news recommendation. + +
TurnHistory News Texts
#1Toddler dancing to celebrate 11 months cancer-free goes viral
#2NFL Week 8 Power Rankings: Old-school football rules the day
#3The 25 US cities where it's easiest to get a mortgage
#4Burning questions for Cowboys vs Giants on "Monday Night Football"
#5Who's the favorite to win 2019 NFL rushing title?
#6Grading all 32 NFL teams heading into the last eight weeks of the 2019 season
#7Jennifer Aniston looks amazing in a makeup-free selfie, plus more news
#8This $12 million "mansion yacht" is made entirely of stainless steel and it's a first for the industry. Take a peek inside
+ +
Candidate News TextsSdSpRClicked
Opinion: Colin Kaepernick is about to get what he deserves: a chance0.3300.4001✓
U.S. Troops Will Die If They Remain in Syria, Bashar Al-Assad Warns0.0240.01110✗
Pete Davidson, Kaia Gerber Are Dating, Trying to Stay "Low Profile"0.0640.0336✗
The Hottest Tech Gifts This Holiday Season0.0500.0278✗
Taylor Swift Rep Hits Back at Big Machine, Claims She's Actually Owed $7.9 Million in Unpaid Royalties0.0460.0387✗
13 Reasons Why's Christian Navarro Slams Disney for Casting "the White Guy" in The Little Mermaid0.0600.0964✓
Some believe Mason Rudolph, hit in head with his own helmet, isn't getting enough blame0.1540.1792✓
South Carolina teen gets life in prison for deadly elementary school shooting0.0660.0465✗
The Unlikely Star of My Family's Thanksgiving Table0.0470.0219✗
Police investigating woman's death after Redskins' player Montae Nicholson took her to hospital0.1580.1493✗
+ +(ii) Qualitative Example-B from news recommendation. + +Table 4: Case analyses of news recommendation. History News Texts are sorted by user-clicked timestamps. $S^d$ , $S^p$ , and $\bar{R}$ are normalized discriminative, perplexity-based scores, and average ranking as described in Appendix B. Clicked denotes the ground truth user-click labels. Note that the experiment history logs are anonymized and delinked, which is always the first priority of the recommendation study. + +
TurnConversation Threading History
#1I own an FJ. It's a great car and even on stockies. It's great offroad.
#2I feel bad for you that you run the risk of being associated with the typical FJ owner.
#3What is a typical FJ owner? I've not heard anything bad about FJ owners.
#4It's like someone who drives a jeep wrangler in NYC. There's no need. Tons of FJ owners do that have it and not use it for what it's made for.
#5God forbid someone likes the design of a car and doesn't use it offroad.
#6Then buy a much more economic environmentalist friendly version. If you buy something and always use it for much less than it's purpose, why buy it?
#7Or people can buy whatever the hell they want because it's their money and not yours.
#8You're entirely right. Just like people can be rude just because you can do it, because you have the ability but why should you ass.
#9I wasn't aware that somebody buying a vehicle that they like and you don't was morally wrong.
#10I love FJs. It's perfectly fine to buy whatever you think looks nice.
+ +
Candidate Quote Texts\( S^d \)\( S^P \)\( \bar{R} \)Ground truth
Beauty is in the eye of the beholder.0.4800.4711✓
A fool and his money are soon parted.0.1760.1402
Form follows function.0.0510.0463
Everything is worth what its purchaser will pay for it.0.0400.0584
Because it's there.0.0380.0295
You can't fix stupid.0.0210.0346
The lady doth protest too much, methinks.0.0220.0137
It's all about the money.0.0200.0138
Anybody driving slower than you is an idiot, and anyone going faster than you is a maniac?0.0120.0189
Opportunity is missed by most people.0.0180.00810
+ +(iii) Qualitative Example-C from quote recommendation. + +
TurnConversation Threading History
#1Society is becoming more efficient, which is a good thing. People should realize there's no point in holding back this technology just for the sake of keeping people employed. If this were beneficial, then calculators and computers shouldn't exist either.
#2One small problem is that people need to pay rent and eat.
#3So we should ditch computers and go back to the typing pool? Should we get rid of heavy earth moving equipment and just use hundreds of guys with hand tools to build everything? It would employ a hell of a lot more people.
#4No one's saying that. I don't think anyone is really against automation, but as it increases, there are soon going to be more people that there are jobs that actually need doing. I actually believe we've already passed this point. So what do we do with the people, who can't get jobs simply because there are none? It's an issue that need assessed immediately.
#5Tons and tons and tons of American jobs have been replaced by new jobs created by technology or in support of technology years ago. An office might have needed people to handle filing paperwork, keeping it in order, and retrieving, where now a document management system has made them completely redundant. The upshot is that to access that DMS, people are out there selling computers, installing computers, servicing computers, and supporting end users building the servers installing, supporting monitoring backing them up, and all that jobs that come in support of those progress is progress. And it advances human efficiency and knowledge. These are just one or two examples, but the answer is not to kill progress. Other countries simply won't. The answer is to push education to the forefront, so people are prepared for these jobs and whatever other challenges the future may bring.
#6This is true. But it's unfortunate technological advances tend to reduce low skill jobs and replace them with high skill jobs. It would feel more fair if the low skilled workers could all do training programs and become high skilled workers. But this isn't really the case. Those jobs end up being taken by someone who had better educational opportunities or someone younger who still has time to take advantage of education.
#7The reality is the reality. Unfortunately or not educating people will create more educated people to handle high skill jobs, and I'll tell you being a desktop support technician isn't high skill. As that's where we push in the future, any amount of hand wringing won't change the facts. We must educate our people if we want to be a global leader in more than homelessness poverty.
#8Education won't matter. We are at the end of the job age at some point in the near future. We are going to have to deal with the fact that getting a job isn't a reality for a significant percentage of the population. Society will have to radically change as it did during the industrial revolution.
#9Much cheaper to heavily discourage having more children free abortions. Then in years there won't be so many useless people who can apparently be replaced by a simple robot.
#10Virtually every job will be replaced by automation name skilled trades that can't be automated. I imagine you'd be surprised at how hard this is. Are pharmacists useless, surgeons, accountants? I'd bet that your job is just as replaceable as these.
+ +
Candidate Quote TextsSdSpRGround truth
There's no such thing as a free lunch.0.3650.4171
I can't predict the future.0.1850.2102✓
I have never let my schooling interfere with my education.0.1040.0593
Prevention is better than cure.0.0440.0834
Knowledge is power.0.0590.0525
Don't let schooling interfere with your education.0.0440.0436
Nature abhors a vacuum.0.0360.0247
There is no substitute for hard work.0.0240.0178
There are three kinds of lies: lies, damned lies, and statistics.0.0220.0139
You can't fix stupid.0.0190.01010
+ +(iv) Qualitative Example-D from quote recommendation. + +Table 5: Case analyses of quote recommendation. We demonstrate the candidate quotes of the top 10 rankings out of all candidates. Note that there is only one ground truth quote for each conversation history. + +A For every submission: + +A1. Did you describe the limitations of your work? Section 5 +A2. Did you discuss any potential risks of your work? We see no concern about potential risks. +A3. Do the abstract and introduction summarize the paper's main claims? Section 1 +A4. Have you used AI writing assistants when working on this paper? Left blank. + +B Did you use or create scientific artifacts? + +The Abstract provides the link to our code. + +B1. Did you cite the creators of artifacts you used? Not applicable. Left blank. +B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Not applicable. Left blank. +B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Not applicable. Left blank. +B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Not applicable. Left blank. +B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? In the Abstract, a Github repository with documentation is released. +B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Appendix A + +C Did you run computational experiments? + +Section 3 + +C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Section 3 + +The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance. + +C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Section 3 +C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Section 3 +C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Not applicable. Left blank. + +D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank. + +D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Not applicable. Left blank. +D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Not applicable. Left blank. +D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Not applicable. Left blank. +D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Not applicable. Left blank. +D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Not applicable. Left blank. \ No newline at end of file diff --git a/2023/UniTRec_ A Unified Text-to-Text Transformer and Joint Contrastive Learning Framework for Text-based Recommendation/images.zip b/2023/UniTRec_ A Unified Text-to-Text Transformer and Joint Contrastive Learning Framework for Text-based Recommendation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e858fb6384b5598ea344a7fbc45ee39a8b8c1797 --- /dev/null +++ b/2023/UniTRec_ A Unified Text-to-Text Transformer and Joint Contrastive Learning Framework for Text-based Recommendation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:be47b330e68f4558ea1b57eb3df93e7ff48884cd4e530295b4b9901033bbfd6e +size 983437 diff --git a/2023/UniTRec_ A Unified Text-to-Text Transformer and Joint Contrastive Learning Framework for Text-based Recommendation/layout.json b/2023/UniTRec_ A Unified Text-to-Text Transformer and Joint Contrastive Learning Framework for Text-based Recommendation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..69ecf850b6d4a5023aee9ccf0c202a3e9769053c --- /dev/null +++ b/2023/UniTRec_ A Unified Text-to-Text Transformer and Joint Contrastive Learning Framework for Text-based Recommendation/layout.json @@ -0,0 +1,7403 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 86, + 68, + 507, + 101 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 68, + 507, + 101 + ], + "spans": [ + { + "bbox": [ + 86, + 68, + 507, + 101 + ], + "type": "text", + "content": "UniTRec: A Unified Text-to-Text Transformer and Joint Contrastive Learning Framework for Text-based Recommendation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 129, + 109, + 470, + 124 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 109, + 470, + 124 + ], + "spans": [ + { + "bbox": [ + 129, + 109, + 470, + 124 + ], + "type": "text", + "content": "Zhiming Mao" + }, + { + "bbox": [ + 129, + 109, + 470, + 124 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 129, + 109, + 470, + 124 + ], + "type": "text", + "content": ", Huimin Wang" + }, + { + "bbox": [ + 129, + 109, + 470, + 124 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 129, + 109, + 470, + 124 + ], + "type": "text", + "content": ", Yiming Du" + }, + { + "bbox": [ + 129, + 109, + 470, + 124 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 129, + 109, + 470, + 124 + ], + "type": "text", + "content": ", Kam-Fai Wong" + }, + { + "bbox": [ + 129, + 109, + 470, + 124 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 154, + 125, + 442, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 125, + 442, + 138 + ], + "spans": [ + { + "bbox": [ + 154, + 125, + 442, + 138 + ], + "type": "text", + "content": "1The Chinese University of Hong Kong, Hong Kong, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 123, + 138, + 474, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 138, + 474, + 153 + ], + "spans": [ + { + "bbox": [ + 123, + 138, + 474, + 153 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 123, + 138, + 474, + 153 + ], + "type": "text", + "content": "MoE Key Laboratory of High Confidence Software Technologies, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 205, + 153, + 391, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 153, + 391, + 166 + ], + "spans": [ + { + "bbox": [ + 205, + 153, + 391, + 166 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 205, + 153, + 391, + 166 + ], + "type": "text", + "content": "Jarvis Lab, Tencent, Shenzhen, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 181, + 168, + 417, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 168, + 417, + 180 + ], + "spans": [ + { + "bbox": [ + 181, + 168, + 417, + 180 + ], + "type": "text", + "content": "{zmmao,ydu,kfwong}@se.cuhk.edu.hk" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 226, + 182, + 371, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 226, + 182, + 371, + 194 + ], + "spans": [ + { + "bbox": [ + 226, + 182, + 371, + 194 + ], + "type": "text", + "content": "hmmmwang@tencent.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 84, + 234, + 274, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 234, + 274, + 497 + ], + "spans": [ + { + "bbox": [ + 84, + 234, + 274, + 497 + ], + "type": "text", + "content": "Prior study has shown that pretrained language models (PLM) can boost the performance of text-based recommendation. In contrast to previous works that either use PLM to encode user history as a whole input text, or impose an additional aggregation network to fuse multi-turn history representations, we propose a unified local- and global-attention Transformer encoder to better model two-level contexts of user history. Moreover, conditioned on user history encoded by Transformer encoders, our framework leverages Transformer decoders to estimate the language perplexity of candidate text items, which can serve as a straightforward yet significant contrastive signal for user-item text matching. Based on this, our framework, UniTRec, unifies the contrastive objectives of discriminative matching scores and candidate text perplexity to jointly enhance text-based recommendation. Extensive evaluation shows that UniTRec delivers SOTA performance on three text-based recommendation tasks." + }, + { + "bbox": [ + 84, + 234, + 274, + 497 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 68, + 507, + 154, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 507, + 154, + 519 + ], + "spans": [ + { + "bbox": [ + 68, + 507, + 154, + 519 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 528, + 291, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 528, + 291, + 649 + ], + "spans": [ + { + "bbox": [ + 67, + 528, + 291, + 649 + ], + "type": "text", + "content": "Text-based recommendation (Li et al., 2010; Gu et al., 2016; Okura et al., 2017; Malkiel et al., 2020) aims to recommend relevant textual content (e.g., news articles, Twitter posts) to people based on their behaviors as represented in historical log texts. For instance, engagement recommendation (Cheng et al., 2022) on social media (e.g., Twitter and Reddit) helps users discover and engage with interested threads by modeling their browsing history." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 650, + 291, + 745 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 650, + 291, + 745 + ], + "spans": [ + { + "bbox": [ + 67, + 650, + 291, + 745 + ], + "type": "text", + "content": "Pretrained language models (Devlin et al., 2019; Brown et al., 2020) have made waves in recent text-based recommendation research (Zhang et al., 2021; Qi et al., 2022; Geng et al., 2022). The most common practice is using PLM encoders (BERT family) to learn representations of user history and candidate item texts. Recommendation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 213, + 526, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 213, + 526, + 266 + ], + "spans": [ + { + "bbox": [ + 302, + 213, + 526, + 266 + ], + "type": "text", + "content": "matching scores are computed over the user and item representations and finally optimized by noise contrastive estimation (NCE) loss (Gutmann and Hyvarinen, 2010) for ranking multiple candidates." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 267, + 526, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 267, + 526, + 539 + ], + "spans": [ + { + "bbox": [ + 302, + 267, + 526, + 539 + ], + "type": "text", + "content": "Unlike encoding single text, using PLM to encode multi-turn texts of user history is nontrivial. Existing works (Malkiel et al., 2020; Qi et al., 2022; Geng et al., 2022) concatenate multi-turn history texts as a whole input text, then use one PLM encoder to learn the holistic user representation. This is a standard PLM encoding manner but ignores the relation among history turns, as all word tokens from different history turns are equally attended2. In contrast, previous studies point out that learning the relation among user history turns is also beneficial (Zeng et al., 2020; Qi et al., 2021). Another approach is using PLM encoders to learn representations from multi-turn history texts, followed by an additional aggregation network to fuse the multi-turn representations (Wu et al., 2021; Li et al., 2022). However, the imposed aggregation networks (with newly initialized parameters) weaken the representation power of PLM encoders which are already pretrained on large-scale corpora." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 539, + 525, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 539, + 525, + 675 + ], + "spans": [ + { + "bbox": [ + 302, + 539, + 525, + 675 + ], + "type": "text", + "content": "This work introduces UniTRec, a Unified text-to-text Transformer framework for text-based Recommendation. In the encoder component of UniTRec, we design local- and global-attention to learn user history representations through tailored attention masking, which aims to jointly model word-level and turn-level relations of user history. UniTRec can utilize the full power of PLM encoders because it preserves the intact structure of PLM encoders without newly imposed parameters." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 676, + 525, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 676, + 525, + 730 + ], + "spans": [ + { + "bbox": [ + 302, + 676, + 525, + 730 + ], + "type": "text", + "content": "Different from most previous works that predict user-candidate matching scores solely based on the representations learned by Transformer encoders, we argue that conditioned on user representations" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 740, + 525, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 740, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 740, + 525, + 772 + ], + "type": "text", + "content": "2There is no inductive bias of turn-level and history-level relations introduced to Transformer self-attention computation, where each token plays an equal role." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 67, + 751, + 291, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 751, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 751, + 291, + 772 + ], + "type": "text", + "content": "1Our code is available at https://github.com/Veason-silverbullet/UniTRec." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "1160" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "spans": [ + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "type": "text", + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 219, + 806, + 374, + 817 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 806, + 374, + 817 + ], + "spans": [ + { + "bbox": [ + 219, + 806, + 374, + 817 + ], + "type": "text", + "content": "Volume 2: Short Papers, pages 1160-1170" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "spans": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "text", + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 70, + 297, + 202 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 297, + 202 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 297, + 202 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 297, + 202 + ], + "type": "image", + "image_path": "40a6b845906692592a6a73d5a2c235d837c405da9535503c9288c81a7c8d3118.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 211, + 290, + 246 + ], + "lines": [ + { + "bbox": [ + 67, + 211, + 290, + 246 + ], + "spans": [ + { + "bbox": [ + 67, + 211, + 290, + 246 + ], + "type": "text", + "content": "Figure 1: An example of perplexity-based ranking for candidate item texts, conditioned on user history. The illustrated task is text-based news recommendation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 269, + 291, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 269, + 291, + 499 + ], + "spans": [ + { + "bbox": [ + 67, + 269, + 291, + 499 + ], + "type": "text", + "content": "learned by Transformer encoders, candidate text perplexity (PPL) estimated by pretrained Transformer decoders is also a straightforward yet significant signal for text-based recommendation. As shown in Figure 1, we hypothesize that the candidate text perplexity estimated by pretrained LM decoders can directly measure the text matching degree between user history and candidate texts. It is because the perplexity estimates the likelihood of candidate texts based on encoder outputs, which naturally indicates the probabilities of candidate texts given the user history. Besides, UniTRec can use the last hidden states of Transformer decoders to directly predict matching scores. Hence, this work unifies the contrastive objectives of discriminative matching scores and candidate text perplexity to jointly enhance text-based recommendation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 501, + 291, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 501, + 291, + 635 + ], + "spans": [ + { + "bbox": [ + 67, + 501, + 291, + 635 + ], + "type": "text", + "content": "The contributions of this work are: (1) We propose local- and global-attention to model two-level relation of user history without additional parameters, which enjoys the full power of PLM encoders. (2) We introduce PLM perplexity to measure user-candidate text matching and unify the objectives of discriminative matching scores and candidate text perplexity to enhance text-based recommendation. (3) Experiments on three text-based recommendation datasets validate the effectiveness of UniTRec." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 648, + 141, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 648, + 141, + 661 + ], + "spans": [ + { + "bbox": [ + 67, + 648, + 141, + 661 + ], + "type": "text", + "content": "2 Approach" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 671, + 238, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 671, + 238, + 684 + ], + "spans": [ + { + "bbox": [ + 67, + 671, + 238, + 684 + ], + "type": "text", + "content": "2.1 Unified User-history Modeling" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 689, + 290, + 758 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 689, + 290, + 758 + ], + "spans": [ + { + "bbox": [ + 67, + 689, + 290, + 758 + ], + "type": "text", + "content": "Formally, multi-turn history of a user is represented as " + }, + { + "bbox": [ + 67, + 689, + 290, + 758 + ], + "type": "inline_equation", + "content": "H = [t_1, t_2, \\dots, t_N]" + }, + { + "bbox": [ + 67, + 689, + 290, + 758 + ], + "type": "text", + "content": ", and each turn text " + }, + { + "bbox": [ + 67, + 689, + 290, + 758 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 67, + 689, + 290, + 758 + ], + "type": "text", + "content": " contains " + }, + { + "bbox": [ + 67, + 689, + 290, + 758 + ], + "type": "inline_equation", + "content": "|t_i|" + }, + { + "bbox": [ + 67, + 689, + 290, + 758 + ], + "type": "text", + "content": " words as " + }, + { + "bbox": [ + 67, + 689, + 290, + 758 + ], + "type": "inline_equation", + "content": "t_i = [x_i^1, x_i^2, \\dots, x_i^{|t_i|}]" + }, + { + "bbox": [ + 67, + 689, + 290, + 758 + ], + "type": "text", + "content": ". UniTRec aims to unify learning word- and turn-level context representations in one Transformer encoder." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 78, + 760, + 290, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 760, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 78, + 760, + 290, + 772 + ], + "type": "text", + "content": "Local attention on word-level context. We first" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 71, + 526, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 180 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 180 + ], + "type": "text", + "content": "concatenate the multi-turn history texts as the input tokens " + }, + { + "bbox": [ + 302, + 71, + 526, + 180 + ], + "type": "inline_equation", + "content": "X = [x_{1}^{1}, x_{1}^{2}, \\dots, x_{1}^{|t_{1}|}, \\dots, x_{N}^{1}, x_{N}^{2}, \\dots, x_{N}^{|t_{N}|}]" + }, + { + "bbox": [ + 302, + 71, + 526, + 180 + ], + "type": "text", + "content": ". Inspired by Dong et al. (2019), we tailor the attention masking in Transformer self-attention to learn the word-level context of each turn. Specifically, we allow word tokens from the same turn to attend to each other, while tokens from different turns are excluded from self-attention computation:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 184, + 521, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 184, + 521, + 219 + ], + "spans": [ + { + "bbox": [ + 302, + 184, + 521, + 219 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{i,j} = \\left\\{ \\begin{array}{ll}0, & \\mathrm{token} x_i\\mathrm{and}x_j\\mathrm{in the same turn}\\\\ -\\infty , & \\mathrm{otherwise} \\end{array} \\right." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 303, + 220, + 525, + 261 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 220, + 525, + 261 + ], + "spans": [ + { + "bbox": [ + 303, + 220, + 525, + 261 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}} + \\mathbf {M}\\right) V \\tag {1}", + "image_path": "3c7a062a9148abe4b3bdb5f7986ae5ef5667a4cab4f1b242ae7dca4915a355f0.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 264, + 525, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 264, + 525, + 343 + ], + "spans": [ + { + "bbox": [ + 302, + 264, + 525, + 343 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 302, + 264, + 525, + 343 + ], + "type": "inline_equation", + "content": "Q, K, V" + }, + { + "bbox": [ + 302, + 264, + 525, + 343 + ], + "type": "text", + "content": " are self-attention query, key, and value in Vaswani et al. (2017), " + }, + { + "bbox": [ + 302, + 264, + 525, + 343 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 302, + 264, + 525, + 343 + ], + "type": "text", + "content": " is the mask matrix to achieve local-attention inside each turn text. The local self-attention blocks consist of " + }, + { + "bbox": [ + 302, + 264, + 525, + 343 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 302, + 264, + 525, + 343 + ], + "type": "text", + "content": " layers, by which original PLM encoders can be adapted to learn word-level context representations of turns." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 345, + 525, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 345, + 525, + 493 + ], + "spans": [ + { + "bbox": [ + 302, + 345, + 525, + 493 + ], + "type": "text", + "content": "Global attention on turn-level context. Over the local self-attention layers, we leverage global self-attention to model the relation among history turns. Specifically, tokens from all turns attend to each other in self-attention computation (by setting the mask matrix " + }, + { + "bbox": [ + 302, + 345, + 525, + 493 + ], + "type": "inline_equation", + "content": "\\mathbf{M} = \\mathbf{0}" + }, + { + "bbox": [ + 302, + 345, + 525, + 493 + ], + "type": "text", + "content": "). In this way, Transformer encoders can perform global interaction among each token (and turn) to learn turn-level context representations of user history. There are " + }, + { + "bbox": [ + 302, + 345, + 525, + 493 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 302, + 345, + 525, + 493 + ], + "type": "text", + "content": " layers in the global self-attention blocks, which can also be inherited from PLM encoders directly." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 502, + 506, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 502, + 506, + 516 + ], + "spans": [ + { + "bbox": [ + 302, + 502, + 506, + 516 + ], + "type": "text", + "content": "2.2 Joint Contrastive Ranking Objectives" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 520, + 525, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 520, + 525, + 600 + ], + "spans": [ + { + "bbox": [ + 302, + 520, + 525, + 600 + ], + "type": "text", + "content": "Conditioned on the history representation, we input the candidate text to Transformer decoders to predict how likely it should be recommended. It is worth noting that Transformer decoders can naturally perform effective cross-attention interaction between history and candidate hidden states." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 608, + 502, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 608, + 502, + 621 + ], + "spans": [ + { + "bbox": [ + 302, + 608, + 502, + 621 + ], + "type": "text", + "content": "2.2.1 Objective on Discriminative Scores" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 624, + 525, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 624, + 525, + 704 + ], + "spans": [ + { + "bbox": [ + 302, + 624, + 525, + 704 + ], + "type": "text", + "content": "Motivated by Lewis et al. (2020), we feed the last hidden state of decoder output " + }, + { + "bbox": [ + 302, + 624, + 525, + 704 + ], + "type": "inline_equation", + "content": "h_{T}" + }, + { + "bbox": [ + 302, + 624, + 525, + 704 + ], + "type": "text", + "content": " to an MLP scorehead which predicts the user-candidate matching score " + }, + { + "bbox": [ + 302, + 624, + 525, + 704 + ], + "type": "inline_equation", + "content": "S^{d} = \\mathrm{ScoreHead}(h_{T})" + }, + { + "bbox": [ + 302, + 624, + 525, + 704 + ], + "type": "text", + "content": ". The matching score is discriminative, as higher scores indicate higher user-candidate matching probabilities." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 302, + 706, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 706, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 706, + 525, + 772 + ], + "type": "text", + "content": "Following previous works (Li et al., 2022; Qi et al., 2022), we adopt negative sampling with NCE loss to optimize matching score prediction. Given the user history and its ground truth matched candidate " + }, + { + "bbox": [ + 302, + 706, + 525, + 772 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 302, + 706, + 525, + 772 + ], + "type": "text", + "content": ", UniTRec predicts the matching score" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 780, + 308, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 308, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 308, + 791 + ], + "type": "text", + "content": "1161" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 118, + 68, + 477, + 216 + ], + "blocks": [ + { + "bbox": [ + 118, + 68, + 477, + 216 + ], + "lines": [ + { + "bbox": [ + 118, + 68, + 477, + 216 + ], + "spans": [ + { + "bbox": [ + 118, + 68, + 477, + 216 + ], + "type": "image", + "image_path": "e95c2a2c75f3e5b5a1f797a305cfcdafd75832efb99e732dd3ad387e4f424bbf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 221, + 526, + 247 + ], + "lines": [ + { + "bbox": [ + 67, + 221, + 526, + 247 + ], + "spans": [ + { + "bbox": [ + 67, + 221, + 526, + 247 + ], + "type": "text", + "content": "Figure 2: Overview of UniTRec. In training, matching scores " + }, + { + "bbox": [ + 67, + 221, + 526, + 247 + ], + "type": "inline_equation", + "content": "S^d" + }, + { + "bbox": [ + 67, + 221, + 526, + 247 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 221, + 526, + 247 + ], + "type": "inline_equation", + "content": "S^p" + }, + { + "bbox": [ + 67, + 221, + 526, + 247 + ], + "type": "text", + "content": " are optimized by the NCE loss, respectively. In inference, " + }, + { + "bbox": [ + 67, + 221, + 526, + 247 + ], + "type": "inline_equation", + "content": "S^d" + }, + { + "bbox": [ + 67, + 221, + 526, + 247 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 221, + 526, + 247 + ], + "type": "inline_equation", + "content": "S^p" + }, + { + "bbox": [ + 67, + 221, + 526, + 247 + ], + "type": "text", + "content": " are normalized and combined to derive the final output ranking." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 264, + 290, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 264, + 290, + 322 + ], + "spans": [ + { + "bbox": [ + 67, + 264, + 290, + 322 + ], + "type": "text", + "content": "as " + }, + { + "bbox": [ + 67, + 264, + 290, + 322 + ], + "type": "inline_equation", + "content": "S_{i}^{d + }" + }, + { + "bbox": [ + 67, + 264, + 290, + 322 + ], + "type": "text", + "content": " . In addition, " + }, + { + "bbox": [ + 67, + 264, + 290, + 322 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 67, + 264, + 290, + 322 + ], + "type": "text", + "content": " unmatched negative candidates " + }, + { + "bbox": [ + 67, + 264, + 290, + 322 + ], + "type": "inline_equation", + "content": "\\{C_j\\}_{j = 1}^K" + }, + { + "bbox": [ + 67, + 264, + 290, + 322 + ], + "type": "text", + "content": " are sampled from the candidate set, and their matching scores are " + }, + { + "bbox": [ + 67, + 264, + 290, + 322 + ], + "type": "inline_equation", + "content": "\\{S_j^{d - }\\}_{j = 1}^K" + }, + { + "bbox": [ + 67, + 264, + 290, + 322 + ], + "type": "text", + "content": " . The NCE loss is represented in a contrastive form:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 78, + 324, + 290, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 324, + 290, + 360 + ], + "spans": [ + { + "bbox": [ + 78, + 324, + 290, + 360 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i} ^ {d} = - \\log \\frac {\\exp (S _ {i} ^ {d +})}{\\exp (S _ {i} ^ {d +}) + \\sum_ {j = 1} ^ {K} \\exp (S _ {j} ^ {d -})} \\quad (2)", + "image_path": "3cef48c6a9f49e4e0473030c1c43da60cc124f85322dc3f67855938bc1c880bd.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 370, + 287, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 370, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 67, + 370, + 287, + 384 + ], + "type": "text", + "content": "2.2.2 Objective on Candidate Text Perplexity" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 391, + 290, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 391, + 290, + 472 + ], + "spans": [ + { + "bbox": [ + 67, + 391, + 290, + 472 + ], + "type": "text", + "content": "As aforementioned, UniTRec leverages perplexity to rank candidate texts. Since lower perplexity indicates higher user-candidate matching probability, regarding the candidate text " + }, + { + "bbox": [ + 67, + 391, + 290, + 472 + ], + "type": "inline_equation", + "content": "Y = [y_{1}, y_{2}, \\dots, y_{T}]" + }, + { + "bbox": [ + 67, + 391, + 290, + 472 + ], + "type": "text", + "content": ", we define the perplexity-based matching score " + }, + { + "bbox": [ + 67, + 391, + 290, + 472 + ], + "type": "inline_equation", + "content": "S^{p}" + }, + { + "bbox": [ + 67, + 391, + 290, + 472 + ], + "type": "text", + "content": " as its negative perplexity:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 74, + 478, + 290, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 478, + 290, + 502 + ], + "spans": [ + { + "bbox": [ + 74, + 478, + 290, + 502 + ], + "type": "interline_equation", + "content": "S ^ {p} = - \\operatorname {P P L} (Y) = \\frac {1}{T} \\sum_ {i = 1} ^ {T} \\log p _ {\\theta} \\left(y _ {i} \\mid y _ {< i}\\right) \\tag {3}", + "image_path": "e95def1425a6070d30e5c312d4cddcf9b1a7a973dc4a2153d9751fdefbaf60cb.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 509, + 290, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 509, + 290, + 604 + ], + "spans": [ + { + "bbox": [ + 67, + 509, + 290, + 604 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 67, + 509, + 290, + 604 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 67, + 509, + 290, + 604 + ], + "type": "text", + "content": " denotes the target probability output from the UniTRec Transformer decoder. Similar to Eq. (2), we optimize the perplexity-based matching score " + }, + { + "bbox": [ + 67, + 509, + 290, + 604 + ], + "type": "inline_equation", + "content": "S^p" + }, + { + "bbox": [ + 67, + 509, + 290, + 604 + ], + "type": "text", + "content": " in the NCE loss form. As perplexity empirically varies in a wide range, we introduce a temperature parameter " + }, + { + "bbox": [ + 67, + 509, + 290, + 604 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 67, + 509, + 290, + 604 + ], + "type": "text", + "content": " to balance the joint NCE loss gradients following Radford et al. (2021)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 70, + 610, + 289, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 610, + 289, + 654 + ], + "spans": [ + { + "bbox": [ + 70, + 610, + 289, + 654 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i} ^ {p} = - \\log \\frac {\\exp \\left(\\tau \\cdot S _ {i} ^ {p +}\\right)}{\\exp \\left(\\tau \\cdot S _ {i} ^ {p +}\\right) + \\sum_ {j = 1} ^ {K} \\exp \\left(\\tau \\cdot S _ {j} ^ {p -}\\right)} \\tag {4}", + "image_path": "e603cdb2b6b013533ca6a288bc4938d81b0bf65be4448eb96b855ce113c9d7aa.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 655, + 290, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 655, + 290, + 695 + ], + "spans": [ + { + "bbox": [ + 67, + 655, + 290, + 695 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 67, + 655, + 290, + 695 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 67, + 655, + 290, + 695 + ], + "type": "text", + "content": " is learnable and initialized to 1. On the training dataset " + }, + { + "bbox": [ + 67, + 655, + 290, + 695 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 67, + 655, + 290, + 695 + ], + "type": "text", + "content": ", the joint contrastive learning objective is formulated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 126, + 698, + 290, + 722 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 698, + 290, + 722 + ], + "spans": [ + { + "bbox": [ + 126, + 698, + 290, + 722 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\sum_ {i = 1} ^ {| \\mathcal {D} |} \\left(\\mathcal {L} _ {i} ^ {d} + \\mathcal {L} _ {i} ^ {p}\\right) \\tag {5}", + "image_path": "d679d97deecded3ff3276a5777675add1164ead12a1765f5164b02aad56afd10.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 266, + 490, + 279 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 266, + 490, + 279 + ], + "spans": [ + { + "bbox": [ + 302, + 266, + 490, + 279 + ], + "type": "text", + "content": "2.3 Model Initialization and Inference" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 283, + 525, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 283, + 525, + 391 + ], + "spans": [ + { + "bbox": [ + 302, + 283, + 525, + 391 + ], + "type": "text", + "content": "As UniTRec is a standard text-to-text Transformer, we initialize the parameters from pretrained BART (Lewis et al., 2020). In inference, UniTRec predicts the discriminative and perplexity-based scores for each candidate item, respectively. The two separate scores " + }, + { + "bbox": [ + 302, + 283, + 525, + 391 + ], + "type": "inline_equation", + "content": "S^d" + }, + { + "bbox": [ + 302, + 283, + 525, + 391 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 302, + 283, + 525, + 391 + ], + "type": "inline_equation", + "content": "S^p" + }, + { + "bbox": [ + 302, + 283, + 525, + 391 + ], + "type": "text", + "content": " are normalized, averaged, and finally ranked as the output. Detailed ranking process is provided in Appendix B." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 401, + 390, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 401, + 390, + 414 + ], + "spans": [ + { + "bbox": [ + 302, + 401, + 390, + 414 + ], + "type": "text", + "content": "3 Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 421, + 525, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 421, + 525, + 583 + ], + "spans": [ + { + "bbox": [ + 302, + 421, + 525, + 583 + ], + "type": "text", + "content": "We evaluate UniTRec on three text-based recommendation tasks: 1) NewsRec, to recommend news articles to users based on their browsing history. We use the MIND-small dataset (Wu et al., 2020) for experiments. 2) QuoteRec, to recommend quotations to users based on their conversation history. We use the Reddit-quotation dataset (Wang et al., 2021) for experiments. 3) EngageRec, to recommend social media posts for users to engage with based on their comment history. We use the dataset released by Zeng et al. (2020) for experiments. Detailed dataset statistics is provided in Appendix A." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 584, + 525, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 584, + 525, + 664 + ], + "spans": [ + { + "bbox": [ + 302, + 584, + 525, + 664 + ], + "type": "text", + "content": "Implementation Details. The UniTRec encoder and decoder both consist of 6 Transformer layers with 768-dimensional hidden states and 12 attention heads. We set " + }, + { + "bbox": [ + 302, + 584, + 525, + 664 + ], + "type": "inline_equation", + "content": "L_{1} = 3" + }, + { + "bbox": [ + 302, + 584, + 525, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 302, + 584, + 525, + 664 + ], + "type": "inline_equation", + "content": "L_{2} = 3" + }, + { + "bbox": [ + 302, + 584, + 525, + 664 + ], + "type": "text", + "content": ". We use AdamW optimizer (Loshchilov and Hutter, 2019) to train UniTRec with cosine learning rate decay." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 665, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 665, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 665, + 525, + 772 + ], + "type": "text", + "content": "Baselines. We compare UniTRec with competitive baselines: 1) GRU4Rec (Balázs et al., 2016) utilizes a GRU network to learn multi-turn history. 2) SASRec (Kang and McAuley, 2018) encodes user history with a self-attention based sequential model. 3) BERT4Rec (Sun et al., 2019) employs bidirectional self-attention to model user history. 4) RoBERTa-Sim, a simple yet strong baseline men" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 730, + 290, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 730, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 730, + 290, + 772 + ], + "type": "text", + "content": "3Note https://huggingface.co/docs/transformers/perplexity for LM perplexity calculation. We empirically discard the outer exponential term in the PPL formula, because it already exists in NCE loss Eq. (4) and does not affect the final ranking." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "1162" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 71, + 67, + 522, + 148 + ], + "blocks": [ + { + "bbox": [ + 71, + 67, + 522, + 148 + ], + "lines": [ + { + "bbox": [ + 71, + 67, + 522, + 148 + ], + "spans": [ + { + "bbox": [ + 71, + 67, + 522, + 148 + ], + "type": "table", + "html": "
ModelNewsRecQuoteRecEngageRec
MRRNDCG@5/10HR@5/10MRRNDCG@5/10HR@5/10MRRNDCG@5/10HR@5/10
GRU4Rec32.9136.20/42.5350.33/68.3534.0834.65/37.9344.45/54.632.121.04/1.511.27/2.65
SASRec32.6036.03/42.3750.63/68.6433.6334.30/37.4944.32/54.202.401.49/1.952.16/3.47
BERT4Rec32.8736.18/42.4050.21/67.9733.5934.26/37.2743.76/53.053.041.98/3.232.81/6.67
RoBERTa-Sim32.9636.47/42.8151.06/69.0837.1337.96/41.1848.14/58.063.742.66/3.754.42/7.70
UNBERT33.0936.53/42.8450.87/68.8239.7540.74/43.6950.90/60.042.831.96/2.673.11/5.24
UniTRec33.7637.63/43.7452.61/69.8941.2442.38/45.3152.87/61.884.063.23/4.294.58/7.68
", + "image_path": "d5616c46fab06466614308db2cbf65cc97437d9fc394f7a6c790be94c198ffa6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 71, + 204, + 522, + 285 + ], + "blocks": [ + { + "bbox": [ + 67, + 157, + 524, + 195 + ], + "lines": [ + { + "bbox": [ + 67, + 157, + 524, + 195 + ], + "spans": [ + { + "bbox": [ + 67, + 157, + 524, + 195 + ], + "type": "text", + "content": "Table 1: Experiment results on three text-based recommendation tasks. MRR denotes mean reciprocal rank, NDCG denotes normalized discounted cumulative gain, and HR denotes hit ratio (presented in percentage). The overall performance of UniTRec is better than other baseline models with " + }, + { + "bbox": [ + 67, + 157, + 524, + 195 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 67, + 157, + 524, + 195 + ], + "type": "text", + "content": "-value " + }, + { + "bbox": [ + 67, + 157, + 524, + 195 + ], + "type": "inline_equation", + "content": "< 0.05" + }, + { + "bbox": [ + 67, + 157, + 524, + 195 + ], + "type": "text", + "content": ", validated by unpaired t-test." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 71, + 204, + 522, + 285 + ], + "lines": [ + { + "bbox": [ + 71, + 204, + 522, + 285 + ], + "spans": [ + { + "bbox": [ + 71, + 204, + 522, + 285 + ], + "type": "table", + "html": "
ModelNewsRecQuoteRecEngageRec
MRRNDCG@5/10HR@5/10MRRNDCG@5/10HR@5/10MRRNDCG@5/10HR@5/10
UniTRec33.7637.63/43.7452.61/69.8941.2442.38/45.3152.87/61.884.063.23/4.294.58/7.68
w/o BART Init30.3133.32/39.6947.55/65.7819.0217.66/20.8022.45/32.162.240.86/1.611.27/3.62
w/o Local-Att33.3437.22/43.3252.28/69.5440.4441.63/44.5652.09/61.153.923.19/4.154.38/7.36
w/o Global-Att33.2237.06/43.1752.14/69.4740.2541.47/44.2652.07/60.763.642.78/3.593.89/6.35
Disc-Score only33.0736.76/43.0351.68/69.4640.5941.81/44.6552.39/61.143.822.99/3.604.49/6.85
PPL-Score only32.8336.39/42.5951.05/68.6740.3141.43/44.4752.13/61.203.292.39/3.033.86/5.66
", + "image_path": "7ce67a8851ac89198c0864f344a456b91b633e6a541a95b1778ca7924257b456.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 161, + 294, + 431, + 306 + ], + "lines": [ + { + "bbox": [ + 161, + 294, + 431, + 306 + ], + "spans": [ + { + "bbox": [ + 161, + 294, + 431, + 306 + ], + "type": "text", + "content": "Table 2: Recommendation performance of ablation model variants." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 327, + 290, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 327, + 290, + 407 + ], + "spans": [ + { + "bbox": [ + 67, + 327, + 290, + 407 + ], + "type": "text", + "content": "tioned in Qi et al. (2022), uses the hidden states of [CLS] tokens to measure user-candidate similarity. 5) UNBERT, implemented as Zhang et al. (2021), concatenates history and candidate texts as the input to BERT and predicts matching scores from the final hidden states of [CLS] tokens." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 409, + 291, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 409, + 291, + 476 + ], + "spans": [ + { + "bbox": [ + 67, + 409, + 291, + 476 + ], + "type": "text", + "content": "Note that we do not consider other methods that use non-text inputs (e.g., user profile, text topic labels). For fair comparison, all baseline models use pretrained 12-layer RoBERTa-base (Liu et al., 2019) as text encoders to learn embeddings of texts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 486, + 157, + 497 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 486, + 157, + 497 + ], + "spans": [ + { + "bbox": [ + 67, + 486, + 157, + 497 + ], + "type": "text", + "content": "3.1 Main Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 503, + 291, + 719 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 503, + 291, + 719 + ], + "spans": [ + { + "bbox": [ + 67, + 503, + 291, + 719 + ], + "type": "text", + "content": "Table 1 shows the performance of experiment models. From the results of NewsRec and QuoteRec, we can see that UniTRec outperforms all baseline models by a clear margin. Also, RoBERTa-Sim and UNBERT that directly use the [CLS] hidden states to represent user history, surpass other baselines that build additional aggregation networks upon the whole RoBERTa outputs. As displayed in the results, EngageRec is the most difficult task. We inspect the dataset and find that the texts on social media contain too much noise (e.g., URL and emoji), and the user history contains less number of turns. Nevertheless, UniTRec achieves better overall performance than other baseline models, validating its robustness on noisy text inputs and limited user history." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 729, + 238, + 741 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 729, + 238, + 741 + ], + "spans": [ + { + "bbox": [ + 67, + 729, + 238, + 741 + ], + "type": "text", + "content": "3.2 Ablation Studies and Analyses" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "type": "text", + "content": "We further conduct ablation studies on UniTRec. The experiment results are reported in Table 2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 327, + 525, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 327, + 525, + 422 + ], + "spans": [ + { + "bbox": [ + 302, + 327, + 525, + 422 + ], + "type": "text", + "content": "Initialization of UniTRec. We train UniTRec from scratch without initialization from pretrained BART (refer to w/o BART Init). The recommendation performance significantly drops in all three tasks, which indicates that acquiring effective text understanding ability from PLM is a necessary key to UniTRec performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 428, + 526, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 428, + 526, + 591 + ], + "spans": [ + { + "bbox": [ + 302, + 428, + 526, + 591 + ], + "type": "text", + "content": "Local and global attention. We investigate the function of two-level attention modules of the UniTRec history encoder. Concretely, we set " + }, + { + "bbox": [ + 302, + 428, + 526, + 591 + ], + "type": "inline_equation", + "content": "L_{1} = 0" + }, + { + "bbox": [ + 302, + 428, + 526, + 591 + ], + "type": "text", + "content": " in w/o Local-Att and " + }, + { + "bbox": [ + 302, + 428, + 526, + 591 + ], + "type": "inline_equation", + "content": "L_{2} = 0" + }, + { + "bbox": [ + 302, + 428, + 526, + 591 + ], + "type": "text", + "content": " in w/o Global-Att, where " + }, + { + "bbox": [ + 302, + 428, + 526, + 591 + ], + "type": "inline_equation", + "content": "L_{1} + L_{2} = 6" + }, + { + "bbox": [ + 302, + 428, + 526, + 591 + ], + "type": "text", + "content": ". We can observe that removing local and global attention from the original UniTRec history encoder both lead to suboptimal performance, while the performance drop is more significant in w/o Global-Att. The results justify the effectiveness of jointly modeling two-level history contexts through adapted Transformer attention masking without additional parameters." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 597, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 597, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 597, + 525, + 772 + ], + "type": "text", + "content": "Discriminative and perplexity-based objectives. We probe into training UniTRec with standalone discriminative (Disc-Score only) and perplexity-based (PPL-Score only) contrastive objectives, respectively. We can see that the discriminative objective yields better performance than the perplexity-based objective. Besides, the model performance on both standalone objectives declines compared to the original joint objective. The results indicate that the discriminative and perplexity-based matching scores are complementary and can jointly provide more accurate signals of user history and candidate text matching for text-based recommendation." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "1163" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 147, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 147, + 83 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 147, + 83 + ], + "type": "text", + "content": "4 Conclusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 93, + 291, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 93, + 291, + 186 + ], + "spans": [ + { + "bbox": [ + 67, + 93, + 291, + 186 + ], + "type": "text", + "content": "We present a unified Transformer UniTRec for text-based recommendation. UniTRec learns two-level contexts of multi-turn user history and jointly exploits discriminative matching scores and candidate text perplexity as matching objectives. Empirical experiments on three text-based recommendation datasets corroborate the effectiveness of UniTRec." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 197, + 149, + 210 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 197, + 149, + 210 + ], + "spans": [ + { + "bbox": [ + 67, + 197, + 149, + 210 + ], + "type": "text", + "content": "5 Limitations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 219, + 292, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 219, + 292, + 435 + ], + "spans": [ + { + "bbox": [ + 67, + 219, + 292, + 435 + ], + "type": "text", + "content": "Our model only focuses on utilizing text information for recommendation, which is a key limitation of this work. In real-world settings, recommender systems are usually required to handle heterogeneous information inputs. UniTRec is a pure text-based recommender modeling user history and candidate texts as inputs. However, incorporating additional side information (e.g., user profile, text topic labels, and dwell time of user behaviors) could further improve the recommendation performance and alleviate the cold start problem. Furthermore, UniTRec only models two-level relations of user behavior history. Nonetheless, incorporating more user behavior information, such as implicit and negative feedback, could further enhance the recommendation performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 447, + 170, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 447, + 170, + 460 + ], + "spans": [ + { + "bbox": [ + 68, + 447, + 170, + 460 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 468, + 291, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 468, + 291, + 521 + ], + "spans": [ + { + "bbox": [ + 67, + 468, + 291, + 521 + ], + "type": "text", + "content": "We appreciate constructive comments from anonymous reviewers. The research described in this paper is partially supported by CUHK under Project No. 3230366." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 545, + 127, + 557 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 545, + 127, + 557 + ], + "spans": [ + { + "bbox": [ + 68, + 545, + 127, + 557 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 565, + 292, + 772 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 69, + 565, + 291, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 565, + 291, + 632 + ], + "spans": [ + { + "bbox": [ + 69, + 565, + 291, + 632 + ], + "type": "text", + "content": "Hidasi Balázs, Karatzoglou Alexandros, Baltrunas Linas, and Tikk Domonkos. 2016. Session-based recommendations with recurrent neural networks. In 4th International Conference on Learning Representations ICLR 2016, San Juan, Puerto Rico, May 2-4, 2016, Conference Track Proceedings." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 640, + 292, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 640, + 292, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 640, + 292, + 772 + ], + "type": "text", + "content": "Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems," + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 303, + 72, + 527, + 772 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 314, + 72, + 525, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 525, + 94 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 525, + 94 + ], + "type": "text", + "content": "volume 33, pages 1877-1901. Curran Associates, Inc." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 103, + 527, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 103, + 527, + 170 + ], + "spans": [ + { + "bbox": [ + 304, + 103, + 527, + 170 + ], + "type": "text", + "content": "Daniel Cheng, Kyle Yan, Phillip Keung, and Noah A. Smith. 2022. The engage corpus: A social media dataset for text-based recommender systems. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 1885-1889, Marseille, France. European Language Resources Association." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 179, + 526, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 179, + 526, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 179, + 526, + 289 + ], + "type": "text", + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 4171-4186. Association for Computational Linguistics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 303, + 297, + 526, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 297, + 526, + 364 + ], + "spans": [ + { + "bbox": [ + 303, + 297, + 526, + 364 + ], + "type": "text", + "content": "Li Dong, Nan Yang, Wenhui Wang, Furu Wei, Xiaodong Liu, Yu Wang, Jianfeng Gao, Ming Zhou, and Hsiao-Wuen Hon. 2019. Unified language model pre-training for natural language understanding and generation. In 33rd Conference on Neural Information Processing Systems (NeurIPS 2019)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 373, + 527, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 527, + 451 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 527, + 451 + ], + "type": "text", + "content": "Shijie Geng, Shuchang Liu, Zuohui Fu, Yingqiang Ge, and Yongfeng Zhang. 2022. Recommendation as language processing (rlp): A unified pretrain, personalized prompt & predict paradigm (p5). In Proceedings of the 16th ACM Conference on Recommender Systems, RecSys '22, page 299-315, New York, NY, USA. Association for Computing Machinery." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 458, + 525, + 526 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 458, + 525, + 526 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 525, + 526 + ], + "type": "text", + "content": "Youyang Gu, Tao Lei, Regina Barzilay, and Tommi Jaakkola. 2016. Learning to refine text based recommendations. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2103-2108, Austin, Texas. Association for Computational Linguistics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 534, + 526, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 526, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 526, + 612 + ], + "type": "text", + "content": "Michael Gutmann and Aapo Hyvarinen. 2010. Noise-contrastive estimation: A new estimation principle for unnormalized statistical models. In Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics, volume 9 of Proceedings of Machine Learning Research, pages 297-304, Chia Laguna Resort, Sardinia, Italy. PMLR." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 619, + 526, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 619, + 526, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 619, + 526, + 665 + ], + "type": "text", + "content": "Wang-Cheng Kang and Julian McAuley. 2018. Self-attentive sequential recommendation. In 2018 IEEE International Conference on Data Mining (ICDM), pages 197-206." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 673, + 527, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 673, + 527, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 673, + 527, + 772 + ], + "type": "text", + "content": "Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871-7880, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "text", + "content": "1164" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 291, + 150 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 291, + 150 + ], + "type": "text", + "content": "Jian Li, Jieming Zhu, Qiwei Bi, Guohao Cai, Lifeng Shang, Zhenhua Dong, Xin Jiang, and Qun Liu. 2022. MINER: Multi-interest matching network for news recommendation. In Findings of the Association for Computational Linguistics: ACL 2022, pages 343-352, Dublin, Ireland. Association for Computational Linguistics." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 158, + 291, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 158, + 291, + 214 + ], + "spans": [ + { + "bbox": [ + 69, + 158, + 291, + 214 + ], + "type": "text", + "content": "Yize Li, Jiazhong Nie, Yi Zhang, Bingqing Wang, Baoshi Yan, and Fuliang Weng. 2010. Contextual recommendation based on text mining. In *Coling* 2010: Posters, pages 692-700, Beijing, China. Coling 2010 Organizing Committee." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 222, + 291, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 222, + 291, + 279 + ], + "spans": [ + { + "bbox": [ + 69, + 222, + 291, + 279 + ], + "type": "text", + "content": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. In arXiv preprint arXiv: 1907.11692. arXiv." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 286, + 291, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 286, + 291, + 341 + ], + "spans": [ + { + "bbox": [ + 69, + 286, + 291, + 341 + ], + "type": "text", + "content": "Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 350, + 291, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 350, + 291, + 417 + ], + "spans": [ + { + "bbox": [ + 69, + 350, + 291, + 417 + ], + "type": "text", + "content": "Itzik Malkiel, Oren Barkan, Avi Caciularu, Noam Razin, Ori Katz, and Noam Koenigstein. 2020. *RecoBERT: A catalog language model for text-based recommendations*. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 426, + 291, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 426, + 291, + 504 + ], + "spans": [ + { + "bbox": [ + 69, + 426, + 291, + 504 + ], + "type": "text", + "content": "Shumpei Okura, Yukihiro Tagami, Shingo Ono, and Akira Tajima. 2017. Embedding-based news recommendation for millions of users. In Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, page 1933-1942, New York, NY, USA. Association for Computing Machinery." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 511, + 291, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 511, + 291, + 590 + ], + "spans": [ + { + "bbox": [ + 69, + 511, + 291, + 590 + ], + "type": "text", + "content": "Fanchao Qi, Yanhui Yang, Jing Yi, Zhili Cheng, Zhiyuan Liu, and Maosong Sun. 2022. QuoteR: A benchmark of quote recommendation for writing. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 336-348, Dublin, Ireland. Association for Computational Linguistics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 597, + 291, + 698 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 597, + 291, + 698 + ], + "spans": [ + { + "bbox": [ + 69, + 597, + 291, + 698 + ], + "type": "text", + "content": "Tao Qi, Fangzhao Wu, Chuhan Wu, Peiru Yang, Yang Yu, Xing Xie, and Yongfeng Huang. 2021. HieRec: Hierarchical user interest modeling for personalized news recommendation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 5446-5456, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 706, + 291, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 706, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 706, + 291, + 772 + ], + "type": "text", + "content": "Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 526, + 716 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 314, + 72, + 524, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 524, + 105 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 524, + 105 + ], + "type": "text", + "content": "Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 8748-8763. PMLR." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 114, + 526, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 114, + 526, + 202 + ], + "spans": [ + { + "bbox": [ + 304, + 114, + 526, + 202 + ], + "type": "text", + "content": "Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang. 2019. Bert4rec: Sequential recommendation with bidirectional encoder representations from transformer. In Proceedings of the 28th ACM International Conference on Information and Knowledge Management, CIKM '19, page 1441-1450, New York, NY, USA. Association for Computing Machinery." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 210, + 526, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 210, + 526, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 210, + 526, + 277 + ], + "type": "text", + "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30, pages 5998-6008. Curran Associates, Inc." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 285, + 526, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 285, + 526, + 374 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 526, + 374 + ], + "type": "text", + "content": "Lingzhi Wang, Xingshan Zeng, and Kam-Fai Wong. 2021. Quotation recommendation and interpretation based on transformation from queries to quotations. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 754-758, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 381, + 526, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 381, + 526, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 381, + 526, + 460 + ], + "type": "text", + "content": "Chuhan Wu, Fangzhao Wu, Tao Qi, and Yongfeng Huang. 2021. Empowering news recommendation with pre-trained language models. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '21, page 1652-1656, New York, NY, USA. Association for Computing Machinery." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 467, + 526, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 467, + 526, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 467, + 526, + 555 + ], + "type": "text", + "content": "Fangzhao Wu, Ying Qiao, Jiun-Hung Chen, Chuhan Wu, Tao Qi, Jianxun Lian, Danyang Liu, Xing Xie, Jianfeng Gao, Winnie Wu, and Ming Zhou. 2020. MIND: A large-scale dataset for news recommendation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3597-3606, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 564, + 526, + 630 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 564, + 526, + 630 + ], + "spans": [ + { + "bbox": [ + 304, + 564, + 526, + 630 + ], + "type": "text", + "content": "Xingshan Zeng, Jing Li, Lu Wang, Zhiming Mao, and Kam-Fai Wong. 2020. Dynamic online conversation recommendation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3331-3341, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 638, + 526, + 716 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 638, + 526, + 716 + ], + "spans": [ + { + "bbox": [ + 304, + 638, + 526, + 716 + ], + "type": "text", + "content": "Qi Zhang, Jingjie Li, Qinglin Jia, Chuyuan Wang, Jieming Zhu, Zhaowei Wang, and Xiuqiang He. 2021. Umbert: User-news matching bert for news recommendation. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-21, pages 3356-3362. International Joint Conferences on Artificial Intelligence Organization. Main Track." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1165" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 68, + 296, + 124 + ], + "blocks": [ + { + "bbox": [ + 70, + 68, + 296, + 124 + ], + "lines": [ + { + "bbox": [ + 70, + 68, + 296, + 124 + ], + "spans": [ + { + "bbox": [ + 70, + 68, + 296, + 124 + ], + "type": "table", + "html": "
DatasetNewsRecQuoteRecEngageRec
Avg. history turns26.094.243.29
Avg. history tokens414.40279.82286.82
Avg. candidates37.2311117163
Avg. candidate tokens16.1519.11102.42
", + "image_path": "cee588b2bc04a0511e0aedc46adfb0b28ffdec05c6da196ac45df5a80b5bc0f0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 132, + 291, + 205 + ], + "lines": [ + { + "bbox": [ + 67, + 132, + 291, + 205 + ], + "spans": [ + { + "bbox": [ + 67, + 132, + 291, + 205 + ], + "type": "text", + "content": "Table 3: Statistics of three text-based recommendation training datasets. History and candidate tokens denote the number of BPE-tokenized tokens. The test set distribution is closed to the training sets (except candidates of EngageRec) and hence omitted. Note that the max length of each history log is truncated to 1024 tokens." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 222, + 180, + 235 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 222, + 180, + 235 + ], + "spans": [ + { + "bbox": [ + 67, + 222, + 180, + 235 + ], + "type": "text", + "content": "A Dataset Statistics" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 243, + 291, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 243, + 291, + 405 + ], + "spans": [ + { + "bbox": [ + 67, + 243, + 291, + 405 + ], + "type": "text", + "content": "The detailed statistics of the three text-based recommendation datasets are displayed in Table 3. Note that we use news titles as the text inputs for NewsRec following Qi et al. (2021). NewsRec regards the user clicked and non-clicked news as candidate texts, while QuoteRec and EngageRec regard all potential quotation texts and post texts as candidates. Different from Zeng et al. (2020) that formulates the task as recommending candidate users to given posts based on post content, we formulate the task as recommending candidate posts to given users based on user history." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 416, + 254, + 429 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 416, + 254, + 429 + ], + "spans": [ + { + "bbox": [ + 69, + 416, + 254, + 429 + ], + "type": "text", + "content": "Algorithm 1 Candidate Ranking Processes" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 433, + 276, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 433, + 276, + 455 + ], + "spans": [ + { + "bbox": [ + 69, + 433, + 276, + 455 + ], + "type": "text", + "content": "Input: discriminative scores " + }, + { + "bbox": [ + 69, + 433, + 276, + 455 + ], + "type": "inline_equation", + "content": "S^d = \\{S_1^d,S_2^d,\\dots,S_M^d\\}" + }, + { + "bbox": [ + 69, + 433, + 276, + 455 + ], + "type": "text", + "content": " perplexity-based scores " + }, + { + "bbox": [ + 69, + 433, + 276, + 455 + ], + "type": "inline_equation", + "content": "S^{p} = \\{S_{1}^{p},S_{2}^{p},\\dots,S_{M}^{p}\\}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 455, + 199, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 455, + 199, + 465 + ], + "spans": [ + { + "bbox": [ + 69, + 455, + 199, + 465 + ], + "type": "text", + "content": "Output: final averaged ranking " + }, + { + "bbox": [ + 69, + 455, + 199, + 465 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 69, + 455, + 199, + 465 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 76, + 465, + 290, + 555 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 465, + 289, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 465, + 289, + 484 + ], + "spans": [ + { + "bbox": [ + 76, + 465, + 289, + 484 + ], + "type": "text", + "content": "1: Derive the normalized discriminative scores " + }, + { + "bbox": [ + 76, + 465, + 289, + 484 + ], + "type": "inline_equation", + "content": "S_{norm}^{d} =" + }, + { + "bbox": [ + 76, + 465, + 289, + 484 + ], + "type": "text", + "content": " softmax(Sd)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 76, + 485, + 289, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 485, + 289, + 504 + ], + "spans": [ + { + "bbox": [ + 76, + 485, + 289, + 504 + ], + "type": "text", + "content": "2: Derive the normalized perplexity-based scores " + }, + { + "bbox": [ + 76, + 485, + 289, + 504 + ], + "type": "inline_equation", + "content": "S_{norm}^{p} =" + }, + { + "bbox": [ + 76, + 485, + 289, + 504 + ], + "type": "text", + "content": " softmax( " + }, + { + "bbox": [ + 76, + 485, + 289, + 504 + ], + "type": "inline_equation", + "content": "S^p" + }, + { + "bbox": [ + 76, + 485, + 289, + 504 + ], + "type": "text", + "content": " )." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 76, + 505, + 290, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 505, + 290, + 525 + ], + "spans": [ + { + "bbox": [ + 76, + 505, + 290, + 525 + ], + "type": "text", + "content": "3: Derive the geometric average scores " + }, + { + "bbox": [ + 76, + 505, + 290, + 525 + ], + "type": "inline_equation", + "content": "\\bar{S} = \\log (S_{norm}^d) + \\log (S_{norm}^p)" + }, + { + "bbox": [ + 76, + 505, + 290, + 525 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 76, + 526, + 289, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 526, + 289, + 545 + ], + "spans": [ + { + "bbox": [ + 76, + 526, + 289, + 545 + ], + "type": "text", + "content": "4: Sort the averaged scores " + }, + { + "bbox": [ + 76, + 526, + 289, + 545 + ], + "type": "inline_equation", + "content": "\\bar{S}" + }, + { + "bbox": [ + 76, + 526, + 289, + 545 + ], + "type": "text", + "content": " by descending order to derive the final ranking: " + }, + { + "bbox": [ + 76, + 526, + 289, + 545 + ], + "type": "inline_equation", + "content": "\\bar{R} \\gets \\mathrm{Rank}_{\\mathrm{des}}(\\bar{S})" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 76, + 546, + 126, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 546, + 126, + 555 + ], + "spans": [ + { + "bbox": [ + 76, + 546, + 126, + 555 + ], + "type": "text", + "content": "5: return " + }, + { + "bbox": [ + 76, + 546, + 126, + 555 + ], + "type": "inline_equation", + "content": "R" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 68, + 579, + 186, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 579, + 186, + 592 + ], + "spans": [ + { + "bbox": [ + 68, + 579, + 186, + 592 + ], + "type": "text", + "content": "B Inference Ranking" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 67, + 600, + 291, + 735 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 600, + 291, + 735 + ], + "spans": [ + { + "bbox": [ + 67, + 600, + 291, + 735 + ], + "type": "text", + "content": "Given the user history and " + }, + { + "bbox": [ + 67, + 600, + 291, + 735 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 67, + 600, + 291, + 735 + ], + "type": "text", + "content": " candidate texts, UniTRec first predicts the discriminative ranking scores " + }, + { + "bbox": [ + 67, + 600, + 291, + 735 + ], + "type": "inline_equation", + "content": "S^d = \\{S_1^d,S_2^d,\\dots,S_M^d\\}" + }, + { + "bbox": [ + 67, + 600, + 291, + 735 + ], + "type": "text", + "content": " and perplexity-based ranking scores " + }, + { + "bbox": [ + 67, + 600, + 291, + 735 + ], + "type": "inline_equation", + "content": "S^{p} = \\{S_{1}^{p},S_{2}^{p},\\dots,S_{M}^{p}\\}" + }, + { + "bbox": [ + 67, + 600, + 291, + 735 + ], + "type": "text", + "content": " of the candidates. Algorithm 1 outlines an approach to aggregate the final ranking based on " + }, + { + "bbox": [ + 67, + 600, + 291, + 735 + ], + "type": "inline_equation", + "content": "S^d" + }, + { + "bbox": [ + 67, + 600, + 291, + 735 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 600, + 291, + 735 + ], + "type": "inline_equation", + "content": "S^p" + }, + { + "bbox": [ + 67, + 600, + 291, + 735 + ], + "type": "text", + "content": ". Note that the function " + }, + { + "bbox": [ + 67, + 600, + 291, + 735 + ], + "type": "inline_equation", + "content": "\\mathrm{Rank}(S)^4" + }, + { + "bbox": [ + 67, + 600, + 291, + 735 + ], + "type": "text", + "content": " denotes outputting the sorted order of elements in a score list " + }, + { + "bbox": [ + 67, + 600, + 291, + 735 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 67, + 600, + 291, + 735 + ], + "type": "text", + "content": ". There exist other ways to average the ranking of " + }, + { + "bbox": [ + 67, + 600, + 291, + 735 + ], + "type": "inline_equation", + "content": "S^d" + }, + { + "bbox": [ + 67, + 600, + 291, + 735 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 600, + 291, + 735 + ], + "type": "inline_equation", + "content": "S^p" + }, + { + "bbox": [ + 67, + 600, + 291, + 735 + ], + "type": "text", + "content": ", which we leave for future work to explore." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 303, + 70, + 430, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 70, + 430, + 84 + ], + "spans": [ + { + "bbox": [ + 303, + 70, + 430, + 84 + ], + "type": "text", + "content": "C Qualitative Analysis" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 302, + 92, + 526, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 92, + 526, + 147 + ], + "spans": [ + { + "bbox": [ + 302, + 92, + 526, + 147 + ], + "type": "text", + "content": "We show randomly sampled outputs of UniTRec, for instance, demonstrated on the news recommendation and quote recommendation tasks. Table 4 and 5 showcase the qualitative samples." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 740, + 290, + 773 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 740, + 290, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 740, + 290, + 773 + ], + "type": "inline_equation", + "content": "^4" + }, + { + "bbox": [ + 67, + 740, + 290, + 773 + ], + "type": "text", + "content": "Rank(S) works similarly to scipy.stats.rankdata(). For example in ascending order, " + }, + { + "bbox": [ + 67, + 740, + 290, + 773 + ], + "type": "inline_equation", + "content": "\\mathrm{Ran_{asc}}(\\{0.2, 0.6, 0.7, 0.4\\}) = \\mathrm{scipy.stats.rankdata}([0.2, 0.6, 0.7, 0.4]) = [1, 3, 4, 2]" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "text", + "content": "1166" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 149, + 523, + 258 + ], + "blocks": [ + { + "bbox": [ + 69, + 149, + 523, + 258 + ], + "lines": [ + { + "bbox": [ + 69, + 149, + 523, + 258 + ], + "spans": [ + { + "bbox": [ + 69, + 149, + 523, + 258 + ], + "type": "table", + "html": "
TurnHistory News Texts
#1Mac Engel: As long as these results are acceptable, Dallas Cowboys will continue to be losers
#2NFL world reacts to officials handing Packers win over Lions
#3Maryland Congressman Elijah Cummings, a Democrat and Chair of House Oversight and Reform Committee, has died: CNN
#4Unprecedented movement detected on California earthquake fault capable of 8.0 temblor
#5Bag Explodes While Being Loaded On Volaris Flight At Midway Airport
#6Orlando Scandrick rips Eagles: They have "accountability issues"
#7Meghan King Edmonds, Jim Edmonds' Nanny Denies Cheating Allegations
#8Nearly $400M worth of cocaine and marijuana intercepted by US Coast Guard
#9Former NBA first-round pick arrested in sex sting operation
#10China's trade with US shrinks in October despite optimism
", + "image_path": "7e5a33c24396675552e585b8b72a78527dafeb15cdd5137b89b50a70b3050531.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 70, + 265, + 523, + 364 + ], + "blocks": [ + { + "bbox": [ + 70, + 265, + 523, + 364 + ], + "lines": [ + { + "bbox": [ + 70, + 265, + 523, + 364 + ], + "spans": [ + { + "bbox": [ + 70, + 265, + 523, + 364 + ], + "type": "table", + "html": "
Candidate News TextsSdSpRClicked
Taylor Swift Rep Hits Back at Big Machine, Claims She's Actually Owed $7.9 Million in Unpaid Royalties0.0950.0694X
Former North Carolina State, NBA player Anthony Grundy dies in stabbing, police say0.1720.1553X
13 Reasons Why's Christian Navarro Slams Disney for Casting "the White Guy" in The Little Mermaid0.0480.0657X
Opinion: Colin Kaepernick is about to get what he deserves: a chance0.3030.2501✓
3 Indiana judges suspended after a night of drinking turned into a White Castle brawl0.0760.0595X
66 Cool Tech Gifts Anyone Would Be Thrilled to Receive0.0090.0059X
Police find 26 children behind false wall at Colorado day care0.0340.1166X
I've been writing about tiny homes for a year and spent 2 nights in a 300-foot home to see what it is all about0.0290.0198X
Report: Police investigating woman's death after Redskins' player Montae Nicholson took her to hospital0.2350.2612✓
", + "image_path": "78d34d5e2448d59210826ce9ff2e37c62420605836ead25f73126d6614991189.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 69, + 400, + 523, + 489 + ], + "blocks": [ + { + "bbox": [ + 185, + 370, + 407, + 380 + ], + "lines": [ + { + "bbox": [ + 185, + 370, + 407, + 380 + ], + "spans": [ + { + "bbox": [ + 185, + 370, + 407, + 380 + ], + "type": "text", + "content": "(i) Qualitative Example-A from news recommendation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 400, + 523, + 489 + ], + "lines": [ + { + "bbox": [ + 69, + 400, + 523, + 489 + ], + "spans": [ + { + "bbox": [ + 69, + 400, + 523, + 489 + ], + "type": "table", + "html": "
TurnHistory News Texts
#1Toddler dancing to celebrate 11 months cancer-free goes viral
#2NFL Week 8 Power Rankings: Old-school football rules the day
#3The 25 US cities where it's easiest to get a mortgage
#4Burning questions for Cowboys vs Giants on "Monday Night Football"
#5Who's the favorite to win 2019 NFL rushing title?
#6Grading all 32 NFL teams heading into the last eight weeks of the 2019 season
#7Jennifer Aniston looks amazing in a makeup-free selfie, plus more news
#8This $12 million "mansion yacht" is made entirely of stainless steel and it's a first for the industry. Take a peek inside
", + "image_path": "623da3869d86d65f0daf66f9f99d3cc4428c59f94f4599544aee48256d8e2177.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 71, + 497, + 521, + 607 + ], + "blocks": [ + { + "bbox": [ + 71, + 497, + 521, + 607 + ], + "lines": [ + { + "bbox": [ + 71, + 497, + 521, + 607 + ], + "spans": [ + { + "bbox": [ + 71, + 497, + 521, + 607 + ], + "type": "table", + "html": "
Candidate News TextsSdSpRClicked
Opinion: Colin Kaepernick is about to get what he deserves: a chance0.3300.4001✓
U.S. Troops Will Die If They Remain in Syria, Bashar Al-Assad Warns0.0240.01110✗
Pete Davidson, Kaia Gerber Are Dating, Trying to Stay "Low Profile"0.0640.0336✗
The Hottest Tech Gifts This Holiday Season0.0500.0278✗
Taylor Swift Rep Hits Back at Big Machine, Claims She's Actually Owed $7.9 Million in Unpaid Royalties0.0460.0387✗
13 Reasons Why's Christian Navarro Slams Disney for Casting "the White Guy" in The Little Mermaid0.0600.0964✓
Some believe Mason Rudolph, hit in head with his own helmet, isn't getting enough blame0.1540.1792✓
South Carolina teen gets life in prison for deadly elementary school shooting0.0660.0465✗
The Unlikely Star of My Family's Thanksgiving Table0.0470.0219✗
Police investigating woman's death after Redskins' player Montae Nicholson took her to hospital0.1580.1493✗
", + "image_path": "3e351c6e27ce61a217bda77205221667d02759d07489a0de62e543df624e0ecc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 183, + 611, + 408, + 622 + ], + "lines": [ + { + "bbox": [ + 183, + 611, + 408, + 622 + ], + "spans": [ + { + "bbox": [ + 183, + 611, + 408, + 622 + ], + "type": "text", + "content": "(ii) Qualitative Example-B from news recommendation." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 640, + 525, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 640, + 525, + 689 + ], + "spans": [ + { + "bbox": [ + 67, + 640, + 525, + 689 + ], + "type": "text", + "content": "Table 4: Case analyses of news recommendation. History News Texts are sorted by user-clicked timestamps. " + }, + { + "bbox": [ + 67, + 640, + 525, + 689 + ], + "type": "inline_equation", + "content": "S^d" + }, + { + "bbox": [ + 67, + 640, + 525, + 689 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 67, + 640, + 525, + 689 + ], + "type": "inline_equation", + "content": "S^p" + }, + { + "bbox": [ + 67, + 640, + 525, + 689 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 67, + 640, + 525, + 689 + ], + "type": "inline_equation", + "content": "\\bar{R}" + }, + { + "bbox": [ + 67, + 640, + 525, + 689 + ], + "type": "text", + "content": " are normalized discriminative, perplexity-based scores, and average ranking as described in Appendix B. Clicked denotes the ground truth user-click labels. Note that the experiment history logs are anonymized and delinked, which is always the first priority of the recommendation study." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "1167" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 68, + 523, + 186 + ], + "blocks": [ + { + "bbox": [ + 69, + 68, + 523, + 186 + ], + "lines": [ + { + "bbox": [ + 69, + 68, + 523, + 186 + ], + "spans": [ + { + "bbox": [ + 69, + 68, + 523, + 186 + ], + "type": "table", + "html": "
TurnConversation Threading History
#1I own an FJ. It's a great car and even on stockies. It's great offroad.
#2I feel bad for you that you run the risk of being associated with the typical FJ owner.
#3What is a typical FJ owner? I've not heard anything bad about FJ owners.
#4It's like someone who drives a jeep wrangler in NYC. There's no need. Tons of FJ owners do that have it and not use it for what it's made for.
#5God forbid someone likes the design of a car and doesn't use it offroad.
#6Then buy a much more economic environmentalist friendly version. If you buy something and always use it for much less than it's purpose, why buy it?
#7Or people can buy whatever the hell they want because it's their money and not yours.
#8You're entirely right. Just like people can be rude just because you can do it, because you have the ability but why should you ass.
#9I wasn't aware that somebody buying a vehicle that they like and you don't was morally wrong.
#10I love FJs. It's perfectly fine to buy whatever you think looks nice.
", + "image_path": "b1f9abb62882335295bc058772916268cad21fe7a1d0a5746ddce5588e3aa348.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 87, + 195, + 505, + 303 + ], + "blocks": [ + { + "bbox": [ + 87, + 195, + 505, + 303 + ], + "lines": [ + { + "bbox": [ + 87, + 195, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 87, + 195, + 505, + 303 + ], + "type": "table", + "html": "
Candidate Quote Texts\\( S^d \\)\\( S^P \\)\\( \\bar{R} \\)Ground truth
Beauty is in the eye of the beholder.0.4800.4711✓
A fool and his money are soon parted.0.1760.1402
Form follows function.0.0510.0463
Everything is worth what its purchaser will pay for it.0.0400.0584
Because it's there.0.0380.0295
You can't fix stupid.0.0210.0346
The lady doth protest too much, methinks.0.0220.0137
It's all about the money.0.0200.0138
Anybody driving slower than you is an idiot, and anyone going faster than you is a maniac?0.0120.0189
Opportunity is missed by most people.0.0180.00810
", + "image_path": "196da27e5d28c749cbc667bcd84304052778a81fecdd12867e526adf4d514a1a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 69, + 339, + 523, + 618 + ], + "blocks": [ + { + "bbox": [ + 181, + 308, + 411, + 319 + ], + "lines": [ + { + "bbox": [ + 181, + 308, + 411, + 319 + ], + "spans": [ + { + "bbox": [ + 181, + 308, + 411, + 319 + ], + "type": "text", + "content": "(iii) Qualitative Example-C from quote recommendation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 339, + 523, + 618 + ], + "lines": [ + { + "bbox": [ + 69, + 339, + 523, + 618 + ], + "spans": [ + { + "bbox": [ + 69, + 339, + 523, + 618 + ], + "type": "table", + "html": "
TurnConversation Threading History
#1Society is becoming more efficient, which is a good thing. People should realize there's no point in holding back this technology just for the sake of keeping people employed. If this were beneficial, then calculators and computers shouldn't exist either.
#2One small problem is that people need to pay rent and eat.
#3So we should ditch computers and go back to the typing pool? Should we get rid of heavy earth moving equipment and just use hundreds of guys with hand tools to build everything? It would employ a hell of a lot more people.
#4No one's saying that. I don't think anyone is really against automation, but as it increases, there are soon going to be more people that there are jobs that actually need doing. I actually believe we've already passed this point. So what do we do with the people, who can't get jobs simply because there are none? It's an issue that need assessed immediately.
#5Tons and tons and tons of American jobs have been replaced by new jobs created by technology or in support of technology years ago. An office might have needed people to handle filing paperwork, keeping it in order, and retrieving, where now a document management system has made them completely redundant. The upshot is that to access that DMS, people are out there selling computers, installing computers, servicing computers, and supporting end users building the servers installing, supporting monitoring backing them up, and all that jobs that come in support of those progress is progress. And it advances human efficiency and knowledge. These are just one or two examples, but the answer is not to kill progress. Other countries simply won't. The answer is to push education to the forefront, so people are prepared for these jobs and whatever other challenges the future may bring.
#6This is true. But it's unfortunate technological advances tend to reduce low skill jobs and replace them with high skill jobs. It would feel more fair if the low skilled workers could all do training programs and become high skilled workers. But this isn't really the case. Those jobs end up being taken by someone who had better educational opportunities or someone younger who still has time to take advantage of education.
#7The reality is the reality. Unfortunately or not educating people will create more educated people to handle high skill jobs, and I'll tell you being a desktop support technician isn't high skill. As that's where we push in the future, any amount of hand wringing won't change the facts. We must educate our people if we want to be a global leader in more than homelessness poverty.
#8Education won't matter. We are at the end of the job age at some point in the near future. We are going to have to deal with the fact that getting a job isn't a reality for a significant percentage of the population. Society will have to radically change as it did during the industrial revolution.
#9Much cheaper to heavily discourage having more children free abortions. Then in years there won't be so many useless people who can apparently be replaced by a simple robot.
#10Virtually every job will be replaced by automation name skilled trades that can't be automated. I imagine you'd be surprised at how hard this is. Are pharmacists useless, surgeons, accountants? I'd bet that your job is just as replaceable as these.
", + "image_path": "50d963228235f11772ef629de780327ca9a770d9d496f5bd0aa2fcdd8470d501.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 133, + 625, + 460, + 735 + ], + "blocks": [ + { + "bbox": [ + 133, + 625, + 460, + 735 + ], + "lines": [ + { + "bbox": [ + 133, + 625, + 460, + 735 + ], + "spans": [ + { + "bbox": [ + 133, + 625, + 460, + 735 + ], + "type": "table", + "html": "
Candidate Quote TextsSdSpRGround truth
There's no such thing as a free lunch.0.3650.4171
I can't predict the future.0.1850.2102✓
I have never let my schooling interfere with my education.0.1040.0593
Prevention is better than cure.0.0440.0834
Knowledge is power.0.0590.0525
Don't let schooling interfere with your education.0.0440.0436
Nature abhors a vacuum.0.0360.0247
There is no substitute for hard work.0.0240.0178
There are three kinds of lies: lies, damned lies, and statistics.0.0220.0139
You can't fix stupid.0.0190.01010
", + "image_path": "895d3eb52d984a32d554e29ffc6f140ebfd2e1ce8023b632c7a3e11ed7c0bd71.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 181, + 739, + 411, + 750 + ], + "lines": [ + { + "bbox": [ + 181, + 739, + 411, + 750 + ], + "spans": [ + { + "bbox": [ + 181, + 739, + 411, + 750 + ], + "type": "text", + "content": "(iv) Qualitative Example-D from quote recommendation." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 769, + 524, + 793 + ], + "lines": [ + { + "bbox": [ + 67, + 769, + 524, + 793 + ], + "spans": [ + { + "bbox": [ + 67, + 769, + 524, + 793 + ], + "type": "text", + "content": "Table 5: Case analyses of quote recommendation. We demonstrate the candidate quotes of the top 10 rankings out of all candidates. Note that there is only one ground truth quote for each conversation history." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "spans": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "type": "text", + "content": "A For every submission:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 106, + 414, + 242 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "spans": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "type": "text", + "content": "A1. Did you describe the limitations of your work? Section 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 143, + 329, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 143, + 329, + 169 + ], + "spans": [ + { + "bbox": [ + 77, + 143, + 329, + 169 + ], + "type": "text", + "content": "A2. Did you discuss any potential risks of your work? We see no concern about potential risks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 179, + 414, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 179, + 414, + 205 + ], + "spans": [ + { + "bbox": [ + 77, + 179, + 414, + 205 + ], + "type": "text", + "content": "A3. Do the abstract and introduction summarize the paper's main claims? Section 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 77, + 215, + 398, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 215, + 398, + 242 + ], + "spans": [ + { + "bbox": [ + 77, + 215, + 398, + 242 + ], + "type": "text", + "content": "A4. Have you used AI writing assistants when working on this paper? Left blank." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 68, + 252, + 291, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 252, + 291, + 266 + ], + "spans": [ + { + "bbox": [ + 68, + 252, + 291, + 266 + ], + "type": "text", + "content": "B Did you use or create scientific artifacts?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 79, + 270, + 267, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 270, + 267, + 283 + ], + "spans": [ + { + "bbox": [ + 79, + 270, + 267, + 283 + ], + "type": "text", + "content": "The Abstract provides the link to our code." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 76, + 292, + 524, + 634 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 292, + 315, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 292, + 315, + 319 + ], + "spans": [ + { + "bbox": [ + 76, + 292, + 315, + 319 + ], + "type": "text", + "content": "B1. Did you cite the creators of artifacts you used? Not applicable. Left blank." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 76, + 328, + 463, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 328, + 463, + 355 + ], + "spans": [ + { + "bbox": [ + 76, + 328, + 463, + 355 + ], + "type": "text", + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Not applicable. Left blank." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 76, + 364, + 524, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 364, + 524, + 432 + ], + "spans": [ + { + "bbox": [ + 76, + 364, + 524, + 432 + ], + "type": "text", + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Not applicable. Left blank." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 76, + 441, + 524, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 441, + 524, + 494 + ], + "spans": [ + { + "bbox": [ + 76, + 441, + 524, + 494 + ], + "type": "text", + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Not applicable. Left blank." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 77, + 503, + 524, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 503, + 524, + 544 + ], + "spans": [ + { + "bbox": [ + 77, + 503, + 524, + 544 + ], + "type": "text", + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? In the Abstract, a Github repository with documentation is released." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 77, + 553, + 524, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 553, + 524, + 634 + ], + "spans": [ + { + "bbox": [ + 77, + 553, + 524, + 634 + ], + "type": "text", + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Appendix A" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "spans": [ + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "type": "text", + "content": "C Did you run computational experiments?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 79, + 662, + 122, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 662, + 122, + 673 + ], + "spans": [ + { + "bbox": [ + 79, + 662, + 122, + 673 + ], + "type": "text", + "content": "Section 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 77, + 683, + 524, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 683, + 524, + 723 + ], + "spans": [ + { + "bbox": [ + 77, + 683, + 524, + 723 + ], + "type": "text", + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Section 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 67, + 729, + 522, + 748 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 729, + 522, + 748 + ], + "spans": [ + { + "bbox": [ + 67, + 729, + 522, + 748 + ], + "type": "text", + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "spans": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "text", + "content": "ACL 2023 Responsible NLP Checklist" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1169" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 76, + 70, + 524, + 238 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 77, + 70, + 523, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 70, + 523, + 110 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 523, + 110 + ], + "type": "text", + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Section 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 77, + 120, + 524, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 120, + 524, + 174 + ], + "spans": [ + { + "bbox": [ + 77, + 120, + 524, + 174 + ], + "type": "text", + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Section 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 76, + 184, + 524, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 184, + 524, + 238 + ], + "spans": [ + { + "bbox": [ + 76, + 184, + 524, + 238 + ], + "type": "text", + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Not applicable. Left blank." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 67, + 247, + 522, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 247, + 522, + 278 + ], + "spans": [ + { + "bbox": [ + 67, + 247, + 522, + 278 + ], + "type": "text", + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 76, + 286, + 524, + 539 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 76, + 286, + 524, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 286, + 524, + 326 + ], + "spans": [ + { + "bbox": [ + 76, + 286, + 524, + 326 + ], + "type": "text", + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Not applicable. Left blank." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 76, + 336, + 524, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 336, + 524, + 390 + ], + "spans": [ + { + "bbox": [ + 76, + 336, + 524, + 390 + ], + "type": "text", + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Not applicable. Left blank." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 76, + 400, + 524, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 400, + 524, + 454 + ], + "spans": [ + { + "bbox": [ + 76, + 400, + 524, + 454 + ], + "type": "text", + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Not applicable. Left blank." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 76, + 462, + 519, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 462, + 519, + 489 + ], + "spans": [ + { + "bbox": [ + 76, + 462, + 519, + 489 + ], + "type": "text", + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Not applicable. Left blank." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 76, + 498, + 524, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 498, + 524, + 539 + ], + "spans": [ + { + "bbox": [ + 76, + 498, + 524, + 539 + ], + "type": "text", + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Not applicable. Left blank." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 780, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 780, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 780, + 310, + 791 + ], + "type": "text", + "content": "1170" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Unsupervised Subtitle Segmentation with Masked Language Models/d5fa5db2-10b5-46c6-ab9c-a92346dcaabb_content_list.json b/2023/Unsupervised Subtitle Segmentation with Masked Language Models/d5fa5db2-10b5-46c6-ab9c-a92346dcaabb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b1f9512ed6f4c5b706a57650f00ba31985418d2b --- /dev/null +++ b/2023/Unsupervised Subtitle Segmentation with Masked Language Models/d5fa5db2-10b5-46c6-ab9c-a92346dcaabb_content_list.json @@ -0,0 +1,1408 @@ +[ + { + "type": "text", + "text": "Unsupervised Subtitle Segmentation with Masked Language Models", + "text_level": 1, + "bbox": [ + 146, + 89, + 852, + 111 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "David Ponce\\*1,2 and Thierry Etchegoyhen\\*1 and Victor Ruiz", + "bbox": [ + 228, + 135, + 769, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Vicomtech Foundation, Basque Research and Technology Alliance (BRTA)", + "bbox": [ + 189, + 153, + 811, + 168 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ University of the Basque Country UPV/EHU", + "bbox": [ + 310, + 170, + 689, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{adponce,tetchegoyhen,vruiz}@vicomtech.org", + "bbox": [ + 250, + 187, + 754, + 203 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 252, + 339, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We describe a novel unsupervised approach to subtitle segmentation, based on pretrained masked language models, where line endings and subtitle breaks are predicted according to the likelihood of punctuation to occur at candidate segmentation points. Our approach obtained competitive results in terms of segmentation accuracy across metrics, while also fully preserving the original text and complying with length constraints. Although supervised models trained on in-domain data and with access to source audio information can provide better segmentation accuracy, our approach is highly portable across languages and domains and may constitute a robust off-the-shelf solution for subtitle segmentation.", + "bbox": [ + 141, + 278, + 460, + 505 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 114, + 517, + 260, + 532 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Subtitling is one of the principal means of providing accessible audiovisual content. With the ever increasing production of audiovisual content in multiple domains and languages, in the current digital era, subtitle provision can benefit from automation support, via Automatic Speech Recognition and/or Machine Translation (Volk et al., 2010; Aliprandi et al., 2014; Etchegoyhen et al., 2014; Tardel, 2020; Bojar et al., 2021).", + "bbox": [ + 112, + 542, + 489, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Subtitles are subject to specific constraints in order to achieve adequate readability, including layout, on-screen duration and text editing. Among these constraints, segmentation addresses the maximum number of characters per line, the number of lines per subtitle, and breaks at natural linguistic frontiers. Segmentation has been shown to be an important readability factor (Perego et al., 2010; Rajendran et al., 2013), with improperly segmented subtitles resulting in increased cognitive effort and reading times for users. Thus, automated subtitle systems need to generate properly segmented subtitles to achieve readability.", + "bbox": [ + 112, + 687, + 489, + 896 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A typical baseline for subtitle segmentation, still used in some production systems, is simple character counting, whereby line breaks are inserted before reaching the maximum allowed number of characters per line. Although simple and fast, this approach does not address the need for linguistically correct segments and, therefore, falls short in terms of readability. Several approaches have been proposed to improve segmentation by automated means. Álvarez et al. (2014) proposed a machine learning method where subtitle breaks are predicted by Support Vector Machine and Linear Regression models trained on professionally-created subtitles. A similar method based on Conditional Random Fields was then shown to improve over these results (Alvarez et al., 2017). Approaches that directly generate subtitle breaks within Neural Machine Translation have also been proposed in recent years (Matusov et al., 2019; Karakanta et al., 2020a). Recently, Papi et al. (2022) developed a multilingual segmenter which generates both text and breaks and may be trained on textual input only, or on joint text and audio data.", + "bbox": [ + 507, + 253, + 884, + 621 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although quality subtitle segmentation may be achieved with the aforementioned approaches, they require supervised training on segmented subtitle corpora. At present, the largest subtitle corpus is Open Subtitles (Lison et al., 2018), which mainly covers entertainment material, contains subtitles mostly created by non-professionals or automatically translated, and does not include line breaks. The MuST-Cinema corpus (Karakanta et al., 2020b), on the other hand, is a multilingual speech translation corpus that includes subtitles breaks, but is only available for 8 languages at the moment. Considering the vast amount of languages and domains in audiovisual content, the lack of segmented training data hinders the development of robust automated subtitleing systems.", + "bbox": [ + 507, + 626, + 884, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we describe a novel unsupervised method based on pretrained masked language mod", + "bbox": [ + 507, + 887, + 882, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*These authors contributed equally to this work.", + "bbox": [ + 136, + 904, + 428, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "771", + "bbox": [ + 485, + 927, + 512, + 940 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics", + "bbox": [ + 226, + 945, + 769, + 957 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Volume 2: Short Papers, pages 771-781", + "bbox": [ + 376, + 958, + 618, + 971 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "July 9-14, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 295, + 972, + 700, + 984 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "els (MLM), where line and subtitle breaks are inserted according to the likelihood of a segment acting as an isolated unit, as approximated by the probability of a punctuation mark occurring at a given segmentation point. In our experiments, this novel approach obtained competitive results on most metrics, while also fully preserving the original text and complying with length constraints. Our system may thus be used as a simple yet efficient subtitle segmenter with any pretrained masked language model, for any language covered by the model.", + "bbox": [ + 112, + 84, + 490, + 261 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Approach", + "text_level": 1, + "bbox": [ + 112, + 275, + 235, + 291 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our approach is based on the standard view that the more appropriate subtitle segments are those that may function as isolated grammatical chunks. We further hypothesise that a relevant approximation for the identification of this type of unit is the likelihood of a punctuation mark being inserted at the end of a candidate segment, as punctuation may mark the closure of a syntactic unit and is often associated with discursive pauses. To test this hypothesis, we compute the likelihood of punctuation marks at different segmentation points, as predicted by a pretrained MLM, and select the insertion point with the highest likelihood.1", + "bbox": [ + 112, + 302, + 489, + 508 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The segmentation candidates are determined under a sliding-window approach over the entire input text. We first generate the list of all pairs $< \\alpha, \\beta>$ over the unprocessed portion of the text, where $\\alpha$ is a segmentation candidate of length under a specified limit $K$ , corresponding to the maximum number of characters per line, and $\\beta$ is the remaining portion of the text to be segmented.", + "bbox": [ + 112, + 511, + 489, + 640 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We then score all segmentation candidates $\\alpha$ with one of the LM scoring variants described below. A segmentation marker, either end-of-line (), or end-of-block indicating the end of a subtitle (), is then appended to the best scoring candidate, and $\\beta$ becomes the input text to be segmented in a recursive iteration of the process.", + "bbox": [ + 112, + 640, + 489, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Since our method does not rely on any additional information, such as an audio source, to determine the segmentation type, an tag is inserted every even segment or when $\\beta$ is empty; otherwise, an tag is inserted. We thus generate subtitles with a maximum of two lines, following a standard recommendation in subtitling. We also define a minimal number of characters (min) in $\\alpha$ for the", + "bbox": [ + 112, + 753, + 489, + 882 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "segmentation process to apply, and do not segment lines that are under the specified character limit.", + "bbox": [ + 507, + 84, + 880, + 115 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluated three approaches to compute segmentation scores over each candidate pair $< \\alpha, \\beta>$ :", + "bbox": [ + 507, + 117, + 882, + 148 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Substitution: The last token of $\\alpha$ is masked and the score is the highest MLM probability among punctuation marks on this mask.", + "- Insertion: A mask is appended to $\\alpha$ and the score is the highest MLM probability among punctuation marks on this mask.", + "- LM-Score: The score is the average of the perplexity of $\\alpha$ and $\\beta$ , as derived from the MLM probabilities for each token in the corresponding sequence." + ], + "bbox": [ + 531, + 158, + 880, + 338 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The first two methods are variants of our core approach. The third method, while also based on the same pretrained MLM, relies instead on the pseudoperplexity of the sequences according to the MLM, computed following Salazar et al. (2020). We included this latter variant to measure the potential of using LM scoring directly, without resorting to the likelihood of punctuation marks.", + "bbox": [ + 507, + 349, + 882, + 476 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3 Experimental Setup", + "text_level": 1, + "bbox": [ + 507, + 489, + 717, + 506 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Corpora. For all experiments, we used the MustST-Cinema corpus (Karakanta et al., 2020b), which is derived from TED talks and contains both line and subtitle break markers. In addition to being publicly available, it also allows for a direct comparison with the supervised models of Papi et al. (2022). We report results of our approach on the 6 MuST-Cinema datasets for which comparative results were available, directly predicting segmentation on the test sets without any training.", + "bbox": [ + 507, + 514, + 882, + 675 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Methods. For our approach, we tested the three variants described in Section 2. We used BERT (Devlin et al., 2019) as our MLM for all languages. Additionally, we included a variant called overt clueing $(OC)$ , where an overt punctuation mark at the end of a candidate segment increments the mask score by 1. We then compared the results of the best LM-based variant with those obtained by alternative approaches. In all cases, our results were computed with $min = 15$ , as this value obtained the best results overall over the development", + "bbox": [ + 507, + 684, + 882, + 860 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "$^{2}$ Our results on all remaining languages of the MuST-Cinema datasets are presented in Appendix B.", + "bbox": [ + 507, + 866, + 882, + 892 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "${}^{3}$ Specifically bert-base-uncased as available on HugginFace (https://huggingface.co/),accessed on November 2022.", + "bbox": [ + 507, + 892, + 882, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1 Throughout our experiments, we used the following punctuation marks: ' ', ' ', ' ?', ' !', ' ' and ' '.", + "bbox": [ + 112, + 891, + 489, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "772", + "bbox": [ + 485, + 927, + 515, + 940 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/30f4ac866468a79f0f6ce81bb1faac41531cef49cb27a9058316e4c83b9fbc3b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
EnglishSpanishGerman
MethodSigmaEOLEOBSigmaEOLEOBSigmaEOLEOB
Substitution71.65+19.86-10.9669.34+12.36-5.7469.31+19.05-7.05
Insertion76.77+19.18-9.9173.47+12.98-4.9170.85+18.53-7.96
LM-Score69.97+21.40-8.6667.70+13.29-5.3764.07+16.45-6.51
", + "bbox": [ + 141, + 80, + 857, + 200 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1: Sigma and break coverage test set results for LM-based segmentation variants", + "bbox": [ + 203, + 209, + 793, + 225 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sets, although the differences were minor with the other values we tested (1, 10 and 20).4", + "bbox": [ + 112, + 249, + 487, + 281 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We used the simple character counting approach (hereafter, CountChars) as baseline, and, as representative supervised methods on the selected datasets, the models described by (Papi et al., 2022). Their core supervised approach is based on a Transformer (Vaswani et al., 2017) architecture with 3 encoder layers and 3 decoder layers, trained on textual MuST-Cinema input only (MC.Text), or on complementary audio data as well via an additional speech encoder with 12 encoder layers (MC.Multi). They trained each variant on either monolingual data alone (mono), or in a multilingual setting (multi). Finally, they also report results for a variant (OS.Text) trained on the Open Subtitles corpus (Lison et al., 2018) for their zero-shot experiments.", + "bbox": [ + 115, + 284, + 489, + 541 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Evaluation. We use the subtitle-oriented metric Sigma (Karakanta et al., 2022), which computes the ratio of achieved BLEU (Papineni et al., 2002) over an approximated upper-bound BLEU score, on text that includes line and subtitle breaks. Sigma is meant to support the evaluation of imperfect texts, i.e. text that differs from the reference when breaks are omitted. Although our approach does not produce imperfect text, achieving perfect BLEU scores when breaks are ignored, we used this metric for comparison purposes. We also report break coverage results (Papi et al., 2022), defined as the ratio of predicted breaks over reference breaks, which we computed separately for the EOL and EOB breaks. Finally, we include length conformity results (CPL), measured as the percentage of subtitle lines whose length is under the maximum number of characters defined by the subtitle guidelines (42 in the TED guidelines $^5$ ).", + "bbox": [ + 112, + 557, + 489, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4 LM-based Segmentation Variants", + "text_level": 1, + "bbox": [ + 507, + 249, + 836, + 266 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first compared the three methods described in Section 2 on the English, Spanish and German datasets, with the results described in Table 1. In terms of Sigma, the Insertion method obtained the best results in all cases. It also obtained the best scores in terms of coverage for the EOL marker, except in Spanish, although all three variants tend to overgenerate end-of-line markers to similar extents. The LM-Score variant obtained the worst results in terms of Sigma, but outperformed the alternatives in terms of EOB coverage, a metric on which the three variants performed markedly better than on EOL coverage. Considering the overall results, we selected the Insertion variant as the most balanced one for all remaining experiments reported below.", + "bbox": [ + 507, + 274, + 885, + 517 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "5 Comparative Results", + "text_level": 1, + "bbox": [ + 507, + 527, + 726, + 545 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In Table 2, we present the results obtained by the selected approaches on the languages for which results were available with supervised models trained on in-domain data. Overall, our approach outperformed the CountChars baseline across the board, and was in turn outperformed by the supervised variants in terms of Sigma scores. Although it is clear from these results that training segmentation models on in-domain data, with or without audio data, provides clear advantages in terms of subtitle segmentation, it is worth noting that Sigma does not, by design, reflect the actual BLEU score without breaks, i.e. the generation of imperfect text, which is a by-product of the above supervised approaches and non-existent in ours. In terms of CPL, all supervised models generate subtitle lines that overflow the limit, to a significant degree, whereas the selected unsupervised models trivially respect the length constraint.", + "bbox": [ + 507, + 554, + 885, + 860 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "The results indicated in Table 3 on unseen data seem to indicate that their MC.Multi model can reach BLEU scores close to 100, thereby limiting the negative impact of imperfect text generation in these cases.", + "bbox": [ + 507, + 869, + 884, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "4See Appendix C for results with different values of the min parameter.", + "bbox": [ + 112, + 878, + 485, + 904 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "5https://www.ted.com/participate/translate/subtitling-tips", + "bbox": [ + 132, + 904, + 482, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "773", + "bbox": [ + 485, + 927, + 515, + 940 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/f30c2fad5dd35b609c4efd393d7c19953b6c96cf42537d7e72eb9e22dfd5f454.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
EnglishFrenchGermanItalian
MethodTrainingSigmaCPLSigmaCPLSigmaCPLSigmaCPL
CountCharsN/A63.71100%62.87100%62.34100%61.49100%
MC.Textmono84.8796.6%83.6896.7%83.6290.9%82.2290.0%
multi85.9888.5%84.5694.3%84.0290.9%83.0491.2%
MC.Multimono85.7694.8%84.2593.9%84.2291.4%82.6289.9%
multi87.4495.0%86.4994.1%86.4089.9%85.3390.0%
MLMN/A76.77100%73.78100%70.85100%71.38100%
MLM+OCN/A77.89100%76.07100%75.63100%74.20100%
", + "bbox": [ + 132, + 80, + 865, + 291 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/5197bb7b925f215a4dcf35c2683653aba92c2ed88e910d73c03d8095ef1bb50d.jpg", + "table_caption": [ + "Table 2: Comparative results between unsupervised methods and supervised approaches trained on in-domain data" + ], + "table_footnote": [], + "table_body": "
Dutch
MethodBLEUSigmaCPLEOLEOB
CountChars10063.2100%-21.2-7.1
OS.Text89.564.471.2%-31.4-51.3
MC.Text61.374.477.8%-23.4-9.9
MC.Multi99.980.391.4%-27.20.4
MLM10068.7100%+20.4-10.0
MLM+OC10073.9100%+21.2-10.0
Spanish
MethodBLEUSigmaCPLEOLEOB
CountChars10063.2100%-24.6-4.4
OS.Text92.664.171.2%-32.3-45.4
MC.Text69.675.870.1%-47.6-19.3
MC.Multi99.678.791.8%-22.44.7
MLM10073.5100%+13.0-4.9
MLM+OC10075.6100%+13.4-4.6
", + "bbox": [ + 129, + 337, + 470, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 3: Comparative results between unsupervised methods and zero-short supervised approaches", + "bbox": [ + 112, + 692, + 485, + 722 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Table 3, we show the comparative results between the selected unsupervised methods and the supervised variants, in languages where zero-shot results were available for the latter approaches. In this scenario, in terms of Sigma our approach obtained results on a par with the supervised MC.Text models trained on in-domain MuST-Cinema data, outperformed the OS.Text models trained on Open Subtitles data, and was surpassed by the MC.Multi model, which exploits additional audio information,", + "bbox": [ + 112, + 758, + 489, + 917 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "by 3.1 and 6.4 points. In terms of break coverage, in most cases our unsupervised method outperformed the supervised variants, to a significant degree compared to the text-based OS.Text and MC.Text models. Regarding BLEU scores without breaks, only the MC.Multi model reaches a score close to the perfect one achieved by the unsupervised models, whereas the MC.Text model is outperformed by 38.7 and 31.4 points in Dutch and Spanish, respectively. In all cases, the CPL scores indicate that none of the supervised approaches fully meet the length constraint, leading to overflowing lines in $8.2\\%$ of the cases at best and $29.9\\%$ at worst. In this scenario as well, the unsupervised approaches fully meet the length constraint, by design.", + "bbox": [ + 507, + 340, + 884, + 582 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overall, overt clueing improved over our core method by an average of 3.12 Sigma points, indicating that some likely punctuation configurations were not properly captured by our MLM approximation. In general, our approach tends to overgenerate EOL markers, whereas the opposite is true for the selected supervised models. Determining which of these tendencies leads to better subtitle readability would require a specific human evaluation which we leave for future research.", + "bbox": [ + 507, + 589, + 884, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although the zero-shot Sigma results obtained by the supervised MC.Multi method show the potential of this approach to provide pretrained models applicable to other languages, two important aspects are worth considering. First, the available zero-shot results were obtained on datasets in the same domain as the data seen to train the supervised models. A more complete assessment of the capabilities of these models in zero-shot settings, which would be the most frequent scenario consid-", + "bbox": [ + 507, + 758, + 885, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "774", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ering the lack of training data across domains and languages, would require specific evaluations in other domains. Secondly, although segmentation is a key aspect for subtitle readability, length conformity is an equally important constraint, if not more so considering that subtitles with lines over the CPL limit are considered invalid in subtitleing. Our proposed unsupervised method can thus be seen as a pragmatic approach which guarantees valid subtitles while also providing quality segmentation across the board. $^7$", + "bbox": [ + 112, + 84, + 489, + 261 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6 Conclusions", + "text_level": 1, + "bbox": [ + 112, + 275, + 253, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We described an unsupervised approach to subtitle segmentation, based on pretrained masked language models, where line or subtitle breaks are inserted according to the likelihood of punctuation occurring at candidate segmentation points.", + "bbox": [ + 112, + 304, + 489, + 384 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although supervised models, trained on indomain data with audio support, were shown to perform better that this simple textual approach in terms of the Sigma metric, they tend to generate imperfect text to varying degrees, while also failing to fully meet length constraints that are essential for subtitling.", + "bbox": [ + 112, + 386, + 489, + 498 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In contrast, our LM-based textual approach outperformed supervised models in most cases in terms of break generation coverage, while also fully preserving the original text, complying with length constraints, and obtaining competitive results in terms of Sigma. This simple approach may thus provide a highly portable complementary solution for subtitle segmentation across languages and domains.", + "bbox": [ + 112, + 500, + 489, + 644 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "7 Limitations", + "text_level": 1, + "bbox": [ + 112, + 659, + 250, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The first clear limitation of our approach is its text-based nature. This prevents important audio information, typically silences in speech patterns, from being exploited to generate subtitle breaks. A more complete system could be devised though, for instance by associating our text-based approach with the information provided by a forced alignment toolkit, whenever audio information is available. A simple method along these lines could be the following: 1. Apply our MLM-based segmentation but only generating a unique segmentation tag SEG; 2. Insert EOB markers wherever the", + "bbox": [ + 112, + 687, + 489, + 879 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "silence between two aligned words is above a specified threshold; 3. Traverse the text sequentially and replace SEG with EOL if there exists a previous marker of type EOB, otherwise replace with EOB. We left this use of our method in combination with audio information for future research, as audio alignment for subtitles typically involves additional factors such as non-literal transcriptions.", + "bbox": [ + 507, + 84, + 884, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Additionally, our method is limited in its adaptability to specific segmentation guidelines, which may be company-specific. The main adaptable parameters of our methods are the minimum and maximum parameters of the segmentation window, and the set of predefined punctuation marks over which masking is computed, neither of which could fully model idiosyncratic segmentation guidelines. However, in our experience at least, segmentation in real professional data tends to display varying degrees of consistency with respect to guidelines, and natural linguistic breaks seem to be the dominant factor for subtitle segmentation. A specific evaluation would be needed on data from varied professional datasets to determine the extent to which our method might deviate from specific guidelines.", + "bbox": [ + 507, + 214, + 884, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, other aspects of subtitling, such as the recommendation in some guidelines for subtitles to appear in a pyramidal view, i.e. with the first line shorter than the second line, have not been taken into consideration in this work. Our aim was to evaluate our core LM-based approach without additional variables that can vary across guidelines and may also have led to results that are more difficult to interpret overall. Our approach could nonetheless be easily augmented with constraints on relative line lengths within subtitles, by incrementing the scores of segmentation candidates that respect this surface-level constraint.", + "bbox": [ + 507, + 473, + 882, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "8 Ethical Considerations", + "text_level": 1, + "bbox": [ + 507, + 697, + 741, + 713 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our approach involves the use of large pretrained language models, whose computational performance is typically higher when deployed in more powerful environments with GPUs. Under such usage, electric consumption and associated carbon footprint are likely to increase and users of our method under these conditions should be aware of this type of impact. However, subtitle segmentation is often performed offline, where efficient processing is less of a concern, and lower-cost CPU deployments are an entirely viable option. All our results were obtained with a single large LM de", + "bbox": [ + 507, + 726, + 884, + 917 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "Examples of segmented subtitles can be found in Appendix A.", + "bbox": [ + 112, + 891, + 489, + 917 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "775", + "bbox": [ + 485, + 927, + 515, + 940 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ployed on CPU, with the aim of reducing energy consumption at inference time.", + "bbox": [ + 112, + 84, + 485, + 116 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Additionally, our method requires no training for the task at hand and thus removes the cost of model training associated with the supervised methods with which we compare our results. For instance, Papi et al. (2022) indicate that they use four K80 GPUs to train their models, which we took as comparison points, with 1 day of training for their text-only models and 1 week for their multimodal segmenters. Therefore, given the large number of potential language pairs and domains in need of segmented subtitle content, our approach can provide competitive results with a comparatively lesser impact on energy resource consumption.", + "bbox": [ + 112, + 118, + 489, + 326 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 114, + 342, + 285, + 357 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We thank the anonymous reviewers for their helpful comments. This work was partially supported by the Department of Economic Development and Competitiveness of the Basque Government (Spri Group) through funding for the StreAmS project (ZL-2021/00700).", + "bbox": [ + 112, + 370, + 487, + 464 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 114, + 495, + 213, + 511 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Carlo Aliprandi, Cristina Scudellari, Isabella Gallucci, Nicola Piccinini, Matteo Raffaelli, Arantza del Pozo, Aitor Alvarez, Haritz Arzelus, Renato Cassaca, Tiago Luis, et al. 2014. Automatic live subtitling: state of the art, expectations and current trends. In Proceedings of NAB Broadcast Engineering Conference: Papers on Advanced Media Technologies, Las Vegas, volume 13.", + "Aitor Álvarez, Haritz Arzelus, and Thierry Etchegoyhen. 2014. Towards customized automatic segmentation of subtitles. In Advances in Speech and Language Technologies for Iberian Languages, pages 229-238. Springer.", + "Aitor Alvarez, Carlos-D Martínez-Hinarejos, Haritz Arzelus, Marina Balenciaga, and Arantza del Pozo. 2017. Improving the automatic segmentation of subtitles through conditional random field. Speech Communication, 88:83-95.", + "Ondrej Bojar, Dominik Machacek, Sangeet Sagar, Otakar Smrz, Jonas Kratochvil, Peter Polak, Ebrahim Ansari, Mohammad Mahmoudi, Rishu Kumar, Dario Franceschini, Chiara Canton, Ivan Simonini, Thai Son Nguyen, Felix Schneider, Sebastian Stüker, Alex Waibel, Barry Haddow, Rico Sennrich, and Philip Williams. 2021. ELITR multilingual live subtitling: Demo and strategy. In Proceedings of the 16th Conference of the European Chapter of the Association" + ], + "bbox": [ + 114, + 521, + 489, + 917 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for Computational Linguistics: System Demonstrations, pages 271-277, Online. Association for Computational Linguistics.", + "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics.", + "Thierry Etchegoyhen, Lindsay Bywood, Mark Fishel, Panayota Georgakopoulou, Jie Jiang, Gerard van Loenhout, Arantza del Pozo, Mirjam Sepesy Maučec, Anja Turner, and Martin Volk. 2014. Machine translation for subtitling: A large-scale evaluation. In Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC'14), pages 46-53, Reykjavik, Iceland. European Language Resources Association (ELRA).", + "Alina Karakanta, François Buet, Mauro Cettolo, and François Yvon. 2022. Evaluating subtitle segmentation for end-to-end generation systems. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 3069-3078, Marseille, France. European Language Resources Association.", + "Alina Karakanta, Matteo Negri, and Marco Turchi. 2020a. Is 42 the answer to everything in subtitling-oriented speech translation? In Proceedings of the 17th International Conference on Spoken Language Translation, pages 209-219, Online. Association for Computational Linguistics.", + "Alina Karakanta, Matteo Negri, and Marco Turchi. 2020b. MuST-cinema: a speech-to-subtitles corpus. In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 3727-3734, Marseille, France. European Language Resources Association.", + "Pierre Lison, Jörg Tiedemann, and Milen Kouylekov. 2018. OpenSubtitles2018: Statistical rescoring of sentence alignments in large, noisy parallel corpora. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources Association (ELRA).", + "Evgeny Matusov, Patrick Wilken, and Yota Georgakopoulou. 2019. Customizing neural machine translation for subtitling. In Proceedings of the Fourth Conference on Machine Translation (Volume 1: Research Papers), pages 82-93, Florence, Italy. Association for Computational Linguistics.", + "Sara Papi, Alina Karakanta, Matteo Negri, and Marco Turchi. 2022. Dodging the data bottleneck: Automatic subtitling with automatically segmented ST corpora. In Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint" + ], + "bbox": [ + 510, + 85, + 884, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "776", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Conference on Natural Language Processing (Volume 2: Short Papers), pages 480-487, Online only. Association for Computational Linguistics.", + "bbox": [ + 132, + 85, + 489, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311-318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics.", + "bbox": [ + 115, + 131, + 489, + 227 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Elisa Perego, Fabio Del Missier, Marco Porta, and Mauro Mosconi. 2010. The cognitive effectiveness of subtitle processing. *Media psychology*, 13(3):243-272.", + "bbox": [ + 115, + 233, + 489, + 287 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dhevi J Rajendran, Andrew T Duchowski, Pilar Orero, Juan Martínez, and Pablo Romero-Fresco. 2013. Effects of text chunking on subtitling: A quantitative and qualitative examination. Perspectives, 21(1):5-21.", + "bbox": [ + 115, + 294, + 489, + 360 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Julian Salazar, Davis Liang, Toan Q. Nguyen, and Katrin Kirchhoff. 2020. Masked language model scoring. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2699-2712, Online. Association for Computational Linguistics.", + "bbox": [ + 115, + 367, + 489, + 447 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Anke Tardel. 2020. Effort in semi-automatized subtitling processes: speech recognition and experience during transcription. Journal of Audiovisual Translation, 3(2):79-102.", + "bbox": [ + 115, + 454, + 489, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30.", + "bbox": [ + 115, + 514, + 487, + 581 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Martin Volk, Rico Sennrich, Christian Hardmeier, and Frida Tidström. 2010. Machine translation of TV subtitles for large scale production. In Proceedings of the Second Joint EM+/CNGL Workshop: Bringing MT to the User: Research on Integrating MT in the Translation Industry, pages 53-62, Denver, Colorado, USA. Association for Machine Translation in the Americas.", + "bbox": [ + 115, + 588, + 489, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A Segmentation Examples", + "text_level": 1, + "bbox": [ + 114, + 701, + 361, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4 provides examples of subtitles in the MuST-Cinema test sets segmented with either the character counting baseline or our LM-based approach, in its insertion variant without resorting to overt punctuation clueing.", + "bbox": [ + 112, + 726, + 489, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In these examples, the MLM approach generates end-of-line and end-of-subtitle breaks that are overall in line with natural linguistic breaks, contrary to the character counting baseline. As such, on either short, medium or longer input, the readability of the generated subtitles is significantly enhanced with our approach.", + "bbox": [ + 112, + 807, + 489, + 917 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "B Extended Results", + "text_level": 1, + "bbox": [ + 509, + 83, + 697, + 99 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results presented in Section 5 were limited to the subset of languages and metrics for which published comparative results were available on the MuST-Cinema datasets. In Table 5, we present the complete list of results obtained with our method, for all languages and metrics. The selected variant of our method is the insertion masking approach, which was selected for the main results in our paper, with a segmentation window starting at 15 characters and ending at 42. We do not include BLEU scores computed over text that includes segmentation breaks, as the results are identical to those obtained with the Sigma metric for our approach, which does not generate imperfect text.", + "bbox": [ + 507, + 109, + 884, + 332 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Across languages, the results are relatively uniform, with the best Sigma scores obtained in English and the lowest in Dutch, for a difference of 4.1 points between the two languages. In terms of break coverage, the best results were obtained for Spanish and the worst for Romanian, although results were also relatively uniform across languages. In all cases, overt clueing, where overt punctuation marks raised the LM score by 1, improved Sigma scores, although it had less of an impact on break coverage results, where both variants performed similarly overall.", + "bbox": [ + 507, + 335, + 884, + 527 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "C Results With Different min Parameters", + "text_level": 1, + "bbox": [ + 507, + 539, + 880, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As noted in Section 3, considering preliminary results over the development set we selected a default value of 15 for the min parameter, which indicates the number of characters after which the segmentation process applies. In Table 6, we present comparative results on the test sets with different min values. In terms of Sigma, values of 15 and 20 led to rather similar results; values of 1 and 10 resulted in slightly lower results, with the lowest results achieved with the former.", + "bbox": [ + 507, + 565, + 882, + 722 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In terms of $\\langle \\mathrm{eol} \\rangle$ and $\\langle \\mathrm{eob} \\rangle$ coverage, the former increases with larger min values, which is expected given the more restricted space to insert these end-of-line markers as the value increases; for $\\langle \\mathrm{eob} \\rangle$ , the restricted insertion space results in increased under-generation, which in turn results in better scores for lower values of the min parameter.", + "bbox": [ + 507, + 726, + 884, + 839 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "777", + "bbox": [ + 485, + 927, + 515, + 940 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b6a7d0fb463c965476b7e4eb46ca26d6a8a8f8078ca9a5f77957f8f62a656370.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CountCharsMLM
They're things you access through your <eol> computer. <eob>They're things you access <eol> through your computer. <eob>
Every row of data is a life whose story <eol> deserves to be told with dignity. <eob>Every row of data is a life <eol> whose story deserves to be told <eob> with dignity. <eob>
During the winter, struggling to get <eol> warm, my neighbors would have no choice <eob> but to bypass the meter after their heat <eol> was shut off, just to keep their family <eob> comfortable for one more day. <eob>During the winter, struggling to get warm, <eol> my neighbors would have no choice <eob> but to bypass the meter <eol> after their heat was shut off, <eob> just to keep their family comfortable <eol> for one more day. <eob>
", + "bbox": [ + 115, + 134, + 884, + 360 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c0115864e1110fdf64a9df994c03b8e27f7acb000e1aa76ebbaa90140b68bbbc.jpg", + "table_caption": [ + "Table 4: Examples of subtitles segmented via character counting and MLM-based mask insertion" + ], + "table_footnote": [], + "table_body": "
LanguageMethodBLEUSigmaEOLEOBCPL
DEMLM10070.8518.53-7.96100%
MLM+OC10075.6319.81-7.78100%
ENMLM10076.7719.18-9.91100%
MLM+OC10077.8919.86-9.73100%
ESMLM10073.4712.98-4.91100%
MLM+OC10075.5913.45-4.63100%
FRMLM10073.7816.51-6.58100%
MLM+OC10076.0717.47-6.12100%
ITMLM10071.3818.49-9.55100%
MLM+OC10074.2020.34-8.57100%
NLMLM10068.7120.37-9.96100%
MLM+OC10073.8821.22-9.96100%
PTMLM10071.5920.03-10.81100%
MLM+OC10075.5019.87-10.02100%
ROMLM10069.4523.37-10.44100%
MLM+OC10074.1323.37-10.09100%
", + "bbox": [ + 236, + 497, + 761, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5: Complete results with MLM mask insertion on the MuST-Cinema test sets (min=15)", + "bbox": [ + 179, + 846, + 815, + 860 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "778", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/30ea88503d9dc83caefc2d1e1026152ad155e5b098d62964a7a68aa6b88b3fd9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
LanguageminBLEUSigmaEOLEOB
DE110072.3128.75-0.18
1010073.9622.68-4.43
1510075.6319.81-7.78
2010075.2814.54-11.21
EN110074.3037.33-0.98
1010077.1424.49-7.77
1510077.8919.86-9.73
2010077.1615.24-12.68
ES110073.0020.870.28
1010074.3218.24-2.04
1510075.5913.45-4.63
2010075.838.66-7.87
FR110073.8924.68-0.73
1010075.2620.83-3.93
1510076.0717.47-6.12
2010076.7512.5-10.05
IT110072.0129.75-3.66
1010073.7524.71-6.61
1510074.2020.34-8.57
2010073.6614.62-11.61
NL110072.1626.83-5.47
1010073.5623.26-8.47
1510073.8821.22-9.96
2010074.4016.81-12.43
PT110072.8726.38-6.24
1010074.5322.15-8.08
1510075.5019.87-10.02
2010074.9814.17-13.36
RO110072.0532.3-4.51
1010073.7626.98-7.52
1510074.1323.37-10.09
2010074.8917.53-12.83
", + "bbox": [ + 302, + 187, + 694, + 785 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 6: Test set results with the MLM+OC method and different values of the min parameter", + "bbox": [ + 181, + 793, + 815, + 809 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "779", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "A For every submission:", + "bbox": [ + 115, + 107, + 322, + 122 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "A1. Did you describe the limitations of your work? 7", + "A2. Did you discuss any potential risks of your work? 8", + "A3. Do the abstract and introduction summarize the paper's main claims?", + "A4. Have you used AI writing assistants when working on this paper? Left blank." + ], + "bbox": [ + 129, + 126, + 695, + 288 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "B Did you use or create scientific artifacts?", + "bbox": [ + 114, + 300, + 490, + 316 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Not applicable. Left blank.", + "bbox": [ + 131, + 321, + 332, + 336 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "B1. Did you cite the creators of artifacts you used? Not applicable. Left blank.", + "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Not applicable. Left blank.", + "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Not applicable. Left blank.", + "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Not applicable. Left blank.", + "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Not applicable. Left blank.", + "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Not applicable. Left blank." + ], + "bbox": [ + 127, + 347, + 880, + 753 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "C Did you run computational experiments?", + "bbox": [ + 114, + 764, + 492, + 781 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "3", + "bbox": [ + 134, + 787, + 146, + 799 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?", + "bbox": [ + 129, + 813, + 880, + 844 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We didn't trained any models for this paper, and inference was performed on CPU.", + "bbox": [ + 149, + 845, + 759, + 860 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.", + "bbox": [ + 112, + 868, + 877, + 892 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "ACL 2023 Responsible NLP Checklist", + "bbox": [ + 132, + 84, + 433, + 99 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "780", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? 3", + "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? 5", + "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?" + ], + "bbox": [ + 129, + 83, + 878, + 279 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?", + "text_level": 1, + "bbox": [ + 112, + 293, + 877, + 309 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 132, + 313, + 213, + 329 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Not applicable. Left blank.", + "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Not applicable. Left blank.", + "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Not applicable. Left blank.", + "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Not applicable. Left blank.", + "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Not applicable. Left blank." + ], + "bbox": [ + 127, + 340, + 878, + 639 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "781", + "bbox": [ + 485, + 928, + 512, + 940 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Unsupervised Subtitle Segmentation with Masked Language Models/d5fa5db2-10b5-46c6-ab9c-a92346dcaabb_model.json b/2023/Unsupervised Subtitle Segmentation with Masked Language Models/d5fa5db2-10b5-46c6-ab9c-a92346dcaabb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..159a52c87f0f3caf594aaa42ffb7cf2214dd6422 --- /dev/null +++ b/2023/Unsupervised Subtitle Segmentation with Masked Language Models/d5fa5db2-10b5-46c6-ab9c-a92346dcaabb_model.json @@ -0,0 +1,1762 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.147, + 0.09, + 0.853, + 0.112 + ], + "angle": 0, + "content": "Unsupervised Subtitle Segmentation with Masked Language Models" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.136, + 0.77, + 0.153 + ], + "angle": 0, + "content": "David Ponce\\*1,2 and Thierry Etchegoyhen\\*1 and Victor Ruiz" + }, + { + "type": "text", + "bbox": [ + 0.19, + 0.154, + 0.813, + 0.17 + ], + "angle": 0, + "content": "1 Vicomtech Foundation, Basque Research and Technology Alliance (BRTA)" + }, + { + "type": "text", + "bbox": [ + 0.312, + 0.171, + 0.69, + 0.187 + ], + "angle": 0, + "content": "\\(^{2}\\) University of the Basque Country UPV/EHU" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.189, + 0.756, + 0.204 + ], + "angle": 0, + "content": "{adponce,tetchegoyhen,vruiz}@vicomtech.org" + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.253, + 0.341, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.279, + 0.461, + 0.506 + ], + "angle": 0, + "content": "We describe a novel unsupervised approach to subtitle segmentation, based on pretrained masked language models, where line endings and subtitle breaks are predicted according to the likelihood of punctuation to occur at candidate segmentation points. Our approach obtained competitive results in terms of segmentation accuracy across metrics, while also fully preserving the original text and complying with length constraints. Although supervised models trained on in-domain data and with access to source audio information can provide better segmentation accuracy, our approach is highly portable across languages and domains and may constitute a robust off-the-shelf solution for subtitle segmentation." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.518, + 0.262, + 0.533 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.543, + 0.49, + 0.687 + ], + "angle": 0, + "content": "Subtitling is one of the principal means of providing accessible audiovisual content. With the ever increasing production of audiovisual content in multiple domains and languages, in the current digital era, subtitle provision can benefit from automation support, via Automatic Speech Recognition and/or Machine Translation (Volk et al., 2010; Aliprandi et al., 2014; Etchegoyhen et al., 2014; Tardel, 2020; Bojar et al., 2021)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.688, + 0.49, + 0.897 + ], + "angle": 0, + "content": "Subtitles are subject to specific constraints in order to achieve adequate readability, including layout, on-screen duration and text editing. Among these constraints, segmentation addresses the maximum number of characters per line, the number of lines per subtitle, and breaks at natural linguistic frontiers. Segmentation has been shown to be an important readability factor (Perego et al., 2010; Rajendran et al., 2013), with improperly segmented subtitles resulting in increased cognitive effort and reading times for users. Thus, automated subtitle systems need to generate properly segmented subtitles to achieve readability." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.254, + 0.885, + 0.623 + ], + "angle": 0, + "content": "A typical baseline for subtitle segmentation, still used in some production systems, is simple character counting, whereby line breaks are inserted before reaching the maximum allowed number of characters per line. Although simple and fast, this approach does not address the need for linguistically correct segments and, therefore, falls short in terms of readability. Several approaches have been proposed to improve segmentation by automated means. Álvarez et al. (2014) proposed a machine learning method where subtitle breaks are predicted by Support Vector Machine and Linear Regression models trained on professionally-created subtitles. A similar method based on Conditional Random Fields was then shown to improve over these results (Alvarez et al., 2017). Approaches that directly generate subtitle breaks within Neural Machine Translation have also been proposed in recent years (Matusov et al., 2019; Karakanta et al., 2020a). Recently, Papi et al. (2022) developed a multilingual segmenter which generates both text and breaks and may be trained on textual input only, or on joint text and audio data." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.627, + 0.885, + 0.884 + ], + "angle": 0, + "content": "Although quality subtitle segmentation may be achieved with the aforementioned approaches, they require supervised training on segmented subtitle corpora. At present, the largest subtitle corpus is Open Subtitles (Lison et al., 2018), which mainly covers entertainment material, contains subtitles mostly created by non-professionals or automatically translated, and does not include line breaks. The MuST-Cinema corpus (Karakanta et al., 2020b), on the other hand, is a multilingual speech translation corpus that includes subtitles breaks, but is only available for 8 languages at the moment. Considering the vast amount of languages and domains in audiovisual content, the lack of segmented training data hinders the development of robust automated subtitleing systems." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.888, + 0.884, + 0.919 + ], + "angle": 0, + "content": "In this work, we describe a novel unsupervised method based on pretrained masked language mod" + }, + { + "type": "page_footnote", + "bbox": [ + 0.137, + 0.905, + 0.43, + 0.919 + ], + "angle": 0, + "content": "*These authors contributed equally to this work." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.928, + 0.514, + 0.941 + ], + "angle": 0, + "content": "771" + }, + { + "type": "footer", + "bbox": [ + 0.227, + 0.946, + 0.77, + 0.958 + ], + "angle": 0, + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + }, + { + "type": "footer", + "bbox": [ + 0.377, + 0.959, + 0.62, + 0.972 + ], + "angle": 0, + "content": "Volume 2: Short Papers, pages 771-781" + }, + { + "type": "footer", + "bbox": [ + 0.297, + 0.973, + 0.702, + 0.985 + ], + "angle": 0, + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.491, + 0.262 + ], + "angle": 0, + "content": "els (MLM), where line and subtitle breaks are inserted according to the likelihood of a segment acting as an isolated unit, as approximated by the probability of a punctuation mark occurring at a given segmentation point. In our experiments, this novel approach obtained competitive results on most metrics, while also fully preserving the original text and complying with length constraints. Our system may thus be used as a simple yet efficient subtitle segmenter with any pretrained masked language model, for any language covered by the model." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.276, + 0.236, + 0.292 + ], + "angle": 0, + "content": "2 Approach" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.303, + 0.49, + 0.51 + ], + "angle": 0, + "content": "Our approach is based on the standard view that the more appropriate subtitle segments are those that may function as isolated grammatical chunks. We further hypothesise that a relevant approximation for the identification of this type of unit is the likelihood of a punctuation mark being inserted at the end of a candidate segment, as punctuation may mark the closure of a syntactic unit and is often associated with discursive pauses. To test this hypothesis, we compute the likelihood of punctuation marks at different segmentation points, as predicted by a pretrained MLM, and select the insertion point with the highest likelihood.1" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.512, + 0.49, + 0.641 + ], + "angle": 0, + "content": "The segmentation candidates are determined under a sliding-window approach over the entire input text. We first generate the list of all pairs \\(< \\alpha, \\beta>\\) over the unprocessed portion of the text, where \\(\\alpha\\) is a segmentation candidate of length under a specified limit \\(K\\), corresponding to the maximum number of characters per line, and \\(\\beta\\) is the remaining portion of the text to be segmented." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.642, + 0.49, + 0.754 + ], + "angle": 0, + "content": "We then score all segmentation candidates \\(\\alpha\\) with one of the LM scoring variants described below. A segmentation marker, either end-of-line (), or end-of-block indicating the end of a subtitle (), is then appended to the best scoring candidate, and \\(\\beta\\) becomes the input text to be segmented in a recursive iteration of the process." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.755, + 0.49, + 0.883 + ], + "angle": 0, + "content": "Since our method does not rely on any additional information, such as an audio source, to determine the segmentation type, an tag is inserted every even segment or when \\(\\beta\\) is empty; otherwise, an tag is inserted. We thus generate subtitles with a maximum of two lines, following a standard recommendation in subtitling. We also define a minimal number of characters (min) in \\(\\alpha\\) for the" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.085, + 0.882, + 0.116 + ], + "angle": 0, + "content": "segmentation process to apply, and do not segment lines that are under the specified character limit." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.118, + 0.884, + 0.149 + ], + "angle": 0, + "content": "We evaluated three approaches to compute segmentation scores over each candidate pair \\(< \\alpha, \\beta>\\):" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.159, + 0.882, + 0.207 + ], + "angle": 0, + "content": "- Substitution: The last token of \\(\\alpha\\) is masked and the score is the highest MLM probability among punctuation marks on this mask." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.218, + 0.882, + 0.265 + ], + "angle": 0, + "content": "- Insertion: A mask is appended to \\(\\alpha\\) and the score is the highest MLM probability among punctuation marks on this mask." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.276, + 0.882, + 0.34 + ], + "angle": 0, + "content": "- LM-Score: The score is the average of the perplexity of \\(\\alpha\\) and \\(\\beta\\), as derived from the MLM probabilities for each token in the corresponding sequence." + }, + { + "type": "list", + "bbox": [ + 0.532, + 0.159, + 0.882, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.35, + 0.884, + 0.478 + ], + "angle": 0, + "content": "The first two methods are variants of our core approach. The third method, while also based on the same pretrained MLM, relies instead on the pseudoperplexity of the sequences according to the MLM, computed following Salazar et al. (2020). We included this latter variant to measure the potential of using LM scoring directly, without resorting to the likelihood of punctuation marks." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.49, + 0.719, + 0.507 + ], + "angle": 0, + "content": "3 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.515, + 0.884, + 0.676 + ], + "angle": 0, + "content": "Corpora. For all experiments, we used the MustST-Cinema corpus (Karakanta et al., 2020b), which is derived from TED talks and contains both line and subtitle break markers. In addition to being publicly available, it also allows for a direct comparison with the supervised models of Papi et al. (2022). We report results of our approach on the 6 MuST-Cinema datasets for which comparative results were available, directly predicting segmentation on the test sets without any training." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.685, + 0.884, + 0.861 + ], + "angle": 0, + "content": "Methods. For our approach, we tested the three variants described in Section 2. We used BERT (Devlin et al., 2019) as our MLM for all languages. Additionally, we included a variant called overt clueing \\((OC)\\), where an overt punctuation mark at the end of a candidate segment increments the mask score by 1. We then compared the results of the best LM-based variant with those obtained by alternative approaches. In all cases, our results were computed with \\(min = 15\\), as this value obtained the best results overall over the development" + }, + { + "type": "page_footnote", + "bbox": [ + 0.509, + 0.868, + 0.883, + 0.894 + ], + "angle": 0, + "content": "\\(^{2}\\)Our results on all remaining languages of the MuST-Cinema datasets are presented in Appendix B." + }, + { + "type": "page_footnote", + "bbox": [ + 0.509, + 0.894, + 0.883, + 0.919 + ], + "angle": 0, + "content": "\\( {}^{3} \\) Specifically bert-base-uncased as available on HugginFace (https://huggingface.co/),accessed on November 2022." + }, + { + "type": "list", + "bbox": [ + 0.509, + 0.868, + 0.883, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.892, + 0.49, + 0.919 + ], + "angle": 0, + "content": "1 Throughout our experiments, we used the following punctuation marks: ' ', ' ', ' ?', ' !', ' ' and ' '." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.928, + 0.517, + 0.941 + ], + "angle": 0, + "content": "772" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.142, + 0.082, + 0.858, + 0.201 + ], + "angle": 0, + "content": "
EnglishSpanishGerman
MethodSigmaEOLEOBSigmaEOLEOBSigmaEOLEOB
Substitution71.65+19.86-10.9669.34+12.36-5.7469.31+19.05-7.05
Insertion76.77+19.18-9.9173.47+12.98-4.9170.85+18.53-7.96
LM-Score69.97+21.40-8.6667.70+13.29-5.3764.07+16.45-6.51
" + }, + { + "type": "table_caption", + "bbox": [ + 0.204, + 0.21, + 0.794, + 0.226 + ], + "angle": 0, + "content": "Table 1: Sigma and break coverage test set results for LM-based segmentation variants" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.25, + 0.489, + 0.282 + ], + "angle": 0, + "content": "sets, although the differences were minor with the other values we tested (1, 10 and 20).4" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.285, + 0.49, + 0.542 + ], + "angle": 0, + "content": "We used the simple character counting approach (hereafter, CountChars) as baseline, and, as representative supervised methods on the selected datasets, the models described by (Papi et al., 2022). Their core supervised approach is based on a Transformer (Vaswani et al., 2017) architecture with 3 encoder layers and 3 decoder layers, trained on textual MuST-Cinema input only (MC.Text), or on complementary audio data as well via an additional speech encoder with 12 encoder layers (MC.Multi). They trained each variant on either monolingual data alone (mono), or in a multilingual setting (multi). Finally, they also report results for a variant (OS.Text) trained on the Open Subtitles corpus (Lison et al., 2018) for their zero-shot experiments." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.558, + 0.49, + 0.864 + ], + "angle": 0, + "content": "Evaluation. We use the subtitle-oriented metric Sigma (Karakanta et al., 2022), which computes the ratio of achieved BLEU (Papineni et al., 2002) over an approximated upper-bound BLEU score, on text that includes line and subtitle breaks. Sigma is meant to support the evaluation of imperfect texts, i.e. text that differs from the reference when breaks are omitted. Although our approach does not produce imperfect text, achieving perfect BLEU scores when breaks are ignored, we used this metric for comparison purposes. We also report break coverage results (Papi et al., 2022), defined as the ratio of predicted breaks over reference breaks, which we computed separately for the EOL and EOB breaks. Finally, we include length conformity results (CPL), measured as the percentage of subtitle lines whose length is under the maximum number of characters defined by the subtitle guidelines (42 in the TED guidelines\\(^5\\))." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.25, + 0.837, + 0.267 + ], + "angle": 0, + "content": "4 LM-based Segmentation Variants" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.275, + 0.886, + 0.518 + ], + "angle": 0, + "content": "We first compared the three methods described in Section 2 on the English, Spanish and German datasets, with the results described in Table 1. In terms of Sigma, the Insertion method obtained the best results in all cases. It also obtained the best scores in terms of coverage for the EOL marker, except in Spanish, although all three variants tend to overgenerate end-of-line markers to similar extents. The LM-Score variant obtained the worst results in terms of Sigma, but outperformed the alternatives in terms of EOB coverage, a metric on which the three variants performed markedly better than on EOL coverage. Considering the overall results, we selected the Insertion variant as the most balanced one for all remaining experiments reported below." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.529, + 0.727, + 0.546 + ], + "angle": 0, + "content": "5 Comparative Results" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.555, + 0.886, + 0.861 + ], + "angle": 0, + "content": "In Table 2, we present the results obtained by the selected approaches on the languages for which results were available with supervised models trained on in-domain data. Overall, our approach outperformed the CountChars baseline across the board, and was in turn outperformed by the supervised variants in terms of Sigma scores. Although it is clear from these results that training segmentation models on in-domain data, with or without audio data, provides clear advantages in terms of subtitle segmentation, it is worth noting that Sigma does not, by design, reflect the actual BLEU score without breaks, i.e. the generation of imperfect text, which is a by-product of the above supervised approaches and non-existent in ours. In terms of CPL, all supervised models generate subtitle lines that overflow the limit, to a significant degree, whereas the selected unsupervised models trivially respect the length constraint." + }, + { + "type": "page_footnote", + "bbox": [ + 0.508, + 0.87, + 0.885, + 0.919 + ], + "angle": 0, + "content": "The results indicated in Table 3 on unseen data seem to indicate that their MC.Multi model can reach BLEU scores close to 100, thereby limiting the negative impact of imperfect text generation in these cases." + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.879, + 0.486, + 0.905 + ], + "angle": 0, + "content": "4See Appendix C for results with different values of the min parameter." + }, + { + "type": "page_footnote", + "bbox": [ + 0.134, + 0.905, + 0.484, + 0.919 + ], + "angle": 0, + "content": "5https://www.ted.com/participate/translate/subtitling-tips" + }, + { + "type": "list", + "bbox": [ + 0.114, + 0.879, + 0.486, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.928, + 0.517, + 0.941 + ], + "angle": 0, + "content": "773" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.133, + 0.082, + 0.867, + 0.292 + ], + "angle": 0, + "content": "
EnglishFrenchGermanItalian
MethodTrainingSigmaCPLSigmaCPLSigmaCPLSigmaCPL
CountCharsN/A63.71100%62.87100%62.34100%61.49100%
MC.Textmono84.8796.6%83.6896.7%83.6290.9%82.2290.0%
multi85.9888.5%84.5694.3%84.0290.9%83.0491.2%
MC.Multimono85.7694.8%84.2593.9%84.2291.4%82.6289.9%
multi87.4495.0%86.4994.1%86.4089.9%85.3390.0%
MLMN/A76.77100%73.78100%70.85100%71.38100%
MLM+OCN/A77.89100%76.07100%75.63100%74.20100%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.115, + 0.301, + 0.882, + 0.316 + ], + "angle": 0, + "content": "Table 2: Comparative results between unsupervised methods and supervised approaches trained on in-domain data" + }, + { + "type": "table", + "bbox": [ + 0.131, + 0.338, + 0.472, + 0.684 + ], + "angle": 0, + "content": "
Dutch
MethodBLEUSigmaCPLEOLEOB
CountChars10063.2100%-21.2-7.1
OS.Text89.564.471.2%-31.4-51.3
MC.Text61.374.477.8%-23.4-9.9
MC.Multi99.980.391.4%-27.20.4
MLM10068.7100%+20.4-10.0
MLM+OC10073.9100%+21.2-10.0
Spanish
MethodBLEUSigmaCPLEOLEOB
CountChars10063.2100%-24.6-4.4
OS.Text92.664.171.2%-32.3-45.4
MC.Text69.675.870.1%-47.6-19.3
MC.Multi99.678.791.8%-22.44.7
MLM10073.5100%+13.0-4.9
MLM+OC10075.6100%+13.4-4.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.693, + 0.486, + 0.723 + ], + "angle": 0, + "content": "Table 3: Comparative results between unsupervised methods and zero-short supervised approaches" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.759, + 0.49, + 0.919 + ], + "angle": 0, + "content": "In Table 3, we show the comparative results between the selected unsupervised methods and the supervised variants, in languages where zero-shot results were available for the latter approaches. In this scenario, in terms of Sigma our approach obtained results on a par with the supervised MC.Text models trained on in-domain MuST-Cinema data, outperformed the OS.Text models trained on Open Subtitles data, and was surpassed by the MC.Multi model, which exploits additional audio information," + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.341, + 0.885, + 0.583 + ], + "angle": 0, + "content": "by 3.1 and 6.4 points. In terms of break coverage, in most cases our unsupervised method outperformed the supervised variants, to a significant degree compared to the text-based OS.Text and MC.Text models. Regarding BLEU scores without breaks, only the MC.Multi model reaches a score close to the perfect one achieved by the unsupervised models, whereas the MC.Text model is outperformed by 38.7 and 31.4 points in Dutch and Spanish, respectively. In all cases, the CPL scores indicate that none of the supervised approaches fully meet the length constraint, leading to overflowing lines in \\(8.2\\%\\) of the cases at best and \\(29.9\\%\\) at worst. In this scenario as well, the unsupervised approaches fully meet the length constraint, by design." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.59, + 0.885, + 0.751 + ], + "angle": 0, + "content": "Overall, overt clueing improved over our core method by an average of 3.12 Sigma points, indicating that some likely punctuation configurations were not properly captured by our MLM approximation. In general, our approach tends to overgenerate EOL markers, whereas the opposite is true for the selected supervised models. Determining which of these tendencies leads to better subtitle readability would require a specific human evaluation which we leave for future research." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.759, + 0.887, + 0.919 + ], + "angle": 0, + "content": "Although the zero-shot Sigma results obtained by the supervised MC.Multi method show the potential of this approach to provide pretrained models applicable to other languages, two important aspects are worth considering. First, the available zero-shot results were obtained on datasets in the same domain as the data seen to train the supervised models. A more complete assessment of the capabilities of these models in zero-shot settings, which would be the most frequent scenario consid-" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.516, + 0.941 + ], + "angle": 0, + "content": "774" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.49, + 0.262 + ], + "angle": 0, + "content": "ering the lack of training data across domains and languages, would require specific evaluations in other domains. Secondly, although segmentation is a key aspect for subtitle readability, length conformity is an equally important constraint, if not more so considering that subtitles with lines over the CPL limit are considered invalid in subtitleing. Our proposed unsupervised method can thus be seen as a pragmatic approach which guarantees valid subtitles while also providing quality segmentation across the board.\\(^7\\)" + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.277, + 0.254, + 0.293 + ], + "angle": 0, + "content": "6 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.305, + 0.49, + 0.385 + ], + "angle": 0, + "content": "We described an unsupervised approach to subtitle segmentation, based on pretrained masked language models, where line or subtitle breaks are inserted according to the likelihood of punctuation occurring at candidate segmentation points." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.387, + 0.49, + 0.499 + ], + "angle": 0, + "content": "Although supervised models, trained on indomain data with audio support, were shown to perform better that this simple textual approach in terms of the Sigma metric, they tend to generate imperfect text to varying degrees, while also failing to fully meet length constraints that are essential for subtitling." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.501, + 0.49, + 0.645 + ], + "angle": 0, + "content": "In contrast, our LM-based textual approach outperformed supervised models in most cases in terms of break generation coverage, while also fully preserving the original text, complying with length constraints, and obtaining competitive results in terms of Sigma. This simple approach may thus provide a highly portable complementary solution for subtitle segmentation across languages and domains." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.66, + 0.251, + 0.675 + ], + "angle": 0, + "content": "7 Limitations" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.688, + 0.49, + 0.881 + ], + "angle": 0, + "content": "The first clear limitation of our approach is its text-based nature. This prevents important audio information, typically silences in speech patterns, from being exploited to generate subtitle breaks. A more complete system could be devised though, for instance by associating our text-based approach with the information provided by a forced alignment toolkit, whenever audio information is available. A simple method along these lines could be the following: 1. Apply our MLM-based segmentation but only generating a unique segmentation tag SEG; 2. Insert EOB markers wherever the" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.213 + ], + "angle": 0, + "content": "silence between two aligned words is above a specified threshold; 3. Traverse the text sequentially and replace SEG with EOL if there exists a previous marker of type EOB, otherwise replace with EOB. We left this use of our method in combination with audio information for future research, as audio alignment for subtitles typically involves additional factors such as non-literal transcriptions." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.215, + 0.885, + 0.472 + ], + "angle": 0, + "content": "Additionally, our method is limited in its adaptability to specific segmentation guidelines, which may be company-specific. The main adaptable parameters of our methods are the minimum and maximum parameters of the segmentation window, and the set of predefined punctuation marks over which masking is computed, neither of which could fully model idiosyncratic segmentation guidelines. However, in our experience at least, segmentation in real professional data tends to display varying degrees of consistency with respect to guidelines, and natural linguistic breaks seem to be the dominant factor for subtitle segmentation. A specific evaluation would be needed on data from varied professional datasets to determine the extent to which our method might deviate from specific guidelines." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.474, + 0.884, + 0.681 + ], + "angle": 0, + "content": "Finally, other aspects of subtitling, such as the recommendation in some guidelines for subtitles to appear in a pyramidal view, i.e. with the first line shorter than the second line, have not been taken into consideration in this work. Our aim was to evaluate our core LM-based approach without additional variables that can vary across guidelines and may also have led to results that are more difficult to interpret overall. Our approach could nonetheless be easily augmented with constraints on relative line lengths within subtitles, by incrementing the scores of segmentation candidates that respect this surface-level constraint." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.698, + 0.742, + 0.714 + ], + "angle": 0, + "content": "8 Ethical Considerations" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.727, + 0.885, + 0.919 + ], + "angle": 0, + "content": "Our approach involves the use of large pretrained language models, whose computational performance is typically higher when deployed in more powerful environments with GPUs. Under such usage, electric consumption and associated carbon footprint are likely to increase and users of our method under these conditions should be aware of this type of impact. However, subtitle segmentation is often performed offline, where efficient processing is less of a concern, and lower-cost CPU deployments are an entirely viable option. All our results were obtained with a single large LM de" + }, + { + "type": "page_footnote", + "bbox": [ + 0.113, + 0.892, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Examples of segmented subtitles can be found in Appendix A." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.928, + 0.516, + 0.941 + ], + "angle": 0, + "content": "775" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.114, + 0.085, + 0.486, + 0.117 + ], + "angle": 0, + "content": "ployed on CPU, with the aim of reducing energy consumption at inference time." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.119, + 0.49, + 0.327 + ], + "angle": 0, + "content": "Additionally, our method requires no training for the task at hand and thus removes the cost of model training associated with the supervised methods with which we compare our results. For instance, Papi et al. (2022) indicate that they use four K80 GPUs to train their models, which we took as comparison points, with 1 day of training for their text-only models and 1 week for their multimodal segmenters. Therefore, given the large number of potential language pairs and domains in need of segmented subtitle content, our approach can provide competitive results with a comparatively lesser impact on energy resource consumption." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.343, + 0.287, + 0.358 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.371, + 0.488, + 0.466 + ], + "angle": 0, + "content": "We thank the anonymous reviewers for their helpful comments. This work was partially supported by the Department of Economic Development and Competitiveness of the Basque Government (Spri Group) through funding for the StreAmS project (ZL-2021/00700)." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.497, + 0.214, + 0.512 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.522, + 0.49, + 0.626 + ], + "angle": 0, + "content": "Carlo Aliprandi, Cristina Scudellari, Isabella Gallucci, Nicola Piccinini, Matteo Raffaelli, Arantza del Pozo, Aitor Alvarez, Haritz Arzelus, Renato Cassaca, Tiago Luis, et al. 2014. Automatic live subtitling: state of the art, expectations and current trends. In Proceedings of NAB Broadcast Engineering Conference: Papers on Advanced Media Technologies, Las Vegas, volume 13." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.641, + 0.49, + 0.708 + ], + "angle": 0, + "content": "Aitor Álvarez, Haritz Arzelus, and Thierry Etchegoyhen. 2014. Towards customized automatic segmentation of subtitles. In Advances in Speech and Language Technologies for Iberian Languages, pages 229-238. Springer." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.721, + 0.49, + 0.787 + ], + "angle": 0, + "content": "Aitor Alvarez, Carlos-D Martínez-Hinarejos, Haritz Arzelus, Marina Balenciaga, and Arantza del Pozo. 2017. Improving the automatic segmentation of subtitles through conditional random field. Speech Communication, 88:83-95." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.801, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Ondrej Bojar, Dominik Machacek, Sangeet Sagar, Otakar Smrz, Jonas Kratochvil, Peter Polak, Ebrahim Ansari, Mohammad Mahmoudi, Rishu Kumar, Dario Franceschini, Chiara Canton, Ivan Simonini, Thai Son Nguyen, Felix Schneider, Sebastian Stüker, Alex Waibel, Barry Haddow, Rico Sennrich, and Philip Williams. 2021. ELITR multilingual live subtitling: Demo and strategy. In Proceedings of the 16th Conference of the European Chapter of the Association" + }, + { + "type": "list", + "bbox": [ + 0.115, + 0.522, + 0.49, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.527, + 0.086, + 0.885, + 0.126 + ], + "angle": 0, + "content": "for Computational Linguistics: System Demonstrations, pages 271-277, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.135, + 0.885, + 0.253 + ], + "angle": 0, + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.261, + 0.885, + 0.38 + ], + "angle": 0, + "content": "Thierry Etchegoyhen, Lindsay Bywood, Mark Fishel, Panayota Georgakopoulou, Jie Jiang, Gerard van Loenhout, Arantza del Pozo, Mirjam Sepesy Maučec, Anja Turner, and Martin Volk. 2014. Machine translation for subtitling: A large-scale evaluation. In Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC'14), pages 46-53, Reykjavik, Iceland. European Language Resources Association (ELRA)." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.388, + 0.884, + 0.468 + ], + "angle": 0, + "content": "Alina Karakanta, François Buet, Mauro Cettolo, and François Yvon. 2022. Evaluating subtitle segmentation for end-to-end generation systems. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 3069-3078, Marseille, France. European Language Resources Association." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.476, + 0.884, + 0.555 + ], + "angle": 0, + "content": "Alina Karakanta, Matteo Negri, and Marco Turchi. 2020a. Is 42 the answer to everything in subtitling-oriented speech translation? In Proceedings of the 17th International Conference on Spoken Language Translation, pages 209-219, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.564, + 0.885, + 0.641 + ], + "angle": 0, + "content": "Alina Karakanta, Matteo Negri, and Marco Turchi. 2020b. MuST-cinema: a speech-to-subtitles corpus. In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 3727-3734, Marseille, France. European Language Resources Association." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.651, + 0.884, + 0.743 + ], + "angle": 0, + "content": "Pierre Lison, Jörg Tiedemann, and Milen Kouylekov. 2018. OpenSubtitles2018: Statistical rescoring of sentence alignments in large, noisy parallel corpora. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources Association (ELRA)." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.752, + 0.884, + 0.831 + ], + "angle": 0, + "content": "Evgeny Matusov, Patrick Wilken, and Yota Georgakopoulou. 2019. Customizing neural machine translation for subtitling. In Proceedings of the Fourth Conference on Machine Translation (Volume 1: Research Papers), pages 82-93, Florence, Italy. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.84, + 0.884, + 0.919 + ], + "angle": 0, + "content": "Sara Papi, Alina Karakanta, Matteo Negri, and Marco Turchi. 2022. Dodging the data bottleneck: Automatic subtitling with automatically segmented ST corpora. In Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint" + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.885, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.517, + 0.941 + ], + "angle": 0, + "content": "776" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.133, + 0.086, + 0.49, + 0.127 + ], + "angle": 0, + "content": "Conference on Natural Language Processing (Volume 2: Short Papers), pages 480-487, Online only. Association for Computational Linguistics." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.133, + 0.49, + 0.228 + ], + "angle": 0, + "content": "Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311-318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.234, + 0.49, + 0.288 + ], + "angle": 0, + "content": "Elisa Perego, Fabio Del Missier, Marco Porta, and Mauro Mosconi. 2010. The cognitive effectiveness of subtitle processing. *Media psychology*, 13(3):243-272." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.295, + 0.49, + 0.361 + ], + "angle": 0, + "content": "Dhevi J Rajendran, Andrew T Duchowski, Pilar Orero, Juan Martínez, and Pablo Romero-Fresco. 2013. Effects of text chunking on subtitling: A quantitative and qualitative examination. Perspectives, 21(1):5-21." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.368, + 0.49, + 0.448 + ], + "angle": 0, + "content": "Julian Salazar, Davis Liang, Toan Q. Nguyen, and Katrin Kirchhoff. 2020. Masked language model scoring. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2699-2712, Online. Association for Computational Linguistics." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.455, + 0.49, + 0.509 + ], + "angle": 0, + "content": "Anke Tardel. 2020. Effort in semi-automatized subtitling processes: speech recognition and experience during transcription. Journal of Audiovisual Translation, 3(2):79-102." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.516, + 0.488, + 0.582 + ], + "angle": 0, + "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.589, + 0.49, + 0.695 + ], + "angle": 0, + "content": "Martin Volk, Rico Sennrich, Christian Hardmeier, and Frida Tidström. 2010. Machine translation of TV subtitles for large scale production. In Proceedings of the Second Joint EM+/CNGL Workshop: Bringing MT to the User: Research on Integrating MT in the Translation Industry, pages 53-62, Denver, Colorado, USA. Association for Machine Translation in the Americas." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.702, + 0.363, + 0.719 + ], + "angle": 0, + "content": "A Segmentation Examples" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.727, + 0.49, + 0.807 + ], + "angle": 0, + "content": "Table 4 provides examples of subtitles in the MuST-Cinema test sets segmented with either the character counting baseline or our LM-based approach, in its insertion variant without resorting to overt punctuation clueing." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.808, + 0.49, + 0.919 + ], + "angle": 0, + "content": "In these examples, the MLM approach generates end-of-line and end-of-subtitle breaks that are overall in line with natural linguistic breaks, contrary to the character counting baseline. As such, on either short, medium or longer input, the readability of the generated subtitles is significantly enhanced with our approach." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.084, + 0.699, + 0.1 + ], + "angle": 0, + "content": "B Extended Results" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.11, + 0.885, + 0.334 + ], + "angle": 0, + "content": "The results presented in Section 5 were limited to the subset of languages and metrics for which published comparative results were available on the MuST-Cinema datasets. In Table 5, we present the complete list of results obtained with our method, for all languages and metrics. The selected variant of our method is the insertion masking approach, which was selected for the main results in our paper, with a segmentation window starting at 15 characters and ending at 42. We do not include BLEU scores computed over text that includes segmentation breaks, as the results are identical to those obtained with the Sigma metric for our approach, which does not generate imperfect text." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.336, + 0.885, + 0.528 + ], + "angle": 0, + "content": "Across languages, the results are relatively uniform, with the best Sigma scores obtained in English and the lowest in Dutch, for a difference of 4.1 points between the two languages. In terms of break coverage, the best results were obtained for Spanish and the worst for Romanian, although results were also relatively uniform across languages. In all cases, overt clueing, where overt punctuation marks raised the LM score by 1, improved Sigma scores, although it had less of an impact on break coverage results, where both variants performed similarly overall." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.54, + 0.882, + 0.557 + ], + "angle": 0, + "content": "C Results With Different min Parameters" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.566, + 0.884, + 0.724 + ], + "angle": 0, + "content": "As noted in Section 3, considering preliminary results over the development set we selected a default value of 15 for the min parameter, which indicates the number of characters after which the segmentation process applies. In Table 6, we present comparative results on the test sets with different min values. In terms of Sigma, values of 15 and 20 led to rather similar results; values of 1 and 10 resulted in slightly lower results, with the lowest results achieved with the former." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.727, + 0.885, + 0.84 + ], + "angle": 0, + "content": "In terms of \\( \\langle \\mathrm{eol} \\rangle \\) and \\( \\langle \\mathrm{eob} \\rangle \\) coverage, the former increases with larger min values, which is expected given the more restricted space to insert these end-of-line markers as the value increases; for \\( \\langle \\mathrm{eob} \\rangle \\), the restricted insertion space results in increased under-generation, which in turn results in better scores for lower values of the min parameter." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.928, + 0.517, + 0.941 + ], + "angle": 0, + "content": "777" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.116, + 0.135, + 0.885, + 0.361 + ], + "angle": 0, + "content": "
CountCharsMLM
They're things you access through your <eol> computer. <eob>They're things you access <eol> through your computer. <eob>
Every row of data is a life whose story <eol> deserves to be told with dignity. <eob>Every row of data is a life <eol> whose story deserves to be told <eob> with dignity. <eob>
During the winter, struggling to get <eol> warm, my neighbors would have no choice <eob> but to bypass the meter after their heat <eol> was shut off, just to keep their family <eob> comfortable for one more day. <eob>During the winter, struggling to get warm, <eol> my neighbors would have no choice <eob> but to bypass the meter <eol> after their heat was shut off, <eob> just to keep their family comfortable <eol> for one more day. <eob>
" + }, + { + "type": "table_caption", + "bbox": [ + 0.17, + 0.369, + 0.825, + 0.384 + ], + "angle": 0, + "content": "Table 4: Examples of subtitles segmented via character counting and MLM-based mask insertion" + }, + { + "type": "table", + "bbox": [ + 0.237, + 0.498, + 0.763, + 0.837 + ], + "angle": 0, + "content": "
LanguageMethodBLEUSigmaEOLEOBCPL
DEMLM10070.8518.53-7.96100%
MLM+OC10075.6319.81-7.78100%
ENMLM10076.7719.18-9.91100%
MLM+OC10077.8919.86-9.73100%
ESMLM10073.4712.98-4.91100%
MLM+OC10075.5913.45-4.63100%
FRMLM10073.7816.51-6.58100%
MLM+OC10076.0717.47-6.12100%
ITMLM10071.3818.49-9.55100%
MLM+OC10074.2020.34-8.57100%
NLMLM10068.7120.37-9.96100%
MLM+OC10073.8821.22-9.96100%
PTMLM10071.5920.03-10.81100%
MLM+OC10075.5019.87-10.02100%
ROMLM10069.4523.37-10.44100%
MLM+OC10074.1323.37-10.09100%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.18, + 0.847, + 0.816, + 0.862 + ], + "angle": 0, + "content": "Table 5: Complete results with MLM mask insertion on the MuST-Cinema test sets (min=15)" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.516, + 0.941 + ], + "angle": 0, + "content": "778" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.304, + 0.188, + 0.695, + 0.786 + ], + "angle": 0, + "content": "
LanguageminBLEUSigmaEOLEOB
DE110072.3128.75-0.18
1010073.9622.68-4.43
1510075.6319.81-7.78
2010075.2814.54-11.21
EN110074.3037.33-0.98
1010077.1424.49-7.77
1510077.8919.86-9.73
2010077.1615.24-12.68
ES110073.0020.870.28
1010074.3218.24-2.04
1510075.5913.45-4.63
2010075.838.66-7.87
FR110073.8924.68-0.73
1010075.2620.83-3.93
1510076.0717.47-6.12
2010076.7512.5-10.05
IT110072.0129.75-3.66
1010073.7524.71-6.61
1510074.2020.34-8.57
2010073.6614.62-11.61
NL110072.1626.83-5.47
1010073.5623.26-8.47
1510073.8821.22-9.96
2010074.4016.81-12.43
PT110072.8726.38-6.24
1010074.5322.15-8.08
1510075.5019.87-10.02
2010074.9814.17-13.36
RO110072.0532.3-4.51
1010073.7626.98-7.52
1510074.1323.37-10.09
2010074.8917.53-12.83
" + }, + { + "type": "table_caption", + "bbox": [ + 0.183, + 0.794, + 0.816, + 0.81 + ], + "angle": 0, + "content": "Table 6: Test set results with the MLM+OC method and different values of the min parameter" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.516, + 0.941 + ], + "angle": 0, + "content": "779" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.133, + 0.085, + 0.435, + 0.1 + ], + "angle": 0, + "content": "ACL 2023 Responsible NLP Checklist" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.108, + 0.323, + 0.123 + ], + "angle": 0, + "content": "A For every submission:" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.127, + 0.533, + 0.158 + ], + "angle": 0, + "content": "A1. Did you describe the limitations of your work? 7" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.171, + 0.553, + 0.201 + ], + "angle": 0, + "content": "A2. Did you discuss any potential risks of your work? 8" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.212, + 0.697, + 0.244 + ], + "angle": 0, + "content": "A3. Do the abstract and introduction summarize the paper's main claims?" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.256, + 0.669, + 0.289 + ], + "angle": 0, + "content": "A4. Have you used AI writing assistants when working on this paper? Left blank." + }, + { + "type": "list", + "bbox": [ + 0.13, + 0.127, + 0.697, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.301, + 0.491, + 0.317 + ], + "angle": 0, + "content": "B Did you use or create scientific artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.132, + 0.322, + 0.334, + 0.337 + ], + "angle": 0, + "content": "Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.348, + 0.531, + 0.38 + ], + "angle": 0, + "content": "B1. Did you cite the creators of artifacts you used? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.391, + 0.779, + 0.423 + ], + "angle": 0, + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.434, + 0.881, + 0.514 + ], + "angle": 0, + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.525, + 0.881, + 0.589 + ], + "angle": 0, + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.6, + 0.881, + 0.648 + ], + "angle": 0, + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.659, + 0.881, + 0.755 + ], + "angle": 0, + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Not applicable. Left blank." + }, + { + "type": "list", + "bbox": [ + 0.129, + 0.348, + 0.881, + 0.755 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.765, + 0.494, + 0.782 + ], + "angle": 0, + "content": "C Did you run computational experiments?" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.788, + 0.147, + 0.8 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.814, + 0.881, + 0.845 + ], + "angle": 0, + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.846, + 0.761, + 0.862 + ], + "angle": 0, + "content": "We didn't trained any models for this paper, and inference was performed on CPU." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.869, + 0.878, + 0.894 + ], + "angle": 0, + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.517, + 0.941 + ], + "angle": 0, + "content": "780" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.13, + 0.084, + 0.88, + 0.131 + ], + "angle": 0, + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? 3" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.143, + 0.88, + 0.205 + ], + "angle": 0, + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? 5" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.218, + 0.88, + 0.28 + ], + "angle": 0, + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?" + }, + { + "type": "list", + "bbox": [ + 0.13, + 0.084, + 0.88, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.294, + 0.878, + 0.31 + ], + "angle": 0, + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.315, + 0.215, + 0.33 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.341, + 0.88, + 0.386 + ], + "angle": 0, + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.4, + 0.88, + 0.461 + ], + "angle": 0, + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.475, + 0.88, + 0.537 + ], + "angle": 0, + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.55, + 0.873, + 0.581 + ], + "angle": 0, + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.593, + 0.88, + 0.64 + ], + "angle": 0, + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Not applicable. Left blank." + }, + { + "type": "list", + "bbox": [ + 0.129, + 0.341, + 0.88, + 0.64 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.514, + 0.941 + ], + "angle": 0, + "content": "781" + } + ] +] \ No newline at end of file diff --git a/2023/Unsupervised Subtitle Segmentation with Masked Language Models/d5fa5db2-10b5-46c6-ab9c-a92346dcaabb_origin.pdf b/2023/Unsupervised Subtitle Segmentation with Masked Language Models/d5fa5db2-10b5-46c6-ab9c-a92346dcaabb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5dff8614a63b42bf9863c9d10e9cbe3a3c779806 --- /dev/null +++ b/2023/Unsupervised Subtitle Segmentation with Masked Language Models/d5fa5db2-10b5-46c6-ab9c-a92346dcaabb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:432af0044627e966910e9599cd35c9226102b3b507964c1bcf8c4a35e1d9bc51 +size 218782 diff --git a/2023/Unsupervised Subtitle Segmentation with Masked Language Models/full.md b/2023/Unsupervised Subtitle Segmentation with Masked Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0dc7e94e3e85fbd7861280fbf4f0fefeb857c766 --- /dev/null +++ b/2023/Unsupervised Subtitle Segmentation with Masked Language Models/full.md @@ -0,0 +1,224 @@ +# Unsupervised Subtitle Segmentation with Masked Language Models + +David Ponce\*1,2 and Thierry Etchegoyhen\*1 and Victor Ruiz + +1 Vicomtech Foundation, Basque Research and Technology Alliance (BRTA) + +$^{2}$ University of the Basque Country UPV/EHU + +{adponce,tetchegoyhen,vruiz}@vicomtech.org + +# Abstract + +We describe a novel unsupervised approach to subtitle segmentation, based on pretrained masked language models, where line endings and subtitle breaks are predicted according to the likelihood of punctuation to occur at candidate segmentation points. Our approach obtained competitive results in terms of segmentation accuracy across metrics, while also fully preserving the original text and complying with length constraints. Although supervised models trained on in-domain data and with access to source audio information can provide better segmentation accuracy, our approach is highly portable across languages and domains and may constitute a robust off-the-shelf solution for subtitle segmentation. + +# 1 Introduction + +Subtitling is one of the principal means of providing accessible audiovisual content. With the ever increasing production of audiovisual content in multiple domains and languages, in the current digital era, subtitle provision can benefit from automation support, via Automatic Speech Recognition and/or Machine Translation (Volk et al., 2010; Aliprandi et al., 2014; Etchegoyhen et al., 2014; Tardel, 2020; Bojar et al., 2021). + +Subtitles are subject to specific constraints in order to achieve adequate readability, including layout, on-screen duration and text editing. Among these constraints, segmentation addresses the maximum number of characters per line, the number of lines per subtitle, and breaks at natural linguistic frontiers. Segmentation has been shown to be an important readability factor (Perego et al., 2010; Rajendran et al., 2013), with improperly segmented subtitles resulting in increased cognitive effort and reading times for users. Thus, automated subtitle systems need to generate properly segmented subtitles to achieve readability. + +A typical baseline for subtitle segmentation, still used in some production systems, is simple character counting, whereby line breaks are inserted before reaching the maximum allowed number of characters per line. Although simple and fast, this approach does not address the need for linguistically correct segments and, therefore, falls short in terms of readability. Several approaches have been proposed to improve segmentation by automated means. Álvarez et al. (2014) proposed a machine learning method where subtitle breaks are predicted by Support Vector Machine and Linear Regression models trained on professionally-created subtitles. A similar method based on Conditional Random Fields was then shown to improve over these results (Alvarez et al., 2017). Approaches that directly generate subtitle breaks within Neural Machine Translation have also been proposed in recent years (Matusov et al., 2019; Karakanta et al., 2020a). Recently, Papi et al. (2022) developed a multilingual segmenter which generates both text and breaks and may be trained on textual input only, or on joint text and audio data. + +Although quality subtitle segmentation may be achieved with the aforementioned approaches, they require supervised training on segmented subtitle corpora. At present, the largest subtitle corpus is Open Subtitles (Lison et al., 2018), which mainly covers entertainment material, contains subtitles mostly created by non-professionals or automatically translated, and does not include line breaks. The MuST-Cinema corpus (Karakanta et al., 2020b), on the other hand, is a multilingual speech translation corpus that includes subtitles breaks, but is only available for 8 languages at the moment. Considering the vast amount of languages and domains in audiovisual content, the lack of segmented training data hinders the development of robust automated subtitleing systems. + +In this work, we describe a novel unsupervised method based on pretrained masked language mod + +els (MLM), where line and subtitle breaks are inserted according to the likelihood of a segment acting as an isolated unit, as approximated by the probability of a punctuation mark occurring at a given segmentation point. In our experiments, this novel approach obtained competitive results on most metrics, while also fully preserving the original text and complying with length constraints. Our system may thus be used as a simple yet efficient subtitle segmenter with any pretrained masked language model, for any language covered by the model. + +# 2 Approach + +Our approach is based on the standard view that the more appropriate subtitle segments are those that may function as isolated grammatical chunks. We further hypothesise that a relevant approximation for the identification of this type of unit is the likelihood of a punctuation mark being inserted at the end of a candidate segment, as punctuation may mark the closure of a syntactic unit and is often associated with discursive pauses. To test this hypothesis, we compute the likelihood of punctuation marks at different segmentation points, as predicted by a pretrained MLM, and select the insertion point with the highest likelihood.1 + +The segmentation candidates are determined under a sliding-window approach over the entire input text. We first generate the list of all pairs $< \alpha, \beta>$ over the unprocessed portion of the text, where $\alpha$ is a segmentation candidate of length under a specified limit $K$ , corresponding to the maximum number of characters per line, and $\beta$ is the remaining portion of the text to be segmented. + +We then score all segmentation candidates $\alpha$ with one of the LM scoring variants described below. A segmentation marker, either end-of-line (), or end-of-block indicating the end of a subtitle (), is then appended to the best scoring candidate, and $\beta$ becomes the input text to be segmented in a recursive iteration of the process. + +Since our method does not rely on any additional information, such as an audio source, to determine the segmentation type, an tag is inserted every even segment or when $\beta$ is empty; otherwise, an tag is inserted. We thus generate subtitles with a maximum of two lines, following a standard recommendation in subtitling. We also define a minimal number of characters (min) in $\alpha$ for the + +segmentation process to apply, and do not segment lines that are under the specified character limit. + +We evaluated three approaches to compute segmentation scores over each candidate pair $< \alpha, \beta>$ : + +- Substitution: The last token of $\alpha$ is masked and the score is the highest MLM probability among punctuation marks on this mask. +- Insertion: A mask is appended to $\alpha$ and the score is the highest MLM probability among punctuation marks on this mask. +- LM-Score: The score is the average of the perplexity of $\alpha$ and $\beta$ , as derived from the MLM probabilities for each token in the corresponding sequence. + +The first two methods are variants of our core approach. The third method, while also based on the same pretrained MLM, relies instead on the pseudoperplexity of the sequences according to the MLM, computed following Salazar et al. (2020). We included this latter variant to measure the potential of using LM scoring directly, without resorting to the likelihood of punctuation marks. + +# 3 Experimental Setup + +Corpora. For all experiments, we used the MustST-Cinema corpus (Karakanta et al., 2020b), which is derived from TED talks and contains both line and subtitle break markers. In addition to being publicly available, it also allows for a direct comparison with the supervised models of Papi et al. (2022). We report results of our approach on the 6 MuST-Cinema datasets for which comparative results were available, directly predicting segmentation on the test sets without any training. + +Methods. For our approach, we tested the three variants described in Section 2. We used BERT (Devlin et al., 2019) as our MLM for all languages. Additionally, we included a variant called overt clueing $(OC)$ , where an overt punctuation mark at the end of a candidate segment increments the mask score by 1. We then compared the results of the best LM-based variant with those obtained by alternative approaches. In all cases, our results were computed with $min = 15$ , as this value obtained the best results overall over the development + +
EnglishSpanishGerman
MethodSigmaEOLEOBSigmaEOLEOBSigmaEOLEOB
Substitution71.65+19.86-10.9669.34+12.36-5.7469.31+19.05-7.05
Insertion76.77+19.18-9.9173.47+12.98-4.9170.85+18.53-7.96
LM-Score69.97+21.40-8.6667.70+13.29-5.3764.07+16.45-6.51
+ +Table 1: Sigma and break coverage test set results for LM-based segmentation variants + +sets, although the differences were minor with the other values we tested (1, 10 and 20).4 + +We used the simple character counting approach (hereafter, CountChars) as baseline, and, as representative supervised methods on the selected datasets, the models described by (Papi et al., 2022). Their core supervised approach is based on a Transformer (Vaswani et al., 2017) architecture with 3 encoder layers and 3 decoder layers, trained on textual MuST-Cinema input only (MC.Text), or on complementary audio data as well via an additional speech encoder with 12 encoder layers (MC.Multi). They trained each variant on either monolingual data alone (mono), or in a multilingual setting (multi). Finally, they also report results for a variant (OS.Text) trained on the Open Subtitles corpus (Lison et al., 2018) for their zero-shot experiments. + +Evaluation. We use the subtitle-oriented metric Sigma (Karakanta et al., 2022), which computes the ratio of achieved BLEU (Papineni et al., 2002) over an approximated upper-bound BLEU score, on text that includes line and subtitle breaks. Sigma is meant to support the evaluation of imperfect texts, i.e. text that differs from the reference when breaks are omitted. Although our approach does not produce imperfect text, achieving perfect BLEU scores when breaks are ignored, we used this metric for comparison purposes. We also report break coverage results (Papi et al., 2022), defined as the ratio of predicted breaks over reference breaks, which we computed separately for the EOL and EOB breaks. Finally, we include length conformity results (CPL), measured as the percentage of subtitle lines whose length is under the maximum number of characters defined by the subtitle guidelines (42 in the TED guidelines $^5$ ). + +# 4 LM-based Segmentation Variants + +We first compared the three methods described in Section 2 on the English, Spanish and German datasets, with the results described in Table 1. In terms of Sigma, the Insertion method obtained the best results in all cases. It also obtained the best scores in terms of coverage for the EOL marker, except in Spanish, although all three variants tend to overgenerate end-of-line markers to similar extents. The LM-Score variant obtained the worst results in terms of Sigma, but outperformed the alternatives in terms of EOB coverage, a metric on which the three variants performed markedly better than on EOL coverage. Considering the overall results, we selected the Insertion variant as the most balanced one for all remaining experiments reported below. + +# 5 Comparative Results + +In Table 2, we present the results obtained by the selected approaches on the languages for which results were available with supervised models trained on in-domain data. Overall, our approach outperformed the CountChars baseline across the board, and was in turn outperformed by the supervised variants in terms of Sigma scores. Although it is clear from these results that training segmentation models on in-domain data, with or without audio data, provides clear advantages in terms of subtitle segmentation, it is worth noting that Sigma does not, by design, reflect the actual BLEU score without breaks, i.e. the generation of imperfect text, which is a by-product of the above supervised approaches and non-existent in ours. In terms of CPL, all supervised models generate subtitle lines that overflow the limit, to a significant degree, whereas the selected unsupervised models trivially respect the length constraint. + +
EnglishFrenchGermanItalian
MethodTrainingSigmaCPLSigmaCPLSigmaCPLSigmaCPL
CountCharsN/A63.71100%62.87100%62.34100%61.49100%
MC.Textmono84.8796.6%83.6896.7%83.6290.9%82.2290.0%
multi85.9888.5%84.5694.3%84.0290.9%83.0491.2%
MC.Multimono85.7694.8%84.2593.9%84.2291.4%82.6289.9%
multi87.4495.0%86.4994.1%86.4089.9%85.3390.0%
MLMN/A76.77100%73.78100%70.85100%71.38100%
MLM+OCN/A77.89100%76.07100%75.63100%74.20100%
+ +Table 2: Comparative results between unsupervised methods and supervised approaches trained on in-domain data + +
Dutch
MethodBLEUSigmaCPLEOLEOB
CountChars10063.2100%-21.2-7.1
OS.Text89.564.471.2%-31.4-51.3
MC.Text61.374.477.8%-23.4-9.9
MC.Multi99.980.391.4%-27.20.4
MLM10068.7100%+20.4-10.0
MLM+OC10073.9100%+21.2-10.0
Spanish
MethodBLEUSigmaCPLEOLEOB
CountChars10063.2100%-24.6-4.4
OS.Text92.664.171.2%-32.3-45.4
MC.Text69.675.870.1%-47.6-19.3
MC.Multi99.678.791.8%-22.44.7
MLM10073.5100%+13.0-4.9
MLM+OC10075.6100%+13.4-4.6
+ +Table 3: Comparative results between unsupervised methods and zero-short supervised approaches + +In Table 3, we show the comparative results between the selected unsupervised methods and the supervised variants, in languages where zero-shot results were available for the latter approaches. In this scenario, in terms of Sigma our approach obtained results on a par with the supervised MC.Text models trained on in-domain MuST-Cinema data, outperformed the OS.Text models trained on Open Subtitles data, and was surpassed by the MC.Multi model, which exploits additional audio information, + +by 3.1 and 6.4 points. In terms of break coverage, in most cases our unsupervised method outperformed the supervised variants, to a significant degree compared to the text-based OS.Text and MC.Text models. Regarding BLEU scores without breaks, only the MC.Multi model reaches a score close to the perfect one achieved by the unsupervised models, whereas the MC.Text model is outperformed by 38.7 and 31.4 points in Dutch and Spanish, respectively. In all cases, the CPL scores indicate that none of the supervised approaches fully meet the length constraint, leading to overflowing lines in $8.2\%$ of the cases at best and $29.9\%$ at worst. In this scenario as well, the unsupervised approaches fully meet the length constraint, by design. + +Overall, overt clueing improved over our core method by an average of 3.12 Sigma points, indicating that some likely punctuation configurations were not properly captured by our MLM approximation. In general, our approach tends to overgenerate EOL markers, whereas the opposite is true for the selected supervised models. Determining which of these tendencies leads to better subtitle readability would require a specific human evaluation which we leave for future research. + +Although the zero-shot Sigma results obtained by the supervised MC.Multi method show the potential of this approach to provide pretrained models applicable to other languages, two important aspects are worth considering. First, the available zero-shot results were obtained on datasets in the same domain as the data seen to train the supervised models. A more complete assessment of the capabilities of these models in zero-shot settings, which would be the most frequent scenario consid- + +ering the lack of training data across domains and languages, would require specific evaluations in other domains. Secondly, although segmentation is a key aspect for subtitle readability, length conformity is an equally important constraint, if not more so considering that subtitles with lines over the CPL limit are considered invalid in subtitleing. Our proposed unsupervised method can thus be seen as a pragmatic approach which guarantees valid subtitles while also providing quality segmentation across the board. $^7$ + +# 6 Conclusions + +We described an unsupervised approach to subtitle segmentation, based on pretrained masked language models, where line or subtitle breaks are inserted according to the likelihood of punctuation occurring at candidate segmentation points. + +Although supervised models, trained on indomain data with audio support, were shown to perform better that this simple textual approach in terms of the Sigma metric, they tend to generate imperfect text to varying degrees, while also failing to fully meet length constraints that are essential for subtitling. + +In contrast, our LM-based textual approach outperformed supervised models in most cases in terms of break generation coverage, while also fully preserving the original text, complying with length constraints, and obtaining competitive results in terms of Sigma. This simple approach may thus provide a highly portable complementary solution for subtitle segmentation across languages and domains. + +# 7 Limitations + +The first clear limitation of our approach is its text-based nature. This prevents important audio information, typically silences in speech patterns, from being exploited to generate subtitle breaks. A more complete system could be devised though, for instance by associating our text-based approach with the information provided by a forced alignment toolkit, whenever audio information is available. A simple method along these lines could be the following: 1. Apply our MLM-based segmentation but only generating a unique segmentation tag SEG; 2. Insert EOB markers wherever the + +silence between two aligned words is above a specified threshold; 3. Traverse the text sequentially and replace SEG with EOL if there exists a previous marker of type EOB, otherwise replace with EOB. We left this use of our method in combination with audio information for future research, as audio alignment for subtitles typically involves additional factors such as non-literal transcriptions. + +Additionally, our method is limited in its adaptability to specific segmentation guidelines, which may be company-specific. The main adaptable parameters of our methods are the minimum and maximum parameters of the segmentation window, and the set of predefined punctuation marks over which masking is computed, neither of which could fully model idiosyncratic segmentation guidelines. However, in our experience at least, segmentation in real professional data tends to display varying degrees of consistency with respect to guidelines, and natural linguistic breaks seem to be the dominant factor for subtitle segmentation. A specific evaluation would be needed on data from varied professional datasets to determine the extent to which our method might deviate from specific guidelines. + +Finally, other aspects of subtitling, such as the recommendation in some guidelines for subtitles to appear in a pyramidal view, i.e. with the first line shorter than the second line, have not been taken into consideration in this work. Our aim was to evaluate our core LM-based approach without additional variables that can vary across guidelines and may also have led to results that are more difficult to interpret overall. Our approach could nonetheless be easily augmented with constraints on relative line lengths within subtitles, by incrementing the scores of segmentation candidates that respect this surface-level constraint. + +# 8 Ethical Considerations + +Our approach involves the use of large pretrained language models, whose computational performance is typically higher when deployed in more powerful environments with GPUs. Under such usage, electric consumption and associated carbon footprint are likely to increase and users of our method under these conditions should be aware of this type of impact. However, subtitle segmentation is often performed offline, where efficient processing is less of a concern, and lower-cost CPU deployments are an entirely viable option. All our results were obtained with a single large LM de + +ployed on CPU, with the aim of reducing energy consumption at inference time. + +Additionally, our method requires no training for the task at hand and thus removes the cost of model training associated with the supervised methods with which we compare our results. For instance, Papi et al. (2022) indicate that they use four K80 GPUs to train their models, which we took as comparison points, with 1 day of training for their text-only models and 1 week for their multimodal segmenters. Therefore, given the large number of potential language pairs and domains in need of segmented subtitle content, our approach can provide competitive results with a comparatively lesser impact on energy resource consumption. + +# Acknowledgements + +We thank the anonymous reviewers for their helpful comments. This work was partially supported by the Department of Economic Development and Competitiveness of the Basque Government (Spri Group) through funding for the StreAmS project (ZL-2021/00700). + +# References + +Carlo Aliprandi, Cristina Scudellari, Isabella Gallucci, Nicola Piccinini, Matteo Raffaelli, Arantza del Pozo, Aitor Alvarez, Haritz Arzelus, Renato Cassaca, Tiago Luis, et al. 2014. Automatic live subtitling: state of the art, expectations and current trends. In Proceedings of NAB Broadcast Engineering Conference: Papers on Advanced Media Technologies, Las Vegas, volume 13. +Aitor Álvarez, Haritz Arzelus, and Thierry Etchegoyhen. 2014. Towards customized automatic segmentation of subtitles. In Advances in Speech and Language Technologies for Iberian Languages, pages 229-238. Springer. +Aitor Alvarez, Carlos-D Martínez-Hinarejos, Haritz Arzelus, Marina Balenciaga, and Arantza del Pozo. 2017. Improving the automatic segmentation of subtitles through conditional random field. Speech Communication, 88:83-95. +Ondrej Bojar, Dominik Machacek, Sangeet Sagar, Otakar Smrz, Jonas Kratochvil, Peter Polak, Ebrahim Ansari, Mohammad Mahmoudi, Rishu Kumar, Dario Franceschini, Chiara Canton, Ivan Simonini, Thai Son Nguyen, Felix Schneider, Sebastian Stüker, Alex Waibel, Barry Haddow, Rico Sennrich, and Philip Williams. 2021. ELITR multilingual live subtitling: Demo and strategy. In Proceedings of the 16th Conference of the European Chapter of the Association + +for Computational Linguistics: System Demonstrations, pages 271-277, Online. Association for Computational Linguistics. +Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics. +Thierry Etchegoyhen, Lindsay Bywood, Mark Fishel, Panayota Georgakopoulou, Jie Jiang, Gerard van Loenhout, Arantza del Pozo, Mirjam Sepesy Maučec, Anja Turner, and Martin Volk. 2014. Machine translation for subtitling: A large-scale evaluation. In Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC'14), pages 46-53, Reykjavik, Iceland. European Language Resources Association (ELRA). +Alina Karakanta, François Buet, Mauro Cettolo, and François Yvon. 2022. Evaluating subtitle segmentation for end-to-end generation systems. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 3069-3078, Marseille, France. European Language Resources Association. +Alina Karakanta, Matteo Negri, and Marco Turchi. 2020a. Is 42 the answer to everything in subtitling-oriented speech translation? In Proceedings of the 17th International Conference on Spoken Language Translation, pages 209-219, Online. Association for Computational Linguistics. +Alina Karakanta, Matteo Negri, and Marco Turchi. 2020b. MuST-cinema: a speech-to-subtitles corpus. In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 3727-3734, Marseille, France. European Language Resources Association. +Pierre Lison, Jörg Tiedemann, and Milen Kouylekov. 2018. OpenSubtitles2018: Statistical rescoring of sentence alignments in large, noisy parallel corpora. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources Association (ELRA). +Evgeny Matusov, Patrick Wilken, and Yota Georgakopoulou. 2019. Customizing neural machine translation for subtitling. In Proceedings of the Fourth Conference on Machine Translation (Volume 1: Research Papers), pages 82-93, Florence, Italy. Association for Computational Linguistics. +Sara Papi, Alina Karakanta, Matteo Negri, and Marco Turchi. 2022. Dodging the data bottleneck: Automatic subtitling with automatically segmented ST corpora. In Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint + +Conference on Natural Language Processing (Volume 2: Short Papers), pages 480-487, Online only. Association for Computational Linguistics. + +Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311-318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics. + +Elisa Perego, Fabio Del Missier, Marco Porta, and Mauro Mosconi. 2010. The cognitive effectiveness of subtitle processing. *Media psychology*, 13(3):243-272. + +Dhevi J Rajendran, Andrew T Duchowski, Pilar Orero, Juan Martínez, and Pablo Romero-Fresco. 2013. Effects of text chunking on subtitling: A quantitative and qualitative examination. Perspectives, 21(1):5-21. + +Julian Salazar, Davis Liang, Toan Q. Nguyen, and Katrin Kirchhoff. 2020. Masked language model scoring. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2699-2712, Online. Association for Computational Linguistics. + +Anke Tardel. 2020. Effort in semi-automatized subtitling processes: speech recognition and experience during transcription. Journal of Audiovisual Translation, 3(2):79-102. + +Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30. + +Martin Volk, Rico Sennrich, Christian Hardmeier, and Frida Tidström. 2010. Machine translation of TV subtitles for large scale production. In Proceedings of the Second Joint EM+/CNGL Workshop: Bringing MT to the User: Research on Integrating MT in the Translation Industry, pages 53-62, Denver, Colorado, USA. Association for Machine Translation in the Americas. + +# A Segmentation Examples + +Table 4 provides examples of subtitles in the MuST-Cinema test sets segmented with either the character counting baseline or our LM-based approach, in its insertion variant without resorting to overt punctuation clueing. + +In these examples, the MLM approach generates end-of-line and end-of-subtitle breaks that are overall in line with natural linguistic breaks, contrary to the character counting baseline. As such, on either short, medium or longer input, the readability of the generated subtitles is significantly enhanced with our approach. + +# B Extended Results + +The results presented in Section 5 were limited to the subset of languages and metrics for which published comparative results were available on the MuST-Cinema datasets. In Table 5, we present the complete list of results obtained with our method, for all languages and metrics. The selected variant of our method is the insertion masking approach, which was selected for the main results in our paper, with a segmentation window starting at 15 characters and ending at 42. We do not include BLEU scores computed over text that includes segmentation breaks, as the results are identical to those obtained with the Sigma metric for our approach, which does not generate imperfect text. + +Across languages, the results are relatively uniform, with the best Sigma scores obtained in English and the lowest in Dutch, for a difference of 4.1 points between the two languages. In terms of break coverage, the best results were obtained for Spanish and the worst for Romanian, although results were also relatively uniform across languages. In all cases, overt clueing, where overt punctuation marks raised the LM score by 1, improved Sigma scores, although it had less of an impact on break coverage results, where both variants performed similarly overall. + +# C Results With Different min Parameters + +As noted in Section 3, considering preliminary results over the development set we selected a default value of 15 for the min parameter, which indicates the number of characters after which the segmentation process applies. In Table 6, we present comparative results on the test sets with different min values. In terms of Sigma, values of 15 and 20 led to rather similar results; values of 1 and 10 resulted in slightly lower results, with the lowest results achieved with the former. + +In terms of $\langle \mathrm{eol} \rangle$ and $\langle \mathrm{eob} \rangle$ coverage, the former increases with larger min values, which is expected given the more restricted space to insert these end-of-line markers as the value increases; for $\langle \mathrm{eob} \rangle$ , the restricted insertion space results in increased under-generation, which in turn results in better scores for lower values of the min parameter. + +
CountCharsMLM
They're things you access through your <eol> computer. <eob>They're things you access <eol> through your computer. <eob>
Every row of data is a life whose story <eol> deserves to be told with dignity. <eob>Every row of data is a life <eol> whose story deserves to be told <eob> with dignity. <eob>
During the winter, struggling to get <eol> warm, my neighbors would have no choice <eob> but to bypass the meter after their heat <eol> was shut off, just to keep their family <eob> comfortable for one more day. <eob>During the winter, struggling to get warm, <eol> my neighbors would have no choice <eob> but to bypass the meter <eol> after their heat was shut off, <eob> just to keep their family comfortable <eol> for one more day. <eob>
+ +Table 4: Examples of subtitles segmented via character counting and MLM-based mask insertion + +
LanguageMethodBLEUSigmaEOLEOBCPL
DEMLM10070.8518.53-7.96100%
MLM+OC10075.6319.81-7.78100%
ENMLM10076.7719.18-9.91100%
MLM+OC10077.8919.86-9.73100%
ESMLM10073.4712.98-4.91100%
MLM+OC10075.5913.45-4.63100%
FRMLM10073.7816.51-6.58100%
MLM+OC10076.0717.47-6.12100%
ITMLM10071.3818.49-9.55100%
MLM+OC10074.2020.34-8.57100%
NLMLM10068.7120.37-9.96100%
MLM+OC10073.8821.22-9.96100%
PTMLM10071.5920.03-10.81100%
MLM+OC10075.5019.87-10.02100%
ROMLM10069.4523.37-10.44100%
MLM+OC10074.1323.37-10.09100%
+ +Table 5: Complete results with MLM mask insertion on the MuST-Cinema test sets (min=15) + +
LanguageminBLEUSigmaEOLEOB
DE110072.3128.75-0.18
1010073.9622.68-4.43
1510075.6319.81-7.78
2010075.2814.54-11.21
EN110074.3037.33-0.98
1010077.1424.49-7.77
1510077.8919.86-9.73
2010077.1615.24-12.68
ES110073.0020.870.28
1010074.3218.24-2.04
1510075.5913.45-4.63
2010075.838.66-7.87
FR110073.8924.68-0.73
1010075.2620.83-3.93
1510076.0717.47-6.12
2010076.7512.5-10.05
IT110072.0129.75-3.66
1010073.7524.71-6.61
1510074.2020.34-8.57
2010073.6614.62-11.61
NL110072.1626.83-5.47
1010073.5623.26-8.47
1510073.8821.22-9.96
2010074.4016.81-12.43
PT110072.8726.38-6.24
1010074.5322.15-8.08
1510075.5019.87-10.02
2010074.9814.17-13.36
RO110072.0532.3-4.51
1010073.7626.98-7.52
1510074.1323.37-10.09
2010074.8917.53-12.83
+ +Table 6: Test set results with the MLM+OC method and different values of the min parameter + +A For every submission: + +A1. Did you describe the limitations of your work? 7 +A2. Did you discuss any potential risks of your work? 8 +A3. Do the abstract and introduction summarize the paper's main claims? +A4. Have you used AI writing assistants when working on this paper? Left blank. + +B Did you use or create scientific artifacts? + +Not applicable. Left blank. + +B1. Did you cite the creators of artifacts you used? Not applicable. Left blank. +B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Not applicable. Left blank. +B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Not applicable. Left blank. +B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Not applicable. Left blank. +B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Not applicable. Left blank. +B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Not applicable. Left blank. + +C Did you run computational experiments? + +3 + +C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? + +We didn't trained any models for this paper, and inference was performed on CPU. + +The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance. + +C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? 3 +C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? 5 +C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? + +# D Did you use human annotators (e.g., crowdworkers) or research with human participants? + +Left blank. + +D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Not applicable. Left blank. +D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Not applicable. Left blank. +D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Not applicable. Left blank. +D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Not applicable. Left blank. +D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Not applicable. Left blank. \ No newline at end of file diff --git a/2023/Unsupervised Subtitle Segmentation with Masked Language Models/images.zip b/2023/Unsupervised Subtitle Segmentation with Masked Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fc479b0f7c17f24139e70361b2797ffaf33e32bd --- /dev/null +++ b/2023/Unsupervised Subtitle Segmentation with Masked Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7fc843b1f83aee3546f2a12c52d2a081a77010f76a68f840601f19ab8afc9b0d +size 492404 diff --git a/2023/Unsupervised Subtitle Segmentation with Masked Language Models/layout.json b/2023/Unsupervised Subtitle Segmentation with Masked Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1d5ffae1d2de90f248af52fee90133ea37c6aea2 --- /dev/null +++ b/2023/Unsupervised Subtitle Segmentation with Masked Language Models/layout.json @@ -0,0 +1,5680 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 87, + 75, + 507, + 94 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 75, + 507, + 94 + ], + "spans": [ + { + "bbox": [ + 87, + 75, + 507, + 94 + ], + "type": "text", + "content": "Unsupervised Subtitle Segmentation with Masked Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 136, + 114, + 458, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 114, + 458, + 128 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 458, + 128 + ], + "type": "text", + "content": "David Ponce\\*1,2 and Thierry Etchegoyhen\\*1 and Victor Ruiz" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 113, + 129, + 483, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 129, + 483, + 142 + ], + "spans": [ + { + "bbox": [ + 113, + 129, + 483, + 142 + ], + "type": "text", + "content": "1 Vicomtech Foundation, Basque Research and Technology Alliance (BRTA)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 185, + 143, + 410, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 143, + 410, + 157 + ], + "spans": [ + { + "bbox": [ + 185, + 143, + 410, + 157 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 185, + 143, + 410, + 157 + ], + "type": "text", + "content": " University of the Basque Country UPV/EHU" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 149, + 158, + 449, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 158, + 449, + 171 + ], + "spans": [ + { + "bbox": [ + 149, + 158, + 449, + 171 + ], + "type": "text", + "content": "{adponce,tetchegoyhen,vruiz}@vicomtech.org" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 84, + 234, + 274, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 234, + 274, + 425 + ], + "spans": [ + { + "bbox": [ + 84, + 234, + 274, + 425 + ], + "type": "text", + "content": "We describe a novel unsupervised approach to subtitle segmentation, based on pretrained masked language models, where line endings and subtitle breaks are predicted according to the likelihood of punctuation to occur at candidate segmentation points. Our approach obtained competitive results in terms of segmentation accuracy across metrics, while also fully preserving the original text and complying with length constraints. Although supervised models trained on in-domain data and with access to source audio information can provide better segmentation accuracy, our approach is highly portable across languages and domains and may constitute a robust off-the-shelf solution for subtitle segmentation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 68, + 435, + 155, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 435, + 155, + 448 + ], + "spans": [ + { + "bbox": [ + 68, + 435, + 155, + 448 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 456, + 291, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 456, + 291, + 577 + ], + "spans": [ + { + "bbox": [ + 67, + 456, + 291, + 577 + ], + "type": "text", + "content": "Subtitling is one of the principal means of providing accessible audiovisual content. With the ever increasing production of audiovisual content in multiple domains and languages, in the current digital era, subtitle provision can benefit from automation support, via Automatic Speech Recognition and/or Machine Translation (Volk et al., 2010; Aliprandi et al., 2014; Etchegoyhen et al., 2014; Tardel, 2020; Bojar et al., 2021)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 578, + 291, + 754 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 578, + 291, + 754 + ], + "spans": [ + { + "bbox": [ + 67, + 578, + 291, + 754 + ], + "type": "text", + "content": "Subtitles are subject to specific constraints in order to achieve adequate readability, including layout, on-screen duration and text editing. Among these constraints, segmentation addresses the maximum number of characters per line, the number of lines per subtitle, and breaks at natural linguistic frontiers. Segmentation has been shown to be an important readability factor (Perego et al., 2010; Rajendran et al., 2013), with improperly segmented subtitles resulting in increased cognitive effort and reading times for users. Thus, automated subtitle systems need to generate properly segmented subtitles to achieve readability." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 213, + 526, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 213, + 526, + 523 + ], + "spans": [ + { + "bbox": [ + 302, + 213, + 526, + 523 + ], + "type": "text", + "content": "A typical baseline for subtitle segmentation, still used in some production systems, is simple character counting, whereby line breaks are inserted before reaching the maximum allowed number of characters per line. Although simple and fast, this approach does not address the need for linguistically correct segments and, therefore, falls short in terms of readability. Several approaches have been proposed to improve segmentation by automated means. Álvarez et al. (2014) proposed a machine learning method where subtitle breaks are predicted by Support Vector Machine and Linear Regression models trained on professionally-created subtitles. A similar method based on Conditional Random Fields was then shown to improve over these results (Alvarez et al., 2017). Approaches that directly generate subtitle breaks within Neural Machine Translation have also been proposed in recent years (Matusov et al., 2019; Karakanta et al., 2020a). Recently, Papi et al. (2022) developed a multilingual segmenter which generates both text and breaks and may be trained on textual input only, or on joint text and audio data." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 527, + 526, + 743 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 527, + 526, + 743 + ], + "spans": [ + { + "bbox": [ + 302, + 527, + 526, + 743 + ], + "type": "text", + "content": "Although quality subtitle segmentation may be achieved with the aforementioned approaches, they require supervised training on segmented subtitle corpora. At present, the largest subtitle corpus is Open Subtitles (Lison et al., 2018), which mainly covers entertainment material, contains subtitles mostly created by non-professionals or automatically translated, and does not include line breaks. The MuST-Cinema corpus (Karakanta et al., 2020b), on the other hand, is a multilingual speech translation corpus that includes subtitles breaks, but is only available for 8 languages at the moment. Considering the vast amount of languages and domains in audiovisual content, the lack of segmented training data hinders the development of robust automated subtitleing systems." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "type": "text", + "content": "In this work, we describe a novel unsupervised method based on pretrained masked language mod" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 81, + 761, + 255, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 761, + 255, + 772 + ], + "spans": [ + { + "bbox": [ + 81, + 761, + 255, + 772 + ], + "type": "text", + "content": "*These authors contributed equally to this work." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 289, + 780, + 305, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 780, + 305, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 780, + 305, + 791 + ], + "type": "text", + "content": "771" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "spans": [ + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "type": "text", + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 224, + 806, + 368, + 817 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 806, + 368, + 817 + ], + "spans": [ + { + "bbox": [ + 224, + 806, + 368, + 817 + ], + "type": "text", + "content": "Volume 2: Short Papers, pages 771-781" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 176, + 818, + 417, + 828 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 818, + 417, + 828 + ], + "spans": [ + { + "bbox": [ + 176, + 818, + 417, + 828 + ], + "type": "text", + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 292, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 292, + 220 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 292, + 220 + ], + "type": "text", + "content": "els (MLM), where line and subtitle breaks are inserted according to the likelihood of a segment acting as an isolated unit, as approximated by the probability of a punctuation mark occurring at a given segmentation point. In our experiments, this novel approach obtained competitive results on most metrics, while also fully preserving the original text and complying with length constraints. Our system may thus be used as a simple yet efficient subtitle segmenter with any pretrained masked language model, for any language covered by the model." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 232, + 140, + 245 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 232, + 140, + 245 + ], + "spans": [ + { + "bbox": [ + 67, + 232, + 140, + 245 + ], + "type": "text", + "content": "2 Approach" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 254, + 291, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 254, + 291, + 428 + ], + "spans": [ + { + "bbox": [ + 67, + 254, + 291, + 428 + ], + "type": "text", + "content": "Our approach is based on the standard view that the more appropriate subtitle segments are those that may function as isolated grammatical chunks. We further hypothesise that a relevant approximation for the identification of this type of unit is the likelihood of a punctuation mark being inserted at the end of a candidate segment, as punctuation may mark the closure of a syntactic unit and is often associated with discursive pauses. To test this hypothesis, we compute the likelihood of punctuation marks at different segmentation points, as predicted by a pretrained MLM, and select the insertion point with the highest likelihood.1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 430, + 291, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 430, + 291, + 539 + ], + "spans": [ + { + "bbox": [ + 67, + 430, + 291, + 539 + ], + "type": "text", + "content": "The segmentation candidates are determined under a sliding-window approach over the entire input text. We first generate the list of all pairs " + }, + { + "bbox": [ + 67, + 430, + 291, + 539 + ], + "type": "inline_equation", + "content": "< \\alpha, \\beta>" + }, + { + "bbox": [ + 67, + 430, + 291, + 539 + ], + "type": "text", + "content": " over the unprocessed portion of the text, where " + }, + { + "bbox": [ + 67, + 430, + 291, + 539 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 67, + 430, + 291, + 539 + ], + "type": "text", + "content": " is a segmentation candidate of length under a specified limit " + }, + { + "bbox": [ + 67, + 430, + 291, + 539 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 67, + 430, + 291, + 539 + ], + "type": "text", + "content": ", corresponding to the maximum number of characters per line, and " + }, + { + "bbox": [ + 67, + 430, + 291, + 539 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 67, + 430, + 291, + 539 + ], + "type": "text", + "content": " is the remaining portion of the text to be segmented." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 539, + 291, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 539, + 291, + 634 + ], + "spans": [ + { + "bbox": [ + 67, + 539, + 291, + 634 + ], + "type": "text", + "content": "We then score all segmentation candidates " + }, + { + "bbox": [ + 67, + 539, + 291, + 634 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 67, + 539, + 291, + 634 + ], + "type": "text", + "content": " with one of the LM scoring variants described below. A segmentation marker, either end-of-line (), or end-of-block indicating the end of a subtitle (), is then appended to the best scoring candidate, and " + }, + { + "bbox": [ + 67, + 539, + 291, + 634 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 67, + 539, + 291, + 634 + ], + "type": "text", + "content": " becomes the input text to be segmented in a recursive iteration of the process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 634, + 291, + 742 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 634, + 291, + 742 + ], + "spans": [ + { + "bbox": [ + 67, + 634, + 291, + 742 + ], + "type": "text", + "content": "Since our method does not rely on any additional information, such as an audio source, to determine the segmentation type, an tag is inserted every even segment or when " + }, + { + "bbox": [ + 67, + 634, + 291, + 742 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 67, + 634, + 291, + 742 + ], + "type": "text", + "content": " is empty; otherwise, an tag is inserted. We thus generate subtitles with a maximum of two lines, following a standard recommendation in subtitling. We also define a minimal number of characters (min) in " + }, + { + "bbox": [ + 67, + 634, + 291, + 742 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 67, + 634, + 291, + 742 + ], + "type": "text", + "content": " for the" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 71, + 524, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 524, + 97 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 524, + 97 + ], + "type": "text", + "content": "segmentation process to apply, and do not segment lines that are under the specified character limit." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 99, + 525, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 99, + 525, + 125 + ], + "spans": [ + { + "bbox": [ + 302, + 99, + 525, + 125 + ], + "type": "text", + "content": "We evaluated three approaches to compute segmentation scores over each candidate pair " + }, + { + "bbox": [ + 302, + 99, + 525, + 125 + ], + "type": "inline_equation", + "content": "< \\alpha, \\beta>" + }, + { + "bbox": [ + 302, + 99, + 525, + 125 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 316, + 133, + 524, + 285 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 316, + 133, + 524, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 133, + 524, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 133, + 524, + 174 + ], + "type": "text", + "content": "- Substitution: The last token of " + }, + { + "bbox": [ + 316, + 133, + 524, + 174 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 316, + 133, + 524, + 174 + ], + "type": "text", + "content": " is masked and the score is the highest MLM probability among punctuation marks on this mask." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 316, + 183, + 524, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 183, + 524, + 222 + ], + "spans": [ + { + "bbox": [ + 316, + 183, + 524, + 222 + ], + "type": "text", + "content": "- Insertion: A mask is appended to " + }, + { + "bbox": [ + 316, + 183, + 524, + 222 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 316, + 183, + 524, + 222 + ], + "type": "text", + "content": " and the score is the highest MLM probability among punctuation marks on this mask." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 316, + 232, + 524, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 232, + 524, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 524, + 285 + ], + "type": "text", + "content": "- LM-Score: The score is the average of the perplexity of " + }, + { + "bbox": [ + 316, + 232, + 524, + 285 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 316, + 232, + 524, + 285 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 316, + 232, + 524, + 285 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 316, + 232, + 524, + 285 + ], + "type": "text", + "content": ", as derived from the MLM probabilities for each token in the corresponding sequence." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 302, + 294, + 525, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 294, + 525, + 401 + ], + "spans": [ + { + "bbox": [ + 302, + 294, + 525, + 401 + ], + "type": "text", + "content": "The first two methods are variants of our core approach. The third method, while also based on the same pretrained MLM, relies instead on the pseudoperplexity of the sequences according to the MLM, computed following Salazar et al. (2020). We included this latter variant to measure the potential of using LM scoring directly, without resorting to the likelihood of punctuation marks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 412, + 427, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 412, + 427, + 426 + ], + "spans": [ + { + "bbox": [ + 302, + 412, + 427, + 426 + ], + "type": "text", + "content": "3 Experimental Setup" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 433, + 525, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 433, + 525, + 568 + ], + "spans": [ + { + "bbox": [ + 302, + 433, + 525, + 568 + ], + "type": "text", + "content": "Corpora. For all experiments, we used the MustST-Cinema corpus (Karakanta et al., 2020b), which is derived from TED talks and contains both line and subtitle break markers. In addition to being publicly available, it also allows for a direct comparison with the supervised models of Papi et al. (2022). We report results of our approach on the 6 MuST-Cinema datasets for which comparative results were available, directly predicting segmentation on the test sets without any training." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 576, + 525, + 724 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 576, + 525, + 724 + ], + "spans": [ + { + "bbox": [ + 302, + 576, + 525, + 724 + ], + "type": "text", + "content": "Methods. For our approach, we tested the three variants described in Section 2. We used BERT (Devlin et al., 2019) as our MLM for all languages. Additionally, we included a variant called overt clueing " + }, + { + "bbox": [ + 302, + 576, + 525, + 724 + ], + "type": "inline_equation", + "content": "(OC)" + }, + { + "bbox": [ + 302, + 576, + 525, + 724 + ], + "type": "text", + "content": ", where an overt punctuation mark at the end of a candidate segment increments the mask score by 1. We then compared the results of the best LM-based variant with those obtained by alternative approaches. In all cases, our results were computed with " + }, + { + "bbox": [ + 302, + 576, + 525, + 724 + ], + "type": "inline_equation", + "content": "min = 15" + }, + { + "bbox": [ + 302, + 576, + 525, + 724 + ], + "type": "text", + "content": ", as this value obtained the best results overall over the development" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 729, + 525, + 751 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 729, + 525, + 751 + ], + "spans": [ + { + "bbox": [ + 302, + 729, + 525, + 751 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 302, + 729, + 525, + 751 + ], + "type": "text", + "content": "Our results on all remaining languages of the MuST-Cinema datasets are presented in Appendix B." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 302, + 751, + 525, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 751, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 525, + 772 + ], + "type": "inline_equation", + "content": "{}^{3}" + }, + { + "bbox": [ + 302, + 751, + 525, + 772 + ], + "type": "text", + "content": " Specifically bert-base-uncased as available on HugginFace (https://huggingface.co/),accessed on November 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 67, + 750, + 291, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 750, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 750, + 291, + 772 + ], + "type": "text", + "content": "1 Throughout our experiments, we used the following punctuation marks: ' ', ' ', ' ?', ' !', ' ' and ' '." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "text", + "content": "772" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 84, + 68, + 510, + 169 + ], + "blocks": [ + { + "bbox": [ + 84, + 68, + 510, + 169 + ], + "lines": [ + { + "bbox": [ + 84, + 68, + 510, + 169 + ], + "spans": [ + { + "bbox": [ + 84, + 68, + 510, + 169 + ], + "type": "table", + "html": "
EnglishSpanishGerman
MethodSigmaEOLEOBSigmaEOLEOBSigmaEOLEOB
Substitution71.65+19.86-10.9669.34+12.36-5.7469.31+19.05-7.05
Insertion76.77+19.18-9.9173.47+12.98-4.9170.85+18.53-7.96
LM-Score69.97+21.40-8.6667.70+13.29-5.3764.07+16.45-6.51
", + "image_path": "30f4ac866468a79f0f6ce81bb1faac41531cef49cb27a9058316e4c83b9fbc3b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 121, + 176, + 472, + 190 + ], + "lines": [ + { + "bbox": [ + 121, + 176, + 472, + 190 + ], + "spans": [ + { + "bbox": [ + 121, + 176, + 472, + 190 + ], + "type": "text", + "content": "Table 1: Sigma and break coverage test set results for LM-based segmentation variants" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 210, + 290, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 210, + 290, + 237 + ], + "spans": [ + { + "bbox": [ + 67, + 210, + 290, + 237 + ], + "type": "text", + "content": "sets, although the differences were minor with the other values we tested (1, 10 and 20).4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 239, + 291, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 239, + 291, + 455 + ], + "spans": [ + { + "bbox": [ + 69, + 239, + 291, + 455 + ], + "type": "text", + "content": "We used the simple character counting approach (hereafter, CountChars) as baseline, and, as representative supervised methods on the selected datasets, the models described by (Papi et al., 2022). Their core supervised approach is based on a Transformer (Vaswani et al., 2017) architecture with 3 encoder layers and 3 decoder layers, trained on textual MuST-Cinema input only (MC.Text), or on complementary audio data as well via an additional speech encoder with 12 encoder layers (MC.Multi). They trained each variant on either monolingual data alone (mono), or in a multilingual setting (multi). Finally, they also report results for a variant (OS.Text) trained on the Open Subtitles corpus (Lison et al., 2018) for their zero-shot experiments." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 469, + 291, + 726 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 469, + 291, + 726 + ], + "spans": [ + { + "bbox": [ + 67, + 469, + 291, + 726 + ], + "type": "text", + "content": "Evaluation. We use the subtitle-oriented metric Sigma (Karakanta et al., 2022), which computes the ratio of achieved BLEU (Papineni et al., 2002) over an approximated upper-bound BLEU score, on text that includes line and subtitle breaks. Sigma is meant to support the evaluation of imperfect texts, i.e. text that differs from the reference when breaks are omitted. Although our approach does not produce imperfect text, achieving perfect BLEU scores when breaks are ignored, we used this metric for comparison purposes. We also report break coverage results (Papi et al., 2022), defined as the ratio of predicted breaks over reference breaks, which we computed separately for the EOL and EOB breaks. Finally, we include length conformity results (CPL), measured as the percentage of subtitle lines whose length is under the maximum number of characters defined by the subtitle guidelines (42 in the TED guidelines" + }, + { + "bbox": [ + 67, + 469, + 291, + 726 + ], + "type": "inline_equation", + "content": "^5" + }, + { + "bbox": [ + 67, + 469, + 291, + 726 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 302, + 210, + 498, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 210, + 498, + 224 + ], + "spans": [ + { + "bbox": [ + 302, + 210, + 498, + 224 + ], + "type": "text", + "content": "4 LM-based Segmentation Variants" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 231, + 527, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 231, + 527, + 435 + ], + "spans": [ + { + "bbox": [ + 302, + 231, + 527, + 435 + ], + "type": "text", + "content": "We first compared the three methods described in Section 2 on the English, Spanish and German datasets, with the results described in Table 1. In terms of Sigma, the Insertion method obtained the best results in all cases. It also obtained the best scores in terms of coverage for the EOL marker, except in Spanish, although all three variants tend to overgenerate end-of-line markers to similar extents. The LM-Score variant obtained the worst results in terms of Sigma, but outperformed the alternatives in terms of EOB coverage, a metric on which the three variants performed markedly better than on EOL coverage. Considering the overall results, we selected the Insertion variant as the most balanced one for all remaining experiments reported below." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 444, + 432, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 444, + 432, + 459 + ], + "spans": [ + { + "bbox": [ + 302, + 444, + 432, + 459 + ], + "type": "text", + "content": "5 Comparative Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 466, + 527, + 724 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 466, + 527, + 724 + ], + "spans": [ + { + "bbox": [ + 302, + 466, + 527, + 724 + ], + "type": "text", + "content": "In Table 2, we present the results obtained by the selected approaches on the languages for which results were available with supervised models trained on in-domain data. Overall, our approach outperformed the CountChars baseline across the board, and was in turn outperformed by the supervised variants in terms of Sigma scores. Although it is clear from these results that training segmentation models on in-domain data, with or without audio data, provides clear advantages in terms of subtitle segmentation, it is worth noting that Sigma does not, by design, reflect the actual BLEU score without breaks, i.e. the generation of imperfect text, which is a by-product of the above supervised approaches and non-existent in ours. In terms of CPL, all supervised models generate subtitle lines that overflow the limit, to a significant degree, whereas the selected unsupervised models trivially respect the length constraint." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 731, + 526, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 731, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 731, + 526, + 772 + ], + "type": "text", + "content": "The results indicated in Table 3 on unseen data seem to indicate that their MC.Multi model can reach BLEU scores close to 100, thereby limiting the negative impact of imperfect text generation in these cases." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 739, + 289, + 761 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 739, + 289, + 761 + ], + "spans": [ + { + "bbox": [ + 67, + 739, + 289, + 761 + ], + "type": "text", + "content": "4See Appendix C for results with different values of the min parameter." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 79, + 761, + 287, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 761, + 287, + 772 + ], + "spans": [ + { + "bbox": [ + 79, + 761, + 287, + 772 + ], + "type": "text", + "content": "5https://www.ted.com/participate/translate/subtitling-tips" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "text", + "content": "773" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 79, + 68, + 515, + 245 + ], + "blocks": [ + { + "bbox": [ + 79, + 68, + 515, + 245 + ], + "lines": [ + { + "bbox": [ + 79, + 68, + 515, + 245 + ], + "spans": [ + { + "bbox": [ + 79, + 68, + 515, + 245 + ], + "type": "table", + "html": "
EnglishFrenchGermanItalian
MethodTrainingSigmaCPLSigmaCPLSigmaCPLSigmaCPL
CountCharsN/A63.71100%62.87100%62.34100%61.49100%
MC.Textmono84.8796.6%83.6896.7%83.6290.9%82.2290.0%
multi85.9888.5%84.5694.3%84.0290.9%83.0491.2%
MC.Multimono85.7694.8%84.2593.9%84.2291.4%82.6289.9%
multi87.4495.0%86.4994.1%86.4089.9%85.3390.0%
MLMN/A76.77100%73.78100%70.85100%71.38100%
MLM+OCN/A77.89100%76.07100%75.63100%74.20100%
", + "image_path": "f30c2fad5dd35b609c4efd393d7c19953b6c96cf42537d7e72eb9e22dfd5f454.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 77, + 284, + 280, + 575 + ], + "blocks": [ + { + "bbox": [ + 68, + 253, + 524, + 265 + ], + "lines": [ + { + "bbox": [ + 68, + 253, + 524, + 265 + ], + "spans": [ + { + "bbox": [ + 68, + 253, + 524, + 265 + ], + "type": "text", + "content": "Table 2: Comparative results between unsupervised methods and supervised approaches trained on in-domain data" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 77, + 284, + 280, + 575 + ], + "lines": [ + { + "bbox": [ + 77, + 284, + 280, + 575 + ], + "spans": [ + { + "bbox": [ + 77, + 284, + 280, + 575 + ], + "type": "table", + "html": "
Dutch
MethodBLEUSigmaCPLEOLEOB
CountChars10063.2100%-21.2-7.1
OS.Text89.564.471.2%-31.4-51.3
MC.Text61.374.477.8%-23.4-9.9
MC.Multi99.980.391.4%-27.20.4
MLM10068.7100%+20.4-10.0
MLM+OC10073.9100%+21.2-10.0
Spanish
MethodBLEUSigmaCPLEOLEOB
CountChars10063.2100%-24.6-4.4
OS.Text92.664.171.2%-32.3-45.4
MC.Text69.675.870.1%-47.6-19.3
MC.Multi99.678.791.8%-22.44.7
MLM10073.5100%+13.0-4.9
MLM+OC10075.6100%+13.4-4.6
", + "image_path": "5197bb7b925f215a4dcf35c2683653aba92c2ed88e910d73c03d8095ef1bb50d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 582, + 289, + 608 + ], + "lines": [ + { + "bbox": [ + 67, + 582, + 289, + 608 + ], + "spans": [ + { + "bbox": [ + 67, + 582, + 289, + 608 + ], + "type": "text", + "content": "Table 3: Comparative results between unsupervised methods and zero-short supervised approaches" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 638, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 638, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 638, + 291, + 772 + ], + "type": "text", + "content": "In Table 3, we show the comparative results between the selected unsupervised methods and the supervised variants, in languages where zero-shot results were available for the latter approaches. In this scenario, in terms of Sigma our approach obtained results on a par with the supervised MC.Text models trained on in-domain MuST-Cinema data, outperformed the OS.Text models trained on Open Subtitles data, and was surpassed by the MC.Multi model, which exploits additional audio information," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 302, + 286, + 526, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 286, + 526, + 490 + ], + "spans": [ + { + "bbox": [ + 302, + 286, + 526, + 490 + ], + "type": "text", + "content": "by 3.1 and 6.4 points. In terms of break coverage, in most cases our unsupervised method outperformed the supervised variants, to a significant degree compared to the text-based OS.Text and MC.Text models. Regarding BLEU scores without breaks, only the MC.Multi model reaches a score close to the perfect one achieved by the unsupervised models, whereas the MC.Text model is outperformed by 38.7 and 31.4 points in Dutch and Spanish, respectively. In all cases, the CPL scores indicate that none of the supervised approaches fully meet the length constraint, leading to overflowing lines in " + }, + { + "bbox": [ + 302, + 286, + 526, + 490 + ], + "type": "inline_equation", + "content": "8.2\\%" + }, + { + "bbox": [ + 302, + 286, + 526, + 490 + ], + "type": "text", + "content": " of the cases at best and " + }, + { + "bbox": [ + 302, + 286, + 526, + 490 + ], + "type": "inline_equation", + "content": "29.9\\%" + }, + { + "bbox": [ + 302, + 286, + 526, + 490 + ], + "type": "text", + "content": " at worst. In this scenario as well, the unsupervised approaches fully meet the length constraint, by design." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 496, + 526, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 496, + 526, + 631 + ], + "spans": [ + { + "bbox": [ + 302, + 496, + 526, + 631 + ], + "type": "text", + "content": "Overall, overt clueing improved over our core method by an average of 3.12 Sigma points, indicating that some likely punctuation configurations were not properly captured by our MLM approximation. In general, our approach tends to overgenerate EOL markers, whereas the opposite is true for the selected supervised models. Determining which of these tendencies leads to better subtitle readability would require a specific human evaluation which we leave for future research." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 638, + 527, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 638, + 527, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 638, + 527, + 772 + ], + "type": "text", + "content": "Although the zero-shot Sigma results obtained by the supervised MC.Multi method show the potential of this approach to provide pretrained models applicable to other languages, two important aspects are worth considering. First, the available zero-shot results were obtained on datasets in the same domain as the data seen to train the supervised models. A more complete assessment of the capabilities of these models in zero-shot settings, which would be the most frequent scenario consid-" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "774" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 291, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 291, + 220 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 291, + 220 + ], + "type": "text", + "content": "ering the lack of training data across domains and languages, would require specific evaluations in other domains. Secondly, although segmentation is a key aspect for subtitle readability, length conformity is an equally important constraint, if not more so considering that subtitles with lines over the CPL limit are considered invalid in subtitleing. Our proposed unsupervised method can thus be seen as a pragmatic approach which guarantees valid subtitles while also providing quality segmentation across the board." + }, + { + "bbox": [ + 67, + 71, + 291, + 220 + ], + "type": "inline_equation", + "content": "^7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 232, + 151, + 246 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 232, + 151, + 246 + ], + "spans": [ + { + "bbox": [ + 67, + 232, + 151, + 246 + ], + "type": "text", + "content": "6 Conclusions" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 256, + 291, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 256, + 291, + 323 + ], + "spans": [ + { + "bbox": [ + 67, + 256, + 291, + 323 + ], + "type": "text", + "content": "We described an unsupervised approach to subtitle segmentation, based on pretrained masked language models, where line or subtitle breaks are inserted according to the likelihood of punctuation occurring at candidate segmentation points." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 325, + 291, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 325, + 291, + 419 + ], + "spans": [ + { + "bbox": [ + 67, + 325, + 291, + 419 + ], + "type": "text", + "content": "Although supervised models, trained on indomain data with audio support, were shown to perform better that this simple textual approach in terms of the Sigma metric, they tend to generate imperfect text to varying degrees, while also failing to fully meet length constraints that are essential for subtitling." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 421, + 291, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 421, + 291, + 542 + ], + "spans": [ + { + "bbox": [ + 67, + 421, + 291, + 542 + ], + "type": "text", + "content": "In contrast, our LM-based textual approach outperformed supervised models in most cases in terms of break generation coverage, while also fully preserving the original text, complying with length constraints, and obtaining competitive results in terms of Sigma. This simple approach may thus provide a highly portable complementary solution for subtitle segmentation across languages and domains." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 555, + 149, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 555, + 149, + 567 + ], + "spans": [ + { + "bbox": [ + 67, + 555, + 149, + 567 + ], + "type": "text", + "content": "7 Limitations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 578, + 291, + 740 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 578, + 291, + 740 + ], + "spans": [ + { + "bbox": [ + 67, + 578, + 291, + 740 + ], + "type": "text", + "content": "The first clear limitation of our approach is its text-based nature. This prevents important audio information, typically silences in speech patterns, from being exploited to generate subtitle breaks. A more complete system could be devised though, for instance by associating our text-based approach with the information provided by a forced alignment toolkit, whenever audio information is available. A simple method along these lines could be the following: 1. Apply our MLM-based segmentation but only generating a unique segmentation tag SEG; 2. Insert EOB markers wherever the" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 71, + 526, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 179 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 179 + ], + "type": "text", + "content": "silence between two aligned words is above a specified threshold; 3. Traverse the text sequentially and replace SEG with EOL if there exists a previous marker of type EOB, otherwise replace with EOB. We left this use of our method in combination with audio information for future research, as audio alignment for subtitles typically involves additional factors such as non-literal transcriptions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 180, + 526, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 180, + 526, + 396 + ], + "spans": [ + { + "bbox": [ + 302, + 180, + 526, + 396 + ], + "type": "text", + "content": "Additionally, our method is limited in its adaptability to specific segmentation guidelines, which may be company-specific. The main adaptable parameters of our methods are the minimum and maximum parameters of the segmentation window, and the set of predefined punctuation marks over which masking is computed, neither of which could fully model idiosyncratic segmentation guidelines. However, in our experience at least, segmentation in real professional data tends to display varying degrees of consistency with respect to guidelines, and natural linguistic breaks seem to be the dominant factor for subtitle segmentation. A specific evaluation would be needed on data from varied professional datasets to determine the extent to which our method might deviate from specific guidelines." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 398, + 525, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 398, + 525, + 572 + ], + "spans": [ + { + "bbox": [ + 302, + 398, + 525, + 572 + ], + "type": "text", + "content": "Finally, other aspects of subtitling, such as the recommendation in some guidelines for subtitles to appear in a pyramidal view, i.e. with the first line shorter than the second line, have not been taken into consideration in this work. Our aim was to evaluate our core LM-based approach without additional variables that can vary across guidelines and may also have led to results that are more difficult to interpret overall. Our approach could nonetheless be easily augmented with constraints on relative line lengths within subtitles, by incrementing the scores of segmentation candidates that respect this surface-level constraint." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 587, + 441, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 587, + 441, + 600 + ], + "spans": [ + { + "bbox": [ + 302, + 587, + 441, + 600 + ], + "type": "text", + "content": "8 Ethical Considerations" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 611, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 611, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 611, + 526, + 772 + ], + "type": "text", + "content": "Our approach involves the use of large pretrained language models, whose computational performance is typically higher when deployed in more powerful environments with GPUs. Under such usage, electric consumption and associated carbon footprint are likely to increase and users of our method under these conditions should be aware of this type of impact. However, subtitle segmentation is often performed offline, where efficient processing is less of a concern, and lower-cost CPU deployments are an entirely viable option. All our results were obtained with a single large LM de" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 750, + 291, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 750, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 750, + 291, + 772 + ], + "type": "text", + "content": "Examples of segmented subtitles can be found in Appendix A." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "text", + "content": "775" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 289, + 98 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 289, + 98 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 289, + 98 + ], + "type": "text", + "content": "ployed on CPU, with the aim of reducing energy consumption at inference time." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 100, + 291, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 100, + 291, + 275 + ], + "spans": [ + { + "bbox": [ + 67, + 100, + 291, + 275 + ], + "type": "text", + "content": "Additionally, our method requires no training for the task at hand and thus removes the cost of model training associated with the supervised methods with which we compare our results. For instance, Papi et al. (2022) indicate that they use four K80 GPUs to train their models, which we took as comparison points, with 1 day of training for their text-only models and 1 week for their multimodal segmenters. Therefore, given the large number of potential language pairs and domains in need of segmented subtitle content, our approach can provide competitive results with a comparatively lesser impact on energy resource consumption." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 68, + 288, + 170, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 288, + 170, + 301 + ], + "spans": [ + { + "bbox": [ + 68, + 288, + 170, + 301 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 312, + 290, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 312, + 290, + 391 + ], + "spans": [ + { + "bbox": [ + 67, + 312, + 290, + 391 + ], + "type": "text", + "content": "We thank the anonymous reviewers for their helpful comments. This work was partially supported by the Department of Economic Development and Competitiveness of the Basque Government (Spri Group) through funding for the StreAmS project (ZL-2021/00700)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 417, + 127, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 417, + 127, + 430 + ], + "spans": [ + { + "bbox": [ + 68, + 417, + 127, + 430 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 439, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 68, + 439, + 291, + 526 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 439, + 291, + 526 + ], + "spans": [ + { + "bbox": [ + 68, + 439, + 291, + 526 + ], + "type": "text", + "content": "Carlo Aliprandi, Cristina Scudellari, Isabella Gallucci, Nicola Piccinini, Matteo Raffaelli, Arantza del Pozo, Aitor Alvarez, Haritz Arzelus, Renato Cassaca, Tiago Luis, et al. 2014. Automatic live subtitling: state of the art, expectations and current trends. In Proceedings of NAB Broadcast Engineering Conference: Papers on Advanced Media Technologies, Las Vegas, volume 13." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 539, + 291, + 595 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 539, + 291, + 595 + ], + "spans": [ + { + "bbox": [ + 69, + 539, + 291, + 595 + ], + "type": "text", + "content": "Aitor Álvarez, Haritz Arzelus, and Thierry Etchegoyhen. 2014. Towards customized automatic segmentation of subtitles. In Advances in Speech and Language Technologies for Iberian Languages, pages 229-238. Springer." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 606, + 291, + 661 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 606, + 291, + 661 + ], + "spans": [ + { + "bbox": [ + 69, + 606, + 291, + 661 + ], + "type": "text", + "content": "Aitor Alvarez, Carlos-D Martínez-Hinarejos, Haritz Arzelus, Marina Balenciaga, and Arantza del Pozo. 2017. Improving the automatic segmentation of subtitles through conditional random field. Speech Communication, 88:83-95." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 673, + 291, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 673, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 673, + 291, + 772 + ], + "type": "text", + "content": "Ondrej Bojar, Dominik Machacek, Sangeet Sagar, Otakar Smrz, Jonas Kratochvil, Peter Polak, Ebrahim Ansari, Mohammad Mahmoudi, Rishu Kumar, Dario Franceschini, Chiara Canton, Ivan Simonini, Thai Son Nguyen, Felix Schneider, Sebastian Stüker, Alex Waibel, Barry Haddow, Rico Sennrich, and Philip Williams. 2021. ELITR multilingual live subtitling: Demo and strategy. In Proceedings of the 16th Conference of the European Chapter of the Association" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 526, + 772 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 313, + 72, + 526, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 526, + 105 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 526, + 105 + ], + "type": "text", + "content": "for Computational Linguistics: System Demonstrations, pages 271-277, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 113, + 526, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 113, + 526, + 212 + ], + "spans": [ + { + "bbox": [ + 304, + 113, + 526, + 212 + ], + "type": "text", + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 219, + 526, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 219, + 526, + 319 + ], + "spans": [ + { + "bbox": [ + 304, + 219, + 526, + 319 + ], + "type": "text", + "content": "Thierry Etchegoyhen, Lindsay Bywood, Mark Fishel, Panayota Georgakopoulou, Jie Jiang, Gerard van Loenhout, Arantza del Pozo, Mirjam Sepesy Maučec, Anja Turner, and Martin Volk. 2014. Machine translation for subtitling: A large-scale evaluation. In Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC'14), pages 46-53, Reykjavik, Iceland. European Language Resources Association (ELRA)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 326, + 525, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 326, + 525, + 393 + ], + "spans": [ + { + "bbox": [ + 304, + 326, + 525, + 393 + ], + "type": "text", + "content": "Alina Karakanta, François Buet, Mauro Cettolo, and François Yvon. 2022. Evaluating subtitle segmentation for end-to-end generation systems. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 3069-3078, Marseille, France. European Language Resources Association." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 400, + 525, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 400, + 525, + 466 + ], + "spans": [ + { + "bbox": [ + 304, + 400, + 525, + 466 + ], + "type": "text", + "content": "Alina Karakanta, Matteo Negri, and Marco Turchi. 2020a. Is 42 the answer to everything in subtitling-oriented speech translation? In Proceedings of the 17th International Conference on Spoken Language Translation, pages 209-219, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 474, + 526, + 539 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 526, + 539 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 526, + 539 + ], + "type": "text", + "content": "Alina Karakanta, Matteo Negri, and Marco Turchi. 2020b. MuST-cinema: a speech-to-subtitles corpus. In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 3727-3734, Marseille, France. European Language Resources Association." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 547, + 525, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 547, + 525, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 547, + 525, + 624 + ], + "type": "text", + "content": "Pierre Lison, Jörg Tiedemann, and Milen Kouylekov. 2018. OpenSubtitles2018: Statistical rescoring of sentence alignments in large, noisy parallel corpora. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources Association (ELRA)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 632, + 525, + 698 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 632, + 525, + 698 + ], + "spans": [ + { + "bbox": [ + 304, + 632, + 525, + 698 + ], + "type": "text", + "content": "Evgeny Matusov, Patrick Wilken, and Yota Georgakopoulou. 2019. Customizing neural machine translation for subtitling. In Proceedings of the Fourth Conference on Machine Translation (Volume 1: Research Papers), pages 82-93, Florence, Italy. Association for Computational Linguistics." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 706, + 525, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 706, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 706, + 525, + 772 + ], + "type": "text", + "content": "Sara Papi, Alina Karakanta, Matteo Negri, and Marco Turchi. 2022. Dodging the data bottleneck: Automatic subtitling with automatically segmented ST corpora. In Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "776" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 79, + 72, + 291, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 72, + 291, + 106 + ], + "spans": [ + { + "bbox": [ + 79, + 72, + 291, + 106 + ], + "type": "text", + "content": "Conference on Natural Language Processing (Volume 2: Short Papers), pages 480-487, Online only. Association for Computational Linguistics." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 111, + 291, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 111, + 291, + 191 + ], + "spans": [ + { + "bbox": [ + 69, + 111, + 291, + 191 + ], + "type": "text", + "content": "Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311-318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 196, + 291, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 196, + 291, + 242 + ], + "spans": [ + { + "bbox": [ + 69, + 196, + 291, + 242 + ], + "type": "text", + "content": "Elisa Perego, Fabio Del Missier, Marco Porta, and Mauro Mosconi. 2010. The cognitive effectiveness of subtitle processing. *Media psychology*, 13(3):243-272." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 248, + 291, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 248, + 291, + 303 + ], + "spans": [ + { + "bbox": [ + 69, + 248, + 291, + 303 + ], + "type": "text", + "content": "Dhevi J Rajendran, Andrew T Duchowski, Pilar Orero, Juan Martínez, and Pablo Romero-Fresco. 2013. Effects of text chunking on subtitling: A quantitative and qualitative examination. Perspectives, 21(1):5-21." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 309, + 291, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 309, + 291, + 376 + ], + "spans": [ + { + "bbox": [ + 69, + 309, + 291, + 376 + ], + "type": "text", + "content": "Julian Salazar, Davis Liang, Toan Q. Nguyen, and Katrin Kirchhoff. 2020. Masked language model scoring. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2699-2712, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 382, + 291, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 382, + 291, + 428 + ], + "spans": [ + { + "bbox": [ + 69, + 382, + 291, + 428 + ], + "type": "text", + "content": "Anke Tardel. 2020. Effort in semi-automatized subtitling processes: speech recognition and experience during transcription. Journal of Audiovisual Translation, 3(2):79-102." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 433, + 290, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 433, + 290, + 489 + ], + "spans": [ + { + "bbox": [ + 69, + 433, + 290, + 489 + ], + "type": "text", + "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 495, + 291, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 495, + 291, + 584 + ], + "spans": [ + { + "bbox": [ + 69, + 495, + 291, + 584 + ], + "type": "text", + "content": "Martin Volk, Rico Sennrich, Christian Hardmeier, and Frida Tidström. 2010. Machine translation of TV subtitles for large scale production. In Proceedings of the Second Joint EM+/CNGL Workshop: Bringing MT to the User: Research on Integrating MT in the Translation Industry, pages 53-62, Denver, Colorado, USA. Association for Machine Translation in the Americas." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 68, + 590, + 215, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 590, + 215, + 604 + ], + "spans": [ + { + "bbox": [ + 68, + 590, + 215, + 604 + ], + "type": "text", + "content": "A Segmentation Examples" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 611, + 291, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 611, + 291, + 678 + ], + "spans": [ + { + "bbox": [ + 67, + 611, + 291, + 678 + ], + "type": "text", + "content": "Table 4 provides examples of subtitles in the MuST-Cinema test sets segmented with either the character counting baseline or our LM-based approach, in its insertion variant without resorting to overt punctuation clueing." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 679, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 679, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 679, + 291, + 772 + ], + "type": "text", + "content": "In these examples, the MLM approach generates end-of-line and end-of-subtitle breaks that are overall in line with natural linguistic breaks, contrary to the character counting baseline. As such, on either short, medium or longer input, the readability of the generated subtitles is significantly enhanced with our approach." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 303, + 70, + 415, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 70, + 415, + 84 + ], + "spans": [ + { + "bbox": [ + 303, + 70, + 415, + 84 + ], + "type": "text", + "content": "B Extended Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 92, + 526, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 92, + 526, + 280 + ], + "spans": [ + { + "bbox": [ + 302, + 92, + 526, + 280 + ], + "type": "text", + "content": "The results presented in Section 5 were limited to the subset of languages and metrics for which published comparative results were available on the MuST-Cinema datasets. In Table 5, we present the complete list of results obtained with our method, for all languages and metrics. The selected variant of our method is the insertion masking approach, which was selected for the main results in our paper, with a segmentation window starting at 15 characters and ending at 42. We do not include BLEU scores computed over text that includes segmentation breaks, as the results are identical to those obtained with the Sigma metric for our approach, which does not generate imperfect text." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 282, + 526, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 282, + 526, + 444 + ], + "spans": [ + { + "bbox": [ + 302, + 282, + 526, + 444 + ], + "type": "text", + "content": "Across languages, the results are relatively uniform, with the best Sigma scores obtained in English and the lowest in Dutch, for a difference of 4.1 points between the two languages. In terms of break coverage, the best results were obtained for Spanish and the worst for Romanian, although results were also relatively uniform across languages. In all cases, overt clueing, where overt punctuation marks raised the LM score by 1, improved Sigma scores, although it had less of an impact on break coverage results, where both variants performed similarly overall." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 454, + 524, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 454, + 524, + 468 + ], + "spans": [ + { + "bbox": [ + 302, + 454, + 524, + 468 + ], + "type": "text", + "content": "C Results With Different min Parameters" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 476, + 525, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 476, + 525, + 608 + ], + "spans": [ + { + "bbox": [ + 302, + 476, + 525, + 608 + ], + "type": "text", + "content": "As noted in Section 3, considering preliminary results over the development set we selected a default value of 15 for the min parameter, which indicates the number of characters after which the segmentation process applies. In Table 6, we present comparative results on the test sets with different min values. In terms of Sigma, values of 15 and 20 led to rather similar results; values of 1 and 10 resulted in slightly lower results, with the lowest results achieved with the former." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 611, + 526, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 611, + 526, + 706 + ], + "spans": [ + { + "bbox": [ + 302, + 611, + 526, + 706 + ], + "type": "text", + "content": "In terms of " + }, + { + "bbox": [ + 302, + 611, + 526, + 706 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{eol} \\rangle" + }, + { + "bbox": [ + 302, + 611, + 526, + 706 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 302, + 611, + 526, + 706 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{eob} \\rangle" + }, + { + "bbox": [ + 302, + 611, + 526, + 706 + ], + "type": "text", + "content": " coverage, the former increases with larger min values, which is expected given the more restricted space to insert these end-of-line markers as the value increases; for " + }, + { + "bbox": [ + 302, + 611, + 526, + 706 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{eob} \\rangle" + }, + { + "bbox": [ + 302, + 611, + 526, + 706 + ], + "type": "text", + "content": ", the restricted insertion space results in increased under-generation, which in turn results in better scores for lower values of the min parameter." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "text", + "content": "777" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 113, + 526, + 303 + ], + "blocks": [ + { + "bbox": [ + 69, + 113, + 526, + 303 + ], + "lines": [ + { + "bbox": [ + 69, + 113, + 526, + 303 + ], + "spans": [ + { + "bbox": [ + 69, + 113, + 526, + 303 + ], + "type": "table", + "html": "
CountCharsMLM
They're things you access through your <eol> computer. <eob>They're things you access <eol> through your computer. <eob>
Every row of data is a life whose story <eol> deserves to be told with dignity. <eob>Every row of data is a life <eol> whose story deserves to be told <eob> with dignity. <eob>
During the winter, struggling to get <eol> warm, my neighbors would have no choice <eob> but to bypass the meter after their heat <eol> was shut off, just to keep their family <eob> comfortable for one more day. <eob>During the winter, struggling to get warm, <eol> my neighbors would have no choice <eob> but to bypass the meter <eol> after their heat was shut off, <eob> just to keep their family comfortable <eol> for one more day. <eob>
", + "image_path": "b6a7d0fb463c965476b7e4eb46ca26d6a8a8f8078ca9a5f77957f8f62a656370.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 141, + 418, + 453, + 703 + ], + "blocks": [ + { + "bbox": [ + 101, + 310, + 490, + 322 + ], + "lines": [ + { + "bbox": [ + 101, + 310, + 490, + 322 + ], + "spans": [ + { + "bbox": [ + 101, + 310, + 490, + 322 + ], + "type": "text", + "content": "Table 4: Examples of subtitles segmented via character counting and MLM-based mask insertion" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 141, + 418, + 453, + 703 + ], + "lines": [ + { + "bbox": [ + 141, + 418, + 453, + 703 + ], + "spans": [ + { + "bbox": [ + 141, + 418, + 453, + 703 + ], + "type": "table", + "html": "
LanguageMethodBLEUSigmaEOLEOBCPL
DEMLM10070.8518.53-7.96100%
MLM+OC10075.6319.81-7.78100%
ENMLM10076.7719.18-9.91100%
MLM+OC10077.8919.86-9.73100%
ESMLM10073.4712.98-4.91100%
MLM+OC10075.5913.45-4.63100%
FRMLM10073.7816.51-6.58100%
MLM+OC10076.0717.47-6.12100%
ITMLM10071.3818.49-9.55100%
MLM+OC10074.2020.34-8.57100%
NLMLM10068.7120.37-9.96100%
MLM+OC10073.8821.22-9.96100%
PTMLM10071.5920.03-10.81100%
MLM+OC10075.5019.87-10.02100%
ROMLM10069.4523.37-10.44100%
MLM+OC10074.1323.37-10.09100%
", + "image_path": "c0115864e1110fdf64a9df994c03b8e27f7acb000e1aa76ebbaa90140b68bbbc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 712, + 485, + 724 + ], + "lines": [ + { + "bbox": [ + 107, + 712, + 485, + 724 + ], + "spans": [ + { + "bbox": [ + 107, + 712, + 485, + 724 + ], + "type": "text", + "content": "Table 5: Complete results with MLM mask insertion on the MuST-Cinema test sets (min=15)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "778" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 180, + 158, + 413, + 661 + ], + "blocks": [ + { + "bbox": [ + 180, + 158, + 413, + 661 + ], + "lines": [ + { + "bbox": [ + 180, + 158, + 413, + 661 + ], + "spans": [ + { + "bbox": [ + 180, + 158, + 413, + 661 + ], + "type": "table", + "html": "
LanguageminBLEUSigmaEOLEOB
DE110072.3128.75-0.18
1010073.9622.68-4.43
1510075.6319.81-7.78
2010075.2814.54-11.21
EN110074.3037.33-0.98
1010077.1424.49-7.77
1510077.8919.86-9.73
2010077.1615.24-12.68
ES110073.0020.870.28
1010074.3218.24-2.04
1510075.5913.45-4.63
2010075.838.66-7.87
FR110073.8924.68-0.73
1010075.2620.83-3.93
1510076.0717.47-6.12
2010076.7512.5-10.05
IT110072.0129.75-3.66
1010073.7524.71-6.61
1510074.2020.34-8.57
2010073.6614.62-11.61
NL110072.1626.83-5.47
1010073.5623.26-8.47
1510073.8821.22-9.96
2010074.4016.81-12.43
PT110072.8726.38-6.24
1010074.5322.15-8.08
1510075.5019.87-10.02
2010074.9814.17-13.36
RO110072.0532.3-4.51
1010073.7626.98-7.52
1510074.1323.37-10.09
2010074.8917.53-12.83
", + "image_path": "30ea88503d9dc83caefc2d1e1026152ad155e5b098d62964a7a68aa6b88b3fd9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 108, + 667, + 485, + 681 + ], + "lines": [ + { + "bbox": [ + 108, + 667, + 485, + 681 + ], + "spans": [ + { + "bbox": [ + 108, + 667, + 485, + 681 + ], + "type": "text", + "content": "Table 6: Test set results with the MLM+OC method and different values of the min parameter" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "779" + } + ] + } + ], + "index": 2 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "spans": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "type": "text", + "content": "A For every submission:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 106, + 414, + 243 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "spans": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "type": "text", + "content": "A1. Did you describe the limitations of your work? 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 143, + 329, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 143, + 329, + 169 + ], + "spans": [ + { + "bbox": [ + 77, + 143, + 329, + 169 + ], + "type": "text", + "content": "A2. Did you discuss any potential risks of your work? 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 178, + 414, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 178, + 414, + 205 + ], + "spans": [ + { + "bbox": [ + 77, + 178, + 414, + 205 + ], + "type": "text", + "content": "A3. Do the abstract and introduction summarize the paper's main claims?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 77, + 215, + 398, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 215, + 398, + 243 + ], + "spans": [ + { + "bbox": [ + 77, + 215, + 398, + 243 + ], + "type": "text", + "content": "A4. Have you used AI writing assistants when working on this paper? Left blank." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 68, + 253, + 292, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 253, + 292, + 266 + ], + "spans": [ + { + "bbox": [ + 68, + 253, + 292, + 266 + ], + "type": "text", + "content": "B Did you use or create scientific artifacts?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 78, + 270, + 198, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 270, + 198, + 283 + ], + "spans": [ + { + "bbox": [ + 78, + 270, + 198, + 283 + ], + "type": "text", + "content": "Not applicable. Left blank." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 76, + 292, + 524, + 634 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 292, + 315, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 292, + 315, + 319 + ], + "spans": [ + { + "bbox": [ + 76, + 292, + 315, + 319 + ], + "type": "text", + "content": "B1. Did you cite the creators of artifacts you used? Not applicable. Left blank." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 76, + 328, + 463, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 328, + 463, + 355 + ], + "spans": [ + { + "bbox": [ + 76, + 328, + 463, + 355 + ], + "type": "text", + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Not applicable. Left blank." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 76, + 364, + 524, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 364, + 524, + 432 + ], + "spans": [ + { + "bbox": [ + 76, + 364, + 524, + 432 + ], + "type": "text", + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Not applicable. Left blank." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 76, + 441, + 524, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 441, + 524, + 495 + ], + "spans": [ + { + "bbox": [ + 76, + 441, + 524, + 495 + ], + "type": "text", + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Not applicable. Left blank." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 76, + 504, + 524, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 504, + 524, + 544 + ], + "spans": [ + { + "bbox": [ + 76, + 504, + 524, + 544 + ], + "type": "text", + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Not applicable. Left blank." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 76, + 554, + 524, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 554, + 524, + 634 + ], + "spans": [ + { + "bbox": [ + 76, + 554, + 524, + 634 + ], + "type": "text", + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Not applicable. Left blank." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "spans": [ + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "type": "text", + "content": "C Did you run computational experiments?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 80, + 662, + 87, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 662, + 87, + 672 + ], + "spans": [ + { + "bbox": [ + 80, + 662, + 87, + 672 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 77, + 684, + 524, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 684, + 524, + 710 + ], + "spans": [ + { + "bbox": [ + 77, + 684, + 524, + 710 + ], + "type": "text", + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 89, + 711, + 452, + 724 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 711, + 452, + 724 + ], + "spans": [ + { + "bbox": [ + 89, + 711, + 452, + 724 + ], + "type": "text", + "content": "We didn't trained any models for this paper, and inference was performed on CPU." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "spans": [ + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "type": "text", + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "spans": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "text", + "content": "ACL 2023 Responsible NLP Checklist" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "780" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 77, + 70, + 523, + 235 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 77, + 70, + 523, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 70, + 523, + 110 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 523, + 110 + ], + "type": "text", + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 77, + 120, + 523, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 120, + 523, + 172 + ], + "spans": [ + { + "bbox": [ + 77, + 120, + 523, + 172 + ], + "type": "text", + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 183, + 523, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 183, + 523, + 235 + ], + "spans": [ + { + "bbox": [ + 77, + 183, + 523, + 235 + ], + "type": "text", + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 67, + 247, + 522, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 247, + 522, + 260 + ], + "spans": [ + { + "bbox": [ + 67, + 247, + 522, + 260 + ], + "type": "text", + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 79, + 264, + 127, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 264, + 127, + 277 + ], + "spans": [ + { + "bbox": [ + 79, + 264, + 127, + 277 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 76, + 286, + 523, + 538 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 76, + 286, + 523, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 286, + 523, + 324 + ], + "spans": [ + { + "bbox": [ + 76, + 286, + 523, + 324 + ], + "type": "text", + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Not applicable. Left blank." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 76, + 336, + 523, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 336, + 523, + 387 + ], + "spans": [ + { + "bbox": [ + 76, + 336, + 523, + 387 + ], + "type": "text", + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Not applicable. Left blank." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 76, + 399, + 523, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 399, + 523, + 451 + ], + "spans": [ + { + "bbox": [ + 76, + 399, + 523, + 451 + ], + "type": "text", + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Not applicable. Left blank." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 76, + 462, + 519, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 462, + 519, + 488 + ], + "spans": [ + { + "bbox": [ + 76, + 462, + 519, + 488 + ], + "type": "text", + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Not applicable. Left blank." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 76, + 498, + 523, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 498, + 523, + 538 + ], + "spans": [ + { + "bbox": [ + 76, + 498, + 523, + 538 + ], + "type": "text", + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Not applicable. Left blank." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 305, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 305, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 305, + 791 + ], + "type": "text", + "content": "781" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Using contradictions improves question answering systems/56951eb2-a64c-4213-acb6-6239f47120c0_content_list.json b/2023/Using contradictions improves question answering systems/56951eb2-a64c-4213-acb6-6239f47120c0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7a173f8029b6f9f2cd7d399c89ee9399add75e79 --- /dev/null +++ b/2023/Using contradictions improves question answering systems/56951eb2-a64c-4213-acb6-6239f47120c0_content_list.json @@ -0,0 +1,1841 @@ +[ + { + "type": "text", + "text": "Using contradictions improves question answering systems", + "text_level": 1, + "bbox": [ + 194, + 90, + 803, + 111 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Étienne Fortier-Dubois", + "text_level": 1, + "bbox": [ + 233, + 141, + 438, + 159 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Domenic Rosati", + "text_level": 1, + "bbox": [ + 596, + 143, + 737, + 158 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "scite.ai", + "bbox": [ + 636, + 162, + 695, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dalhousie University", + "bbox": [ + 579, + 178, + 752, + 193 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 252, + 339, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This work examines the use of contradiction in natural language inference (NLI) for question answering (QA). Typically, NLI systems help answer questions by determining if a potential answer is entailed (supported) by some background context. But is it useful to also determine if an answer contradicts the context? We test this in two settings, multiple choice and extractive QA, and find that systems that incorporate contradiction can do slightly better than entailment-only systems on certain datasets. However, the best performances come from using contradiction, entailment, and QA model confidence scores together. This has implications for the deployment of QA systems in domains such as medicine and science where safety is an issue.", + "bbox": [ + 141, + 275, + 460, + 517 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 114, + 524, + 260, + 539 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Safety in NLP systems is unresolved, particularly in biomedical and scientific contexts where hallucination, overconfidence, and other problems are major obstacles to deployment (Ji et al., 2022; Kell et al., 2021). One active area of research to solve these issues is natural language inference (NLI) (Li et al., 2022). NLI is the task of determining whether a hypothesis is true (entailed), false (contradicted), or undetermined (neutral) given some premise.", + "bbox": [ + 112, + 549, + 489, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current NLI systems typically focus only on entailment to verify hypotheses—they calculate the degree to which a hypothesis is supported by the premise. But the premise can provide another signal: contradiction. Regardless of how well a hypothesis is entailed by the context, it can also be more or less contradicted, which could affect whether it is accepted or rejected. Contradictions are an important signal indicating whether some statement might be unacceptable given a premise. In some cases where we might not know if a statement is supported, we should still ensure we are rejecting statements that are outright contradicted.", + "bbox": [ + 112, + 709, + 489, + 919 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0782e5fe3e33d197559d16540fd53b83aff03fb7da05f392f5edf22449885b76.jpg", + "image_caption": [ + "Figure 1: A QA model is used to produce answers which are reformulated as hypotheses to determine if they are entailed or contradicted by a premise. The answers are ranked by NLI class scores to select the best answer." + ], + "image_footnote": [], + "bbox": [ + 510, + 249, + 884, + 397 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We wondered if adding this signal to a question answering (QA) system might improve performance and safety. To this end, we propose a method that reformulates answers from the QA system as hypotheses for NLI, calculates the entailment, contradiction, and neutrality of each hypothesis, and then selects the best one based on a combination of these results (Figure 1). We show that across 16 QA datasets (9 multiple choice and 7 extractive), the best approach is to use entailment, contradiction, and confidence scores together. Using only contradiction is roughly on par with, and sometimes better than, using only entailment.", + "bbox": [ + 507, + 493, + 884, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1.1 Related work", + "text_level": 1, + "bbox": [ + 509, + 717, + 662, + 732 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "NLI for question answering has been explored by several authors in various settings; see Paramasi-vam and Nirmala (2021) for an overview.", + "bbox": [ + 507, + 741, + 882, + 788 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One of these settings is selective question answering for extractive QA, where selective refers to abstention when the system is not confident enough in its answer (Kamath et al., 2020). Chen et al. (2021) have found that NLI systems are able to verify the predictions made by a QA system in this setting, but their result is limited to only selecting a top $k\\%$ of answers. Moreover, they", + "bbox": [ + 507, + 790, + 884, + 919 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "827", + "bbox": [ + 485, + 927, + 515, + 940 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics", + "bbox": [ + 226, + 945, + 771, + 957 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Volume 2: Short Papers, pages 827-840", + "bbox": [ + 376, + 958, + 621, + 971 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "July 9-14, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 295, + 972, + 700, + 984 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "do not provide an approach for improving overall performance, nor do they show the effect of incorporating contradiction directly (but do so indirectly by analyzing non-entailed passages).", + "bbox": [ + 112, + 84, + 487, + 148 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the related setting of multiple choice QA and fact checking, Mishra et al. (2021) have explored the use of entailment, finding that NLI models do well at these tasks by themselves, but can perform even better when they are adapted to in-domain data and longer premises. Yet their method uses only a two-class NLI set up (entailed or not entailed), which doesn't tell us much about directly using the contradiction signal. Pujari and Goldwasser (2019) does incorporate the contradiction signal showing the power of contradiction to improve machine comprehension but does not analyze its effects separately from entailment.", + "bbox": [ + 115, + 149, + 489, + 357 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Other QA settings in which NLI has been used include open domain (Harabagiu and Hickl, 2006) and multi-hop (Trivedi et al., 2019). Thus far, approaches tend to focus on entailment. To our knowledge, our work is the first to directly assess using contradictions for QA isolated from entailment.", + "bbox": [ + 112, + 359, + 487, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Outside of question answering, a domain that uses contradictions is factual consistency—the task of ensuring that a collection of utterances is faithful to a source document. Li et al. (2022) provide an overview. Typically, entailment is still the main focus, but Laban et al. (2022) propose an NLI-based method to ensure the consistency of a summary with a source document using contradiction and neutral scores in addition to entailment, beating out previous systems.", + "bbox": [ + 112, + 457, + 487, + 617 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Other researchers have used contradictions to identify consistency errors across Wikipedia (Schuster et al., 2022; Hsu et al., 2021) or generate credible character dialogue (Nie et al., 2021; Song et al., 2020).", + "bbox": [ + 112, + 619, + 487, + 697 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Methods", + "text_level": 1, + "bbox": [ + 112, + 713, + 223, + 728 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We tested the effect of contradictions in two QA settings and a total of sixteen question-answer datasets. Our approach is broadly similar to both Chen et al. (2021) and Mishra et al. (2021) in that we use most of the same datasets for evaluating NLI reranking for multiple choice QA and extractive QA. Unlike both, we incorporate contradiction directly as a signal for reranking answers.", + "bbox": [ + 112, + 741, + 487, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Briefly, for each dataset, we used pretrained QA models to produce answers and confidence scores for the dataset's questions. We refer to the confi", + "bbox": [ + 112, + 871, + 487, + 917 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "dence scores below as QA. We then trained QA2D models (where QA2D stands for \"question-answer to declarative\") to turn the answers into the declarative hypothesis format required for NLI. For example, the question-answer pair \"What is the most abundant metal in the Earth crust? Copper.\" might be rephrased as \"The most abundant metal in the Earth crust is copper\" (see Appendix D for more details).", + "bbox": [ + 507, + 84, + 884, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the question contexts as premises, we then used NLI models to classify every premise-hypothesis pair into three classes, each with an associated score: entailed (E), contradicted (C), and neutral (N). After that, we trained logistic regression calibration models to find which linear combination of the four scores—QA, E, C, and N—was best able to pick the answers accurately.", + "bbox": [ + 507, + 230, + 884, + 357 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "When evaluating performance, we applied the selective QA approach from Kamath et al. (2020) to rank answers using combinations of the four scores, and then consider only those that the model was most confident in answering. We compared selecting the top $20\\%$ and $50\\%$ . In the multiple choice setting, it was also possible to rank all potential answers according to the four scores, unlike in the extractive QA setting where the QA model produced only one answer per question, so we evaluated performance with that approach as well (see appendix A for details).", + "bbox": [ + 507, + 360, + 884, + 552 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3 Experimental setting", + "text_level": 1, + "bbox": [ + 507, + 567, + 724, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the multiple choice setting, we tested 9 datasets. Two of them are in-domain, since the pretrained QA models we used were finetuned on them. Specifically, we used a RoBERTa large model (Liu et al., 2019) finetuned on the RACE dataset (Lai et al., 2017), as well as two DeBERTa v3 variants, base and xsmall (He et al., 2021a), finetuned on the SciQ dataset (Welbl et al., 2017).", + "bbox": [ + 507, + 596, + 882, + 722 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the extractive QA setting, we used 7 datasets: five from the MRQA 2019 task (Fisch et al., 2019), as well as SQuAD 2.0 (Rajpurkar et al., 2018) and SQuAD adversarial (Jia and Liang, 2017). The SQuAD model is the in-domain dataset: it was used to pretrain (Rajpurkar et al., 2016) the two QA models we used, DistillBERT (Sanh et al., 2020) and BERT-Large (Devlin et al., 2019). Like Chen et al. (2021), we used the Natural Questions dataset for calibration.", + "bbox": [ + 507, + 725, + 882, + 884 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In both settings, all datasets contain the relevant context that can be used by the QA models to select", + "bbox": [ + 507, + 887, + 882, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "828", + "bbox": [ + 485, + 928, + 515, + 939 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "answers. More detail on the datasets and QA models is available in appendices B and C respectively.", + "bbox": [ + 112, + 84, + 489, + 116 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "See appendices D, E, and F for details on the QA2D, NLI, and calibration models. Our models follow the setups described in Kamath et al. (2020), Chen et al. (2021), and Mishra et al. (2021). The main interesting detail is that the calibration models were trained on a holdout set of 100 samples from a single domain, using logistic regression, as in Chen et al. (2021).", + "bbox": [ + 112, + 117, + 489, + 244 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4 Results", + "text_level": 1, + "bbox": [ + 112, + 258, + 213, + 273 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1 Multiple choice setting", + "text_level": 1, + "bbox": [ + 112, + 284, + 339, + 300 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For most multiple choice datasets, the best accuracy—when ranking all potential answers—is attained when using a calibrated model combining QA confidence, entailment, and contradiction (QA+E+C in Table 1). Only for the in-domain case (RACE-C) does the uncalibrated RoBERTa-RACE model perform on par with that. Using QA scores combined with either entailment (QA+E) or contradiction (QA+C) achieves similar performance, with contradiction winning by a small margin: $84.33\\%$ average accuracy compared to $84.31\\%$ .", + "bbox": [ + 112, + 305, + 489, + 482 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To inspect these trends further, we performed a correlation analysis of the NLI classes and QA confidence scores with the correct answer (appendix G). We found that besides QA confidence, it is the contradiction score that has the strongest correlation with the correct answer. The analysis also showed that the neutral class score (N) had almost no effect, which is why it is omitted in all results.", + "bbox": [ + 112, + 483, + 489, + 611 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "When using the selective QA approach and evaluating only the $20\\%$ of $50\\%$ most confident answers, the best performance is attained with the $\\mathbf{QA} + \\mathbf{C}$ combination (Table 2). This model is the only one that beats just using the QA confidence score on average. It is stronger than $\\mathbf{QA} + \\mathbf{E} + \\mathbf{C}$ and $\\mathbf{QA} + \\mathbf{E}$ for both coverage percentages.", + "bbox": [ + 112, + 612, + 489, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Contradiction alone, without QA confidence scores (C), also beats both entailment alone (E) and entailment with contradiction $(\\mathbf{E} + \\mathbf{C})$ for both coverages. These results match our intuition that the less contradicted an answer, the more likely it is correct, even in cases where there is uncertainty about its entailment.", + "bbox": [ + 112, + 725, + 489, + 837 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.2 Extractive QA setting", + "text_level": 1, + "bbox": [ + 112, + 848, + 332, + 866 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similar results occur when evaluating the extractive QA datasets with $20\\%$ and $50\\%$ selective coverage (Table 3). The $\\mathbf{QA} + \\mathbf{C}$ model does better than QA", + "bbox": [ + 112, + 871, + 489, + 919 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "alone, and $\\mathbf{C}$ alone does better than $\\mathbf{E} + \\mathbf{C}$ or $\\mathbf{E}$ alone, indicating the importance of the contradiction signal here too. However, entailment seems to matter more for extractive QA, as the best F1 score overall was from $\\mathbf{QA} + \\mathbf{E}$ in the $20\\%$ coverage case, and $\\mathbf{QA} + \\mathbf{E} + \\mathbf{C}$ in the $50\\%$ case.", + "bbox": [ + 507, + 84, + 884, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "5 Discussion", + "text_level": 1, + "bbox": [ + 507, + 193, + 636, + 209 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Contradiction with background context is a useful signal that NLP systems can use to infer answers to questions. This is not necessarily a superior strategy to using entailment, but our results show that combining these two signals can improve performance beyond what QA models can achieve on their own. These results are interesting because using contradictions comes with potential benefits for the safety of NLP systems and, as a result, their deployment in domains such as medicine or science. Namely, that there are many potential cases where we are not sure if a statement is entailed directly by a background context but we may be sure that the statement is not refuted by a background context. In two-class NLI settings where we focus only on entailment, neutral and contradiction are collapsed together and we don't have this guarantee.", + "bbox": [ + 507, + 219, + 884, + 493 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "6 Limitations", + "text_level": 1, + "bbox": [ + 507, + 506, + 645, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our work comes with some limitations. It is uncertain whether our results in two specific settings, multiple choice and extractive QA, would extend to more general settings for NLI, although the use of contradictions for factual consistency by Laban et al. (2022) suggests that they could. Additionally, 3-class NLI is not sufficient to capture all the natural language relations that might be needed to verify an answer. As such more challenging datasets in other settings and more granular NLI settings should be attempted.", + "bbox": [ + 507, + 533, + 884, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Another limitation involves answer ranking and the associated computational cost. The main reason we did not test answer ranking in extractive QA is that we did not generate diverse outputs, but another reason is that such a procedure grows prohibitively expensive as the domain becomes more open. In a fully open domain, ranking would require a quadratic evaluation for each context passage against each reformulated answer candidate (Schuster et al., 2022). Future work should look at comparison approaches that amortize this cost, such as NLI-based dense passage retrieval (Reimers and Gurevych, 2019).", + "bbox": [ + 507, + 709, + 884, + 919 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "829", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/08e65eb7bd5e6d98700618a8fa6707566b577a331716d35e543b9e2ed6369cfb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
QA ModelCosmosDREAMMCSMCS2MCTQASCRACERACE-CSciQAverage
SciQ-base18.4643.8061.9963.7144.7693.4130.9727.3995.2853.30
SciQ-small25.4648.2660.2866.0459.7690.6035.5630.6298.0957.18
QA64.2282.5689.7086.9890.4898.1676.9369.8097.9684.08
QA+E+C64.72*83.19*90.06*87.59*91.43*98.6077.53*69.80*98.2184.57
QA+E64.3282.85*89.92*87.29*91.0798.49*77.1869.6698.0984.31
QA+C64.8282.75*89.88*87.29*90.8398.3877.1669.8098.0984.33
", + "bbox": [ + 127, + 92, + 868, + 180 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/da55983854c26153070ce30967fc533987e904bda5a2d352b0f691d187b60e8a.jpg", + "table_caption": [ + "Table 1: Multiple choice setting. Accuracy scores (best per column in bold, second best underlined, statistical significance (pairwise students t-test) is indicated by asterix) after answer ranking with the mnli-large NLI model. The top three rows show the accuracy of using only the QA models' confidence score; \"QA\" refers to the scores of the RoBERTa-RACE model, which was used for calibration. The bottom rows add the entailment and/or contradiction scores to the RoBERTa-RACE score. For other NLI models, and for just E, C, and $\\mathrm{E + C}$ without calibration with RoBERTa-RACE, see Table 8 in the appendix." + ], + "table_footnote": [], + "table_body": "
DatasetQA+E+CQA+CQA+EE+CECQA
20%CosmosQA77.5591.1276.8869.1868.3483.2588.61
DREAM98.2898.7798.2896.3296.3296.8198.28
MCScript99.8299.4699.8299.6499.6499.4699.82
MCScript-2.099.5899.7299.4599.1799.0397.3799.58
MCTest10099.4010010010099.4098.81
QASC100100100100100100100
RACE94.9396.6994.7292.4492.2490.1798.24
RACE-C88.7392.9689.4485.2185.9286.6293.66
SciQ100100100100100100100
Average95.4397.5795.4093.5593.5094.7997.45
50%CosmosQA80.2981.7076.9475.8070.6480.6376.47
DREAM95.1096.8694.9093.6393.6393.6396.67
MCScript98.5798.6498.2898.0097.9397.1498.78
MCScript-2.096.4098.2395.8494.6894.4096.0198.01
MCTest99.5299.7699.5299.0599.0599.7699.52
QASC10010010099.7899.7899.78100
RACE90.1192.6889.9987.7187.3885.2393.88
RACE-C85.1184.8385.3978.3778.3777.2587.36
SciQ10010010010010099.74100
Average93.9094.7493.4391.8991.2492.1394.52
", + "bbox": [ + 211, + 305, + 786, + 560 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/50157353668011eb304df5c595fe8b61a7f01acc31185099d156a6efa8867692.jpg", + "table_caption": [ + "Table 2: Multiple choice setting. Accuracy scores (best per row in bold, second best underlined) for selective QA with $20\\%$ and $50\\%$ coverage of the dataset. Calibrations and QA confidence are all from RoBERTa-RACE, where RACE is the in-domain dataset." + ], + "table_footnote": [], + "table_body": "
DatasetQA+E+CQA+CQA+EE+CECQA
20%BioASQ85.0483.1085.0674.2274.2275.4782.99
HotpotQA86.6285.8986.6980.6080.6079.8285.33
Natural Questions91.8492.1891.6879.8979.8782.0990.98
SQuAD98.2698.7692.3798.1792.4890.8899.04
SQuAD-adv43.9943.5743.9843.7443.6042.8139.83
SQuAD237.6436.0737.5637.4337.3137.6830.52
TriviaQA81.3380.3681.2165.5365.2569.1380.68
Average74.9674.1974.9967.6867.6268.2772.77
50%BioASQ76.1375.5176.0471.4971.4972.9775.49
HotpotQA79.3778.9579.3077.4377.4377.3178.74
Natural Questions84.5383.2484.4874.9674.9378.6282.47
SQuAD96.9897.0196.9791.5891.5291.1997.00
SQuAD-adv41.8041.4941.1642.7642.7942.0340.26
SQuAD229.4128.7728.4534.4334.1434.3926.18
TriviaQA74.3074.2374.3765.0564.9368.0874.21
Average68.9368.4668.6865.3965.3266.3767.76
", + "bbox": [ + 201, + 643, + 796, + 850 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 3: Extractive QA setting. F1 scores (best per row in bold, second best underlined) for selective QA with $20\\%$ and $50\\%$ coverage of the dataset. Calibrations and QA confidence are from the BERT-large model, where SQuAD is the in-domain dataset. For similar results on the smaller DistillBERT model, see Table 10 in the appendix.", + "bbox": [ + 112, + 860, + 882, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "830", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 115, + 84, + 213, + 98 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Jifan Chen, Eunsol Choi, and Greg Durrett. 2021. Can NLI Models Verify QA Systems' Predictions? In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 3841-3854, Punta Cana, Dominican Republic. Association for Computational Linguistics.", + "Dorottya Demszky, Kelvin Guu, and Percy Liang. 2018. Transforming Question Answering Datasets Into Natural Language Inference Datasets. Technical Report arXiv:1809.02922, arXiv. ArXiv:1809.02922 [cs] type: article.", + "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics.", + "Adam Fisch, Alon Talmor, Robin Jia, Minjoon Seo, Eunsol Choi, and Danqi Chen. 2019. MRQA 2019 Shared Task: Evaluating Generalization in Reading Comprehension. In Proceedings of the 2nd Workshop on Machine Reading for Question Answering, pages 1-13, Hong Kong, China. Association for Computational Linguistics.", + "Sanda Harabagiu and Andrew Hickl. 2006. Methods for Using Textual Entailment in Open-Domain Question Answering. In Proceedings of the 21st International Conference on Computational Linguistics and 44th Annual Meeting of the Association for Computational Linguistics, pages 905-912, Sydney, Australia. Association for Computational Linguistics.", + "Pengcheng He, Jianfeng Gao, and Weizhu Chen. 2021a. DeBERTaV3: Improving DeBERTa using ELECTRA-Style Pre-Training with Gradient-Disentangled Embedding Sharing. Number: arXiv:2111.09543 arXiv:2111.09543 [cs].", + "Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. 2021b. DeBERTa: Decoding-enhanced BERT with Disentangled Attention. Number: arXiv:2006.03654 arXiv:2006.03654 [cs].", + "Cheng Hsu, Cheng-Te Li, Diego Saez-Trumper, and Yi-Zhan Hsu. 2021. WikiContradiction: Detecting Self-Contradiction Articles on Wikipedia. Technical Report arXiv:2111.08543, arXiv. ArXiv:2111.08543 [cs] type: article.", + "Lifu Huang, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. 2019. Cosmos QA: Machine Reading Comprehension with Contextual Commonsense Reasoning. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages" + ], + "bbox": [ + 115, + 107, + 487, + 917 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "2391-2401, Hong Kong, China. Association for Computational Linguistics.", + "Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Yejin Bang, Andrea Madotto, and Pascale Fung. 2022. Survey of Hallucination in Natural Language Generation. Number: arXiv:2202.03629 arXiv:2202.03629 [cs].", + "Robin Jia and Percy Liang. 2017. Adversarial Examples for Evaluating Reading Comprehension Systems. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2021-2031, Copenhagen, Denmark. Association for Computational Linguistics.", + "Amita Kamath, Robin Jia, and Percy Liang. 2020. Selective Question Answering under Domain Shift. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5684-5696, Online. Association for Computational Linguistics.", + "Gregory Kell, Iain Marshall, Byron Wallace, and Andre Jaun. 2021. What Would it Take to get Biomedical QA Systems into Practice? In Proceedings of the 3rd Workshop on Machine Reading for Question Answering, pages 28-41, Punta Cana, Dominican Republic. Association for Computational Linguistics.", + "Tushar Khot, Peter Clark, Michal Guerquin, Peter Jansen, and Ashish Sabharwal. 2020. QASC: A Dataset for Question Answering via Sentence Composition. Proceedings of the AAAI Conference on Artificial Intelligence, 34(05):8082-8090. Number: 05.", + "Philippe Laban, Tobias Schnabel, Paul N. Bennett, and Marti A. Hearst. 2022. SummaC: Re-Visiting NLIBased Models for Inconsistency Detection in Summarization. Transactions of the Association for Computational Linguistics, 10:163-177.", + "Guokun Lai, Qizhe Xie, Hanxiao Liu, Yiming Yang, and Eduard Hovy. 2017. RACE: Large-scale ReAding Comprehension Dataset From Examinations. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 785-794, Copenhagen, Denmark. Association for Computational Linguistics.", + "Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. 2019. Albert: A lite bert for self-supervised learning of language representations.", + "Wei Li, Wenhao Wu, Moye Chen, Jiachen Liu, Xinyan Xiao, and Hua Wu. 2022. Faithfulness in Natural Language Generation: A Systematic Survey of Analysis, Evaluation and Optimization Methods. Number: arXiv:2203.05227 arXiv:2203.05227 [cs].", + "Yichan Liang, Jianheng Li, and Jian Yin. 2019. A New Multi-choice Reading Comprehension Dataset for Curriculum Learning. In Proceedings of The Eleventh Asian Conference on Machine Learning, pages 742-757. PMLR. ISSN: 2640-3498." + ], + "bbox": [ + 510, + 85, + 882, + 917 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "831", + "bbox": [ + 485, + 928, + 512, + 940 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. RoBERTa: A Robustly Optimized BERT Pretraining Approach. Number: arXiv:1907.11692 arXiv:1907.11692 [cs].", + "Anshuman Mishra, Dhruvesh Patel, Aparna Vijayakumar, Xiang Lorraine Li, Pavan Kapanipathi, and Kartik Talamadupula. 2021. Looking Beyond Sentence-Level Natural Language Inference for Question Answering and Text Summarization. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1322-1336, Online. Association for Computational Linguistics.", + "Yixin Nie, Adina Williams, Emily Dinan, Mohit Bansal, Jason Weston, and Douwe Kiela. 2020. Adversarial NLI: A New Benchmark for Natural Language Understanding. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4885-4901, Online. Association for Computational Linguistics.", + "Yixin Nie, Mary Williamson, Mohit Bansal, Douwe Kiela, and Jason Weston. 2021. I like fish, especially dolphins: Addressing Contradictions in Dialogue Modeling. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1699-1713, Online. Association for Computational Linguistics.", + "Simon Ostermann, Ashutosh Modi, Michael Roth, Stefan Thater, and Manfred Pinkal. 2018. MCScript: A Novel Dataset for Assessing Machine Comprehension Using Script Knowledge. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources Association (ELRA).", + "Simon Ostermann, Michael Roth, and Manfred Pinkal. 2019. MCScript2.0: A Machine Comprehension Corpus Focused on Script Events and Participants. In Proceedings of the Eighth Joint Conference on Lexical and Computational Semantics (*SEM* 2019), pages 103-117, Minneapolis, Minnesota. Association for Computational Linguistics.", + "Aarthi Paramasivam and S. Jaya Nirmala. 2021. A survey on textual entailment based question answering. Journal of King Saud University - Computer and Information Sciences.", + "Rajkumar Pujari and Dan Goldwasser. 2019. Using natural language relations between answer choices for machine comprehension. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4010-4015, Minneapolis, Minnesota. Association for Computational Linguistics." + ], + "bbox": [ + 117, + 85, + 485, + 917 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer. Journal of Machine Learning Research, 21(140):1-67.", + "Pranav Rajpurkar, Robin Jia, and Percy Liang. 2018. Know What You Don't Know: Unanswerable Questions for SQuAD. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 784-789, Melbourne, Australia. Association for Computational Linguistics.", + "Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD: 100,000+ Questions for Machine Comprehension of Text. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2383-2392, Austin, Texas. Association for Computational Linguistics.", + "Nils Reimers and Iryna Gurevych. 2019. SentenceBERT: Sentence Embeddings using Siamese BERT-Networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3982-3992, Hong Kong, China. Association for Computational Linguistics.", + "Matthew Richardson, Christopher J.C. Burges, and Erin Renshaw. 2013. MCTest: A Challenge Dataset for the Open-Domain Machine Comprehension of Text. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 193-203, Seattle, Washington, USA. Association for Computational Linguistics.", + "Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. 2020. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. Number: arXiv:1910.01108 arXiv:1910.01108 [cs].", + "Tal Schuster, Sihao Chen, Senaka Buthpitiya, Alex Fabrikant, and Donald Metzler. 2022. Stretching Sentence-pair NLI Models to Reason over Long Documents and Clusters. Number: arXiv:2204.07447 arXiv:2204.07447 [cs].", + "Haoyu Song, Wei-Nan Zhang, Jingwen Hu, and Ting Liu. 2020. Generating Persona Consistent Dialogues by Exploiting Natural Language Inference. Proceedings of the AAAI Conference on Artificial Intelligence, 34(05):8878-8885. Number: 05.", + "Kai Sun, Dian Yu, Jianshu Chen, Dong Yu, Yejin Choi, and Claire Cardie. 2019. DREAM: A Challenge Data Set and Models for Dialogue-Based Reading Comprehension. Transactions of the Association for Computational Linguistics, 7:217-231. Place: Cambridge, MA Publisher: MIT Press.", + "Harsh Trivedi, Heeyoung Kwon, Tushar Khot, Ashish Sabharwal, and Niranjan Balasubramanian. 2019." + ], + "bbox": [ + 510, + 85, + 880, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "832", + "bbox": [ + 485, + 928, + 515, + 939 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Repurposing Entailment for Multi-Hop Question Answering Tasks. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2948-2958, Minneapolis, Minnesota. Association for Computational Linguistics.", + "bbox": [ + 131, + 85, + 489, + 178 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Johannes Welbl, Nelson F. Liu, and Matt Gardner. 2017. Crowdsourcing Multiple Choice Science Questions. In NUT@EMNLP.", + "bbox": [ + 114, + 187, + 489, + 227 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Adina Williams, Nikita Nangia, and Samuel Bowman. 2018. A broad-coverage challenge corpus for sentence understanding through inference. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1112-1122, New Orleans, Louisiana. Association for Computational Linguistics.", + "bbox": [ + 115, + 237, + 489, + 355 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander M. Rush. 2020. HuggingFace's Transformers: State-of-the-art Natural Language Processing. Number: arXiv:1910.03771 arXiv:1910.03771 [cs].", + "bbox": [ + 115, + 363, + 489, + 495 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A Answer ranking procedure", + "text_level": 1, + "bbox": [ + 114, + 508, + 386, + 526 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the multiple choice setting, we performed an answer ranking procedure to pick the answer to a given question among a set of alternative answers $N$ , using both NLI class scores and QA confidence scores. (This is distinct from the selection procedure for the top 20% or 50% of answers we used in both settings.)", + "bbox": [ + 112, + 535, + 487, + 646 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Similar to Harabagiu and Hickl (2006), answers are ranked based on the highest probability from the calibration model $\\sigma$ , given a linear combination of the QA or NLI scores given an answer $n \\in N$ answer set. When a single feature is used, such as an NLI class or the QA score, no calibration is made and $\\sigma$ is simply the identity of the confidence score. In the case of contradiction only, $\\sigma$ is the inverse of the contradiction confidence score, indicating the least contradicted answer is being selected. Formally, our approach can be described as:", + "bbox": [ + 112, + 646, + 487, + 838 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname * {a r g m a x} _ {N} \\sigma (\\mathrm {Q A} _ {n}; \\mathrm {N L I} _ {n})\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 840, + 389, + 865 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathrm{QA}_n$ is the QA model confidence score for answer $n$ , and $\\mathrm{NLI}_n$ represents the various NLI class scores for $n$ .", + "bbox": [ + 112, + 871, + 487, + 917 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We did not use this approach in extractive QA, because we found that asking the model for the top $K = 4$ answer produced almost the same four answer alternatives with slightly different spans each time.", + "bbox": [ + 507, + 84, + 884, + 162 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "B Datasets", + "text_level": 1, + "bbox": [ + 509, + 175, + 621, + 191 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tables 4 (multiple choice) and 5 (extractive QA) outline the datasets we used. Additional details such as train size and preprocessing steps are available in the references provided. When space doesn't allow CosmosQA is aliased to Cosmos, MCScript to MCS, MCScript-2.0 to MCS2, and MCTest to MCT. The only preprocessing step we performed was to filter out questions where no context passage is provided. Validation splits (as opposed to test splits) are used in the CosmosQA and QASC cases, since context passages or gold standard answers are not available for these datasets.", + "bbox": [ + 507, + 200, + 884, + 394 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "C QA models", + "text_level": 1, + "bbox": [ + 507, + 406, + 645, + 422 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6 outlines the pretrained QA models that we used and the datasets they are trained on. All these models are publicly available on the Hugging Face hub under the locations listed. Where space doesn't allow, RoBERTa-RACE is aliased as RACE.", + "bbox": [ + 507, + 431, + 882, + 511 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We trained the two DeBERTa-v3 models (xsmall and base) as shown in Table 7. They were trained using the Hugging Face trainer API (Wolf et al., 2020) with an Adam optimizer at a learning rate of $5.60\\mathrm{e - }05$ with weight decay of 0.01. All models and inference were performed on 1 Tesla P100 GPU. Full instructions on reproducibility as well as trained models are provided in the publicly available code, including directions to weights and biases to inspect the training runs, full parameter set, and evaluation suites.", + "bbox": [ + 507, + 512, + 882, + 688 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "D QA2D models", + "text_level": 1, + "bbox": [ + 507, + 700, + 670, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A QA2D model reformulates a question-answer pair to a declarative statement (Demszky et al., 2018). As noted in Chen et al. (2021) and Mishra et al. (2021), the QA2D reformulation is critical to using NLI models in QA since the proposed answer needs to match the format of NLI. We trained a T5-small model (Raffel et al., 2020) on the dataset proposed by Demszky et al. (2018) for QA2D since we found almost no noticeable differences in performance in larger models. This used the same setup as the DeBERTa-v3 models xsmall and base (see Table 7).", + "bbox": [ + 507, + 726, + 884, + 917 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "833", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/95ff9adeab6c72a1bc22cef52c0589dd4940a5fdda9a97baeed88dc83f6f4017.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetSplitSizeReference
CosmosQAvalidation2985Huang et al. (2019)
DREAMtest2041Sun et al. (2019)
MCScripttest2797Ostermann et al. (2018)
MCScript-2.0test3610Ostermann et al. (2019)
MCTesttest840Richardson et al. (2013)
QASCvalidation926Khot et al. (2020)
RACEtest4934Lai et al. (2017)
RACE-Ctest712Liang et al. (2019)
SciQtest884Welbl et al. (2017)
", + "bbox": [ + 262, + 80, + 736, + 247 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ec3116058e0f8a9362002140c917858818ecd846c2d3ed929efaefe44bc4765d.jpg", + "table_caption": [ + "Table 4: Datasets used for the multiple choice setting, including split used and sample size. Validation splits were used for CosmosQA since the test split is not publicly available, and for QASC since context passages or gold answers are not available." + ], + "table_footnote": [], + "table_body": "
DatasetSizeReference
BioASQ1504Fisch et al. (2019)
TriviaQA7785
HotpotQA5901
SQuAD10506
Natural Questions12836
SQuAD211871Rajpurkar et al. (2018)
SQuAD-adv5347Jia and Liang (2017)
", + "bbox": [ + 292, + 311, + 707, + 445 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5: Extractive QA datasets used. Validation sets are used on the SQuAD2.0 and SQuAD adversarial datasets. MRQA 2019 dev sets are used for the other five datasets.", + "bbox": [ + 112, + 455, + 882, + 483 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Unlike Chen et al. (2021), we found that regardless of size, these QA2D models struggled with long questions or questions with complex syntax and would often leave the answer out of the statement. In order to solve this, constrained decoding that required the answer to be in the statement was tried. However, this often produced ungrammatical or nonsensical statements. We settled with the following heuristic to postprocess QA2D outputs: If less than $50\\%$ of the tokens in the answer were in the statement then we appended the answer to the end of the statement. $50\\%$ was used to account for rephrasing the answer or swapping pronouns. While some statements resulted in answer redundancy, this was better than having hypotheses which left out the answer.", + "bbox": [ + 112, + 508, + 489, + 765 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Future work on QA2D should focus on how these models can be used outside of the domains in the dataset provided by Demszky et al. (2018). Finally it is important to note that erroneous QA2D outputs could effect the quality of the whole pipeline see Chen et al. (2021) for a more detailed analysis of this.", + "bbox": [ + 112, + 806, + 489, + 919 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "E NLI models", + "text_level": 1, + "bbox": [ + 509, + 508, + 650, + 524 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "NLI is used to classify whether the reformulated answer is contradicted, entailed, or neutral with respect to a context passage. We used the whole context, as Schuster et al. (2022) and Mishra et al. (2021) demonstrated that long premises still performed adequate though not as well as sentence-length premises. Using the whole context avoids needing to use decontextualization as is required in Chen et al. (2021).", + "bbox": [ + 507, + 533, + 884, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We used two DeBERTa-based models (He et al., 2021b) trained on the MNLI dataset (Williams et al., 2018) (called mnli-base and mnli-large) and an ALBERT model (Lan et al., 2019) trained on the ANLI dataset in addition to various other NLI datasets (called albert-anli) (Nie et al., 2020). Table 6 contains the Hugging Face references to the NLI models. After inference, the confidence scores are used for answer selection and performance evaluation.", + "bbox": [ + 507, + 678, + 885, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "E.1 Model size and approach performance analysis", + "text_level": 1, + "bbox": [ + 507, + 850, + 860, + 882 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8 mirrors Table 1 in the main text, but shows the accuracy results for uncalibrated E, C, and $\\mathbf{E} + \\mathbf{C}$", + "bbox": [ + 507, + 887, + 882, + 917 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "834", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/aabe6847b74a5c1f9194a6b3c8dffb8261573554b3ed4595b5ad315827774805.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Hugging FaceName
LIAMF-USP/roberta-large-finetuned-RACERoBERTa-RACE
bert-large-uncased-whole-word-masking-finetuned-squadBERT-Large
distilbert-base-uncased-distilled-squadDistillBERT
ynie/albert-xxlarge-v2-snli_mnli_fever_anli_R1_R2_R3-nlialbert-anli
microsoft/deberta-base-mnlimnli-base
microsoft/deberta-v2-xxlarge-mnlimnli-large
", + "bbox": [ + 193, + 111, + 803, + 229 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/ff65b4f2ddde834b1a916e0e2db5de3c4c3e5d630d5aaf41816de55145f7042c.jpg", + "table_caption": [ + "Table 6: Pretrained QA and NLI models used." + ], + "table_footnote": [], + "table_body": "
ModelDatasetEpochsScore
t5-smallDemszky et al. (2018)20Rogue190.73
deberta-v3-xsmallWelbl et al. (2017)6Accuracy93.99
deberta-v3-baseWelbl et al. (2017)6Accuracy91.79
", + "bbox": [ + 215, + 322, + 783, + 391 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/7b086607a861841b750a1ee7a6e869cd48ea33a90061f475cc41082373f894a9.jpg", + "table_caption": [ + "Table 7: The models we trained for or setups with evaluation scores and number of epochs trained." + ], + "table_footnote": [], + "table_body": "
QA ModelCosmosDREAMMCSMCS2MCTQASCRACERACE-CSciQAverage
SciQ-base18.4643.8061.9963.7144.7693.4130.9727.3995.2853.31
SciQ-small25.4648.2660.2866.0459.7690.6035.5630.6298.0957.19
RACE64.2282.5689.7086.9890.4898.1676.9369.8097.9684.09
mnli-large
E+C44.3680.9485.5284.9990.6096.4464.2951.4092.4776.77
E36.1879.0386.0279.7289.8895.9062.1449.7291.9674.50
C59.2678.9883.1284.4389.2992.7662.7447.0591.5876.58
mnli-base
QA + E + C64.3282.6689.6387.0190.7198.2776.9569.8098.0984.16
QA + E64.2582.6689.6386.9890.7198.2776.9569.8097.9684.14
QA + C64.2982.5689.6387.0190.6098.1676.9369.8097.9684.1
E + C33.0362.2776.7672.1168.5792.6645.1634.4188.0163.66
E27.8162.4779.3771.9468.8192.6643.4834.4188.0163.22
C43.4559.1970.1869.9767.5081.8641.8132.5887.3761.55
albert-anli
QA + E + C64.1982.5689.7087.0690.4898.1676.9369.8097.9684.09
QA + E64.1982.5689.7087.0690.6098.1676.9369.8097.9684.11
QA + C64.2282.5689.7086.9890.4898.1676.9369.8097.9684.09
E + C35.7168.2079.5573.8877.5091.7949.0539.4790.8267.33
E33.6768.3579.9173.1977.3891.9049.0739.1990.9467.07
C45.1663.7473.5872.7173.3377.8646.3438.2087.2464.24
", + "bbox": [ + 115, + 483, + 884, + 845 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 8: Accuracy scores in the multiple choice setting for various NLI models used. Calibration was with the RoBERTA-RACE model.", + "bbox": [ + 112, + 854, + 882, + 883 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "835", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "in the main mnli-large model, as well as the results with the other NLI models, mnli-base and albertanli. Table 9 shows selective QA accuracy in the multiple choice setting where answer selection is done through ranking before we rank answers for selective QA. Selective QA on extractive QA using DistillBERT (table 10) shows that $\\mathbf{QA} + \\mathbf{E} + \\mathbf{C}$ does best in all cases and contradiction only does second best at $50\\%$ coverage.", + "bbox": [ + 112, + 84, + 487, + 229 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "F Calibration models", + "text_level": 1, + "bbox": [ + 112, + 242, + 317, + 256 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Like Kamath et al. (2020) and Chen et al. (2021) we developed a set of calibration models in order to perform answer ranking. A calibration model is trained on a set of posterior probabilities from downstream models to predict whether an answer is correct.", + "bbox": [ + 112, + 268, + 487, + 362 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To compare the effect of using different combinations of NLI class confidence scores, we trained a logistic regression model on linear combinations of the following features: QA indicates that the QA model confidence score is being used, $\\mathbf{E}$ indicates the entailment score, $\\mathbf{C}$ indicates the contradiction score, and $\\mathbf{N}$ indicates the neutral score. Like in Chen et al. (2021), all calibration models are trained on a holdout set of 100 samples from a single domain using logistic regression which predicts, given the confidence scores of the downstream models, whether the answer is correct. A multi-domain calibration approach like in Kamath et al. (2020) was not used since the focus was a minimum experiment to test the viability of leveraging different NLI classifications.", + "bbox": [ + 115, + 365, + 489, + 621 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "F.1 Regression Analysis", + "text_level": 1, + "bbox": [ + 112, + 634, + 317, + 649 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To illustrate the characteristics of the calibration models, we present a regression analysis for the multiple choice setting (Table 11). The results indicate that as the mnli model gets larger, the calibration model uses its NLI confidence scores more. Importantly, entailment coefficients are stronger than contradiction coefficients in all cases.", + "bbox": [ + 112, + 655, + 487, + 765 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "G Correlation Analysis", + "text_level": 1, + "bbox": [ + 112, + 781, + 332, + 796 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Since we are using the NLI and QA model scores to construct the setups above, it is useful to know how these factors correlate with the correct answer. Table 13 shows how each NLI class correlates both by score and by actual classification (score $>50\\%$ ) as compared against QA model confidence score. The multiple choice analysis shows", + "bbox": [ + 112, + 806, + 487, + 917 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "answers from the RoBERTa-RACE model and the extractive QA analysis shows answers from the BERT-large model trained on SQuAD. The correlation analysis presents Spearman rank correlations.", + "bbox": [ + 507, + 84, + 880, + 148 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "What we see is that in the multiple choice setting, the confidence score has a strong correlation with the correct answer, which makes sense given the confidence score is a softmax over the multiple choice classes. Extractive QA confidence scores have a much weaker correlation and tend to have less correlation than entailment has with the correct answer. Despite the results presented above, contradiction only has a notable correlation with the correct answer when the score is used rather than the classification. This is a point in favor of our approach of using confidence scores for NLI rather than classifications.", + "bbox": [ + 507, + 149, + 882, + 356 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Interestingly, in the extractive QA case, the neutral class is more negatively correlated when selecting for contradiction when using classification. Our conjecture would be that in the extractive QA case, we don't have much to compare against. When looking at the per dataset correlations for the multiple choice setting (Table 12), we see that in most cases, other than the QA confidence scores, the contradiction scores have the strongest correlations with the correct answer out of any NLI class and neutral, as we would expect, tends to have very weak correlations. We do not present the per dataset correlation for extractive QA as they are very weak, which we again hypothesize comes from having no answers to compare with.", + "bbox": [ + 507, + 357, + 882, + 599 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "836", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/408f34b65fbab8808d9bfcfc9d848a79c7d6b25d673bce83a108edd2bb832999.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetQA+E+CQA+EQA+CE+CECQA
20%CosmosQA77.5567.1783.2520.1027.4767.5088.61
DREAM98.2896.3296.8181.1391.9193.8798.28
MCScript99.8299.6499.4693.0298.9396.9699.82
MCScript-2.099.5899.0397.3792.2497.3795.0199.58
MCTest10010099.4085.1297.0297.0298.81
QASC10010010097.3010099.46100
RACE94.9392.1390.1762.7376.7175.0598.24
RACE-C88.7385.2186.6271.1374.6569.0193.66
SciQ10010010082.0510096.15100
Avg95.4393.2894.7976.0984.9087.7897.45
50%CosmosQA80.2970.7880.7032.1734.7264.8876.47
DREAM95.1093.6393.6385.2089.4188.3396.67
MCScript98.5797.8597.1494.7195.9992.7098.78
MCScript-2.096.4094.4696.0791.0291.7591.6998.01
MCTest99.5298.8199.7691.4395.2496.1999.52
QASC10099.7899.7898.2798.7098.49100
RACE90.1187.2285.2367.8971.7068.1893.88
RACE-C85.1178.0977.2566.5766.8555.0687.36
SciQ10010099.7489.0396.4396.43100
Avg93.9091.1892.1479.5982.3183.5594.52
", + "bbox": [ + 173, + 112, + 823, + 454 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/5c718e56704d68138c5deda131e4c1792915577d0978111b534fc7b34580456f.jpg", + "table_caption": [ + "Table 9: Selective QA accuracies in the multiple choice setting where answer selection is done through ranking before we rank answers for selective QA." + ], + "table_footnote": [], + "table_body": "
DatasetQA+E+CQA+EQA+CE+CECQA
20%BioASQ70.9770.4171.5574.0774.0774.3468.99
HotpotQA73.4473.0870.8871.5971.5170.4169.41
Natural Questions85.5985.2985.4578.4678.4680.5383.27
SQuAD96.2296.4595.7783.1583.0981.3797.15
SQuAD-adv40.3939.7539.4940.0739.5640.5931.98
SQuAD235.4635.2433.6436.3636.1336.6625.95
TriviaQA64.9664.6864.5552.6752.0952.5663.98
Avg66.7266.4165.9062.3462.1362.3562.96
50%BioASQ65.9665.9264.3763.5363.5366.9564.79
HotpotQA64.4264.2163.6565.8865.8566.9162.81
Natural Questions72.2871.9970.8267.5467.5174.1869.95
SQuAD92.5692.5792.3481.8682.2180.9592.54
SQuAD-adv33.6932.9033.4538.7438.2238.5231.89
SQuAD226.6825.7026.0032.9532.6132.8323.52
TriviaQA58.4058.4158.2551.4351.1852.9958.25
Avg59.1458.8158.4157.4257.3059.0557.68
", + "bbox": [ + 154, + 564, + 842, + 843 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 10: SelectiveQA on extractive QA using DistillBERT. Note that QA+E+C does best in all cases and contradiction only does second best at $50\\%$ coverage.", + "bbox": [ + 112, + 853, + 882, + 883 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "837", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/a850ae80ec18a1b5bc704fa8d01ae7903f77bd6be016fb12557705a1a902681f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
QA ModelNLI ModelCombinationConfidenceEntailmentContradictionAcc
SciQmnli-baseQA + C4.13-1.060.99
QA + E3.901.370.99
QA + E + C3.831.22-0.760.99
E + C2.56-1.470.86
mnli-largeQA + C3.98-1.320.99
QA + E3.781.550.99
QA + E + C3.651.31-0.970.99
E + C2.63-1.720.91
RACEmnli-baseQA + C3.04-0.150.89
QA + E3.030.270.89
QA + E + C3.020.26-0.140.89
E + C0.73-0.460.75
mnli-largeQA + C2.970.00-0.810.89
QA + E2.910.980.89
QA + E + C2.850.92-0.750.89
E + C1.76-1.120.78
", + "bbox": [ + 147, + 104, + 848, + 382 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/fe73c7096c9f26a9a8032c7ee38eaa69d298bb3a701d1ef341f645204ae60247.jpg", + "table_caption": [ + "Table 11: Regression analysis for each mnli-based nli model with each QA model used calibration with logistic regression multiple choice settings. Accuracy is the evaluation metric used." + ], + "table_footnote": [], + "table_body": "
ContradictionEntailmentNeutral
DatasetQAScoreClassScoreClassScoreClass
CosmosQA0.53-0.34-0.170.05-0.010.210.16
DREAM0.72-0.57-0.350.540.50-0.11-0.13
MCScript0.80-0.59-0.420.590.50-0.04-0.08
MCScript20.77-0.50-0.320.410.37-0.04-0.05
MCTest0.73-0.65-0.470.640.69-0.20-0.15
QASC0.57-0.54-0.280.550.67-0.50-0.26
RACE0.65-0.37-0.200.350.34-0.11-0.11
RACE-C0.59-0.24-0.130.180.25-0.09-0.11
SciQ0.75-0.69-0.470.680.67-0.42-0.19
", + "bbox": [ + 179, + 476, + 821, + 659 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/2d7dde0f008ecc0fe2efb805caf08f130341d5c5c37e4f70bc309a98dfb1f46f.jpg", + "table_caption": [ + "Table 12: Correlation analysis (Spearman rank correlation) per dataset in the multiple choice setting. RoBERTaRACE is used for the QA scores." + ], + "table_footnote": [], + "table_body": "
ContradictionEntailmentNeutralQA
multiple choiceScore-0.470.37-0.060.71
Class-0.280.38-0.06
extractive QAScore-0.160.31-0.120.19
Class-0.150.39-0.29
", + "bbox": [ + 223, + 752, + 774, + 839 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 13: Correlation analysis (Spearman rank correlation) in the multiple choice and extractive QA settings. RoBERTa-RACE is the QA model used for multiple choice QA scores and BERT-large is used for the extractive QA scores.", + "bbox": [ + 112, + 847, + 882, + 890 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "838", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A For every submission:", + "text_level": 1, + "bbox": [ + 114, + 107, + 322, + 122 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A1. Did you describe the limitations of your work?", + "bbox": [ + 127, + 127, + 532, + 143 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 149, + 143, + 231, + 159 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A2. Did you discuss any potential risks of your work?", + "bbox": [ + 127, + 170, + 552, + 186 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 149, + 187, + 231, + 200 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A3. Do the abstract and introduction summarize the paper's main claims?", + "bbox": [ + 127, + 212, + 695, + 229 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 149, + 230, + 231, + 244 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "□ A4. Have you used AI writing assistants when working on this paper?", + "bbox": [ + 127, + 255, + 668, + 272 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 149, + 273, + 231, + 287 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "B Did you use or create scientific artifacts?", + "text_level": 1, + "bbox": [ + 114, + 300, + 490, + 316 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 132, + 321, + 215, + 336 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "B1. Did you cite the creators of artifacts you used?", + "bbox": [ + 127, + 347, + 529, + 363 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 149, + 363, + 231, + 379 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?", + "bbox": [ + 127, + 390, + 778, + 406 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 149, + 407, + 231, + 422 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?", + "bbox": [ + 127, + 432, + 880, + 495 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 149, + 498, + 231, + 513 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?", + "bbox": [ + 127, + 524, + 880, + 571 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 149, + 573, + 231, + 588 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?", + "bbox": [ + 127, + 599, + 880, + 631 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 149, + 632, + 231, + 646 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be.", + "bbox": [ + 127, + 658, + 880, + 739 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 149, + 740, + 231, + 753 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "C Did you run computational experiments?", + "text_level": 1, + "bbox": [ + 114, + 765, + 495, + 781 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 132, + 785, + 215, + 801 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?", + "bbox": [ + 127, + 813, + 880, + 845 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 149, + 846, + 231, + 860 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.", + "bbox": [ + 112, + 868, + 877, + 892 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "ACL 2023 Responsible NLP Checklist", + "bbox": [ + 132, + 84, + 433, + 99 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "839", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Left blank.", + "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Left blank.", + "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Left blank." + ], + "bbox": [ + 127, + 84, + 880, + 282 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?", + "text_level": 1, + "bbox": [ + 112, + 293, + 877, + 309 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 132, + 313, + 213, + 329 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Left blank.", + "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Left blank.", + "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Left blank.", + "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Left blank.", + "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Left blank." + ], + "bbox": [ + 127, + 340, + 880, + 640 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "840", + "bbox": [ + 485, + 928, + 515, + 939 + ], + "page_idx": 13 + } +] \ No newline at end of file diff --git a/2023/Using contradictions improves question answering systems/56951eb2-a64c-4213-acb6-6239f47120c0_model.json b/2023/Using contradictions improves question answering systems/56951eb2-a64c-4213-acb6-6239f47120c0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..baa8033bb51003f658ec5d2481e66eeb1942c5ce --- /dev/null +++ b/2023/Using contradictions improves question answering systems/56951eb2-a64c-4213-acb6-6239f47120c0_model.json @@ -0,0 +1,2329 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.196, + 0.091, + 0.805, + 0.112 + ], + "angle": 0, + "content": "Using contradictions improves question answering systems" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.142, + 0.44, + 0.16 + ], + "angle": 0, + "content": "Étienne Fortier-Dubois" + }, + { + "type": "title", + "bbox": [ + 0.597, + 0.144, + 0.738, + 0.159 + ], + "angle": 0, + "content": "Domenic Rosati" + }, + { + "type": "text", + "bbox": [ + 0.638, + 0.163, + 0.696, + 0.176 + ], + "angle": 0, + "content": "scite.ai" + }, + { + "type": "text", + "bbox": [ + 0.581, + 0.179, + 0.753, + 0.194 + ], + "angle": 0, + "content": "Dalhousie University" + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.253, + 0.341, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.276, + 0.461, + 0.518 + ], + "angle": 0, + "content": "This work examines the use of contradiction in natural language inference (NLI) for question answering (QA). Typically, NLI systems help answer questions by determining if a potential answer is entailed (supported) by some background context. But is it useful to also determine if an answer contradicts the context? We test this in two settings, multiple choice and extractive QA, and find that systems that incorporate contradiction can do slightly better than entailment-only systems on certain datasets. However, the best performances come from using contradiction, entailment, and QA model confidence scores together. This has implications for the deployment of QA systems in domains such as medicine and science where safety is an issue." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.525, + 0.262, + 0.54 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.55, + 0.49, + 0.71 + ], + "angle": 0, + "content": "Safety in NLP systems is unresolved, particularly in biomedical and scientific contexts where hallucination, overconfidence, and other problems are major obstacles to deployment (Ji et al., 2022; Kell et al., 2021). One active area of research to solve these issues is natural language inference (NLI) (Li et al., 2022). NLI is the task of determining whether a hypothesis is true (entailed), false (contradicted), or undetermined (neutral) given some premise." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.711, + 0.49, + 0.92 + ], + "angle": 0, + "content": "Current NLI systems typically focus only on entailment to verify hypotheses—they calculate the degree to which a hypothesis is supported by the premise. But the premise can provide another signal: contradiction. Regardless of how well a hypothesis is entailed by the context, it can also be more or less contradicted, which could affect whether it is accepted or rejected. Contradictions are an important signal indicating whether some statement might be unacceptable given a premise. In some cases where we might not know if a statement is supported, we should still ensure we are rejecting statements that are outright contradicted." + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.25, + 0.885, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.409, + 0.884, + 0.466 + ], + "angle": 0, + "content": "Figure 1: A QA model is used to produce answers which are reformulated as hypotheses to determine if they are entailed or contradicted by a premise. The answers are ranked by NLI class scores to select the best answer." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.494, + 0.885, + 0.703 + ], + "angle": 0, + "content": "We wondered if adding this signal to a question answering (QA) system might improve performance and safety. To this end, we propose a method that reformulates answers from the QA system as hypotheses for NLI, calculates the entailment, contradiction, and neutrality of each hypothesis, and then selects the best one based on a combination of these results (Figure 1). We show that across 16 QA datasets (9 multiple choice and 7 extractive), the best approach is to use entailment, contradiction, and confidence scores together. Using only contradiction is roughly on par with, and sometimes better than, using only entailment." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.718, + 0.663, + 0.733 + ], + "angle": 0, + "content": "1.1 Related work" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.742, + 0.884, + 0.789 + ], + "angle": 0, + "content": "NLI for question answering has been explored by several authors in various settings; see Paramasi-vam and Nirmala (2021) for an overview." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.791, + 0.885, + 0.92 + ], + "angle": 0, + "content": "One of these settings is selective question answering for extractive QA, where selective refers to abstention when the system is not confident enough in its answer (Kamath et al., 2020). Chen et al. (2021) have found that NLI systems are able to verify the predictions made by a QA system in this setting, but their result is limited to only selecting a top \\( k\\% \\) of answers. Moreover, they" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.928, + 0.516, + 0.941 + ], + "angle": 0, + "content": "827" + }, + { + "type": "footer", + "bbox": [ + 0.227, + 0.946, + 0.772, + 0.958 + ], + "angle": 0, + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + }, + { + "type": "footer", + "bbox": [ + 0.378, + 0.959, + 0.622, + 0.972 + ], + "angle": 0, + "content": "Volume 2: Short Papers, pages 827-840" + }, + { + "type": "footer", + "bbox": [ + 0.297, + 0.973, + 0.702, + 0.985 + ], + "angle": 0, + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.489, + 0.149 + ], + "angle": 0, + "content": "do not provide an approach for improving overall performance, nor do they show the effect of incorporating contradiction directly (but do so indirectly by analyzing non-entailed passages)." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.151, + 0.49, + 0.358 + ], + "angle": 0, + "content": "In the related setting of multiple choice QA and fact checking, Mishra et al. (2021) have explored the use of entailment, finding that NLI models do well at these tasks by themselves, but can perform even better when they are adapted to in-domain data and longer premises. Yet their method uses only a two-class NLI set up (entailed or not entailed), which doesn't tell us much about directly using the contradiction signal. Pujari and Goldwasser (2019) does incorporate the contradiction signal showing the power of contradiction to improve machine comprehension but does not analyze its effects separately from entailment." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.36, + 0.489, + 0.456 + ], + "angle": 0, + "content": "Other QA settings in which NLI has been used include open domain (Harabagiu and Hickl, 2006) and multi-hop (Trivedi et al., 2019). Thus far, approaches tend to focus on entailment. To our knowledge, our work is the first to directly assess using contradictions for QA isolated from entailment." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.458, + 0.489, + 0.618 + ], + "angle": 0, + "content": "Outside of question answering, a domain that uses contradictions is factual consistency—the task of ensuring that a collection of utterances is faithful to a source document. Li et al. (2022) provide an overview. Typically, entailment is still the main focus, but Laban et al. (2022) propose an NLI-based method to ensure the consistency of a summary with a source document using contradiction and neutral scores in addition to entailment, beating out previous systems." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.62, + 0.489, + 0.699 + ], + "angle": 0, + "content": "Other researchers have used contradictions to identify consistency errors across Wikipedia (Schuster et al., 2022; Hsu et al., 2021) or generate credible character dialogue (Nie et al., 2021; Song et al., 2020)." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.714, + 0.225, + 0.729 + ], + "angle": 0, + "content": "2 Methods" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.742, + 0.489, + 0.87 + ], + "angle": 0, + "content": "We tested the effect of contradictions in two QA settings and a total of sixteen question-answer datasets. Our approach is broadly similar to both Chen et al. (2021) and Mishra et al. (2021) in that we use most of the same datasets for evaluating NLI reranking for multiple choice QA and extractive QA. Unlike both, we incorporate contradiction directly as a signal for reranking answers." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.872, + 0.489, + 0.919 + ], + "angle": 0, + "content": "Briefly, for each dataset, we used pretrained QA models to produce answers and confidence scores for the dataset's questions. We refer to the confi" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.228 + ], + "angle": 0, + "content": "dence scores below as QA. We then trained QA2D models (where QA2D stands for \"question-answer to declarative\") to turn the answers into the declarative hypothesis format required for NLI. For example, the question-answer pair \"What is the most abundant metal in the Earth crust? Copper.\" might be rephrased as \"The most abundant metal in the Earth crust is copper\" (see Appendix D for more details)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.231, + 0.885, + 0.359 + ], + "angle": 0, + "content": "With the question contexts as premises, we then used NLI models to classify every premise-hypothesis pair into three classes, each with an associated score: entailed (E), contradicted (C), and neutral (N). After that, we trained logistic regression calibration models to find which linear combination of the four scores—QA, E, C, and N—was best able to pick the answers accurately." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.361, + 0.885, + 0.553 + ], + "angle": 0, + "content": "When evaluating performance, we applied the selective QA approach from Kamath et al. (2020) to rank answers using combinations of the four scores, and then consider only those that the model was most confident in answering. We compared selecting the top \\(20\\%\\) and \\(50\\%\\). In the multiple choice setting, it was also possible to rank all potential answers according to the four scores, unlike in the extractive QA setting where the QA model produced only one answer per question, so we evaluated performance with that approach as well (see appendix A for details)." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.568, + 0.725, + 0.585 + ], + "angle": 0, + "content": "3 Experimental setting" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.597, + 0.884, + 0.723 + ], + "angle": 0, + "content": "In the multiple choice setting, we tested 9 datasets. Two of them are in-domain, since the pretrained QA models we used were finetuned on them. Specifically, we used a RoBERTa large model (Liu et al., 2019) finetuned on the RACE dataset (Lai et al., 2017), as well as two DeBERTa v3 variants, base and xsmall (He et al., 2021a), finetuned on the SciQ dataset (Welbl et al., 2017)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.726, + 0.884, + 0.885 + ], + "angle": 0, + "content": "In the extractive QA setting, we used 7 datasets: five from the MRQA 2019 task (Fisch et al., 2019), as well as SQuAD 2.0 (Rajpurkar et al., 2018) and SQuAD adversarial (Jia and Liang, 2017). The SQuAD model is the in-domain dataset: it was used to pretrain (Rajpurkar et al., 2016) the two QA models we used, DistillBERT (Sanh et al., 2020) and BERT-Large (Devlin et al., 2019). Like Chen et al. (2021), we used the Natural Questions dataset for calibration." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.888, + 0.883, + 0.919 + ], + "angle": 0, + "content": "In both settings, all datasets contain the relevant context that can be used by the QA models to select" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.516, + 0.94 + ], + "angle": 0, + "content": "828" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.49, + 0.117 + ], + "angle": 0, + "content": "answers. More detail on the datasets and QA models is available in appendices B and C respectively." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.118, + 0.49, + 0.246 + ], + "angle": 0, + "content": "See appendices D, E, and F for details on the QA2D, NLI, and calibration models. Our models follow the setups described in Kamath et al. (2020), Chen et al. (2021), and Mishra et al. (2021). The main interesting detail is that the calibration models were trained on a holdout set of 100 samples from a single domain, using logistic regression, as in Chen et al. (2021)." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.259, + 0.214, + 0.274 + ], + "angle": 0, + "content": "4 Results" + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.285, + 0.34, + 0.302 + ], + "angle": 0, + "content": "4.1 Multiple choice setting" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.306, + 0.49, + 0.483 + ], + "angle": 0, + "content": "For most multiple choice datasets, the best accuracy—when ranking all potential answers—is attained when using a calibrated model combining QA confidence, entailment, and contradiction (QA+E+C in Table 1). Only for the in-domain case (RACE-C) does the uncalibrated RoBERTa-RACE model perform on par with that. Using QA scores combined with either entailment (QA+E) or contradiction (QA+C) achieves similar performance, with contradiction winning by a small margin: \\(84.33\\%\\) average accuracy compared to \\(84.31\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.485, + 0.49, + 0.612 + ], + "angle": 0, + "content": "To inspect these trends further, we performed a correlation analysis of the NLI classes and QA confidence scores with the correct answer (appendix G). We found that besides QA confidence, it is the contradiction score that has the strongest correlation with the correct answer. The analysis also showed that the neutral class score (N) had almost no effect, which is why it is omitted in all results." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.613, + 0.49, + 0.725 + ], + "angle": 0, + "content": "When using the selective QA approach and evaluating only the \\(20\\%\\) of \\(50\\%\\) most confident answers, the best performance is attained with the \\(\\mathbf{QA} + \\mathbf{C}\\) combination (Table 2). This model is the only one that beats just using the QA confidence score on average. It is stronger than \\(\\mathbf{QA} + \\mathbf{E} + \\mathbf{C}\\) and \\(\\mathbf{QA} + \\mathbf{E}\\) for both coverage percentages." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.726, + 0.49, + 0.838 + ], + "angle": 0, + "content": "Contradiction alone, without QA confidence scores (C), also beats both entailment alone (E) and entailment with contradiction \\((\\mathbf{E} + \\mathbf{C})\\) for both coverages. These results match our intuition that the less contradicted an answer, the more likely it is correct, even in cases where there is uncertainty about its entailment." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.85, + 0.334, + 0.867 + ], + "angle": 0, + "content": "4.2 Extractive QA setting" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.872, + 0.49, + 0.92 + ], + "angle": 0, + "content": "Similar results occur when evaluating the extractive QA datasets with \\(20\\%\\) and \\(50\\%\\) selective coverage (Table 3). The \\(\\mathbf{QA} + \\mathbf{C}\\) model does better than QA" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.182 + ], + "angle": 0, + "content": "alone, and \\(\\mathbf{C}\\) alone does better than \\(\\mathbf{E} + \\mathbf{C}\\) or \\(\\mathbf{E}\\) alone, indicating the importance of the contradiction signal here too. However, entailment seems to matter more for extractive QA, as the best F1 score overall was from \\(\\mathbf{QA} + \\mathbf{E}\\) in the \\(20\\%\\) coverage case, and \\(\\mathbf{QA} + \\mathbf{E} + \\mathbf{C}\\) in the \\(50\\%\\) case." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.194, + 0.637, + 0.21 + ], + "angle": 0, + "content": "5 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.22, + 0.885, + 0.494 + ], + "angle": 0, + "content": "Contradiction with background context is a useful signal that NLP systems can use to infer answers to questions. This is not necessarily a superior strategy to using entailment, but our results show that combining these two signals can improve performance beyond what QA models can achieve on their own. These results are interesting because using contradictions comes with potential benefits for the safety of NLP systems and, as a result, their deployment in domains such as medicine or science. Namely, that there are many potential cases where we are not sure if a statement is entailed directly by a background context but we may be sure that the statement is not refuted by a background context. In two-class NLI settings where we focus only on entailment, neutral and contradiction are collapsed together and we don't have this guarantee." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.507, + 0.646, + 0.523 + ], + "angle": 0, + "content": "6 Limitations" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.534, + 0.885, + 0.71 + ], + "angle": 0, + "content": "Our work comes with some limitations. It is uncertain whether our results in two specific settings, multiple choice and extractive QA, would extend to more general settings for NLI, although the use of contradictions for factual consistency by Laban et al. (2022) suggests that they could. Additionally, 3-class NLI is not sufficient to capture all the natural language relations that might be needed to verify an answer. As such more challenging datasets in other settings and more granular NLI settings should be attempted." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.711, + 0.885, + 0.92 + ], + "angle": 0, + "content": "Another limitation involves answer ranking and the associated computational cost. The main reason we did not test answer ranking in extractive QA is that we did not generate diverse outputs, but another reason is that such a procedure grows prohibitively expensive as the domain becomes more open. In a fully open domain, ranking would require a quadratic evaluation for each context passage against each reformulated answer candidate (Schuster et al., 2022). Future work should look at comparison approaches that amortize this cost, such as NLI-based dense passage retrieval (Reimers and Gurevych, 2019)." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.517, + 0.941 + ], + "angle": 0, + "content": "829" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.129, + 0.093, + 0.87, + 0.181 + ], + "angle": 0, + "content": "
QA ModelCosmosDREAMMCSMCS2MCTQASCRACERACE-CSciQAverage
SciQ-base18.4643.8061.9963.7144.7693.4130.9727.3995.2853.30
SciQ-small25.4648.2660.2866.0459.7690.6035.5630.6298.0957.18
QA64.2282.5689.7086.9890.4898.1676.9369.8097.9684.08
QA+E+C64.72*83.19*90.06*87.59*91.43*98.6077.53*69.80*98.2184.57
QA+E64.3282.85*89.92*87.29*91.0798.49*77.1869.6698.0984.31
QA+C64.8282.75*89.88*87.29*90.8398.3877.1669.8098.0984.33
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.191, + 0.885, + 0.278 + ], + "angle": 0, + "content": "Table 1: Multiple choice setting. Accuracy scores (best per column in bold, second best underlined, statistical significance (pairwise students t-test) is indicated by asterix) after answer ranking with the mnli-large NLI model. The top three rows show the accuracy of using only the QA models' confidence score; \"QA\" refers to the scores of the RoBERTa-RACE model, which was used for calibration. The bottom rows add the entailment and/or contradiction scores to the RoBERTa-RACE score. For other NLI models, and for just E, C, and \\(\\mathrm{E + C}\\) without calibration with RoBERTa-RACE, see Table 8 in the appendix." + }, + { + "type": "table", + "bbox": [ + 0.212, + 0.306, + 0.787, + 0.561 + ], + "angle": 0, + "content": "
DatasetQA+E+CQA+CQA+EE+CECQA
20%CosmosQA77.5591.1276.8869.1868.3483.2588.61
DREAM98.2898.7798.2896.3296.3296.8198.28
MCScript99.8299.4699.8299.6499.6499.4699.82
MCScript-2.099.5899.7299.4599.1799.0397.3799.58
MCTest10099.4010010010099.4098.81
QASC100100100100100100100
RACE94.9396.6994.7292.4492.2490.1798.24
RACE-C88.7392.9689.4485.2185.9286.6293.66
SciQ100100100100100100100
Average95.4397.5795.4093.5593.5094.7997.45
50%CosmosQA80.2981.7076.9475.8070.6480.6376.47
DREAM95.1096.8694.9093.6393.6393.6396.67
MCScript98.5798.6498.2898.0097.9397.1498.78
MCScript-2.096.4098.2395.8494.6894.4096.0198.01
MCTest99.5299.7699.5299.0599.0599.7699.52
QASC10010010099.7899.7899.78100
RACE90.1192.6889.9987.7187.3885.2393.88
RACE-C85.1184.8385.3978.3778.3777.2587.36
SciQ10010010010010099.74100
Average93.9094.7493.4391.8991.2492.1394.52
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.571, + 0.884, + 0.613 + ], + "angle": 0, + "content": "Table 2: Multiple choice setting. Accuracy scores (best per row in bold, second best underlined) for selective QA with \\(20\\%\\) and \\(50\\%\\) coverage of the dataset. Calibrations and QA confidence are all from RoBERTa-RACE, where RACE is the in-domain dataset." + }, + { + "type": "table", + "bbox": [ + 0.202, + 0.644, + 0.797, + 0.851 + ], + "angle": 0, + "content": "
DatasetQA+E+CQA+CQA+EE+CECQA
20%BioASQ85.0483.1085.0674.2274.2275.4782.99
HotpotQA86.6285.8986.6980.6080.6079.8285.33
Natural Questions91.8492.1891.6879.8979.8782.0990.98
SQuAD98.2698.7692.3798.1792.4890.8899.04
SQuAD-adv43.9943.5743.9843.7443.6042.8139.83
SQuAD237.6436.0737.5637.4337.3137.6830.52
TriviaQA81.3380.3681.2165.5365.2569.1380.68
Average74.9674.1974.9967.6867.6268.2772.77
50%BioASQ76.1375.5176.0471.4971.4972.9775.49
HotpotQA79.3778.9579.3077.4377.4377.3178.74
Natural Questions84.5383.2484.4874.9674.9378.6282.47
SQuAD96.9897.0196.9791.5891.5291.1997.00
SQuAD-adv41.8041.4941.1642.7642.7942.0340.26
SQuAD229.4128.7728.4534.4334.1434.3926.18
TriviaQA74.3074.2374.3765.0564.9368.0874.21
Average68.9368.4668.6865.3965.3266.3767.76
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.861, + 0.884, + 0.903 + ], + "angle": 0, + "content": "Table 3: Extractive QA setting. F1 scores (best per row in bold, second best underlined) for selective QA with \\(20\\%\\) and \\(50\\%\\) coverage of the dataset. Calibrations and QA confidence are from the BERT-large model, where SQuAD is the in-domain dataset. For similar results on the smaller DistillBERT model, see Table 10 in the appendix." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.517, + 0.941 + ], + "angle": 0, + "content": "830" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.116, + 0.085, + 0.214, + 0.099 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.108, + 0.489, + 0.187 + ], + "angle": 0, + "content": "Jifan Chen, Eunsol Choi, and Greg Durrett. 2021. Can NLI Models Verify QA Systems' Predictions? In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 3841-3854, Punta Cana, Dominican Republic. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.198, + 0.489, + 0.264 + ], + "angle": 0, + "content": "Dorottya Demszky, Kelvin Guu, and Percy Liang. 2018. Transforming Question Answering Datasets Into Natural Language Inference Datasets. Technical Report arXiv:1809.02922, arXiv. ArXiv:1809.02922 [cs] type: article." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.274, + 0.489, + 0.393 + ], + "angle": 0, + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.403, + 0.489, + 0.495 + ], + "angle": 0, + "content": "Adam Fisch, Alon Talmor, Robin Jia, Minjoon Seo, Eunsol Choi, and Danqi Chen. 2019. MRQA 2019 Shared Task: Evaluating Generalization in Reading Comprehension. In Proceedings of the 2nd Workshop on Machine Reading for Question Answering, pages 1-13, Hong Kong, China. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.506, + 0.489, + 0.599 + ], + "angle": 0, + "content": "Sanda Harabagiu and Andrew Hickl. 2006. Methods for Using Textual Entailment in Open-Domain Question Answering. In Proceedings of the 21st International Conference on Computational Linguistics and 44th Annual Meeting of the Association for Computational Linguistics, pages 905-912, Sydney, Australia. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.609, + 0.489, + 0.675 + ], + "angle": 0, + "content": "Pengcheng He, Jianfeng Gao, and Weizhu Chen. 2021a. DeBERTaV3: Improving DeBERTa using ELECTRA-Style Pre-Training with Gradient-Disentangled Embedding Sharing. Number: arXiv:2111.09543 arXiv:2111.09543 [cs]." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.686, + 0.489, + 0.739 + ], + "angle": 0, + "content": "Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. 2021b. DeBERTa: Decoding-enhanced BERT with Disentangled Attention. Number: arXiv:2006.03654 arXiv:2006.03654 [cs]." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.75, + 0.489, + 0.816 + ], + "angle": 0, + "content": "Cheng Hsu, Cheng-Te Li, Diego Saez-Trumper, and Yi-Zhan Hsu. 2021. WikiContradiction: Detecting Self-Contradiction Articles on Wikipedia. Technical Report arXiv:2111.08543, arXiv. ArXiv:2111.08543 [cs] type: article." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.826, + 0.489, + 0.919 + ], + "angle": 0, + "content": "Lifu Huang, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. 2019. Cosmos QA: Machine Reading Comprehension with Contextual Commonsense Reasoning. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages" + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.108, + 0.489, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.53, + 0.086, + 0.884, + 0.113 + ], + "angle": 0, + "content": "2391-2401, Hong Kong, China. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.121, + 0.884, + 0.187 + ], + "angle": 0, + "content": "Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Yejin Bang, Andrea Madotto, and Pascale Fung. 2022. Survey of Hallucination in Natural Language Generation. Number: arXiv:2202.03629 arXiv:2202.03629 [cs]." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.195, + 0.884, + 0.275 + ], + "angle": 0, + "content": "Robin Jia and Percy Liang. 2017. Adversarial Examples for Evaluating Reading Comprehension Systems. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2021-2031, Copenhagen, Denmark. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.282, + 0.884, + 0.361 + ], + "angle": 0, + "content": "Amita Kamath, Robin Jia, and Percy Liang. 2020. Selective Question Answering under Domain Shift. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5684-5696, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.369, + 0.884, + 0.449 + ], + "angle": 0, + "content": "Gregory Kell, Iain Marshall, Byron Wallace, and Andre Jaun. 2021. What Would it Take to get Biomedical QA Systems into Practice? In Proceedings of the 3rd Workshop on Machine Reading for Question Answering, pages 28-41, Punta Cana, Dominican Republic. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.456, + 0.884, + 0.534 + ], + "angle": 0, + "content": "Tushar Khot, Peter Clark, Michal Guerquin, Peter Jansen, and Ashish Sabharwal. 2020. QASC: A Dataset for Question Answering via Sentence Composition. Proceedings of the AAAI Conference on Artificial Intelligence, 34(05):8082-8090. Number: 05." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.543, + 0.884, + 0.609 + ], + "angle": 0, + "content": "Philippe Laban, Tobias Schnabel, Paul N. Bennett, and Marti A. Hearst. 2022. SummaC: Re-Visiting NLIBased Models for Inconsistency Detection in Summarization. Transactions of the Association for Computational Linguistics, 10:163-177." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.617, + 0.884, + 0.71 + ], + "angle": 0, + "content": "Guokun Lai, Qizhe Xie, Hanxiao Liu, Yiming Yang, and Eduard Hovy. 2017. RACE: Large-scale ReAding Comprehension Dataset From Examinations. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 785-794, Copenhagen, Denmark. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.717, + 0.884, + 0.771 + ], + "angle": 0, + "content": "Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. 2019. Albert: A lite bert for self-supervised learning of language representations." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.778, + 0.884, + 0.845 + ], + "angle": 0, + "content": "Wei Li, Wenhao Wu, Moye Chen, Jiachen Liu, Xinyan Xiao, and Hua Wu. 2022. Faithfulness in Natural Language Generation: A Systematic Survey of Analysis, Evaluation and Optimization Methods. Number: arXiv:2203.05227 arXiv:2203.05227 [cs]." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.852, + 0.884, + 0.919 + ], + "angle": 0, + "content": "Yichan Liang, Jianheng Li, and Jian Yin. 2019. A New Multi-choice Reading Comprehension Dataset for Curriculum Learning. In Proceedings of The Eleventh Asian Conference on Machine Learning, pages 742-757. PMLR. ISSN: 2640-3498." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.884, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.514, + 0.941 + ], + "angle": 0, + "content": "831" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.086, + 0.487, + 0.165 + ], + "angle": 0, + "content": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. RoBERTa: A Robustly Optimized BERT Pretraining Approach. Number: arXiv:1907.11692 arXiv:1907.11692 [cs]." + }, + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.176, + 0.487, + 0.293 + ], + "angle": 0, + "content": "Anshuman Mishra, Dhruvesh Patel, Aparna Vijayakumar, Xiang Lorraine Li, Pavan Kapanipathi, and Kartik Talamadupula. 2021. Looking Beyond Sentence-Level Natural Language Inference for Question Answering and Text Summarization. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1322-1336, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.304, + 0.487, + 0.395 + ], + "angle": 0, + "content": "Yixin Nie, Adina Williams, Emily Dinan, Mohit Bansal, Jason Weston, and Douwe Kiela. 2020. Adversarial NLI: A New Benchmark for Natural Language Understanding. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4885-4901, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.406, + 0.487, + 0.523 + ], + "angle": 0, + "content": "Yixin Nie, Mary Williamson, Mohit Bansal, Douwe Kiela, and Jason Weston. 2021. I like fish, especially dolphins: Addressing Contradictions in Dialogue Modeling. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1699-1713, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.534, + 0.487, + 0.638 + ], + "angle": 0, + "content": "Simon Ostermann, Ashutosh Modi, Michael Roth, Stefan Thater, and Manfred Pinkal. 2018. MCScript: A Novel Dataset for Assessing Machine Comprehension Using Script Knowledge. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources Association (ELRA)." + }, + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.649, + 0.487, + 0.74 + ], + "angle": 0, + "content": "Simon Ostermann, Michael Roth, and Manfred Pinkal. 2019. MCScript2.0: A Machine Comprehension Corpus Focused on Script Events and Participants. In Proceedings of the Eighth Joint Conference on Lexical and Computational Semantics (*SEM* 2019), pages 103-117, Minneapolis, Minnesota. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.751, + 0.487, + 0.803 + ], + "angle": 0, + "content": "Aarthi Paramasivam and S. Jaya Nirmala. 2021. A survey on textual entailment based question answering. Journal of King Saud University - Computer and Information Sciences." + }, + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.814, + 0.487, + 0.918 + ], + "angle": 0, + "content": "Rajkumar Pujari and Dan Goldwasser. 2019. Using natural language relations between answer choices for machine comprehension. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4010-4015, Minneapolis, Minnesota. Association for Computational Linguistics." + }, + { + "type": "list", + "bbox": [ + 0.118, + 0.086, + 0.487, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.086, + 0.882, + 0.165 + ], + "angle": 0, + "content": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer. Journal of Machine Learning Research, 21(140):1-67." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.176, + 0.882, + 0.267 + ], + "angle": 0, + "content": "Pranav Rajpurkar, Robin Jia, and Percy Liang. 2018. Know What You Don't Know: Unanswerable Questions for SQuAD. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 784-789, Melbourne, Australia. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.279, + 0.882, + 0.357 + ], + "angle": 0, + "content": "Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD: 100,000+ Questions for Machine Comprehension of Text. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2383-2392, Austin, Texas. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.368, + 0.882, + 0.473 + ], + "angle": 0, + "content": "Nils Reimers and Iryna Gurevych. 2019. SentenceBERT: Sentence Embeddings using Siamese BERT-Networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3982-3992, Hong Kong, China. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.484, + 0.882, + 0.575 + ], + "angle": 0, + "content": "Matthew Richardson, Christopher J.C. Burges, and Erin Renshaw. 2013. MCTest: A Challenge Dataset for the Open-Domain Machine Comprehension of Text. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 193-203, Seattle, Washington, USA. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.586, + 0.882, + 0.638 + ], + "angle": 0, + "content": "Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. 2020. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. Number: arXiv:1910.01108 arXiv:1910.01108 [cs]." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.649, + 0.882, + 0.715 + ], + "angle": 0, + "content": "Tal Schuster, Sihao Chen, Senaka Buthpitiya, Alex Fabrikant, and Donald Metzler. 2022. Stretching Sentence-pair NLI Models to Reason over Long Documents and Clusters. Number: arXiv:2204.07447 arXiv:2204.07447 [cs]." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.726, + 0.882, + 0.791 + ], + "angle": 0, + "content": "Haoyu Song, Wei-Nan Zhang, Jingwen Hu, and Ting Liu. 2020. Generating Persona Consistent Dialogues by Exploiting Natural Language Inference. Proceedings of the AAAI Conference on Artificial Intelligence, 34(05):8878-8885. Number: 05." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.803, + 0.882, + 0.881 + ], + "angle": 0, + "content": "Kai Sun, Dian Yu, Jianshu Chen, Dong Yu, Yejin Choi, and Claire Cardie. 2019. DREAM: A Challenge Data Set and Models for Dialogue-Based Reading Comprehension. Transactions of the Association for Computational Linguistics, 7:217-231. Place: Cambridge, MA Publisher: MIT Press." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.892, + 0.882, + 0.918 + ], + "angle": 0, + "content": "Harsh Trivedi, Heeyoung Kwon, Tushar Khot, Ashish Sabharwal, and Niranjan Balasubramanian. 2019." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.882, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.516, + 0.94 + ], + "angle": 0, + "content": "832" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.132, + 0.086, + 0.49, + 0.179 + ], + "angle": 0, + "content": "Repurposing Entailment for Multi-Hop Question Answering Tasks. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2948-2958, Minneapolis, Minnesota. Association for Computational Linguistics." + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.188, + 0.49, + 0.228 + ], + "angle": 0, + "content": "Johannes Welbl, Nelson F. Liu, and Matt Gardner. 2017. Crowdsourcing Multiple Choice Science Questions. In NUT@EMNLP." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.238, + 0.49, + 0.356 + ], + "angle": 0, + "content": "Adina Williams, Nikita Nangia, and Samuel Bowman. 2018. A broad-coverage challenge corpus for sentence understanding through inference. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1112-1122, New Orleans, Louisiana. Association for Computational Linguistics." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.365, + 0.49, + 0.497 + ], + "angle": 0, + "content": "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander M. Rush. 2020. HuggingFace's Transformers: State-of-the-art Natural Language Processing. Number: arXiv:1910.03771 arXiv:1910.03771 [cs]." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.51, + 0.388, + 0.527 + ], + "angle": 0, + "content": "A Answer ranking procedure" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.536, + 0.489, + 0.648 + ], + "angle": 0, + "content": "In the multiple choice setting, we performed an answer ranking procedure to pick the answer to a given question among a set of alternative answers \\( N \\), using both NLI class scores and QA confidence scores. (This is distinct from the selection procedure for the top 20% or 50% of answers we used in both settings.)" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.648, + 0.489, + 0.839 + ], + "angle": 0, + "content": "Similar to Harabagiu and Hickl (2006), answers are ranked based on the highest probability from the calibration model \\(\\sigma\\), given a linear combination of the QA or NLI scores given an answer \\(n \\in N\\) answer set. When a single feature is used, such as an NLI class or the QA score, no calibration is made and \\(\\sigma\\) is simply the identity of the confidence score. In the case of contradiction only, \\(\\sigma\\) is the inverse of the contradiction confidence score, indicating the least contradicted answer is being selected. Formally, our approach can be described as:" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.841, + 0.39, + 0.866 + ], + "angle": 0, + "content": "\\[\n\\operatorname * {a r g m a x} _ {N} \\sigma (\\mathrm {Q A} _ {n}; \\mathrm {N L I} _ {n})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.872, + 0.489, + 0.919 + ], + "angle": 0, + "content": "where \\(\\mathrm{QA}_n\\) is the QA model confidence score for answer \\(n\\), and \\(\\mathrm{NLI}_n\\) represents the various NLI class scores for \\(n\\)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.164 + ], + "angle": 0, + "content": "We did not use this approach in extractive QA, because we found that asking the model for the top \\( K = 4 \\) answer produced almost the same four answer alternatives with slightly different spans each time." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.177, + 0.623, + 0.192 + ], + "angle": 0, + "content": "B Datasets" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.202, + 0.885, + 0.395 + ], + "angle": 0, + "content": "Tables 4 (multiple choice) and 5 (extractive QA) outline the datasets we used. Additional details such as train size and preprocessing steps are available in the references provided. When space doesn't allow CosmosQA is aliased to Cosmos, MCScript to MCS, MCScript-2.0 to MCS2, and MCTest to MCT. The only preprocessing step we performed was to filter out questions where no context passage is provided. Validation splits (as opposed to test splits) are used in the CosmosQA and QASC cases, since context passages or gold standard answers are not available for these datasets." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.407, + 0.646, + 0.423 + ], + "angle": 0, + "content": "C QA models" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.432, + 0.884, + 0.512 + ], + "angle": 0, + "content": "Table 6 outlines the pretrained QA models that we used and the datasets they are trained on. All these models are publicly available on the Hugging Face hub under the locations listed. Where space doesn't allow, RoBERTa-RACE is aliased as RACE." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.513, + 0.884, + 0.689 + ], + "angle": 0, + "content": "We trained the two DeBERTa-v3 models (xsmall and base) as shown in Table 7. They were trained using the Hugging Face trainer API (Wolf et al., 2020) with an Adam optimizer at a learning rate of \\(5.60\\mathrm{e - }05\\) with weight decay of 0.01. All models and inference were performed on 1 Tesla P100 GPU. Full instructions on reproducibility as well as trained models are provided in the publicly available code, including directions to weights and biases to inspect the training runs, full parameter set, and evaluation suites." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.701, + 0.671, + 0.717 + ], + "angle": 0, + "content": "D QA2D models" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.727, + 0.885, + 0.919 + ], + "angle": 0, + "content": "A QA2D model reformulates a question-answer pair to a declarative statement (Demszky et al., 2018). As noted in Chen et al. (2021) and Mishra et al. (2021), the QA2D reformulation is critical to using NLI models in QA since the proposed answer needs to match the format of NLI. We trained a T5-small model (Raffel et al., 2020) on the dataset proposed by Demszky et al. (2018) for QA2D since we found almost no noticeable differences in performance in larger models. This used the same setup as the DeBERTa-v3 models xsmall and base (see Table 7)." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.516, + 0.941 + ], + "angle": 0, + "content": "833" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.263, + 0.081, + 0.737, + 0.248 + ], + "angle": 0, + "content": "
DatasetSplitSizeReference
CosmosQAvalidation2985Huang et al. (2019)
DREAMtest2041Sun et al. (2019)
MCScripttest2797Ostermann et al. (2018)
MCScript-2.0test3610Ostermann et al. (2019)
MCTesttest840Richardson et al. (2013)
QASCvalidation926Khot et al. (2020)
RACEtest4934Lai et al. (2017)
RACE-Ctest712Liang et al. (2019)
SciQtest884Welbl et al. (2017)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.257, + 0.885, + 0.3 + ], + "angle": 0, + "content": "Table 4: Datasets used for the multiple choice setting, including split used and sample size. Validation splits were used for CosmosQA since the test split is not publicly available, and for QASC since context passages or gold answers are not available." + }, + { + "type": "table", + "bbox": [ + 0.294, + 0.312, + 0.709, + 0.447 + ], + "angle": 0, + "content": "
DatasetSizeReference
BioASQ1504Fisch et al. (2019)
TriviaQA7785
HotpotQA5901
SQuAD10506
Natural Questions12836
SQuAD211871Rajpurkar et al. (2018)
SQuAD-adv5347Jia and Liang (2017)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.456, + 0.884, + 0.485 + ], + "angle": 0, + "content": "Table 5: Extractive QA datasets used. Validation sets are used on the SQuAD2.0 and SQuAD adversarial datasets. MRQA 2019 dev sets are used for the other five datasets." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.51, + 0.49, + 0.766 + ], + "angle": 0, + "content": "Unlike Chen et al. (2021), we found that regardless of size, these QA2D models struggled with long questions or questions with complex syntax and would often leave the answer out of the statement. In order to solve this, constrained decoding that required the answer to be in the statement was tried. However, this often produced ungrammatical or nonsensical statements. We settled with the following heuristic to postprocess QA2D outputs: If less than \\(50\\%\\) of the tokens in the answer were in the statement then we appended the answer to the end of the statement. \\(50\\%\\) was used to account for rephrasing the answer or swapping pronouns. While some statements resulted in answer redundancy, this was better than having hypotheses which left out the answer." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.807, + 0.49, + 0.92 + ], + "angle": 0, + "content": "Future work on QA2D should focus on how these models can be used outside of the domains in the dataset provided by Demszky et al. (2018). Finally it is important to note that erroneous QA2D outputs could effect the quality of the whole pipeline see Chen et al. (2021) for a more detailed analysis of this." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.51, + 0.652, + 0.525 + ], + "angle": 0, + "content": "E NLI models" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.534, + 0.885, + 0.679 + ], + "angle": 0, + "content": "NLI is used to classify whether the reformulated answer is contradicted, entailed, or neutral with respect to a context passage. We used the whole context, as Schuster et al. (2022) and Mishra et al. (2021) demonstrated that long premises still performed adequate though not as well as sentence-length premises. Using the whole context avoids needing to use decontextualization as is required in Chen et al. (2021)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.68, + 0.887, + 0.839 + ], + "angle": 0, + "content": "We used two DeBERTa-based models (He et al., 2021b) trained on the MNLI dataset (Williams et al., 2018) (called mnli-base and mnli-large) and an ALBERT model (Lan et al., 2019) trained on the ANLI dataset in addition to various other NLI datasets (called albert-anli) (Nie et al., 2020). Table 6 contains the Hugging Face references to the NLI models. After inference, the confidence scores are used for answer selection and performance evaluation." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.851, + 0.862, + 0.883 + ], + "angle": 0, + "content": "E.1 Model size and approach performance analysis" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.888, + 0.884, + 0.919 + ], + "angle": 0, + "content": "Table 8 mirrors Table 1 in the main text, but shows the accuracy results for uncalibrated E, C, and \\(\\mathbf{E} + \\mathbf{C}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.517, + 0.941 + ], + "angle": 0, + "content": "834" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.194, + 0.112, + 0.805, + 0.23 + ], + "angle": 0, + "content": "
Hugging FaceName
LIAMF-USP/roberta-large-finetuned-RACERoBERTa-RACE
bert-large-uncased-whole-word-masking-finetuned-squadBERT-Large
distilbert-base-uncased-distilled-squadDistillBERT
ynie/albert-xxlarge-v2-snli_mnli_fever_anli_R1_R2_R3-nlialbert-anli
microsoft/deberta-base-mnlimnli-base
microsoft/deberta-v2-xxlarge-mnlimnli-large
" + }, + { + "type": "table_caption", + "bbox": [ + 0.341, + 0.239, + 0.655, + 0.253 + ], + "angle": 0, + "content": "Table 6: Pretrained QA and NLI models used." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.323, + 0.784, + 0.392 + ], + "angle": 0, + "content": "
ModelDatasetEpochsScore
t5-smallDemszky et al. (2018)20Rogue190.73
deberta-v3-xsmallWelbl et al. (2017)6Accuracy93.99
deberta-v3-baseWelbl et al. (2017)6Accuracy91.79
" + }, + { + "type": "table_caption", + "bbox": [ + 0.165, + 0.402, + 0.832, + 0.417 + ], + "angle": 0, + "content": "Table 7: The models we trained for or setups with evaluation scores and number of epochs trained." + }, + { + "type": "table", + "bbox": [ + 0.116, + 0.485, + 0.885, + 0.846 + ], + "angle": 0, + "content": "
QA ModelCosmosDREAMMCSMCS2MCTQASCRACERACE-CSciQAverage
SciQ-base18.4643.8061.9963.7144.7693.4130.9727.3995.2853.31
SciQ-small25.4648.2660.2866.0459.7690.6035.5630.6298.0957.19
RACE64.2282.5689.7086.9890.4898.1676.9369.8097.9684.09
mnli-large
E+C44.3680.9485.5284.9990.6096.4464.2951.4092.4776.77
E36.1879.0386.0279.7289.8895.9062.1449.7291.9674.50
C59.2678.9883.1284.4389.2992.7662.7447.0591.5876.58
mnli-base
QA + E + C64.3282.6689.6387.0190.7198.2776.9569.8098.0984.16
QA + E64.2582.6689.6386.9890.7198.2776.9569.8097.9684.14
QA + C64.2982.5689.6387.0190.6098.1676.9369.8097.9684.1
E + C33.0362.2776.7672.1168.5792.6645.1634.4188.0163.66
E27.8162.4779.3771.9468.8192.6643.4834.4188.0163.22
C43.4559.1970.1869.9767.5081.8641.8132.5887.3761.55
albert-anli
QA + E + C64.1982.5689.7087.0690.4898.1676.9369.8097.9684.09
QA + E64.1982.5689.7087.0690.6098.1676.9369.8097.9684.11
QA + C64.2282.5689.7086.9890.4898.1676.9369.8097.9684.09
E + C35.7168.2079.5573.8877.5091.7949.0539.4790.8267.33
E33.6768.3579.9173.1977.3891.9049.0739.1990.9467.07
C45.1663.7473.5872.7173.3377.8646.3438.2087.2464.24
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.855, + 0.883, + 0.884 + ], + "angle": 0, + "content": "Table 8: Accuracy scores in the multiple choice setting for various NLI models used. Calibration was with the RoBERTA-RACE model." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.516, + 0.941 + ], + "angle": 0, + "content": "835" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.488, + 0.23 + ], + "angle": 0, + "content": "in the main mnli-large model, as well as the results with the other NLI models, mnli-base and albertanli. Table 9 shows selective QA accuracy in the multiple choice setting where answer selection is done through ranking before we rank answers for selective QA. Selective QA on extractive QA using DistillBERT (table 10) shows that \\(\\mathbf{QA} + \\mathbf{E} + \\mathbf{C}\\) does best in all cases and contradiction only does second best at \\(50\\%\\) coverage." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.243, + 0.318, + 0.258 + ], + "angle": 0, + "content": "F Calibration models" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.269, + 0.488, + 0.363 + ], + "angle": 0, + "content": "Like Kamath et al. (2020) and Chen et al. (2021) we developed a set of calibration models in order to perform answer ranking. A calibration model is trained on a set of posterior probabilities from downstream models to predict whether an answer is correct." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.366, + 0.49, + 0.622 + ], + "angle": 0, + "content": "To compare the effect of using different combinations of NLI class confidence scores, we trained a logistic regression model on linear combinations of the following features: QA indicates that the QA model confidence score is being used, \\(\\mathbf{E}\\) indicates the entailment score, \\(\\mathbf{C}\\) indicates the contradiction score, and \\(\\mathbf{N}\\) indicates the neutral score. Like in Chen et al. (2021), all calibration models are trained on a holdout set of 100 samples from a single domain using logistic regression which predicts, given the confidence scores of the downstream models, whether the answer is correct. A multi-domain calibration approach like in Kamath et al. (2020) was not used since the focus was a minimum experiment to test the viability of leveraging different NLI classifications." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.635, + 0.319, + 0.65 + ], + "angle": 0, + "content": "F.1 Regression Analysis" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.656, + 0.488, + 0.766 + ], + "angle": 0, + "content": "To illustrate the characteristics of the calibration models, we present a regression analysis for the multiple choice setting (Table 11). The results indicate that as the mnli model gets larger, the calibration model uses its NLI confidence scores more. Importantly, entailment coefficients are stronger than contradiction coefficients in all cases." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.782, + 0.334, + 0.797 + ], + "angle": 0, + "content": "G Correlation Analysis" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.807, + 0.488, + 0.919 + ], + "angle": 0, + "content": "Since we are using the NLI and QA model scores to construct the setups above, it is useful to know how these factors correlate with the correct answer. Table 13 shows how each NLI class correlates both by score and by actual classification (score \\(>50\\%\\)) as compared against QA model confidence score. The multiple choice analysis shows" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.882, + 0.149 + ], + "angle": 0, + "content": "answers from the RoBERTa-RACE model and the extractive QA analysis shows answers from the BERT-large model trained on SQuAD. The correlation analysis presents Spearman rank correlations." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.15, + 0.884, + 0.357 + ], + "angle": 0, + "content": "What we see is that in the multiple choice setting, the confidence score has a strong correlation with the correct answer, which makes sense given the confidence score is a softmax over the multiple choice classes. Extractive QA confidence scores have a much weaker correlation and tend to have less correlation than entailment has with the correct answer. Despite the results presented above, contradiction only has a notable correlation with the correct answer when the score is used rather than the classification. This is a point in favor of our approach of using confidence scores for NLI rather than classifications." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.359, + 0.884, + 0.6 + ], + "angle": 0, + "content": "Interestingly, in the extractive QA case, the neutral class is more negatively correlated when selecting for contradiction when using classification. Our conjecture would be that in the extractive QA case, we don't have much to compare against. When looking at the per dataset correlations for the multiple choice setting (Table 12), we see that in most cases, other than the QA confidence scores, the contradiction scores have the strongest correlations with the correct answer out of any NLI class and neutral, as we would expect, tends to have very weak correlations. We do not present the per dataset correlation for extractive QA as they are very weak, which we again hypothesize comes from having no answers to compare with." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.517, + 0.941 + ], + "angle": 0, + "content": "836" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.174, + 0.113, + 0.825, + 0.455 + ], + "angle": 0, + "content": "
DatasetQA+E+CQA+EQA+CE+CECQA
20%CosmosQA77.5567.1783.2520.1027.4767.5088.61
DREAM98.2896.3296.8181.1391.9193.8798.28
MCScript99.8299.6499.4693.0298.9396.9699.82
MCScript-2.099.5899.0397.3792.2497.3795.0199.58
MCTest10010099.4085.1297.0297.0298.81
QASC10010010097.3010099.46100
RACE94.9392.1390.1762.7376.7175.0598.24
RACE-C88.7385.2186.6271.1374.6569.0193.66
SciQ10010010082.0510096.15100
Avg95.4393.2894.7976.0984.9087.7897.45
50%CosmosQA80.2970.7880.7032.1734.7264.8876.47
DREAM95.1093.6393.6385.2089.4188.3396.67
MCScript98.5797.8597.1494.7195.9992.7098.78
MCScript-2.096.4094.4696.0791.0291.7591.6998.01
MCTest99.5298.8199.7691.4395.2496.1999.52
QASC10099.7899.7898.2798.7098.49100
RACE90.1187.2285.2367.8971.7068.1893.88
RACE-C85.1178.0977.2566.5766.8555.0687.36
SciQ10010099.7489.0396.4396.43100
Avg93.9091.1892.1479.5982.3183.5594.52
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.466, + 0.882, + 0.496 + ], + "angle": 0, + "content": "Table 9: Selective QA accuracies in the multiple choice setting where answer selection is done through ranking before we rank answers for selective QA." + }, + { + "type": "table", + "bbox": [ + 0.156, + 0.565, + 0.843, + 0.844 + ], + "angle": 0, + "content": "
DatasetQA+E+CQA+EQA+CE+CECQA
20%BioASQ70.9770.4171.5574.0774.0774.3468.99
HotpotQA73.4473.0870.8871.5971.5170.4169.41
Natural Questions85.5985.2985.4578.4678.4680.5383.27
SQuAD96.2296.4595.7783.1583.0981.3797.15
SQuAD-adv40.3939.7539.4940.0739.5640.5931.98
SQuAD235.4635.2433.6436.3636.1336.6625.95
TriviaQA64.9664.6864.5552.6752.0952.5663.98
Avg66.7266.4165.9062.3462.1362.3562.96
50%BioASQ65.9665.9264.3763.5363.5366.9564.79
HotpotQA64.4264.2163.6565.8865.8566.9162.81
Natural Questions72.2871.9970.8267.5467.5174.1869.95
SQuAD92.5692.5792.3481.8682.2180.9592.54
SQuAD-adv33.6932.9033.4538.7438.2238.5231.89
SQuAD226.6825.7026.0032.9532.6132.8323.52
TriviaQA58.4058.4158.2551.4351.1852.9958.25
Avg59.1458.8158.4157.4257.3059.0557.68
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.854, + 0.884, + 0.884 + ], + "angle": 0, + "content": "Table 10: SelectiveQA on extractive QA using DistillBERT. Note that QA+E+C does best in all cases and contradiction only does second best at \\(50\\%\\) coverage." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.516, + 0.941 + ], + "angle": 0, + "content": "837" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.149, + 0.105, + 0.849, + 0.384 + ], + "angle": 0, + "content": "
QA ModelNLI ModelCombinationConfidenceEntailmentContradictionAcc
SciQmnli-baseQA + C4.13-1.060.99
QA + E3.901.370.99
QA + E + C3.831.22-0.760.99
E + C2.56-1.470.86
mnli-largeQA + C3.98-1.320.99
QA + E3.781.550.99
QA + E + C3.651.31-0.970.99
E + C2.63-1.720.91
RACEmnli-baseQA + C3.04-0.150.89
QA + E3.030.270.89
QA + E + C3.020.26-0.140.89
E + C0.73-0.460.75
mnli-largeQA + C2.970.00-0.810.89
QA + E2.910.980.89
QA + E + C2.850.92-0.750.89
E + C1.76-1.120.78
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.393, + 0.884, + 0.424 + ], + "angle": 0, + "content": "Table 11: Regression analysis for each mnli-based nli model with each QA model used calibration with logistic regression multiple choice settings. Accuracy is the evaluation metric used." + }, + { + "type": "table", + "bbox": [ + 0.18, + 0.477, + 0.822, + 0.66 + ], + "angle": 0, + "content": "
ContradictionEntailmentNeutral
DatasetQAScoreClassScoreClassScoreClass
CosmosQA0.53-0.34-0.170.05-0.010.210.16
DREAM0.72-0.57-0.350.540.50-0.11-0.13
MCScript0.80-0.59-0.420.590.50-0.04-0.08
MCScript20.77-0.50-0.320.410.37-0.04-0.05
MCTest0.73-0.65-0.470.640.69-0.20-0.15
QASC0.57-0.54-0.280.550.67-0.50-0.26
RACE0.65-0.37-0.200.350.34-0.11-0.11
RACE-C0.59-0.24-0.130.180.25-0.09-0.11
SciQ0.75-0.69-0.470.680.67-0.42-0.19
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.669, + 0.884, + 0.699 + ], + "angle": 0, + "content": "Table 12: Correlation analysis (Spearman rank correlation) per dataset in the multiple choice setting. RoBERTaRACE is used for the QA scores." + }, + { + "type": "table", + "bbox": [ + 0.225, + 0.753, + 0.776, + 0.84 + ], + "angle": 0, + "content": "
ContradictionEntailmentNeutralQA
multiple choiceScore-0.470.37-0.060.71
Class-0.280.38-0.06
extractive QAScore-0.160.31-0.120.19
Class-0.150.39-0.29
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.848, + 0.884, + 0.891 + ], + "angle": 0, + "content": "Table 13: Correlation analysis (Spearman rank correlation) in the multiple choice and extractive QA settings. RoBERTa-RACE is the QA model used for multiple choice QA scores and BERT-large is used for the extractive QA scores." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.517, + 0.941 + ], + "angle": 0, + "content": "838" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.134, + 0.085, + 0.435, + 0.1 + ], + "angle": 0, + "content": "ACL 2023 Responsible NLP Checklist" + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.108, + 0.323, + 0.123 + ], + "angle": 0, + "content": "A For every submission:" + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.128, + 0.533, + 0.144 + ], + "angle": 0, + "content": "A1. Did you describe the limitations of your work?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.145, + 0.233, + 0.16 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.171, + 0.553, + 0.187 + ], + "angle": 0, + "content": "A2. Did you discuss any potential risks of your work?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.188, + 0.233, + 0.202 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.214, + 0.696, + 0.23 + ], + "angle": 0, + "content": "A3. Do the abstract and introduction summarize the paper's main claims?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.231, + 0.233, + 0.246 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.256, + 0.669, + 0.273 + ], + "angle": 0, + "content": "□ A4. Have you used AI writing assistants when working on this paper?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.274, + 0.233, + 0.288 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.301, + 0.492, + 0.317 + ], + "angle": 0, + "content": "B Did you use or create scientific artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.322, + 0.216, + 0.337 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.348, + 0.531, + 0.364 + ], + "angle": 0, + "content": "B1. Did you cite the creators of artifacts you used?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.365, + 0.233, + 0.38 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.391, + 0.779, + 0.407 + ], + "angle": 0, + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.408, + 0.233, + 0.423 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.434, + 0.882, + 0.497 + ], + "angle": 0, + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.499, + 0.233, + 0.514 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.525, + 0.882, + 0.573 + ], + "angle": 0, + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.574, + 0.233, + 0.589 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.6, + 0.882, + 0.632 + ], + "angle": 0, + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.633, + 0.233, + 0.648 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.659, + 0.882, + 0.74 + ], + "angle": 0, + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be." + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.741, + 0.233, + 0.755 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.766, + 0.496, + 0.782 + ], + "angle": 0, + "content": "C Did you run computational experiments?" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.787, + 0.216, + 0.802 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.814, + 0.882, + 0.846 + ], + "angle": 0, + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.847, + 0.233, + 0.861 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.869, + 0.878, + 0.893 + ], + "angle": 0, + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.517, + 0.941 + ], + "angle": 0, + "content": "839" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.129, + 0.085, + 0.88, + 0.133 + ], + "angle": 0, + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.144, + 0.881, + 0.207 + ], + "angle": 0, + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.219, + 0.881, + 0.283 + ], + "angle": 0, + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Left blank." + }, + { + "type": "list", + "bbox": [ + 0.129, + 0.085, + 0.881, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.294, + 0.878, + 0.31 + ], + "angle": 0, + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.315, + 0.215, + 0.33 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.341, + 0.881, + 0.388 + ], + "angle": 0, + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.4, + 0.881, + 0.464 + ], + "angle": 0, + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.475, + 0.881, + 0.54 + ], + "angle": 0, + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.55, + 0.873, + 0.582 + ], + "angle": 0, + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.593, + 0.881, + 0.641 + ], + "angle": 0, + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Left blank." + }, + { + "type": "list", + "bbox": [ + 0.129, + 0.341, + 0.881, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.516, + 0.94 + ], + "angle": 0, + "content": "840" + } + ] +] \ No newline at end of file diff --git a/2023/Using contradictions improves question answering systems/56951eb2-a64c-4213-acb6-6239f47120c0_origin.pdf b/2023/Using contradictions improves question answering systems/56951eb2-a64c-4213-acb6-6239f47120c0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1a85a304b0782e0d3beb4430ce58be2a69647a0c --- /dev/null +++ b/2023/Using contradictions improves question answering systems/56951eb2-a64c-4213-acb6-6239f47120c0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0d3c3144a1e6f2ba10fe991ebd31aa158f9f22088c4a48b49577edf1979d18c3 +size 339184 diff --git a/2023/Using contradictions improves question answering systems/full.md b/2023/Using contradictions improves question answering systems/full.md new file mode 100644 index 0000000000000000000000000000000000000000..582e7c82b3e2a019160e9d3cb5fc72d8f9e55a8d --- /dev/null +++ b/2023/Using contradictions improves question answering systems/full.md @@ -0,0 +1,330 @@ +# Using contradictions improves question answering systems + +# Étienne Fortier-Dubois + +# Domenic Rosati + +scite.ai + +Dalhousie University + +# Abstract + +This work examines the use of contradiction in natural language inference (NLI) for question answering (QA). Typically, NLI systems help answer questions by determining if a potential answer is entailed (supported) by some background context. But is it useful to also determine if an answer contradicts the context? We test this in two settings, multiple choice and extractive QA, and find that systems that incorporate contradiction can do slightly better than entailment-only systems on certain datasets. However, the best performances come from using contradiction, entailment, and QA model confidence scores together. This has implications for the deployment of QA systems in domains such as medicine and science where safety is an issue. + +# 1 Introduction + +Safety in NLP systems is unresolved, particularly in biomedical and scientific contexts where hallucination, overconfidence, and other problems are major obstacles to deployment (Ji et al., 2022; Kell et al., 2021). One active area of research to solve these issues is natural language inference (NLI) (Li et al., 2022). NLI is the task of determining whether a hypothesis is true (entailed), false (contradicted), or undetermined (neutral) given some premise. + +Current NLI systems typically focus only on entailment to verify hypotheses—they calculate the degree to which a hypothesis is supported by the premise. But the premise can provide another signal: contradiction. Regardless of how well a hypothesis is entailed by the context, it can also be more or less contradicted, which could affect whether it is accepted or rejected. Contradictions are an important signal indicating whether some statement might be unacceptable given a premise. In some cases where we might not know if a statement is supported, we should still ensure we are rejecting statements that are outright contradicted. + +![](images/0782e5fe3e33d197559d16540fd53b83aff03fb7da05f392f5edf22449885b76.jpg) +Figure 1: A QA model is used to produce answers which are reformulated as hypotheses to determine if they are entailed or contradicted by a premise. The answers are ranked by NLI class scores to select the best answer. + +We wondered if adding this signal to a question answering (QA) system might improve performance and safety. To this end, we propose a method that reformulates answers from the QA system as hypotheses for NLI, calculates the entailment, contradiction, and neutrality of each hypothesis, and then selects the best one based on a combination of these results (Figure 1). We show that across 16 QA datasets (9 multiple choice and 7 extractive), the best approach is to use entailment, contradiction, and confidence scores together. Using only contradiction is roughly on par with, and sometimes better than, using only entailment. + +# 1.1 Related work + +NLI for question answering has been explored by several authors in various settings; see Paramasi-vam and Nirmala (2021) for an overview. + +One of these settings is selective question answering for extractive QA, where selective refers to abstention when the system is not confident enough in its answer (Kamath et al., 2020). Chen et al. (2021) have found that NLI systems are able to verify the predictions made by a QA system in this setting, but their result is limited to only selecting a top $k\%$ of answers. Moreover, they + +do not provide an approach for improving overall performance, nor do they show the effect of incorporating contradiction directly (but do so indirectly by analyzing non-entailed passages). + +In the related setting of multiple choice QA and fact checking, Mishra et al. (2021) have explored the use of entailment, finding that NLI models do well at these tasks by themselves, but can perform even better when they are adapted to in-domain data and longer premises. Yet their method uses only a two-class NLI set up (entailed or not entailed), which doesn't tell us much about directly using the contradiction signal. Pujari and Goldwasser (2019) does incorporate the contradiction signal showing the power of contradiction to improve machine comprehension but does not analyze its effects separately from entailment. + +Other QA settings in which NLI has been used include open domain (Harabagiu and Hickl, 2006) and multi-hop (Trivedi et al., 2019). Thus far, approaches tend to focus on entailment. To our knowledge, our work is the first to directly assess using contradictions for QA isolated from entailment. + +Outside of question answering, a domain that uses contradictions is factual consistency—the task of ensuring that a collection of utterances is faithful to a source document. Li et al. (2022) provide an overview. Typically, entailment is still the main focus, but Laban et al. (2022) propose an NLI-based method to ensure the consistency of a summary with a source document using contradiction and neutral scores in addition to entailment, beating out previous systems. + +Other researchers have used contradictions to identify consistency errors across Wikipedia (Schuster et al., 2022; Hsu et al., 2021) or generate credible character dialogue (Nie et al., 2021; Song et al., 2020). + +# 2 Methods + +We tested the effect of contradictions in two QA settings and a total of sixteen question-answer datasets. Our approach is broadly similar to both Chen et al. (2021) and Mishra et al. (2021) in that we use most of the same datasets for evaluating NLI reranking for multiple choice QA and extractive QA. Unlike both, we incorporate contradiction directly as a signal for reranking answers. + +Briefly, for each dataset, we used pretrained QA models to produce answers and confidence scores for the dataset's questions. We refer to the confi + +dence scores below as QA. We then trained QA2D models (where QA2D stands for "question-answer to declarative") to turn the answers into the declarative hypothesis format required for NLI. For example, the question-answer pair "What is the most abundant metal in the Earth crust? Copper." might be rephrased as "The most abundant metal in the Earth crust is copper" (see Appendix D for more details). + +With the question contexts as premises, we then used NLI models to classify every premise-hypothesis pair into three classes, each with an associated score: entailed (E), contradicted (C), and neutral (N). After that, we trained logistic regression calibration models to find which linear combination of the four scores—QA, E, C, and N—was best able to pick the answers accurately. + +When evaluating performance, we applied the selective QA approach from Kamath et al. (2020) to rank answers using combinations of the four scores, and then consider only those that the model was most confident in answering. We compared selecting the top $20\%$ and $50\%$ . In the multiple choice setting, it was also possible to rank all potential answers according to the four scores, unlike in the extractive QA setting where the QA model produced only one answer per question, so we evaluated performance with that approach as well (see appendix A for details). + +# 3 Experimental setting + +In the multiple choice setting, we tested 9 datasets. Two of them are in-domain, since the pretrained QA models we used were finetuned on them. Specifically, we used a RoBERTa large model (Liu et al., 2019) finetuned on the RACE dataset (Lai et al., 2017), as well as two DeBERTa v3 variants, base and xsmall (He et al., 2021a), finetuned on the SciQ dataset (Welbl et al., 2017). + +In the extractive QA setting, we used 7 datasets: five from the MRQA 2019 task (Fisch et al., 2019), as well as SQuAD 2.0 (Rajpurkar et al., 2018) and SQuAD adversarial (Jia and Liang, 2017). The SQuAD model is the in-domain dataset: it was used to pretrain (Rajpurkar et al., 2016) the two QA models we used, DistillBERT (Sanh et al., 2020) and BERT-Large (Devlin et al., 2019). Like Chen et al. (2021), we used the Natural Questions dataset for calibration. + +In both settings, all datasets contain the relevant context that can be used by the QA models to select + +answers. More detail on the datasets and QA models is available in appendices B and C respectively. + +See appendices D, E, and F for details on the QA2D, NLI, and calibration models. Our models follow the setups described in Kamath et al. (2020), Chen et al. (2021), and Mishra et al. (2021). The main interesting detail is that the calibration models were trained on a holdout set of 100 samples from a single domain, using logistic regression, as in Chen et al. (2021). + +# 4 Results + +# 4.1 Multiple choice setting + +For most multiple choice datasets, the best accuracy—when ranking all potential answers—is attained when using a calibrated model combining QA confidence, entailment, and contradiction (QA+E+C in Table 1). Only for the in-domain case (RACE-C) does the uncalibrated RoBERTa-RACE model perform on par with that. Using QA scores combined with either entailment (QA+E) or contradiction (QA+C) achieves similar performance, with contradiction winning by a small margin: $84.33\%$ average accuracy compared to $84.31\%$ . + +To inspect these trends further, we performed a correlation analysis of the NLI classes and QA confidence scores with the correct answer (appendix G). We found that besides QA confidence, it is the contradiction score that has the strongest correlation with the correct answer. The analysis also showed that the neutral class score (N) had almost no effect, which is why it is omitted in all results. + +When using the selective QA approach and evaluating only the $20\%$ of $50\%$ most confident answers, the best performance is attained with the $\mathbf{QA} + \mathbf{C}$ combination (Table 2). This model is the only one that beats just using the QA confidence score on average. It is stronger than $\mathbf{QA} + \mathbf{E} + \mathbf{C}$ and $\mathbf{QA} + \mathbf{E}$ for both coverage percentages. + +Contradiction alone, without QA confidence scores (C), also beats both entailment alone (E) and entailment with contradiction $(\mathbf{E} + \mathbf{C})$ for both coverages. These results match our intuition that the less contradicted an answer, the more likely it is correct, even in cases where there is uncertainty about its entailment. + +# 4.2 Extractive QA setting + +Similar results occur when evaluating the extractive QA datasets with $20\%$ and $50\%$ selective coverage (Table 3). The $\mathbf{QA} + \mathbf{C}$ model does better than QA + +alone, and $\mathbf{C}$ alone does better than $\mathbf{E} + \mathbf{C}$ or $\mathbf{E}$ alone, indicating the importance of the contradiction signal here too. However, entailment seems to matter more for extractive QA, as the best F1 score overall was from $\mathbf{QA} + \mathbf{E}$ in the $20\%$ coverage case, and $\mathbf{QA} + \mathbf{E} + \mathbf{C}$ in the $50\%$ case. + +# 5 Discussion + +Contradiction with background context is a useful signal that NLP systems can use to infer answers to questions. This is not necessarily a superior strategy to using entailment, but our results show that combining these two signals can improve performance beyond what QA models can achieve on their own. These results are interesting because using contradictions comes with potential benefits for the safety of NLP systems and, as a result, their deployment in domains such as medicine or science. Namely, that there are many potential cases where we are not sure if a statement is entailed directly by a background context but we may be sure that the statement is not refuted by a background context. In two-class NLI settings where we focus only on entailment, neutral and contradiction are collapsed together and we don't have this guarantee. + +# 6 Limitations + +Our work comes with some limitations. It is uncertain whether our results in two specific settings, multiple choice and extractive QA, would extend to more general settings for NLI, although the use of contradictions for factual consistency by Laban et al. (2022) suggests that they could. Additionally, 3-class NLI is not sufficient to capture all the natural language relations that might be needed to verify an answer. As such more challenging datasets in other settings and more granular NLI settings should be attempted. + +Another limitation involves answer ranking and the associated computational cost. The main reason we did not test answer ranking in extractive QA is that we did not generate diverse outputs, but another reason is that such a procedure grows prohibitively expensive as the domain becomes more open. In a fully open domain, ranking would require a quadratic evaluation for each context passage against each reformulated answer candidate (Schuster et al., 2022). Future work should look at comparison approaches that amortize this cost, such as NLI-based dense passage retrieval (Reimers and Gurevych, 2019). + +
QA ModelCosmosDREAMMCSMCS2MCTQASCRACERACE-CSciQAverage
SciQ-base18.4643.8061.9963.7144.7693.4130.9727.3995.2853.30
SciQ-small25.4648.2660.2866.0459.7690.6035.5630.6298.0957.18
QA64.2282.5689.7086.9890.4898.1676.9369.8097.9684.08
QA+E+C64.72*83.19*90.06*87.59*91.43*98.6077.53*69.80*98.2184.57
QA+E64.3282.85*89.92*87.29*91.0798.49*77.1869.6698.0984.31
QA+C64.8282.75*89.88*87.29*90.8398.3877.1669.8098.0984.33
+ +Table 1: Multiple choice setting. Accuracy scores (best per column in bold, second best underlined, statistical significance (pairwise students t-test) is indicated by asterix) after answer ranking with the mnli-large NLI model. The top three rows show the accuracy of using only the QA models' confidence score; "QA" refers to the scores of the RoBERTa-RACE model, which was used for calibration. The bottom rows add the entailment and/or contradiction scores to the RoBERTa-RACE score. For other NLI models, and for just E, C, and $\mathrm{E + C}$ without calibration with RoBERTa-RACE, see Table 8 in the appendix. + +
DatasetQA+E+CQA+CQA+EE+CECQA
20%CosmosQA77.5591.1276.8869.1868.3483.2588.61
DREAM98.2898.7798.2896.3296.3296.8198.28
MCScript99.8299.4699.8299.6499.6499.4699.82
MCScript-2.099.5899.7299.4599.1799.0397.3799.58
MCTest10099.4010010010099.4098.81
QASC100100100100100100100
RACE94.9396.6994.7292.4492.2490.1798.24
RACE-C88.7392.9689.4485.2185.9286.6293.66
SciQ100100100100100100100
Average95.4397.5795.4093.5593.5094.7997.45
50%CosmosQA80.2981.7076.9475.8070.6480.6376.47
DREAM95.1096.8694.9093.6393.6393.6396.67
MCScript98.5798.6498.2898.0097.9397.1498.78
MCScript-2.096.4098.2395.8494.6894.4096.0198.01
MCTest99.5299.7699.5299.0599.0599.7699.52
QASC10010010099.7899.7899.78100
RACE90.1192.6889.9987.7187.3885.2393.88
RACE-C85.1184.8385.3978.3778.3777.2587.36
SciQ10010010010010099.74100
Average93.9094.7493.4391.8991.2492.1394.52
+ +Table 2: Multiple choice setting. Accuracy scores (best per row in bold, second best underlined) for selective QA with $20\%$ and $50\%$ coverage of the dataset. Calibrations and QA confidence are all from RoBERTa-RACE, where RACE is the in-domain dataset. + +
DatasetQA+E+CQA+CQA+EE+CECQA
20%BioASQ85.0483.1085.0674.2274.2275.4782.99
HotpotQA86.6285.8986.6980.6080.6079.8285.33
Natural Questions91.8492.1891.6879.8979.8782.0990.98
SQuAD98.2698.7692.3798.1792.4890.8899.04
SQuAD-adv43.9943.5743.9843.7443.6042.8139.83
SQuAD237.6436.0737.5637.4337.3137.6830.52
TriviaQA81.3380.3681.2165.5365.2569.1380.68
Average74.9674.1974.9967.6867.6268.2772.77
50%BioASQ76.1375.5176.0471.4971.4972.9775.49
HotpotQA79.3778.9579.3077.4377.4377.3178.74
Natural Questions84.5383.2484.4874.9674.9378.6282.47
SQuAD96.9897.0196.9791.5891.5291.1997.00
SQuAD-adv41.8041.4941.1642.7642.7942.0340.26
SQuAD229.4128.7728.4534.4334.1434.3926.18
TriviaQA74.3074.2374.3765.0564.9368.0874.21
Average68.9368.4668.6865.3965.3266.3767.76
+ +Table 3: Extractive QA setting. F1 scores (best per row in bold, second best underlined) for selective QA with $20\%$ and $50\%$ coverage of the dataset. Calibrations and QA confidence are from the BERT-large model, where SQuAD is the in-domain dataset. For similar results on the smaller DistillBERT model, see Table 10 in the appendix. + +# References + +Jifan Chen, Eunsol Choi, and Greg Durrett. 2021. Can NLI Models Verify QA Systems' Predictions? In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 3841-3854, Punta Cana, Dominican Republic. Association for Computational Linguistics. +Dorottya Demszky, Kelvin Guu, and Percy Liang. 2018. Transforming Question Answering Datasets Into Natural Language Inference Datasets. Technical Report arXiv:1809.02922, arXiv. ArXiv:1809.02922 [cs] type: article. +Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics. +Adam Fisch, Alon Talmor, Robin Jia, Minjoon Seo, Eunsol Choi, and Danqi Chen. 2019. MRQA 2019 Shared Task: Evaluating Generalization in Reading Comprehension. In Proceedings of the 2nd Workshop on Machine Reading for Question Answering, pages 1-13, Hong Kong, China. Association for Computational Linguistics. +Sanda Harabagiu and Andrew Hickl. 2006. Methods for Using Textual Entailment in Open-Domain Question Answering. In Proceedings of the 21st International Conference on Computational Linguistics and 44th Annual Meeting of the Association for Computational Linguistics, pages 905-912, Sydney, Australia. Association for Computational Linguistics. +Pengcheng He, Jianfeng Gao, and Weizhu Chen. 2021a. DeBERTaV3: Improving DeBERTa using ELECTRA-Style Pre-Training with Gradient-Disentangled Embedding Sharing. Number: arXiv:2111.09543 arXiv:2111.09543 [cs]. +Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. 2021b. DeBERTa: Decoding-enhanced BERT with Disentangled Attention. Number: arXiv:2006.03654 arXiv:2006.03654 [cs]. +Cheng Hsu, Cheng-Te Li, Diego Saez-Trumper, and Yi-Zhan Hsu. 2021. WikiContradiction: Detecting Self-Contradiction Articles on Wikipedia. Technical Report arXiv:2111.08543, arXiv. ArXiv:2111.08543 [cs] type: article. +Lifu Huang, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. 2019. Cosmos QA: Machine Reading Comprehension with Contextual Commonsense Reasoning. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages + +2391-2401, Hong Kong, China. Association for Computational Linguistics. +Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Yejin Bang, Andrea Madotto, and Pascale Fung. 2022. Survey of Hallucination in Natural Language Generation. Number: arXiv:2202.03629 arXiv:2202.03629 [cs]. +Robin Jia and Percy Liang. 2017. Adversarial Examples for Evaluating Reading Comprehension Systems. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2021-2031, Copenhagen, Denmark. Association for Computational Linguistics. +Amita Kamath, Robin Jia, and Percy Liang. 2020. Selective Question Answering under Domain Shift. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5684-5696, Online. Association for Computational Linguistics. +Gregory Kell, Iain Marshall, Byron Wallace, and Andre Jaun. 2021. What Would it Take to get Biomedical QA Systems into Practice? In Proceedings of the 3rd Workshop on Machine Reading for Question Answering, pages 28-41, Punta Cana, Dominican Republic. Association for Computational Linguistics. +Tushar Khot, Peter Clark, Michal Guerquin, Peter Jansen, and Ashish Sabharwal. 2020. QASC: A Dataset for Question Answering via Sentence Composition. Proceedings of the AAAI Conference on Artificial Intelligence, 34(05):8082-8090. Number: 05. +Philippe Laban, Tobias Schnabel, Paul N. Bennett, and Marti A. Hearst. 2022. SummaC: Re-Visiting NLIBased Models for Inconsistency Detection in Summarization. Transactions of the Association for Computational Linguistics, 10:163-177. +Guokun Lai, Qizhe Xie, Hanxiao Liu, Yiming Yang, and Eduard Hovy. 2017. RACE: Large-scale ReAding Comprehension Dataset From Examinations. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 785-794, Copenhagen, Denmark. Association for Computational Linguistics. +Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. 2019. Albert: A lite bert for self-supervised learning of language representations. +Wei Li, Wenhao Wu, Moye Chen, Jiachen Liu, Xinyan Xiao, and Hua Wu. 2022. Faithfulness in Natural Language Generation: A Systematic Survey of Analysis, Evaluation and Optimization Methods. Number: arXiv:2203.05227 arXiv:2203.05227 [cs]. +Yichan Liang, Jianheng Li, and Jian Yin. 2019. A New Multi-choice Reading Comprehension Dataset for Curriculum Learning. In Proceedings of The Eleventh Asian Conference on Machine Learning, pages 742-757. PMLR. ISSN: 2640-3498. + +Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. RoBERTa: A Robustly Optimized BERT Pretraining Approach. Number: arXiv:1907.11692 arXiv:1907.11692 [cs]. +Anshuman Mishra, Dhruvesh Patel, Aparna Vijayakumar, Xiang Lorraine Li, Pavan Kapanipathi, and Kartik Talamadupula. 2021. Looking Beyond Sentence-Level Natural Language Inference for Question Answering and Text Summarization. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1322-1336, Online. Association for Computational Linguistics. +Yixin Nie, Adina Williams, Emily Dinan, Mohit Bansal, Jason Weston, and Douwe Kiela. 2020. Adversarial NLI: A New Benchmark for Natural Language Understanding. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4885-4901, Online. Association for Computational Linguistics. +Yixin Nie, Mary Williamson, Mohit Bansal, Douwe Kiela, and Jason Weston. 2021. I like fish, especially dolphins: Addressing Contradictions in Dialogue Modeling. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1699-1713, Online. Association for Computational Linguistics. +Simon Ostermann, Ashutosh Modi, Michael Roth, Stefan Thater, and Manfred Pinkal. 2018. MCScript: A Novel Dataset for Assessing Machine Comprehension Using Script Knowledge. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources Association (ELRA). +Simon Ostermann, Michael Roth, and Manfred Pinkal. 2019. MCScript2.0: A Machine Comprehension Corpus Focused on Script Events and Participants. In Proceedings of the Eighth Joint Conference on Lexical and Computational Semantics (*SEM* 2019), pages 103-117, Minneapolis, Minnesota. Association for Computational Linguistics. +Aarthi Paramasivam and S. Jaya Nirmala. 2021. A survey on textual entailment based question answering. Journal of King Saud University - Computer and Information Sciences. +Rajkumar Pujari and Dan Goldwasser. 2019. Using natural language relations between answer choices for machine comprehension. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4010-4015, Minneapolis, Minnesota. Association for Computational Linguistics. + +Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer. Journal of Machine Learning Research, 21(140):1-67. +Pranav Rajpurkar, Robin Jia, and Percy Liang. 2018. Know What You Don't Know: Unanswerable Questions for SQuAD. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 784-789, Melbourne, Australia. Association for Computational Linguistics. +Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD: 100,000+ Questions for Machine Comprehension of Text. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2383-2392, Austin, Texas. Association for Computational Linguistics. +Nils Reimers and Iryna Gurevych. 2019. SentenceBERT: Sentence Embeddings using Siamese BERT-Networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3982-3992, Hong Kong, China. Association for Computational Linguistics. +Matthew Richardson, Christopher J.C. Burges, and Erin Renshaw. 2013. MCTest: A Challenge Dataset for the Open-Domain Machine Comprehension of Text. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 193-203, Seattle, Washington, USA. Association for Computational Linguistics. +Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. 2020. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. Number: arXiv:1910.01108 arXiv:1910.01108 [cs]. +Tal Schuster, Sihao Chen, Senaka Buthpitiya, Alex Fabrikant, and Donald Metzler. 2022. Stretching Sentence-pair NLI Models to Reason over Long Documents and Clusters. Number: arXiv:2204.07447 arXiv:2204.07447 [cs]. +Haoyu Song, Wei-Nan Zhang, Jingwen Hu, and Ting Liu. 2020. Generating Persona Consistent Dialogues by Exploiting Natural Language Inference. Proceedings of the AAAI Conference on Artificial Intelligence, 34(05):8878-8885. Number: 05. +Kai Sun, Dian Yu, Jianshu Chen, Dong Yu, Yejin Choi, and Claire Cardie. 2019. DREAM: A Challenge Data Set and Models for Dialogue-Based Reading Comprehension. Transactions of the Association for Computational Linguistics, 7:217-231. Place: Cambridge, MA Publisher: MIT Press. +Harsh Trivedi, Heeyoung Kwon, Tushar Khot, Ashish Sabharwal, and Niranjan Balasubramanian. 2019. + +Repurposing Entailment for Multi-Hop Question Answering Tasks. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2948-2958, Minneapolis, Minnesota. Association for Computational Linguistics. + +Johannes Welbl, Nelson F. Liu, and Matt Gardner. 2017. Crowdsourcing Multiple Choice Science Questions. In NUT@EMNLP. + +Adina Williams, Nikita Nangia, and Samuel Bowman. 2018. A broad-coverage challenge corpus for sentence understanding through inference. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1112-1122, New Orleans, Louisiana. Association for Computational Linguistics. + +Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander M. Rush. 2020. HuggingFace's Transformers: State-of-the-art Natural Language Processing. Number: arXiv:1910.03771 arXiv:1910.03771 [cs]. + +# A Answer ranking procedure + +In the multiple choice setting, we performed an answer ranking procedure to pick the answer to a given question among a set of alternative answers $N$ , using both NLI class scores and QA confidence scores. (This is distinct from the selection procedure for the top 20% or 50% of answers we used in both settings.) + +Similar to Harabagiu and Hickl (2006), answers are ranked based on the highest probability from the calibration model $\sigma$ , given a linear combination of the QA or NLI scores given an answer $n \in N$ answer set. When a single feature is used, such as an NLI class or the QA score, no calibration is made and $\sigma$ is simply the identity of the confidence score. In the case of contradiction only, $\sigma$ is the inverse of the contradiction confidence score, indicating the least contradicted answer is being selected. Formally, our approach can be described as: + +$$ +\operatorname * {a r g m a x} _ {N} \sigma (\mathrm {Q A} _ {n}; \mathrm {N L I} _ {n}) +$$ + +where $\mathrm{QA}_n$ is the QA model confidence score for answer $n$ , and $\mathrm{NLI}_n$ represents the various NLI class scores for $n$ . + +We did not use this approach in extractive QA, because we found that asking the model for the top $K = 4$ answer produced almost the same four answer alternatives with slightly different spans each time. + +# B Datasets + +Tables 4 (multiple choice) and 5 (extractive QA) outline the datasets we used. Additional details such as train size and preprocessing steps are available in the references provided. When space doesn't allow CosmosQA is aliased to Cosmos, MCScript to MCS, MCScript-2.0 to MCS2, and MCTest to MCT. The only preprocessing step we performed was to filter out questions where no context passage is provided. Validation splits (as opposed to test splits) are used in the CosmosQA and QASC cases, since context passages or gold standard answers are not available for these datasets. + +# C QA models + +Table 6 outlines the pretrained QA models that we used and the datasets they are trained on. All these models are publicly available on the Hugging Face hub under the locations listed. Where space doesn't allow, RoBERTa-RACE is aliased as RACE. + +We trained the two DeBERTa-v3 models (xsmall and base) as shown in Table 7. They were trained using the Hugging Face trainer API (Wolf et al., 2020) with an Adam optimizer at a learning rate of $5.60\mathrm{e - }05$ with weight decay of 0.01. All models and inference were performed on 1 Tesla P100 GPU. Full instructions on reproducibility as well as trained models are provided in the publicly available code, including directions to weights and biases to inspect the training runs, full parameter set, and evaluation suites. + +# D QA2D models + +A QA2D model reformulates a question-answer pair to a declarative statement (Demszky et al., 2018). As noted in Chen et al. (2021) and Mishra et al. (2021), the QA2D reformulation is critical to using NLI models in QA since the proposed answer needs to match the format of NLI. We trained a T5-small model (Raffel et al., 2020) on the dataset proposed by Demszky et al. (2018) for QA2D since we found almost no noticeable differences in performance in larger models. This used the same setup as the DeBERTa-v3 models xsmall and base (see Table 7). + +
DatasetSplitSizeReference
CosmosQAvalidation2985Huang et al. (2019)
DREAMtest2041Sun et al. (2019)
MCScripttest2797Ostermann et al. (2018)
MCScript-2.0test3610Ostermann et al. (2019)
MCTesttest840Richardson et al. (2013)
QASCvalidation926Khot et al. (2020)
RACEtest4934Lai et al. (2017)
RACE-Ctest712Liang et al. (2019)
SciQtest884Welbl et al. (2017)
+ +Table 4: Datasets used for the multiple choice setting, including split used and sample size. Validation splits were used for CosmosQA since the test split is not publicly available, and for QASC since context passages or gold answers are not available. + +
DatasetSizeReference
BioASQ1504Fisch et al. (2019)
TriviaQA7785
HotpotQA5901
SQuAD10506
Natural Questions12836
SQuAD211871Rajpurkar et al. (2018)
SQuAD-adv5347Jia and Liang (2017)
+ +Table 5: Extractive QA datasets used. Validation sets are used on the SQuAD2.0 and SQuAD adversarial datasets. MRQA 2019 dev sets are used for the other five datasets. + +Unlike Chen et al. (2021), we found that regardless of size, these QA2D models struggled with long questions or questions with complex syntax and would often leave the answer out of the statement. In order to solve this, constrained decoding that required the answer to be in the statement was tried. However, this often produced ungrammatical or nonsensical statements. We settled with the following heuristic to postprocess QA2D outputs: If less than $50\%$ of the tokens in the answer were in the statement then we appended the answer to the end of the statement. $50\%$ was used to account for rephrasing the answer or swapping pronouns. While some statements resulted in answer redundancy, this was better than having hypotheses which left out the answer. + +Future work on QA2D should focus on how these models can be used outside of the domains in the dataset provided by Demszky et al. (2018). Finally it is important to note that erroneous QA2D outputs could effect the quality of the whole pipeline see Chen et al. (2021) for a more detailed analysis of this. + +# E NLI models + +NLI is used to classify whether the reformulated answer is contradicted, entailed, or neutral with respect to a context passage. We used the whole context, as Schuster et al. (2022) and Mishra et al. (2021) demonstrated that long premises still performed adequate though not as well as sentence-length premises. Using the whole context avoids needing to use decontextualization as is required in Chen et al. (2021). + +We used two DeBERTa-based models (He et al., 2021b) trained on the MNLI dataset (Williams et al., 2018) (called mnli-base and mnli-large) and an ALBERT model (Lan et al., 2019) trained on the ANLI dataset in addition to various other NLI datasets (called albert-anli) (Nie et al., 2020). Table 6 contains the Hugging Face references to the NLI models. After inference, the confidence scores are used for answer selection and performance evaluation. + +# E.1 Model size and approach performance analysis + +Table 8 mirrors Table 1 in the main text, but shows the accuracy results for uncalibrated E, C, and $\mathbf{E} + \mathbf{C}$ + +
Hugging FaceName
LIAMF-USP/roberta-large-finetuned-RACERoBERTa-RACE
bert-large-uncased-whole-word-masking-finetuned-squadBERT-Large
distilbert-base-uncased-distilled-squadDistillBERT
ynie/albert-xxlarge-v2-snli_mnli_fever_anli_R1_R2_R3-nlialbert-anli
microsoft/deberta-base-mnlimnli-base
microsoft/deberta-v2-xxlarge-mnlimnli-large
+ +Table 6: Pretrained QA and NLI models used. + +
ModelDatasetEpochsScore
t5-smallDemszky et al. (2018)20Rogue190.73
deberta-v3-xsmallWelbl et al. (2017)6Accuracy93.99
deberta-v3-baseWelbl et al. (2017)6Accuracy91.79
+ +Table 7: The models we trained for or setups with evaluation scores and number of epochs trained. + +
QA ModelCosmosDREAMMCSMCS2MCTQASCRACERACE-CSciQAverage
SciQ-base18.4643.8061.9963.7144.7693.4130.9727.3995.2853.31
SciQ-small25.4648.2660.2866.0459.7690.6035.5630.6298.0957.19
RACE64.2282.5689.7086.9890.4898.1676.9369.8097.9684.09
mnli-large
E+C44.3680.9485.5284.9990.6096.4464.2951.4092.4776.77
E36.1879.0386.0279.7289.8895.9062.1449.7291.9674.50
C59.2678.9883.1284.4389.2992.7662.7447.0591.5876.58
mnli-base
QA + E + C64.3282.6689.6387.0190.7198.2776.9569.8098.0984.16
QA + E64.2582.6689.6386.9890.7198.2776.9569.8097.9684.14
QA + C64.2982.5689.6387.0190.6098.1676.9369.8097.9684.1
E + C33.0362.2776.7672.1168.5792.6645.1634.4188.0163.66
E27.8162.4779.3771.9468.8192.6643.4834.4188.0163.22
C43.4559.1970.1869.9767.5081.8641.8132.5887.3761.55
albert-anli
QA + E + C64.1982.5689.7087.0690.4898.1676.9369.8097.9684.09
QA + E64.1982.5689.7087.0690.6098.1676.9369.8097.9684.11
QA + C64.2282.5689.7086.9890.4898.1676.9369.8097.9684.09
E + C35.7168.2079.5573.8877.5091.7949.0539.4790.8267.33
E33.6768.3579.9173.1977.3891.9049.0739.1990.9467.07
C45.1663.7473.5872.7173.3377.8646.3438.2087.2464.24
+ +Table 8: Accuracy scores in the multiple choice setting for various NLI models used. Calibration was with the RoBERTA-RACE model. + +in the main mnli-large model, as well as the results with the other NLI models, mnli-base and albertanli. Table 9 shows selective QA accuracy in the multiple choice setting where answer selection is done through ranking before we rank answers for selective QA. Selective QA on extractive QA using DistillBERT (table 10) shows that $\mathbf{QA} + \mathbf{E} + \mathbf{C}$ does best in all cases and contradiction only does second best at $50\%$ coverage. + +# F Calibration models + +Like Kamath et al. (2020) and Chen et al. (2021) we developed a set of calibration models in order to perform answer ranking. A calibration model is trained on a set of posterior probabilities from downstream models to predict whether an answer is correct. + +To compare the effect of using different combinations of NLI class confidence scores, we trained a logistic regression model on linear combinations of the following features: QA indicates that the QA model confidence score is being used, $\mathbf{E}$ indicates the entailment score, $\mathbf{C}$ indicates the contradiction score, and $\mathbf{N}$ indicates the neutral score. Like in Chen et al. (2021), all calibration models are trained on a holdout set of 100 samples from a single domain using logistic regression which predicts, given the confidence scores of the downstream models, whether the answer is correct. A multi-domain calibration approach like in Kamath et al. (2020) was not used since the focus was a minimum experiment to test the viability of leveraging different NLI classifications. + +# F.1 Regression Analysis + +To illustrate the characteristics of the calibration models, we present a regression analysis for the multiple choice setting (Table 11). The results indicate that as the mnli model gets larger, the calibration model uses its NLI confidence scores more. Importantly, entailment coefficients are stronger than contradiction coefficients in all cases. + +# G Correlation Analysis + +Since we are using the NLI and QA model scores to construct the setups above, it is useful to know how these factors correlate with the correct answer. Table 13 shows how each NLI class correlates both by score and by actual classification (score $>50\%$ ) as compared against QA model confidence score. The multiple choice analysis shows + +answers from the RoBERTa-RACE model and the extractive QA analysis shows answers from the BERT-large model trained on SQuAD. The correlation analysis presents Spearman rank correlations. + +What we see is that in the multiple choice setting, the confidence score has a strong correlation with the correct answer, which makes sense given the confidence score is a softmax over the multiple choice classes. Extractive QA confidence scores have a much weaker correlation and tend to have less correlation than entailment has with the correct answer. Despite the results presented above, contradiction only has a notable correlation with the correct answer when the score is used rather than the classification. This is a point in favor of our approach of using confidence scores for NLI rather than classifications. + +Interestingly, in the extractive QA case, the neutral class is more negatively correlated when selecting for contradiction when using classification. Our conjecture would be that in the extractive QA case, we don't have much to compare against. When looking at the per dataset correlations for the multiple choice setting (Table 12), we see that in most cases, other than the QA confidence scores, the contradiction scores have the strongest correlations with the correct answer out of any NLI class and neutral, as we would expect, tends to have very weak correlations. We do not present the per dataset correlation for extractive QA as they are very weak, which we again hypothesize comes from having no answers to compare with. + +
DatasetQA+E+CQA+EQA+CE+CECQA
20%CosmosQA77.5567.1783.2520.1027.4767.5088.61
DREAM98.2896.3296.8181.1391.9193.8798.28
MCScript99.8299.6499.4693.0298.9396.9699.82
MCScript-2.099.5899.0397.3792.2497.3795.0199.58
MCTest10010099.4085.1297.0297.0298.81
QASC10010010097.3010099.46100
RACE94.9392.1390.1762.7376.7175.0598.24
RACE-C88.7385.2186.6271.1374.6569.0193.66
SciQ10010010082.0510096.15100
Avg95.4393.2894.7976.0984.9087.7897.45
50%CosmosQA80.2970.7880.7032.1734.7264.8876.47
DREAM95.1093.6393.6385.2089.4188.3396.67
MCScript98.5797.8597.1494.7195.9992.7098.78
MCScript-2.096.4094.4696.0791.0291.7591.6998.01
MCTest99.5298.8199.7691.4395.2496.1999.52
QASC10099.7899.7898.2798.7098.49100
RACE90.1187.2285.2367.8971.7068.1893.88
RACE-C85.1178.0977.2566.5766.8555.0687.36
SciQ10010099.7489.0396.4396.43100
Avg93.9091.1892.1479.5982.3183.5594.52
+ +Table 9: Selective QA accuracies in the multiple choice setting where answer selection is done through ranking before we rank answers for selective QA. + +
DatasetQA+E+CQA+EQA+CE+CECQA
20%BioASQ70.9770.4171.5574.0774.0774.3468.99
HotpotQA73.4473.0870.8871.5971.5170.4169.41
Natural Questions85.5985.2985.4578.4678.4680.5383.27
SQuAD96.2296.4595.7783.1583.0981.3797.15
SQuAD-adv40.3939.7539.4940.0739.5640.5931.98
SQuAD235.4635.2433.6436.3636.1336.6625.95
TriviaQA64.9664.6864.5552.6752.0952.5663.98
Avg66.7266.4165.9062.3462.1362.3562.96
50%BioASQ65.9665.9264.3763.5363.5366.9564.79
HotpotQA64.4264.2163.6565.8865.8566.9162.81
Natural Questions72.2871.9970.8267.5467.5174.1869.95
SQuAD92.5692.5792.3481.8682.2180.9592.54
SQuAD-adv33.6932.9033.4538.7438.2238.5231.89
SQuAD226.6825.7026.0032.9532.6132.8323.52
TriviaQA58.4058.4158.2551.4351.1852.9958.25
Avg59.1458.8158.4157.4257.3059.0557.68
+ +Table 10: SelectiveQA on extractive QA using DistillBERT. Note that QA+E+C does best in all cases and contradiction only does second best at $50\%$ coverage. + +
QA ModelNLI ModelCombinationConfidenceEntailmentContradictionAcc
SciQmnli-baseQA + C4.13-1.060.99
QA + E3.901.370.99
QA + E + C3.831.22-0.760.99
E + C2.56-1.470.86
mnli-largeQA + C3.98-1.320.99
QA + E3.781.550.99
QA + E + C3.651.31-0.970.99
E + C2.63-1.720.91
RACEmnli-baseQA + C3.04-0.150.89
QA + E3.030.270.89
QA + E + C3.020.26-0.140.89
E + C0.73-0.460.75
mnli-largeQA + C2.970.00-0.810.89
QA + E2.910.980.89
QA + E + C2.850.92-0.750.89
E + C1.76-1.120.78
+ +Table 11: Regression analysis for each mnli-based nli model with each QA model used calibration with logistic regression multiple choice settings. Accuracy is the evaluation metric used. + +
ContradictionEntailmentNeutral
DatasetQAScoreClassScoreClassScoreClass
CosmosQA0.53-0.34-0.170.05-0.010.210.16
DREAM0.72-0.57-0.350.540.50-0.11-0.13
MCScript0.80-0.59-0.420.590.50-0.04-0.08
MCScript20.77-0.50-0.320.410.37-0.04-0.05
MCTest0.73-0.65-0.470.640.69-0.20-0.15
QASC0.57-0.54-0.280.550.67-0.50-0.26
RACE0.65-0.37-0.200.350.34-0.11-0.11
RACE-C0.59-0.24-0.130.180.25-0.09-0.11
SciQ0.75-0.69-0.470.680.67-0.42-0.19
+ +Table 12: Correlation analysis (Spearman rank correlation) per dataset in the multiple choice setting. RoBERTaRACE is used for the QA scores. + +
ContradictionEntailmentNeutralQA
multiple choiceScore-0.470.37-0.060.71
Class-0.280.38-0.06
extractive QAScore-0.160.31-0.120.19
Class-0.150.39-0.29
+ +Table 13: Correlation analysis (Spearman rank correlation) in the multiple choice and extractive QA settings. RoBERTa-RACE is the QA model used for multiple choice QA scores and BERT-large is used for the extractive QA scores. + +# A For every submission: + +A1. Did you describe the limitations of your work? + +Left blank. + +A2. Did you discuss any potential risks of your work? + +Left blank. + +A3. Do the abstract and introduction summarize the paper's main claims? + +Left blank. + +□ A4. Have you used AI writing assistants when working on this paper? + +Left blank. + +# B Did you use or create scientific artifacts? + +Left blank. + +B1. Did you cite the creators of artifacts you used? + +Left blank. + +B2. Did you discuss the license or terms for use and / or distribution of any artifacts? + +Left blank. + +B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? + +Left blank. + +B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? + +Left blank. + +B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? + +Left blank. + +B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. + +Left blank. + +# C Did you run computational experiments? + +Left blank. + +C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? + +Left blank. + +The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance. + +C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Left blank. +C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Left blank. +C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Left blank. + +# D Did you use human annotators (e.g., crowdworkers) or research with human participants? + +Left blank. + +D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Left blank. +D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Left blank. +D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Left blank. +D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Left blank. +D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Left blank. \ No newline at end of file diff --git a/2023/Using contradictions improves question answering systems/images.zip b/2023/Using contradictions improves question answering systems/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3edf3a3e8bbaec2f6cb710ee79951936fc8cb584 --- /dev/null +++ b/2023/Using contradictions improves question answering systems/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a328fa8ed09e5fc36cfa4b56cbd01b2ae9d50eb8a93aa920390f61bfa52e0443 +size 1103959 diff --git a/2023/Using contradictions improves question answering systems/layout.json b/2023/Using contradictions improves question answering systems/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d815e13a8ceee2ffa806140344bc98fd81767b30 --- /dev/null +++ b/2023/Using contradictions improves question answering systems/layout.json @@ -0,0 +1,8055 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 116, + 76, + 478, + 94 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 76, + 478, + 94 + ], + "spans": [ + { + "bbox": [ + 116, + 76, + 478, + 94 + ], + "type": "text", + "content": "Using contradictions improves question answering systems" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 139, + 119, + 261, + 134 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 119, + 261, + 134 + ], + "spans": [ + { + "bbox": [ + 139, + 119, + 261, + 134 + ], + "type": "text", + "content": "Étienne Fortier-Dubois" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 355, + 121, + 439, + 133 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 121, + 439, + 133 + ], + "spans": [ + { + "bbox": [ + 355, + 121, + 439, + 133 + ], + "type": "text", + "content": "Domenic Rosati" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 379, + 137, + 414, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 137, + 414, + 148 + ], + "spans": [ + { + "bbox": [ + 379, + 137, + 414, + 148 + ], + "type": "text", + "content": "scite.ai" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 345, + 150, + 448, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 150, + 448, + 163 + ], + "spans": [ + { + "bbox": [ + 345, + 150, + 448, + 163 + ], + "type": "text", + "content": "Dalhousie University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 84, + 232, + 274, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 232, + 274, + 435 + ], + "spans": [ + { + "bbox": [ + 84, + 232, + 274, + 435 + ], + "type": "text", + "content": "This work examines the use of contradiction in natural language inference (NLI) for question answering (QA). Typically, NLI systems help answer questions by determining if a potential answer is entailed (supported) by some background context. But is it useful to also determine if an answer contradicts the context? We test this in two settings, multiple choice and extractive QA, and find that systems that incorporate contradiction can do slightly better than entailment-only systems on certain datasets. However, the best performances come from using contradiction, entailment, and QA model confidence scores together. This has implications for the deployment of QA systems in domains such as medicine and science where safety is an issue." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 68, + 441, + 155, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 441, + 155, + 454 + ], + "spans": [ + { + "bbox": [ + 68, + 441, + 155, + 454 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 462, + 291, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 462, + 291, + 597 + ], + "spans": [ + { + "bbox": [ + 67, + 462, + 291, + 597 + ], + "type": "text", + "content": "Safety in NLP systems is unresolved, particularly in biomedical and scientific contexts where hallucination, overconfidence, and other problems are major obstacles to deployment (Ji et al., 2022; Kell et al., 2021). One active area of research to solve these issues is natural language inference (NLI) (Li et al., 2022). NLI is the task of determining whether a hypothesis is true (entailed), false (contradicted), or undetermined (neutral) given some premise." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 597, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 597, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 597, + 291, + 773 + ], + "type": "text", + "content": "Current NLI systems typically focus only on entailment to verify hypotheses—they calculate the degree to which a hypothesis is supported by the premise. But the premise can provide another signal: contradiction. Regardless of how well a hypothesis is entailed by the context, it can also be more or less contradicted, which could affect whether it is accepted or rejected. Contradictions are an important signal indicating whether some statement might be unacceptable given a premise. In some cases where we might not know if a statement is supported, we should still ensure we are rejecting statements that are outright contradicted." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 304, + 210, + 526, + 334 + ], + "blocks": [ + { + "bbox": [ + 304, + 210, + 526, + 334 + ], + "lines": [ + { + "bbox": [ + 304, + 210, + 526, + 334 + ], + "spans": [ + { + "bbox": [ + 304, + 210, + 526, + 334 + ], + "type": "image", + "image_path": "0782e5fe3e33d197559d16540fd53b83aff03fb7da05f392f5edf22449885b76.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 343, + 525, + 391 + ], + "lines": [ + { + "bbox": [ + 302, + 343, + 525, + 391 + ], + "spans": [ + { + "bbox": [ + 302, + 343, + 525, + 391 + ], + "type": "text", + "content": "Figure 1: A QA model is used to produce answers which are reformulated as hypotheses to determine if they are entailed or contradicted by a premise. The answers are ranked by NLI class scores to select the best answer." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 415, + 526, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 415, + 526, + 591 + ], + "spans": [ + { + "bbox": [ + 302, + 415, + 526, + 591 + ], + "type": "text", + "content": "We wondered if adding this signal to a question answering (QA) system might improve performance and safety. To this end, we propose a method that reformulates answers from the QA system as hypotheses for NLI, calculates the entailment, contradiction, and neutrality of each hypothesis, and then selects the best one based on a combination of these results (Figure 1). We show that across 16 QA datasets (9 multiple choice and 7 extractive), the best approach is to use entailment, contradiction, and confidence scores together. Using only contradiction is roughly on par with, and sometimes better than, using only entailment." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 303, + 603, + 394, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 603, + 394, + 616 + ], + "spans": [ + { + "bbox": [ + 303, + 603, + 394, + 616 + ], + "type": "text", + "content": "1.1 Related work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 624, + 525, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 624, + 525, + 663 + ], + "spans": [ + { + "bbox": [ + 302, + 624, + 525, + 663 + ], + "type": "text", + "content": "NLI for question answering has been explored by several authors in various settings; see Paramasi-vam and Nirmala (2021) for an overview." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 665, + 526, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 665, + 526, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 665, + 526, + 773 + ], + "type": "text", + "content": "One of these settings is selective question answering for extractive QA, where selective refers to abstention when the system is not confident enough in its answer (Kamath et al., 2020). Chen et al. (2021) have found that NLI systems are able to verify the predictions made by a QA system in this setting, but their result is limited to only selecting a top " + }, + { + "bbox": [ + 302, + 665, + 526, + 773 + ], + "type": "inline_equation", + "content": "k\\%" + }, + { + "bbox": [ + 302, + 665, + 526, + 773 + ], + "type": "text", + "content": " of answers. Moreover, they" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "text", + "content": "827" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 135, + 795, + 459, + 805 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 795, + 459, + 805 + ], + "spans": [ + { + "bbox": [ + 135, + 795, + 459, + 805 + ], + "type": "text", + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 224, + 806, + 370, + 817 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 806, + 370, + 817 + ], + "spans": [ + { + "bbox": [ + 224, + 806, + 370, + 817 + ], + "type": "text", + "content": "Volume 2: Short Papers, pages 827-840" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 176, + 818, + 417, + 828 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 818, + 417, + 828 + ], + "spans": [ + { + "bbox": [ + 176, + 818, + 417, + 828 + ], + "type": "text", + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 290, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 290, + 125 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 290, + 125 + ], + "type": "text", + "content": "do not provide an approach for improving overall performance, nor do they show the effect of incorporating contradiction directly (but do so indirectly by analyzing non-entailed passages)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 126, + 291, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 126, + 291, + 301 + ], + "spans": [ + { + "bbox": [ + 69, + 126, + 291, + 301 + ], + "type": "text", + "content": "In the related setting of multiple choice QA and fact checking, Mishra et al. (2021) have explored the use of entailment, finding that NLI models do well at these tasks by themselves, but can perform even better when they are adapted to in-domain data and longer premises. Yet their method uses only a two-class NLI set up (entailed or not entailed), which doesn't tell us much about directly using the contradiction signal. Pujari and Goldwasser (2019) does incorporate the contradiction signal showing the power of contradiction to improve machine comprehension but does not analyze its effects separately from entailment." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 302, + 290, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 302, + 290, + 383 + ], + "spans": [ + { + "bbox": [ + 67, + 302, + 290, + 383 + ], + "type": "text", + "content": "Other QA settings in which NLI has been used include open domain (Harabagiu and Hickl, 2006) and multi-hop (Trivedi et al., 2019). Thus far, approaches tend to focus on entailment. To our knowledge, our work is the first to directly assess using contradictions for QA isolated from entailment." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 385, + 290, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 385, + 290, + 519 + ], + "spans": [ + { + "bbox": [ + 67, + 385, + 290, + 519 + ], + "type": "text", + "content": "Outside of question answering, a domain that uses contradictions is factual consistency—the task of ensuring that a collection of utterances is faithful to a source document. Li et al. (2022) provide an overview. Typically, entailment is still the main focus, but Laban et al. (2022) propose an NLI-based method to ensure the consistency of a summary with a source document using contradiction and neutral scores in addition to entailment, beating out previous systems." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 521, + 290, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 521, + 290, + 587 + ], + "spans": [ + { + "bbox": [ + 67, + 521, + 290, + 587 + ], + "type": "text", + "content": "Other researchers have used contradictions to identify consistency errors across Wikipedia (Schuster et al., 2022; Hsu et al., 2021) or generate credible character dialogue (Nie et al., 2021; Song et al., 2020)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 600, + 133, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 600, + 133, + 613 + ], + "spans": [ + { + "bbox": [ + 67, + 600, + 133, + 613 + ], + "type": "text", + "content": "2 Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 624, + 290, + 731 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 624, + 290, + 731 + ], + "spans": [ + { + "bbox": [ + 67, + 624, + 290, + 731 + ], + "type": "text", + "content": "We tested the effect of contradictions in two QA settings and a total of sixteen question-answer datasets. Our approach is broadly similar to both Chen et al. (2021) and Mishra et al. (2021) in that we use most of the same datasets for evaluating NLI reranking for multiple choice QA and extractive QA. Unlike both, we incorporate contradiction directly as a signal for reranking answers." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 733, + 290, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 733, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 733, + 290, + 772 + ], + "type": "text", + "content": "Briefly, for each dataset, we used pretrained QA models to produce answers and confidence scores for the dataset's questions. We refer to the confi" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 71, + 526, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 191 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 191 + ], + "type": "text", + "content": "dence scores below as QA. We then trained QA2D models (where QA2D stands for \"question-answer to declarative\") to turn the answers into the declarative hypothesis format required for NLI. For example, the question-answer pair \"What is the most abundant metal in the Earth crust? Copper.\" might be rephrased as \"The most abundant metal in the Earth crust is copper\" (see Appendix D for more details)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 194, + 526, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 194, + 526, + 301 + ], + "spans": [ + { + "bbox": [ + 302, + 194, + 526, + 301 + ], + "type": "text", + "content": "With the question contexts as premises, we then used NLI models to classify every premise-hypothesis pair into three classes, each with an associated score: entailed (E), contradicted (C), and neutral (N). After that, we trained logistic regression calibration models to find which linear combination of the four scores—QA, E, C, and N—was best able to pick the answers accurately." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 303, + 526, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 303, + 526, + 465 + ], + "spans": [ + { + "bbox": [ + 302, + 303, + 526, + 465 + ], + "type": "text", + "content": "When evaluating performance, we applied the selective QA approach from Kamath et al. (2020) to rank answers using combinations of the four scores, and then consider only those that the model was most confident in answering. We compared selecting the top " + }, + { + "bbox": [ + 302, + 303, + 526, + 465 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 302, + 303, + 526, + 465 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 302, + 303, + 526, + 465 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 302, + 303, + 526, + 465 + ], + "type": "text", + "content": ". In the multiple choice setting, it was also possible to rank all potential answers according to the four scores, unlike in the extractive QA setting where the QA model produced only one answer per question, so we evaluated performance with that approach as well (see appendix A for details)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 477, + 431, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 477, + 431, + 491 + ], + "spans": [ + { + "bbox": [ + 302, + 477, + 431, + 491 + ], + "type": "text", + "content": "3 Experimental setting" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 502, + 525, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 502, + 525, + 608 + ], + "spans": [ + { + "bbox": [ + 302, + 502, + 525, + 608 + ], + "type": "text", + "content": "In the multiple choice setting, we tested 9 datasets. Two of them are in-domain, since the pretrained QA models we used were finetuned on them. Specifically, we used a RoBERTa large model (Liu et al., 2019) finetuned on the RACE dataset (Lai et al., 2017), as well as two DeBERTa v3 variants, base and xsmall (He et al., 2021a), finetuned on the SciQ dataset (Welbl et al., 2017)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 610, + 525, + 744 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 610, + 525, + 744 + ], + "spans": [ + { + "bbox": [ + 302, + 610, + 525, + 744 + ], + "type": "text", + "content": "In the extractive QA setting, we used 7 datasets: five from the MRQA 2019 task (Fisch et al., 2019), as well as SQuAD 2.0 (Rajpurkar et al., 2018) and SQuAD adversarial (Jia and Liang, 2017). The SQuAD model is the in-domain dataset: it was used to pretrain (Rajpurkar et al., 2016) the two QA models we used, DistillBERT (Sanh et al., 2020) and BERT-Large (Devlin et al., 2019). Like Chen et al. (2021), we used the Natural Questions dataset for calibration." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "type": "text", + "content": "In both settings, all datasets contain the relevant context that can be used by the QA models to select" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 790 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 790 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 790 + ], + "type": "text", + "content": "828" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 291, + 98 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 291, + 98 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 291, + 98 + ], + "type": "text", + "content": "answers. More detail on the datasets and QA models is available in appendices B and C respectively." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 99, + 291, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 99, + 291, + 206 + ], + "spans": [ + { + "bbox": [ + 67, + 99, + 291, + 206 + ], + "type": "text", + "content": "See appendices D, E, and F for details on the QA2D, NLI, and calibration models. Our models follow the setups described in Kamath et al. (2020), Chen et al. (2021), and Mishra et al. (2021). The main interesting detail is that the calibration models were trained on a holdout set of 100 samples from a single domain, using logistic regression, as in Chen et al. (2021)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 217, + 127, + 230 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 217, + 127, + 230 + ], + "spans": [ + { + "bbox": [ + 67, + 217, + 127, + 230 + ], + "type": "text", + "content": "4 Results" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 239, + 202, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 239, + 202, + 253 + ], + "spans": [ + { + "bbox": [ + 67, + 239, + 202, + 253 + ], + "type": "text", + "content": "4.1 Multiple choice setting" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 257, + 291, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 257, + 291, + 406 + ], + "spans": [ + { + "bbox": [ + 67, + 257, + 291, + 406 + ], + "type": "text", + "content": "For most multiple choice datasets, the best accuracy—when ranking all potential answers—is attained when using a calibrated model combining QA confidence, entailment, and contradiction (QA+E+C in Table 1). Only for the in-domain case (RACE-C) does the uncalibrated RoBERTa-RACE model perform on par with that. Using QA scores combined with either entailment (QA+E) or contradiction (QA+C) achieves similar performance, with contradiction winning by a small margin: " + }, + { + "bbox": [ + 67, + 257, + 291, + 406 + ], + "type": "inline_equation", + "content": "84.33\\%" + }, + { + "bbox": [ + 67, + 257, + 291, + 406 + ], + "type": "text", + "content": " average accuracy compared to " + }, + { + "bbox": [ + 67, + 257, + 291, + 406 + ], + "type": "inline_equation", + "content": "84.31\\%" + }, + { + "bbox": [ + 67, + 257, + 291, + 406 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 407, + 291, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 407, + 291, + 514 + ], + "spans": [ + { + "bbox": [ + 67, + 407, + 291, + 514 + ], + "type": "text", + "content": "To inspect these trends further, we performed a correlation analysis of the NLI classes and QA confidence scores with the correct answer (appendix G). We found that besides QA confidence, it is the contradiction score that has the strongest correlation with the correct answer. The analysis also showed that the neutral class score (N) had almost no effect, which is why it is omitted in all results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 515, + 291, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 515, + 291, + 609 + ], + "spans": [ + { + "bbox": [ + 67, + 515, + 291, + 609 + ], + "type": "text", + "content": "When using the selective QA approach and evaluating only the " + }, + { + "bbox": [ + 67, + 515, + 291, + 609 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 67, + 515, + 291, + 609 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 67, + 515, + 291, + 609 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 67, + 515, + 291, + 609 + ], + "type": "text", + "content": " most confident answers, the best performance is attained with the " + }, + { + "bbox": [ + 67, + 515, + 291, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{QA} + \\mathbf{C}" + }, + { + "bbox": [ + 67, + 515, + 291, + 609 + ], + "type": "text", + "content": " combination (Table 2). This model is the only one that beats just using the QA confidence score on average. It is stronger than " + }, + { + "bbox": [ + 67, + 515, + 291, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{QA} + \\mathbf{E} + \\mathbf{C}" + }, + { + "bbox": [ + 67, + 515, + 291, + 609 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 515, + 291, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{QA} + \\mathbf{E}" + }, + { + "bbox": [ + 67, + 515, + 291, + 609 + ], + "type": "text", + "content": " for both coverage percentages." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 610, + 291, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 610, + 291, + 704 + ], + "spans": [ + { + "bbox": [ + 67, + 610, + 291, + 704 + ], + "type": "text", + "content": "Contradiction alone, without QA confidence scores (C), also beats both entailment alone (E) and entailment with contradiction " + }, + { + "bbox": [ + 67, + 610, + 291, + 704 + ], + "type": "inline_equation", + "content": "(\\mathbf{E} + \\mathbf{C})" + }, + { + "bbox": [ + 67, + 610, + 291, + 704 + ], + "type": "text", + "content": " for both coverages. These results match our intuition that the less contradicted an answer, the more likely it is correct, even in cases where there is uncertainty about its entailment." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 714, + 198, + 729 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 714, + 198, + 729 + ], + "spans": [ + { + "bbox": [ + 67, + 714, + 198, + 729 + ], + "type": "text", + "content": "4.2 Extractive QA setting" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 733, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 733, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 733, + 291, + 773 + ], + "type": "text", + "content": "Similar results occur when evaluating the extractive QA datasets with " + }, + { + "bbox": [ + 67, + 733, + 291, + 773 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 67, + 733, + 291, + 773 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 733, + 291, + 773 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 67, + 733, + 291, + 773 + ], + "type": "text", + "content": " selective coverage (Table 3). The " + }, + { + "bbox": [ + 67, + 733, + 291, + 773 + ], + "type": "inline_equation", + "content": "\\mathbf{QA} + \\mathbf{C}" + }, + { + "bbox": [ + 67, + 733, + 291, + 773 + ], + "type": "text", + "content": " model does better than QA" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 71, + 526, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 153 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 153 + ], + "type": "text", + "content": "alone, and " + }, + { + "bbox": [ + 302, + 71, + 526, + 153 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 302, + 71, + 526, + 153 + ], + "type": "text", + "content": " alone does better than " + }, + { + "bbox": [ + 302, + 71, + 526, + 153 + ], + "type": "inline_equation", + "content": "\\mathbf{E} + \\mathbf{C}" + }, + { + "bbox": [ + 302, + 71, + 526, + 153 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 302, + 71, + 526, + 153 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 302, + 71, + 526, + 153 + ], + "type": "text", + "content": " alone, indicating the importance of the contradiction signal here too. However, entailment seems to matter more for extractive QA, as the best F1 score overall was from " + }, + { + "bbox": [ + 302, + 71, + 526, + 153 + ], + "type": "inline_equation", + "content": "\\mathbf{QA} + \\mathbf{E}" + }, + { + "bbox": [ + 302, + 71, + 526, + 153 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 302, + 71, + 526, + 153 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 302, + 71, + 526, + 153 + ], + "type": "text", + "content": " coverage case, and " + }, + { + "bbox": [ + 302, + 71, + 526, + 153 + ], + "type": "inline_equation", + "content": "\\mathbf{QA} + \\mathbf{E} + \\mathbf{C}" + }, + { + "bbox": [ + 302, + 71, + 526, + 153 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 302, + 71, + 526, + 153 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 302, + 71, + 526, + 153 + ], + "type": "text", + "content": " case." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 163, + 379, + 176 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 163, + 379, + 176 + ], + "spans": [ + { + "bbox": [ + 302, + 163, + 379, + 176 + ], + "type": "text", + "content": "5 Discussion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 185, + 526, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 185, + 526, + 415 + ], + "spans": [ + { + "bbox": [ + 302, + 185, + 526, + 415 + ], + "type": "text", + "content": "Contradiction with background context is a useful signal that NLP systems can use to infer answers to questions. This is not necessarily a superior strategy to using entailment, but our results show that combining these two signals can improve performance beyond what QA models can achieve on their own. These results are interesting because using contradictions comes with potential benefits for the safety of NLP systems and, as a result, their deployment in domains such as medicine or science. Namely, that there are many potential cases where we are not sure if a statement is entailed directly by a background context but we may be sure that the statement is not refuted by a background context. In two-class NLI settings where we focus only on entailment, neutral and contradiction are collapsed together and we don't have this guarantee." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 426, + 384, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 426, + 384, + 439 + ], + "spans": [ + { + "bbox": [ + 302, + 426, + 384, + 439 + ], + "type": "text", + "content": "6 Limitations" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 449, + 526, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 449, + 526, + 597 + ], + "spans": [ + { + "bbox": [ + 302, + 449, + 526, + 597 + ], + "type": "text", + "content": "Our work comes with some limitations. It is uncertain whether our results in two specific settings, multiple choice and extractive QA, would extend to more general settings for NLI, although the use of contradictions for factual consistency by Laban et al. (2022) suggests that they could. Additionally, 3-class NLI is not sufficient to capture all the natural language relations that might be needed to verify an answer. As such more challenging datasets in other settings and more granular NLI settings should be attempted." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 597, + 526, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 597, + 526, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 597, + 526, + 773 + ], + "type": "text", + "content": "Another limitation involves answer ranking and the associated computational cost. The main reason we did not test answer ranking in extractive QA is that we did not generate diverse outputs, but another reason is that such a procedure grows prohibitively expensive as the domain becomes more open. In a fully open domain, ranking would require a quadratic evaluation for each context passage against each reformulated answer candidate (Schuster et al., 2022). Future work should look at comparison approaches that amortize this cost, such as NLI-based dense passage retrieval (Reimers and Gurevych, 2019)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "829" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 76, + 78, + 517, + 152 + ], + "blocks": [ + { + "bbox": [ + 76, + 78, + 517, + 152 + ], + "lines": [ + { + "bbox": [ + 76, + 78, + 517, + 152 + ], + "spans": [ + { + "bbox": [ + 76, + 78, + 517, + 152 + ], + "type": "table", + "html": "
QA ModelCosmosDREAMMCSMCS2MCTQASCRACERACE-CSciQAverage
SciQ-base18.4643.8061.9963.7144.7693.4130.9727.3995.2853.30
SciQ-small25.4648.2660.2866.0459.7690.6035.5630.6298.0957.18
QA64.2282.5689.7086.9890.4898.1676.9369.8097.9684.08
QA+E+C64.72*83.19*90.06*87.59*91.43*98.6077.53*69.80*98.2184.57
QA+E64.3282.85*89.92*87.29*91.0798.49*77.1869.6698.0984.31
QA+C64.8282.75*89.88*87.29*90.8398.3877.1669.8098.0984.33
", + "image_path": "08e65eb7bd5e6d98700618a8fa6707566b577a331716d35e543b9e2ed6369cfb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 126, + 257, + 468, + 471 + ], + "blocks": [ + { + "bbox": [ + 67, + 160, + 526, + 233 + ], + "lines": [ + { + "bbox": [ + 67, + 160, + 526, + 233 + ], + "spans": [ + { + "bbox": [ + 67, + 160, + 526, + 233 + ], + "type": "text", + "content": "Table 1: Multiple choice setting. Accuracy scores (best per column in bold, second best underlined, statistical significance (pairwise students t-test) is indicated by asterix) after answer ranking with the mnli-large NLI model. The top three rows show the accuracy of using only the QA models' confidence score; \"QA\" refers to the scores of the RoBERTa-RACE model, which was used for calibration. The bottom rows add the entailment and/or contradiction scores to the RoBERTa-RACE score. For other NLI models, and for just E, C, and " + }, + { + "bbox": [ + 67, + 160, + 526, + 233 + ], + "type": "inline_equation", + "content": "\\mathrm{E + C}" + }, + { + "bbox": [ + 67, + 160, + 526, + 233 + ], + "type": "text", + "content": " without calibration with RoBERTa-RACE, see Table 8 in the appendix." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 126, + 257, + 468, + 471 + ], + "lines": [ + { + "bbox": [ + 126, + 257, + 468, + 471 + ], + "spans": [ + { + "bbox": [ + 126, + 257, + 468, + 471 + ], + "type": "table", + "html": "
DatasetQA+E+CQA+CQA+EE+CECQA
20%CosmosQA77.5591.1276.8869.1868.3483.2588.61
DREAM98.2898.7798.2896.3296.3296.8198.28
MCScript99.8299.4699.8299.6499.6499.4699.82
MCScript-2.099.5899.7299.4599.1799.0397.3799.58
MCTest10099.4010010010099.4098.81
QASC100100100100100100100
RACE94.9396.6994.7292.4492.2490.1798.24
RACE-C88.7392.9689.4485.2185.9286.6293.66
SciQ100100100100100100100
Average95.4397.5795.4093.5593.5094.7997.45
50%CosmosQA80.2981.7076.9475.8070.6480.6376.47
DREAM95.1096.8694.9093.6393.6393.6396.67
MCScript98.5798.6498.2898.0097.9397.1498.78
MCScript-2.096.4098.2395.8494.6894.4096.0198.01
MCTest99.5299.7699.5299.0599.0599.7699.52
QASC10010010099.7899.7899.78100
RACE90.1192.6889.9987.7187.3885.2393.88
RACE-C85.1184.8385.3978.3778.3777.2587.36
SciQ10010010010010099.74100
Average93.9094.7493.4391.8991.2492.1394.52
", + "image_path": "da55983854c26153070ce30967fc533987e904bda5a2d352b0f691d187b60e8a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 120, + 541, + 474, + 715 + ], + "blocks": [ + { + "bbox": [ + 67, + 480, + 525, + 515 + ], + "lines": [ + { + "bbox": [ + 67, + 480, + 525, + 515 + ], + "spans": [ + { + "bbox": [ + 67, + 480, + 525, + 515 + ], + "type": "text", + "content": "Table 2: Multiple choice setting. Accuracy scores (best per row in bold, second best underlined) for selective QA with " + }, + { + "bbox": [ + 67, + 480, + 525, + 515 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 67, + 480, + 525, + 515 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 480, + 525, + 515 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 67, + 480, + 525, + 515 + ], + "type": "text", + "content": " coverage of the dataset. Calibrations and QA confidence are all from RoBERTa-RACE, where RACE is the in-domain dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 120, + 541, + 474, + 715 + ], + "lines": [ + { + "bbox": [ + 120, + 541, + 474, + 715 + ], + "spans": [ + { + "bbox": [ + 120, + 541, + 474, + 715 + ], + "type": "table", + "html": "
DatasetQA+E+CQA+CQA+EE+CECQA
20%BioASQ85.0483.1085.0674.2274.2275.4782.99
HotpotQA86.6285.8986.6980.6080.6079.8285.33
Natural Questions91.8492.1891.6879.8979.8782.0990.98
SQuAD98.2698.7692.3798.1792.4890.8899.04
SQuAD-adv43.9943.5743.9843.7443.6042.8139.83
SQuAD237.6436.0737.5637.4337.3137.6830.52
TriviaQA81.3380.3681.2165.5365.2569.1380.68
Average74.9674.1974.9967.6867.6268.2772.77
50%BioASQ76.1375.5176.0471.4971.4972.9775.49
HotpotQA79.3778.9579.3077.4377.4377.3178.74
Natural Questions84.5383.2484.4874.9674.9378.6282.47
SQuAD96.9897.0196.9791.5891.5291.1997.00
SQuAD-adv41.8041.4941.1642.7642.7942.0340.26
SQuAD229.4128.7728.4534.4334.1434.3926.18
TriviaQA74.3074.2374.3765.0564.9368.0874.21
Average68.9368.4668.6865.3965.3266.3767.76
", + "image_path": "50157353668011eb304df5c595fe8b61a7f01acc31185099d156a6efa8867692.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 724, + 525, + 759 + ], + "lines": [ + { + "bbox": [ + 67, + 724, + 525, + 759 + ], + "spans": [ + { + "bbox": [ + 67, + 724, + 525, + 759 + ], + "type": "text", + "content": "Table 3: Extractive QA setting. F1 scores (best per row in bold, second best underlined) for selective QA with " + }, + { + "bbox": [ + 67, + 724, + 525, + 759 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 67, + 724, + 525, + 759 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 724, + 525, + 759 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 67, + 724, + 525, + 759 + ], + "type": "text", + "content": " coverage of the dataset. Calibrations and QA confidence are from the BERT-large model, where SQuAD is the in-domain dataset. For similar results on the smaller DistillBERT model, see Table 10 in the appendix." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "830" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 71, + 127, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 71, + 127, + 83 + ], + "spans": [ + { + "bbox": [ + 69, + 71, + 127, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 90, + 290, + 772 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 69, + 90, + 290, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 90, + 290, + 157 + ], + "spans": [ + { + "bbox": [ + 69, + 90, + 290, + 157 + ], + "type": "text", + "content": "Jifan Chen, Eunsol Choi, and Greg Durrett. 2021. Can NLI Models Verify QA Systems' Predictions? In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 3841-3854, Punta Cana, Dominican Republic. Association for Computational Linguistics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 166, + 290, + 222 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 166, + 290, + 222 + ], + "spans": [ + { + "bbox": [ + 69, + 166, + 290, + 222 + ], + "type": "text", + "content": "Dorottya Demszky, Kelvin Guu, and Percy Liang. 2018. Transforming Question Answering Datasets Into Natural Language Inference Datasets. Technical Report arXiv:1809.02922, arXiv. ArXiv:1809.02922 [cs] type: article." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 230, + 290, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 230, + 290, + 330 + ], + "spans": [ + { + "bbox": [ + 69, + 230, + 290, + 330 + ], + "type": "text", + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 338, + 290, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 338, + 290, + 416 + ], + "spans": [ + { + "bbox": [ + 69, + 338, + 290, + 416 + ], + "type": "text", + "content": "Adam Fisch, Alon Talmor, Robin Jia, Minjoon Seo, Eunsol Choi, and Danqi Chen. 2019. MRQA 2019 Shared Task: Evaluating Generalization in Reading Comprehension. In Proceedings of the 2nd Workshop on Machine Reading for Question Answering, pages 1-13, Hong Kong, China. Association for Computational Linguistics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 425, + 290, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 425, + 290, + 503 + ], + "spans": [ + { + "bbox": [ + 69, + 425, + 290, + 503 + ], + "type": "text", + "content": "Sanda Harabagiu and Andrew Hickl. 2006. Methods for Using Textual Entailment in Open-Domain Question Answering. In Proceedings of the 21st International Conference on Computational Linguistics and 44th Annual Meeting of the Association for Computational Linguistics, pages 905-912, Sydney, Australia. Association for Computational Linguistics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 512, + 290, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 512, + 290, + 567 + ], + "spans": [ + { + "bbox": [ + 69, + 512, + 290, + 567 + ], + "type": "text", + "content": "Pengcheng He, Jianfeng Gao, and Weizhu Chen. 2021a. DeBERTaV3: Improving DeBERTa using ELECTRA-Style Pre-Training with Gradient-Disentangled Embedding Sharing. Number: arXiv:2111.09543 arXiv:2111.09543 [cs]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 576, + 290, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 576, + 290, + 621 + ], + "spans": [ + { + "bbox": [ + 69, + 576, + 290, + 621 + ], + "type": "text", + "content": "Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. 2021b. DeBERTa: Decoding-enhanced BERT with Disentangled Attention. Number: arXiv:2006.03654 arXiv:2006.03654 [cs]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 630, + 290, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 630, + 290, + 686 + ], + "spans": [ + { + "bbox": [ + 69, + 630, + 290, + 686 + ], + "type": "text", + "content": "Cheng Hsu, Cheng-Te Li, Diego Saez-Trumper, and Yi-Zhan Hsu. 2021. WikiContradiction: Detecting Self-Contradiction Articles on Wikipedia. Technical Report arXiv:2111.08543, arXiv. ArXiv:2111.08543 [cs] type: article." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 694, + 290, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 694, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 694, + 290, + 772 + ], + "type": "text", + "content": "Lifu Huang, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. 2019. Cosmos QA: Machine Reading Comprehension with Contextual Commonsense Reasoning. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 525, + 772 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 315, + 72, + 525, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 525, + 95 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 525, + 95 + ], + "type": "text", + "content": "2391-2401, Hong Kong, China. Association for Computational Linguistics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 101, + 525, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 101, + 525, + 157 + ], + "spans": [ + { + "bbox": [ + 304, + 101, + 525, + 157 + ], + "type": "text", + "content": "Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Yejin Bang, Andrea Madotto, and Pascale Fung. 2022. Survey of Hallucination in Natural Language Generation. Number: arXiv:2202.03629 arXiv:2202.03629 [cs]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 163, + 525, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 163, + 525, + 231 + ], + "spans": [ + { + "bbox": [ + 304, + 163, + 525, + 231 + ], + "type": "text", + "content": "Robin Jia and Percy Liang. 2017. Adversarial Examples for Evaluating Reading Comprehension Systems. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2021-2031, Copenhagen, Denmark. Association for Computational Linguistics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 237, + 525, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 237, + 525, + 303 + ], + "spans": [ + { + "bbox": [ + 304, + 237, + 525, + 303 + ], + "type": "text", + "content": "Amita Kamath, Robin Jia, and Percy Liang. 2020. Selective Question Answering under Domain Shift. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5684-5696, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 310, + 525, + 377 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 310, + 525, + 377 + ], + "spans": [ + { + "bbox": [ + 304, + 310, + 525, + 377 + ], + "type": "text", + "content": "Gregory Kell, Iain Marshall, Byron Wallace, and Andre Jaun. 2021. What Would it Take to get Biomedical QA Systems into Practice? In Proceedings of the 3rd Workshop on Machine Reading for Question Answering, pages 28-41, Punta Cana, Dominican Republic. Association for Computational Linguistics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 383, + 525, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 383, + 525, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 383, + 525, + 449 + ], + "type": "text", + "content": "Tushar Khot, Peter Clark, Michal Guerquin, Peter Jansen, and Ashish Sabharwal. 2020. QASC: A Dataset for Question Answering via Sentence Composition. Proceedings of the AAAI Conference on Artificial Intelligence, 34(05):8082-8090. Number: 05." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 456, + 525, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 525, + 512 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 525, + 512 + ], + "type": "text", + "content": "Philippe Laban, Tobias Schnabel, Paul N. Bennett, and Marti A. Hearst. 2022. SummaC: Re-Visiting NLIBased Models for Inconsistency Detection in Summarization. Transactions of the Association for Computational Linguistics, 10:163-177." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 518, + 525, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 525, + 597 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 525, + 597 + ], + "type": "text", + "content": "Guokun Lai, Qizhe Xie, Hanxiao Liu, Yiming Yang, and Eduard Hovy. 2017. RACE: Large-scale ReAding Comprehension Dataset From Examinations. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 785-794, Copenhagen, Denmark. Association for Computational Linguistics." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 602, + 525, + 648 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 602, + 525, + 648 + ], + "spans": [ + { + "bbox": [ + 304, + 602, + 525, + 648 + ], + "type": "text", + "content": "Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. 2019. Albert: A lite bert for self-supervised learning of language representations." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 654, + 525, + 710 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 525, + 710 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 525, + 710 + ], + "type": "text", + "content": "Wei Li, Wenhao Wu, Moye Chen, Jiachen Liu, Xinyan Xiao, and Hua Wu. 2022. Faithfulness in Natural Language Generation: A Systematic Survey of Analysis, Evaluation and Optimization Methods. Number: arXiv:2203.05227 arXiv:2203.05227 [cs]." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 716, + 525, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 716, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 716, + 525, + 772 + ], + "type": "text", + "content": "Yichan Liang, Jianheng Li, and Jian Yin. 2019. A New Multi-choice Reading Comprehension Dataset for Curriculum Learning. In Proceedings of The Eleventh Asian Conference on Machine Learning, pages 742-757. PMLR. ISSN: 2640-3498." + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 305, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 305, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 305, + 791 + ], + "type": "text", + "content": "831" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 70, + 72, + 289, + 772 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 70, + 72, + 289, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 72, + 289, + 138 + ], + "spans": [ + { + "bbox": [ + 70, + 72, + 289, + 138 + ], + "type": "text", + "content": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. RoBERTa: A Robustly Optimized BERT Pretraining Approach. Number: arXiv:1907.11692 arXiv:1907.11692 [cs]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 70, + 148, + 289, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 148, + 289, + 246 + ], + "spans": [ + { + "bbox": [ + 70, + 148, + 289, + 246 + ], + "type": "text", + "content": "Anshuman Mishra, Dhruvesh Patel, Aparna Vijayakumar, Xiang Lorraine Li, Pavan Kapanipathi, and Kartik Talamadupula. 2021. Looking Beyond Sentence-Level Natural Language Inference for Question Answering and Text Summarization. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1322-1336, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 255, + 289, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 255, + 289, + 332 + ], + "spans": [ + { + "bbox": [ + 70, + 255, + 289, + 332 + ], + "type": "text", + "content": "Yixin Nie, Adina Williams, Emily Dinan, Mohit Bansal, Jason Weston, and Douwe Kiela. 2020. Adversarial NLI: A New Benchmark for Natural Language Understanding. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4885-4901, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 341, + 289, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 341, + 289, + 439 + ], + "spans": [ + { + "bbox": [ + 70, + 341, + 289, + 439 + ], + "type": "text", + "content": "Yixin Nie, Mary Williamson, Mohit Bansal, Douwe Kiela, and Jason Weston. 2021. I like fish, especially dolphins: Addressing Contradictions in Dialogue Modeling. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1699-1713, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 449, + 289, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 449, + 289, + 536 + ], + "spans": [ + { + "bbox": [ + 70, + 449, + 289, + 536 + ], + "type": "text", + "content": "Simon Ostermann, Ashutosh Modi, Michael Roth, Stefan Thater, and Manfred Pinkal. 2018. MCScript: A Novel Dataset for Assessing Machine Comprehension Using Script Knowledge. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources Association (ELRA)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 545, + 289, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 545, + 289, + 622 + ], + "spans": [ + { + "bbox": [ + 70, + 545, + 289, + 622 + ], + "type": "text", + "content": "Simon Ostermann, Michael Roth, and Manfred Pinkal. 2019. MCScript2.0: A Machine Comprehension Corpus Focused on Script Events and Participants. In Proceedings of the Eighth Joint Conference on Lexical and Computational Semantics (*SEM* 2019), pages 103-117, Minneapolis, Minnesota. Association for Computational Linguistics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 70, + 631, + 289, + 675 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 631, + 289, + 675 + ], + "spans": [ + { + "bbox": [ + 70, + 631, + 289, + 675 + ], + "type": "text", + "content": "Aarthi Paramasivam and S. Jaya Nirmala. 2021. A survey on textual entailment based question answering. Journal of King Saud University - Computer and Information Sciences." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 70, + 684, + 289, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 684, + 289, + 772 + ], + "spans": [ + { + "bbox": [ + 70, + 684, + 289, + 772 + ], + "type": "text", + "content": "Rajkumar Pujari and Dan Goldwasser. 2019. Using natural language relations between answer choices for machine comprehension. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4010-4015, Minneapolis, Minnesota. Association for Computational Linguistics." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 524, + 772 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 304, + 72, + 524, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 524, + 138 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 524, + 138 + ], + "type": "text", + "content": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer. Journal of Machine Learning Research, 21(140):1-67." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 148, + 524, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 148, + 524, + 224 + ], + "spans": [ + { + "bbox": [ + 304, + 148, + 524, + 224 + ], + "type": "text", + "content": "Pranav Rajpurkar, Robin Jia, and Percy Liang. 2018. Know What You Don't Know: Unanswerable Questions for SQuAD. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 784-789, Melbourne, Australia. Association for Computational Linguistics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 234, + 524, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 234, + 524, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 234, + 524, + 300 + ], + "type": "text", + "content": "Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD: 100,000+ Questions for Machine Comprehension of Text. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2383-2392, Austin, Texas. Association for Computational Linguistics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 309, + 524, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 309, + 524, + 397 + ], + "spans": [ + { + "bbox": [ + 304, + 309, + 524, + 397 + ], + "type": "text", + "content": "Nils Reimers and Iryna Gurevych. 2019. SentenceBERT: Sentence Embeddings using Siamese BERT-Networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3982-3992, Hong Kong, China. Association for Computational Linguistics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 407, + 524, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 524, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 524, + 483 + ], + "type": "text", + "content": "Matthew Richardson, Christopher J.C. Burges, and Erin Renshaw. 2013. MCTest: A Challenge Dataset for the Open-Domain Machine Comprehension of Text. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 193-203, Seattle, Washington, USA. Association for Computational Linguistics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 492, + 524, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 492, + 524, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 492, + 524, + 536 + ], + "type": "text", + "content": "Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. 2020. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. Number: arXiv:1910.01108 arXiv:1910.01108 [cs]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 545, + 524, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 524, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 524, + 601 + ], + "type": "text", + "content": "Tal Schuster, Sihao Chen, Senaka Buthpitiya, Alex Fabrikant, and Donald Metzler. 2022. Stretching Sentence-pair NLI Models to Reason over Long Documents and Clusters. Number: arXiv:2204.07447 arXiv:2204.07447 [cs]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 610, + 524, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 610, + 524, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 610, + 524, + 665 + ], + "type": "text", + "content": "Haoyu Song, Wei-Nan Zhang, Jingwen Hu, and Ting Liu. 2020. Generating Persona Consistent Dialogues by Exploiting Natural Language Inference. Proceedings of the AAAI Conference on Artificial Intelligence, 34(05):8878-8885. Number: 05." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 675, + 524, + 740 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 675, + 524, + 740 + ], + "spans": [ + { + "bbox": [ + 304, + 675, + 524, + 740 + ], + "type": "text", + "content": "Kai Sun, Dian Yu, Jianshu Chen, Dong Yu, Yejin Choi, and Claire Cardie. 2019. DREAM: A Challenge Data Set and Models for Dialogue-Based Reading Comprehension. Transactions of the Association for Computational Linguistics, 7:217-231. Place: Cambridge, MA Publisher: MIT Press." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 750, + 524, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 750, + 524, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 750, + 524, + 772 + ], + "type": "text", + "content": "Harsh Trivedi, Heeyoung Kwon, Tushar Khot, Ashish Sabharwal, and Niranjan Balasubramanian. 2019." + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 790 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 790 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 790 + ], + "type": "text", + "content": "832" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 78, + 72, + 291, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 72, + 291, + 150 + ], + "spans": [ + { + "bbox": [ + 78, + 72, + 291, + 150 + ], + "type": "text", + "content": "Repurposing Entailment for Multi-Hop Question Answering Tasks. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2948-2958, Minneapolis, Minnesota. Association for Computational Linguistics." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 158, + 291, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 158, + 291, + 191 + ], + "spans": [ + { + "bbox": [ + 68, + 158, + 291, + 191 + ], + "type": "text", + "content": "Johannes Welbl, Nelson F. Liu, and Matt Gardner. 2017. Crowdsourcing Multiple Choice Science Questions. In NUT@EMNLP." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 200, + 291, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 200, + 291, + 299 + ], + "spans": [ + { + "bbox": [ + 69, + 200, + 291, + 299 + ], + "type": "text", + "content": "Adina Williams, Nikita Nangia, and Samuel Bowman. 2018. A broad-coverage challenge corpus for sentence understanding through inference. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1112-1122, New Orleans, Louisiana. Association for Computational Linguistics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 306, + 291, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 306, + 291, + 417 + ], + "spans": [ + { + "bbox": [ + 69, + 306, + 291, + 417 + ], + "type": "text", + "content": "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander M. Rush. 2020. HuggingFace's Transformers: State-of-the-art Natural Language Processing. Number: arXiv:1910.03771 arXiv:1910.03771 [cs]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 428, + 230, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 428, + 230, + 443 + ], + "spans": [ + { + "bbox": [ + 68, + 428, + 230, + 443 + ], + "type": "text", + "content": "A Answer ranking procedure" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 450, + 290, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 450, + 290, + 544 + ], + "spans": [ + { + "bbox": [ + 67, + 450, + 290, + 544 + ], + "type": "text", + "content": "In the multiple choice setting, we performed an answer ranking procedure to pick the answer to a given question among a set of alternative answers " + }, + { + "bbox": [ + 67, + 450, + 290, + 544 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 67, + 450, + 290, + 544 + ], + "type": "text", + "content": ", using both NLI class scores and QA confidence scores. (This is distinct from the selection procedure for the top 20% or 50% of answers we used in both settings.)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 544, + 290, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 544, + 290, + 705 + ], + "spans": [ + { + "bbox": [ + 67, + 544, + 290, + 705 + ], + "type": "text", + "content": "Similar to Harabagiu and Hickl (2006), answers are ranked based on the highest probability from the calibration model " + }, + { + "bbox": [ + 67, + 544, + 290, + 705 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 67, + 544, + 290, + 705 + ], + "type": "text", + "content": ", given a linear combination of the QA or NLI scores given an answer " + }, + { + "bbox": [ + 67, + 544, + 290, + 705 + ], + "type": "inline_equation", + "content": "n \\in N" + }, + { + "bbox": [ + 67, + 544, + 290, + 705 + ], + "type": "text", + "content": " answer set. When a single feature is used, such as an NLI class or the QA score, no calibration is made and " + }, + { + "bbox": [ + 67, + 544, + 290, + 705 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 67, + 544, + 290, + 705 + ], + "type": "text", + "content": " is simply the identity of the confidence score. In the case of contradiction only, " + }, + { + "bbox": [ + 67, + 544, + 290, + 705 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 67, + 544, + 290, + 705 + ], + "type": "text", + "content": " is the inverse of the contradiction confidence score, indicating the least contradicted answer is being selected. Formally, our approach can be described as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 125, + 707, + 232, + 728 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 707, + 232, + 728 + ], + "spans": [ + { + "bbox": [ + 125, + 707, + 232, + 728 + ], + "type": "interline_equation", + "content": "\\operatorname * {a r g m a x} _ {N} \\sigma (\\mathrm {Q A} _ {n}; \\mathrm {N L I} _ {n})", + "image_path": "dffa2b3d6e2b6f58d5e25c1cd0abcdd5c4bf1f1d04dc26861f519f6d4d7fc057.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 733, + 290, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 733, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 733, + 290, + 772 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 67, + 733, + 290, + 772 + ], + "type": "inline_equation", + "content": "\\mathrm{QA}_n" + }, + { + "bbox": [ + 67, + 733, + 290, + 772 + ], + "type": "text", + "content": " is the QA model confidence score for answer " + }, + { + "bbox": [ + 67, + 733, + 290, + 772 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 67, + 733, + 290, + 772 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 67, + 733, + 290, + 772 + ], + "type": "inline_equation", + "content": "\\mathrm{NLI}_n" + }, + { + "bbox": [ + 67, + 733, + 290, + 772 + ], + "type": "text", + "content": " represents the various NLI class scores for " + }, + { + "bbox": [ + 67, + 733, + 290, + 772 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 67, + 733, + 290, + 772 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 71, + 526, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 137 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 137 + ], + "type": "text", + "content": "We did not use this approach in extractive QA, because we found that asking the model for the top " + }, + { + "bbox": [ + 302, + 71, + 526, + 137 + ], + "type": "inline_equation", + "content": "K = 4" + }, + { + "bbox": [ + 302, + 71, + 526, + 137 + ], + "type": "text", + "content": " answer produced almost the same four answer alternatives with slightly different spans each time." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 303, + 148, + 370, + 161 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 148, + 370, + 161 + ], + "spans": [ + { + "bbox": [ + 303, + 148, + 370, + 161 + ], + "type": "text", + "content": "B Datasets" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 169, + 526, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 169, + 526, + 332 + ], + "spans": [ + { + "bbox": [ + 302, + 169, + 526, + 332 + ], + "type": "text", + "content": "Tables 4 (multiple choice) and 5 (extractive QA) outline the datasets we used. Additional details such as train size and preprocessing steps are available in the references provided. When space doesn't allow CosmosQA is aliased to Cosmos, MCScript to MCS, MCScript-2.0 to MCS2, and MCTest to MCT. The only preprocessing step we performed was to filter out questions where no context passage is provided. Validation splits (as opposed to test splits) are used in the CosmosQA and QASC cases, since context passages or gold standard answers are not available for these datasets." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 342, + 384, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 342, + 384, + 355 + ], + "spans": [ + { + "bbox": [ + 302, + 342, + 384, + 355 + ], + "type": "text", + "content": "C QA models" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 363, + 525, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 363, + 525, + 430 + ], + "spans": [ + { + "bbox": [ + 302, + 363, + 525, + 430 + ], + "type": "text", + "content": "Table 6 outlines the pretrained QA models that we used and the datasets they are trained on. All these models are publicly available on the Hugging Face hub under the locations listed. Where space doesn't allow, RoBERTa-RACE is aliased as RACE." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 431, + 525, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 431, + 525, + 579 + ], + "spans": [ + { + "bbox": [ + 302, + 431, + 525, + 579 + ], + "type": "text", + "content": "We trained the two DeBERTa-v3 models (xsmall and base) as shown in Table 7. They were trained using the Hugging Face trainer API (Wolf et al., 2020) with an Adam optimizer at a learning rate of " + }, + { + "bbox": [ + 302, + 431, + 525, + 579 + ], + "type": "inline_equation", + "content": "5.60\\mathrm{e - }05" + }, + { + "bbox": [ + 302, + 431, + 525, + 579 + ], + "type": "text", + "content": " with weight decay of 0.01. All models and inference were performed on 1 Tesla P100 GPU. Full instructions on reproducibility as well as trained models are provided in the publicly available code, including directions to weights and biases to inspect the training runs, full parameter set, and evaluation suites." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 589, + 399, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 589, + 399, + 602 + ], + "spans": [ + { + "bbox": [ + 302, + 589, + 399, + 602 + ], + "type": "text", + "content": "D QA2D models" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 611, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 611, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 611, + 526, + 772 + ], + "type": "text", + "content": "A QA2D model reformulates a question-answer pair to a declarative statement (Demszky et al., 2018). As noted in Chen et al. (2021) and Mishra et al. (2021), the QA2D reformulation is critical to using NLI models in QA since the proposed answer needs to match the format of NLI. We trained a T5-small model (Raffel et al., 2020) on the dataset proposed by Demszky et al. (2018) for QA2D since we found almost no noticeable differences in performance in larger models. This used the same setup as the DeBERTa-v3 models xsmall and base (see Table 7)." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "833" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 156, + 68, + 438, + 208 + ], + "blocks": [ + { + "bbox": [ + 156, + 68, + 438, + 208 + ], + "lines": [ + { + "bbox": [ + 156, + 68, + 438, + 208 + ], + "spans": [ + { + "bbox": [ + 156, + 68, + 438, + 208 + ], + "type": "table", + "html": "
DatasetSplitSizeReference
CosmosQAvalidation2985Huang et al. (2019)
DREAMtest2041Sun et al. (2019)
MCScripttest2797Ostermann et al. (2018)
MCScript-2.0test3610Ostermann et al. (2019)
MCTesttest840Richardson et al. (2013)
QASCvalidation926Khot et al. (2020)
RACEtest4934Lai et al. (2017)
RACE-Ctest712Liang et al. (2019)
SciQtest884Welbl et al. (2017)
", + "image_path": "95ff9adeab6c72a1bc22cef52c0589dd4940a5fdda9a97baeed88dc83f6f4017.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 174, + 262, + 421, + 375 + ], + "blocks": [ + { + "bbox": [ + 67, + 216, + 526, + 252 + ], + "lines": [ + { + "bbox": [ + 67, + 216, + 526, + 252 + ], + "spans": [ + { + "bbox": [ + 67, + 216, + 526, + 252 + ], + "type": "text", + "content": "Table 4: Datasets used for the multiple choice setting, including split used and sample size. Validation splits were used for CosmosQA since the test split is not publicly available, and for QASC since context passages or gold answers are not available." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 174, + 262, + 421, + 375 + ], + "lines": [ + { + "bbox": [ + 174, + 262, + 421, + 375 + ], + "spans": [ + { + "bbox": [ + 174, + 262, + 421, + 375 + ], + "type": "table", + "html": "
DatasetSizeReference
BioASQ1504Fisch et al. (2019)
TriviaQA7785
HotpotQA5901
SQuAD10506
Natural Questions12836
SQuAD211871Rajpurkar et al. (2018)
SQuAD-adv5347Jia and Liang (2017)
", + "image_path": "ec3116058e0f8a9362002140c917858818ecd846c2d3ed929efaefe44bc4765d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 383, + 525, + 407 + ], + "lines": [ + { + "bbox": [ + 67, + 383, + 525, + 407 + ], + "spans": [ + { + "bbox": [ + 67, + 383, + 525, + 407 + ], + "type": "text", + "content": "Table 5: Extractive QA datasets used. Validation sets are used on the SQuAD2.0 and SQuAD adversarial datasets. MRQA 2019 dev sets are used for the other five datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 428, + 291, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 428, + 291, + 644 + ], + "spans": [ + { + "bbox": [ + 67, + 428, + 291, + 644 + ], + "type": "text", + "content": "Unlike Chen et al. (2021), we found that regardless of size, these QA2D models struggled with long questions or questions with complex syntax and would often leave the answer out of the statement. In order to solve this, constrained decoding that required the answer to be in the statement was tried. However, this often produced ungrammatical or nonsensical statements. We settled with the following heuristic to postprocess QA2D outputs: If less than " + }, + { + "bbox": [ + 67, + 428, + 291, + 644 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 67, + 428, + 291, + 644 + ], + "type": "text", + "content": " of the tokens in the answer were in the statement then we appended the answer to the end of the statement. " + }, + { + "bbox": [ + 67, + 428, + 291, + 644 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 67, + 428, + 291, + 644 + ], + "type": "text", + "content": " was used to account for rephrasing the answer or swapping pronouns. While some statements resulted in answer redundancy, this was better than having hypotheses which left out the answer." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 678, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 678, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 678, + 291, + 773 + ], + "type": "text", + "content": "Future work on QA2D should focus on how these models can be used outside of the domains in the dataset provided by Demszky et al. (2018). Finally it is important to note that erroneous QA2D outputs could effect the quality of the whole pipeline see Chen et al. (2021) for a more detailed analysis of this." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 303, + 428, + 387, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 428, + 387, + 441 + ], + "spans": [ + { + "bbox": [ + 303, + 428, + 387, + 441 + ], + "type": "text", + "content": "E NLI models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 449, + 526, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 449, + 526, + 571 + ], + "spans": [ + { + "bbox": [ + 302, + 449, + 526, + 571 + ], + "type": "text", + "content": "NLI is used to classify whether the reformulated answer is contradicted, entailed, or neutral with respect to a context passage. We used the whole context, as Schuster et al. (2022) and Mishra et al. (2021) demonstrated that long premises still performed adequate though not as well as sentence-length premises. Using the whole context avoids needing to use decontextualization as is required in Chen et al. (2021)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 571, + 527, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 571, + 527, + 705 + ], + "spans": [ + { + "bbox": [ + 302, + 571, + 527, + 705 + ], + "type": "text", + "content": "We used two DeBERTa-based models (He et al., 2021b) trained on the MNLI dataset (Williams et al., 2018) (called mnli-base and mnli-large) and an ALBERT model (Lan et al., 2019) trained on the ANLI dataset in addition to various other NLI datasets (called albert-anli) (Nie et al., 2020). Table 6 contains the Hugging Face references to the NLI models. After inference, the confidence scores are used for answer selection and performance evaluation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 715, + 512, + 742 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 715, + 512, + 742 + ], + "spans": [ + { + "bbox": [ + 302, + 715, + 512, + 742 + ], + "type": "text", + "content": "E.1 Model size and approach performance analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "type": "text", + "content": "Table 8 mirrors Table 1 in the main text, but shows the accuracy results for uncalibrated E, C, and " + }, + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "type": "inline_equation", + "content": "\\mathbf{E} + \\mathbf{C}" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "834" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 115, + 94, + 478, + 193 + ], + "blocks": [ + { + "bbox": [ + 115, + 94, + 478, + 193 + ], + "lines": [ + { + "bbox": [ + 115, + 94, + 478, + 193 + ], + "spans": [ + { + "bbox": [ + 115, + 94, + 478, + 193 + ], + "type": "table", + "html": "
Hugging FaceName
LIAMF-USP/roberta-large-finetuned-RACERoBERTa-RACE
bert-large-uncased-whole-word-masking-finetuned-squadBERT-Large
distilbert-base-uncased-distilled-squadDistillBERT
ynie/albert-xxlarge-v2-snli_mnli_fever_anli_R1_R2_R3-nlialbert-anli
microsoft/deberta-base-mnlimnli-base
microsoft/deberta-v2-xxlarge-mnlimnli-large
", + "image_path": "aabe6847b74a5c1f9194a6b3c8dffb8261573554b3ed4595b5ad315827774805.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 128, + 271, + 466, + 329 + ], + "blocks": [ + { + "bbox": [ + 202, + 200, + 389, + 212 + ], + "lines": [ + { + "bbox": [ + 202, + 200, + 389, + 212 + ], + "spans": [ + { + "bbox": [ + 202, + 200, + 389, + 212 + ], + "type": "text", + "content": "Table 6: Pretrained QA and NLI models used." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 128, + 271, + 466, + 329 + ], + "lines": [ + { + "bbox": [ + 128, + 271, + 466, + 329 + ], + "spans": [ + { + "bbox": [ + 128, + 271, + 466, + 329 + ], + "type": "table", + "html": "
ModelDatasetEpochsScore
t5-smallDemszky et al. (2018)20Rogue190.73
deberta-v3-xsmallWelbl et al. (2017)6Accuracy93.99
deberta-v3-baseWelbl et al. (2017)6Accuracy91.79
", + "image_path": "ff65b4f2ddde834b1a916e0e2db5de3c4c3e5d630d5aaf41816de55145f7042c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 69, + 407, + 526, + 711 + ], + "blocks": [ + { + "bbox": [ + 98, + 338, + 495, + 350 + ], + "lines": [ + { + "bbox": [ + 98, + 338, + 495, + 350 + ], + "spans": [ + { + "bbox": [ + 98, + 338, + 495, + 350 + ], + "type": "text", + "content": "Table 7: The models we trained for or setups with evaluation scores and number of epochs trained." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 407, + 526, + 711 + ], + "lines": [ + { + "bbox": [ + 69, + 407, + 526, + 711 + ], + "spans": [ + { + "bbox": [ + 69, + 407, + 526, + 711 + ], + "type": "table", + "html": "
QA ModelCosmosDREAMMCSMCS2MCTQASCRACERACE-CSciQAverage
SciQ-base18.4643.8061.9963.7144.7693.4130.9727.3995.2853.31
SciQ-small25.4648.2660.2866.0459.7690.6035.5630.6298.0957.19
RACE64.2282.5689.7086.9890.4898.1676.9369.8097.9684.09
mnli-large
E+C44.3680.9485.5284.9990.6096.4464.2951.4092.4776.77
E36.1879.0386.0279.7289.8895.9062.1449.7291.9674.50
C59.2678.9883.1284.4389.2992.7662.7447.0591.5876.58
mnli-base
QA + E + C64.3282.6689.6387.0190.7198.2776.9569.8098.0984.16
QA + E64.2582.6689.6386.9890.7198.2776.9569.8097.9684.14
QA + C64.2982.5689.6387.0190.6098.1676.9369.8097.9684.1
E + C33.0362.2776.7672.1168.5792.6645.1634.4188.0163.66
E27.8162.4779.3771.9468.8192.6643.4834.4188.0163.22
C43.4559.1970.1869.9767.5081.8641.8132.5887.3761.55
albert-anli
QA + E + C64.1982.5689.7087.0690.4898.1676.9369.8097.9684.09
QA + E64.1982.5689.7087.0690.6098.1676.9369.8097.9684.11
QA + C64.2282.5689.7086.9890.4898.1676.9369.8097.9684.09
E + C35.7168.2079.5573.8877.5091.7949.0539.4790.8267.33
E33.6768.3579.9173.1977.3891.9049.0739.1990.9467.07
C45.1663.7473.5872.7173.3377.8646.3438.2087.2464.24
", + "image_path": "7b086607a861841b750a1ee7a6e869cd48ea33a90061f475cc41082373f894a9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 719, + 525, + 743 + ], + "lines": [ + { + "bbox": [ + 67, + 719, + 525, + 743 + ], + "spans": [ + { + "bbox": [ + 67, + 719, + 525, + 743 + ], + "type": "text", + "content": "Table 8: Accuracy scores in the multiple choice setting for various NLI models used. Calibration was with the RoBERTA-RACE model." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "835" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 290, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 290, + 193 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 290, + 193 + ], + "type": "text", + "content": "in the main mnli-large model, as well as the results with the other NLI models, mnli-base and albertanli. Table 9 shows selective QA accuracy in the multiple choice setting where answer selection is done through ranking before we rank answers for selective QA. Selective QA on extractive QA using DistillBERT (table 10) shows that " + }, + { + "bbox": [ + 67, + 71, + 290, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{QA} + \\mathbf{E} + \\mathbf{C}" + }, + { + "bbox": [ + 67, + 71, + 290, + 193 + ], + "type": "text", + "content": " does best in all cases and contradiction only does second best at " + }, + { + "bbox": [ + 67, + 71, + 290, + 193 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 67, + 71, + 290, + 193 + ], + "type": "text", + "content": " coverage." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 204, + 189, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 204, + 189, + 216 + ], + "spans": [ + { + "bbox": [ + 67, + 204, + 189, + 216 + ], + "type": "text", + "content": "F Calibration models" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 226, + 290, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 226, + 290, + 305 + ], + "spans": [ + { + "bbox": [ + 67, + 226, + 290, + 305 + ], + "type": "text", + "content": "Like Kamath et al. (2020) and Chen et al. (2021) we developed a set of calibration models in order to perform answer ranking. A calibration model is trained on a set of posterior probabilities from downstream models to predict whether an answer is correct." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 307, + 291, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 307, + 291, + 523 + ], + "spans": [ + { + "bbox": [ + 69, + 307, + 291, + 523 + ], + "type": "text", + "content": "To compare the effect of using different combinations of NLI class confidence scores, we trained a logistic regression model on linear combinations of the following features: QA indicates that the QA model confidence score is being used, " + }, + { + "bbox": [ + 69, + 307, + 291, + 523 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 69, + 307, + 291, + 523 + ], + "type": "text", + "content": " indicates the entailment score, " + }, + { + "bbox": [ + 69, + 307, + 291, + 523 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 69, + 307, + 291, + 523 + ], + "type": "text", + "content": " indicates the contradiction score, and " + }, + { + "bbox": [ + 69, + 307, + 291, + 523 + ], + "type": "inline_equation", + "content": "\\mathbf{N}" + }, + { + "bbox": [ + 69, + 307, + 291, + 523 + ], + "type": "text", + "content": " indicates the neutral score. Like in Chen et al. (2021), all calibration models are trained on a holdout set of 100 samples from a single domain using logistic regression which predicts, given the confidence scores of the downstream models, whether the answer is correct. A multi-domain calibration approach like in Kamath et al. (2020) was not used since the focus was a minimum experiment to test the viability of leveraging different NLI classifications." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 534, + 189, + 546 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 534, + 189, + 546 + ], + "spans": [ + { + "bbox": [ + 67, + 534, + 189, + 546 + ], + "type": "text", + "content": "F.1 Regression Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 551, + 290, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 551, + 290, + 644 + ], + "spans": [ + { + "bbox": [ + 67, + 551, + 290, + 644 + ], + "type": "text", + "content": "To illustrate the characteristics of the calibration models, we present a regression analysis for the multiple choice setting (Table 11). The results indicate that as the mnli model gets larger, the calibration model uses its NLI confidence scores more. Importantly, entailment coefficients are stronger than contradiction coefficients in all cases." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 657, + 198, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 657, + 198, + 670 + ], + "spans": [ + { + "bbox": [ + 67, + 657, + 198, + 670 + ], + "type": "text", + "content": "G Correlation Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 678, + 290, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 678, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 678, + 290, + 772 + ], + "type": "text", + "content": "Since we are using the NLI and QA model scores to construct the setups above, it is useful to know how these factors correlate with the correct answer. Table 13 shows how each NLI class correlates both by score and by actual classification (score " + }, + { + "bbox": [ + 67, + 678, + 290, + 772 + ], + "type": "inline_equation", + "content": ">50\\%" + }, + { + "bbox": [ + 67, + 678, + 290, + 772 + ], + "type": "text", + "content": ") as compared against QA model confidence score. The multiple choice analysis shows" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 71, + 524, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 524, + 125 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 524, + 125 + ], + "type": "text", + "content": "answers from the RoBERTa-RACE model and the extractive QA analysis shows answers from the BERT-large model trained on SQuAD. The correlation analysis presents Spearman rank correlations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 126, + 525, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 126, + 525, + 300 + ], + "spans": [ + { + "bbox": [ + 302, + 126, + 525, + 300 + ], + "type": "text", + "content": "What we see is that in the multiple choice setting, the confidence score has a strong correlation with the correct answer, which makes sense given the confidence score is a softmax over the multiple choice classes. Extractive QA confidence scores have a much weaker correlation and tend to have less correlation than entailment has with the correct answer. Despite the results presented above, contradiction only has a notable correlation with the correct answer when the score is used rather than the classification. This is a point in favor of our approach of using confidence scores for NLI rather than classifications." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 301, + 525, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 301, + 525, + 504 + ], + "spans": [ + { + "bbox": [ + 302, + 301, + 525, + 504 + ], + "type": "text", + "content": "Interestingly, in the extractive QA case, the neutral class is more negatively correlated when selecting for contradiction when using classification. Our conjecture would be that in the extractive QA case, we don't have much to compare against. When looking at the per dataset correlations for the multiple choice setting (Table 12), we see that in most cases, other than the QA confidence scores, the contradiction scores have the strongest correlations with the correct answer out of any NLI class and neutral, as we would expect, tends to have very weak correlations. We do not present the per dataset correlation for extractive QA as they are very weak, which we again hypothesize comes from having no answers to compare with." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "836" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 103, + 95, + 490, + 382 + ], + "blocks": [ + { + "bbox": [ + 103, + 95, + 490, + 382 + ], + "lines": [ + { + "bbox": [ + 103, + 95, + 490, + 382 + ], + "spans": [ + { + "bbox": [ + 103, + 95, + 490, + 382 + ], + "type": "table", + "html": "
DatasetQA+E+CQA+EQA+CE+CECQA
20%CosmosQA77.5567.1783.2520.1027.4767.5088.61
DREAM98.2896.3296.8181.1391.9193.8798.28
MCScript99.8299.6499.4693.0298.9396.9699.82
MCScript-2.099.5899.0397.3792.2497.3795.0199.58
MCTest10010099.4085.1297.0297.0298.81
QASC10010010097.3010099.46100
RACE94.9392.1390.1762.7376.7175.0598.24
RACE-C88.7385.2186.6271.1374.6569.0193.66
SciQ10010010082.0510096.15100
Avg95.4393.2894.7976.0984.9087.7897.45
50%CosmosQA80.2970.7880.7032.1734.7264.8876.47
DREAM95.1093.6393.6385.2089.4188.3396.67
MCScript98.5797.8597.1494.7195.9992.7098.78
MCScript-2.096.4094.4696.0791.0291.7591.6998.01
MCTest99.5298.8199.7691.4395.2496.1999.52
QASC10099.7899.7898.2798.7098.49100
RACE90.1187.2285.2367.8971.7068.1893.88
RACE-C85.1178.0977.2566.5766.8555.0687.36
SciQ10010099.7489.0396.4396.43100
Avg93.9091.1892.1479.5982.3183.5594.52
", + "image_path": "408f34b65fbab8808d9bfcfc9d848a79c7d6b25d673bce83a108edd2bb832999.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 92, + 475, + 501, + 709 + ], + "blocks": [ + { + "bbox": [ + 67, + 391, + 524, + 417 + ], + "lines": [ + { + "bbox": [ + 67, + 391, + 524, + 417 + ], + "spans": [ + { + "bbox": [ + 67, + 391, + 524, + 417 + ], + "type": "text", + "content": "Table 9: Selective QA accuracies in the multiple choice setting where answer selection is done through ranking before we rank answers for selective QA." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 92, + 475, + 501, + 709 + ], + "lines": [ + { + "bbox": [ + 92, + 475, + 501, + 709 + ], + "spans": [ + { + "bbox": [ + 92, + 475, + 501, + 709 + ], + "type": "table", + "html": "
DatasetQA+E+CQA+EQA+CE+CECQA
20%BioASQ70.9770.4171.5574.0774.0774.3468.99
HotpotQA73.4473.0870.8871.5971.5170.4169.41
Natural Questions85.5985.2985.4578.4678.4680.5383.27
SQuAD96.2296.4595.7783.1583.0981.3797.15
SQuAD-adv40.3939.7539.4940.0739.5640.5931.98
SQuAD235.4635.2433.6436.3636.1336.6625.95
TriviaQA64.9664.6864.5552.6752.0952.5663.98
Avg66.7266.4165.9062.3462.1362.3562.96
50%BioASQ65.9665.9264.3763.5363.5366.9564.79
HotpotQA64.4264.2163.6565.8865.8566.9162.81
Natural Questions72.2871.9970.8267.5467.5174.1869.95
SQuAD92.5692.5792.3481.8682.2180.9592.54
SQuAD-adv33.6932.9033.4538.7438.2238.5231.89
SQuAD226.6825.7026.0032.9532.6132.8323.52
TriviaQA58.4058.4158.2551.4351.1852.9958.25
Avg59.1458.8158.4157.4257.3059.0557.68
", + "image_path": "5c718e56704d68138c5deda131e4c1792915577d0978111b534fc7b34580456f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 718, + 525, + 743 + ], + "lines": [ + { + "bbox": [ + 67, + 718, + 525, + 743 + ], + "spans": [ + { + "bbox": [ + 67, + 718, + 525, + 743 + ], + "type": "text", + "content": "Table 10: SelectiveQA on extractive QA using DistillBERT. Note that QA+E+C does best in all cases and contradiction only does second best at " + }, + { + "bbox": [ + 67, + 718, + 525, + 743 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 67, + 718, + 525, + 743 + ], + "type": "text", + "content": " coverage." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "837" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 88, + 88, + 505, + 322 + ], + "blocks": [ + { + "bbox": [ + 88, + 88, + 505, + 322 + ], + "lines": [ + { + "bbox": [ + 88, + 88, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 88, + 88, + 505, + 322 + ], + "type": "table", + "html": "
QA ModelNLI ModelCombinationConfidenceEntailmentContradictionAcc
SciQmnli-baseQA + C4.13-1.060.99
QA + E3.901.370.99
QA + E + C3.831.22-0.760.99
E + C2.56-1.470.86
mnli-largeQA + C3.98-1.320.99
QA + E3.781.550.99
QA + E + C3.651.31-0.970.99
E + C2.63-1.720.91
RACEmnli-baseQA + C3.04-0.150.89
QA + E3.030.270.89
QA + E + C3.020.26-0.140.89
E + C0.73-0.460.75
mnli-largeQA + C2.970.00-0.810.89
QA + E2.910.980.89
QA + E + C2.850.92-0.750.89
E + C1.76-1.120.78
", + "image_path": "a850ae80ec18a1b5bc704fa8d01ae7903f77bd6be016fb12557705a1a902681f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 107, + 401, + 489, + 555 + ], + "blocks": [ + { + "bbox": [ + 67, + 330, + 525, + 356 + ], + "lines": [ + { + "bbox": [ + 67, + 330, + 525, + 356 + ], + "spans": [ + { + "bbox": [ + 67, + 330, + 525, + 356 + ], + "type": "text", + "content": "Table 11: Regression analysis for each mnli-based nli model with each QA model used calibration with logistic regression multiple choice settings. Accuracy is the evaluation metric used." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 107, + 401, + 489, + 555 + ], + "lines": [ + { + "bbox": [ + 107, + 401, + 489, + 555 + ], + "spans": [ + { + "bbox": [ + 107, + 401, + 489, + 555 + ], + "type": "table", + "html": "
ContradictionEntailmentNeutral
DatasetQAScoreClassScoreClassScoreClass
CosmosQA0.53-0.34-0.170.05-0.010.210.16
DREAM0.72-0.57-0.350.540.50-0.11-0.13
MCScript0.80-0.59-0.420.590.50-0.04-0.08
MCScript20.77-0.50-0.320.410.37-0.04-0.05
MCTest0.73-0.65-0.470.640.69-0.20-0.15
QASC0.57-0.54-0.280.550.67-0.50-0.26
RACE0.65-0.37-0.200.350.34-0.11-0.11
RACE-C0.59-0.24-0.130.180.25-0.09-0.11
SciQ0.75-0.69-0.470.680.67-0.42-0.19
", + "image_path": "fe73c7096c9f26a9a8032c7ee38eaa69d298bb3a701d1ef341f645204ae60247.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 133, + 633, + 461, + 706 + ], + "blocks": [ + { + "bbox": [ + 67, + 562, + 525, + 587 + ], + "lines": [ + { + "bbox": [ + 67, + 562, + 525, + 587 + ], + "spans": [ + { + "bbox": [ + 67, + 562, + 525, + 587 + ], + "type": "text", + "content": "Table 12: Correlation analysis (Spearman rank correlation) per dataset in the multiple choice setting. RoBERTaRACE is used for the QA scores." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 633, + 461, + 706 + ], + "lines": [ + { + "bbox": [ + 133, + 633, + 461, + 706 + ], + "spans": [ + { + "bbox": [ + 133, + 633, + 461, + 706 + ], + "type": "table", + "html": "
ContradictionEntailmentNeutralQA
multiple choiceScore-0.470.37-0.060.71
Class-0.280.38-0.06
extractive QAScore-0.160.31-0.120.19
Class-0.150.39-0.29
", + "image_path": "2d7dde0f008ecc0fe2efb805caf08f130341d5c5c37e4f70bc309a98dfb1f46f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 713, + 525, + 749 + ], + "lines": [ + { + "bbox": [ + 67, + 713, + 525, + 749 + ], + "spans": [ + { + "bbox": [ + 67, + 713, + 525, + 749 + ], + "type": "text", + "content": "Table 13: Correlation analysis (Spearman rank correlation) in the multiple choice and extractive QA settings. RoBERTa-RACE is the QA model used for multiple choice QA scores and BERT-large is used for the extractive QA scores." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "838" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 90, + 192, + 103 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 90, + 192, + 103 + ], + "spans": [ + { + "bbox": [ + 68, + 90, + 192, + 103 + ], + "type": "text", + "content": "A For every submission:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 76, + 107, + 317, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 107, + 317, + 121 + ], + "spans": [ + { + "bbox": [ + 76, + 107, + 317, + 121 + ], + "type": "text", + "content": "A1. Did you describe the limitations of your work?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 121, + 138, + 134 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 121, + 138, + 134 + ], + "spans": [ + { + "bbox": [ + 89, + 121, + 138, + 134 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 76, + 143, + 329, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 143, + 329, + 157 + ], + "spans": [ + { + "bbox": [ + 76, + 143, + 329, + 157 + ], + "type": "text", + "content": "A2. Did you discuss any potential risks of your work?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 158, + 138, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 158, + 138, + 169 + ], + "spans": [ + { + "bbox": [ + 89, + 158, + 138, + 169 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 76, + 179, + 414, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 179, + 414, + 193 + ], + "spans": [ + { + "bbox": [ + 76, + 179, + 414, + 193 + ], + "type": "text", + "content": "A3. Do the abstract and introduction summarize the paper's main claims?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 89, + 194, + 138, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 194, + 138, + 206 + ], + "spans": [ + { + "bbox": [ + 89, + 194, + 138, + 206 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 76, + 215, + 398, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 215, + 398, + 229 + ], + "spans": [ + { + "bbox": [ + 76, + 215, + 398, + 229 + ], + "type": "text", + "content": "□ A4. Have you used AI writing assistants when working on this paper?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 89, + 230, + 138, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 230, + 138, + 242 + ], + "spans": [ + { + "bbox": [ + 89, + 230, + 138, + 242 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 68, + 253, + 292, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 253, + 292, + 266 + ], + "spans": [ + { + "bbox": [ + 68, + 253, + 292, + 266 + ], + "type": "text", + "content": "B Did you use or create scientific artifacts?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 79, + 270, + 128, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 270, + 128, + 283 + ], + "spans": [ + { + "bbox": [ + 79, + 270, + 128, + 283 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 76, + 292, + 315, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 292, + 315, + 306 + ], + "spans": [ + { + "bbox": [ + 76, + 292, + 315, + 306 + ], + "type": "text", + "content": "B1. Did you cite the creators of artifacts you used?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 89, + 306, + 138, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 306, + 138, + 319 + ], + "spans": [ + { + "bbox": [ + 89, + 306, + 138, + 319 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 76, + 328, + 463, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 328, + 463, + 342 + ], + "spans": [ + { + "bbox": [ + 76, + 328, + 463, + 342 + ], + "type": "text", + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 89, + 343, + 138, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 343, + 138, + 355 + ], + "spans": [ + { + "bbox": [ + 89, + 343, + 138, + 355 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 76, + 364, + 524, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 364, + 524, + 417 + ], + "spans": [ + { + "bbox": [ + 76, + 364, + 524, + 417 + ], + "type": "text", + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 89, + 419, + 138, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 419, + 138, + 432 + ], + "spans": [ + { + "bbox": [ + 89, + 419, + 138, + 432 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 76, + 441, + 524, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 441, + 524, + 481 + ], + "spans": [ + { + "bbox": [ + 76, + 441, + 524, + 481 + ], + "type": "text", + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 89, + 482, + 138, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 482, + 138, + 495 + ], + "spans": [ + { + "bbox": [ + 89, + 482, + 138, + 495 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 76, + 504, + 524, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 504, + 524, + 531 + ], + "spans": [ + { + "bbox": [ + 76, + 504, + 524, + 531 + ], + "type": "text", + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 89, + 532, + 138, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 532, + 138, + 544 + ], + "spans": [ + { + "bbox": [ + 89, + 532, + 138, + 544 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 76, + 554, + 524, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 554, + 524, + 622 + ], + "spans": [ + { + "bbox": [ + 76, + 554, + 524, + 622 + ], + "type": "text", + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 89, + 623, + 138, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 623, + 138, + 634 + ], + "spans": [ + { + "bbox": [ + 89, + 623, + 138, + 634 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 68, + 644, + 295, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 644, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 68, + 644, + 295, + 657 + ], + "type": "text", + "content": "C Did you run computational experiments?" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 79, + 661, + 128, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 661, + 128, + 674 + ], + "spans": [ + { + "bbox": [ + 79, + 661, + 128, + 674 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 76, + 684, + 524, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 684, + 524, + 711 + ], + "spans": [ + { + "bbox": [ + 76, + 684, + 524, + 711 + ], + "type": "text", + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 89, + 712, + 138, + 724 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 712, + 138, + 724 + ], + "spans": [ + { + "bbox": [ + 89, + 712, + 138, + 724 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "spans": [ + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "type": "text", + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "spans": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "text", + "content": "ACL 2023 Responsible NLP Checklist" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "839" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 76, + 71, + 524, + 238 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 76, + 71, + 523, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 71, + 523, + 111 + ], + "spans": [ + { + "bbox": [ + 76, + 71, + 523, + 111 + ], + "type": "text", + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Left blank." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 76, + 121, + 524, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 121, + 524, + 174 + ], + "spans": [ + { + "bbox": [ + 76, + 121, + 524, + 174 + ], + "type": "text", + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Left blank." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 76, + 184, + 524, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 184, + 524, + 238 + ], + "spans": [ + { + "bbox": [ + 76, + 184, + 524, + 238 + ], + "type": "text", + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Left blank." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 67, + 247, + 522, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 247, + 522, + 260 + ], + "spans": [ + { + "bbox": [ + 67, + 247, + 522, + 260 + ], + "type": "text", + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 79, + 264, + 127, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 264, + 127, + 277 + ], + "spans": [ + { + "bbox": [ + 79, + 264, + 127, + 277 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 76, + 286, + 524, + 539 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 76, + 286, + 524, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 286, + 524, + 326 + ], + "spans": [ + { + "bbox": [ + 76, + 286, + 524, + 326 + ], + "type": "text", + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Left blank." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 76, + 336, + 524, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 336, + 524, + 390 + ], + "spans": [ + { + "bbox": [ + 76, + 336, + 524, + 390 + ], + "type": "text", + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Left blank." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 76, + 399, + 524, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 399, + 524, + 454 + ], + "spans": [ + { + "bbox": [ + 76, + 399, + 524, + 454 + ], + "type": "text", + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Left blank." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 76, + 462, + 519, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 462, + 519, + 489 + ], + "spans": [ + { + "bbox": [ + 76, + 462, + 519, + 489 + ], + "type": "text", + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Left blank." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 76, + 498, + 524, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 498, + 524, + 539 + ], + "spans": [ + { + "bbox": [ + 76, + 498, + 524, + 539 + ], + "type": "text", + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Left blank." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 790 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 790 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 790 + ], + "type": "text", + "content": "840" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 13 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/When to Use Efficient Self Attention_ Profiling Text, Speech and Image Transformer Variants/8b2664b3-8c7e-4af7-bd6b-26db06c033c3_content_list.json b/2023/When to Use Efficient Self Attention_ Profiling Text, Speech and Image Transformer Variants/8b2664b3-8c7e-4af7-bd6b-26db06c033c3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a4508929517dcf9854ce50d1d5324bfeca66ef52 --- /dev/null +++ b/2023/When to Use Efficient Self Attention_ Profiling Text, Speech and Image Transformer Variants/8b2664b3-8c7e-4af7-bd6b-26db06c033c3_content_list.json @@ -0,0 +1,2027 @@ +[ + { + "type": "text", + "text": "When to Use Efficient Self Attention? Profiling Text, Speech and Image Transformer Variants", + "text_level": 1, + "bbox": [ + 210, + 90, + 786, + 130 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anuj Diwan, Eunsol Choi, David Harwath", + "bbox": [ + 315, + 149, + 684, + 165 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Department of Computer Science", + "bbox": [ + 363, + 167, + 638, + 181 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The University of Texas at Austin", + "bbox": [ + 363, + 183, + 638, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{anuj.diwan, eunsol, harwath}@utexas.edu", + "bbox": [ + 299, + 200, + 704, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 252, + 339, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present the first unified study of the efficiency of self-attention-based Transformer variants spanning text, speech and vision. We identify input length thresholds (tipping points) at which efficient Transformer variants become more efficient than vanilla models, using a variety of efficiency metrics (latency, throughput, and memory). To conduct this analysis for speech, we introduce L-HuBERT, a novel local-attention variant of a self-supervised speech model. We observe that these thresholds are (a) much higher than typical dataset sequence lengths and (b) dependent on the metric and modality, showing that choosing the right model depends on modality, task type (long-form vs. typical context) and resource constraints (time vs. memory). By visualising the breakdown of the computational costs for transformer components, we also show that non-self-attention components exhibit significant computational costs. We release our profiling toolkit at https://github.com/ajd12342/profiling-transformers.", + "bbox": [ + 141, + 279, + 460, + 605 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction and Related Work", + "text_level": 1, + "bbox": [ + 114, + 618, + 421, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Transformers (Vaswani et al., 2017) are widely adopted across NLP (Devlin et al., 2019; Brown et al., 2020), Speech Processing (Mohamed et al., 2022) and Computer Vision (Dosovitskiy et al., 2021). Studies have shown that scaling models up improves performance (Chowdhery et al., 2022), making efficiency an important research topic. Many Transformer variants focus on improving the efficiency of self-attention, motivated by its asymptotic quadratic time/space complexity with respect to the input sequence length. While these Transformer variants are designed be asymptotically faster, in practice they may actually be slower, especially given modest input lengths that are typical of many tasks.", + "bbox": [ + 112, + 643, + 489, + 885 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our paper presents two main analyses. First, we visualize the layerwise efficiency of such models to locate bottlenecks and attempt to answer the question \"is self-attention the true bottleneck?\" We find that in the non-asymptotic case, non-self-attention layers contribute significantly to the overall cost, especially for speech architectures due to the input waveform tokenizer in models like HuBERT (Hsu et al., 2021). Second, when should we use self-attention-based efficient Transformers? Comparing efficient variants with vanilla models at different input lengths, we find that this tipping point where efficient variants outperform vanilla architectures is much higher than typical input lengths of existing benchmarks across all modalities, calling into question the efficacy of using such efficient Transformers and requiring new benchmarks. We introduce a local-attention variant of a speech Transformer, HuBERT, to conduct this analysis. Together, our analyses suggest that current approaches that focus on improving self-attention might not be the most effective for improving efficiency.", + "bbox": [ + 507, + 253, + 884, + 607 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Efficiency Metrics", + "text_level": 1, + "bbox": [ + 509, + 619, + 700, + 634 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Model efficiency is an umbrella term for a suite of efficiency metrics, which do not always correlate with, and sometimes contradict, each other (Dehghani et al., 2022). Further, different metrics are relevant to different end use-cases. To cover most use-cases, we evaluate a set of four metrics; two for computational time and two for memory usage: Throughput: Number of examples processed per sec, given inputs of a given sequence length, using the maximum possible batch size for a given GPU. Latency-Inference: Time (in ms) to run inference for 1 unbatched input of a given sequence length. Max-Memory: The allocated GPU memory (MiB) for processing 1 input of a given sequence length. Parameter Count: Number of model parameters.", + "bbox": [ + 507, + 644, + 884, + 885 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We profile models in all modalities in training mode and inference mode. For training, while", + "bbox": [ + 507, + 887, + 882, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1We refer the readers to Tay et al. (2022) for a comprehensive overview of efficient Transformers.", + "bbox": [ + 112, + 891, + 489, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1639", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics", + "bbox": [ + 226, + 945, + 771, + 957 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Volume 2: Short Papers, pages 1639-1650", + "bbox": [ + 368, + 958, + 630, + 971 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "July 9-14, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 295, + 972, + 700, + 985 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6366a7ff2881849f4c23a64787c4e27fa3d208d30074e37d3446ccd27d22e935.jpg", + "image_caption": [ + "Figure 1: Transformer layer types profiled in our layerwise efficiency profiling experiments." + ], + "image_footnote": [], + "bbox": [ + 163, + 93, + 371, + 331 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Transformer architectures often use prediction heads with a larger output space (e.g., for text generation), we choose a lightweight binary classification head for profiling.", + "bbox": [ + 112, + 388, + 489, + 453 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Layerwise Efficiency Metrics We also profile some metrics and models in a layerwise fashion to locate their efficiency bottlenecks. Our goal is twofold: a) provide an empirical approach to efficient model design, as an alternative to theoretical analyses or mental models (e.g. self-attention is $O(n^{2})$ ) and b) empirically answer the question \"to what degree is self-attention the bottleneck?\"", + "bbox": [ + 112, + 462, + 487, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We identify important layer types (Self-Attention, Feedforward, etc.) and profile the Latency-Inference and Parameter Count metrics per-layer-type to obtain a fine-grained understanding of which layer types and indices (layer 0 vs 11) contribute the most to model efficiency costs. We use param counts as a proxy for memory (profiling real layerwise memory usage is non-trivial due to Pytorch memory allocation intricacies). We profile the layers depicted in Figure 1; more details in Appendix E.", + "bbox": [ + 112, + 590, + 489, + 769 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3 Local-Attention Speech Model", + "text_level": 1, + "bbox": [ + 112, + 780, + 413, + 797 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Efficient transformers (Xiong et al., 2021; Ma et al., 2021) have not received as much attention in Speech as they have in NLP and CV, perhaps due to two reasons. First, there is a relative lack of long-context speech benchmarks as compared to those in NLP (LRA (Tay et al., 2021) and QuALITY (Pang et al., 2022)). Second, when performing speech", + "bbox": [ + 112, + 806, + 489, + 919 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/7c3aa4125714a0fa17abbecfdd53d5338c032f90bbee14c8456026120ab1c736.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelWER ↓WER (w/ FT) ↓
HuBERT Base7.093.4
L-HuBERT (32 | 100)21.06 | 14.488.52 | 7.39
", + "bbox": [ + 515, + 80, + 877, + 134 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1: WERs on the SUPERB ASR task.", + "bbox": [ + 547, + 143, + 842, + 158 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tasks like automatic speech recognition (ASR), it is typical to segment a long speech signal into small individual utterances and perform ASR independently on each. For example, most Librispeech examples are less than 5 seconds. Many popular speech models like HuBERT (Hsu et al., 2021) tokenize the waveform at 50 tokens per second, implying that a typical utterance has only several hundred tokens; far below the number of tokens in long-context NLP tasks. Nevertheless, textless speech models (Lakhotia et al., 2021) are becoming more feasible, motivating the modelling of long speech utterances.", + "bbox": [ + 507, + 166, + 884, + 376 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Local HuBERT Model To investigate the efficiency of the self-attention layer in speech models, we introduce the Local HuBERT model which replaces HuBERT's self-attention with the Longformer (Beltagy et al., 2020) sliding-window self-attention. In this attention mechanism, every token attends to tokens within a local window context, rather than the full token sequence. Our model is similar to the temporally windowed-attention Transformer acoustic model proposed by Alastruey et al. (2021) for speech translation; our approach differs by using the self-supervised HuBERT model as our basis, and we evaluate on ASR. Choosing the appropriate window size for the local attention context is key; we explore 32 and 100 token contexts, corresponding to $640~\\mathrm{ms}$ and $2\\mathrm{s}$ , inspired by phone recognition models that typically incorporate similar context sizes (Peddinti et al., 2015; feng Yeh et al., 2019).", + "bbox": [ + 507, + 385, + 884, + 692 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ASR Results We initialize the L-HuBERT model with pretrained HuBERT Base weights (pretrained with full self-attention), and then replace self-attention with sliding-window self-attention; due to limited compute, we did not pretrain L-HuBERT from scratch using sliding-window attention. We then evaluate L-HuBERT on Librispeech (Panayotov et al., 2015) ASR via the SUPERB (Yang et al., 2021) benchmark under two settings; a) Freeze: freezing the model and only training projection weights and b) Finetune: fully finetune the model. We use the default S3PRL2 hyperparams; but we", + "bbox": [ + 507, + 701, + 885, + 896 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "2https://github.com/s3pr1/s3pr1", + "bbox": [ + 529, + 903, + 769, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1640", + "bbox": [ + 482, + 927, + 521, + 940 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/1f0876a34676d6c3cd5854b59abf5785d826528b49fc2572289352904ed667b4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelEmbPosSAIntermOutputOthers
BERT23.8M-29M28.3M28.3M0.6M
HuBERT4.2M5.1M29M28.3M28.3M0.2M
ViT0.6M-29M28.3M28.3M0.6M
", + "bbox": [ + 134, + 80, + 467, + 146 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 2: Layer-wise parameter counts. Emb: Input Embedding, Pos: Positional Emb. SA: Self-Attention, Interm: Intermediate.", + "bbox": [ + 112, + 155, + 487, + 198 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "train for 200k steps for Freeze and 104k steps for Finetune. Both models converge by 104k steps; we train Freeze for longer to eke out as much performance as possible, while we stop training Finetune due to limited compute.", + "bbox": [ + 112, + 212, + 487, + 294 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We report Word Error Rate (WER) on Librispeech test-clean in Table 1; lower is better. In the frozen setting (middle column), we see a large WER increase over HuBERT; we hypothesize that this is due to the attention layer mismatch since we initialize L-HuBERT with HuBERT weights that were pretrained with full self attention, rather than pretraining L-HuBERT from scratch. However, in the finetuning setting, the gap between HuBERT Base and L-HuBERT narrows considerably and using a larger window size achieves better performance. As our L-HuBERT model is a reasonable architecture capable of moderate ASR performance, we can continue to study its computational efficiency (we profile the window-100 variant).", + "bbox": [ + 115, + 294, + 489, + 536 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4 Methods and Implementation", + "text_level": 1, + "bbox": [ + 112, + 550, + 405, + 567 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We analyze the Base versions of the BERT (Devlin et al., 2019), Longformer (Beltagy et al., 2020) and Nystromformer (Xiong et al., 2021) models for text; the HuBERT (Hsu et al., 2021) and L-HuBERT (Section 3) models for speech; and Vision Transformer (Dosovitskiy et al., 2021) and Swin Transformer (Liu et al., 2021) models for vision; BERT, HuBERT and ViT are standard Transformer encoder architectures. Longformer, L-HuBERT and Swin use fixed-pattern self-attention while Nystromformer uses approximate self-attention.", + "bbox": [ + 112, + 577, + 489, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1 Sequence Length Ranges", + "text_level": 1, + "bbox": [ + 112, + 768, + 356, + 784 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We profile our models on a wide range of input sequence lengths to cover both avg. sequence lengths of commonly used contemporary datasets (Table 3) and typical sequence lengths of long-context tasks. Details about how we compute sequence lengths in Table 3 can be found in Appendix B. Most image datasets use images resized to 224 or 512 pixels. Below, $\\text{range}(a, b, c)$ means a range from $a$ to $b$", + "bbox": [ + 112, + 790, + 489, + 917 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "in steps of $c$ . Since there is no difference between synthetic and real inputs from a computational complexity standpoint, we use synthetic inputs to more easily control for their sequence lengths.", + "bbox": [ + 507, + 84, + 882, + 148 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text Modality The input is 'This is a sentence.' repeated $n$ times, $n \\in \\text{range}(10, 560, 10)$ i.e. range(62, 3362, 60) tokens for all tokenizers.", + "bbox": [ + 507, + 149, + 884, + 197 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Speech Modality The inputs have durations in range(1, 50, 0.5) sec i.e. range(50, 2500, 25) tokens for all featurizers (CNNs with 20 ms framerate). Our sampling strategy is in Appendix A.", + "bbox": [ + 507, + 199, + 882, + 262 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image Modality We use square inputs of dimension in range(32, 1024, 32) pixels by rescaling a fixed image. The # tokens depend on featurizer patch size, which is different for different models.", + "bbox": [ + 507, + 263, + 882, + 326 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.2 Implementational Details", + "text_level": 1, + "bbox": [ + 507, + 342, + 754, + 357 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We profile time-based metrics (latency/throughput) using Pytorch CUDA Events3 by executing 20 iterations sequentially. The first few iterations serve as GPU warm-start; thus, we report the average of the last 10. We record Max-Memory with torch.cuda.max_memory_allocated() and param counts with torchinfo (Yep, 2020).", + "bbox": [ + 507, + 363, + 882, + 476 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To profile throughput, we approximate the max batch size that fits on a single GPU using a linear estimator; more details in Appendix C. Finally, we profile the layerwise Latency-Inference metric using torchprof (Wong, 2020). We attach profiling hooks to modules of interest (e.g. Self-Attention, Embedding), giving us execution times of their forward() functions (other modules/functions are not profiled). We use the Huggingface (Wolf et al., 2020) implementations of text and image models and fairseq (Ott et al., 2019) implementations for speech models; more details in Appendix D.", + "bbox": [ + 507, + 478, + 882, + 671 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "5 Profiling Results", + "text_level": 1, + "bbox": [ + 507, + 684, + 687, + 701 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "5.1 Layerwise Profiling Results", + "text_level": 1, + "bbox": [ + 507, + 713, + 769, + 728 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 2 shows the layerwise Latency-Inference for all 3 vanilla architectures in each modality. Figures for efficient models are in Appendix F. Color darkness represents the layer index (layer 0 is darkest). Table 2 shows the layerwise param count.", + "bbox": [ + 507, + 734, + 882, + 815 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Asymptotically, self-attention dominates the computation. However, since the average seq length for most text and speech tasks is less than 1000 tokens and most image datasets are used at", + "bbox": [ + 507, + 816, + 882, + 879 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "3https://pytorch.org/docs/stable/generated/torch.cuda.Event.html", + "bbox": [ + 509, + 891, + 857, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1641", + "bbox": [ + 482, + 928, + 517, + 940 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/39ef649a7c060b17c65f7004c4d275cb416c10082ab41b5665f8a47b47e52ca7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TextSpeech
Dataset # of tokensSST 23MNLI 36SQ 177ON 506CNN 863HPQA 1,316TQA 6,589TEDL 301LJS 328VoxC 390Libri 615S-SQuAD 3080Spotify 101400
", + "bbox": [ + 129, + 80, + 868, + 133 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 3: Average token sequence lengths. Left to right: Stanford Sentiment Treebank, MultiNLI, SQuAD2.0, OntoNotes, CNN-DailyMail, HotpotQA, TriviaQA, TEDLIUM, LJSpeech, VoxCeleb Speaker Recognition, Librispeech, Spoken SQuAD, Spotify Podcasts.", + "bbox": [ + 112, + 137, + 884, + 183 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2dc8d48cf017d296c3e1dce24a493e6eefd6a3ba25dee3586c4abcf985ed3a99.jpg", + "image_caption": [ + "Figure 2: Layerwise latency of different vanilla Transformer architectures in inference mode." + ], + "image_footnote": [], + "bbox": [ + 159, + 193, + 868, + 354 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8cae2bfbc1133a9c1217e49b9329d11ff210a9f328dc8e2a38b8574465775c33.jpg", + "image_caption": [ + "Figure 3: Overall Inference-time Profiling Results. Text and speech models in first row, vision models in second." + ], + "image_footnote": [], + "bbox": [ + 189, + 376, + 806, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "a max dimension of 512, at these points, non-self-attention components take up $35\\%$ , $58.8\\%$ and $43.75\\%$ latency for NLP, speech and images. Additionally, parameter counts of SA are also comparable to Interm/Output layers. This shows that it is also important to direct efficiency efforts for other model components.", + "bbox": [ + 112, + 695, + 489, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While the latency associated with embedding layers is minimal for BERT, they are sizable for HuBERT. HuBERT uses a CNN feature extractor with different strides and kernel sizes and consumes more time in the earlier CNN layers as opposed to later ones, as is visible in Figure 2, which shows", + "bbox": [ + 112, + 822, + 489, + 919 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "darker shades i.e. earlier layers dominating the computation. Optimal efficiency strategies can thus differ across modalities, e.g. Wu et al. (2022) slims down this CNN feature extractor embedding layer. On the other hand, embedding layers take up a lot of parameters in BERT; thus, it may be helpful to shrink the BERT embedding layer for memory purposes (as opposed to latency for HuBERT). Finally, analyzing Transformer variants (Appendix F), we see that self-attention in Longformer, Swin and L-HuBERT encouragingly scales latency linearly, but with large overhead for smaller inputs.", + "bbox": [ + 507, + 695, + 884, + 889 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1642", + "bbox": [ + 482, + 928, + 521, + 940 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c8dfd713d4b58aeab8460506b9f6b169349523c936c974429fc483c1dcbdf442.jpg", + "image_caption": [ + "Training-Throughput" + ], + "image_footnote": [], + "bbox": [ + 250, + 112, + 485, + 233 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/885ee6c268af0864b801e34d825ed27426d4a43a17dd6a3873856d4e459defeb.jpg", + "image_caption": [ + "Training-Max-Memory" + ], + "image_footnote": [], + "bbox": [ + 510, + 115, + 744, + 233 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/285c81d7a81a18a21d06935526ec682e6da54ebea76ece4d543d9fee41a78179.jpg", + "image_caption": [ + "Figure 4: Overall Training-time Profiling Results. Text and speech models in first row, vision models in second." + ], + "image_footnote": [], + "bbox": [ + 252, + 235, + 485, + 355 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c19367041d0a197782c686d10827d610859db53f3505c2e1a485e6c1377afce1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 510, + 237, + 744, + 355 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2 Overall Profiling Results", + "text_level": 1, + "bbox": [ + 112, + 394, + 354, + 410 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our profiling results are in Figures 3 and 4. Inference Throughput is in the Appendix at Figure 6, exhibiting similar trends as training Throughput.", + "bbox": [ + 112, + 414, + 489, + 464 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Tipping Point Analysis We see that most variants are slower and more memory hungry than vanilla models for input lengths of typical-context tasks. We define the tipping point for each modality: the input length at which the variant becomes more efficient than the vanilla model. For text and speech, it is $1750 - 2000$ tokens for inference latency and max-memory, greater than typical input lengths (Table 3). However, while the tipping point for training max-memory is $\\approx 1500$ tokens for text (still a large number), it is $\\approx 0 - 250$ for speech, an encouraging result. For images, it is $500 - 700$ pixels for all metrics apart from throughput. This is less reasonable for 224 pixel datasets but good for high resolution image datasets (512/1024). All variants are either worse or comparable than vanilla models across modalities for throughput.", + "bbox": [ + 112, + 474, + 487, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We hypothesize that some efficient models suffer from additional overheads; while vanilla attention benefits from highly optimized matrix multiplication, windowed attention requires complex reshaping and preprocessing.", + "bbox": [ + 112, + 747, + 489, + 829 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Choosing the Right Model Depends on Resource Constraints Our results show that the choice of the right model depends on resource constraints. Suppose that one is training models under a time constraint; then, throughput is the bottleneck and", + "bbox": [ + 112, + 839, + 489, + 917 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "efficient models would not be a good fit. On the other hand, efficient models are useful for long context memory-constrained inference.", + "bbox": [ + 507, + 394, + 882, + 442 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Local Attention and Excessive Padding The Longformer pads input lengths to be a multiple of 512 and Swin requires input dimension to be a multiple of 224. This slows shorter inputs down and results in extremely low performance (measured by all 3 metrics) as compared to vanilla models.", + "bbox": [ + 507, + 457, + 884, + 552 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparing Parameter Counts The Longformer uses more parameters compared to vanilla BERT (148M vs. 109M) because it uses two sets of Q,K,V projection matrices for its global and local attention operations; sharing these may decrease its memory usage. For other modalities, efficient models do not incur more parameters.", + "bbox": [ + 507, + 567, + 882, + 678 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 507, + 697, + 640, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We present an empirical efficiency analysis of vanilla Transformers and their self-attention-based efficient variants across modalities, metrics and input context sizes. We find substantial differences across modalities and metrics when analyzing the tipping point for efficient variants. Finally, the layerwise analysis finds that self-attention is not the only bottleneck. We recommend that all efficient model papers should report such cross-modal, layerwise profiling results on multiple efficiency metrics covering a variety of use-cases to provide a full picture of the benefits of the model.", + "bbox": [ + 507, + 726, + 884, + 917 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "■ BERT Nyströmformer Longformer HuBERT L-HuBERT ViT Swin", + "bbox": [ + 191, + 80, + 803, + 95 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1643", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Limitations", + "text_level": 1, + "bbox": [ + 114, + 84, + 220, + 99 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We focus primarily on comparing model efficiencies using a variety of efficiency metrics and do not consider model performance; one can perform a more elaborate analysis of performance-efficiency tradeoffs, which we did not do here.", + "bbox": [ + 112, + 110, + 489, + 187 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We only profile a total of seven models across three modalities while there are more efficient variants and vanilla Transformers proposed in the literature. While we choose our models to be as representative of each modality and efficiency technique as possible, we cannot extrapolate results to other model variants and other modalities. In particular, modalities like video and genomics and efficiency approaches like quantization would be interesting to profile, which we did not do.", + "bbox": [ + 112, + 191, + 489, + 350 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 114, + 363, + 285, + 380 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We thank the reviewers and the meta-reviewer of the ACL community for helpful feedback on the draft. This work was partially funded by a grant from UT Machine Learning Lab.", + "bbox": [ + 112, + 390, + 489, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 114, + 481, + 213, + 495 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Belen Alastruey, Gerard I. Gållego, and Marta R. Costajussa. 2021. Efficient Transformer for Direct Speech Translation.", + "Iz Beltagy, Matthew E. Peters, and Arman Cohan. 2020. Longformer: The Long-Document Transformer. ArXiv preprint, abs/2004.05150.", + "Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language Models are Few-Shot Learners. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual.", + "Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob" + ], + "bbox": [ + 115, + 504, + 489, + 919 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan Firat, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. 2022. PaLM: Scaling Language Modeling with Pathways.", + "Ann Clifton, Sravana Reddy, Yongze Yu, Aasish Pappu, Rezvaneh Rezapour, Hamed Bonab, Maria Eskevich, Gareth Jones, Jussi Karlgren, Ben Carterette, and Rosie Jones. 2020. 100,000 Podcasts: A Spoken English Document Corpus. In Proceedings of the 28th International Conference on Computational Linguistics, pages 5903-5917, Barcelona, Spain (Online). International Committee on Computational Linguistics.", + "Mostafa Dehghani, Yi Tay, Anurag Arnab, Lucas Beyer, and Ashish Vaswani. 2022. The Efficiency Mismonomer. In *The Tenth International Conference on Learning Representations*, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net.", + "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics.", + "Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. 2021. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net.", + "Ching feng Yeh, Jay Mahadeokar, Kaustubh Kalgaonkar, Yongqiang Wang, Duc Le, Mahaveer Jain, Kjell Schubert, Christian Fuegen, and Michael L. Seltzer. 2019. Transformer-transducer: End-to-end speech recognition with self-attention. ArXiv, abs/1910.12977.", + "Karl Moritz Hermann, Tomás Kocisky, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa Suleyman, and Phil Blunsom. 2015. Teaching Machines to Read and Comprehend. In Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December" + ], + "bbox": [ + 510, + 85, + 884, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1644", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "7-12, 2015, Montreal, Quebec, Canada, pages 1693-1701.", + "François Hernandez, Vincent Nguyen, Sahar Ghannay, Natalia Tomashenko, and Yannick Esteve. Ted-lium 3: Twice as much data and corpus repartition for experiments on speaker adaptation. In Speech and Computer, pages 198-208. Springer International Publishing.", + "Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. 2021. HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 29:3451-3460.", + "Keith Ito and Linda Johnson. 2017. The LJ Speech Dataset. https://keithito.com/LJ-Speech-Dataset/.", + "Mandar Joshi, Eunsol Choi, Daniel Weld, and Luke Zettlemoyer. 2017. TriviaQA: A large scale distantly supervised challenge dataset for reading comprehension. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1601-1611, Vancouver, Canada. Association for Computational Linguistics.", + "Kushal Lakhotia, Eugene Kharitonov, Wei-Ning Hsu, Yossi Adi, Adam Polyak, Benjamin Bolte, Tu-Anh Nguyen, Jade Copet, Alexei Baevski, Abdelrahman Mohamed, and Emmanuel Dupoux. 2021. On generative spoken language modeling from raw audio. Transactions of the Association for Computational Linguistics, 9:1336-1354.", + "Chia-Hsuan Li, Szu-Lin Wu, Chi-Liang Liu, and Hung-yi Lee. 2018. Spoken SQuAD: A Study of Mitigating the Impact of Speech Recognition Errors on Listening Comprehension. In *Interspeech* 2018, 19th Annual Conference of the International Speech Communication Association, Hyderabad, India, 2-6 September 2018, pages 3459-3463. ISCA.", + "Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. 2021. Swin Transformer: Hierarchical Vision Transformer using Shfted Windows. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 9992-10002. IEEE.", + "Xuezhe Ma, Xiang Kong, Sinong Wang, Chunting Zhou, Jonathan May, Hao Ma, and Luke Zettlemoyer. 2021. Luna: Linear Unified Nested Attention. In NeurIPS.", + "Abdelrahman Mohamed, Hung yi Lee, Lasse Borgholt, Jakob D. Havtorn, Joakim Edin, Christian Igel, Katrin Kirchhoff, Shang-Wen Li, Karen Livescu, Lars Maaloe, Tara N. Sainath, and Shinji Watanabe. 2022. Self-Supervised Speech Representation Learning: A Review. IEEE Journal of Selected Topics in Signal Processing, 16(6):1179-1210." + ], + "bbox": [ + 115, + 85, + 489, + 917 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Arsha Nagrani, Joon Son Chung, and Andrew Zisserman. 2017. VoxCeleb: A Large-Scale Speaker Identification Dataset. In *Interspeech* 2017, 18th Annual Conference of the International Speech Communication Association, Stockholm, Sweden, August 20-24, 2017, pages 2616-2620. ISCA.", + "Myle Ott, Sergey Edunov, Alexei Baevski, Angela Fan, Sam Gross, Nathan Ng, David Grangier, and Michael Auli. 2019. *fairoseq: A fast, extensible toolkit for sequence modeling.* In *Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics* (Demonstrations), pages 48-53, Minneapolis, Minnesota. Association for Computational Linguistics.", + "Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. 2015. Librispeech: An ASR corpus based on public domain audio books. In 2015 IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP 2015, South Brisbane, Queensland, Australia, April 19-24, 2015, pages 5206-5210. IEEE.", + "Richard Yuanzhe Pang, Alicia Parrish, Nitish Joshi, Nikita Nangia, Jason Phang, Angelica Chen, Vishakh Padmakumar, Johnny Ma, Jana Thompson, He He, and Samuel Bowman. 2022. Quality: Question answering with long input texts, yes! In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5336-5358, Seattle, United States. Association for Computational Linguistics.", + "Vijayaditya Peddinti, Daniel Povey, and Sanjeev Khudanpur. 2015. A time delay neural network architecture for efficient modeling of long temporal contexts. In Proc. Interspeech 2015, pages 3214-3218.", + "Sameer S. Pradhan and Nianwen Xue. 2009. OntoNotes: The $90\\%$ solution. In Proceedings of Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the Association for Computational Linguistics, Companion Volume: Tutorial Abstracts, pages 11-12, Boulder, Colorado. Association for Computational Linguistics.", + "Pranav Rajpurkar, Robin Jia, and Percy Liang. 2018. Know what you don't know: Unanswerable questions for SQuAD. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 784-789, Melbourne, Australia. Association for Computational Linguistics.", + "Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1631-1642, Seattle, Washington, USA. Association for Computational Linguistics." + ], + "bbox": [ + 510, + 85, + 882, + 917 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1645", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Yi Tay, Mostafa Dehghani, Samira Abnar, Yikang Shen, Dara Bahri, Philip Pham, Jinfeng Rao, Liu Yang, Sebastian Ruder, and Donald Metzler. 2021. Long Range Arena: A Benchmark for Efficient Transformers. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net.", + "Yi Tay, Mostafa Dehghani, Dara Bahri, and Donald Metzler. 2022. Efficient Transformers: A Survey. In ACM Comput. Surv., volume 55, New York, NY, USA. Association for Computing Machinery.", + "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is All You Need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 5998-6008.", + "Adina Williams, Nikita Nangia, and Samuel Bowman. 2018. A broad-coverage challenge corpus for sentence understanding through inference. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1112-1122, New Orleans, Louisiana. Association for Computational Linguistics.", + "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online. Association for Computational Linguistics.", + "Alexander William Wong. 2020. torchprof. https://github.com/awwong1/torchprof.", + "Felix Wu, Kwangyoun Kim, Jing Pan, Kyu J. Han, Kilian Q. Weinberger, and Yoav Artzi. 2022. Performance-Efficiency Trade-Offs in Unsupervised Pre-Training for Speech Recognition. In ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7667-7671.", + "Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, and Vikas Singh. 2021. Nystromformer: A Nystrom-based Algorithm for Approximating Self-Attention. In Thirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2021, Thirty-Third Conference on Innovative Applications of Artificial Intelligence, IAAI 2021, The Eleventh Symposium on Educational Advances in Artificial Intelligence, EAAI 2021, Virtual Event, February 2-9, 2021, pages 14138-14148. AAAI Press." + ], + "bbox": [ + 115, + 85, + 489, + 917 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Shu-Wen Yang, Po-Han Chi, Yung-Sung Chuang, Cheng-I Jeff Lai, Kushal Lakhotia, Yist Y. Lin, Andy T. Liu, Jiatong Shi, Xuankai Chang, Guanting Lin, Tzu-Hsien Huang, Wei-Cheng Tseng, Kotik Lee, Da-Rong Liu, Zili Huang, Shuyan Dong, Shang-Wen Li, Shinji Watanabe, Abdelrahman Mohamed, and Hung-yi Lee. 2021. SUPERB: Speech Processing Universal PERformance Benchmark. In Interspeech 2021, 22nd Annual Conference of the International Speech Communication Association, Brno, Czechia, 30 August - 3 September 2021, pages 1194-1198. ISCA.", + "Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. 2018. HotpotQA: A dataset for diverse, explainable multi-hop question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2369-2380, Brussels, Belgium. Association for Computational Linguistics.", + "Tyler Yep. 2020. torchinfo. https://github.com/ TylerYep/torchinfo." + ], + "bbox": [ + 509, + 85, + 884, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A Sampling Speech Utterances for Profiling", + "text_level": 1, + "bbox": [ + 509, + 405, + 823, + 439 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To obtain speech inputs of length $i$ seconds to $i + 0.5$ seconds for all $i$ less than 12 seconds, we sample 5 speech utterances from the training set of the Librispeech dataset (Panayotov et al., 2015) whose lengths fall within this range and compute aggregate metrics over these 5 utterances. Since the Librispeech dataset does not contain extremely long speech utterances, for $i$ of length greater than 12 seconds, we adopt a different approach to generate inputs. To generate such an input utterance of length between $i$ and $i + 0.5$ seconds, we first sample 5 speech utterances from the Librispeech training set of input length ranging from $\\frac{i}{5}$ to $\\frac{i + 0.5}{5}$ and concatenate them to obtain utterances of length ranging from $i$ to $i + 0.5$ as desired. We do this 5 times to get 5 different utterances and compute aggregate metrics over these 5 utterances.", + "bbox": [ + 507, + 447, + 882, + 721 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "B Computing Token Lengths for NLP and Speech Datasets", + "text_level": 1, + "bbox": [ + 509, + 732, + 853, + 765 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We compute average sequence token lengths for 7 NLP datasets and 6 speech datasets. For all speech datasets, we compute mean utterance durations and multiply durations by 50 to obtain number of tokens (model framereates are $20\\mathrm{ms}$ i.e. $\\times 50$ ). For TEDLIUM (Hernandez et al.), LJSpeech (Ito and Johnson, 2017), VoxCeleb Speaker Recognition Dataset (Nagrani et al., 2017) and Librispeech (Panayotov et al., 2015), we compute", + "bbox": [ + 507, + 774, + 884, + 919 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1646", + "bbox": [ + 482, + 928, + 521, + 940 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "mean validation-set utterance durations; for Spoken SQuAD (Li et al., 2018), we report mean validation-set paragraph duration and for the Spotify English Podcasts dataset (Clifton et al., 2020), we report mean podcast duration directly obtained from Clifton et al. (2020).", + "bbox": [ + 112, + 84, + 487, + 179 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "SST (Socher et al., 2013). We use test-set sentences. We use the HuggingFace BERTTokenizer.", + "bbox": [ + 112, + 181, + 489, + 212 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "MNLI (Williams et al., 2018). We use validation-matched-set examples by concatenating the premise and the hypothesis. We use the HuggingFace BERTTokenizer.", + "bbox": [ + 112, + 214, + 487, + 275 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "SQuAD2.0 (Rajpurkar et al., 2018). We use validation-set examples by concatenating the context and the question. We use the HuggingFace BERTTokenizer.", + "bbox": [ + 112, + 278, + 487, + 341 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "OntoNotes (Pradhan and Xue, 2009). We obtain this number from the Longformer (Beltagy et al., 2020) paper.", + "bbox": [ + 112, + 343, + 487, + 390 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "CNN-Dailymail (Hermann et al., 2015). We use the 3.0.0 version of the dataset and use test-set articles. We use the HuggingFace BERTTokenizer.", + "bbox": [ + 112, + 391, + 487, + 438 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "HotpotQA (Yang et al., 2018). We obtain this number from the Longformer (Beltagy et al., 2020) paper.", + "bbox": [ + 112, + 439, + 487, + 488 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "TriviaQA (Joshi et al., 2017). We obtain this number from the Longformer (Beltagy et al., 2020) paper.", + "bbox": [ + 112, + 488, + 487, + 536 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "C Implementing Throughput Profiling", + "text_level": 1, + "bbox": [ + 112, + 548, + 465, + 565 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To profile Throughput, we need to compute the max batch size that can fit on a single GPU. We approximately predict this using a linear estimator as follows. We first record the memory $B$ reserved on the GPU after just loading the model. Next, we independently run batches of sizes 1 and 2 and record memory usages $M_{1}$ and $M_{2}$ . We use an NVIDIA Quadro RTX 8000 GPU with a maximum memory of 45000 MiB. Thus, assuming a linear relationship between batch size and memory consumption, we predict a maximum batch size of $bsz = \\frac{45000 - B}{M_2 - M_1}$ . In practice, this is an overestimate; we keep decreasing the batch size by a factor of 0.9 until no OOM errors occur and this is our final estimate.", + "bbox": [ + 112, + 575, + 489, + 799 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "D Implementational Details for Models", + "text_level": 1, + "bbox": [ + 112, + 812, + 470, + 829 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We use the following HuggingFace configurations: bert-base-uncased for BERT, allenai/longformer-base-4096 for Longformer, uw-madison/nystromformer-4096 for Nyströmformer,", + "bbox": [ + 112, + 839, + 489, + 917 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "google/vit-base-patch16-224 for ViT and microsoft/swin-base-patch4-window7-224 for Swin. The BERT model natively supports a maximum of 512 tokens as input because it has 512 positional embeddings; we modify the positional embedding computation to allow an arbitrarily long input to be provided. The Longformer internally pads all input lengths to a multiple of 512. For Swin, we pad images to have an input dimension that is a multiple of 224; this is necessary due to the windowed attention mechanism in Swin. In fact, the Swin model natively supports only a $224 \\times 224$ resolution; we make a small modification in order to support resolutions that are multiples of 224. We use the HuBERT Base model for both HuBERT and L-HuBERT.", + "bbox": [ + 507, + 84, + 882, + 356 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "E Transformer Layer Types", + "text_level": 1, + "bbox": [ + 507, + 370, + 769, + 386 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Input Embedding Layer. (red) Maps the input sequence into fixed-dimensional embeddings. This is a linear layer for text and a CNN featurizer for image/speech.", + "bbox": [ + 507, + 394, + 882, + 458 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Positional Embedding Layer. (fuchsia) For text and image models this is part of the input embedding layer. For speech models, this is a very wide convolution layer.", + "bbox": [ + 507, + 458, + 882, + 523 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Self Attention Layer.(■/blue) The multi-head self attention block, which computes self-attention outputs and maps the result to the model dimension.", + "bbox": [ + 507, + 523, + 882, + 571 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Intermediate Layer. (yellow) Linear layer of the feedforward block that maps the output from the Self Attention block into the 'feedforward dimension' (typically 4x the model dimension).", + "bbox": [ + 507, + 571, + 882, + 634 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Output Layer.(green) Second linear layer of the feedforward block, which maps the output from Intermediate layer back to the model dimension.", + "bbox": [ + 507, + 636, + 880, + 684 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Other Layers. (black) Other modules (activations, layer normalizations, other linear layers, etc.) not covered by the above components.", + "bbox": [ + 507, + 684, + 882, + 733 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "F Additional Profiling Analyses", + "text_level": 1, + "bbox": [ + 507, + 744, + 800, + 760 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We report layerwise profiling runs for efficient self-attention variants and inference-time throughput profiling runs for all variants in this section at Figures 5 and 6.", + "bbox": [ + 507, + 770, + 882, + 832 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1647", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/9e371d3b233e2bae162bc004d5f03449c08a5e0e27a7f339c86179214cf85375.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 163, + 128, + 835, + 318 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/9249c4691a988c422a39f4111211abfb69ad3db9893bf3ae77ed73eb23fc5cef.jpg", + "image_caption": [ + "Figure 5: Layerwise latency of different Transformer variants in inference mode." + ], + "image_footnote": [], + "bbox": [ + 174, + 328, + 489, + 519 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/bd011fcec959415932d91ba35217eeade8d4fc19c1c0bac82ca1fb888ab942e2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 328, + 821, + 520 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/ba8dc71a28326aa76ba3e2b655d2801599299bbb27d1924138fa59aa3d23b009.jpg", + "image_caption": [ + "Figure 6: Throughput Profiling Results in inference mode." + ], + "image_footnote": [], + "bbox": [ + 178, + 661, + 494, + 838 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/48fd5506f2e1274e7f1b551029d302ef8116dddbc94099479951753dd4d971d1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 505, + 661, + 821, + 838 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1648", + "bbox": [ + 482, + 928, + 521, + 940 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "A For every submission:", + "bbox": [ + 114, + 107, + 322, + 122 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "A1. Did you describe the limitations of your work?", + "bbox": [ + 129, + 126, + 532, + 142 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The Limitations section", + "bbox": [ + 149, + 143, + 327, + 156 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "A2. Did you discuss any potential risks of your work?", + "bbox": [ + 129, + 168, + 552, + 185 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The Limitations section", + "bbox": [ + 149, + 186, + 327, + 200 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "A3. Do the abstract and introduction summarize the paper's main claims?", + "bbox": [ + 129, + 211, + 695, + 227 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Section 1", + "bbox": [ + 151, + 229, + 221, + 241 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "A4. Have you used AI writing assistants when working on this paper?", + "bbox": [ + 129, + 254, + 668, + 269 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 151, + 272, + 231, + 285 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "B Did you use or create scientific artifacts?", + "bbox": [ + 114, + 297, + 487, + 313 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Section 3, 4", + "bbox": [ + 132, + 319, + 223, + 332 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "B1. Did you cite the creators of artifacts you used?", + "bbox": [ + 129, + 343, + 529, + 359 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Section 3, 4", + "bbox": [ + 151, + 361, + 242, + 374 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?", + "bbox": [ + 129, + 387, + 778, + 401 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Not explicitly, since we use publicly available Huggingface and Fairseq models that are intended for research use", + "bbox": [ + 149, + 403, + 880, + 432 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?", + "bbox": [ + 129, + 444, + 880, + 508 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We use publicly available Huggingface and Fairseq models that are intended for research use", + "bbox": [ + 149, + 510, + 840, + 525 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?", + "bbox": [ + 129, + 536, + 880, + 583 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Not applicable. Left blank.", + "bbox": [ + 149, + 585, + 349, + 600 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?", + "bbox": [ + 129, + 609, + 880, + 643 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Section 4 and 4.2, Appendices B,D", + "bbox": [ + 151, + 643, + 410, + 658 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be.", + "bbox": [ + 129, + 668, + 880, + 749 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We only use datasets to profile models over different sequence lengths, but don't use the content of the dataset itself. Thus we report the relevant statistic i.e. dataset sequence length.", + "bbox": [ + 149, + 750, + 884, + 781 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "C Did you run computational experiments?", + "bbox": [ + 114, + 790, + 492, + 808 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Section 3, 4.", + "bbox": [ + 132, + 813, + 226, + 827 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?", + "bbox": [ + 129, + 837, + 880, + 870 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Section 4.1, 4.2, Appendix C.", + "bbox": [ + 149, + 871, + 366, + 885 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "ACL 2023 Responsible NLP Checklist", + "bbox": [ + 132, + 84, + 433, + 99 + ], + "page_idx": 10 + }, + { + "type": "footer", + "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.", + "bbox": [ + 112, + 892, + 877, + 917 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1649", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?", + "bbox": [ + 129, + 83, + 878, + 115 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Section 4.2", + "bbox": [ + 149, + 117, + 236, + 130 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?", + "bbox": [ + 129, + 142, + 882, + 190 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Section 4.2", + "bbox": [ + 149, + 192, + 236, + 205 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?", + "bbox": [ + 129, + 217, + 882, + 265 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Section 3, 4", + "bbox": [ + 149, + 267, + 240, + 280 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?", + "bbox": [ + 112, + 293, + 877, + 309 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 132, + 313, + 213, + 329 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?", + "bbox": [ + 127, + 340, + 882, + 372 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 374, + 248, + 388 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?", + "bbox": [ + 127, + 399, + 882, + 447 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 449, + 248, + 463 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?", + "bbox": [ + 127, + 475, + 882, + 521 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 524, + 248, + 538 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?", + "bbox": [ + 127, + 549, + 873, + 564 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 565, + 248, + 581 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?", + "bbox": [ + 127, + 592, + 880, + 623 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 626, + 248, + 640 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "1650", + "bbox": [ + 482, + 928, + 521, + 940 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/When to Use Efficient Self Attention_ Profiling Text, Speech and Image Transformer Variants/8b2664b3-8c7e-4af7-bd6b-26db06c033c3_model.json b/2023/When to Use Efficient Self Attention_ Profiling Text, Speech and Image Transformer Variants/8b2664b3-8c7e-4af7-bd6b-26db06c033c3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..daff1b10d488a57754ba1ea7395d55cfe1dbfac9 --- /dev/null +++ b/2023/When to Use Efficient Self Attention_ Profiling Text, Speech and Image Transformer Variants/8b2664b3-8c7e-4af7-bd6b-26db06c033c3_model.json @@ -0,0 +1,2457 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.211, + 0.091, + 0.788, + 0.131 + ], + "angle": 0, + "content": "When to Use Efficient Self Attention? Profiling Text, Speech and Image Transformer Variants" + }, + { + "type": "text", + "bbox": [ + 0.317, + 0.15, + 0.685, + 0.166 + ], + "angle": 0, + "content": "Anuj Diwan, Eunsol Choi, David Harwath" + }, + { + "type": "text", + "bbox": [ + 0.364, + 0.168, + 0.639, + 0.183 + ], + "angle": 0, + "content": "Department of Computer Science" + }, + { + "type": "text", + "bbox": [ + 0.364, + 0.184, + 0.64, + 0.199 + ], + "angle": 0, + "content": "The University of Texas at Austin" + }, + { + "type": "text", + "bbox": [ + 0.3, + 0.201, + 0.705, + 0.216 + ], + "angle": 0, + "content": "{anuj.diwan, eunsol, harwath}@utexas.edu" + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.253, + 0.341, + 0.269 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.28, + 0.461, + 0.606 + ], + "angle": 0, + "content": "We present the first unified study of the efficiency of self-attention-based Transformer variants spanning text, speech and vision. We identify input length thresholds (tipping points) at which efficient Transformer variants become more efficient than vanilla models, using a variety of efficiency metrics (latency, throughput, and memory). To conduct this analysis for speech, we introduce L-HuBERT, a novel local-attention variant of a self-supervised speech model. We observe that these thresholds are (a) much higher than typical dataset sequence lengths and (b) dependent on the metric and modality, showing that choosing the right model depends on modality, task type (long-form vs. typical context) and resource constraints (time vs. memory). By visualising the breakdown of the computational costs for transformer components, we also show that non-self-attention components exhibit significant computational costs. We release our profiling toolkit at https://github.com/ajd12342/profiling-transformers." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.619, + 0.422, + 0.634 + ], + "angle": 0, + "content": "1 Introduction and Related Work" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.644, + 0.49, + 0.886 + ], + "angle": 0, + "content": "Transformers (Vaswani et al., 2017) are widely adopted across NLP (Devlin et al., 2019; Brown et al., 2020), Speech Processing (Mohamed et al., 2022) and Computer Vision (Dosovitskiy et al., 2021). Studies have shown that scaling models up improves performance (Chowdhery et al., 2022), making efficiency an important research topic. Many Transformer variants focus on improving the efficiency of self-attention, motivated by its asymptotic quadratic time/space complexity with respect to the input sequence length. While these Transformer variants are designed be asymptotically faster, in practice they may actually be slower, especially given modest input lengths that are typical of many tasks." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.254, + 0.885, + 0.608 + ], + "angle": 0, + "content": "Our paper presents two main analyses. First, we visualize the layerwise efficiency of such models to locate bottlenecks and attempt to answer the question \"is self-attention the true bottleneck?\" We find that in the non-asymptotic case, non-self-attention layers contribute significantly to the overall cost, especially for speech architectures due to the input waveform tokenizer in models like HuBERT (Hsu et al., 2021). Second, when should we use self-attention-based efficient Transformers? Comparing efficient variants with vanilla models at different input lengths, we find that this tipping point where efficient variants outperform vanilla architectures is much higher than typical input lengths of existing benchmarks across all modalities, calling into question the efficacy of using such efficient Transformers and requiring new benchmarks. We introduce a local-attention variant of a speech Transformer, HuBERT, to conduct this analysis. Together, our analyses suggest that current approaches that focus on improving self-attention might not be the most effective for improving efficiency." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.62, + 0.702, + 0.636 + ], + "angle": 0, + "content": "2 Efficiency Metrics" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.645, + 0.885, + 0.887 + ], + "angle": 0, + "content": "Model efficiency is an umbrella term for a suite of efficiency metrics, which do not always correlate with, and sometimes contradict, each other (Dehghani et al., 2022). Further, different metrics are relevant to different end use-cases. To cover most use-cases, we evaluate a set of four metrics; two for computational time and two for memory usage: Throughput: Number of examples processed per sec, given inputs of a given sequence length, using the maximum possible batch size for a given GPU. Latency-Inference: Time (in ms) to run inference for 1 unbatched input of a given sequence length. Max-Memory: The allocated GPU memory (MiB) for processing 1 input of a given sequence length. Parameter Count: Number of model parameters." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.888, + 0.884, + 0.919 + ], + "angle": 0, + "content": "We profile models in all modalities in training mode and inference mode. For training, while" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.892, + 0.49, + 0.918 + ], + "angle": 0, + "content": "1We refer the readers to Tay et al. (2022) for a comprehensive overview of efficient Transformers." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1639" + }, + { + "type": "footer", + "bbox": [ + 0.227, + 0.946, + 0.772, + 0.958 + ], + "angle": 0, + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + }, + { + "type": "footer", + "bbox": [ + 0.369, + 0.959, + 0.631, + 0.972 + ], + "angle": 0, + "content": "Volume 2: Short Papers, pages 1639-1650" + }, + { + "type": "footer", + "bbox": [ + 0.297, + 0.973, + 0.702, + 0.986 + ], + "angle": 0, + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.164, + 0.094, + 0.372, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.348, + 0.49, + 0.379 + ], + "angle": 0, + "content": "Figure 1: Transformer layer types profiled in our layerwise efficiency profiling experiments." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.39, + 0.49, + 0.454 + ], + "angle": 0, + "content": "Transformer architectures often use prediction heads with a larger output space (e.g., for text generation), we choose a lightweight binary classification head for profiling." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.463, + 0.489, + 0.591 + ], + "angle": 0, + "content": "Layerwise Efficiency Metrics We also profile some metrics and models in a layerwise fashion to locate their efficiency bottlenecks. Our goal is twofold: a) provide an empirical approach to efficient model design, as an alternative to theoretical analyses or mental models (e.g. self-attention is \\( O(n^{2}) \\)) and b) empirically answer the question \"to what degree is self-attention the bottleneck?\"" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.592, + 0.49, + 0.77 + ], + "angle": 0, + "content": "We identify important layer types (Self-Attention, Feedforward, etc.) and profile the Latency-Inference and Parameter Count metrics per-layer-type to obtain a fine-grained understanding of which layer types and indices (layer 0 vs 11) contribute the most to model efficiency costs. We use param counts as a proxy for memory (profiling real layerwise memory usage is non-trivial due to Pytorch memory allocation intricacies). We profile the layers depicted in Figure 1; more details in Appendix E." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.781, + 0.414, + 0.798 + ], + "angle": 0, + "content": "3 Local-Attention Speech Model" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.807, + 0.49, + 0.92 + ], + "angle": 0, + "content": "Efficient transformers (Xiong et al., 2021; Ma et al., 2021) have not received as much attention in Speech as they have in NLP and CV, perhaps due to two reasons. First, there is a relative lack of long-context speech benchmarks as compared to those in NLP (LRA (Tay et al., 2021) and QuALITY (Pang et al., 2022)). Second, when performing speech" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.082, + 0.878, + 0.135 + ], + "angle": 0, + "content": "
ModelWER ↓WER (w/ FT) ↓
HuBERT Base7.093.4
L-HuBERT (32 | 100)21.06 | 14.488.52 | 7.39
" + }, + { + "type": "table_caption", + "bbox": [ + 0.548, + 0.145, + 0.843, + 0.159 + ], + "angle": 0, + "content": "Table 1: WERs on the SUPERB ASR task." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.167, + 0.885, + 0.378 + ], + "angle": 0, + "content": "tasks like automatic speech recognition (ASR), it is typical to segment a long speech signal into small individual utterances and perform ASR independently on each. For example, most Librispeech examples are less than 5 seconds. Many popular speech models like HuBERT (Hsu et al., 2021) tokenize the waveform at 50 tokens per second, implying that a typical utterance has only several hundred tokens; far below the number of tokens in long-context NLP tasks. Nevertheless, textless speech models (Lakhotia et al., 2021) are becoming more feasible, motivating the modelling of long speech utterances." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.386, + 0.885, + 0.693 + ], + "angle": 0, + "content": "Local HuBERT Model To investigate the efficiency of the self-attention layer in speech models, we introduce the Local HuBERT model which replaces HuBERT's self-attention with the Longformer (Beltagy et al., 2020) sliding-window self-attention. In this attention mechanism, every token attends to tokens within a local window context, rather than the full token sequence. Our model is similar to the temporally windowed-attention Transformer acoustic model proposed by Alastruey et al. (2021) for speech translation; our approach differs by using the self-supervised HuBERT model as our basis, and we evaluate on ASR. Choosing the appropriate window size for the local attention context is key; we explore 32 and 100 token contexts, corresponding to \\(640~\\mathrm{ms}\\) and \\(2\\mathrm{s}\\), inspired by phone recognition models that typically incorporate similar context sizes (Peddinti et al., 2015; feng Yeh et al., 2019)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.702, + 0.887, + 0.897 + ], + "angle": 0, + "content": "ASR Results We initialize the L-HuBERT model with pretrained HuBERT Base weights (pretrained with full self-attention), and then replace self-attention with sliding-window self-attention; due to limited compute, we did not pretrain L-HuBERT from scratch using sliding-window attention. We then evaluate L-HuBERT on Librispeech (Panayotov et al., 2015) ASR via the SUPERB (Yang et al., 2021) benchmark under two settings; a) Freeze: freezing the model and only training projection weights and b) Finetune: fully finetune the model. We use the default S3PRL2 hyperparams; but we" + }, + { + "type": "page_footnote", + "bbox": [ + 0.53, + 0.904, + 0.771, + 0.919 + ], + "angle": 0, + "content": "2https://github.com/s3pr1/s3pr1" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.928, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1640" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.136, + 0.082, + 0.468, + 0.147 + ], + "angle": 0, + "content": "
ModelEmbPosSAIntermOutputOthers
BERT23.8M-29M28.3M28.3M0.6M
HuBERT4.2M5.1M29M28.3M28.3M0.2M
ViT0.6M-29M28.3M28.3M0.6M
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.156, + 0.489, + 0.199 + ], + "angle": 0, + "content": "Table 2: Layer-wise parameter counts. Emb: Input Embedding, Pos: Positional Emb. SA: Self-Attention, Interm: Intermediate." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.214, + 0.489, + 0.295 + ], + "angle": 0, + "content": "train for 200k steps for Freeze and 104k steps for Finetune. Both models converge by 104k steps; we train Freeze for longer to eke out as much performance as possible, while we stop training Finetune due to limited compute." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.296, + 0.49, + 0.537 + ], + "angle": 0, + "content": "We report Word Error Rate (WER) on Librispeech test-clean in Table 1; lower is better. In the frozen setting (middle column), we see a large WER increase over HuBERT; we hypothesize that this is due to the attention layer mismatch since we initialize L-HuBERT with HuBERT weights that were pretrained with full self attention, rather than pretraining L-HuBERT from scratch. However, in the finetuning setting, the gap between HuBERT Base and L-HuBERT narrows considerably and using a larger window size achieves better performance. As our L-HuBERT model is a reasonable architecture capable of moderate ASR performance, we can continue to study its computational efficiency (we profile the window-100 variant)." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.551, + 0.406, + 0.568 + ], + "angle": 0, + "content": "4 Methods and Implementation" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.579, + 0.49, + 0.756 + ], + "angle": 0, + "content": "We analyze the Base versions of the BERT (Devlin et al., 2019), Longformer (Beltagy et al., 2020) and Nystromformer (Xiong et al., 2021) models for text; the HuBERT (Hsu et al., 2021) and L-HuBERT (Section 3) models for speech; and Vision Transformer (Dosovitskiy et al., 2021) and Swin Transformer (Liu et al., 2021) models for vision; BERT, HuBERT and ViT are standard Transformer encoder architectures. Longformer, L-HuBERT and Swin use fixed-pattern self-attention while Nystromformer uses approximate self-attention." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.769, + 0.357, + 0.785 + ], + "angle": 0, + "content": "4.1 Sequence Length Ranges" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.791, + 0.49, + 0.919 + ], + "angle": 0, + "content": "We profile our models on a wide range of input sequence lengths to cover both avg. sequence lengths of commonly used contemporary datasets (Table 3) and typical sequence lengths of long-context tasks. Details about how we compute sequence lengths in Table 3 can be found in Appendix B. Most image datasets use images resized to 224 or 512 pixels. Below, \\( \\text{range}(a, b, c) \\) means a range from \\( a \\) to \\( b \\)" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.883, + 0.149 + ], + "angle": 0, + "content": "in steps of \\( c \\). Since there is no difference between synthetic and real inputs from a computational complexity standpoint, we use synthetic inputs to more easily control for their sequence lengths." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.15, + 0.885, + 0.198 + ], + "angle": 0, + "content": "Text Modality The input is 'This is a sentence.' repeated \\( n \\) times, \\( n \\in \\text{range}(10, 560, 10) \\) i.e. range(62, 3362, 60) tokens for all tokenizers." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.2, + 0.884, + 0.263 + ], + "angle": 0, + "content": "Speech Modality The inputs have durations in range(1, 50, 0.5) sec i.e. range(50, 2500, 25) tokens for all featurizers (CNNs with 20 ms framerate). Our sampling strategy is in Appendix A." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.265, + 0.884, + 0.328 + ], + "angle": 0, + "content": "Image Modality We use square inputs of dimension in range(32, 1024, 32) pixels by rescaling a fixed image. The # tokens depend on featurizer patch size, which is different for different models." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.343, + 0.755, + 0.358 + ], + "angle": 0, + "content": "4.2 Implementational Details" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.365, + 0.884, + 0.477 + ], + "angle": 0, + "content": "We profile time-based metrics (latency/throughput) using Pytorch CUDA Events3 by executing 20 iterations sequentially. The first few iterations serve as GPU warm-start; thus, we report the average of the last 10. We record Max-Memory with torch.cuda.max_memory_allocated() and param counts with torchinfo (Yep, 2020)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.479, + 0.884, + 0.672 + ], + "angle": 0, + "content": "To profile throughput, we approximate the max batch size that fits on a single GPU using a linear estimator; more details in Appendix C. Finally, we profile the layerwise Latency-Inference metric using torchprof (Wong, 2020). We attach profiling hooks to modules of interest (e.g. Self-Attention, Embedding), giving us execution times of their forward() functions (other modules/functions are not profiled). We use the Huggingface (Wolf et al., 2020) implementations of text and image models and fairseq (Ott et al., 2019) implementations for speech models; more details in Appendix D." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.686, + 0.688, + 0.702 + ], + "angle": 0, + "content": "5 Profiling Results" + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.714, + 0.771, + 0.73 + ], + "angle": 0, + "content": "5.1 Layerwise Profiling Results" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.736, + 0.884, + 0.816 + ], + "angle": 0, + "content": "Figure 2 shows the layerwise Latency-Inference for all 3 vanilla architectures in each modality. Figures for efficient models are in Appendix F. Color darkness represents the layer index (layer 0 is darkest). Table 2 shows the layerwise param count." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.818, + 0.883, + 0.881 + ], + "angle": 0, + "content": "Asymptotically, self-attention dominates the computation. However, since the average seq length for most text and speech tasks is less than 1000 tokens and most image datasets are used at" + }, + { + "type": "page_footnote", + "bbox": [ + 0.51, + 0.892, + 0.858, + 0.918 + ], + "angle": 0, + "content": "3https://pytorch.org/docs/stable/generated/torch.cuda.Event.html" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.519, + 0.941 + ], + "angle": 0, + "content": "1641" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.13, + 0.082, + 0.87, + 0.134 + ], + "angle": 0, + "content": "
TextSpeech
Dataset # of tokensSST 23MNLI 36SQ 177ON 506CNN 863HPQA 1,316TQA 6,589TEDL 301LJS 328VoxC 390Libri 615S-SQuAD 3080Spotify 101400
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.139, + 0.885, + 0.184 + ], + "angle": 0, + "content": "Table 3: Average token sequence lengths. Left to right: Stanford Sentiment Treebank, MultiNLI, SQuAD2.0, OntoNotes, CNN-DailyMail, HotpotQA, TriviaQA, TEDLIUM, LJSpeech, VoxCeleb Speaker Recognition, Librispeech, Spoken SQuAD, Spotify Podcasts." + }, + { + "type": "image", + "bbox": [ + 0.16, + 0.194, + 0.87, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.183, + 0.358, + 0.813, + 0.373 + ], + "angle": 0, + "content": "Figure 2: Layerwise latency of different vanilla Transformer architectures in inference mode." + }, + { + "type": "image", + "bbox": [ + 0.191, + 0.378, + 0.807, + 0.658 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.67, + 0.879, + 0.687 + ], + "angle": 0, + "content": "Figure 3: Overall Inference-time Profiling Results. Text and speech models in first row, vision models in second." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.696, + 0.49, + 0.809 + ], + "angle": 0, + "content": "a max dimension of 512, at these points, non-self-attention components take up \\(35\\%\\), \\(58.8\\%\\) and \\(43.75\\%\\) latency for NLP, speech and images. Additionally, parameter counts of SA are also comparable to Interm/Output layers. This shows that it is also important to direct efficiency efforts for other model components." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.823, + 0.49, + 0.92 + ], + "angle": 0, + "content": "While the latency associated with embedding layers is minimal for BERT, they are sizable for HuBERT. HuBERT uses a CNN feature extractor with different strides and kernel sizes and consumes more time in the earlier CNN layers as opposed to later ones, as is visible in Figure 2, which shows" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.696, + 0.885, + 0.89 + ], + "angle": 0, + "content": "darker shades i.e. earlier layers dominating the computation. Optimal efficiency strategies can thus differ across modalities, e.g. Wu et al. (2022) slims down this CNN feature extractor embedding layer. On the other hand, embedding layers take up a lot of parameters in BERT; thus, it may be helpful to shrink the BERT embedding layer for memory purposes (as opposed to latency for HuBERT). Finally, analyzing Transformer variants (Appendix F), we see that self-attention in Longformer, Swin and L-HuBERT encouragingly scales latency linearly, but with large overhead for smaller inputs." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1642" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.192, + 0.081, + 0.805, + 0.096 + ], + "angle": 0, + "content": "■ BERT Nyströmformer Longformer HuBERT L-HuBERT ViT Swin" + }, + { + "type": "image_caption", + "bbox": [ + 0.304, + 0.101, + 0.436, + 0.113 + ], + "angle": 0, + "content": "Training-Throughput" + }, + { + "type": "image", + "bbox": [ + 0.252, + 0.113, + 0.487, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.101, + 0.7, + 0.114 + ], + "angle": 0, + "content": "Training-Max-Memory" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.116, + 0.745, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.236, + 0.486, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.238, + 0.746, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.12, + 0.369, + 0.875, + 0.384 + ], + "angle": 0, + "content": "Figure 4: Overall Training-time Profiling Results. Text and speech models in first row, vision models in second." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.395, + 0.356, + 0.411 + ], + "angle": 0, + "content": "5.2 Overall Profiling Results" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.416, + 0.49, + 0.465 + ], + "angle": 0, + "content": "Our profiling results are in Figures 3 and 4. Inference Throughput is in the Appendix at Figure 6, exhibiting similar trends as training Throughput." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.475, + 0.489, + 0.748 + ], + "angle": 0, + "content": "Tipping Point Analysis We see that most variants are slower and more memory hungry than vanilla models for input lengths of typical-context tasks. We define the tipping point for each modality: the input length at which the variant becomes more efficient than the vanilla model. For text and speech, it is \\(1750 - 2000\\) tokens for inference latency and max-memory, greater than typical input lengths (Table 3). However, while the tipping point for training max-memory is \\(\\approx 1500\\) tokens for text (still a large number), it is \\(\\approx 0 - 250\\) for speech, an encouraging result. For images, it is \\(500 - 700\\) pixels for all metrics apart from throughput. This is less reasonable for 224 pixel datasets but good for high resolution image datasets (512/1024). All variants are either worse or comparable than vanilla models across modalities for throughput." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.749, + 0.49, + 0.83 + ], + "angle": 0, + "content": "We hypothesize that some efficient models suffer from additional overheads; while vanilla attention benefits from highly optimized matrix multiplication, windowed attention requires complex reshaping and preprocessing." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.84, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Choosing the Right Model Depends on Resource Constraints Our results show that the choice of the right model depends on resource constraints. Suppose that one is training models under a time constraint; then, throughput is the bottleneck and" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.395, + 0.883, + 0.443 + ], + "angle": 0, + "content": "efficient models would not be a good fit. On the other hand, efficient models are useful for long context memory-constrained inference." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.458, + 0.885, + 0.554 + ], + "angle": 0, + "content": "Local Attention and Excessive Padding The Longformer pads input lengths to be a multiple of 512 and Swin requires input dimension to be a multiple of 224. This slows shorter inputs down and results in extremely low performance (measured by all 3 metrics) as compared to vanilla models." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.568, + 0.884, + 0.68 + ], + "angle": 0, + "content": "Comparing Parameter Counts The Longformer uses more parameters compared to vanilla BERT (148M vs. 109M) because it uses two sets of Q,K,V projection matrices for its global and local attention operations; sharing these may decrease its memory usage. For other modalities, efficient models do not incur more parameters." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.698, + 0.642, + 0.713 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.727, + 0.885, + 0.919 + ], + "angle": 0, + "content": "We present an empirical efficiency analysis of vanilla Transformers and their self-attention-based efficient variants across modalities, metrics and input context sizes. We find substantial differences across modalities and metrics when analyzing the tipping point for efficient variants. Finally, the layerwise analysis finds that self-attention is not the only bottleneck. We recommend that all efficient model papers should report such cross-modal, layerwise profiling results on multiple efficiency metrics covering a variety of use-cases to provide a full picture of the benefits of the model." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1643" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.115, + 0.085, + 0.221, + 0.1 + ], + "angle": 0, + "content": "Limitations" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.111, + 0.49, + 0.189 + ], + "angle": 0, + "content": "We focus primarily on comparing model efficiencies using a variety of efficiency metrics and do not consider model performance; one can perform a more elaborate analysis of performance-efficiency tradeoffs, which we did not do here." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.192, + 0.49, + 0.351 + ], + "angle": 0, + "content": "We only profile a total of seven models across three modalities while there are more efficient variants and vanilla Transformers proposed in the literature. While we choose our models to be as representative of each modality and efficiency technique as possible, we cannot extrapolate results to other model variants and other modalities. In particular, modalities like video and genomics and efficiency approaches like quantization would be interesting to profile, which we did not do." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.365, + 0.287, + 0.381 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.391, + 0.49, + 0.456 + ], + "angle": 0, + "content": "We thank the reviewers and the meta-reviewer of the ACL community for helpful feedback on the draft. This work was partially funded by a grant from UT Machine Learning Lab." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.482, + 0.214, + 0.497 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.505, + 0.489, + 0.545 + ], + "angle": 0, + "content": "Belen Alastruey, Gerard I. Gållego, and Marta R. Costajussa. 2021. Efficient Transformer for Direct Speech Translation." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.556, + 0.49, + 0.597 + ], + "angle": 0, + "content": "Iz Beltagy, Matthew E. Peters, and Arman Cohan. 2020. Longformer: The Long-Document Transformer. ArXiv preprint, abs/2004.05150." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.607, + 0.49, + 0.803 + ], + "angle": 0, + "content": "Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language Models are Few-Shot Learners. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.814, + 0.49, + 0.92 + ], + "angle": 0, + "content": "Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob" + }, + { + "type": "list", + "bbox": [ + 0.116, + 0.505, + 0.49, + 0.92 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.527, + 0.086, + 0.885, + 0.283 + ], + "angle": 0, + "content": "Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan Firat, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. 2022. PaLM: Scaling Language Modeling with Pathways." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.294, + 0.885, + 0.412 + ], + "angle": 0, + "content": "Ann Clifton, Sravana Reddy, Yongze Yu, Aasish Pappu, Rezvaneh Rezapour, Hamed Bonab, Maria Eskevich, Gareth Jones, Jussi Karlgren, Ben Carterette, and Rosie Jones. 2020. 100,000 Podcasts: A Spoken English Document Corpus. In Proceedings of the 28th International Conference on Computational Linguistics, pages 5903-5917, Barcelona, Spain (Online). International Committee on Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.424, + 0.884, + 0.491 + ], + "angle": 0, + "content": "Mostafa Dehghani, Yi Tay, Anurag Arnab, Lucas Beyer, and Ashish Vaswani. 2022. The Efficiency Mismonomer. In *The Tenth International Conference on Learning Representations*, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.502, + 0.884, + 0.621 + ], + "angle": 0, + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.632, + 0.883, + 0.75 + ], + "angle": 0, + "content": "Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. 2021. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.762, + 0.884, + 0.828 + ], + "angle": 0, + "content": "Ching feng Yeh, Jay Mahadeokar, Kaustubh Kalgaonkar, Yongqiang Wang, Duc Le, Mahaveer Jain, Kjell Schubert, Christian Fuegen, and Michael L. Seltzer. 2019. Transformer-transducer: End-to-end speech recognition with self-attention. ArXiv, abs/1910.12977." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.84, + 0.884, + 0.919 + ], + "angle": 0, + "content": "Karl Moritz Hermann, Tomás Kocisky, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa Suleyman, and Phil Blunsom. 2015. Teaching Machines to Read and Comprehend. In Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December" + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.885, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1644" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.135, + 0.086, + 0.49, + 0.113 + ], + "angle": 0, + "content": "7-12, 2015, Montreal, Quebec, Canada, pages 1693-1701." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.124, + 0.488, + 0.203 + ], + "angle": 0, + "content": "François Hernandez, Vincent Nguyen, Sahar Ghannay, Natalia Tomashenko, and Yannick Esteve. Ted-lium 3: Twice as much data and corpus repartition for experiments on speaker adaptation. In Speech and Computer, pages 198-208. Springer International Publishing." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.213, + 0.489, + 0.304 + ], + "angle": 0, + "content": "Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. 2021. HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 29:3451-3460." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.315, + 0.488, + 0.355 + ], + "angle": 0, + "content": "Keith Ito and Linda Johnson. 2017. The LJ Speech Dataset. https://keithito.com/LJ-Speech-Dataset/." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.366, + 0.489, + 0.458 + ], + "angle": 0, + "content": "Mandar Joshi, Eunsol Choi, Daniel Weld, and Luke Zettlemoyer. 2017. TriviaQA: A large scale distantly supervised challenge dataset for reading comprehension. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1601-1611, Vancouver, Canada. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.469, + 0.489, + 0.56 + ], + "angle": 0, + "content": "Kushal Lakhotia, Eugene Kharitonov, Wei-Ning Hsu, Yossi Adi, Adam Polyak, Benjamin Bolte, Tu-Anh Nguyen, Jade Copet, Alexei Baevski, Abdelrahman Mohamed, and Emmanuel Dupoux. 2021. On generative spoken language modeling from raw audio. Transactions of the Association for Computational Linguistics, 9:1336-1354." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.571, + 0.489, + 0.662 + ], + "angle": 0, + "content": "Chia-Hsuan Li, Szu-Lin Wu, Chi-Liang Liu, and Hung-yi Lee. 2018. Spoken SQuAD: A Study of Mitigating the Impact of Speech Recognition Errors on Listening Comprehension. In *Interspeech* 2018, 19th Annual Conference of the International Speech Communication Association, Hyderabad, India, 2-6 September 2018, pages 3459-3463. ISCA." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.673, + 0.489, + 0.764 + ], + "angle": 0, + "content": "Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. 2021. Swin Transformer: Hierarchical Vision Transformer using Shfted Windows. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 9992-10002. IEEE." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.776, + 0.489, + 0.815 + ], + "angle": 0, + "content": "Xuezhe Ma, Xiang Kong, Sinong Wang, Chunting Zhou, Jonathan May, Hao Ma, and Luke Zettlemoyer. 2021. Luna: Linear Unified Nested Attention. In NeurIPS." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.826, + 0.489, + 0.918 + ], + "angle": 0, + "content": "Abdelrahman Mohamed, Hung yi Lee, Lasse Borgholt, Jakob D. Havtorn, Joakim Edin, Christian Igel, Katrin Kirchhoff, Shang-Wen Li, Karen Livescu, Lars Maaloe, Tara N. Sainath, and Shinji Watanabe. 2022. Self-Supervised Speech Representation Learning: A Review. IEEE Journal of Selected Topics in Signal Processing, 16(6):1179-1210." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.513, + 0.086, + 0.883, + 0.165 + ], + "angle": 0, + "content": "Arsha Nagrani, Joon Son Chung, and Andrew Zisserman. 2017. VoxCeleb: A Large-Scale Speaker Identification Dataset. In *Interspeech* 2017, 18th Annual Conference of the International Speech Communication Association, Stockholm, Sweden, August 20-24, 2017, pages 2616-2620. ISCA." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.177, + 0.883, + 0.282 + ], + "angle": 0, + "content": "Myle Ott, Sergey Edunov, Alexei Baevski, Angela Fan, Sam Gross, Nathan Ng, David Grangier, and Michael Auli. 2019. *fairoseq: A fast, extensible toolkit for sequence modeling.* In *Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics* (Demonstrations), pages 48-53, Minneapolis, Minnesota. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.294, + 0.883, + 0.386 + ], + "angle": 0, + "content": "Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. 2015. Librispeech: An ASR corpus based on public domain audio books. In 2015 IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP 2015, South Brisbane, Queensland, Australia, April 19-24, 2015, pages 5206-5210. IEEE." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.398, + 0.883, + 0.529 + ], + "angle": 0, + "content": "Richard Yuanzhe Pang, Alicia Parrish, Nitish Joshi, Nikita Nangia, Jason Phang, Angelica Chen, Vishakh Padmakumar, Johnny Ma, Jana Thompson, He He, and Samuel Bowman. 2022. Quality: Question answering with long input texts, yes! In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5336-5358, Seattle, United States. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.541, + 0.883, + 0.594 + ], + "angle": 0, + "content": "Vijayaditya Peddinti, Daniel Povey, and Sanjeev Khudanpur. 2015. A time delay neural network architecture for efficient modeling of long temporal contexts. In Proc. Interspeech 2015, pages 3214-3218." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.606, + 0.883, + 0.698 + ], + "angle": 0, + "content": "Sameer S. Pradhan and Nianwen Xue. 2009. OntoNotes: The \\(90\\%\\) solution. In Proceedings of Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the Association for Computational Linguistics, Companion Volume: Tutorial Abstracts, pages 11-12, Boulder, Colorado. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.71, + 0.883, + 0.801 + ], + "angle": 0, + "content": "Pranav Rajpurkar, Robin Jia, and Percy Liang. 2018. Know what you don't know: Unanswerable questions for SQuAD. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 784-789, Melbourne, Australia. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.814, + 0.883, + 0.918 + ], + "angle": 0, + "content": "Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1631-1642, Seattle, Washington, USA. Association for Computational Linguistics." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.883, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1645" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.179 + ], + "angle": 0, + "content": "Yi Tay, Mostafa Dehghani, Samira Abnar, Yikang Shen, Dara Bahri, Philip Pham, Jinfeng Rao, Liu Yang, Sebastian Ruder, and Donald Metzler. 2021. Long Range Arena: A Benchmark for Efficient Transformers. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.188, + 0.49, + 0.243 + ], + "angle": 0, + "content": "Yi Tay, Mostafa Dehghani, Dara Bahri, and Donald Metzler. 2022. Efficient Transformers: A Survey. In ACM Comput. Surv., volume 55, New York, NY, USA. Association for Computing Machinery." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.251, + 0.49, + 0.344 + ], + "angle": 0, + "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is All You Need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 5998-6008." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.353, + 0.49, + 0.471 + ], + "angle": 0, + "content": "Adina Williams, Nikita Nangia, and Samuel Bowman. 2018. A broad-coverage challenge corpus for sentence understanding through inference. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1112-1122, New Orleans, Louisiana. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.481, + 0.49, + 0.639 + ], + "angle": 0, + "content": "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.648, + 0.489, + 0.676 + ], + "angle": 0, + "content": "Alexander William Wong. 2020. torchprof. https://github.com/awwong1/torchprof." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.685, + 0.49, + 0.778 + ], + "angle": 0, + "content": "Felix Wu, Kwangyoun Kim, Jing Pan, Kyu J. Han, Kilian Q. Weinberger, and Yoav Artzi. 2022. Performance-Efficiency Trade-Offs in Unsupervised Pre-Training for Speech Recognition. In ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7667-7671." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.787, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, and Vikas Singh. 2021. Nystromformer: A Nystrom-based Algorithm for Approximating Self-Attention. In Thirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2021, Thirty-Third Conference on Innovative Applications of Artificial Intelligence, IAAI 2021, The Eleventh Symposium on Educational Advances in Artificial Intelligence, EAAI 2021, Virtual Event, February 2-9, 2021, pages 14138-14148. AAAI Press." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.086, + 0.885, + 0.243 + ], + "angle": 0, + "content": "Shu-Wen Yang, Po-Han Chi, Yung-Sung Chuang, Cheng-I Jeff Lai, Kushal Lakhotia, Yist Y. Lin, Andy T. Liu, Jiatong Shi, Xuankai Chang, Guanting Lin, Tzu-Hsien Huang, Wei-Cheng Tseng, Kotik Lee, Da-Rong Liu, Zili Huang, Shuyan Dong, Shang-Wen Li, Shinji Watanabe, Abdelrahman Mohamed, and Hung-yi Lee. 2021. SUPERB: Speech Processing Universal PERformance Benchmark. In Interspeech 2021, 22nd Annual Conference of the International Speech Communication Association, Brno, Czechia, 30 August - 3 September 2021, pages 1194-1198. ISCA." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.252, + 0.885, + 0.358 + ], + "angle": 0, + "content": "Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. 2018. HotpotQA: A dataset for diverse, explainable multi-hop question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2369-2380, Brussels, Belgium. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.367, + 0.884, + 0.394 + ], + "angle": 0, + "content": "Tyler Yep. 2020. torchinfo. https://github.com/ TylerYep/torchinfo." + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.086, + 0.885, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.406, + 0.825, + 0.44 + ], + "angle": 0, + "content": "A Sampling Speech Utterances for Profiling" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.448, + 0.884, + 0.722 + ], + "angle": 0, + "content": "To obtain speech inputs of length \\( i \\) seconds to \\( i + 0.5 \\) seconds for all \\( i \\) less than 12 seconds, we sample 5 speech utterances from the training set of the Librispeech dataset (Panayotov et al., 2015) whose lengths fall within this range and compute aggregate metrics over these 5 utterances. Since the Librispeech dataset does not contain extremely long speech utterances, for \\( i \\) of length greater than 12 seconds, we adopt a different approach to generate inputs. To generate such an input utterance of length between \\( i \\) and \\( i + 0.5 \\) seconds, we first sample 5 speech utterances from the Librispeech training set of input length ranging from \\( \\frac{i}{5} \\) to \\( \\frac{i + 0.5}{5} \\) and concatenate them to obtain utterances of length ranging from \\( i \\) to \\( i + 0.5 \\) as desired. We do this 5 times to get 5 different utterances and compute aggregate metrics over these 5 utterances." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.733, + 0.855, + 0.766 + ], + "angle": 0, + "content": "B Computing Token Lengths for NLP and Speech Datasets" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.775, + 0.885, + 0.92 + ], + "angle": 0, + "content": "We compute average sequence token lengths for 7 NLP datasets and 6 speech datasets. For all speech datasets, we compute mean utterance durations and multiply durations by 50 to obtain number of tokens (model framereates are \\(20\\mathrm{ms}\\) i.e. \\(\\times 50\\) ). For TEDLIUM (Hernandez et al.), LJSpeech (Ito and Johnson, 2017), VoxCeleb Speaker Recognition Dataset (Nagrani et al., 2017) and Librispeech (Panayotov et al., 2015), we compute" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1646" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.489, + 0.18 + ], + "angle": 0, + "content": "mean validation-set utterance durations; for Spoken SQuAD (Li et al., 2018), we report mean validation-set paragraph duration and for the Spotify English Podcasts dataset (Clifton et al., 2020), we report mean podcast duration directly obtained from Clifton et al. (2020)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.182, + 0.49, + 0.213 + ], + "angle": 0, + "content": "SST (Socher et al., 2013). We use test-set sentences. We use the HuggingFace BERTTokenizer." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.215, + 0.489, + 0.277 + ], + "angle": 0, + "content": "MNLI (Williams et al., 2018). We use validation-matched-set examples by concatenating the premise and the hypothesis. We use the HuggingFace BERTTokenizer." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.279, + 0.489, + 0.342 + ], + "angle": 0, + "content": "SQuAD2.0 (Rajpurkar et al., 2018). We use validation-set examples by concatenating the context and the question. We use the HuggingFace BERTTokenizer." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.344, + 0.489, + 0.391 + ], + "angle": 0, + "content": "OntoNotes (Pradhan and Xue, 2009). We obtain this number from the Longformer (Beltagy et al., 2020) paper." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.392, + 0.489, + 0.439 + ], + "angle": 0, + "content": "CNN-Dailymail (Hermann et al., 2015). We use the 3.0.0 version of the dataset and use test-set articles. We use the HuggingFace BERTTokenizer." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.441, + 0.488, + 0.489 + ], + "angle": 0, + "content": "HotpotQA (Yang et al., 2018). We obtain this number from the Longformer (Beltagy et al., 2020) paper." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.489, + 0.489, + 0.537 + ], + "angle": 0, + "content": "TriviaQA (Joshi et al., 2017). We obtain this number from the Longformer (Beltagy et al., 2020) paper." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.549, + 0.466, + 0.566 + ], + "angle": 0, + "content": "C Implementing Throughput Profiling" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.576, + 0.49, + 0.8 + ], + "angle": 0, + "content": "To profile Throughput, we need to compute the max batch size that can fit on a single GPU. We approximately predict this using a linear estimator as follows. We first record the memory \\( B \\) reserved on the GPU after just loading the model. Next, we independently run batches of sizes 1 and 2 and record memory usages \\( M_{1} \\) and \\( M_{2} \\). We use an NVIDIA Quadro RTX 8000 GPU with a maximum memory of 45000 MiB. Thus, assuming a linear relationship between batch size and memory consumption, we predict a maximum batch size of \\( bsz = \\frac{45000 - B}{M_2 - M_1} \\). In practice, this is an overestimate; we keep decreasing the batch size by a factor of 0.9 until no OOM errors occur and this is our final estimate." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.813, + 0.471, + 0.83 + ], + "angle": 0, + "content": "D Implementational Details for Models" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.84, + 0.49, + 0.918 + ], + "angle": 0, + "content": "We use the following HuggingFace configurations: bert-base-uncased for BERT, allenai/longformer-base-4096 for Longformer, uw-madison/nystromformer-4096 for Nyströmformer," + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.883, + 0.357 + ], + "angle": 0, + "content": "google/vit-base-patch16-224 for ViT and microsoft/swin-base-patch4-window7-224 for Swin. The BERT model natively supports a maximum of 512 tokens as input because it has 512 positional embeddings; we modify the positional embedding computation to allow an arbitrarily long input to be provided. The Longformer internally pads all input lengths to a multiple of 512. For Swin, we pad images to have an input dimension that is a multiple of 224; this is necessary due to the windowed attention mechanism in Swin. In fact, the Swin model natively supports only a \\(224 \\times 224\\) resolution; we make a small modification in order to support resolutions that are multiples of 224. We use the HuBERT Base model for both HuBERT and L-HuBERT." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.371, + 0.771, + 0.387 + ], + "angle": 0, + "content": "E Transformer Layer Types" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.395, + 0.883, + 0.459 + ], + "angle": 0, + "content": "Input Embedding Layer. (red) Maps the input sequence into fixed-dimensional embeddings. This is a linear layer for text and a CNN featurizer for image/speech." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.46, + 0.884, + 0.524 + ], + "angle": 0, + "content": "Positional Embedding Layer. (fuchsia) For text and image models this is part of the input embedding layer. For speech models, this is a very wide convolution layer." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.524, + 0.883, + 0.572 + ], + "angle": 0, + "content": "Self Attention Layer.(■/blue) The multi-head self attention block, which computes self-attention outputs and maps the result to the model dimension." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.573, + 0.883, + 0.636 + ], + "angle": 0, + "content": "Intermediate Layer. (yellow) Linear layer of the feedforward block that maps the output from the Self Attention block into the 'feedforward dimension' (typically 4x the model dimension)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.637, + 0.882, + 0.685 + ], + "angle": 0, + "content": "Output Layer.(green) Second linear layer of the feedforward block, which maps the output from Intermediate layer back to the model dimension." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.686, + 0.883, + 0.734 + ], + "angle": 0, + "content": "Other Layers. (black) Other modules (activations, layer normalizations, other linear layers, etc.) not covered by the above components." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.745, + 0.8, + 0.762 + ], + "angle": 0, + "content": "F Additional Profiling Analyses" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.771, + 0.884, + 0.833 + ], + "angle": 0, + "content": "We report layerwise profiling runs for efficient self-attention variants and inference-time throughput profiling runs for all variants in this section at Figures 5 and 6." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1647" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.164, + 0.129, + 0.836, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.175, + 0.329, + 0.49, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.329, + 0.822, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.544, + 0.772, + 0.56 + ], + "angle": 0, + "content": "Figure 5: Layerwise latency of different Transformer variants in inference mode." + }, + { + "type": "image", + "bbox": [ + 0.179, + 0.662, + 0.495, + 0.839 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.662, + 0.822, + 0.839 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.295, + 0.852, + 0.7, + 0.867 + ], + "angle": 0, + "content": "Figure 6: Throughput Profiling Results in inference mode." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1648" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.133, + 0.085, + 0.435, + 0.1 + ], + "angle": 0, + "content": "ACL 2023 Responsible NLP Checklist" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.108, + 0.323, + 0.123 + ], + "angle": 0, + "content": "A For every submission:" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.127, + 0.533, + 0.143 + ], + "angle": 0, + "content": "A1. Did you describe the limitations of your work?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.145, + 0.328, + 0.158 + ], + "angle": 0, + "content": "The Limitations section" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.169, + 0.553, + 0.186 + ], + "angle": 0, + "content": "A2. Did you discuss any potential risks of your work?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.187, + 0.328, + 0.201 + ], + "angle": 0, + "content": "The Limitations section" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.212, + 0.696, + 0.228 + ], + "angle": 0, + "content": "A3. Do the abstract and introduction summarize the paper's main claims?" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.23, + 0.223, + 0.242 + ], + "angle": 0, + "content": "Section 1" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.255, + 0.669, + 0.271 + ], + "angle": 0, + "content": "A4. Have you used AI writing assistants when working on this paper?" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.273, + 0.233, + 0.286 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.298, + 0.489, + 0.314 + ], + "angle": 0, + "content": "B Did you use or create scientific artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.32, + 0.225, + 0.333 + ], + "angle": 0, + "content": "Section 3, 4" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.344, + 0.53, + 0.36 + ], + "angle": 0, + "content": "B1. Did you cite the creators of artifacts you used?" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.362, + 0.243, + 0.375 + ], + "angle": 0, + "content": "Section 3, 4" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.388, + 0.779, + 0.403 + ], + "angle": 0, + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.404, + 0.882, + 0.434 + ], + "angle": 0, + "content": "Not explicitly, since we use publicly available Huggingface and Fairseq models that are intended for research use" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.445, + 0.882, + 0.51 + ], + "angle": 0, + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.511, + 0.841, + 0.526 + ], + "angle": 0, + "content": "We use publicly available Huggingface and Fairseq models that are intended for research use" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.537, + 0.882, + 0.585 + ], + "angle": 0, + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.586, + 0.351, + 0.601 + ], + "angle": 0, + "content": "Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.611, + 0.882, + 0.644 + ], + "angle": 0, + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.644, + 0.411, + 0.659 + ], + "angle": 0, + "content": "Section 4 and 4.2, Appendices B,D" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.669, + 0.882, + 0.75 + ], + "angle": 0, + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be." + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.751, + 0.885, + 0.782 + ], + "angle": 0, + "content": "We only use datasets to profile models over different sequence lengths, but don't use the content of the dataset itself. Thus we report the relevant statistic i.e. dataset sequence length." + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.791, + 0.494, + 0.809 + ], + "angle": 0, + "content": "C Did you run computational experiments?" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.814, + 0.227, + 0.828 + ], + "angle": 0, + "content": "Section 3, 4." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.838, + 0.882, + 0.871 + ], + "angle": 0, + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.872, + 0.368, + 0.887 + ], + "angle": 0, + "content": "Section 4.1, 4.2, Appendix C." + }, + { + "type": "footer", + "bbox": [ + 0.114, + 0.893, + 0.878, + 0.918 + ], + "angle": 0, + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1649" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.13, + 0.084, + 0.88, + 0.116 + ], + "angle": 0, + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.118, + 0.237, + 0.131 + ], + "angle": 0, + "content": "Section 4.2" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.143, + 0.884, + 0.191 + ], + "angle": 0, + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.193, + 0.237, + 0.206 + ], + "angle": 0, + "content": "Section 4.2" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.218, + 0.884, + 0.266 + ], + "angle": 0, + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.268, + 0.242, + 0.281 + ], + "angle": 0, + "content": "Section 3, 4" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.294, + 0.878, + 0.31 + ], + "angle": 0, + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.315, + 0.215, + 0.33 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.341, + 0.883, + 0.373 + ], + "angle": 0, + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.375, + 0.249, + 0.389 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.4, + 0.884, + 0.448 + ], + "angle": 0, + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.45, + 0.249, + 0.464 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.476, + 0.884, + 0.523 + ], + "angle": 0, + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.525, + 0.249, + 0.539 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.55, + 0.875, + 0.565 + ], + "angle": 0, + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.567, + 0.249, + 0.582 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.593, + 0.881, + 0.624 + ], + "angle": 0, + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.627, + 0.249, + 0.641 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1650" + } + ] +] \ No newline at end of file diff --git a/2023/When to Use Efficient Self Attention_ Profiling Text, Speech and Image Transformer Variants/8b2664b3-8c7e-4af7-bd6b-26db06c033c3_origin.pdf b/2023/When to Use Efficient Self Attention_ Profiling Text, Speech and Image Transformer Variants/8b2664b3-8c7e-4af7-bd6b-26db06c033c3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..97019081e2a0a69e3ad514c41e85a5c7f8bc4755 --- /dev/null +++ b/2023/When to Use Efficient Self Attention_ Profiling Text, Speech and Image Transformer Variants/8b2664b3-8c7e-4af7-bd6b-26db06c033c3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eb7401f820dfef654c3591eae025ddc04c281096514f76d5cd4972588340f669 +size 871922 diff --git a/2023/When to Use Efficient Self Attention_ Profiling Text, Speech and Image Transformer Variants/full.md b/2023/When to Use Efficient Self Attention_ Profiling Text, Speech and Image Transformer Variants/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9b1e4fbb34d0894b01b59906699224f060d469bd --- /dev/null +++ b/2023/When to Use Efficient Self Attention_ Profiling Text, Speech and Image Transformer Variants/full.md @@ -0,0 +1,345 @@ +# When to Use Efficient Self Attention? Profiling Text, Speech and Image Transformer Variants + +Anuj Diwan, Eunsol Choi, David Harwath + +Department of Computer Science + +The University of Texas at Austin + +{anuj.diwan, eunsol, harwath}@utexas.edu + +# Abstract + +We present the first unified study of the efficiency of self-attention-based Transformer variants spanning text, speech and vision. We identify input length thresholds (tipping points) at which efficient Transformer variants become more efficient than vanilla models, using a variety of efficiency metrics (latency, throughput, and memory). To conduct this analysis for speech, we introduce L-HuBERT, a novel local-attention variant of a self-supervised speech model. We observe that these thresholds are (a) much higher than typical dataset sequence lengths and (b) dependent on the metric and modality, showing that choosing the right model depends on modality, task type (long-form vs. typical context) and resource constraints (time vs. memory). By visualising the breakdown of the computational costs for transformer components, we also show that non-self-attention components exhibit significant computational costs. We release our profiling toolkit at https://github.com/ajd12342/profiling-transformers. + +# 1 Introduction and Related Work + +Transformers (Vaswani et al., 2017) are widely adopted across NLP (Devlin et al., 2019; Brown et al., 2020), Speech Processing (Mohamed et al., 2022) and Computer Vision (Dosovitskiy et al., 2021). Studies have shown that scaling models up improves performance (Chowdhery et al., 2022), making efficiency an important research topic. Many Transformer variants focus on improving the efficiency of self-attention, motivated by its asymptotic quadratic time/space complexity with respect to the input sequence length. While these Transformer variants are designed be asymptotically faster, in practice they may actually be slower, especially given modest input lengths that are typical of many tasks. + +Our paper presents two main analyses. First, we visualize the layerwise efficiency of such models to locate bottlenecks and attempt to answer the question "is self-attention the true bottleneck?" We find that in the non-asymptotic case, non-self-attention layers contribute significantly to the overall cost, especially for speech architectures due to the input waveform tokenizer in models like HuBERT (Hsu et al., 2021). Second, when should we use self-attention-based efficient Transformers? Comparing efficient variants with vanilla models at different input lengths, we find that this tipping point where efficient variants outperform vanilla architectures is much higher than typical input lengths of existing benchmarks across all modalities, calling into question the efficacy of using such efficient Transformers and requiring new benchmarks. We introduce a local-attention variant of a speech Transformer, HuBERT, to conduct this analysis. Together, our analyses suggest that current approaches that focus on improving self-attention might not be the most effective for improving efficiency. + +# 2 Efficiency Metrics + +Model efficiency is an umbrella term for a suite of efficiency metrics, which do not always correlate with, and sometimes contradict, each other (Dehghani et al., 2022). Further, different metrics are relevant to different end use-cases. To cover most use-cases, we evaluate a set of four metrics; two for computational time and two for memory usage: Throughput: Number of examples processed per sec, given inputs of a given sequence length, using the maximum possible batch size for a given GPU. Latency-Inference: Time (in ms) to run inference for 1 unbatched input of a given sequence length. Max-Memory: The allocated GPU memory (MiB) for processing 1 input of a given sequence length. Parameter Count: Number of model parameters. + +We profile models in all modalities in training mode and inference mode. For training, while + +![](images/6366a7ff2881849f4c23a64787c4e27fa3d208d30074e37d3446ccd27d22e935.jpg) +Figure 1: Transformer layer types profiled in our layerwise efficiency profiling experiments. + +Transformer architectures often use prediction heads with a larger output space (e.g., for text generation), we choose a lightweight binary classification head for profiling. + +Layerwise Efficiency Metrics We also profile some metrics and models in a layerwise fashion to locate their efficiency bottlenecks. Our goal is twofold: a) provide an empirical approach to efficient model design, as an alternative to theoretical analyses or mental models (e.g. self-attention is $O(n^{2})$ ) and b) empirically answer the question "to what degree is self-attention the bottleneck?" + +We identify important layer types (Self-Attention, Feedforward, etc.) and profile the Latency-Inference and Parameter Count metrics per-layer-type to obtain a fine-grained understanding of which layer types and indices (layer 0 vs 11) contribute the most to model efficiency costs. We use param counts as a proxy for memory (profiling real layerwise memory usage is non-trivial due to Pytorch memory allocation intricacies). We profile the layers depicted in Figure 1; more details in Appendix E. + +# 3 Local-Attention Speech Model + +Efficient transformers (Xiong et al., 2021; Ma et al., 2021) have not received as much attention in Speech as they have in NLP and CV, perhaps due to two reasons. First, there is a relative lack of long-context speech benchmarks as compared to those in NLP (LRA (Tay et al., 2021) and QuALITY (Pang et al., 2022)). Second, when performing speech + +
ModelWER ↓WER (w/ FT) ↓
HuBERT Base7.093.4
L-HuBERT (32 | 100)21.06 | 14.488.52 | 7.39
+ +Table 1: WERs on the SUPERB ASR task. + +tasks like automatic speech recognition (ASR), it is typical to segment a long speech signal into small individual utterances and perform ASR independently on each. For example, most Librispeech examples are less than 5 seconds. Many popular speech models like HuBERT (Hsu et al., 2021) tokenize the waveform at 50 tokens per second, implying that a typical utterance has only several hundred tokens; far below the number of tokens in long-context NLP tasks. Nevertheless, textless speech models (Lakhotia et al., 2021) are becoming more feasible, motivating the modelling of long speech utterances. + +Local HuBERT Model To investigate the efficiency of the self-attention layer in speech models, we introduce the Local HuBERT model which replaces HuBERT's self-attention with the Longformer (Beltagy et al., 2020) sliding-window self-attention. In this attention mechanism, every token attends to tokens within a local window context, rather than the full token sequence. Our model is similar to the temporally windowed-attention Transformer acoustic model proposed by Alastruey et al. (2021) for speech translation; our approach differs by using the self-supervised HuBERT model as our basis, and we evaluate on ASR. Choosing the appropriate window size for the local attention context is key; we explore 32 and 100 token contexts, corresponding to $640~\mathrm{ms}$ and $2\mathrm{s}$ , inspired by phone recognition models that typically incorporate similar context sizes (Peddinti et al., 2015; feng Yeh et al., 2019). + +ASR Results We initialize the L-HuBERT model with pretrained HuBERT Base weights (pretrained with full self-attention), and then replace self-attention with sliding-window self-attention; due to limited compute, we did not pretrain L-HuBERT from scratch using sliding-window attention. We then evaluate L-HuBERT on Librispeech (Panayotov et al., 2015) ASR via the SUPERB (Yang et al., 2021) benchmark under two settings; a) Freeze: freezing the model and only training projection weights and b) Finetune: fully finetune the model. We use the default S3PRL2 hyperparams; but we + +
ModelEmbPosSAIntermOutputOthers
BERT23.8M-29M28.3M28.3M0.6M
HuBERT4.2M5.1M29M28.3M28.3M0.2M
ViT0.6M-29M28.3M28.3M0.6M
+ +Table 2: Layer-wise parameter counts. Emb: Input Embedding, Pos: Positional Emb. SA: Self-Attention, Interm: Intermediate. + +train for 200k steps for Freeze and 104k steps for Finetune. Both models converge by 104k steps; we train Freeze for longer to eke out as much performance as possible, while we stop training Finetune due to limited compute. + +We report Word Error Rate (WER) on Librispeech test-clean in Table 1; lower is better. In the frozen setting (middle column), we see a large WER increase over HuBERT; we hypothesize that this is due to the attention layer mismatch since we initialize L-HuBERT with HuBERT weights that were pretrained with full self attention, rather than pretraining L-HuBERT from scratch. However, in the finetuning setting, the gap between HuBERT Base and L-HuBERT narrows considerably and using a larger window size achieves better performance. As our L-HuBERT model is a reasonable architecture capable of moderate ASR performance, we can continue to study its computational efficiency (we profile the window-100 variant). + +# 4 Methods and Implementation + +We analyze the Base versions of the BERT (Devlin et al., 2019), Longformer (Beltagy et al., 2020) and Nystromformer (Xiong et al., 2021) models for text; the HuBERT (Hsu et al., 2021) and L-HuBERT (Section 3) models for speech; and Vision Transformer (Dosovitskiy et al., 2021) and Swin Transformer (Liu et al., 2021) models for vision; BERT, HuBERT and ViT are standard Transformer encoder architectures. Longformer, L-HuBERT and Swin use fixed-pattern self-attention while Nystromformer uses approximate self-attention. + +# 4.1 Sequence Length Ranges + +We profile our models on a wide range of input sequence lengths to cover both avg. sequence lengths of commonly used contemporary datasets (Table 3) and typical sequence lengths of long-context tasks. Details about how we compute sequence lengths in Table 3 can be found in Appendix B. Most image datasets use images resized to 224 or 512 pixels. Below, $\text{range}(a, b, c)$ means a range from $a$ to $b$ + +in steps of $c$ . Since there is no difference between synthetic and real inputs from a computational complexity standpoint, we use synthetic inputs to more easily control for their sequence lengths. + +Text Modality The input is 'This is a sentence.' repeated $n$ times, $n \in \text{range}(10, 560, 10)$ i.e. range(62, 3362, 60) tokens for all tokenizers. + +Speech Modality The inputs have durations in range(1, 50, 0.5) sec i.e. range(50, 2500, 25) tokens for all featurizers (CNNs with 20 ms framerate). Our sampling strategy is in Appendix A. + +Image Modality We use square inputs of dimension in range(32, 1024, 32) pixels by rescaling a fixed image. The # tokens depend on featurizer patch size, which is different for different models. + +# 4.2 Implementational Details + +We profile time-based metrics (latency/throughput) using Pytorch CUDA Events3 by executing 20 iterations sequentially. The first few iterations serve as GPU warm-start; thus, we report the average of the last 10. We record Max-Memory with torch.cuda.max_memory_allocated() and param counts with torchinfo (Yep, 2020). + +To profile throughput, we approximate the max batch size that fits on a single GPU using a linear estimator; more details in Appendix C. Finally, we profile the layerwise Latency-Inference metric using torchprof (Wong, 2020). We attach profiling hooks to modules of interest (e.g. Self-Attention, Embedding), giving us execution times of their forward() functions (other modules/functions are not profiled). We use the Huggingface (Wolf et al., 2020) implementations of text and image models and fairseq (Ott et al., 2019) implementations for speech models; more details in Appendix D. + +# 5 Profiling Results + +# 5.1 Layerwise Profiling Results + +Figure 2 shows the layerwise Latency-Inference for all 3 vanilla architectures in each modality. Figures for efficient models are in Appendix F. Color darkness represents the layer index (layer 0 is darkest). Table 2 shows the layerwise param count. + +Asymptotically, self-attention dominates the computation. However, since the average seq length for most text and speech tasks is less than 1000 tokens and most image datasets are used at + +
TextSpeech
Dataset # of tokensSST 23MNLI 36SQ 177ON 506CNN 863HPQA 1,316TQA 6,589TEDL 301LJS 328VoxC 390Libri 615S-SQuAD 3080Spotify 101400
+ +Table 3: Average token sequence lengths. Left to right: Stanford Sentiment Treebank, MultiNLI, SQuAD2.0, OntoNotes, CNN-DailyMail, HotpotQA, TriviaQA, TEDLIUM, LJSpeech, VoxCeleb Speaker Recognition, Librispeech, Spoken SQuAD, Spotify Podcasts. + +![](images/2dc8d48cf017d296c3e1dce24a493e6eefd6a3ba25dee3586c4abcf985ed3a99.jpg) +Figure 2: Layerwise latency of different vanilla Transformer architectures in inference mode. + +![](images/8cae2bfbc1133a9c1217e49b9329d11ff210a9f328dc8e2a38b8574465775c33.jpg) +Figure 3: Overall Inference-time Profiling Results. Text and speech models in first row, vision models in second. + +a max dimension of 512, at these points, non-self-attention components take up $35\%$ , $58.8\%$ and $43.75\%$ latency for NLP, speech and images. Additionally, parameter counts of SA are also comparable to Interm/Output layers. This shows that it is also important to direct efficiency efforts for other model components. + +While the latency associated with embedding layers is minimal for BERT, they are sizable for HuBERT. HuBERT uses a CNN feature extractor with different strides and kernel sizes and consumes more time in the earlier CNN layers as opposed to later ones, as is visible in Figure 2, which shows + +darker shades i.e. earlier layers dominating the computation. Optimal efficiency strategies can thus differ across modalities, e.g. Wu et al. (2022) slims down this CNN feature extractor embedding layer. On the other hand, embedding layers take up a lot of parameters in BERT; thus, it may be helpful to shrink the BERT embedding layer for memory purposes (as opposed to latency for HuBERT). Finally, analyzing Transformer variants (Appendix F), we see that self-attention in Longformer, Swin and L-HuBERT encouragingly scales latency linearly, but with large overhead for smaller inputs. + +![](images/c8dfd713d4b58aeab8460506b9f6b169349523c936c974429fc483c1dcbdf442.jpg) +Training-Throughput + +![](images/885ee6c268af0864b801e34d825ed27426d4a43a17dd6a3873856d4e459defeb.jpg) +Training-Max-Memory + +![](images/285c81d7a81a18a21d06935526ec682e6da54ebea76ece4d543d9fee41a78179.jpg) +Figure 4: Overall Training-time Profiling Results. Text and speech models in first row, vision models in second. + +![](images/c19367041d0a197782c686d10827d610859db53f3505c2e1a485e6c1377afce1.jpg) + +# 5.2 Overall Profiling Results + +Our profiling results are in Figures 3 and 4. Inference Throughput is in the Appendix at Figure 6, exhibiting similar trends as training Throughput. + +Tipping Point Analysis We see that most variants are slower and more memory hungry than vanilla models for input lengths of typical-context tasks. We define the tipping point for each modality: the input length at which the variant becomes more efficient than the vanilla model. For text and speech, it is $1750 - 2000$ tokens for inference latency and max-memory, greater than typical input lengths (Table 3). However, while the tipping point for training max-memory is $\approx 1500$ tokens for text (still a large number), it is $\approx 0 - 250$ for speech, an encouraging result. For images, it is $500 - 700$ pixels for all metrics apart from throughput. This is less reasonable for 224 pixel datasets but good for high resolution image datasets (512/1024). All variants are either worse or comparable than vanilla models across modalities for throughput. + +We hypothesize that some efficient models suffer from additional overheads; while vanilla attention benefits from highly optimized matrix multiplication, windowed attention requires complex reshaping and preprocessing. + +Choosing the Right Model Depends on Resource Constraints Our results show that the choice of the right model depends on resource constraints. Suppose that one is training models under a time constraint; then, throughput is the bottleneck and + +efficient models would not be a good fit. On the other hand, efficient models are useful for long context memory-constrained inference. + +Local Attention and Excessive Padding The Longformer pads input lengths to be a multiple of 512 and Swin requires input dimension to be a multiple of 224. This slows shorter inputs down and results in extremely low performance (measured by all 3 metrics) as compared to vanilla models. + +Comparing Parameter Counts The Longformer uses more parameters compared to vanilla BERT (148M vs. 109M) because it uses two sets of Q,K,V projection matrices for its global and local attention operations; sharing these may decrease its memory usage. For other modalities, efficient models do not incur more parameters. + +# 6 Conclusion + +We present an empirical efficiency analysis of vanilla Transformers and their self-attention-based efficient variants across modalities, metrics and input context sizes. We find substantial differences across modalities and metrics when analyzing the tipping point for efficient variants. Finally, the layerwise analysis finds that self-attention is not the only bottleneck. We recommend that all efficient model papers should report such cross-modal, layerwise profiling results on multiple efficiency metrics covering a variety of use-cases to provide a full picture of the benefits of the model. + +# Limitations + +We focus primarily on comparing model efficiencies using a variety of efficiency metrics and do not consider model performance; one can perform a more elaborate analysis of performance-efficiency tradeoffs, which we did not do here. + +We only profile a total of seven models across three modalities while there are more efficient variants and vanilla Transformers proposed in the literature. While we choose our models to be as representative of each modality and efficiency technique as possible, we cannot extrapolate results to other model variants and other modalities. In particular, modalities like video and genomics and efficiency approaches like quantization would be interesting to profile, which we did not do. + +# Acknowledgements + +We thank the reviewers and the meta-reviewer of the ACL community for helpful feedback on the draft. This work was partially funded by a grant from UT Machine Learning Lab. + +# References + +Belen Alastruey, Gerard I. Gållego, and Marta R. Costajussa. 2021. Efficient Transformer for Direct Speech Translation. +Iz Beltagy, Matthew E. Peters, and Arman Cohan. 2020. Longformer: The Long-Document Transformer. ArXiv preprint, abs/2004.05150. +Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language Models are Few-Shot Learners. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual. +Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob + +Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan Firat, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. 2022. PaLM: Scaling Language Modeling with Pathways. +Ann Clifton, Sravana Reddy, Yongze Yu, Aasish Pappu, Rezvaneh Rezapour, Hamed Bonab, Maria Eskevich, Gareth Jones, Jussi Karlgren, Ben Carterette, and Rosie Jones. 2020. 100,000 Podcasts: A Spoken English Document Corpus. In Proceedings of the 28th International Conference on Computational Linguistics, pages 5903-5917, Barcelona, Spain (Online). International Committee on Computational Linguistics. +Mostafa Dehghani, Yi Tay, Anurag Arnab, Lucas Beyer, and Ashish Vaswani. 2022. The Efficiency Mismonomer. In *The Tenth International Conference on Learning Representations*, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net. +Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics. +Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. 2021. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net. +Ching feng Yeh, Jay Mahadeokar, Kaustubh Kalgaonkar, Yongqiang Wang, Duc Le, Mahaveer Jain, Kjell Schubert, Christian Fuegen, and Michael L. Seltzer. 2019. Transformer-transducer: End-to-end speech recognition with self-attention. ArXiv, abs/1910.12977. +Karl Moritz Hermann, Tomás Kocisky, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa Suleyman, and Phil Blunsom. 2015. Teaching Machines to Read and Comprehend. In Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December + +7-12, 2015, Montreal, Quebec, Canada, pages 1693-1701. +François Hernandez, Vincent Nguyen, Sahar Ghannay, Natalia Tomashenko, and Yannick Esteve. Ted-lium 3: Twice as much data and corpus repartition for experiments on speaker adaptation. In Speech and Computer, pages 198-208. Springer International Publishing. +Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. 2021. HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 29:3451-3460. +Keith Ito and Linda Johnson. 2017. The LJ Speech Dataset. https://keithito.com/LJ-Speech-Dataset/. +Mandar Joshi, Eunsol Choi, Daniel Weld, and Luke Zettlemoyer. 2017. TriviaQA: A large scale distantly supervised challenge dataset for reading comprehension. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1601-1611, Vancouver, Canada. Association for Computational Linguistics. +Kushal Lakhotia, Eugene Kharitonov, Wei-Ning Hsu, Yossi Adi, Adam Polyak, Benjamin Bolte, Tu-Anh Nguyen, Jade Copet, Alexei Baevski, Abdelrahman Mohamed, and Emmanuel Dupoux. 2021. On generative spoken language modeling from raw audio. Transactions of the Association for Computational Linguistics, 9:1336-1354. +Chia-Hsuan Li, Szu-Lin Wu, Chi-Liang Liu, and Hung-yi Lee. 2018. Spoken SQuAD: A Study of Mitigating the Impact of Speech Recognition Errors on Listening Comprehension. In *Interspeech* 2018, 19th Annual Conference of the International Speech Communication Association, Hyderabad, India, 2-6 September 2018, pages 3459-3463. ISCA. +Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. 2021. Swin Transformer: Hierarchical Vision Transformer using Shfted Windows. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 9992-10002. IEEE. +Xuezhe Ma, Xiang Kong, Sinong Wang, Chunting Zhou, Jonathan May, Hao Ma, and Luke Zettlemoyer. 2021. Luna: Linear Unified Nested Attention. In NeurIPS. +Abdelrahman Mohamed, Hung yi Lee, Lasse Borgholt, Jakob D. Havtorn, Joakim Edin, Christian Igel, Katrin Kirchhoff, Shang-Wen Li, Karen Livescu, Lars Maaloe, Tara N. Sainath, and Shinji Watanabe. 2022. Self-Supervised Speech Representation Learning: A Review. IEEE Journal of Selected Topics in Signal Processing, 16(6):1179-1210. + +Arsha Nagrani, Joon Son Chung, and Andrew Zisserman. 2017. VoxCeleb: A Large-Scale Speaker Identification Dataset. In *Interspeech* 2017, 18th Annual Conference of the International Speech Communication Association, Stockholm, Sweden, August 20-24, 2017, pages 2616-2620. ISCA. +Myle Ott, Sergey Edunov, Alexei Baevski, Angela Fan, Sam Gross, Nathan Ng, David Grangier, and Michael Auli. 2019. *fairoseq: A fast, extensible toolkit for sequence modeling.* In *Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics* (Demonstrations), pages 48-53, Minneapolis, Minnesota. Association for Computational Linguistics. +Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. 2015. Librispeech: An ASR corpus based on public domain audio books. In 2015 IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP 2015, South Brisbane, Queensland, Australia, April 19-24, 2015, pages 5206-5210. IEEE. +Richard Yuanzhe Pang, Alicia Parrish, Nitish Joshi, Nikita Nangia, Jason Phang, Angelica Chen, Vishakh Padmakumar, Johnny Ma, Jana Thompson, He He, and Samuel Bowman. 2022. Quality: Question answering with long input texts, yes! In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5336-5358, Seattle, United States. Association for Computational Linguistics. +Vijayaditya Peddinti, Daniel Povey, and Sanjeev Khudanpur. 2015. A time delay neural network architecture for efficient modeling of long temporal contexts. In Proc. Interspeech 2015, pages 3214-3218. +Sameer S. Pradhan and Nianwen Xue. 2009. OntoNotes: The $90\%$ solution. In Proceedings of Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the Association for Computational Linguistics, Companion Volume: Tutorial Abstracts, pages 11-12, Boulder, Colorado. Association for Computational Linguistics. +Pranav Rajpurkar, Robin Jia, and Percy Liang. 2018. Know what you don't know: Unanswerable questions for SQuAD. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 784-789, Melbourne, Australia. Association for Computational Linguistics. +Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1631-1642, Seattle, Washington, USA. Association for Computational Linguistics. + +Yi Tay, Mostafa Dehghani, Samira Abnar, Yikang Shen, Dara Bahri, Philip Pham, Jinfeng Rao, Liu Yang, Sebastian Ruder, and Donald Metzler. 2021. Long Range Arena: A Benchmark for Efficient Transformers. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net. +Yi Tay, Mostafa Dehghani, Dara Bahri, and Donald Metzler. 2022. Efficient Transformers: A Survey. In ACM Comput. Surv., volume 55, New York, NY, USA. Association for Computing Machinery. +Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is All You Need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 5998-6008. +Adina Williams, Nikita Nangia, and Samuel Bowman. 2018. A broad-coverage challenge corpus for sentence understanding through inference. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1112-1122, New Orleans, Louisiana. Association for Computational Linguistics. +Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online. Association for Computational Linguistics. +Alexander William Wong. 2020. torchprof. https://github.com/awwong1/torchprof. +Felix Wu, Kwangyoun Kim, Jing Pan, Kyu J. Han, Kilian Q. Weinberger, and Yoav Artzi. 2022. Performance-Efficiency Trade-Offs in Unsupervised Pre-Training for Speech Recognition. In ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7667-7671. +Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, and Vikas Singh. 2021. Nystromformer: A Nystrom-based Algorithm for Approximating Self-Attention. In Thirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2021, Thirty-Third Conference on Innovative Applications of Artificial Intelligence, IAAI 2021, The Eleventh Symposium on Educational Advances in Artificial Intelligence, EAAI 2021, Virtual Event, February 2-9, 2021, pages 14138-14148. AAAI Press. + +Shu-Wen Yang, Po-Han Chi, Yung-Sung Chuang, Cheng-I Jeff Lai, Kushal Lakhotia, Yist Y. Lin, Andy T. Liu, Jiatong Shi, Xuankai Chang, Guanting Lin, Tzu-Hsien Huang, Wei-Cheng Tseng, Kotik Lee, Da-Rong Liu, Zili Huang, Shuyan Dong, Shang-Wen Li, Shinji Watanabe, Abdelrahman Mohamed, and Hung-yi Lee. 2021. SUPERB: Speech Processing Universal PERformance Benchmark. In Interspeech 2021, 22nd Annual Conference of the International Speech Communication Association, Brno, Czechia, 30 August - 3 September 2021, pages 1194-1198. ISCA. +Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. 2018. HotpotQA: A dataset for diverse, explainable multi-hop question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2369-2380, Brussels, Belgium. Association for Computational Linguistics. +Tyler Yep. 2020. torchinfo. https://github.com/ TylerYep/torchinfo. + +# A Sampling Speech Utterances for Profiling + +To obtain speech inputs of length $i$ seconds to $i + 0.5$ seconds for all $i$ less than 12 seconds, we sample 5 speech utterances from the training set of the Librispeech dataset (Panayotov et al., 2015) whose lengths fall within this range and compute aggregate metrics over these 5 utterances. Since the Librispeech dataset does not contain extremely long speech utterances, for $i$ of length greater than 12 seconds, we adopt a different approach to generate inputs. To generate such an input utterance of length between $i$ and $i + 0.5$ seconds, we first sample 5 speech utterances from the Librispeech training set of input length ranging from $\frac{i}{5}$ to $\frac{i + 0.5}{5}$ and concatenate them to obtain utterances of length ranging from $i$ to $i + 0.5$ as desired. We do this 5 times to get 5 different utterances and compute aggregate metrics over these 5 utterances. + +# B Computing Token Lengths for NLP and Speech Datasets + +We compute average sequence token lengths for 7 NLP datasets and 6 speech datasets. For all speech datasets, we compute mean utterance durations and multiply durations by 50 to obtain number of tokens (model framereates are $20\mathrm{ms}$ i.e. $\times 50$ ). For TEDLIUM (Hernandez et al.), LJSpeech (Ito and Johnson, 2017), VoxCeleb Speaker Recognition Dataset (Nagrani et al., 2017) and Librispeech (Panayotov et al., 2015), we compute + +mean validation-set utterance durations; for Spoken SQuAD (Li et al., 2018), we report mean validation-set paragraph duration and for the Spotify English Podcasts dataset (Clifton et al., 2020), we report mean podcast duration directly obtained from Clifton et al. (2020). + +SST (Socher et al., 2013). We use test-set sentences. We use the HuggingFace BERTTokenizer. + +MNLI (Williams et al., 2018). We use validation-matched-set examples by concatenating the premise and the hypothesis. We use the HuggingFace BERTTokenizer. + +SQuAD2.0 (Rajpurkar et al., 2018). We use validation-set examples by concatenating the context and the question. We use the HuggingFace BERTTokenizer. + +OntoNotes (Pradhan and Xue, 2009). We obtain this number from the Longformer (Beltagy et al., 2020) paper. + +CNN-Dailymail (Hermann et al., 2015). We use the 3.0.0 version of the dataset and use test-set articles. We use the HuggingFace BERTTokenizer. + +HotpotQA (Yang et al., 2018). We obtain this number from the Longformer (Beltagy et al., 2020) paper. + +TriviaQA (Joshi et al., 2017). We obtain this number from the Longformer (Beltagy et al., 2020) paper. + +# C Implementing Throughput Profiling + +To profile Throughput, we need to compute the max batch size that can fit on a single GPU. We approximately predict this using a linear estimator as follows. We first record the memory $B$ reserved on the GPU after just loading the model. Next, we independently run batches of sizes 1 and 2 and record memory usages $M_{1}$ and $M_{2}$ . We use an NVIDIA Quadro RTX 8000 GPU with a maximum memory of 45000 MiB. Thus, assuming a linear relationship between batch size and memory consumption, we predict a maximum batch size of $bsz = \frac{45000 - B}{M_2 - M_1}$ . In practice, this is an overestimate; we keep decreasing the batch size by a factor of 0.9 until no OOM errors occur and this is our final estimate. + +# D Implementational Details for Models + +We use the following HuggingFace configurations: bert-base-uncased for BERT, allenai/longformer-base-4096 for Longformer, uw-madison/nystromformer-4096 for Nyströmformer, + +google/vit-base-patch16-224 for ViT and microsoft/swin-base-patch4-window7-224 for Swin. The BERT model natively supports a maximum of 512 tokens as input because it has 512 positional embeddings; we modify the positional embedding computation to allow an arbitrarily long input to be provided. The Longformer internally pads all input lengths to a multiple of 512. For Swin, we pad images to have an input dimension that is a multiple of 224; this is necessary due to the windowed attention mechanism in Swin. In fact, the Swin model natively supports only a $224 \times 224$ resolution; we make a small modification in order to support resolutions that are multiples of 224. We use the HuBERT Base model for both HuBERT and L-HuBERT. + +# E Transformer Layer Types + +Input Embedding Layer. (red) Maps the input sequence into fixed-dimensional embeddings. This is a linear layer for text and a CNN featurizer for image/speech. + +Positional Embedding Layer. (fuchsia) For text and image models this is part of the input embedding layer. For speech models, this is a very wide convolution layer. + +Self Attention Layer.(■/blue) The multi-head self attention block, which computes self-attention outputs and maps the result to the model dimension. + +Intermediate Layer. (yellow) Linear layer of the feedforward block that maps the output from the Self Attention block into the 'feedforward dimension' (typically 4x the model dimension). + +Output Layer.(green) Second linear layer of the feedforward block, which maps the output from Intermediate layer back to the model dimension. + +Other Layers. (black) Other modules (activations, layer normalizations, other linear layers, etc.) not covered by the above components. + +# F Additional Profiling Analyses + +We report layerwise profiling runs for efficient self-attention variants and inference-time throughput profiling runs for all variants in this section at Figures 5 and 6. + +![](images/9e371d3b233e2bae162bc004d5f03449c08a5e0e27a7f339c86179214cf85375.jpg) + +![](images/9249c4691a988c422a39f4111211abfb69ad3db9893bf3ae77ed73eb23fc5cef.jpg) +Figure 5: Layerwise latency of different Transformer variants in inference mode. + +![](images/bd011fcec959415932d91ba35217eeade8d4fc19c1c0bac82ca1fb888ab942e2.jpg) + +![](images/ba8dc71a28326aa76ba3e2b655d2801599299bbb27d1924138fa59aa3d23b009.jpg) +Figure 6: Throughput Profiling Results in inference mode. + +![](images/48fd5506f2e1274e7f1b551029d302ef8116dddbc94099479951753dd4d971d1.jpg) + +A For every submission: + +A1. Did you describe the limitations of your work? + +The Limitations section + +A2. Did you discuss any potential risks of your work? + +The Limitations section + +A3. Do the abstract and introduction summarize the paper's main claims? + +Section 1 + +A4. Have you used AI writing assistants when working on this paper? + +Left blank. + +B Did you use or create scientific artifacts? + +Section 3, 4 + +B1. Did you cite the creators of artifacts you used? + +Section 3, 4 + +B2. Did you discuss the license or terms for use and / or distribution of any artifacts? + +Not explicitly, since we use publicly available Huggingface and Fairseq models that are intended for research use + +B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? + +We use publicly available Huggingface and Fairseq models that are intended for research use + +B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? + +Not applicable. Left blank. + +B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? + +Section 4 and 4.2, Appendices B,D + +B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. + +We only use datasets to profile models over different sequence lengths, but don't use the content of the dataset itself. Thus we report the relevant statistic i.e. dataset sequence length. + +C Did you run computational experiments? + +Section 3, 4. + +C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? + +Section 4.1, 4.2, Appendix C. + +C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? + +Section 4.2 + +C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? + +Section 4.2 + +C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? + +Section 3, 4 + +D Did you use human annotators (e.g., crowdworkers) or research with human participants? + +Left blank. + +D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? + +No response. + +D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? + +No response. + +D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? + +No response. + +D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? + +No response. + +D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? + +No response. \ No newline at end of file diff --git a/2023/When to Use Efficient Self Attention_ Profiling Text, Speech and Image Transformer Variants/images.zip b/2023/When to Use Efficient Self Attention_ Profiling Text, Speech and Image Transformer Variants/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..56871dcda2f984de2d41e690cce0120be4d40fc4 --- /dev/null +++ b/2023/When to Use Efficient Self Attention_ Profiling Text, Speech and Image Transformer Variants/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:03d5504780d4f637d0c7ea822b0c368aeaaf68b94960e7766287640a286f276d +size 356455 diff --git a/2023/When to Use Efficient Self Attention_ Profiling Text, Speech and Image Transformer Variants/layout.json b/2023/When to Use Efficient Self Attention_ Profiling Text, Speech and Image Transformer Variants/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1d9b5f155f2eb4339766fdeec133984223b5dfcf --- /dev/null +++ b/2023/When to Use Efficient Self Attention_ Profiling Text, Speech and Image Transformer Variants/layout.json @@ -0,0 +1,8170 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 125, + 76, + 468, + 110 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 76, + 468, + 110 + ], + "spans": [ + { + "bbox": [ + 125, + 76, + 468, + 110 + ], + "type": "text", + "content": "When to Use Efficient Self Attention? Profiling Text, Speech and Image Transformer Variants" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 188, + 126, + 407, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 126, + 407, + 139 + ], + "spans": [ + { + "bbox": [ + 188, + 126, + 407, + 139 + ], + "type": "text", + "content": "Anuj Diwan, Eunsol Choi, David Harwath" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 216, + 141, + 380, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 141, + 380, + 153 + ], + "spans": [ + { + "bbox": [ + 216, + 141, + 380, + 153 + ], + "type": "text", + "content": "Department of Computer Science" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 216, + 154, + 380, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 154, + 380, + 167 + ], + "spans": [ + { + "bbox": [ + 216, + 154, + 380, + 167 + ], + "type": "text", + "content": "The University of Texas at Austin" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 178, + 169, + 419, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 169, + 419, + 181 + ], + "spans": [ + { + "bbox": [ + 178, + 169, + 419, + 181 + ], + "type": "text", + "content": "{anuj.diwan, eunsol, harwath}@utexas.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 155, + 212, + 202, + 226 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 212, + 202, + 226 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 202, + 226 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 84, + 235, + 274, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 235, + 274, + 509 + ], + "spans": [ + { + "bbox": [ + 84, + 235, + 274, + 509 + ], + "type": "text", + "content": "We present the first unified study of the efficiency of self-attention-based Transformer variants spanning text, speech and vision. We identify input length thresholds (tipping points) at which efficient Transformer variants become more efficient than vanilla models, using a variety of efficiency metrics (latency, throughput, and memory). To conduct this analysis for speech, we introduce L-HuBERT, a novel local-attention variant of a self-supervised speech model. We observe that these thresholds are (a) much higher than typical dataset sequence lengths and (b) dependent on the metric and modality, showing that choosing the right model depends on modality, task type (long-form vs. typical context) and resource constraints (time vs. memory). By visualising the breakdown of the computational costs for transformer components, we also show that non-self-attention components exhibit significant computational costs. We release our profiling toolkit at https://github.com/ajd12342/profiling-transformers." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 68, + 520, + 251, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 520, + 251, + 533 + ], + "spans": [ + { + "bbox": [ + 68, + 520, + 251, + 533 + ], + "type": "text", + "content": "1 Introduction and Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 541, + 291, + 745 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 541, + 291, + 745 + ], + "spans": [ + { + "bbox": [ + 67, + 541, + 291, + 745 + ], + "type": "text", + "content": "Transformers (Vaswani et al., 2017) are widely adopted across NLP (Devlin et al., 2019; Brown et al., 2020), Speech Processing (Mohamed et al., 2022) and Computer Vision (Dosovitskiy et al., 2021). Studies have shown that scaling models up improves performance (Chowdhery et al., 2022), making efficiency an important research topic. Many Transformer variants focus on improving the efficiency of self-attention, motivated by its asymptotic quadratic time/space complexity with respect to the input sequence length. While these Transformer variants are designed be asymptotically faster, in practice they may actually be slower, especially given modest input lengths that are typical of many tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 213, + 526, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 213, + 526, + 511 + ], + "spans": [ + { + "bbox": [ + 302, + 213, + 526, + 511 + ], + "type": "text", + "content": "Our paper presents two main analyses. First, we visualize the layerwise efficiency of such models to locate bottlenecks and attempt to answer the question \"is self-attention the true bottleneck?\" We find that in the non-asymptotic case, non-self-attention layers contribute significantly to the overall cost, especially for speech architectures due to the input waveform tokenizer in models like HuBERT (Hsu et al., 2021). Second, when should we use self-attention-based efficient Transformers? Comparing efficient variants with vanilla models at different input lengths, we find that this tipping point where efficient variants outperform vanilla architectures is much higher than typical input lengths of existing benchmarks across all modalities, calling into question the efficacy of using such efficient Transformers and requiring new benchmarks. We introduce a local-attention variant of a speech Transformer, HuBERT, to conduct this analysis. Together, our analyses suggest that current approaches that focus on improving self-attention might not be the most effective for improving efficiency." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 303, + 521, + 417, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 521, + 417, + 534 + ], + "spans": [ + { + "bbox": [ + 303, + 521, + 417, + 534 + ], + "type": "text", + "content": "2 Efficiency Metrics" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 542, + 526, + 745 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 542, + 526, + 745 + ], + "spans": [ + { + "bbox": [ + 302, + 542, + 526, + 745 + ], + "type": "text", + "content": "Model efficiency is an umbrella term for a suite of efficiency metrics, which do not always correlate with, and sometimes contradict, each other (Dehghani et al., 2022). Further, different metrics are relevant to different end use-cases. To cover most use-cases, we evaluate a set of four metrics; two for computational time and two for memory usage: Throughput: Number of examples processed per sec, given inputs of a given sequence length, using the maximum possible batch size for a given GPU. Latency-Inference: Time (in ms) to run inference for 1 unbatched input of a given sequence length. Max-Memory: The allocated GPU memory (MiB) for processing 1 input of a given sequence length. Parameter Count: Number of model parameters." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "type": "text", + "content": "We profile models in all modalities in training mode and inference mode. For training, while" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 750, + 291, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 750, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 750, + 291, + 772 + ], + "type": "text", + "content": "1We refer the readers to Tay et al. (2022) for a comprehensive overview of efficient Transformers." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "1639" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 135, + 795, + 459, + 805 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 795, + 459, + 805 + ], + "spans": [ + { + "bbox": [ + 135, + 795, + 459, + 805 + ], + "type": "text", + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 219, + 806, + 375, + 817 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 806, + 375, + 817 + ], + "spans": [ + { + "bbox": [ + 219, + 806, + 375, + 817 + ], + "type": "text", + "content": "Volume 2: Short Papers, pages 1639-1650" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "spans": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "text", + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 79, + 221, + 279 + ], + "blocks": [ + { + "bbox": [ + 97, + 79, + 221, + 279 + ], + "lines": [ + { + "bbox": [ + 97, + 79, + 221, + 279 + ], + "spans": [ + { + "bbox": [ + 97, + 79, + 221, + 279 + ], + "type": "image", + "image_path": "6366a7ff2881849f4c23a64787c4e27fa3d208d30074e37d3446ccd27d22e935.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 292, + 291, + 318 + ], + "lines": [ + { + "bbox": [ + 67, + 292, + 291, + 318 + ], + "spans": [ + { + "bbox": [ + 67, + 292, + 291, + 318 + ], + "type": "text", + "content": "Figure 1: Transformer layer types profiled in our layerwise efficiency profiling experiments." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 327, + 291, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 327, + 291, + 381 + ], + "spans": [ + { + "bbox": [ + 67, + 327, + 291, + 381 + ], + "type": "text", + "content": "Transformer architectures often use prediction heads with a larger output space (e.g., for text generation), we choose a lightweight binary classification head for profiling." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 389, + 290, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 389, + 290, + 497 + ], + "spans": [ + { + "bbox": [ + 67, + 389, + 290, + 497 + ], + "type": "text", + "content": "Layerwise Efficiency Metrics We also profile some metrics and models in a layerwise fashion to locate their efficiency bottlenecks. Our goal is twofold: a) provide an empirical approach to efficient model design, as an alternative to theoretical analyses or mental models (e.g. self-attention is " + }, + { + "bbox": [ + 67, + 389, + 290, + 497 + ], + "type": "inline_equation", + "content": "O(n^{2})" + }, + { + "bbox": [ + 67, + 389, + 290, + 497 + ], + "type": "text", + "content": ") and b) empirically answer the question \"to what degree is self-attention the bottleneck?\"" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 497, + 291, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 497, + 291, + 647 + ], + "spans": [ + { + "bbox": [ + 67, + 497, + 291, + 647 + ], + "type": "text", + "content": "We identify important layer types (Self-Attention, Feedforward, etc.) and profile the Latency-Inference and Parameter Count metrics per-layer-type to obtain a fine-grained understanding of which layer types and indices (layer 0 vs 11) contribute the most to model efficiency costs. We use param counts as a proxy for memory (profiling real layerwise memory usage is non-trivial due to Pytorch memory allocation intricacies). We profile the layers depicted in Figure 1; more details in Appendix E." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 656, + 246, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 656, + 246, + 671 + ], + "spans": [ + { + "bbox": [ + 67, + 656, + 246, + 671 + ], + "type": "text", + "content": "3 Local-Attention Speech Model" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 678, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 678, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 678, + 291, + 773 + ], + "type": "text", + "content": "Efficient transformers (Xiong et al., 2021; Ma et al., 2021) have not received as much attention in Speech as they have in NLP and CV, perhaps due to two reasons. First, there is a relative lack of long-context speech benchmarks as compared to those in NLP (LRA (Tay et al., 2021) and QuALITY (Pang et al., 2022)). Second, when performing speech" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 307, + 68, + 522, + 113 + ], + "blocks": [ + { + "bbox": [ + 307, + 68, + 522, + 113 + ], + "lines": [ + { + "bbox": [ + 307, + 68, + 522, + 113 + ], + "spans": [ + { + "bbox": [ + 307, + 68, + 522, + 113 + ], + "type": "table", + "html": "
ModelWER ↓WER (w/ FT) ↓
HuBERT Base7.093.4
L-HuBERT (32 | 100)21.06 | 14.488.52 | 7.39
", + "image_path": "7c3aa4125714a0fa17abbecfdd53d5338c032f90bbee14c8456026120ab1c736.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 326, + 121, + 501, + 133 + ], + "lines": [ + { + "bbox": [ + 326, + 121, + 501, + 133 + ], + "spans": [ + { + "bbox": [ + 326, + 121, + 501, + 133 + ], + "type": "text", + "content": "Table 1: WERs on the SUPERB ASR task." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 302, + 140, + 526, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 140, + 526, + 317 + ], + "spans": [ + { + "bbox": [ + 302, + 140, + 526, + 317 + ], + "type": "text", + "content": "tasks like automatic speech recognition (ASR), it is typical to segment a long speech signal into small individual utterances and perform ASR independently on each. For example, most Librispeech examples are less than 5 seconds. Many popular speech models like HuBERT (Hsu et al., 2021) tokenize the waveform at 50 tokens per second, implying that a typical utterance has only several hundred tokens; far below the number of tokens in long-context NLP tasks. Nevertheless, textless speech models (Lakhotia et al., 2021) are becoming more feasible, motivating the modelling of long speech utterances." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 324, + 526, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 324, + 526, + 582 + ], + "spans": [ + { + "bbox": [ + 302, + 324, + 526, + 582 + ], + "type": "text", + "content": "Local HuBERT Model To investigate the efficiency of the self-attention layer in speech models, we introduce the Local HuBERT model which replaces HuBERT's self-attention with the Longformer (Beltagy et al., 2020) sliding-window self-attention. In this attention mechanism, every token attends to tokens within a local window context, rather than the full token sequence. Our model is similar to the temporally windowed-attention Transformer acoustic model proposed by Alastruey et al. (2021) for speech translation; our approach differs by using the self-supervised HuBERT model as our basis, and we evaluate on ASR. Choosing the appropriate window size for the local attention context is key; we explore 32 and 100 token contexts, corresponding to " + }, + { + "bbox": [ + 302, + 324, + 526, + 582 + ], + "type": "inline_equation", + "content": "640~\\mathrm{ms}" + }, + { + "bbox": [ + 302, + 324, + 526, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 302, + 324, + 526, + 582 + ], + "type": "inline_equation", + "content": "2\\mathrm{s}" + }, + { + "bbox": [ + 302, + 324, + 526, + 582 + ], + "type": "text", + "content": ", inspired by phone recognition models that typically incorporate similar context sizes (Peddinti et al., 2015; feng Yeh et al., 2019)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 590, + 527, + 754 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 590, + 527, + 754 + ], + "spans": [ + { + "bbox": [ + 302, + 590, + 527, + 754 + ], + "type": "text", + "content": "ASR Results We initialize the L-HuBERT model with pretrained HuBERT Base weights (pretrained with full self-attention), and then replace self-attention with sliding-window self-attention; due to limited compute, we did not pretrain L-HuBERT from scratch using sliding-window attention. We then evaluate L-HuBERT on Librispeech (Panayotov et al., 2015) ASR via the SUPERB (Yang et al., 2021) benchmark under two settings; a) Freeze: freezing the model and only training projection weights and b) Finetune: fully finetune the model. We use the default S3PRL2 hyperparams; but we" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 760, + 458, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 760, + 458, + 772 + ], + "spans": [ + { + "bbox": [ + 315, + 760, + 458, + 772 + ], + "type": "text", + "content": "2https://github.com/s3pr1/s3pr1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 287, + 780, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 780, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 780, + 310, + 791 + ], + "type": "text", + "content": "1640" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 80, + 68, + 278, + 123 + ], + "blocks": [ + { + "bbox": [ + 80, + 68, + 278, + 123 + ], + "lines": [ + { + "bbox": [ + 80, + 68, + 278, + 123 + ], + "spans": [ + { + "bbox": [ + 80, + 68, + 278, + 123 + ], + "type": "table", + "html": "
ModelEmbPosSAIntermOutputOthers
BERT23.8M-29M28.3M28.3M0.6M
HuBERT4.2M5.1M29M28.3M28.3M0.2M
ViT0.6M-29M28.3M28.3M0.6M
", + "image_path": "1f0876a34676d6c3cd5854b59abf5785d826528b49fc2572289352904ed667b4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 131, + 290, + 167 + ], + "lines": [ + { + "bbox": [ + 67, + 131, + 290, + 167 + ], + "spans": [ + { + "bbox": [ + 67, + 131, + 290, + 167 + ], + "type": "text", + "content": "Table 2: Layer-wise parameter counts. Emb: Input Embedding, Pos: Positional Emb. SA: Self-Attention, Interm: Intermediate." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 179, + 290, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 179, + 290, + 248 + ], + "spans": [ + { + "bbox": [ + 67, + 179, + 290, + 248 + ], + "type": "text", + "content": "train for 200k steps for Freeze and 104k steps for Finetune. Both models converge by 104k steps; we train Freeze for longer to eke out as much performance as possible, while we stop training Finetune due to limited compute." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 248, + 291, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 248, + 291, + 451 + ], + "spans": [ + { + "bbox": [ + 69, + 248, + 291, + 451 + ], + "type": "text", + "content": "We report Word Error Rate (WER) on Librispeech test-clean in Table 1; lower is better. In the frozen setting (middle column), we see a large WER increase over HuBERT; we hypothesize that this is due to the attention layer mismatch since we initialize L-HuBERT with HuBERT weights that were pretrained with full self attention, rather than pretraining L-HuBERT from scratch. However, in the finetuning setting, the gap between HuBERT Base and L-HuBERT narrows considerably and using a larger window size achieves better performance. As our L-HuBERT model is a reasonable architecture capable of moderate ASR performance, we can continue to study its computational efficiency (we profile the window-100 variant)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 463, + 241, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 463, + 241, + 477 + ], + "spans": [ + { + "bbox": [ + 67, + 463, + 241, + 477 + ], + "type": "text", + "content": "4 Methods and Implementation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 486, + 291, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 486, + 291, + 635 + ], + "spans": [ + { + "bbox": [ + 67, + 486, + 291, + 635 + ], + "type": "text", + "content": "We analyze the Base versions of the BERT (Devlin et al., 2019), Longformer (Beltagy et al., 2020) and Nystromformer (Xiong et al., 2021) models for text; the HuBERT (Hsu et al., 2021) and L-HuBERT (Section 3) models for speech; and Vision Transformer (Dosovitskiy et al., 2021) and Swin Transformer (Liu et al., 2021) models for vision; BERT, HuBERT and ViT are standard Transformer encoder architectures. Longformer, L-HuBERT and Swin use fixed-pattern self-attention while Nystromformer uses approximate self-attention." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 646, + 212, + 660 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 646, + 212, + 660 + ], + "spans": [ + { + "bbox": [ + 67, + 646, + 212, + 660 + ], + "type": "text", + "content": "4.1 Sequence Length Ranges" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 665, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 665, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 665, + 291, + 772 + ], + "type": "text", + "content": "We profile our models on a wide range of input sequence lengths to cover both avg. sequence lengths of commonly used contemporary datasets (Table 3) and typical sequence lengths of long-context tasks. Details about how we compute sequence lengths in Table 3 can be found in Appendix B. Most image datasets use images resized to 224 or 512 pixels. Below, " + }, + { + "bbox": [ + 67, + 665, + 291, + 772 + ], + "type": "inline_equation", + "content": "\\text{range}(a, b, c)" + }, + { + "bbox": [ + 67, + 665, + 291, + 772 + ], + "type": "text", + "content": " means a range from " + }, + { + "bbox": [ + 67, + 665, + 291, + 772 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 67, + 665, + 291, + 772 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 67, + 665, + 291, + 772 + ], + "type": "inline_equation", + "content": "b" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 71, + 525, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 525, + 125 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 525, + 125 + ], + "type": "text", + "content": "in steps of " + }, + { + "bbox": [ + 302, + 71, + 525, + 125 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 302, + 71, + 525, + 125 + ], + "type": "text", + "content": ". Since there is no difference between synthetic and real inputs from a computational complexity standpoint, we use synthetic inputs to more easily control for their sequence lengths." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 126, + 526, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 126, + 526, + 166 + ], + "spans": [ + { + "bbox": [ + 302, + 126, + 526, + 166 + ], + "type": "text", + "content": "Text Modality The input is 'This is a sentence.' repeated " + }, + { + "bbox": [ + 302, + 126, + 526, + 166 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 302, + 126, + 526, + 166 + ], + "type": "text", + "content": " times, " + }, + { + "bbox": [ + 302, + 126, + 526, + 166 + ], + "type": "inline_equation", + "content": "n \\in \\text{range}(10, 560, 10)" + }, + { + "bbox": [ + 302, + 126, + 526, + 166 + ], + "type": "text", + "content": " i.e. range(62, 3362, 60) tokens for all tokenizers." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 168, + 525, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 168, + 525, + 221 + ], + "spans": [ + { + "bbox": [ + 302, + 168, + 525, + 221 + ], + "type": "text", + "content": "Speech Modality The inputs have durations in range(1, 50, 0.5) sec i.e. range(50, 2500, 25) tokens for all featurizers (CNNs with 20 ms framerate). Our sampling strategy is in Appendix A." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 222, + 525, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 222, + 525, + 275 + ], + "spans": [ + { + "bbox": [ + 302, + 222, + 525, + 275 + ], + "type": "text", + "content": "Image Modality We use square inputs of dimension in range(32, 1024, 32) pixels by rescaling a fixed image. The # tokens depend on featurizer patch size, which is different for different models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 288, + 449, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 288, + 449, + 301 + ], + "spans": [ + { + "bbox": [ + 302, + 288, + 449, + 301 + ], + "type": "text", + "content": "4.2 Implementational Details" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 306, + 525, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 306, + 525, + 401 + ], + "spans": [ + { + "bbox": [ + 302, + 306, + 525, + 401 + ], + "type": "text", + "content": "We profile time-based metrics (latency/throughput) using Pytorch CUDA Events3 by executing 20 iterations sequentially. The first few iterations serve as GPU warm-start; thus, we report the average of the last 10. We record Max-Memory with torch.cuda.max_memory_allocated() and param counts with torchinfo (Yep, 2020)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 402, + 525, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 402, + 525, + 565 + ], + "spans": [ + { + "bbox": [ + 302, + 402, + 525, + 565 + ], + "type": "text", + "content": "To profile throughput, we approximate the max batch size that fits on a single GPU using a linear estimator; more details in Appendix C. Finally, we profile the layerwise Latency-Inference metric using torchprof (Wong, 2020). We attach profiling hooks to modules of interest (e.g. Self-Attention, Embedding), giving us execution times of their forward() functions (other modules/functions are not profiled). We use the Huggingface (Wolf et al., 2020) implementations of text and image models and fairseq (Ott et al., 2019) implementations for speech models; more details in Appendix D." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 576, + 409, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 576, + 409, + 590 + ], + "spans": [ + { + "bbox": [ + 302, + 576, + 409, + 590 + ], + "type": "text", + "content": "5 Profiling Results" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 600, + 458, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 600, + 458, + 613 + ], + "spans": [ + { + "bbox": [ + 302, + 600, + 458, + 613 + ], + "type": "text", + "content": "5.1 Layerwise Profiling Results" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 302, + 618, + 525, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 618, + 525, + 686 + ], + "spans": [ + { + "bbox": [ + 302, + 618, + 525, + 686 + ], + "type": "text", + "content": "Figure 2 shows the layerwise Latency-Inference for all 3 vanilla architectures in each modality. Figures for efficient models are in Appendix F. Color darkness represents the layer index (layer 0 is darkest). Table 2 shows the layerwise param count." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 302, + 687, + 525, + 740 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 687, + 525, + 740 + ], + "spans": [ + { + "bbox": [ + 302, + 687, + 525, + 740 + ], + "type": "text", + "content": "Asymptotically, self-attention dominates the computation. However, since the average seq length for most text and speech tasks is less than 1000 tokens and most image datasets are used at" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 303, + 750, + 510, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 750, + 510, + 772 + ], + "spans": [ + { + "bbox": [ + 303, + 750, + 510, + 772 + ], + "type": "text", + "content": "3https://pytorch.org/docs/stable/generated/torch.cuda.Event.html" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 287, + 781, + 308, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 308, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 308, + 791 + ], + "type": "text", + "content": "1641" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 77, + 68, + 517, + 112 + ], + "blocks": [ + { + "bbox": [ + 77, + 68, + 517, + 112 + ], + "lines": [ + { + "bbox": [ + 77, + 68, + 517, + 112 + ], + "spans": [ + { + "bbox": [ + 77, + 68, + 517, + 112 + ], + "type": "table", + "html": "
TextSpeech
Dataset # of tokensSST 23MNLI 36SQ 177ON 506CNN 863HPQA 1,316TQA 6,589TEDL 301LJS 328VoxC 390Libri 615S-SQuAD 3080Spotify 101400
", + "image_path": "39ef649a7c060b17c65f7004c4d275cb416c10082ab41b5665f8a47b47e52ca7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 116, + 526, + 154 + ], + "lines": [ + { + "bbox": [ + 67, + 116, + 526, + 154 + ], + "spans": [ + { + "bbox": [ + 67, + 116, + 526, + 154 + ], + "type": "text", + "content": "Table 3: Average token sequence lengths. Left to right: Stanford Sentiment Treebank, MultiNLI, SQuAD2.0, OntoNotes, CNN-DailyMail, HotpotQA, TriviaQA, TEDLIUM, LJSpeech, VoxCeleb Speaker Recognition, Librispeech, Spoken SQuAD, Spotify Podcasts." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 95, + 163, + 517, + 298 + ], + "blocks": [ + { + "bbox": [ + 95, + 163, + 517, + 298 + ], + "lines": [ + { + "bbox": [ + 95, + 163, + 517, + 298 + ], + "spans": [ + { + "bbox": [ + 95, + 163, + 517, + 298 + ], + "type": "image", + "image_path": "2dc8d48cf017d296c3e1dce24a493e6eefd6a3ba25dee3586c4abcf985ed3a99.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 108, + 301, + 483, + 313 + ], + "lines": [ + { + "bbox": [ + 108, + 301, + 483, + 313 + ], + "spans": [ + { + "bbox": [ + 108, + 301, + 483, + 313 + ], + "type": "text", + "content": "Figure 2: Layerwise latency of different vanilla Transformer architectures in inference mode." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 113, + 317, + 480, + 553 + ], + "blocks": [ + { + "bbox": [ + 113, + 317, + 480, + 553 + ], + "lines": [ + { + "bbox": [ + 113, + 317, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 113, + 317, + 480, + 553 + ], + "type": "image", + "image_path": "8cae2bfbc1133a9c1217e49b9329d11ff210a9f328dc8e2a38b8574465775c33.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 563, + 523, + 577 + ], + "lines": [ + { + "bbox": [ + 69, + 563, + 523, + 577 + ], + "spans": [ + { + "bbox": [ + 69, + 563, + 523, + 577 + ], + "type": "text", + "content": "Figure 3: Overall Inference-time Profiling Results. Text and speech models in first row, vision models in second." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 585, + 291, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 585, + 291, + 680 + ], + "spans": [ + { + "bbox": [ + 67, + 585, + 291, + 680 + ], + "type": "text", + "content": "a max dimension of 512, at these points, non-self-attention components take up " + }, + { + "bbox": [ + 67, + 585, + 291, + 680 + ], + "type": "inline_equation", + "content": "35\\%" + }, + { + "bbox": [ + 67, + 585, + 291, + 680 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 67, + 585, + 291, + 680 + ], + "type": "inline_equation", + "content": "58.8\\%" + }, + { + "bbox": [ + 67, + 585, + 291, + 680 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 585, + 291, + 680 + ], + "type": "inline_equation", + "content": "43.75\\%" + }, + { + "bbox": [ + 67, + 585, + 291, + 680 + ], + "type": "text", + "content": " latency for NLP, speech and images. Additionally, parameter counts of SA are also comparable to Interm/Output layers. This shows that it is also important to direct efficiency efforts for other model components." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 692, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 692, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 692, + 291, + 773 + ], + "type": "text", + "content": "While the latency associated with embedding layers is minimal for BERT, they are sizable for HuBERT. HuBERT uses a CNN feature extractor with different strides and kernel sizes and consumes more time in the earlier CNN layers as opposed to later ones, as is visible in Figure 2, which shows" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 585, + 526, + 748 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 585, + 526, + 748 + ], + "spans": [ + { + "bbox": [ + 302, + 585, + 526, + 748 + ], + "type": "text", + "content": "darker shades i.e. earlier layers dominating the computation. Optimal efficiency strategies can thus differ across modalities, e.g. Wu et al. (2022) slims down this CNN feature extractor embedding layer. On the other hand, embedding layers take up a lot of parameters in BERT; thus, it may be helpful to shrink the BERT embedding layer for memory purposes (as opposed to latency for HuBERT). Finally, analyzing Transformer variants (Appendix F), we see that self-attention in Longformer, Swin and L-HuBERT encouragingly scales latency linearly, but with large overhead for smaller inputs." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "text", + "content": "1642" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 95, + 289, + 196 + ], + "blocks": [ + { + "bbox": [ + 180, + 84, + 259, + 95 + ], + "lines": [ + { + "bbox": [ + 180, + 84, + 259, + 95 + ], + "spans": [ + { + "bbox": [ + 180, + 84, + 259, + 95 + ], + "type": "text", + "content": "Training-Throughput" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 149, + 95, + 289, + 196 + ], + "lines": [ + { + "bbox": [ + 149, + 95, + 289, + 196 + ], + "spans": [ + { + "bbox": [ + 149, + 95, + 289, + 196 + ], + "type": "image", + "image_path": "c8dfd713d4b58aeab8460506b9f6b169349523c936c974429fc483c1dcbdf442.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 304, + 97, + 443, + 196 + ], + "blocks": [ + { + "bbox": [ + 330, + 84, + 416, + 95 + ], + "lines": [ + { + "bbox": [ + 330, + 84, + 416, + 95 + ], + "spans": [ + { + "bbox": [ + 330, + 84, + 416, + 95 + ], + "type": "text", + "content": "Training-Max-Memory" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 97, + 443, + 196 + ], + "lines": [ + { + "bbox": [ + 304, + 97, + 443, + 196 + ], + "spans": [ + { + "bbox": [ + 304, + 97, + 443, + 196 + ], + "type": "image", + "image_path": "885ee6c268af0864b801e34d825ed27426d4a43a17dd6a3873856d4e459defeb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 150, + 198, + 289, + 299 + ], + "blocks": [ + { + "bbox": [ + 150, + 198, + 289, + 299 + ], + "lines": [ + { + "bbox": [ + 150, + 198, + 289, + 299 + ], + "spans": [ + { + "bbox": [ + 150, + 198, + 289, + 299 + ], + "type": "image", + "image_path": "285c81d7a81a18a21d06935526ec682e6da54ebea76ece4d543d9fee41a78179.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 310, + 520, + 322 + ], + "lines": [ + { + "bbox": [ + 71, + 310, + 520, + 322 + ], + "spans": [ + { + "bbox": [ + 71, + 310, + 520, + 322 + ], + "type": "text", + "content": "Figure 4: Overall Training-time Profiling Results. Text and speech models in first row, vision models in second." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 304, + 200, + 443, + 299 + ], + "blocks": [ + { + "bbox": [ + 304, + 200, + 443, + 299 + ], + "lines": [ + { + "bbox": [ + 304, + 200, + 443, + 299 + ], + "spans": [ + { + "bbox": [ + 304, + 200, + 443, + 299 + ], + "type": "image", + "image_path": "c19367041d0a197782c686d10827d610859db53f3505c2e1a485e6c1377afce1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 332, + 211, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 332, + 211, + 345 + ], + "spans": [ + { + "bbox": [ + 67, + 332, + 211, + 345 + ], + "type": "text", + "content": "5.2 Overall Profiling Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 349, + 291, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 349, + 291, + 391 + ], + "spans": [ + { + "bbox": [ + 67, + 349, + 291, + 391 + ], + "type": "text", + "content": "Our profiling results are in Figures 3 and 4. Inference Throughput is in the Appendix at Figure 6, exhibiting similar trends as training Throughput." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 399, + 290, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 399, + 290, + 629 + ], + "spans": [ + { + "bbox": [ + 67, + 399, + 290, + 629 + ], + "type": "text", + "content": "Tipping Point Analysis We see that most variants are slower and more memory hungry than vanilla models for input lengths of typical-context tasks. We define the tipping point for each modality: the input length at which the variant becomes more efficient than the vanilla model. For text and speech, it is " + }, + { + "bbox": [ + 67, + 399, + 290, + 629 + ], + "type": "inline_equation", + "content": "1750 - 2000" + }, + { + "bbox": [ + 67, + 399, + 290, + 629 + ], + "type": "text", + "content": " tokens for inference latency and max-memory, greater than typical input lengths (Table 3). However, while the tipping point for training max-memory is " + }, + { + "bbox": [ + 67, + 399, + 290, + 629 + ], + "type": "inline_equation", + "content": "\\approx 1500" + }, + { + "bbox": [ + 67, + 399, + 290, + 629 + ], + "type": "text", + "content": " tokens for text (still a large number), it is " + }, + { + "bbox": [ + 67, + 399, + 290, + 629 + ], + "type": "inline_equation", + "content": "\\approx 0 - 250" + }, + { + "bbox": [ + 67, + 399, + 290, + 629 + ], + "type": "text", + "content": " for speech, an encouraging result. For images, it is " + }, + { + "bbox": [ + 67, + 399, + 290, + 629 + ], + "type": "inline_equation", + "content": "500 - 700" + }, + { + "bbox": [ + 67, + 399, + 290, + 629 + ], + "type": "text", + "content": " pixels for all metrics apart from throughput. This is less reasonable for 224 pixel datasets but good for high resolution image datasets (512/1024). All variants are either worse or comparable than vanilla models across modalities for throughput." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 629, + 291, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 629, + 291, + 698 + ], + "spans": [ + { + "bbox": [ + 67, + 629, + 291, + 698 + ], + "type": "text", + "content": "We hypothesize that some efficient models suffer from additional overheads; while vanilla attention benefits from highly optimized matrix multiplication, windowed attention requires complex reshaping and preprocessing." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 67, + 706, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 706, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 706, + 291, + 772 + ], + "type": "text", + "content": "Choosing the Right Model Depends on Resource Constraints Our results show that the choice of the right model depends on resource constraints. Suppose that one is training models under a time constraint; then, throughput is the bottleneck and" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 332, + 525, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 332, + 525, + 372 + ], + "spans": [ + { + "bbox": [ + 302, + 332, + 525, + 372 + ], + "type": "text", + "content": "efficient models would not be a good fit. On the other hand, efficient models are useful for long context memory-constrained inference." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 385, + 526, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 385, + 526, + 465 + ], + "spans": [ + { + "bbox": [ + 302, + 385, + 526, + 465 + ], + "type": "text", + "content": "Local Attention and Excessive Padding The Longformer pads input lengths to be a multiple of 512 and Swin requires input dimension to be a multiple of 224. This slows shorter inputs down and results in extremely low performance (measured by all 3 metrics) as compared to vanilla models." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 477, + 525, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 477, + 525, + 571 + ], + "spans": [ + { + "bbox": [ + 302, + 477, + 525, + 571 + ], + "type": "text", + "content": "Comparing Parameter Counts The Longformer uses more parameters compared to vanilla BERT (148M vs. 109M) because it uses two sets of Q,K,V projection matrices for its global and local attention operations; sharing these may decrease its memory usage. For other modalities, efficient models do not incur more parameters." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 587, + 381, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 587, + 381, + 599 + ], + "spans": [ + { + "bbox": [ + 302, + 587, + 381, + 599 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 302, + 611, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 611, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 611, + 526, + 772 + ], + "type": "text", + "content": "We present an empirical efficiency analysis of vanilla Transformers and their self-attention-based efficient variants across modalities, metrics and input context sizes. We find substantial differences across modalities and metrics when analyzing the tipping point for efficient variants. Finally, the layerwise analysis finds that self-attention is not the only bottleneck. We recommend that all efficient model papers should report such cross-modal, layerwise profiling results on multiple efficiency metrics covering a variety of use-cases to provide a full picture of the benefits of the model." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 114, + 68, + 478, + 80 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 68, + 478, + 80 + ], + "spans": [ + { + "bbox": [ + 114, + 68, + 478, + 80 + ], + "type": "text", + "content": "■ BERT Nyströmformer Longformer HuBERT L-HuBERT ViT Swin" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1643" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 71, + 131, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 71, + 131, + 84 + ], + "spans": [ + { + "bbox": [ + 68, + 71, + 131, + 84 + ], + "type": "text", + "content": "Limitations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 93, + 291, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 93, + 291, + 158 + ], + "spans": [ + { + "bbox": [ + 67, + 93, + 291, + 158 + ], + "type": "text", + "content": "We focus primarily on comparing model efficiencies using a variety of efficiency metrics and do not consider model performance; one can perform a more elaborate analysis of performance-efficiency tradeoffs, which we did not do here." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 161, + 291, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 161, + 291, + 295 + ], + "spans": [ + { + "bbox": [ + 67, + 161, + 291, + 295 + ], + "type": "text", + "content": "We only profile a total of seven models across three modalities while there are more efficient variants and vanilla Transformers proposed in the literature. While we choose our models to be as representative of each modality and efficiency technique as possible, we cannot extrapolate results to other model variants and other modalities. In particular, modalities like video and genomics and efficiency approaches like quantization would be interesting to profile, which we did not do." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 306, + 170, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 306, + 170, + 320 + ], + "spans": [ + { + "bbox": [ + 68, + 306, + 170, + 320 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 328, + 291, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 328, + 291, + 383 + ], + "spans": [ + { + "bbox": [ + 67, + 328, + 291, + 383 + ], + "type": "text", + "content": "We thank the reviewers and the meta-reviewer of the ACL community for helpful feedback on the draft. This work was partially funded by a grant from UT Machine Learning Lab." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 405, + 127, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 405, + 127, + 417 + ], + "spans": [ + { + "bbox": [ + 68, + 405, + 127, + 417 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 424, + 291, + 773 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 69, + 424, + 290, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 424, + 290, + 458 + ], + "spans": [ + { + "bbox": [ + 69, + 424, + 290, + 458 + ], + "type": "text", + "content": "Belen Alastruey, Gerard I. Gållego, and Marta R. Costajussa. 2021. Efficient Transformer for Direct Speech Translation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 467, + 291, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 467, + 291, + 502 + ], + "spans": [ + { + "bbox": [ + 69, + 467, + 291, + 502 + ], + "type": "text", + "content": "Iz Beltagy, Matthew E. Peters, and Arman Cohan. 2020. Longformer: The Long-Document Transformer. ArXiv preprint, abs/2004.05150." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 510, + 291, + 675 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 510, + 291, + 675 + ], + "spans": [ + { + "bbox": [ + 69, + 510, + 291, + 675 + ], + "type": "text", + "content": "Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language Models are Few-Shot Learners. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 684, + 291, + 773 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 684, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 69, + 684, + 291, + 773 + ], + "type": "text", + "content": "Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 526, + 772 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 313, + 72, + 526, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 526, + 238 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 526, + 238 + ], + "type": "text", + "content": "Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan Firat, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. 2022. PaLM: Scaling Language Modeling with Pathways." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 247, + 526, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 526, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 526, + 346 + ], + "type": "text", + "content": "Ann Clifton, Sravana Reddy, Yongze Yu, Aasish Pappu, Rezvaneh Rezapour, Hamed Bonab, Maria Eskevich, Gareth Jones, Jussi Karlgren, Ben Carterette, and Rosie Jones. 2020. 100,000 Podcasts: A Spoken English Document Corpus. In Proceedings of the 28th International Conference on Computational Linguistics, pages 5903-5917, Barcelona, Spain (Online). International Committee on Computational Linguistics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 356, + 525, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 356, + 525, + 412 + ], + "spans": [ + { + "bbox": [ + 304, + 356, + 525, + 412 + ], + "type": "text", + "content": "Mostafa Dehghani, Yi Tay, Anurag Arnab, Lucas Beyer, and Ashish Vaswani. 2022. The Efficiency Mismonomer. In *The Tenth International Conference on Learning Representations*, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 422, + 525, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 525, + 522 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 525, + 522 + ], + "type": "text", + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 531, + 525, + 630 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 525, + 630 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 525, + 630 + ], + "type": "text", + "content": "Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. 2021. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 640, + 525, + 696 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 640, + 525, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 640, + 525, + 696 + ], + "type": "text", + "content": "Ching feng Yeh, Jay Mahadeokar, Kaustubh Kalgaonkar, Yongqiang Wang, Duc Le, Mahaveer Jain, Kjell Schubert, Christian Fuegen, and Michael L. Seltzer. 2019. Transformer-transducer: End-to-end speech recognition with self-attention. ArXiv, abs/1910.12977." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 706, + 525, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 706, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 706, + 525, + 772 + ], + "type": "text", + "content": "Karl Moritz Hermann, Tomás Kocisky, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa Suleyman, and Phil Blunsom. 2015. Teaching Machines to Read and Comprehend. In Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1644" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 80, + 72, + 291, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 72, + 291, + 95 + ], + "spans": [ + { + "bbox": [ + 80, + 72, + 291, + 95 + ], + "type": "text", + "content": "7-12, 2015, Montreal, Quebec, Canada, pages 1693-1701." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 104, + 290, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 104, + 290, + 170 + ], + "spans": [ + { + "bbox": [ + 69, + 104, + 290, + 170 + ], + "type": "text", + "content": "François Hernandez, Vincent Nguyen, Sahar Ghannay, Natalia Tomashenko, and Yannick Esteve. Ted-lium 3: Twice as much data and corpus repartition for experiments on speaker adaptation. In Speech and Computer, pages 198-208. Springer International Publishing." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 179, + 290, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 179, + 290, + 255 + ], + "spans": [ + { + "bbox": [ + 69, + 179, + 290, + 255 + ], + "type": "text", + "content": "Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. 2021. HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 29:3451-3460." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 264, + 290, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 264, + 290, + 298 + ], + "spans": [ + { + "bbox": [ + 69, + 264, + 290, + 298 + ], + "type": "text", + "content": "Keith Ito and Linda Johnson. 2017. The LJ Speech Dataset. https://keithito.com/LJ-Speech-Dataset/." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 307, + 290, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 307, + 290, + 385 + ], + "spans": [ + { + "bbox": [ + 69, + 307, + 290, + 385 + ], + "type": "text", + "content": "Mandar Joshi, Eunsol Choi, Daniel Weld, and Luke Zettlemoyer. 2017. TriviaQA: A large scale distantly supervised challenge dataset for reading comprehension. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1601-1611, Vancouver, Canada. Association for Computational Linguistics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 394, + 290, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 394, + 290, + 470 + ], + "spans": [ + { + "bbox": [ + 69, + 394, + 290, + 470 + ], + "type": "text", + "content": "Kushal Lakhotia, Eugene Kharitonov, Wei-Ning Hsu, Yossi Adi, Adam Polyak, Benjamin Bolte, Tu-Anh Nguyen, Jade Copet, Alexei Baevski, Abdelrahman Mohamed, and Emmanuel Dupoux. 2021. On generative spoken language modeling from raw audio. Transactions of the Association for Computational Linguistics, 9:1336-1354." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 480, + 290, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 480, + 290, + 556 + ], + "spans": [ + { + "bbox": [ + 69, + 480, + 290, + 556 + ], + "type": "text", + "content": "Chia-Hsuan Li, Szu-Lin Wu, Chi-Liang Liu, and Hung-yi Lee. 2018. Spoken SQuAD: A Study of Mitigating the Impact of Speech Recognition Errors on Listening Comprehension. In *Interspeech* 2018, 19th Annual Conference of the International Speech Communication Association, Hyderabad, India, 2-6 September 2018, pages 3459-3463. ISCA." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 565, + 290, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 565, + 290, + 642 + ], + "spans": [ + { + "bbox": [ + 69, + 565, + 290, + 642 + ], + "type": "text", + "content": "Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. 2021. Swin Transformer: Hierarchical Vision Transformer using Shfted Windows. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 9992-10002. IEEE." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 652, + 290, + 685 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 652, + 290, + 685 + ], + "spans": [ + { + "bbox": [ + 69, + 652, + 290, + 685 + ], + "type": "text", + "content": "Xuezhe Ma, Xiang Kong, Sinong Wang, Chunting Zhou, Jonathan May, Hao Ma, and Luke Zettlemoyer. 2021. Luna: Linear Unified Nested Attention. In NeurIPS." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 694, + 290, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 694, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 694, + 290, + 772 + ], + "type": "text", + "content": "Abdelrahman Mohamed, Hung yi Lee, Lasse Borgholt, Jakob D. Havtorn, Joakim Edin, Christian Igel, Katrin Kirchhoff, Shang-Wen Li, Karen Livescu, Lars Maaloe, Tara N. Sainath, and Shinji Watanabe. 2022. Self-Supervised Speech Representation Learning: A Review. IEEE Journal of Selected Topics in Signal Processing, 16(6):1179-1210." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 525, + 772 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 305, + 72, + 525, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 525, + 138 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 525, + 138 + ], + "type": "text", + "content": "Arsha Nagrani, Joon Son Chung, and Andrew Zisserman. 2017. VoxCeleb: A Large-Scale Speaker Identification Dataset. In *Interspeech* 2017, 18th Annual Conference of the International Speech Communication Association, Stockholm, Sweden, August 20-24, 2017, pages 2616-2620. ISCA." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 148, + 525, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 148, + 525, + 237 + ], + "spans": [ + { + "bbox": [ + 304, + 148, + 525, + 237 + ], + "type": "text", + "content": "Myle Ott, Sergey Edunov, Alexei Baevski, Angela Fan, Sam Gross, Nathan Ng, David Grangier, and Michael Auli. 2019. *fairoseq: A fast, extensible toolkit for sequence modeling.* In *Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics* (Demonstrations), pages 48-53, Minneapolis, Minnesota. Association for Computational Linguistics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 247, + 525, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 525, + 324 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 525, + 324 + ], + "type": "text", + "content": "Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. 2015. Librispeech: An ASR corpus based on public domain audio books. In 2015 IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP 2015, South Brisbane, Queensland, Australia, April 19-24, 2015, pages 5206-5210. IEEE." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 334, + 525, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 334, + 525, + 444 + ], + "spans": [ + { + "bbox": [ + 304, + 334, + 525, + 444 + ], + "type": "text", + "content": "Richard Yuanzhe Pang, Alicia Parrish, Nitish Joshi, Nikita Nangia, Jason Phang, Angelica Chen, Vishakh Padmakumar, Johnny Ma, Jana Thompson, He He, and Samuel Bowman. 2022. Quality: Question answering with long input texts, yes! In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5336-5358, Seattle, United States. Association for Computational Linguistics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 454, + 525, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 454, + 525, + 499 + ], + "spans": [ + { + "bbox": [ + 304, + 454, + 525, + 499 + ], + "type": "text", + "content": "Vijayaditya Peddinti, Daniel Povey, and Sanjeev Khudanpur. 2015. A time delay neural network architecture for efficient modeling of long temporal contexts. In Proc. Interspeech 2015, pages 3214-3218." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 509, + 525, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 525, + 587 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 525, + 587 + ], + "type": "text", + "content": "Sameer S. Pradhan and Nianwen Xue. 2009. OntoNotes: The " + }, + { + "bbox": [ + 304, + 509, + 525, + 587 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 304, + 509, + 525, + 587 + ], + "type": "text", + "content": " solution. In Proceedings of Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the Association for Computational Linguistics, Companion Volume: Tutorial Abstracts, pages 11-12, Boulder, Colorado. Association for Computational Linguistics." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 597, + 525, + 673 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 597, + 525, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 597, + 525, + 673 + ], + "type": "text", + "content": "Pranav Rajpurkar, Robin Jia, and Percy Liang. 2018. Know what you don't know: Unanswerable questions for SQuAD. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 784-789, Melbourne, Australia. Association for Computational Linguistics." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 684, + 525, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 684, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 684, + 525, + 772 + ], + "type": "text", + "content": "Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1631-1642, Seattle, Washington, USA. Association for Computational Linguistics." + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1645" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 291, + 150 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 291, + 150 + ], + "type": "text", + "content": "Yi Tay, Mostafa Dehghani, Samira Abnar, Yikang Shen, Dara Bahri, Philip Pham, Jinfeng Rao, Liu Yang, Sebastian Ruder, and Donald Metzler. 2021. Long Range Arena: A Benchmark for Efficient Transformers. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 158, + 291, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 158, + 291, + 204 + ], + "spans": [ + { + "bbox": [ + 69, + 158, + 291, + 204 + ], + "type": "text", + "content": "Yi Tay, Mostafa Dehghani, Dara Bahri, and Donald Metzler. 2022. Efficient Transformers: A Survey. In ACM Comput. Surv., volume 55, New York, NY, USA. Association for Computing Machinery." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 211, + 291, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 211, + 291, + 289 + ], + "spans": [ + { + "bbox": [ + 69, + 211, + 291, + 289 + ], + "type": "text", + "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is All You Need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 5998-6008." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 296, + 291, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 296, + 291, + 396 + ], + "spans": [ + { + "bbox": [ + 69, + 296, + 291, + 396 + ], + "type": "text", + "content": "Adina Williams, Nikita Nangia, and Samuel Bowman. 2018. A broad-coverage challenge corpus for sentence understanding through inference. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1112-1122, New Orleans, Louisiana. Association for Computational Linguistics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 404, + 291, + 537 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 404, + 291, + 537 + ], + "spans": [ + { + "bbox": [ + 69, + 404, + 291, + 537 + ], + "type": "text", + "content": "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 544, + 290, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 544, + 290, + 568 + ], + "spans": [ + { + "bbox": [ + 69, + 544, + 290, + 568 + ], + "type": "text", + "content": "Alexander William Wong. 2020. torchprof. https://github.com/awwong1/torchprof." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 576, + 291, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 576, + 291, + 654 + ], + "spans": [ + { + "bbox": [ + 69, + 576, + 291, + 654 + ], + "type": "text", + "content": "Felix Wu, Kwangyoun Kim, Jing Pan, Kyu J. Han, Kilian Q. Weinberger, and Yoav Artzi. 2022. Performance-Efficiency Trade-Offs in Unsupervised Pre-Training for Speech Recognition. In ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7667-7671." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 661, + 291, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 661, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 661, + 291, + 772 + ], + "type": "text", + "content": "Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, and Vikas Singh. 2021. Nystromformer: A Nystrom-based Algorithm for Approximating Self-Attention. In Thirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2021, Thirty-Third Conference on Innovative Applications of Artificial Intelligence, IAAI 2021, The Eleventh Symposium on Educational Advances in Artificial Intelligence, EAAI 2021, Virtual Event, February 2-9, 2021, pages 14138-14148. AAAI Press." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 303, + 72, + 526, + 331 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 304, + 72, + 526, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 526, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 526, + 204 + ], + "type": "text", + "content": "Shu-Wen Yang, Po-Han Chi, Yung-Sung Chuang, Cheng-I Jeff Lai, Kushal Lakhotia, Yist Y. Lin, Andy T. Liu, Jiatong Shi, Xuankai Chang, Guanting Lin, Tzu-Hsien Huang, Wei-Cheng Tseng, Kotik Lee, Da-Rong Liu, Zili Huang, Shuyan Dong, Shang-Wen Li, Shinji Watanabe, Abdelrahman Mohamed, and Hung-yi Lee. 2021. SUPERB: Speech Processing Universal PERformance Benchmark. In Interspeech 2021, 22nd Annual Conference of the International Speech Communication Association, Brno, Czechia, 30 August - 3 September 2021, pages 1194-1198. ISCA." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 303, + 211, + 526, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 211, + 526, + 301 + ], + "spans": [ + { + "bbox": [ + 303, + 211, + 526, + 301 + ], + "type": "text", + "content": "Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. 2018. HotpotQA: A dataset for diverse, explainable multi-hop question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2369-2380, Brussels, Belgium. Association for Computational Linguistics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 303, + 308, + 525, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 308, + 525, + 331 + ], + "spans": [ + { + "bbox": [ + 303, + 308, + 525, + 331 + ], + "type": "text", + "content": "Tyler Yep. 2020. torchinfo. https://github.com/ TylerYep/torchinfo." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 303, + 341, + 490, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 341, + 490, + 370 + ], + "spans": [ + { + "bbox": [ + 303, + 341, + 490, + 370 + ], + "type": "text", + "content": "A Sampling Speech Utterances for Profiling" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 376, + 525, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 376, + 525, + 607 + ], + "spans": [ + { + "bbox": [ + 302, + 376, + 525, + 607 + ], + "type": "text", + "content": "To obtain speech inputs of length " + }, + { + "bbox": [ + 302, + 376, + 525, + 607 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 302, + 376, + 525, + 607 + ], + "type": "text", + "content": " seconds to " + }, + { + "bbox": [ + 302, + 376, + 525, + 607 + ], + "type": "inline_equation", + "content": "i + 0.5" + }, + { + "bbox": [ + 302, + 376, + 525, + 607 + ], + "type": "text", + "content": " seconds for all " + }, + { + "bbox": [ + 302, + 376, + 525, + 607 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 302, + 376, + 525, + 607 + ], + "type": "text", + "content": " less than 12 seconds, we sample 5 speech utterances from the training set of the Librispeech dataset (Panayotov et al., 2015) whose lengths fall within this range and compute aggregate metrics over these 5 utterances. Since the Librispeech dataset does not contain extremely long speech utterances, for " + }, + { + "bbox": [ + 302, + 376, + 525, + 607 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 302, + 376, + 525, + 607 + ], + "type": "text", + "content": " of length greater than 12 seconds, we adopt a different approach to generate inputs. To generate such an input utterance of length between " + }, + { + "bbox": [ + 302, + 376, + 525, + 607 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 302, + 376, + 525, + 607 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 302, + 376, + 525, + 607 + ], + "type": "inline_equation", + "content": "i + 0.5" + }, + { + "bbox": [ + 302, + 376, + 525, + 607 + ], + "type": "text", + "content": " seconds, we first sample 5 speech utterances from the Librispeech training set of input length ranging from " + }, + { + "bbox": [ + 302, + 376, + 525, + 607 + ], + "type": "inline_equation", + "content": "\\frac{i}{5}" + }, + { + "bbox": [ + 302, + 376, + 525, + 607 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 302, + 376, + 525, + 607 + ], + "type": "inline_equation", + "content": "\\frac{i + 0.5}{5}" + }, + { + "bbox": [ + 302, + 376, + 525, + 607 + ], + "type": "text", + "content": " and concatenate them to obtain utterances of length ranging from " + }, + { + "bbox": [ + 302, + 376, + 525, + 607 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 302, + 376, + 525, + 607 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 302, + 376, + 525, + 607 + ], + "type": "inline_equation", + "content": "i + 0.5" + }, + { + "bbox": [ + 302, + 376, + 525, + 607 + ], + "type": "text", + "content": " as desired. We do this 5 times to get 5 different utterances and compute aggregate metrics over these 5 utterances." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 303, + 616, + 508, + 644 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 616, + 508, + 644 + ], + "spans": [ + { + "bbox": [ + 303, + 616, + 508, + 644 + ], + "type": "text", + "content": "B Computing Token Lengths for NLP and Speech Datasets" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 651, + 526, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 651, + 526, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 651, + 526, + 773 + ], + "type": "text", + "content": "We compute average sequence token lengths for 7 NLP datasets and 6 speech datasets. For all speech datasets, we compute mean utterance durations and multiply durations by 50 to obtain number of tokens (model framereates are " + }, + { + "bbox": [ + 302, + 651, + 526, + 773 + ], + "type": "inline_equation", + "content": "20\\mathrm{ms}" + }, + { + "bbox": [ + 302, + 651, + 526, + 773 + ], + "type": "text", + "content": " i.e. " + }, + { + "bbox": [ + 302, + 651, + 526, + 773 + ], + "type": "inline_equation", + "content": "\\times 50" + }, + { + "bbox": [ + 302, + 651, + 526, + 773 + ], + "type": "text", + "content": " ). For TEDLIUM (Hernandez et al.), LJSpeech (Ito and Johnson, 2017), VoxCeleb Speaker Recognition Dataset (Nagrani et al., 2017) and Librispeech (Panayotov et al., 2015), we compute" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "text", + "content": "1646" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 290, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 290, + 151 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 290, + 151 + ], + "type": "text", + "content": "mean validation-set utterance durations; for Spoken SQuAD (Li et al., 2018), we report mean validation-set paragraph duration and for the Spotify English Podcasts dataset (Clifton et al., 2020), we report mean podcast duration directly obtained from Clifton et al. (2020)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 153, + 291, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 153, + 291, + 179 + ], + "spans": [ + { + "bbox": [ + 67, + 153, + 291, + 179 + ], + "type": "text", + "content": "SST (Socher et al., 2013). We use test-set sentences. We use the HuggingFace BERTTokenizer." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 180, + 290, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 180, + 290, + 232 + ], + "spans": [ + { + "bbox": [ + 67, + 180, + 290, + 232 + ], + "type": "text", + "content": "MNLI (Williams et al., 2018). We use validation-matched-set examples by concatenating the premise and the hypothesis. We use the HuggingFace BERTTokenizer." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 234, + 290, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 234, + 290, + 287 + ], + "spans": [ + { + "bbox": [ + 67, + 234, + 290, + 287 + ], + "type": "text", + "content": "SQuAD2.0 (Rajpurkar et al., 2018). We use validation-set examples by concatenating the context and the question. We use the HuggingFace BERTTokenizer." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 289, + 290, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 289, + 290, + 328 + ], + "spans": [ + { + "bbox": [ + 67, + 289, + 290, + 328 + ], + "type": "text", + "content": "OntoNotes (Pradhan and Xue, 2009). We obtain this number from the Longformer (Beltagy et al., 2020) paper." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 329, + 290, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 329, + 290, + 369 + ], + "spans": [ + { + "bbox": [ + 67, + 329, + 290, + 369 + ], + "type": "text", + "content": "CNN-Dailymail (Hermann et al., 2015). We use the 3.0.0 version of the dataset and use test-set articles. We use the HuggingFace BERTTokenizer." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 370, + 290, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 370, + 290, + 411 + ], + "spans": [ + { + "bbox": [ + 67, + 370, + 290, + 411 + ], + "type": "text", + "content": "HotpotQA (Yang et al., 2018). We obtain this number from the Longformer (Beltagy et al., 2020) paper." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 411, + 290, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 411, + 290, + 451 + ], + "spans": [ + { + "bbox": [ + 67, + 411, + 290, + 451 + ], + "type": "text", + "content": "TriviaQA (Joshi et al., 2017). We obtain this number from the Longformer (Beltagy et al., 2020) paper." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 461, + 277, + 476 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 461, + 277, + 476 + ], + "spans": [ + { + "bbox": [ + 67, + 461, + 277, + 476 + ], + "type": "text", + "content": "C Implementing Throughput Profiling" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 484, + 291, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 484, + 291, + 672 + ], + "spans": [ + { + "bbox": [ + 67, + 484, + 291, + 672 + ], + "type": "text", + "content": "To profile Throughput, we need to compute the max batch size that can fit on a single GPU. We approximately predict this using a linear estimator as follows. We first record the memory " + }, + { + "bbox": [ + 67, + 484, + 291, + 672 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 67, + 484, + 291, + 672 + ], + "type": "text", + "content": " reserved on the GPU after just loading the model. Next, we independently run batches of sizes 1 and 2 and record memory usages " + }, + { + "bbox": [ + 67, + 484, + 291, + 672 + ], + "type": "inline_equation", + "content": "M_{1}" + }, + { + "bbox": [ + 67, + 484, + 291, + 672 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 484, + 291, + 672 + ], + "type": "inline_equation", + "content": "M_{2}" + }, + { + "bbox": [ + 67, + 484, + 291, + 672 + ], + "type": "text", + "content": ". We use an NVIDIA Quadro RTX 8000 GPU with a maximum memory of 45000 MiB. Thus, assuming a linear relationship between batch size and memory consumption, we predict a maximum batch size of " + }, + { + "bbox": [ + 67, + 484, + 291, + 672 + ], + "type": "inline_equation", + "content": "bsz = \\frac{45000 - B}{M_2 - M_1}" + }, + { + "bbox": [ + 67, + 484, + 291, + 672 + ], + "type": "text", + "content": ". In practice, this is an overestimate; we keep decreasing the batch size by a factor of 0.9 until no OOM errors occur and this is our final estimate." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 683, + 280, + 698 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 683, + 280, + 698 + ], + "spans": [ + { + "bbox": [ + 67, + 683, + 280, + 698 + ], + "type": "text", + "content": "D Implementational Details for Models" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 706, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 706, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 706, + 291, + 772 + ], + "type": "text", + "content": "We use the following HuggingFace configurations: bert-base-uncased for BERT, allenai/longformer-base-4096 for Longformer, uw-madison/nystromformer-4096 for Nyströmformer," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 71, + 525, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 525, + 300 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 525, + 300 + ], + "type": "text", + "content": "google/vit-base-patch16-224 for ViT and microsoft/swin-base-patch4-window7-224 for Swin. The BERT model natively supports a maximum of 512 tokens as input because it has 512 positional embeddings; we modify the positional embedding computation to allow an arbitrarily long input to be provided. The Longformer internally pads all input lengths to a multiple of 512. For Swin, we pad images to have an input dimension that is a multiple of 224; this is necessary due to the windowed attention mechanism in Swin. In fact, the Swin model natively supports only a " + }, + { + "bbox": [ + 302, + 71, + 525, + 300 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 302, + 71, + 525, + 300 + ], + "type": "text", + "content": " resolution; we make a small modification in order to support resolutions that are multiples of 224. We use the HuBERT Base model for both HuBERT and L-HuBERT." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 312, + 458, + 325 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 312, + 458, + 325 + ], + "spans": [ + { + "bbox": [ + 302, + 312, + 458, + 325 + ], + "type": "text", + "content": "E Transformer Layer Types" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 332, + 525, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 332, + 525, + 386 + ], + "spans": [ + { + "bbox": [ + 302, + 332, + 525, + 386 + ], + "type": "text", + "content": "Input Embedding Layer. (red) Maps the input sequence into fixed-dimensional embeddings. This is a linear layer for text and a CNN featurizer for image/speech." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 386, + 525, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 386, + 525, + 440 + ], + "spans": [ + { + "bbox": [ + 302, + 386, + 525, + 440 + ], + "type": "text", + "content": "Positional Embedding Layer. (fuchsia) For text and image models this is part of the input embedding layer. For speech models, this is a very wide convolution layer." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 440, + 525, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 440, + 525, + 481 + ], + "spans": [ + { + "bbox": [ + 302, + 440, + 525, + 481 + ], + "type": "text", + "content": "Self Attention Layer.(■/blue) The multi-head self attention block, which computes self-attention outputs and maps the result to the model dimension." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 302, + 481, + 525, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 481, + 525, + 534 + ], + "spans": [ + { + "bbox": [ + 302, + 481, + 525, + 534 + ], + "type": "text", + "content": "Intermediate Layer. (yellow) Linear layer of the feedforward block that maps the output from the Self Attention block into the 'feedforward dimension' (typically 4x the model dimension)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 302, + 535, + 524, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 535, + 524, + 576 + ], + "spans": [ + { + "bbox": [ + 302, + 535, + 524, + 576 + ], + "type": "text", + "content": "Output Layer.(green) Second linear layer of the feedforward block, which maps the output from Intermediate layer back to the model dimension." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 302, + 576, + 525, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 576, + 525, + 617 + ], + "spans": [ + { + "bbox": [ + 302, + 576, + 525, + 617 + ], + "type": "text", + "content": "Other Layers. (black) Other modules (activations, layer normalizations, other linear layers, etc.) not covered by the above components." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 302, + 626, + 476, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 626, + 476, + 640 + ], + "spans": [ + { + "bbox": [ + 302, + 626, + 476, + 640 + ], + "type": "text", + "content": "F Additional Profiling Analyses" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 302, + 648, + 525, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 648, + 525, + 700 + ], + "spans": [ + { + "bbox": [ + 302, + 648, + 525, + 700 + ], + "type": "text", + "content": "We report layerwise profiling runs for efficient self-attention variants and inference-time throughput profiling runs for all variants in this section at Figures 5 and 6." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1647" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 108, + 497, + 268 + ], + "blocks": [ + { + "bbox": [ + 97, + 108, + 497, + 268 + ], + "lines": [ + { + "bbox": [ + 97, + 108, + 497, + 268 + ], + "spans": [ + { + "bbox": [ + 97, + 108, + 497, + 268 + ], + "type": "image", + "image_path": "9e371d3b233e2bae162bc004d5f03449c08a5e0e27a7f339c86179214cf85375.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 104, + 276, + 291, + 437 + ], + "blocks": [ + { + "bbox": [ + 104, + 276, + 291, + 437 + ], + "lines": [ + { + "bbox": [ + 104, + 276, + 291, + 437 + ], + "spans": [ + { + "bbox": [ + 104, + 276, + 291, + 437 + ], + "type": "image", + "image_path": "9249c4691a988c422a39f4111211abfb69ad3db9893bf3ae77ed73eb23fc5cef.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 457, + 459, + 470 + ], + "lines": [ + { + "bbox": [ + 133, + 457, + 459, + 470 + ], + "spans": [ + { + "bbox": [ + 133, + 457, + 459, + 470 + ], + "type": "text", + "content": "Figure 5: Layerwise latency of different Transformer variants in inference mode." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 298, + 276, + 489, + 438 + ], + "blocks": [ + { + "bbox": [ + 298, + 276, + 489, + 438 + ], + "lines": [ + { + "bbox": [ + 298, + 276, + 489, + 438 + ], + "spans": [ + { + "bbox": [ + 298, + 276, + 489, + 438 + ], + "type": "image", + "image_path": "bd011fcec959415932d91ba35217eeade8d4fc19c1c0bac82ca1fb888ab942e2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 106, + 556, + 294, + 705 + ], + "blocks": [ + { + "bbox": [ + 106, + 556, + 294, + 705 + ], + "lines": [ + { + "bbox": [ + 106, + 556, + 294, + 705 + ], + "spans": [ + { + "bbox": [ + 106, + 556, + 294, + 705 + ], + "type": "image", + "image_path": "ba8dc71a28326aa76ba3e2b655d2801599299bbb27d1924138fa59aa3d23b009.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 175, + 716, + 416, + 729 + ], + "lines": [ + { + "bbox": [ + 175, + 716, + 416, + 729 + ], + "spans": [ + { + "bbox": [ + 175, + 716, + 416, + 729 + ], + "type": "text", + "content": "Figure 6: Throughput Profiling Results in inference mode." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 301, + 556, + 489, + 705 + ], + "blocks": [ + { + "bbox": [ + 301, + 556, + 489, + 705 + ], + "lines": [ + { + "bbox": [ + 301, + 556, + 489, + 705 + ], + "spans": [ + { + "bbox": [ + 301, + 556, + 489, + 705 + ], + "type": "image", + "image_path": "48fd5506f2e1274e7f1b551029d302ef8116dddbc94099479951753dd4d971d1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "text", + "content": "1648" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 90, + 192, + 103 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 90, + 192, + 103 + ], + "spans": [ + { + "bbox": [ + 68, + 90, + 192, + 103 + ], + "type": "text", + "content": "A For every submission:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 106, + 317, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 106, + 317, + 120 + ], + "spans": [ + { + "bbox": [ + 77, + 106, + 317, + 120 + ], + "type": "text", + "content": "A1. Did you describe the limitations of your work?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 121, + 195, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 121, + 195, + 132 + ], + "spans": [ + { + "bbox": [ + 89, + 121, + 195, + 132 + ], + "type": "text", + "content": "The Limitations section" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 142, + 329, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 142, + 329, + 156 + ], + "spans": [ + { + "bbox": [ + 77, + 142, + 329, + 156 + ], + "type": "text", + "content": "A2. Did you discuss any potential risks of your work?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 157, + 195, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 157, + 195, + 169 + ], + "spans": [ + { + "bbox": [ + 89, + 157, + 195, + 169 + ], + "type": "text", + "content": "The Limitations section" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 77, + 178, + 414, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 178, + 414, + 191 + ], + "spans": [ + { + "bbox": [ + 77, + 178, + 414, + 191 + ], + "type": "text", + "content": "A3. Do the abstract and introduction summarize the paper's main claims?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 90, + 193, + 132, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 193, + 132, + 203 + ], + "spans": [ + { + "bbox": [ + 90, + 193, + 132, + 203 + ], + "type": "text", + "content": "Section 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 77, + 214, + 398, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 214, + 398, + 227 + ], + "spans": [ + { + "bbox": [ + 77, + 214, + 398, + 227 + ], + "type": "text", + "content": "A4. Have you used AI writing assistants when working on this paper?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 90, + 229, + 138, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 229, + 138, + 240 + ], + "spans": [ + { + "bbox": [ + 90, + 229, + 138, + 240 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 68, + 250, + 290, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 250, + 290, + 264 + ], + "spans": [ + { + "bbox": [ + 68, + 250, + 290, + 264 + ], + "type": "text", + "content": "B Did you use or create scientific artifacts?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 79, + 269, + 133, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 269, + 133, + 280 + ], + "spans": [ + { + "bbox": [ + 79, + 269, + 133, + 280 + ], + "type": "text", + "content": "Section 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 77, + 289, + 315, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 289, + 315, + 302 + ], + "spans": [ + { + "bbox": [ + 77, + 289, + 315, + 302 + ], + "type": "text", + "content": "B1. Did you cite the creators of artifacts you used?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 90, + 304, + 144, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 304, + 144, + 315 + ], + "spans": [ + { + "bbox": [ + 90, + 304, + 144, + 315 + ], + "type": "text", + "content": "Section 3, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 77, + 326, + 463, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 326, + 463, + 338 + ], + "spans": [ + { + "bbox": [ + 77, + 326, + 463, + 338 + ], + "type": "text", + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 89, + 339, + 524, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 339, + 524, + 364 + ], + "spans": [ + { + "bbox": [ + 89, + 339, + 524, + 364 + ], + "type": "text", + "content": "Not explicitly, since we use publicly available Huggingface and Fairseq models that are intended for research use" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 77, + 374, + 524, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 374, + 524, + 428 + ], + "spans": [ + { + "bbox": [ + 77, + 374, + 524, + 428 + ], + "type": "text", + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 89, + 429, + 500, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 429, + 500, + 442 + ], + "spans": [ + { + "bbox": [ + 89, + 429, + 500, + 442 + ], + "type": "text", + "content": "We use publicly available Huggingface and Fairseq models that are intended for research use" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 77, + 451, + 524, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 451, + 524, + 491 + ], + "spans": [ + { + "bbox": [ + 77, + 451, + 524, + 491 + ], + "type": "text", + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 89, + 492, + 208, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 492, + 208, + 505 + ], + "spans": [ + { + "bbox": [ + 89, + 492, + 208, + 505 + ], + "type": "text", + "content": "Not applicable. Left blank." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 77, + 513, + 524, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 513, + 524, + 541 + ], + "spans": [ + { + "bbox": [ + 77, + 513, + 524, + 541 + ], + "type": "text", + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 90, + 541, + 244, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 541, + 244, + 554 + ], + "spans": [ + { + "bbox": [ + 90, + 541, + 244, + 554 + ], + "type": "text", + "content": "Section 4 and 4.2, Appendices B,D" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 77, + 562, + 524, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 562, + 524, + 630 + ], + "spans": [ + { + "bbox": [ + 77, + 562, + 524, + 630 + ], + "type": "text", + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 89, + 631, + 526, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 631, + 526, + 657 + ], + "spans": [ + { + "bbox": [ + 89, + 631, + 526, + 657 + ], + "type": "text", + "content": "We only use datasets to profile models over different sequence lengths, but don't use the content of the dataset itself. Thus we report the relevant statistic i.e. dataset sequence length." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 68, + 665, + 293, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 665, + 293, + 680 + ], + "spans": [ + { + "bbox": [ + 68, + 665, + 293, + 680 + ], + "type": "text", + "content": "C Did you run computational experiments?" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 79, + 684, + 135, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 684, + 135, + 696 + ], + "spans": [ + { + "bbox": [ + 79, + 684, + 135, + 696 + ], + "type": "text", + "content": "Section 3, 4." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 77, + 704, + 524, + 732 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 704, + 524, + 732 + ], + "spans": [ + { + "bbox": [ + 77, + 704, + 524, + 732 + ], + "type": "text", + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 89, + 733, + 218, + 745 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 733, + 218, + 745 + ], + "spans": [ + { + "bbox": [ + 89, + 733, + 218, + 745 + ], + "type": "text", + "content": "Section 4.1, 4.2, Appendix C." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "spans": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "text", + "content": "ACL 2023 Responsible NLP Checklist" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 751, + 522, + 772 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 751, + 522, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 751, + 522, + 772 + ], + "type": "text", + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1649" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 77, + 70, + 523, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 70, + 523, + 97 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 523, + 97 + ], + "type": "text", + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 89, + 99, + 141, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 99, + 141, + 110 + ], + "spans": [ + { + "bbox": [ + 89, + 99, + 141, + 110 + ], + "type": "text", + "content": "Section 4.2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 120, + 525, + 160 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 120, + 525, + 160 + ], + "spans": [ + { + "bbox": [ + 77, + 120, + 525, + 160 + ], + "type": "text", + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 162, + 141, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 162, + 141, + 173 + ], + "spans": [ + { + "bbox": [ + 89, + 162, + 141, + 173 + ], + "type": "text", + "content": "Section 4.2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 183, + 525, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 183, + 525, + 223 + ], + "spans": [ + { + "bbox": [ + 77, + 183, + 525, + 223 + ], + "type": "text", + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 225, + 143, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 225, + 143, + 236 + ], + "spans": [ + { + "bbox": [ + 89, + 225, + 143, + 236 + ], + "type": "text", + "content": "Section 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 247, + 522, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 247, + 522, + 260 + ], + "spans": [ + { + "bbox": [ + 67, + 247, + 522, + 260 + ], + "type": "text", + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 79, + 264, + 127, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 264, + 127, + 277 + ], + "spans": [ + { + "bbox": [ + 79, + 264, + 127, + 277 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 76, + 286, + 525, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 286, + 525, + 313 + ], + "spans": [ + { + "bbox": [ + 76, + 286, + 525, + 313 + ], + "type": "text", + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 89, + 315, + 148, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 315, + 148, + 327 + ], + "spans": [ + { + "bbox": [ + 89, + 315, + 148, + 327 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 76, + 336, + 525, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 336, + 525, + 376 + ], + "spans": [ + { + "bbox": [ + 76, + 336, + 525, + 376 + ], + "type": "text", + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 89, + 378, + 148, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 378, + 148, + 390 + ], + "spans": [ + { + "bbox": [ + 89, + 378, + 148, + 390 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 76, + 400, + 525, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 400, + 525, + 439 + ], + "spans": [ + { + "bbox": [ + 76, + 400, + 525, + 439 + ], + "type": "text", + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 89, + 441, + 148, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 441, + 148, + 453 + ], + "spans": [ + { + "bbox": [ + 89, + 441, + 148, + 453 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 76, + 462, + 520, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 462, + 520, + 475 + ], + "spans": [ + { + "bbox": [ + 76, + 462, + 520, + 475 + ], + "type": "text", + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 89, + 476, + 148, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 476, + 148, + 489 + ], + "spans": [ + { + "bbox": [ + 89, + 476, + 148, + 489 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 76, + 498, + 524, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 498, + 524, + 524 + ], + "spans": [ + { + "bbox": [ + 76, + 498, + 524, + 524 + ], + "type": "text", + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 89, + 527, + 148, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 527, + 148, + 539 + ], + "spans": [ + { + "bbox": [ + 89, + 527, + 148, + 539 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "text", + "content": "1650" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/With a Little Push, NLI Models can Robustly and Efficiently Predict Faithfulness/d3da2b49-e96a-43e5-ae6f-e575c826018a_content_list.json b/2023/With a Little Push, NLI Models can Robustly and Efficiently Predict Faithfulness/d3da2b49-e96a-43e5-ae6f-e575c826018a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a1c3ccc06702bb9a8727c051063c8fb0ebe1e983 --- /dev/null +++ b/2023/With a Little Push, NLI Models can Robustly and Efficiently Predict Faithfulness/d3da2b49-e96a-43e5-ae6f-e575c826018a_content_list.json @@ -0,0 +1,1909 @@ +[ + { + "type": "text", + "text": "With a Little Push, NLI Models can Robustly and Efficiently Predict Faithfulness", + "text_level": 1, + "bbox": [ + 144, + 87, + 852, + 127 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Julius Steen Juri Opitz Anette Frank Katja Markert", + "bbox": [ + 258, + 142, + 742, + 158 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Department of Computational Linguistics", + "bbox": [ + 331, + 160, + 670, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Heidelberg University", + "bbox": [ + 410, + 177, + 591, + 192 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "69120 Heidelberg, Germany", + "bbox": [ + 384, + 193, + 616, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(steen|opitz|frank|markert)@cl.uni-heidelberg.de", + "bbox": [ + 258, + 210, + 744, + 225 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 252, + 342, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Conditional language models still generate unfaithful output that is not supported by their input. These unfaithful generations jeopardize trust in real-world applications such as summarization or human-machine interaction, motivating a need for automatic faithfulness metrics. To implement such metrics, NLI models seem attractive, since they solve a strongly related task that comes with a wealth of prior research and data. But recent research suggests that NLI models require costly additional machinery to perform reliably across datasets, e.g., by running inference on a cartesian product of input and generated sentences, or supporting them with a question-generation/answering step.", + "bbox": [ + 141, + 282, + 460, + 495 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work we show that pure NLI models can outperform more complex metrics when combining task-adaptive data augmentation with robust inference procedures. We propose: (1) Augmenting NLI training data to adapt NL inferences to the specificities of faithfulness prediction in dialogue; (2) Making use of both entailment and contradiction probabilities in NLI, and (3) Using Monte-Carlo dropout during inference. Applied to the TRUE benchmark, which combines faithfulness datasets across diverse domains and tasks, our approach strongly improves a vanilla NLI model and significantly outperforms previous work, while showing favourable computational cost.", + "bbox": [ + 141, + 502, + 460, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 114, + 730, + 260, + 746 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Conditional language models suffer from a tendency to hallucinate information (Maynez et al., 2020), resulting in generations that are not faithful to their input documents, which limits the trustworthiness of such models. This raises a need for automatic faithfulness metrics. In this context, models trained on natural language inference (NLI) (Bowman et al., 2015) are attractive since, intuitively, a generation being faithful implies it must be entailed by the source (Falke et al., 2019).", + "bbox": [ + 112, + 758, + 490, + 919 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, pure NLI models have seen mixed success in faithfulness evaluation (Falke et al., 2019; Kryscinski et al., 2020; Wang et al., 2020; Maynez et al., 2020). While in recent evaluation on the TRUE benchmark (Honovich et al., 2022), which contains datasets from knowledge-grounded dialogue, summarization and paraphrasing, NLI-derived metrics perform best overall, they require impractically large models, or costly additional machinery such as question generation and answering models at inference, while still showing robustness issues. Thus we ask: What is still needed for pure NLI models to perform robustly across faithfulness datasets – while remaining cheap enough to serve as a lean and practical evaluation tool?", + "bbox": [ + 507, + 253, + 885, + 494 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We enhance a relatively small NLI model to make it work robustly across tasks in three ways:", + "bbox": [ + 507, + 495, + 882, + 527 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Task-Adaptive Data Augmentation. In NLI, a hypothesis must be fully entailed by its supporting premise. However, in faithfulness, not all parts of the generation always need to be grounded. We identify an instance of this phenomenon in dialogue where parts of a turn can fulfill communicative functions such as hedging or establishing emotional connection and are often disregarded in faithfulness annotation. Hence, when applying NLI models to complete dialogue turns that may include statements irrelevant for grounding, we run a risk of producing incorrect unfaithfulness predictions.", + "bbox": [ + 507, + 530, + 885, + 722 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To alleviate this issue, we propose a simple data augmentation method to adapt NLI models to genres where they need to be aware of statements that must be exempt from NLI-based faithfulness evaluation. Our approach is computationally attractive, as it avoids an increase of cost at inference time.", + "bbox": [ + 507, + 724, + 885, + 820 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Integration of NLI Contradiction Scores. Existing NLI faithfulness metrics typically use the entailment score for their predictions (Honovich et al., 2022; Falke et al., 2019; Kryscinski et al., 2020). However, Chen and Eger (2022) show that subtracting the contradiction score from the entail", + "bbox": [ + 507, + 822, + 885, + 919 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "914", + "bbox": [ + 485, + 927, + 515, + 940 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics", + "bbox": [ + 226, + 945, + 769, + 957 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Volume 2: Short Papers, pages 914-924", + "bbox": [ + 376, + 958, + 620, + 971 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "July 9-14, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 295, + 972, + 700, + 985 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ment score (referred to as $e-c$ ) can improve NLI performance in certain evaluation tasks. We show that there also is a strong positive effect of $e-c$ for faithfulness prediction, and demonstrate that this is due to a high contradiction probability being a more reliable predictor of unfaithfulness than low entailment probability.", + "bbox": [ + 112, + 84, + 487, + 197 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Monte-Carlo Dropout Inference. Applying NLI models to faithfulness prediction involves a domain shift from largely human-written data to automatically generated text. To make NLI model scores more robust under this shift, we propose to use Monte-Carlo dropout during inference (Srivastava et al., 2014). This essentially creates a cheap ensemble and has been shown to deal better with noisy labels (Goel and Chen, 2021). This approach leads to consistent score improvements in our tasks.", + "bbox": [ + 112, + 199, + 485, + 359 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The combination of all modifications not only strongly improves over a baseline NLI model, but also outperforms all other metrics on TRUE, on average, while being cheaper and smaller.", + "bbox": [ + 112, + 361, + 485, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Method Details", + "text_level": 1, + "bbox": [ + 114, + 444, + 280, + 458 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 Task-adaptive Data Augmentation", + "text_level": 1, + "bbox": [ + 114, + 475, + 428, + 491 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To illustrate that task requirements can be incompatible between faithfulness and NLI, consider the following instance from the Q2 dialogue corpus (Honovich et al., 2021) that is labelled as faithful:", + "bbox": [ + 112, + 500, + 487, + 563 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Grounding: American pancakes are similar to Scotch pancakes or drop scones.", + "bbox": [ + 149, + 582, + 450, + 614 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generation: yes, i love american pancakes, they are like scotch pancakes", + "bbox": [ + 149, + 615, + 452, + 646 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "From an NLI perspective, the generation is clearly not entailed, since the statement \"I love american pancakes\" is not supported by the input.", + "bbox": [ + 112, + 665, + 485, + 714 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To better prepare an NLI system for such genre or task-specific cases, we manually curate a small list of statements that should not influence the faithfulness prediction. We augment NLI data from the ANLI corpus (Nie et al., 2020) by adding a randomly chosen phrase from this set to each instance, while preserving the label. We then train an already fine-tuned NLI model on a concatenation of these augmented samples and original ANLI data. For training details see Appendix A.", + "bbox": [ + 112, + 715, + 487, + 876 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2 Monte-Carlo Dropout", + "text_level": 1, + "bbox": [ + 507, + 84, + 727, + 99 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To compute scores under Monte-Carlo dropout, we randomly sample $k$ dropout masks and compute the average of the model predictions. We set $k = 15$ , since preliminary experiments showed that performance did not profit from additional samples.", + "bbox": [ + 507, + 105, + 882, + 186 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3 Experimental Setup", + "text_level": 1, + "bbox": [ + 507, + 198, + 717, + 215 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We run experiments on TRUE (Honovich et al., 2022), a benchmark that compiles a wide variety of faithfulness tasks in a standardized format. It contains summarization (Pagnoni et al., 2021; Maynez et al., 2020; Wang et al., 2020; Fabbri et al., 2021), knowledge-grounded dialog (Honovich et al., 2021; Gupta et al., 2022; Dziri et al., 2022) and paraphrasing (Zhang et al., 2019) datasets. Following recommendations in TRUE, we evaluate using Area under the ROC Curve (AUC).", + "bbox": [ + 507, + 223, + 882, + 384 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As our BASE model, we use the DeBERTa-large (He et al., 2020) model of Laurer et al. (2022), trained on MultiNLI (Williams et al., 2018), FeverNLI (Thorne et al., 2018), ANLI (Nie et al., 2020), LingNLI (Parrish et al., 2021) and WANLI (Liu et al., 2022). The metric A11 uses all three of our proposed modifications to Base. We also investigate a variant without MC dropout inference $(-MC)$ as a more cost efficient alternative.", + "bbox": [ + 507, + 385, + 882, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We compare to the strongest models on TRUE:", + "bbox": [ + 527, + 530, + 878, + 545 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "T5 ANLI (Honovich et al., 2022) is a T5-11B (Raffel et al., 2020) model trained on ANLI.4", + "bbox": [ + 507, + 546, + 880, + 577 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SummacZS (Laban et al., 2022) evaluates an NLI model on all pairs of input and generated sentences and then averages maximum entailment probabilities for each generated sentence.", + "bbox": [ + 507, + 579, + 882, + 642 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Q2 (Honovich et al., 2021) combines a question generation/answering pipeline with an NLI score.", + "bbox": [ + 507, + 643, + 880, + 675 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, Honovich et al. (2022) introduce a strong ensemble of these 3 methods (Eorig). To further verify our approach, we construct a new ensemble (Eour) by replacing T5 with A11.", + "bbox": [ + 507, + 676, + 880, + 740 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "4 Results", + "text_level": 1, + "bbox": [ + 507, + 752, + 606, + 766 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1 shows the AUC scores for each metric. Our model A11 not only significantly improves over", + "bbox": [ + 507, + 778, + 882, + 810 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "2TRUE uses an earlier variant of BEGIN that is described in https://arxiv.org/pdf/2105.00071v1.pdf", + "bbox": [ + 507, + 819, + 880, + 844 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "3TRUE also has a fact-checking part, which was not included in average metric performance. We also exclude it here, as our base NLI model was trained on parts of it.", + "bbox": [ + 507, + 844, + 882, + 881 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "4The original T5 model is also pretrained on GLUE (Wang et al., 2018) and SuperGLUE (Wang et al., 2019) data, which contains additional NLI data.", + "bbox": [ + 507, + 881, + 880, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1All code is available at https://github.com/julmaxi/ with_a_little.push", + "bbox": [ + 112, + 892, + 487, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "915", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/a810c5362656a8a5f192e99dc3bd0de08207345413b7c6f8cda9841998e9c07a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodQ2SummacZST5 ANLIBase-MCA11EorigEour
Summarization
Frank85.487.890.086.789.191.187.389.491.283.185.688.084.286.6†88.985.587.7†89.889.491.293.089.791.593.2
MNBM65.668.717.768.671.374.175.577.980.271.774.677.470.173.576.671.374.577.474.076.679.473.676.479.2
SummEval75.978.881.479.481.783.978.080.583.069.672.875.872.375.2†78.173.276.1†78.880.482.985.480.383.085.3
QAGS-X65.570.976.273.178.182.979.583.888.276.981.686.577.782.286.876.381.185.480.484.888.979.483.888.0
QAGS-C79.183.587.976.380.985.277.582.186.768.774.179.373.078.4†82.973.278.0†82.983.587.791.383.186.790.3
Dialogue
BEGIN77.279.782.279.282.084.680.382.685.177.580.482.975.778.581.476.479.382.384.186.288.282.184.787.1
DialFact85.486.186.883.384.184.876.877.778.681.081.8*82.591.391.8*†x92.392.092.5*†x93.089.990.491.094.194.5x94.9
Q278.880.983.074.977.479.770.372.775.277.579.8*†82.087.288.9*†x90.387.889.4*†x90.980.882.884.986.888.5x90.1
Paraphrasing
PAWS89.189.790.387.588.288.785.786.487.187.287.8*†88.488.489.0*†89.689.490.0*†90.590.791.291.791.892.3x92.8
Avg79.780.781.780.481.482.380.681.582.478.879.880.881.782.7†83.682.283.2*†84.185.186.086.886.086.8x87.7
", + "bbox": [ + 117, + 80, + 877, + 265 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Base on six out of nine corpora, but also significantly outperforms all other competitors on average, while being more computationally efficient.", + "bbox": [ + 112, + 343, + 489, + 391 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As expected, we find the biggest gains in dialogue, where the A11 model even outperforms Eorig on 2 out of 3 corpora. We do not improve on BEGIN, which is likely due to bias in the dataset construction, which we elaborate on in Section 5.1. On the summarization part, A11 improves significantly over Base on 3 out of 5 corpora, while not significantly harming performance on any corpus. However, it still falls short of the best models in TRUE. The strong showing of T5 on these corpora suggests that this might be alleviated with a stronger base model.", + "bbox": [ + 112, + 394, + 489, + 586 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overall, a very similar behaviour is exhibited by -MC, presenting an attractive option when the added overhead of multiple samples is undesirable.", + "bbox": [ + 110, + 589, + 487, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Eour is on par with Eorig, despite massively reduced costs; it even significantly outperforms it on two dialog and the paraphrasing corpora.", + "bbox": [ + 112, + 640, + 487, + 689 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We also investigate the performance of each individual modification to our model (Table 2). They all improve average scores, while only leading to a notable decrease on BEGIN for both $e-c$ and dialogue augmentations and on MNBM for $e-c$ .", + "bbox": [ + 112, + 690, + 489, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Outside of dialogue, we find that the augmentation methods have a positive impact on PAWS, as well as all summarization corpora that are at least partially based on summaries for the CNN/DM dataset (Hermann et al., 2015) (Frank, QAGS-C, and SummEval). While we do not have a definitive explanation for this phenomenon, we hypothesize that on these datasets our augmentations aid in making the model robust in the presence of noise", + "bbox": [ + 112, + 774, + 489, + 919 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/8723d87d40c6c72534961f8eeb1eafa0627425e13c1a21b91273974cc78ebd85.jpg", + "table_caption": [ + "Table 1: AUC scores for all models on TRUE. Small numbers indicate $95\\%$ CIs computed via bootstrap. * indicates statistically significant improvement over T5; †: statistically sign. improvement over Base; $x$ : statistically sign. improvement over Eorig ( $p \\leq 0.05$ , approximate randomization test). Best non-ensemble models in bold." + ], + "table_footnote": [], + "table_body": "
Corpus+e-c+MC+Aug.
Frank-0.0+0.3+0.5+0.1+0.9+1.8+0.3+1.0+1.7
MNBM-2.1-0.8+0.5+1.4+2.1+2.9-0.4+0.0+0.6
SummEval+0.7+1.0+1.3+0.1+1.2+2.3+0.6+1.6+2.6
QAGS-X-0.4+0.3+0.9-1.5-0.2+1.1-0.3+0.9+2.1
QAGS-C+0.5+1.2+2.0-1.6-0.1+1.5+2.2+3.5+5.0
BEGIN-3.0-1.1+0.6+0.0+0.6+1.3-1.6-1.0-0.5
DialFact+8.3+9.1+9.9+1.1+1.3+1.5+3.1+3.3+3.5
Q2+5.1+6.5+7.9-0.4-0.0+0.4+3.5+4.2+5.0
PAWS+0.3+0.4+0.5+1.1+1.3+1.4+0.8+0.9+1.0
Avg+1.6+1.9+2.2+0.5+0.8+1.1+1.4+1.6+1.9
", + "bbox": [ + 522, + 340, + 870, + 478 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 2: AUC differences for individual modifications of Base. Small numbers: ${95}\\%$ CIs (bootstrap resampling).", + "bbox": [ + 507, + 488, + 882, + 517 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "or irrelevant context since our augmentations are label-neutral and must similarly be 'ignored' during training.", + "bbox": [ + 507, + 544, + 882, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "5 Analysis", + "text_level": 1, + "bbox": [ + 507, + 606, + 618, + 621 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "5.1 Effect of Dialogue Adaptation", + "text_level": 1, + "bbox": [ + 507, + 633, + 789, + 649 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We investigate whether the improvements via our augmentation approach are indeed due to them improving the handling of personal statements.", + "bbox": [ + 507, + 655, + 882, + 703 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We use the occurrences of the pronoun $I$ in a generation as a proxy measure5 and compute its correlation with human labels and metrics (see Table 3). On both Q2 and Dialfact, our proxy measure, while uncorrelated with human labels, is strongly correlated with the scores of both Base and T5. This indicates these metrics indeed tend to incorrectly reject generations with personal statements. A11 on the other hand reduces this dependency.", + "bbox": [ + 507, + 703, + 884, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our results also help explain why A11 fails to improve on BEGIN, since BEGIN gold labels are", + "bbox": [ + 507, + 848, + 882, + 881 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "5We use spacy (spacy.io) for POS tagging to identify pronouns.", + "bbox": [ + 507, + 892, + 882, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "916", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/c469ebc43607dcd002a83e393457f0b0ee8e910d2cba7126c683378bbfb53b7d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method(Begin)Q2DialFact
T5(-0.27)-0.40-0.13
Base(-0.28)-0.32-0.10
A11(-0.19)-0.190.04
Gold Label(-0.35)-0.030.05
", + "bbox": [ + 132, + 80, + 470, + 170 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b68b2418de9c7cb2e08cd7e712ac85c3d6d574b69e559527ee4fb02f463cd2d8.jpg", + "image_caption": [ + "Figure 1: Histogram of the score distributions with and without $e-c$ for faithful and non-faithful instances." + ], + "image_footnote": [], + "bbox": [ + 112, + 247, + 487, + 376 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "negatively correlated with first person pronouns. This is likely due to a bias in dataset construction: The BEGIN dataset used in TRUE has generations from two models, one of which is both more likely to generate pronouns and more likely to generate unfaithful output (see Appendix B).", + "bbox": [ + 112, + 438, + 489, + 536 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.2 Effect of integrating contradiction scores", + "text_level": 1, + "bbox": [ + 112, + 545, + 480, + 562 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To isolate the effect of $e-c$ we compare score distributions of Base and $\\text{Base} + e-c$ in Figure 1. The left-hand side of the figure shows that in Base ca. 2700 faithful instances are predicted as non-entailed (i.e., $e$ -score near 0), which implies they are labelled as contradictory or neutral. $e-c$ , on the other hand, further differentiates these instances into instances with high contradiction (negative $e-c$ score) and high neutral probability ( $e-c$ score near 0). We observe that almost all low-scoring faithful generations are classified as neutral, whereas nearly all instances that are classified as contradictory are indeed unfaithful. Where Base has no way to make use of this information, $e-c$ allows to reliably label contradictory instances as unfaithful.", + "bbox": [ + 112, + 565, + 489, + 806 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.3 Cost comparison to other approaches", + "text_level": 1, + "bbox": [ + 112, + 818, + 453, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "There is increasing awareness of the resource-hungriness of deep learning (Strubell et al., 2019). Especially for faithfulness, cheap and reliable metrics are critical, given rising demands for NLG in research and industry. Table 4 shows that our model", + "bbox": [ + 112, + 839, + 490, + 917 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/3754aea18a0759d7fb7dbf67927c53a87c0803e369bca6412e2786b90d04c23f.jpg", + "table_caption": [ + "Table 3: Kendall's $\\tau$ correlations of gold labels/system scores with first person pronoun occurrence. BEGIN shows a strong negative correlation which we attribute to model-induced dataset bias (see Appendix B)." + ], + "table_footnote": [], + "table_body": "
MethodAUC↑Param·106↓Model calls↓
SummacZS80.7355#snt×#snt
T5 ANLI81.511,0001
Q281.4220 + 355 + 355#Q × (Q1 + 2)
-MC82.73501
A1183.235015
", + "bbox": [ + 512, + 80, + 880, + 165 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/35502bb19b402effb419aed66d1b150a7d8baad5213f8d83296bbe14c13d1a6f.jpg", + "table_caption": [ + "Table 4: Performance vs. cost analysis" + ], + "table_footnote": [], + "table_body": "
Datasetw/ Five AugmentationsNo Aug. Avg.
Avg.Std.MinMax
Frank86.7-1.00.485.887.686.2
MBNM74.4-0.10.473.774.975.1
SummEval75.2-0.90.574.576.074.3
QAGS-X81.6+0.50.580.882.480.7
QAGS-C76.4-1.60.874.777.975.2
DialFact92.1-0.40.291.592.391.2
BEGIN79.6+0.30.579.080.680.9
Q288.8-0.60.388.189.286.3
PAWS89.7-0.30.189.590.089.3
Avg.82.7-0.50.282.382.982.1
", + "bbox": [ + 510, + 203, + 880, + 353 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 5: Results of our phrase selection robustness analysis. For each run, we sample five phrases, recreated our dataset and retrain our model. We repeat this process ten times and report the average, as well as the standard deviation, minimum and maximum scores of the runs. Small numbers indicate difference to the original scores. All results were computed using $e-c$ and MC dropout. For better comparison, we also report the scores of a model without any augmentation (i.e. without any additional training) with $e-c$ and MC dropout.", + "bbox": [ + 507, + 363, + 884, + 506 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "requires fewer parameters than any other metric, including a more than $30\\mathrm{x}$ reduction compared to T5. During inference our model always requires a constant number of calls which can be reduced to a single call when ablating MC dropout. On the other hand, the number of calls in SummacZS scales with the number of input and output sentences. Q2 needs to generate questions by calling an auto-regressive QG model $n$ times, where $n$ factors in the amount and length of questions $(\\# \\mathbf{Q}\\times \\mathbf{Q}1)$ , answer #Q questions with the QA model and finally check #Q answers with an NLI model $(\\# \\mathbf{Q}\\times 2)$ .", + "bbox": [ + 507, + 532, + 882, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In sum, our model compares favourably with other approaches, while also allowing for a performance/cost tradeoff by forgiving MC dropout.", + "bbox": [ + 507, + 725, + 882, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.4 Phrase Selection Robustness", + "text_level": 1, + "bbox": [ + 507, + 785, + 778, + 799 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To ensure that our augmentation is robust and not overly reliant on any particular choice of phrases, we repeat our dataset augmentation process multiple times with five randomly chosen augmentation phrases out of the original ten. We sample ten such datasets and retrain our model for each. Table 5 shows the average score, minimum and maxi", + "bbox": [ + 507, + 806, + 884, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "917", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "mum score, as well as the standard deviation of the scores. We also report results of a model with both MC dropout and $e-c$ but without any additional training and augmentations to directly quantify whether the augmentations are still helpful in their reduced form. This corresponds to applying MC dropout and $e-c$ to Base.", + "bbox": [ + 112, + 84, + 487, + 197 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As expected, we find that reducing the variety of available phrases leads to a drop in performance across almost all datasets, compared to A11. The only exception is BEGIN, where we instead see a slight improvement. This is likely to be related to the construction of BEGIN (see the discussion in Section 5.1).", + "bbox": [ + 112, + 198, + 487, + 309 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When comparing our limited augmentation models to the non-augmented model, we find that they still outperform the non-augmented model in almost all cases. In particular for Q2 and DialFact, for which we expect the strongest impact of our augmentations, we find that even the worst run still outperforms non-augmented model. This suggests that our augmentations can robustly adapt the model to the dialogue task.", + "bbox": [ + 112, + 312, + 489, + 456 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we observe a relatively large drop in scores for all datasets that are at (least partially) derived from CNN/DM (Frank, SummEval and QAGS-C). This mirrors our earlier observation in Section 4 that these datasets profit from our augmentation procedure.", + "bbox": [ + 112, + 458, + 489, + 552 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6 Related Work", + "text_level": 1, + "bbox": [ + 112, + 568, + 268, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Previous work on the utility of NLI for faithfulness led to mixed conclusions. In summarization, Falke et al. (2019) and Kryscinski et al. (2020) find out-of-the-box models have only limited utility in a faithfulness setting. In Wang et al. (2020), an NLI model is outperformed by a question generation/answering (QA/QG)-based method. In contrast, Maynez et al. (2020) find that a similar NLI model vastly outperforms a QA/QG metric on their data. In knowledge-grounded dialogue, Dziri et al. (2022), Gupta et al. (2022) and Honovich et al. (2021) find out-of-the-box models underperform.", + "bbox": [ + 112, + 596, + 489, + 788 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To improve NLI models for faithfulness in summarization, Kryscinski et al. (2020) propose FactCC, which is trained on artificially noised summaries. Utama et al. (2022) propose a controllable generation model to generate artificial faithfulness data. In knowledge-grounded dialogue, Dziri et al. (2022) and Gupta et al. (2022) combine noising techniques to generate additional training data for", + "bbox": [ + 112, + 791, + 489, + 919 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "NLI-based faithfulness models. In contrast to our work, these approaches a) generate training data from external sources, instead of directly augmenting NLI data, and b) do not explicitly focus on reconciling differences between NLI and faithfulness with their augmentation. Outside of augmentation-based approaches, Goyal and Durrett (2020) propose to train NLI models to label faithfulness at the dependency arc level.", + "bbox": [ + 507, + 84, + 884, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "7 Conclusion", + "text_level": 1, + "bbox": [ + 509, + 240, + 640, + 255 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We have demonstrated that with a small number of focused adaptations, even a relatively small NLI model can robustly predict faithfulness. We have:", + "bbox": [ + 507, + 265, + 882, + 313 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Shown that NLI-based metrics can be incompatible with task-specific requirements and identified and fixed one such incompatibility in dialogue with an augmentation strategy.", + "2. Demonstrated the importance of contradiction probability for scoring and that the underlying mechanism is the high reliability of NLI contradiction scores for detecting unfaithfulness", + "3. Shown that using Monte-Carlo dropout improves metric performance." + ], + "bbox": [ + 524, + 323, + 882, + 502 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our improved NLI model significantly improves over its baseline across many corpora and outperforms all competitors in average score on TRUE, while being much more efficient at inference.", + "bbox": [ + 507, + 512, + 882, + 576 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our work suggests that strong improvements are possible for NLI-based faithfulness metrics, by combining data augmentation with adapted NLI score computation. We hope this finding will spurn advances in cheap and robust NLI for faithfulness.", + "bbox": [ + 507, + 577, + 882, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "8 Limitations", + "text_level": 1, + "bbox": [ + 507, + 668, + 643, + 684 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Some of the summarization datasets annotated for faithfulness are relatively small, which makes score estimates uncertain. Furthermore, many datasets contain only output from a limited number of generation systems, which makes it hard to properly account for potential biases towards certain generation systems that may confound scores (see Pagnoni et al. (2021)). These concerns are, however, alleviated to some extent since we study trends across many independently created datasets, which makes it less likely for a single bias to persist in all of them. Furthermore the availability of generation and thus annotated faithfulness data limits our experiments to English. Finally, it remains", + "bbox": [ + 507, + 694, + 884, + 917 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "918", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "unclear whether our results would still provide advantages when applied to larger models such as T5-11B, whose parameter count makes experimentation infeasible on the hardware available to us.", + "bbox": [ + 112, + 84, + 489, + 149 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "9 Ethics Statement", + "text_level": 1, + "bbox": [ + 112, + 159, + 295, + 174 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Faithfulness metrics help reduce the amount of incorrect information generated by NLG systems, reducing the risk associated which such generations. However, faulty or unreliable faithfulness metrics might cause harm by incorrectly classifying faithful content as unfaithful and vice versa.", + "bbox": [ + 112, + 184, + 490, + 279 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We run all experiments on publicly available data that has been specifically constructed for faithfulness evaluation. The underlying publication has been published at a conference whose review process involved an ethics review. For a specific discussion of the human effort involved in creation of the datasets we refer the reader to the original publications.", + "bbox": [ + 112, + 280, + 489, + 409 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 114, + 435, + 213, + 450 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Samuel R. Bowman, Gabor Angeli, Christopher Potts, and Christopher D. Manning. 2015. A large annotated corpus for learning natural language inference. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 632-642, Lisbon, Portugal. Association for Computational Linguistics.", + "Yanran Chen and Steffen Eger. 2022. Menli: Robust evaluation metrics from natural language inference. arXiv preprint arXiv:2208.07316.", + "Emily Dinan, Stephen Roller, Kurt Shuster, Angela Fan, Michael Auli, and Jason Weston. 2019. Wizard of wikipedia: Knowledge-powered conversational agents. In International Conference on Learning Representations.", + "Nouha Dziri, Hannah Rashkin, Tal Linzen, and David Reitter. 2022. Evaluating Attribution in Dialogue Systems: The BEGIN Benchmark. Transactions of the Association for Computational Linguistics, 10:1066-1083. Note: TRUE uses an earlier version of the BEGIN dataset. The version used in TRUE is described in an earlier preprint at https://arxiv.org/pdf/2105.00071v1.pdf.", + "Alexander R. Fabbri, Wojciech Krysciński, Bryan McCann, Caiming Xiong, Richard Socher, and Dragomir Radev. 2021. SummEval: Re-evaluating summarization evaluation. Transactions of the Association for Computational Linguistics, 9:391-409.", + "Tobias Falke, Leonardo F. R. Ribeiro, Prasetya Ajie Utama, Ido Dagan, and Iryna Gurevych. 2019. Ranking generated summaries by correctness: An interesting but challenging application for natural language" + ], + "bbox": [ + 115, + 456, + 489, + 919 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "inference. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2214-2220, Florence, Italy. Association for Computational Linguistics.", + "Purvi Goel and Li Chen. 2021. On the robustness of monte carlo dropout trained with noisy labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 2219-2228.", + "Tanya Goyal and Greg Durrett. 2020. Evaluating factuality in generation with dependency-level entailment. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 3592-3603, Online. Association for Computational Linguistics.", + "Prakhar Gupta, Chien-Sheng Wu, Wenhao Liu, and Caiming Xiong. 2022. DialFact: A benchmark for fact-checking in dialogue. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3785-3801, Dublin, Ireland. Association for Computational Linguistics.", + "Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. 2020. Deberta: Decoding-enhanced bert with disentangled attention. In International Conference on Learning Representations.", + "Karl Moritz Hermann, Tomáš Kočisky, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa Suleyman, and Phil Blunsom. 2015. Teaching machines to read and comprehend. In Proceedings of the 28th International Conference on Neural Information Processing Systems - Volume 1, NIPS'15, page 1693-1701, Cambridge, MA, USA. MIT Press.", + "Or Honovich, Roee Aharoni, Jonathan Herzig, Hagai Taitelbaum, Doron Kukliansy, Vered Cohen, Thomas Scialom, Idan Szpektor, Avinatan Hassidim, and Yossi Matias. 2022. TRUE: Re-evaluating factual consistency evaluation. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3905-3920, Seattle, United States. Association for Computational Linguistics.", + "Or Honovich, Leshem Choshen, Roee Aharoni, Ella Neeman, Idan Szpektor, and Omri Abend. 2021. $q^2$ : Evaluating factual consistency in knowledge-grounded dialogues via question generation and question answering. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7856-7870, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.", + "Wojciech Kryscinski, Bryan McCann, Caiming Xiong, and Richard Socher. 2020. Evaluating the factual consistency of abstractive text summarization. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 9332-9346, Online. Association for Computational Linguistics." + ], + "bbox": [ + 510, + 85, + 884, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "919", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Philippe Laban, Tobias Schnabel, Paul N. Bennett, and Marti A. Hearst. 2022. SummaC: Re-visiting NLIBased models for inconsistency detection in summarization. Transactions of the Association for Computational Linguistics, 10:163-177.", + "Moritz Laurer, W v Atteveldt, Andreu Casas, and Kasper Welbers. 2022. Less annotating, more classifying-addressing the data scarcity issue of supervised machine learning with deep transfer learning and bert-nli.", + "Alisa Liu, Swabha Swayamdipta, Noah A Smith, and Yejin Choi. 2022. Wanli: Worker and ai collaboration for natural language inference dataset creation. arXiv preprint arXiv:2201.05955.", + "Joshua Maynez, Shashi Narayan, Bernd Bohnet, and Ryan McDonald. 2020. On faithfulness and factuality in abstractive summarization. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1906-1919, Online. Association for Computational Linguistics.", + "Yixin Nie, Adina Williams, Emily Dinan, Mohit Bansal, Jason Weston, and Douwe Kiela. 2020. Adversarial NLI: A new benchmark for natural language understanding. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4885-4901, Online. Association for Computational Linguistics.", + "Artidoro Pagnoni, Vidhisha Balachandran, and Yulia Tsvetkov. 2021. Understanding factuality in abstractive summarization with FRANK: A benchmark for factuality metrics. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4812-4829, Online. Association for Computational Linguistics.", + "Alicia Parrish, William Huang, Omar Agha, Soo-Hwan Lee, Nikita Nangia, Alexia Warstadt, Karmanya Aggarwal, Emily Allaway, Tal Linzen, and Samuel R. Bowman. 2021. Does putting a linguist in the loop improve NLU data collection? In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 4886-4901, Punta Cana, Dominican Republic. Association for Computational Linguistics.", + "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9.", + "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J Liu, et al. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(140):1-67.", + "Hannah Rashkin, David Reitter, Gaurav Singh Tomar, and Dipanjan Das. 2021. Increasing faithfulness in knowledge-grounded dialogue with controllable features. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics" + ], + "bbox": [ + 115, + 85, + 489, + 917 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 704-718, Online. Association for Computational Linguistics.", + "Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. 2014. Dropout: A simple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 15(56):1929-1958.", + "Emma Strubell, Ananya Ganesh, and Andrew McCallum. 2019. Energy and policy considerations for deep learning in NLP. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3645-3650, Florence, Italy. Association for Computational Linguistics.", + "James Thorne, Andreas Vlachos, Christos Christodoulopoulos, and Arpit Mittal. 2018. FEVER: a large-scale dataset for fact extraction and VERIFICATION. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 809-819, New Orleans, Louisiana. Association for Computational Linguistics.", + "Prasetya Utama, Joshua Bambrick, Nafise Moosavi, and Iryna Gurevych. 2022. Falsesum: Generating document-level NLI examples for recognizing factual inconsistency in summarization. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2763-2776, Seattle, United States. Association for Computational Linguistics.", + "Alex Wang, Kyunghyun Cho, and Mike Lewis. 2020. Asking and answering questions to evaluate the factual consistency of summaries. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5008-5020, Online. Association for Computational Linguistics.", + "Alex Wang, Yada Pruksachatkun, Nikita Nangia, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. 2019. Superglue: A stickier benchmark for general-purpose language understanding systems. Advances in neural information processing systems, 32.", + "Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. 2018. GLUE: A multi-task benchmark and analysis platform for natural language understanding. In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 353-355, Brussels, Belgium. Association for Computational Linguistics.", + "Adina Williams, Nikita Nangia, and Samuel Bowman. 2018. A broad-coverage challenge corpus for sentence understanding through inference. In Proceedings of the 2018 Conference of the North American" + ], + "bbox": [ + 510, + 85, + 882, + 917 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "920", + "bbox": [ + 485, + 928, + 515, + 939 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/06ab63d2277531f27567d7eca39e01fd8f66b688b9ec31cb55590a15c6e7cba3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Introductory Statements
Here is what I know:
yep. Also
Sure! Here is what I know:
Hedging
I am not sure, but
I am not sure but I do know that
I do not have information on this but
I think
I believe
Sentiment
I love that!
I like that!
", + "bbox": [ + 154, + 80, + 448, + 305 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1112-1122, New Orleans, Louisiana. Association for Computational Linguistics.", + "bbox": [ + 131, + 354, + 489, + 420 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online. Association for Computational Linguistics.", + "bbox": [ + 115, + 429, + 489, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Yuan Zhang, Jason Baldridge, and Luheng He. 2019. PAWS: Paraphrase adversaries from word scrambling. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 1298-1308, Minneapolis, Minnesota. Association for Computational Linguistics.", + "bbox": [ + 115, + 595, + 489, + 702 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A Augmentation Training Details", + "text_level": 1, + "bbox": [ + 114, + 712, + 420, + 728 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A.1 Augmentation Phrases", + "text_level": 1, + "bbox": [ + 114, + 737, + 344, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6 lists our manually curated list of phrases inserted during data augmentation. All phrases were derived via a small manual error analysis on the Base model.", + "bbox": [ + 112, + 758, + 487, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We broadly divide our phrases into three categories: introductory statements, hedging, and sentiment statements. For each instance in ANLI, one random phrase from the list is presupended to the hypothesis. We use all three rounds of ANLI annotations. This results in 162,865 augmented instances", + "bbox": [ + 112, + 822, + 489, + 917 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6f282655f4d8975ecd11d6244cc838b609584d1676fa0d44545c8bb9b26ea6cc.jpg", + "table_caption": [ + "Table 6: Manually curated list of dialogue phrases" + ], + "table_footnote": [], + "table_body": "
ParameterVal.
Warmup Ratio0.06
Weight Decay0.01
Effective Batch Size64
", + "bbox": [ + 583, + 80, + 808, + 149 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7: Hyperparameters", + "bbox": [ + 605, + 160, + 784, + 174 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "which, together with the original ANLI instances, leads to a total of 325,730 training instances.", + "bbox": [ + 507, + 199, + 882, + 231 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A.2 Hyperparameters", + "text_level": 1, + "bbox": [ + 509, + 241, + 700, + 256 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7 lists the hyperparameter settings for our model. We use the same optimizer hyperparameters as Laurer et al. (2022) except for an increased batch size and the learning rate. For the latter we tested three learning rates $(5e - 6, 5e - 2, 5e - 1)$ and select the one that provided the best loss on the augmented ANLI validation set. We initially ran models for 10,000 steps with a checkpoint every 1,000 steps and selected the checkpoint with the lowest loss on the augmented ANLI validation set. Later we reduced the number of training steps to 2,000 since we found we would usually select an early checkpoint as validation loss increased later in training, likely related to overfitting on the augmented data.", + "bbox": [ + 507, + 262, + 884, + 502 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A.3 Training", + "text_level": 1, + "bbox": [ + 509, + 514, + 628, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We use the DeBERTa implementation in the huggingface transformers library (Wolf et al., 2020) and trained our model on a single node using two RX6800 GPUs, with one training run taking about three hours. Later experiments with fewer steps cut that time by $80\\%$ .", + "bbox": [ + 507, + 533, + 882, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "B Dataset Bias in BEGIN", + "text_level": 1, + "bbox": [ + 509, + 643, + 746, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "BEGIN is the only dialogue corpus on which first person pronoun occurrence shows a strong (negative) correlation with faithfulness (see Table 3). Since there is nothing in the annotation guidelines that would explain this correlation, we instead hypothesize that this is the consequence of a model induced bias in the data. Specifically, we hypothesize that one of the two models in BEGIN is (1) more likely to generate personal statements and (2) less likely to generate faithful responses.", + "bbox": [ + 507, + 667, + 882, + 828 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To avoid confusion in the remainder of this section, we highlight that there are two variants of BEGIN:", + "bbox": [ + 507, + 829, + 882, + 875 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "BEGIN-v1 is the variant used in TRUE. It contains labeled generations by a fine-tuned GPT", + "bbox": [ + 509, + 887, + 882, + 917 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "921", + "bbox": [ + 485, + 928, + 512, + 940 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "2 base (Radford et al., 2019) and a fine-tuned T5 base model (Raffel et al., 2020) on the Wizard of Wikipedia dataset (Dinan et al., 2019).6", + "bbox": [ + 149, + 84, + 487, + 131 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "BEGIN-v2 is a more recent variant of BEGIN that is not part of TRUE. In addition to new instances generated by T5 and GPT-2 it contains outputs from two additional models. It also has a revised annotation procedure. When we refer to BEGIN-v2, we exclusively mean the Wizard of Wikipedia subset.", + "bbox": [ + 115, + 142, + 489, + 255 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Unfortunately, BEGIN-v1 does not allow us to retrieve which model generated which instance. This makes it impossible to directly investigate for model bias. However, BEGIN-v2 includes outputs by the same two models, fine-tuned on the same data. Since we only need corpus level statistics to verify our assumptions, we conduct our analysis on the GPT-2 and T5 instances in BEGIN-v2.", + "bbox": [ + 112, + 266, + 487, + 393 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To verify (1), we compute the correlation between a binary variable indicating which model generated each instance (T5: 0, GPT-2: 1) and first-person pronoun occurrence. We find a positive correlation (Kendall's $\\tau$ wrt. to $I$ -pronoun occurrence: $0.18, p < 0.001$ ), indicating that GPT-2 generates outputs including more first-person pronouns.", + "bbox": [ + 112, + 394, + 487, + 507 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To investigate whether GPT-2 is also more likely to be unfaithful, i.e. to verify (2), we compute the correlation between the binary model indicator variable and a faithfulness variable that is 1 when the output is labelled as Fully attributable and 0 otherwise. We find a negative correlation (Kendall's $\\tau$ wrt. to Faithfulness: $-0.25$ , $p < 0.001$ ), supporting our hypothesis that GPT-2 is also overall less faithful. To ensure that this is not an effect of additional personal statements leading to more unfaithful generations, we conduct the same analysis only on instances where we identify no first-person pronouns. We find a similarly strong negative correlation of $-0.29$ ( $p < 0.001$ ).", + "bbox": [ + 112, + 507, + 487, + 731 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our analysis shows that GPT-2 produces both overall less faithful outputs and more first-person pronouns than T5. Since BEGIN-v1 contains only outputs from T5 and GPT-2 this suggests that the root cause for the negative correlation between faithfulness label and first-person pronoun occurrence in BEGIN-v1 is model bias confounding faithfulness and first-person pronoun occurrence.", + "bbox": [ + 112, + 732, + 487, + 860 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/bbabbe41630dc254d59f33a0ce15224ff234b93320468df5f2729a95508bda9a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CorpusFaith.Non. FaithTotal
Frank223 (33.2%)448 (66.8%)671
MNBM255 (10.2%)2245 (89.8%)2500
SummEval1306 (81.6%)294 (18.4%)1600
QAGS-X116 (48.5%)123 (51.5%)239
QAGS-C113 (48.1%)122 (51.9%)235
BEGIN282 (33.7%)554 (66.3%)836
DialFact3341 (38.5%)5348 (61.5%)8689
Q2628 (57.7%)460 (42.3%)1088
PAWS3539 (44.2%)4461 (55.8%)8000
", + "bbox": [ + 522, + 80, + 870, + 204 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 8: Dataset statistics for all constituent corpora in TRUE", + "bbox": [ + 507, + 214, + 880, + 242 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "B.1 Dataset Bias in BEGIN-v2", + "text_level": 1, + "bbox": [ + 507, + 269, + 764, + 282 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We conduct a preliminary study to investigate whether similar biases also exist in BEGIN-v2.", + "bbox": [ + 507, + 290, + 880, + 319 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We observe that while BEGIN-v2 uses data from four dialogue systems, a majority of faithful generations is produced by a single system called CTRL-DIALOG (Rashkin et al., 2021). CTRL-DIALOG is specifically trained to generate less subjective text, which we hypothesize might result in fewer first person pronouns. Since CTRL-DIALOG also produces more faithful texts, this would lead to a negative correlation between faithfulness and first person pronouns, similar to what we observe on BEGIN-v1.", + "bbox": [ + 507, + 322, + 882, + 495 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We verify this assumption by computing the correlation of a binary variable indicating an instance has been generated by CTRL-DIALOG with a) the faithfulness labels on BEGIN-v2 and b) first-person pronoun occurrence. We find that an instance being generated by CTRL-DIALOG is positively correlated with it having a faithful label (Kendall $\\tau$ w.r.t. faithfulness: 0.48, $p < 0.001$ ) while being negatively correlated with the number of pronouns (Kendall $\\tau$ w.r.t. I-pronoun occurrence: -0.34, $p < 0.001$ ). This suggests future evaluations on the BEGIN-v2 might run into similar bias issues.", + "bbox": [ + 507, + 499, + 882, + 706 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "C Dataset Statistics", + "text_level": 1, + "bbox": [ + 507, + 720, + 695, + 734 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We report the number of instances, as well as the class distribution of TRUE in Table 8.", + "bbox": [ + 507, + 745, + 880, + 776 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "The relevant data can be found at https://raw. githubusercontent.com/google/BEGIN-dataset/ 5fa0cb0dde0e653d2016724a52a5ca27fe8b6a3f/dev_05_ 24_21.tsv", + "bbox": [ + 112, + 869, + 487, + 917 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "922", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "A For every submission:", + "bbox": [ + 115, + 107, + 322, + 122 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "A1. Did you describe the limitations of your work? 8", + "A2. Did you discuss any potential risks of your work? 9", + "A3. Do the abstract and introduction summarize the paper's main claims?", + "A4. Have you used AI writing assistants when working on this paper? Left blank." + ], + "bbox": [ + 129, + 126, + 695, + 285 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "B Did you use or create scientific artifacts?", + "bbox": [ + 115, + 296, + 487, + 313 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "1,3", + "bbox": [ + 134, + 319, + 159, + 332 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "B1. Did you cite the creators of artifacts you used? 1,3", + "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? 1,9", + "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? 9", + "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?" + ], + "bbox": [ + 129, + 343, + 880, + 567 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Most data is machine generated and thus unlikely to reveal personal information. All data is also already publicly available and has been introduced in peer-reviewed publications, providing an additional safeguard.", + "bbox": [ + 147, + 569, + 880, + 615 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? We discuss the limitation to English in Section 9.", + "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be." + ], + "bbox": [ + 129, + 626, + 880, + 764 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Appendix C", + "bbox": [ + 149, + 766, + 240, + 781 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "C Did you run computational experiments?", + "bbox": [ + 115, + 790, + 492, + 807 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "3,4", + "bbox": [ + 134, + 813, + 159, + 826 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?", + "bbox": [ + 129, + 837, + 880, + 869 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Appendix A", + "bbox": [ + 149, + 871, + 238, + 885 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "ACL 2023 Responsible NLP Checklist", + "bbox": [ + 132, + 84, + 433, + 99 + ], + "page_idx": 9 + }, + { + "type": "footer", + "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.", + "bbox": [ + 112, + 892, + 877, + 916 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "923", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Appendix A", + "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?", + "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? 5.2,Appendix A" + ], + "bbox": [ + 129, + 83, + 878, + 282 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank.", + "bbox": [ + 112, + 293, + 877, + 330 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response.", + "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response.", + "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response.", + "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response.", + "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response." + ], + "bbox": [ + 127, + 340, + 878, + 640 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "924", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/With a Little Push, NLI Models can Robustly and Efficiently Predict Faithfulness/d3da2b49-e96a-43e5-ae6f-e575c826018a_model.json b/2023/With a Little Push, NLI Models can Robustly and Efficiently Predict Faithfulness/d3da2b49-e96a-43e5-ae6f-e575c826018a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..43ff1b188c06602cb9f630181309faab8e21794d --- /dev/null +++ b/2023/With a Little Push, NLI Models can Robustly and Efficiently Predict Faithfulness/d3da2b49-e96a-43e5-ae6f-e575c826018a_model.json @@ -0,0 +1,2466 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.146, + 0.089, + 0.853, + 0.128 + ], + "angle": 0, + "content": "With a Little Push, NLI Models can Robustly and Efficiently Predict Faithfulness" + }, + { + "type": "text", + "bbox": [ + 0.259, + 0.143, + 0.744, + 0.159 + ], + "angle": 0, + "content": "Julius Steen Juri Opitz Anette Frank Katja Markert" + }, + { + "type": "text", + "bbox": [ + 0.332, + 0.161, + 0.671, + 0.177 + ], + "angle": 0, + "content": "Department of Computational Linguistics" + }, + { + "type": "text", + "bbox": [ + 0.411, + 0.178, + 0.593, + 0.193 + ], + "angle": 0, + "content": "Heidelberg University" + }, + { + "type": "text", + "bbox": [ + 0.385, + 0.194, + 0.617, + 0.21 + ], + "angle": 0, + "content": "69120 Heidelberg, Germany" + }, + { + "type": "text", + "bbox": [ + 0.26, + 0.211, + 0.745, + 0.227 + ], + "angle": 0, + "content": "(steen|opitz|frank|markert)@cl.uni-heidelberg.de" + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.253, + 0.343, + 0.269 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.283, + 0.461, + 0.497 + ], + "angle": 0, + "content": "Conditional language models still generate unfaithful output that is not supported by their input. These unfaithful generations jeopardize trust in real-world applications such as summarization or human-machine interaction, motivating a need for automatic faithfulness metrics. To implement such metrics, NLI models seem attractive, since they solve a strongly related task that comes with a wealth of prior research and data. But recent research suggests that NLI models require costly additional machinery to perform reliably across datasets, e.g., by running inference on a cartesian product of input and generated sentences, or supporting them with a question-generation/answering step." + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.503, + 0.462, + 0.717 + ], + "angle": 0, + "content": "In this work we show that pure NLI models can outperform more complex metrics when combining task-adaptive data augmentation with robust inference procedures. We propose: (1) Augmenting NLI training data to adapt NL inferences to the specificities of faithfulness prediction in dialogue; (2) Making use of both entailment and contradiction probabilities in NLI, and (3) Using Monte-Carlo dropout during inference. Applied to the TRUE benchmark, which combines faithfulness datasets across diverse domains and tasks, our approach strongly improves a vanilla NLI model and significantly outperforms previous work, while showing favourable computational cost." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.731, + 0.262, + 0.747 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.759, + 0.491, + 0.92 + ], + "angle": 0, + "content": "Conditional language models suffer from a tendency to hallucinate information (Maynez et al., 2020), resulting in generations that are not faithful to their input documents, which limits the trustworthiness of such models. This raises a need for automatic faithfulness metrics. In this context, models trained on natural language inference (NLI) (Bowman et al., 2015) are attractive since, intuitively, a generation being faithful implies it must be entailed by the source (Falke et al., 2019)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.254, + 0.886, + 0.495 + ], + "angle": 0, + "content": "However, pure NLI models have seen mixed success in faithfulness evaluation (Falke et al., 2019; Kryscinski et al., 2020; Wang et al., 2020; Maynez et al., 2020). While in recent evaluation on the TRUE benchmark (Honovich et al., 2022), which contains datasets from knowledge-grounded dialogue, summarization and paraphrasing, NLI-derived metrics perform best overall, they require impractically large models, or costly additional machinery such as question generation and answering models at inference, while still showing robustness issues. Thus we ask: What is still needed for pure NLI models to perform robustly across faithfulness datasets – while remaining cheap enough to serve as a lean and practical evaluation tool?" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.497, + 0.884, + 0.529 + ], + "angle": 0, + "content": "We enhance a relatively small NLI model to make it work robustly across tasks in three ways:" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.531, + 0.886, + 0.723 + ], + "angle": 0, + "content": "Task-Adaptive Data Augmentation. In NLI, a hypothesis must be fully entailed by its supporting premise. However, in faithfulness, not all parts of the generation always need to be grounded. We identify an instance of this phenomenon in dialogue where parts of a turn can fulfill communicative functions such as hedging or establishing emotional connection and are often disregarded in faithfulness annotation. Hence, when applying NLI models to complete dialogue turns that may include statements irrelevant for grounding, we run a risk of producing incorrect unfaithfulness predictions." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.725, + 0.886, + 0.821 + ], + "angle": 0, + "content": "To alleviate this issue, we propose a simple data augmentation method to adapt NLI models to genres where they need to be aware of statements that must be exempt from NLI-based faithfulness evaluation. Our approach is computationally attractive, as it avoids an increase of cost at inference time." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.824, + 0.886, + 0.92 + ], + "angle": 0, + "content": "Integration of NLI Contradiction Scores. Existing NLI faithfulness metrics typically use the entailment score for their predictions (Honovich et al., 2022; Falke et al., 2019; Kryscinski et al., 2020). However, Chen and Eger (2022) show that subtracting the contradiction score from the entail" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.928, + 0.517, + 0.941 + ], + "angle": 0, + "content": "914" + }, + { + "type": "footer", + "bbox": [ + 0.227, + 0.946, + 0.771, + 0.958 + ], + "angle": 0, + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + }, + { + "type": "footer", + "bbox": [ + 0.377, + 0.959, + 0.621, + 0.972 + ], + "angle": 0, + "content": "Volume 2: Short Papers, pages 914-924" + }, + { + "type": "footer", + "bbox": [ + 0.297, + 0.973, + 0.702, + 0.986 + ], + "angle": 0, + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.488, + 0.198 + ], + "angle": 0, + "content": "ment score (referred to as \\( e-c \\)) can improve NLI performance in certain evaluation tasks. We show that there also is a strong positive effect of \\( e-c \\) for faithfulness prediction, and demonstrate that this is due to a high contradiction probability being a more reliable predictor of unfaithfulness than low entailment probability." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.2, + 0.487, + 0.36 + ], + "angle": 0, + "content": "Monte-Carlo Dropout Inference. Applying NLI models to faithfulness prediction involves a domain shift from largely human-written data to automatically generated text. To make NLI model scores more robust under this shift, we propose to use Monte-Carlo dropout during inference (Srivastava et al., 2014). This essentially creates a cheap ensemble and has been shown to deal better with noisy labels (Goel and Chen, 2021). This approach leads to consistent score improvements in our tasks." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.362, + 0.487, + 0.427 + ], + "angle": 0, + "content": "The combination of all modifications not only strongly improves over a baseline NLI model, but also outperforms all other metrics on TRUE, on average, while being cheaper and smaller." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.445, + 0.282, + 0.459 + ], + "angle": 0, + "content": "2 Method Details" + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.476, + 0.429, + 0.492 + ], + "angle": 0, + "content": "2.1 Task-adaptive Data Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.501, + 0.489, + 0.564 + ], + "angle": 0, + "content": "To illustrate that task requirements can be incompatible between faithfulness and NLI, consider the following instance from the Q2 dialogue corpus (Honovich et al., 2021) that is labelled as faithful:" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.583, + 0.452, + 0.615 + ], + "angle": 0, + "content": "Grounding: American pancakes are similar to Scotch pancakes or drop scones." + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.616, + 0.453, + 0.647 + ], + "angle": 0, + "content": "Generation: yes, i love american pancakes, they are like scotch pancakes" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.666, + 0.487, + 0.715 + ], + "angle": 0, + "content": "From an NLI perspective, the generation is clearly not entailed, since the statement \"I love american pancakes\" is not supported by the input." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.717, + 0.489, + 0.877 + ], + "angle": 0, + "content": "To better prepare an NLI system for such genre or task-specific cases, we manually curate a small list of statements that should not influence the faithfulness prediction. We augment NLI data from the ANLI corpus (Nie et al., 2020) by adding a randomly chosen phrase from this set to each instance, while preserving the label. We then train an already fine-tuned NLI model on a concatenation of these augmented samples and original ANLI data. For training details see Appendix A." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.085, + 0.729, + 0.101 + ], + "angle": 0, + "content": "2.2 Monte-Carlo Dropout" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.106, + 0.884, + 0.187 + ], + "angle": 0, + "content": "To compute scores under Monte-Carlo dropout, we randomly sample \\( k \\) dropout masks and compute the average of the model predictions. We set \\( k = 15 \\), since preliminary experiments showed that performance did not profit from additional samples." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.199, + 0.719, + 0.216 + ], + "angle": 0, + "content": "3 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.224, + 0.884, + 0.385 + ], + "angle": 0, + "content": "We run experiments on TRUE (Honovich et al., 2022), a benchmark that compiles a wide variety of faithfulness tasks in a standardized format. It contains summarization (Pagnoni et al., 2021; Maynez et al., 2020; Wang et al., 2020; Fabbri et al., 2021), knowledge-grounded dialog (Honovich et al., 2021; Gupta et al., 2022; Dziri et al., 2022) and paraphrasing (Zhang et al., 2019) datasets. Following recommendations in TRUE, we evaluate using Area under the ROC Curve (AUC)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.386, + 0.884, + 0.529 + ], + "angle": 0, + "content": "As our BASE model, we use the DeBERTa-large (He et al., 2020) model of Laurer et al. (2022), trained on MultiNLI (Williams et al., 2018), FeverNLI (Thorne et al., 2018), ANLI (Nie et al., 2020), LingNLI (Parrish et al., 2021) and WANLI (Liu et al., 2022). The metric A11 uses all three of our proposed modifications to Base. We also investigate a variant without MC dropout inference \\((-MC)\\) as a more cost efficient alternative." + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.531, + 0.88, + 0.546 + ], + "angle": 0, + "content": "We compare to the strongest models on TRUE:" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.547, + 0.881, + 0.578 + ], + "angle": 0, + "content": "T5 ANLI (Honovich et al., 2022) is a T5-11B (Raffel et al., 2020) model trained on ANLI.4" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.58, + 0.884, + 0.643 + ], + "angle": 0, + "content": "SummacZS (Laban et al., 2022) evaluates an NLI model on all pairs of input and generated sentences and then averages maximum entailment probabilities for each generated sentence." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.644, + 0.881, + 0.676 + ], + "angle": 0, + "content": "Q2 (Honovich et al., 2021) combines a question generation/answering pipeline with an NLI score." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.677, + 0.881, + 0.741 + ], + "angle": 0, + "content": "Finally, Honovich et al. (2022) introduce a strong ensemble of these 3 methods (Eorig). To further verify our approach, we construct a new ensemble (Eour) by replacing T5 with A11." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.753, + 0.608, + 0.768 + ], + "angle": 0, + "content": "4 Results" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.78, + 0.883, + 0.811 + ], + "angle": 0, + "content": "Table 1 shows the AUC scores for each metric. Our model A11 not only significantly improves over" + }, + { + "type": "page_footnote", + "bbox": [ + 0.508, + 0.82, + 0.882, + 0.845 + ], + "angle": 0, + "content": "2TRUE uses an earlier variant of BEGIN that is described in https://arxiv.org/pdf/2105.00071v1.pdf" + }, + { + "type": "page_footnote", + "bbox": [ + 0.508, + 0.845, + 0.883, + 0.882 + ], + "angle": 0, + "content": "3TRUE also has a fact-checking part, which was not included in average metric performance. We also exclude it here, as our base NLI model was trained on parts of it." + }, + { + "type": "page_footnote", + "bbox": [ + 0.508, + 0.882, + 0.882, + 0.918 + ], + "angle": 0, + "content": "4The original T5 model is also pretrained on GLUE (Wang et al., 2018) and SuperGLUE (Wang et al., 2019) data, which contains additional NLI data." + }, + { + "type": "list", + "bbox": [ + 0.508, + 0.82, + 0.883, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.113, + 0.893, + 0.488, + 0.919 + ], + "angle": 0, + "content": "1All code is available at https://github.com/julmaxi/ with_a_little.push" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.516, + 0.941 + ], + "angle": 0, + "content": "915" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.118, + 0.082, + 0.878, + 0.266 + ], + "angle": 0, + "content": "
MethodQ2SummacZST5 ANLIBase-MCA11EorigEour
Summarization
Frank85.487.890.086.789.191.187.389.491.283.185.688.084.286.6†88.985.587.7†89.889.491.293.089.791.593.2
MNBM65.668.717.768.671.374.175.577.980.271.774.677.470.173.576.671.374.577.474.076.679.473.676.479.2
SummEval75.978.881.479.481.783.978.080.583.069.672.875.872.375.2†78.173.276.1†78.880.482.985.480.383.085.3
QAGS-X65.570.976.273.178.182.979.583.888.276.981.686.577.782.286.876.381.185.480.484.888.979.483.888.0
QAGS-C79.183.587.976.380.985.277.582.186.768.774.179.373.078.4†82.973.278.0†82.983.587.791.383.186.790.3
Dialogue
BEGIN77.279.782.279.282.084.680.382.685.177.580.482.975.778.581.476.479.382.384.186.288.282.184.787.1
DialFact85.486.186.883.384.184.876.877.778.681.081.8*82.591.391.8*†x92.392.092.5*†x93.089.990.491.094.194.5x94.9
Q278.880.983.074.977.479.770.372.775.277.579.8*†82.087.288.9*†x90.387.889.4*†x90.980.882.884.986.888.5x90.1
Paraphrasing
PAWS89.189.790.387.588.288.785.786.487.187.287.8*†88.488.489.0*†89.689.490.0*†90.590.791.291.791.892.3x92.8
Avg79.780.781.780.481.482.380.681.582.478.879.880.881.782.7†83.682.283.2*†84.185.186.086.886.086.8x87.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.275, + 0.885, + 0.319 + ], + "angle": 0, + "content": "Table 1: AUC scores for all models on TRUE. Small numbers indicate \\(95\\%\\) CIs computed via bootstrap. * indicates statistically significant improvement over T5; †: statistically sign. improvement over Base; \\(x\\): statistically sign. improvement over Eorig (\\(p \\leq 0.05\\), approximate randomization test). Best non-ensemble models in bold." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.344, + 0.49, + 0.392 + ], + "angle": 0, + "content": "Base on six out of nine corpora, but also significantly outperforms all other competitors on average, while being more computationally efficient." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.395, + 0.49, + 0.587 + ], + "angle": 0, + "content": "As expected, we find the biggest gains in dialogue, where the A11 model even outperforms Eorig on 2 out of 3 corpora. We do not improve on BEGIN, which is likely due to bias in the dataset construction, which we elaborate on in Section 5.1. On the summarization part, A11 improves significantly over Base on 3 out of 5 corpora, while not significantly harming performance on any corpus. However, it still falls short of the best models in TRUE. The strong showing of T5 on these corpora suggests that this might be alleviated with a stronger base model." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.59, + 0.489, + 0.639 + ], + "angle": 0, + "content": "Overall, a very similar behaviour is exhibited by -MC, presenting an attractive option when the added overhead of multiple samples is undesirable." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.641, + 0.489, + 0.69 + ], + "angle": 0, + "content": "Eour is on par with Eorig, despite massively reduced costs; it even significantly outperforms it on two dialog and the paraphrasing corpora." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.692, + 0.49, + 0.773 + ], + "angle": 0, + "content": "We also investigate the performance of each individual modification to our model (Table 2). They all improve average scores, while only leading to a notable decrease on BEGIN for both \\( e-c \\) and dialogue augmentations and on MNBM for \\( e-c \\)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.775, + 0.49, + 0.92 + ], + "angle": 0, + "content": "Outside of dialogue, we find that the augmentation methods have a positive impact on PAWS, as well as all summarization corpora that are at least partially based on summaries for the CNN/DM dataset (Hermann et al., 2015) (Frank, QAGS-C, and SummEval). While we do not have a definitive explanation for this phenomenon, we hypothesize that on these datasets our augmentations aid in making the model robust in the presence of noise" + }, + { + "type": "table", + "bbox": [ + 0.523, + 0.341, + 0.871, + 0.479 + ], + "angle": 0, + "content": "
Corpus+e-c+MC+Aug.
Frank-0.0+0.3+0.5+0.1+0.9+1.8+0.3+1.0+1.7
MNBM-2.1-0.8+0.5+1.4+2.1+2.9-0.4+0.0+0.6
SummEval+0.7+1.0+1.3+0.1+1.2+2.3+0.6+1.6+2.6
QAGS-X-0.4+0.3+0.9-1.5-0.2+1.1-0.3+0.9+2.1
QAGS-C+0.5+1.2+2.0-1.6-0.1+1.5+2.2+3.5+5.0
BEGIN-3.0-1.1+0.6+0.0+0.6+1.3-1.6-1.0-0.5
DialFact+8.3+9.1+9.9+1.1+1.3+1.5+3.1+3.3+3.5
Q2+5.1+6.5+7.9-0.4-0.0+0.4+3.5+4.2+5.0
PAWS+0.3+0.4+0.5+1.1+1.3+1.4+0.8+0.9+1.0
Avg+1.6+1.9+2.2+0.5+0.8+1.1+1.4+1.6+1.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.509, + 0.489, + 0.884, + 0.518 + ], + "angle": 0, + "content": "Table 2: AUC differences for individual modifications of Base. Small numbers: \\( {95}\\% \\) CIs (bootstrap resampling)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.545, + 0.884, + 0.593 + ], + "angle": 0, + "content": "or irrelevant context since our augmentations are label-neutral and must similarly be 'ignored' during training." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.607, + 0.619, + 0.623 + ], + "angle": 0, + "content": "5 Analysis" + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.634, + 0.791, + 0.65 + ], + "angle": 0, + "content": "5.1 Effect of Dialogue Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.656, + 0.884, + 0.704 + ], + "angle": 0, + "content": "We investigate whether the improvements via our augmentation approach are indeed due to them improving the handling of personal statements." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.705, + 0.885, + 0.849 + ], + "angle": 0, + "content": "We use the occurrences of the pronoun \\( I \\) in a generation as a proxy measure5 and compute its correlation with human labels and metrics (see Table 3). On both Q2 and Dialfact, our proxy measure, while uncorrelated with human labels, is strongly correlated with the scores of both Base and T5. This indicates these metrics indeed tend to incorrectly reject generations with personal statements. A11 on the other hand reduces this dependency." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.85, + 0.884, + 0.882 + ], + "angle": 0, + "content": "Our results also help explain why A11 fails to improve on BEGIN, since BEGIN gold labels are" + }, + { + "type": "page_footnote", + "bbox": [ + 0.509, + 0.893, + 0.883, + 0.919 + ], + "angle": 0, + "content": "5We use spacy (spacy.io) for POS tagging to identify pronouns." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.517, + 0.941 + ], + "angle": 0, + "content": "916" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.134, + 0.082, + 0.472, + 0.171 + ], + "angle": 0, + "content": "
Method(Begin)Q2DialFact
T5(-0.27)-0.40-0.13
Base(-0.28)-0.32-0.10
A11(-0.19)-0.190.04
Gold Label(-0.35)-0.030.05
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.18, + 0.489, + 0.238 + ], + "angle": 0, + "content": "Table 3: Kendall's \\(\\tau\\) correlations of gold labels/system scores with first person pronoun occurrence. BEGIN shows a strong negative correlation which we attribute to model-induced dataset bias (see Appendix B)." + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.248, + 0.489, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.385, + 0.489, + 0.414 + ], + "angle": 0, + "content": "Figure 1: Histogram of the score distributions with and without \\( e-c \\) for faithful and non-faithful instances." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.439, + 0.49, + 0.537 + ], + "angle": 0, + "content": "negatively correlated with first person pronouns. This is likely due to a bias in dataset construction: The BEGIN dataset used in TRUE has generations from two models, one of which is both more likely to generate pronouns and more likely to generate unfaithful output (see Appendix B)." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.546, + 0.482, + 0.563 + ], + "angle": 0, + "content": "5.2 Effect of integrating contradiction scores" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.567, + 0.49, + 0.807 + ], + "angle": 0, + "content": "To isolate the effect of \\( e-c \\) we compare score distributions of Base and \\( \\text{Base} + e-c \\) in Figure 1. The left-hand side of the figure shows that in Base ca. 2700 faithful instances are predicted as non-entailed (i.e., \\( e \\)-score near 0), which implies they are labelled as contradictory or neutral. \\( e-c \\), on the other hand, further differentiates these instances into instances with high contradiction (negative \\( e-c \\) score) and high neutral probability (\\( e-c \\) score near 0). We observe that almost all low-scoring faithful generations are classified as neutral, whereas nearly all instances that are classified as contradictory are indeed unfaithful. Where Base has no way to make use of this information, \\( e-c \\) allows to reliably label contradictory instances as unfaithful." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.819, + 0.455, + 0.835 + ], + "angle": 0, + "content": "5.3 Cost comparison to other approaches" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.84, + 0.491, + 0.919 + ], + "angle": 0, + "content": "There is increasing awareness of the resource-hungriness of deep learning (Strubell et al., 2019). Especially for faithfulness, cheap and reliable metrics are critical, given rising demands for NLG in research and industry. Table 4 shows that our model" + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.082, + 0.881, + 0.166 + ], + "angle": 0, + "content": "
MethodAUC↑Param·106↓Model calls↓
SummacZS80.7355#snt×#snt
T5 ANLI81.511,0001
Q281.4220 + 355 + 355#Q × (Q1 + 2)
-MC82.73501
A1183.235015
" + }, + { + "type": "table_caption", + "bbox": [ + 0.565, + 0.176, + 0.828, + 0.191 + ], + "angle": 0, + "content": "Table 4: Performance vs. cost analysis" + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.204, + 0.881, + 0.354 + ], + "angle": 0, + "content": "
Datasetw/ Five AugmentationsNo Aug. Avg.
Avg.Std.MinMax
Frank86.7-1.00.485.887.686.2
MBNM74.4-0.10.473.774.975.1
SummEval75.2-0.90.574.576.074.3
QAGS-X81.6+0.50.580.882.480.7
QAGS-C76.4-1.60.874.777.975.2
DialFact92.1-0.40.291.592.391.2
BEGIN79.6+0.30.579.080.680.9
Q288.8-0.60.388.189.286.3
PAWS89.7-0.30.189.590.089.3
Avg.82.7-0.50.282.382.982.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.508, + 0.364, + 0.885, + 0.507 + ], + "angle": 0, + "content": "Table 5: Results of our phrase selection robustness analysis. For each run, we sample five phrases, recreated our dataset and retrain our model. We repeat this process ten times and report the average, as well as the standard deviation, minimum and maximum scores of the runs. Small numbers indicate difference to the original scores. All results were computed using \\( e-c \\) and MC dropout. For better comparison, we also report the scores of a model without any augmentation (i.e. without any additional training) with \\( e-c \\) and MC dropout." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.533, + 0.884, + 0.725 + ], + "angle": 0, + "content": "requires fewer parameters than any other metric, including a more than \\(30\\mathrm{x}\\) reduction compared to T5. During inference our model always requires a constant number of calls which can be reduced to a single call when ablating MC dropout. On the other hand, the number of calls in SummacZS scales with the number of input and output sentences. Q2 needs to generate questions by calling an auto-regressive QG model \\(n\\) times, where \\(n\\) factors in the amount and length of questions \\((\\# \\mathbf{Q}\\times \\mathbf{Q}1)\\), answer #Q questions with the QA model and finally check #Q answers with an NLI model \\((\\# \\mathbf{Q}\\times 2)\\)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.726, + 0.884, + 0.774 + ], + "angle": 0, + "content": "In sum, our model compares favourably with other approaches, while also allowing for a performance/cost tradeoff by forgiving MC dropout." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.786, + 0.779, + 0.8 + ], + "angle": 0, + "content": "5.4 Phrase Selection Robustness" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.807, + 0.885, + 0.919 + ], + "angle": 0, + "content": "To ensure that our augmentation is robust and not overly reliant on any particular choice of phrases, we repeat our dataset augmentation process multiple times with five randomly chosen augmentation phrases out of the original ten. We sample ten such datasets and retrain our model for each. Table 5 shows the average score, minimum and maxi" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.516, + 0.941 + ], + "angle": 0, + "content": "917" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.489, + 0.198 + ], + "angle": 0, + "content": "mum score, as well as the standard deviation of the scores. We also report results of a model with both MC dropout and \\( e-c \\) but without any additional training and augmentations to directly quantify whether the augmentations are still helpful in their reduced form. This corresponds to applying MC dropout and \\( e-c \\) to Base." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.199, + 0.489, + 0.31 + ], + "angle": 0, + "content": "As expected, we find that reducing the variety of available phrases leads to a drop in performance across almost all datasets, compared to A11. The only exception is BEGIN, where we instead see a slight improvement. This is likely to be related to the construction of BEGIN (see the discussion in Section 5.1)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.313, + 0.49, + 0.457 + ], + "angle": 0, + "content": "When comparing our limited augmentation models to the non-augmented model, we find that they still outperform the non-augmented model in almost all cases. In particular for Q2 and DialFact, for which we expect the strongest impact of our augmentations, we find that even the worst run still outperforms non-augmented model. This suggests that our augmentations can robustly adapt the model to the dialogue task." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.459, + 0.49, + 0.554 + ], + "angle": 0, + "content": "Finally, we observe a relatively large drop in scores for all datasets that are at (least partially) derived from CNN/DM (Frank, SummEval and QAGS-C). This mirrors our earlier observation in Section 4 that these datasets profit from our augmentation procedure." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.569, + 0.27, + 0.584 + ], + "angle": 0, + "content": "6 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.597, + 0.49, + 0.789 + ], + "angle": 0, + "content": "Previous work on the utility of NLI for faithfulness led to mixed conclusions. In summarization, Falke et al. (2019) and Kryscinski et al. (2020) find out-of-the-box models have only limited utility in a faithfulness setting. In Wang et al. (2020), an NLI model is outperformed by a question generation/answering (QA/QG)-based method. In contrast, Maynez et al. (2020) find that a similar NLI model vastly outperforms a QA/QG metric on their data. In knowledge-grounded dialogue, Dziri et al. (2022), Gupta et al. (2022) and Honovich et al. (2021) find out-of-the-box models underperform." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.792, + 0.49, + 0.92 + ], + "angle": 0, + "content": "To improve NLI models for faithfulness in summarization, Kryscinski et al. (2020) propose FactCC, which is trained on artificially noised summaries. Utama et al. (2022) propose a controllable generation model to generate artificial faithfulness data. In knowledge-grounded dialogue, Dziri et al. (2022) and Gupta et al. (2022) combine noising techniques to generate additional training data for" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.23 + ], + "angle": 0, + "content": "NLI-based faithfulness models. In contrast to our work, these approaches a) generate training data from external sources, instead of directly augmenting NLI data, and b) do not explicitly focus on reconciling differences between NLI and faithfulness with their augmentation. Outside of augmentation-based approaches, Goyal and Durrett (2020) propose to train NLI models to label faithfulness at the dependency arc level." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.241, + 0.642, + 0.256 + ], + "angle": 0, + "content": "7 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.266, + 0.883, + 0.314 + ], + "angle": 0, + "content": "We have demonstrated that with a small number of focused adaptations, even a relatively small NLI model can robustly predict faithfulness. We have:" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.324, + 0.883, + 0.388 + ], + "angle": 0, + "content": "1. Shown that NLI-based metrics can be incompatible with task-specific requirements and identified and fixed one such incompatibility in dialogue with an augmentation strategy." + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.398, + 0.883, + 0.461 + ], + "angle": 0, + "content": "2. Demonstrated the importance of contradiction probability for scoring and that the underlying mechanism is the high reliability of NLI contradiction scores for detecting unfaithfulness" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.472, + 0.883, + 0.504 + ], + "angle": 0, + "content": "3. Shown that using Monte-Carlo dropout improves metric performance." + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.324, + 0.883, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.513, + 0.883, + 0.577 + ], + "angle": 0, + "content": "Our improved NLI model significantly improves over its baseline across many corpora and outperforms all competitors in average score on TRUE, while being much more efficient at inference." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.579, + 0.883, + 0.658 + ], + "angle": 0, + "content": "Our work suggests that strong improvements are possible for NLI-based faithfulness metrics, by combining data augmentation with adapted NLI score computation. We hope this finding will spurn advances in cheap and robust NLI for faithfulness." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.669, + 0.645, + 0.685 + ], + "angle": 0, + "content": "8 Limitations" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.695, + 0.885, + 0.919 + ], + "angle": 0, + "content": "Some of the summarization datasets annotated for faithfulness are relatively small, which makes score estimates uncertain. Furthermore, many datasets contain only output from a limited number of generation systems, which makes it hard to properly account for potential biases towards certain generation systems that may confound scores (see Pagnoni et al. (2021)). These concerns are, however, alleviated to some extent since we study trends across many independently created datasets, which makes it less likely for a single bias to persist in all of them. Furthermore the availability of generation and thus annotated faithfulness data limits our experiments to English. Finally, it remains" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.517, + 0.941 + ], + "angle": 0, + "content": "918" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.49, + 0.15 + ], + "angle": 0, + "content": "unclear whether our results would still provide advantages when applied to larger models such as T5-11B, whose parameter count makes experimentation infeasible on the hardware available to us." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.16, + 0.296, + 0.175 + ], + "angle": 0, + "content": "9 Ethics Statement" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.185, + 0.491, + 0.28 + ], + "angle": 0, + "content": "Faithfulness metrics help reduce the amount of incorrect information generated by NLG systems, reducing the risk associated which such generations. However, faulty or unreliable faithfulness metrics might cause harm by incorrectly classifying faithful content as unfaithful and vice versa." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.281, + 0.49, + 0.41 + ], + "angle": 0, + "content": "We run all experiments on publicly available data that has been specifically constructed for faithfulness evaluation. The underlying publication has been published at a conference whose review process involved an ethics review. For a specific discussion of the human effort involved in creation of the datasets we refer the reader to the original publications." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.436, + 0.214, + 0.451 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.457, + 0.49, + 0.55 + ], + "angle": 0, + "content": "Samuel R. Bowman, Gabor Angeli, Christopher Potts, and Christopher D. Manning. 2015. A large annotated corpus for learning natural language inference. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 632-642, Lisbon, Portugal. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.557, + 0.489, + 0.598 + ], + "angle": 0, + "content": "Yanran Chen and Steffen Eger. 2022. Menli: Robust evaluation metrics from natural language inference. arXiv preprint arXiv:2208.07316." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.605, + 0.488, + 0.671 + ], + "angle": 0, + "content": "Emily Dinan, Stephen Roller, Kurt Shuster, Angela Fan, Michael Auli, and Jason Weston. 2019. Wizard of wikipedia: Knowledge-powered conversational agents. In International Conference on Learning Representations." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.679, + 0.489, + 0.784 + ], + "angle": 0, + "content": "Nouha Dziri, Hannah Rashkin, Tal Linzen, and David Reitter. 2022. Evaluating Attribution in Dialogue Systems: The BEGIN Benchmark. Transactions of the Association for Computational Linguistics, 10:1066-1083. Note: TRUE uses an earlier version of the BEGIN dataset. The version used in TRUE is described in an earlier preprint at https://arxiv.org/pdf/2105.00071v1.pdf." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.792, + 0.49, + 0.858 + ], + "angle": 0, + "content": "Alexander R. Fabbri, Wojciech Krysciński, Bryan McCann, Caiming Xiong, Richard Socher, and Dragomir Radev. 2021. SummEval: Re-evaluating summarization evaluation. Transactions of the Association for Computational Linguistics, 9:391-409." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.866, + 0.49, + 0.92 + ], + "angle": 0, + "content": "Tobias Falke, Leonardo F. R. Ribeiro, Prasetya Ajie Utama, Ido Dagan, and Iryna Gurevych. 2019. Ranking generated summaries by correctness: An interesting but challenging application for natural language" + }, + { + "type": "list", + "bbox": [ + 0.116, + 0.457, + 0.49, + 0.92 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.529, + 0.086, + 0.885, + 0.14 + ], + "angle": 0, + "content": "inference. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2214-2220, Florence, Italy. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.147, + 0.885, + 0.214 + ], + "angle": 0, + "content": "Purvi Goel and Li Chen. 2021. On the robustness of monte carlo dropout trained with noisy labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 2219-2228." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.222, + 0.885, + 0.289 + ], + "angle": 0, + "content": "Tanya Goyal and Greg Durrett. 2020. Evaluating factuality in generation with dependency-level entailment. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 3592-3603, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.297, + 0.885, + 0.389 + ], + "angle": 0, + "content": "Prakhar Gupta, Chien-Sheng Wu, Wenhao Liu, and Caiming Xiong. 2022. DialFact: A benchmark for fact-checking in dialogue. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3785-3801, Dublin, Ireland. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.398, + 0.884, + 0.452 + ], + "angle": 0, + "content": "Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. 2020. Deberta: Decoding-enhanced bert with disentangled attention. In International Conference on Learning Representations." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.459, + 0.885, + 0.552 + ], + "angle": 0, + "content": "Karl Moritz Hermann, Tomáš Kočisky, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa Suleyman, and Phil Blunsom. 2015. Teaching machines to read and comprehend. In Proceedings of the 28th International Conference on Neural Information Processing Systems - Volume 1, NIPS'15, page 1693-1701, Cambridge, MA, USA. MIT Press." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.56, + 0.884, + 0.692 + ], + "angle": 0, + "content": "Or Honovich, Roee Aharoni, Jonathan Herzig, Hagai Taitelbaum, Doron Kukliansy, Vered Cohen, Thomas Scialom, Idan Szpektor, Avinatan Hassidim, and Yossi Matias. 2022. TRUE: Re-evaluating factual consistency evaluation. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3905-3920, Seattle, United States. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.7, + 0.885, + 0.818 + ], + "angle": 0, + "content": "Or Honovich, Leshem Choshen, Roee Aharoni, Ella Neeman, Idan Szpektor, and Omri Abend. 2021. \\(q^2\\): Evaluating factual consistency in knowledge-grounded dialogues via question generation and question answering. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7856-7870, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.826, + 0.885, + 0.919 + ], + "angle": 0, + "content": "Wojciech Kryscinski, Bryan McCann, Caiming Xiong, and Richard Socher. 2020. Evaluating the factual consistency of abstractive text summarization. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 9332-9346, Online. Association for Computational Linguistics." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.885, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.517, + 0.941 + ], + "angle": 0, + "content": "919" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.152 + ], + "angle": 0, + "content": "Philippe Laban, Tobias Schnabel, Paul N. Bennett, and Marti A. Hearst. 2022. SummaC: Re-visiting NLIBased models for inconsistency detection in summarization. Transactions of the Association for Computational Linguistics, 10:163-177." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.161, + 0.488, + 0.226 + ], + "angle": 0, + "content": "Moritz Laurer, W v Atteveldt, Andreu Casas, and Kasper Welbers. 2022. Less annotating, more classifying-addressing the data scarcity issue of supervised machine learning with deep transfer learning and bert-nli." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.236, + 0.487, + 0.289 + ], + "angle": 0, + "content": "Alisa Liu, Swabha Swayamdipta, Noah A Smith, and Yejin Choi. 2022. Wanli: Worker and ai collaboration for natural language inference dataset creation. arXiv preprint arXiv:2201.05955." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.298, + 0.488, + 0.377 + ], + "angle": 0, + "content": "Joshua Maynez, Shashi Narayan, Bernd Bohnet, and Ryan McDonald. 2020. On faithfulness and factuality in abstractive summarization. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1906-1919, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.386, + 0.488, + 0.478 + ], + "angle": 0, + "content": "Yixin Nie, Adina Williams, Emily Dinan, Mohit Bansal, Jason Weston, and Douwe Kiela. 2020. Adversarial NLI: A new benchmark for natural language understanding. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4885-4901, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.487, + 0.488, + 0.592 + ], + "angle": 0, + "content": "Artidoro Pagnoni, Vidhisha Balachandran, and Yulia Tsvetkov. 2021. Understanding factuality in abstractive summarization with FRANK: A benchmark for factuality metrics. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4812-4829, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.602, + 0.488, + 0.707 + ], + "angle": 0, + "content": "Alicia Parrish, William Huang, Omar Agha, Soo-Hwan Lee, Nikita Nangia, Alexia Warstadt, Karmanya Aggarwal, Emily Allaway, Tal Linzen, and Samuel R. Bowman. 2021. Does putting a linguist in the loop improve NLU data collection? In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 4886-4901, Punta Cana, Dominican Republic. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.716, + 0.488, + 0.768 + ], + "angle": 0, + "content": "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.777, + 0.488, + 0.843 + ], + "angle": 0, + "content": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J Liu, et al. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(140):1-67." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.852, + 0.488, + 0.918 + ], + "angle": 0, + "content": "Hannah Rashkin, David Reitter, Gaurav Singh Tomar, and Dipanjan Das. 2021. Increasing faithfulness in knowledge-grounded dialogue with controllable features. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics" + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.529, + 0.086, + 0.884, + 0.139 + ], + "angle": 0, + "content": "and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 704-718, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.149, + 0.884, + 0.214 + ], + "angle": 0, + "content": "Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. 2014. Dropout: A simple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 15(56):1929-1958." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.226, + 0.884, + 0.305 + ], + "angle": 0, + "content": "Emma Strubell, Ananya Ganesh, and Andrew McCallum. 2019. Energy and policy considerations for deep learning in NLP. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3645-3650, Florence, Italy. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.315, + 0.884, + 0.433 + ], + "angle": 0, + "content": "James Thorne, Andreas Vlachos, Christos Christodoulopoulos, and Arpit Mittal. 2018. FEVER: a large-scale dataset for fact extraction and VERIFICATION. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 809-819, New Orleans, Louisiana. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.443, + 0.884, + 0.561 + ], + "angle": 0, + "content": "Prasetya Utama, Joshua Bambrick, Nafise Moosavi, and Iryna Gurevych. 2022. Falsesum: Generating document-level NLI examples for recognizing factual inconsistency in summarization. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2763-2776, Seattle, United States. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.572, + 0.884, + 0.65 + ], + "angle": 0, + "content": "Alex Wang, Kyunghyun Cho, and Mike Lewis. 2020. Asking and answering questions to evaluate the factual consistency of summaries. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5008-5020, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.661, + 0.884, + 0.74 + ], + "angle": 0, + "content": "Alex Wang, Yada Pruksachatkun, Nikita Nangia, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. 2019. Superglue: A stickier benchmark for general-purpose language understanding systems. Advances in neural information processing systems, 32." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.75, + 0.884, + 0.856 + ], + "angle": 0, + "content": "Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. 2018. GLUE: A multi-task benchmark and analysis platform for natural language understanding. In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 353-355, Brussels, Belgium. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.866, + 0.884, + 0.918 + ], + "angle": 0, + "content": "Adina Williams, Nikita Nangia, and Samuel Bowman. 2018. A broad-coverage challenge corpus for sentence understanding through inference. In Proceedings of the 2018 Conference of the North American" + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.884, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.517, + 0.94 + ], + "angle": 0, + "content": "920" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.155, + 0.082, + 0.449, + 0.306 + ], + "angle": 0, + "content": "
Introductory Statements
Here is what I know:
yep. Also
Sure! Here is what I know:
Hedging
I am not sure, but
I am not sure but I do know that
I do not have information on this but
I think
I believe
Sentiment
I love that!
I like that!
" + }, + { + "type": "table_caption", + "bbox": [ + 0.131, + 0.315, + 0.471, + 0.331 + ], + "angle": 0, + "content": "Table 6: Manually curated list of dialogue phrases" + }, + { + "type": "text", + "bbox": [ + 0.132, + 0.355, + 0.49, + 0.422 + ], + "angle": 0, + "content": "Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1112-1122, New Orleans, Louisiana. Association for Computational Linguistics." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.43, + 0.49, + 0.589 + ], + "angle": 0, + "content": "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online. Association for Computational Linguistics." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.596, + 0.49, + 0.703 + ], + "angle": 0, + "content": "Yuan Zhang, Jason Baldridge, and Luheng He. 2019. PAWS: Paraphrase adversaries from word scrambling. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 1298-1308, Minneapolis, Minnesota. Association for Computational Linguistics." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.713, + 0.421, + 0.73 + ], + "angle": 0, + "content": "A Augmentation Training Details" + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.738, + 0.345, + 0.754 + ], + "angle": 0, + "content": "A.1 Augmentation Phrases" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.759, + 0.489, + 0.822 + ], + "angle": 0, + "content": "Table 6 lists our manually curated list of phrases inserted during data augmentation. All phrases were derived via a small manual error analysis on the Base model." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.823, + 0.49, + 0.919 + ], + "angle": 0, + "content": "We broadly divide our phrases into three categories: introductory statements, hedging, and sentiment statements. For each instance in ANLI, one random phrase from the list is presupended to the hypothesis. We use all three rounds of ANLI annotations. This results in 162,865 augmented instances" + }, + { + "type": "table", + "bbox": [ + 0.584, + 0.082, + 0.81, + 0.151 + ], + "angle": 0, + "content": "
ParameterVal.
Warmup Ratio0.06
Weight Decay0.01
Effective Batch Size64
" + }, + { + "type": "table_caption", + "bbox": [ + 0.606, + 0.161, + 0.786, + 0.175 + ], + "angle": 0, + "content": "Table 7: Hyperparameters" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.2, + 0.884, + 0.232 + ], + "angle": 0, + "content": "which, together with the original ANLI instances, leads to a total of 325,730 training instances." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.242, + 0.702, + 0.258 + ], + "angle": 0, + "content": "A.2 Hyperparameters" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.263, + 0.885, + 0.504 + ], + "angle": 0, + "content": "Table 7 lists the hyperparameter settings for our model. We use the same optimizer hyperparameters as Laurer et al. (2022) except for an increased batch size and the learning rate. For the latter we tested three learning rates \\((5e - 6, 5e - 2, 5e - 1)\\) and select the one that provided the best loss on the augmented ANLI validation set. We initially ran models for 10,000 steps with a checkpoint every 1,000 steps and selected the checkpoint with the lowest loss on the augmented ANLI validation set. Later we reduced the number of training steps to 2,000 since we found we would usually select an early checkpoint as validation loss increased later in training, likely related to overfitting on the augmented data." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.515, + 0.63, + 0.53 + ], + "angle": 0, + "content": "A.3 Training" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.535, + 0.884, + 0.631 + ], + "angle": 0, + "content": "We use the DeBERTa implementation in the huggingface transformers library (Wolf et al., 2020) and trained our model on a single node using two RX6800 GPUs, with one training run taking about three hours. Later experiments with fewer steps cut that time by \\(80\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.644, + 0.747, + 0.658 + ], + "angle": 0, + "content": "B Dataset Bias in BEGIN" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.668, + 0.884, + 0.829 + ], + "angle": 0, + "content": "BEGIN is the only dialogue corpus on which first person pronoun occurrence shows a strong (negative) correlation with faithfulness (see Table 3). Since there is nothing in the annotation guidelines that would explain this correlation, we instead hypothesize that this is the consequence of a model induced bias in the data. Specifically, we hypothesize that one of the two models in BEGIN is (1) more likely to generate personal statements and (2) less likely to generate faithful responses." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.83, + 0.884, + 0.876 + ], + "angle": 0, + "content": "To avoid confusion in the remainder of this section, we highlight that there are two variants of BEGIN:" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.888, + 0.884, + 0.919 + ], + "angle": 0, + "content": "BEGIN-v1 is the variant used in TRUE. It contains labeled generations by a fine-tuned GPT" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.514, + 0.941 + ], + "angle": 0, + "content": "921" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.15, + 0.085, + 0.488, + 0.133 + ], + "angle": 0, + "content": "2 base (Radford et al., 2019) and a fine-tuned T5 base model (Raffel et al., 2020) on the Wizard of Wikipedia dataset (Dinan et al., 2019).6" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.143, + 0.49, + 0.256 + ], + "angle": 0, + "content": "BEGIN-v2 is a more recent variant of BEGIN that is not part of TRUE. In addition to new instances generated by T5 and GPT-2 it contains outputs from two additional models. It also has a revised annotation procedure. When we refer to BEGIN-v2, we exclusively mean the Wizard of Wikipedia subset." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.267, + 0.489, + 0.394 + ], + "angle": 0, + "content": "Unfortunately, BEGIN-v1 does not allow us to retrieve which model generated which instance. This makes it impossible to directly investigate for model bias. However, BEGIN-v2 includes outputs by the same two models, fine-tuned on the same data. Since we only need corpus level statistics to verify our assumptions, we conduct our analysis on the GPT-2 and T5 instances in BEGIN-v2." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.395, + 0.489, + 0.508 + ], + "angle": 0, + "content": "To verify (1), we compute the correlation between a binary variable indicating which model generated each instance (T5: 0, GPT-2: 1) and first-person pronoun occurrence. We find a positive correlation (Kendall's \\(\\tau\\) wrt. to \\(I\\)-pronoun occurrence: \\(0.18, p < 0.001\\)), indicating that GPT-2 generates outputs including more first-person pronouns." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.508, + 0.489, + 0.732 + ], + "angle": 0, + "content": "To investigate whether GPT-2 is also more likely to be unfaithful, i.e. to verify (2), we compute the correlation between the binary model indicator variable and a faithfulness variable that is 1 when the output is labelled as Fully attributable and 0 otherwise. We find a negative correlation (Kendall's \\(\\tau\\) wrt. to Faithfulness: \\(-0.25\\), \\(p < 0.001\\)), supporting our hypothesis that GPT-2 is also overall less faithful. To ensure that this is not an effect of additional personal statements leading to more unfaithful generations, we conduct the same analysis only on instances where we identify no first-person pronouns. We find a similarly strong negative correlation of \\(-0.29\\) (\\(p < 0.001\\))." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.733, + 0.489, + 0.862 + ], + "angle": 0, + "content": "Our analysis shows that GPT-2 produces both overall less faithful outputs and more first-person pronouns than T5. Since BEGIN-v1 contains only outputs from T5 and GPT-2 this suggests that the root cause for the negative correlation between faithfulness label and first-person pronoun occurrence in BEGIN-v1 is model bias confounding faithfulness and first-person pronoun occurrence." + }, + { + "type": "table", + "bbox": [ + 0.523, + 0.082, + 0.871, + 0.205 + ], + "angle": 0, + "content": "
CorpusFaith.Non. FaithTotal
Frank223 (33.2%)448 (66.8%)671
MNBM255 (10.2%)2245 (89.8%)2500
SummEval1306 (81.6%)294 (18.4%)1600
QAGS-X116 (48.5%)123 (51.5%)239
QAGS-C113 (48.1%)122 (51.9%)235
BEGIN282 (33.7%)554 (66.3%)836
DialFact3341 (38.5%)5348 (61.5%)8689
Q2628 (57.7%)460 (42.3%)1088
PAWS3539 (44.2%)4461 (55.8%)8000
" + }, + { + "type": "table_caption", + "bbox": [ + 0.508, + 0.215, + 0.882, + 0.243 + ], + "angle": 0, + "content": "Table 8: Dataset statistics for all constituent corpora in TRUE" + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.27, + 0.766, + 0.284 + ], + "angle": 0, + "content": "B.1 Dataset Bias in BEGIN-v2" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.291, + 0.882, + 0.321 + ], + "angle": 0, + "content": "We conduct a preliminary study to investigate whether similar biases also exist in BEGIN-v2." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.323, + 0.884, + 0.497 + ], + "angle": 0, + "content": "We observe that while BEGIN-v2 uses data from four dialogue systems, a majority of faithful generations is produced by a single system called CTRL-DIALOG (Rashkin et al., 2021). CTRL-DIALOG is specifically trained to generate less subjective text, which we hypothesize might result in fewer first person pronouns. Since CTRL-DIALOG also produces more faithful texts, this would lead to a negative correlation between faithfulness and first person pronouns, similar to what we observe on BEGIN-v1." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.5, + 0.884, + 0.707 + ], + "angle": 0, + "content": "We verify this assumption by computing the correlation of a binary variable indicating an instance has been generated by CTRL-DIALOG with a) the faithfulness labels on BEGIN-v2 and b) first-person pronoun occurrence. We find that an instance being generated by CTRL-DIALOG is positively correlated with it having a faithful label (Kendall \\(\\tau\\) w.r.t. faithfulness: 0.48, \\(p < 0.001\\)) while being negatively correlated with the number of pronouns (Kendall \\(\\tau\\) w.r.t. I-pronoun occurrence: -0.34, \\(p < 0.001\\)). This suggests future evaluations on the BEGIN-v2 might run into similar bias issues." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.721, + 0.697, + 0.736 + ], + "angle": 0, + "content": "C Dataset Statistics" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.746, + 0.882, + 0.777 + ], + "angle": 0, + "content": "We report the number of instances, as well as the class distribution of TRUE in Table 8." + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.87, + 0.489, + 0.918 + ], + "angle": 0, + "content": "The relevant data can be found at https://raw. githubusercontent.com/google/BEGIN-dataset/ 5fa0cb0dde0e653d2016724a52a5ca27fe8b6a3f/dev_05_ 24_21.tsv" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.516, + 0.941 + ], + "angle": 0, + "content": "922" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.133, + 0.085, + 0.435, + 0.1 + ], + "angle": 0, + "content": "ACL 2023 Responsible NLP Checklist" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.108, + 0.323, + 0.123 + ], + "angle": 0, + "content": "A For every submission:" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.127, + 0.533, + 0.158 + ], + "angle": 0, + "content": "A1. Did you describe the limitations of your work? 8" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.169, + 0.553, + 0.2 + ], + "angle": 0, + "content": "A2. Did you discuss any potential risks of your work? 9" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.212, + 0.696, + 0.242 + ], + "angle": 0, + "content": "A3. Do the abstract and introduction summarize the paper's main claims?" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.255, + 0.669, + 0.286 + ], + "angle": 0, + "content": "A4. Have you used AI writing assistants when working on this paper? Left blank." + }, + { + "type": "list", + "bbox": [ + 0.13, + 0.127, + 0.696, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.297, + 0.489, + 0.314 + ], + "angle": 0, + "content": "B Did you use or create scientific artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.136, + 0.32, + 0.161, + 0.333 + ], + "angle": 0, + "content": "1,3" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.344, + 0.53, + 0.375 + ], + "angle": 0, + "content": "B1. Did you cite the creators of artifacts you used? 1,3" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.386, + 0.779, + 0.417 + ], + "angle": 0, + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? 1,9" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.429, + 0.881, + 0.507 + ], + "angle": 0, + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? 9" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.521, + 0.881, + 0.568 + ], + "angle": 0, + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?" + }, + { + "type": "list", + "bbox": [ + 0.13, + 0.344, + 0.881, + 0.568 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.57, + 0.881, + 0.617 + ], + "angle": 0, + "content": "Most data is machine generated and thus unlikely to reveal personal information. All data is also already publicly available and has been introduced in peer-reviewed publications, providing an additional safeguard." + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.627, + 0.881, + 0.675 + ], + "angle": 0, + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? We discuss the limitation to English in Section 9." + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.685, + 0.881, + 0.765 + ], + "angle": 0, + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be." + }, + { + "type": "list", + "bbox": [ + 0.131, + 0.627, + 0.881, + 0.765 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.768, + 0.241, + 0.782 + ], + "angle": 0, + "content": "Appendix C" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.791, + 0.494, + 0.808 + ], + "angle": 0, + "content": "C Did you run computational experiments?" + }, + { + "type": "text", + "bbox": [ + 0.136, + 0.814, + 0.161, + 0.827 + ], + "angle": 0, + "content": "3,4" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.838, + 0.881, + 0.87 + ], + "angle": 0, + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.872, + 0.24, + 0.887 + ], + "angle": 0, + "content": "Appendix A" + }, + { + "type": "footer", + "bbox": [ + 0.114, + 0.893, + 0.878, + 0.917 + ], + "angle": 0, + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.516, + 0.941 + ], + "angle": 0, + "content": "923" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.13, + 0.084, + 0.88, + 0.133 + ], + "angle": 0, + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Appendix A" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.143, + 0.88, + 0.205 + ], + "angle": 0, + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.218, + 0.88, + 0.283 + ], + "angle": 0, + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? 5.2,Appendix A" + }, + { + "type": "list", + "bbox": [ + 0.13, + 0.084, + 0.88, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.294, + 0.878, + 0.331 + ], + "angle": 0, + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.341, + 0.88, + 0.388 + ], + "angle": 0, + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.4, + 0.88, + 0.464 + ], + "angle": 0, + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.476, + 0.88, + 0.539 + ], + "angle": 0, + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.55, + 0.873, + 0.582 + ], + "angle": 0, + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.593, + 0.88, + 0.641 + ], + "angle": 0, + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response." + }, + { + "type": "list", + "bbox": [ + 0.128, + 0.341, + 0.88, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.516, + 0.941 + ], + "angle": 0, + "content": "924" + } + ] +] \ No newline at end of file diff --git a/2023/With a Little Push, NLI Models can Robustly and Efficiently Predict Faithfulness/d3da2b49-e96a-43e5-ae6f-e575c826018a_origin.pdf b/2023/With a Little Push, NLI Models can Robustly and Efficiently Predict Faithfulness/d3da2b49-e96a-43e5-ae6f-e575c826018a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8c287b8e00fcec01dcb3251ef0fd98b095a26db8 --- /dev/null +++ b/2023/With a Little Push, NLI Models can Robustly and Efficiently Predict Faithfulness/d3da2b49-e96a-43e5-ae6f-e575c826018a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:250856e028811e29247762742937232a5047c1b80a5de34427337d51353b3ede +size 336184 diff --git a/2023/With a Little Push, NLI Models can Robustly and Efficiently Predict Faithfulness/full.md b/2023/With a Little Push, NLI Models can Robustly and Efficiently Predict Faithfulness/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1f922de9b4c63781ca2d37fb1659a0ef4af86ef0 --- /dev/null +++ b/2023/With a Little Push, NLI Models can Robustly and Efficiently Predict Faithfulness/full.md @@ -0,0 +1,331 @@ +# With a Little Push, NLI Models can Robustly and Efficiently Predict Faithfulness + +Julius Steen Juri Opitz Anette Frank Katja Markert + +Department of Computational Linguistics + +Heidelberg University + +69120 Heidelberg, Germany + +(steen|opitz|frank|markert)@cl.uni-heidelberg.de + +# Abstract + +Conditional language models still generate unfaithful output that is not supported by their input. These unfaithful generations jeopardize trust in real-world applications such as summarization or human-machine interaction, motivating a need for automatic faithfulness metrics. To implement such metrics, NLI models seem attractive, since they solve a strongly related task that comes with a wealth of prior research and data. But recent research suggests that NLI models require costly additional machinery to perform reliably across datasets, e.g., by running inference on a cartesian product of input and generated sentences, or supporting them with a question-generation/answering step. + +In this work we show that pure NLI models can outperform more complex metrics when combining task-adaptive data augmentation with robust inference procedures. We propose: (1) Augmenting NLI training data to adapt NL inferences to the specificities of faithfulness prediction in dialogue; (2) Making use of both entailment and contradiction probabilities in NLI, and (3) Using Monte-Carlo dropout during inference. Applied to the TRUE benchmark, which combines faithfulness datasets across diverse domains and tasks, our approach strongly improves a vanilla NLI model and significantly outperforms previous work, while showing favourable computational cost. + +# 1 Introduction + +Conditional language models suffer from a tendency to hallucinate information (Maynez et al., 2020), resulting in generations that are not faithful to their input documents, which limits the trustworthiness of such models. This raises a need for automatic faithfulness metrics. In this context, models trained on natural language inference (NLI) (Bowman et al., 2015) are attractive since, intuitively, a generation being faithful implies it must be entailed by the source (Falke et al., 2019). + +However, pure NLI models have seen mixed success in faithfulness evaluation (Falke et al., 2019; Kryscinski et al., 2020; Wang et al., 2020; Maynez et al., 2020). While in recent evaluation on the TRUE benchmark (Honovich et al., 2022), which contains datasets from knowledge-grounded dialogue, summarization and paraphrasing, NLI-derived metrics perform best overall, they require impractically large models, or costly additional machinery such as question generation and answering models at inference, while still showing robustness issues. Thus we ask: What is still needed for pure NLI models to perform robustly across faithfulness datasets – while remaining cheap enough to serve as a lean and practical evaluation tool? + +We enhance a relatively small NLI model to make it work robustly across tasks in three ways: + +Task-Adaptive Data Augmentation. In NLI, a hypothesis must be fully entailed by its supporting premise. However, in faithfulness, not all parts of the generation always need to be grounded. We identify an instance of this phenomenon in dialogue where parts of a turn can fulfill communicative functions such as hedging or establishing emotional connection and are often disregarded in faithfulness annotation. Hence, when applying NLI models to complete dialogue turns that may include statements irrelevant for grounding, we run a risk of producing incorrect unfaithfulness predictions. + +To alleviate this issue, we propose a simple data augmentation method to adapt NLI models to genres where they need to be aware of statements that must be exempt from NLI-based faithfulness evaluation. Our approach is computationally attractive, as it avoids an increase of cost at inference time. + +Integration of NLI Contradiction Scores. Existing NLI faithfulness metrics typically use the entailment score for their predictions (Honovich et al., 2022; Falke et al., 2019; Kryscinski et al., 2020). However, Chen and Eger (2022) show that subtracting the contradiction score from the entail + +ment score (referred to as $e-c$ ) can improve NLI performance in certain evaluation tasks. We show that there also is a strong positive effect of $e-c$ for faithfulness prediction, and demonstrate that this is due to a high contradiction probability being a more reliable predictor of unfaithfulness than low entailment probability. + +Monte-Carlo Dropout Inference. Applying NLI models to faithfulness prediction involves a domain shift from largely human-written data to automatically generated text. To make NLI model scores more robust under this shift, we propose to use Monte-Carlo dropout during inference (Srivastava et al., 2014). This essentially creates a cheap ensemble and has been shown to deal better with noisy labels (Goel and Chen, 2021). This approach leads to consistent score improvements in our tasks. + +The combination of all modifications not only strongly improves over a baseline NLI model, but also outperforms all other metrics on TRUE, on average, while being cheaper and smaller. + +# 2 Method Details + +# 2.1 Task-adaptive Data Augmentation + +To illustrate that task requirements can be incompatible between faithfulness and NLI, consider the following instance from the Q2 dialogue corpus (Honovich et al., 2021) that is labelled as faithful: + +Grounding: American pancakes are similar to Scotch pancakes or drop scones. + +Generation: yes, i love american pancakes, they are like scotch pancakes + +From an NLI perspective, the generation is clearly not entailed, since the statement "I love american pancakes" is not supported by the input. + +To better prepare an NLI system for such genre or task-specific cases, we manually curate a small list of statements that should not influence the faithfulness prediction. We augment NLI data from the ANLI corpus (Nie et al., 2020) by adding a randomly chosen phrase from this set to each instance, while preserving the label. We then train an already fine-tuned NLI model on a concatenation of these augmented samples and original ANLI data. For training details see Appendix A. + +# 2.2 Monte-Carlo Dropout + +To compute scores under Monte-Carlo dropout, we randomly sample $k$ dropout masks and compute the average of the model predictions. We set $k = 15$ , since preliminary experiments showed that performance did not profit from additional samples. + +# 3 Experimental Setup + +We run experiments on TRUE (Honovich et al., 2022), a benchmark that compiles a wide variety of faithfulness tasks in a standardized format. It contains summarization (Pagnoni et al., 2021; Maynez et al., 2020; Wang et al., 2020; Fabbri et al., 2021), knowledge-grounded dialog (Honovich et al., 2021; Gupta et al., 2022; Dziri et al., 2022) and paraphrasing (Zhang et al., 2019) datasets. Following recommendations in TRUE, we evaluate using Area under the ROC Curve (AUC). + +As our BASE model, we use the DeBERTa-large (He et al., 2020) model of Laurer et al. (2022), trained on MultiNLI (Williams et al., 2018), FeverNLI (Thorne et al., 2018), ANLI (Nie et al., 2020), LingNLI (Parrish et al., 2021) and WANLI (Liu et al., 2022). The metric A11 uses all three of our proposed modifications to Base. We also investigate a variant without MC dropout inference $(-MC)$ as a more cost efficient alternative. + +We compare to the strongest models on TRUE: + +T5 ANLI (Honovich et al., 2022) is a T5-11B (Raffel et al., 2020) model trained on ANLI.4 + +SummacZS (Laban et al., 2022) evaluates an NLI model on all pairs of input and generated sentences and then averages maximum entailment probabilities for each generated sentence. + +Q2 (Honovich et al., 2021) combines a question generation/answering pipeline with an NLI score. + +Finally, Honovich et al. (2022) introduce a strong ensemble of these 3 methods (Eorig). To further verify our approach, we construct a new ensemble (Eour) by replacing T5 with A11. + +# 4 Results + +Table 1 shows the AUC scores for each metric. Our model A11 not only significantly improves over + +
MethodQ2SummacZST5 ANLIBase-MCA11EorigEour
Summarization
Frank85.487.890.086.789.191.187.389.491.283.185.688.084.286.6†88.985.587.7†89.889.491.293.089.791.593.2
MNBM65.668.717.768.671.374.175.577.980.271.774.677.470.173.576.671.374.577.474.076.679.473.676.479.2
SummEval75.978.881.479.481.783.978.080.583.069.672.875.872.375.2†78.173.276.1†78.880.482.985.480.383.085.3
QAGS-X65.570.976.273.178.182.979.583.888.276.981.686.577.782.286.876.381.185.480.484.888.979.483.888.0
QAGS-C79.183.587.976.380.985.277.582.186.768.774.179.373.078.4†82.973.278.0†82.983.587.791.383.186.790.3
Dialogue
BEGIN77.279.782.279.282.084.680.382.685.177.580.482.975.778.581.476.479.382.384.186.288.282.184.787.1
DialFact85.486.186.883.384.184.876.877.778.681.081.8*82.591.391.8*†x92.392.092.5*†x93.089.990.491.094.194.5x94.9
Q278.880.983.074.977.479.770.372.775.277.579.8*†82.087.288.9*†x90.387.889.4*†x90.980.882.884.986.888.5x90.1
Paraphrasing
PAWS89.189.790.387.588.288.785.786.487.187.287.8*†88.488.489.0*†89.689.490.0*†90.590.791.291.791.892.3x92.8
Avg79.780.781.780.481.482.380.681.582.478.879.880.881.782.7†83.682.283.2*†84.185.186.086.886.086.8x87.7
+ +Base on six out of nine corpora, but also significantly outperforms all other competitors on average, while being more computationally efficient. + +As expected, we find the biggest gains in dialogue, where the A11 model even outperforms Eorig on 2 out of 3 corpora. We do not improve on BEGIN, which is likely due to bias in the dataset construction, which we elaborate on in Section 5.1. On the summarization part, A11 improves significantly over Base on 3 out of 5 corpora, while not significantly harming performance on any corpus. However, it still falls short of the best models in TRUE. The strong showing of T5 on these corpora suggests that this might be alleviated with a stronger base model. + +Overall, a very similar behaviour is exhibited by -MC, presenting an attractive option when the added overhead of multiple samples is undesirable. + +Eour is on par with Eorig, despite massively reduced costs; it even significantly outperforms it on two dialog and the paraphrasing corpora. + +We also investigate the performance of each individual modification to our model (Table 2). They all improve average scores, while only leading to a notable decrease on BEGIN for both $e-c$ and dialogue augmentations and on MNBM for $e-c$ . + +Outside of dialogue, we find that the augmentation methods have a positive impact on PAWS, as well as all summarization corpora that are at least partially based on summaries for the CNN/DM dataset (Hermann et al., 2015) (Frank, QAGS-C, and SummEval). While we do not have a definitive explanation for this phenomenon, we hypothesize that on these datasets our augmentations aid in making the model robust in the presence of noise + +Table 1: AUC scores for all models on TRUE. Small numbers indicate $95\%$ CIs computed via bootstrap. * indicates statistically significant improvement over T5; †: statistically sign. improvement over Base; $x$ : statistically sign. improvement over Eorig ( $p \leq 0.05$ , approximate randomization test). Best non-ensemble models in bold. + +
Corpus+e-c+MC+Aug.
Frank-0.0+0.3+0.5+0.1+0.9+1.8+0.3+1.0+1.7
MNBM-2.1-0.8+0.5+1.4+2.1+2.9-0.4+0.0+0.6
SummEval+0.7+1.0+1.3+0.1+1.2+2.3+0.6+1.6+2.6
QAGS-X-0.4+0.3+0.9-1.5-0.2+1.1-0.3+0.9+2.1
QAGS-C+0.5+1.2+2.0-1.6-0.1+1.5+2.2+3.5+5.0
BEGIN-3.0-1.1+0.6+0.0+0.6+1.3-1.6-1.0-0.5
DialFact+8.3+9.1+9.9+1.1+1.3+1.5+3.1+3.3+3.5
Q2+5.1+6.5+7.9-0.4-0.0+0.4+3.5+4.2+5.0
PAWS+0.3+0.4+0.5+1.1+1.3+1.4+0.8+0.9+1.0
Avg+1.6+1.9+2.2+0.5+0.8+1.1+1.4+1.6+1.9
+ +Table 2: AUC differences for individual modifications of Base. Small numbers: ${95}\%$ CIs (bootstrap resampling). + +or irrelevant context since our augmentations are label-neutral and must similarly be 'ignored' during training. + +# 5 Analysis + +# 5.1 Effect of Dialogue Adaptation + +We investigate whether the improvements via our augmentation approach are indeed due to them improving the handling of personal statements. + +We use the occurrences of the pronoun $I$ in a generation as a proxy measure5 and compute its correlation with human labels and metrics (see Table 3). On both Q2 and Dialfact, our proxy measure, while uncorrelated with human labels, is strongly correlated with the scores of both Base and T5. This indicates these metrics indeed tend to incorrectly reject generations with personal statements. A11 on the other hand reduces this dependency. + +Our results also help explain why A11 fails to improve on BEGIN, since BEGIN gold labels are + +
Method(Begin)Q2DialFact
T5(-0.27)-0.40-0.13
Base(-0.28)-0.32-0.10
A11(-0.19)-0.190.04
Gold Label(-0.35)-0.030.05
+ +![](images/b68b2418de9c7cb2e08cd7e712ac85c3d6d574b69e559527ee4fb02f463cd2d8.jpg) +Figure 1: Histogram of the score distributions with and without $e-c$ for faithful and non-faithful instances. + +negatively correlated with first person pronouns. This is likely due to a bias in dataset construction: The BEGIN dataset used in TRUE has generations from two models, one of which is both more likely to generate pronouns and more likely to generate unfaithful output (see Appendix B). + +# 5.2 Effect of integrating contradiction scores + +To isolate the effect of $e-c$ we compare score distributions of Base and $\text{Base} + e-c$ in Figure 1. The left-hand side of the figure shows that in Base ca. 2700 faithful instances are predicted as non-entailed (i.e., $e$ -score near 0), which implies they are labelled as contradictory or neutral. $e-c$ , on the other hand, further differentiates these instances into instances with high contradiction (negative $e-c$ score) and high neutral probability ( $e-c$ score near 0). We observe that almost all low-scoring faithful generations are classified as neutral, whereas nearly all instances that are classified as contradictory are indeed unfaithful. Where Base has no way to make use of this information, $e-c$ allows to reliably label contradictory instances as unfaithful. + +# 5.3 Cost comparison to other approaches + +There is increasing awareness of the resource-hungriness of deep learning (Strubell et al., 2019). Especially for faithfulness, cheap and reliable metrics are critical, given rising demands for NLG in research and industry. Table 4 shows that our model + +Table 3: Kendall's $\tau$ correlations of gold labels/system scores with first person pronoun occurrence. BEGIN shows a strong negative correlation which we attribute to model-induced dataset bias (see Appendix B). + +
MethodAUC↑Param·106↓Model calls↓
SummacZS80.7355#snt×#snt
T5 ANLI81.511,0001
Q281.4220 + 355 + 355#Q × (Q1 + 2)
-MC82.73501
A1183.235015
+ +Table 4: Performance vs. cost analysis + +
Datasetw/ Five AugmentationsNo Aug. Avg.
Avg.Std.MinMax
Frank86.7-1.00.485.887.686.2
MBNM74.4-0.10.473.774.975.1
SummEval75.2-0.90.574.576.074.3
QAGS-X81.6+0.50.580.882.480.7
QAGS-C76.4-1.60.874.777.975.2
DialFact92.1-0.40.291.592.391.2
BEGIN79.6+0.30.579.080.680.9
Q288.8-0.60.388.189.286.3
PAWS89.7-0.30.189.590.089.3
Avg.82.7-0.50.282.382.982.1
+ +Table 5: Results of our phrase selection robustness analysis. For each run, we sample five phrases, recreated our dataset and retrain our model. We repeat this process ten times and report the average, as well as the standard deviation, minimum and maximum scores of the runs. Small numbers indicate difference to the original scores. All results were computed using $e-c$ and MC dropout. For better comparison, we also report the scores of a model without any augmentation (i.e. without any additional training) with $e-c$ and MC dropout. + +requires fewer parameters than any other metric, including a more than $30\mathrm{x}$ reduction compared to T5. During inference our model always requires a constant number of calls which can be reduced to a single call when ablating MC dropout. On the other hand, the number of calls in SummacZS scales with the number of input and output sentences. Q2 needs to generate questions by calling an auto-regressive QG model $n$ times, where $n$ factors in the amount and length of questions $(\# \mathbf{Q}\times \mathbf{Q}1)$ , answer #Q questions with the QA model and finally check #Q answers with an NLI model $(\# \mathbf{Q}\times 2)$ . + +In sum, our model compares favourably with other approaches, while also allowing for a performance/cost tradeoff by forgiving MC dropout. + +# 5.4 Phrase Selection Robustness + +To ensure that our augmentation is robust and not overly reliant on any particular choice of phrases, we repeat our dataset augmentation process multiple times with five randomly chosen augmentation phrases out of the original ten. We sample ten such datasets and retrain our model for each. Table 5 shows the average score, minimum and maxi + +mum score, as well as the standard deviation of the scores. We also report results of a model with both MC dropout and $e-c$ but without any additional training and augmentations to directly quantify whether the augmentations are still helpful in their reduced form. This corresponds to applying MC dropout and $e-c$ to Base. + +As expected, we find that reducing the variety of available phrases leads to a drop in performance across almost all datasets, compared to A11. The only exception is BEGIN, where we instead see a slight improvement. This is likely to be related to the construction of BEGIN (see the discussion in Section 5.1). + +When comparing our limited augmentation models to the non-augmented model, we find that they still outperform the non-augmented model in almost all cases. In particular for Q2 and DialFact, for which we expect the strongest impact of our augmentations, we find that even the worst run still outperforms non-augmented model. This suggests that our augmentations can robustly adapt the model to the dialogue task. + +Finally, we observe a relatively large drop in scores for all datasets that are at (least partially) derived from CNN/DM (Frank, SummEval and QAGS-C). This mirrors our earlier observation in Section 4 that these datasets profit from our augmentation procedure. + +# 6 Related Work + +Previous work on the utility of NLI for faithfulness led to mixed conclusions. In summarization, Falke et al. (2019) and Kryscinski et al. (2020) find out-of-the-box models have only limited utility in a faithfulness setting. In Wang et al. (2020), an NLI model is outperformed by a question generation/answering (QA/QG)-based method. In contrast, Maynez et al. (2020) find that a similar NLI model vastly outperforms a QA/QG metric on their data. In knowledge-grounded dialogue, Dziri et al. (2022), Gupta et al. (2022) and Honovich et al. (2021) find out-of-the-box models underperform. + +To improve NLI models for faithfulness in summarization, Kryscinski et al. (2020) propose FactCC, which is trained on artificially noised summaries. Utama et al. (2022) propose a controllable generation model to generate artificial faithfulness data. In knowledge-grounded dialogue, Dziri et al. (2022) and Gupta et al. (2022) combine noising techniques to generate additional training data for + +NLI-based faithfulness models. In contrast to our work, these approaches a) generate training data from external sources, instead of directly augmenting NLI data, and b) do not explicitly focus on reconciling differences between NLI and faithfulness with their augmentation. Outside of augmentation-based approaches, Goyal and Durrett (2020) propose to train NLI models to label faithfulness at the dependency arc level. + +# 7 Conclusion + +We have demonstrated that with a small number of focused adaptations, even a relatively small NLI model can robustly predict faithfulness. We have: + +1. Shown that NLI-based metrics can be incompatible with task-specific requirements and identified and fixed one such incompatibility in dialogue with an augmentation strategy. +2. Demonstrated the importance of contradiction probability for scoring and that the underlying mechanism is the high reliability of NLI contradiction scores for detecting unfaithfulness +3. Shown that using Monte-Carlo dropout improves metric performance. + +Our improved NLI model significantly improves over its baseline across many corpora and outperforms all competitors in average score on TRUE, while being much more efficient at inference. + +Our work suggests that strong improvements are possible for NLI-based faithfulness metrics, by combining data augmentation with adapted NLI score computation. We hope this finding will spurn advances in cheap and robust NLI for faithfulness. + +# 8 Limitations + +Some of the summarization datasets annotated for faithfulness are relatively small, which makes score estimates uncertain. Furthermore, many datasets contain only output from a limited number of generation systems, which makes it hard to properly account for potential biases towards certain generation systems that may confound scores (see Pagnoni et al. (2021)). These concerns are, however, alleviated to some extent since we study trends across many independently created datasets, which makes it less likely for a single bias to persist in all of them. Furthermore the availability of generation and thus annotated faithfulness data limits our experiments to English. Finally, it remains + +unclear whether our results would still provide advantages when applied to larger models such as T5-11B, whose parameter count makes experimentation infeasible on the hardware available to us. + +# 9 Ethics Statement + +Faithfulness metrics help reduce the amount of incorrect information generated by NLG systems, reducing the risk associated which such generations. However, faulty or unreliable faithfulness metrics might cause harm by incorrectly classifying faithful content as unfaithful and vice versa. + +We run all experiments on publicly available data that has been specifically constructed for faithfulness evaluation. The underlying publication has been published at a conference whose review process involved an ethics review. For a specific discussion of the human effort involved in creation of the datasets we refer the reader to the original publications. + +# References + +Samuel R. Bowman, Gabor Angeli, Christopher Potts, and Christopher D. Manning. 2015. A large annotated corpus for learning natural language inference. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 632-642, Lisbon, Portugal. Association for Computational Linguistics. +Yanran Chen and Steffen Eger. 2022. Menli: Robust evaluation metrics from natural language inference. arXiv preprint arXiv:2208.07316. +Emily Dinan, Stephen Roller, Kurt Shuster, Angela Fan, Michael Auli, and Jason Weston. 2019. Wizard of wikipedia: Knowledge-powered conversational agents. In International Conference on Learning Representations. +Nouha Dziri, Hannah Rashkin, Tal Linzen, and David Reitter. 2022. Evaluating Attribution in Dialogue Systems: The BEGIN Benchmark. Transactions of the Association for Computational Linguistics, 10:1066-1083. Note: TRUE uses an earlier version of the BEGIN dataset. The version used in TRUE is described in an earlier preprint at https://arxiv.org/pdf/2105.00071v1.pdf. +Alexander R. Fabbri, Wojciech Krysciński, Bryan McCann, Caiming Xiong, Richard Socher, and Dragomir Radev. 2021. SummEval: Re-evaluating summarization evaluation. Transactions of the Association for Computational Linguistics, 9:391-409. +Tobias Falke, Leonardo F. R. Ribeiro, Prasetya Ajie Utama, Ido Dagan, and Iryna Gurevych. 2019. Ranking generated summaries by correctness: An interesting but challenging application for natural language + +inference. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2214-2220, Florence, Italy. Association for Computational Linguistics. +Purvi Goel and Li Chen. 2021. On the robustness of monte carlo dropout trained with noisy labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 2219-2228. +Tanya Goyal and Greg Durrett. 2020. Evaluating factuality in generation with dependency-level entailment. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 3592-3603, Online. Association for Computational Linguistics. +Prakhar Gupta, Chien-Sheng Wu, Wenhao Liu, and Caiming Xiong. 2022. DialFact: A benchmark for fact-checking in dialogue. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3785-3801, Dublin, Ireland. Association for Computational Linguistics. +Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. 2020. Deberta: Decoding-enhanced bert with disentangled attention. In International Conference on Learning Representations. +Karl Moritz Hermann, Tomáš Kočisky, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa Suleyman, and Phil Blunsom. 2015. Teaching machines to read and comprehend. In Proceedings of the 28th International Conference on Neural Information Processing Systems - Volume 1, NIPS'15, page 1693-1701, Cambridge, MA, USA. MIT Press. +Or Honovich, Roee Aharoni, Jonathan Herzig, Hagai Taitelbaum, Doron Kukliansy, Vered Cohen, Thomas Scialom, Idan Szpektor, Avinatan Hassidim, and Yossi Matias. 2022. TRUE: Re-evaluating factual consistency evaluation. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3905-3920, Seattle, United States. Association for Computational Linguistics. +Or Honovich, Leshem Choshen, Roee Aharoni, Ella Neeman, Idan Szpektor, and Omri Abend. 2021. $q^2$ : Evaluating factual consistency in knowledge-grounded dialogues via question generation and question answering. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7856-7870, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics. +Wojciech Kryscinski, Bryan McCann, Caiming Xiong, and Richard Socher. 2020. Evaluating the factual consistency of abstractive text summarization. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 9332-9346, Online. Association for Computational Linguistics. + +Philippe Laban, Tobias Schnabel, Paul N. Bennett, and Marti A. Hearst. 2022. SummaC: Re-visiting NLIBased models for inconsistency detection in summarization. Transactions of the Association for Computational Linguistics, 10:163-177. +Moritz Laurer, W v Atteveldt, Andreu Casas, and Kasper Welbers. 2022. Less annotating, more classifying-addressing the data scarcity issue of supervised machine learning with deep transfer learning and bert-nli. +Alisa Liu, Swabha Swayamdipta, Noah A Smith, and Yejin Choi. 2022. Wanli: Worker and ai collaboration for natural language inference dataset creation. arXiv preprint arXiv:2201.05955. +Joshua Maynez, Shashi Narayan, Bernd Bohnet, and Ryan McDonald. 2020. On faithfulness and factuality in abstractive summarization. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1906-1919, Online. Association for Computational Linguistics. +Yixin Nie, Adina Williams, Emily Dinan, Mohit Bansal, Jason Weston, and Douwe Kiela. 2020. Adversarial NLI: A new benchmark for natural language understanding. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4885-4901, Online. Association for Computational Linguistics. +Artidoro Pagnoni, Vidhisha Balachandran, and Yulia Tsvetkov. 2021. Understanding factuality in abstractive summarization with FRANK: A benchmark for factuality metrics. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4812-4829, Online. Association for Computational Linguistics. +Alicia Parrish, William Huang, Omar Agha, Soo-Hwan Lee, Nikita Nangia, Alexia Warstadt, Karmanya Aggarwal, Emily Allaway, Tal Linzen, and Samuel R. Bowman. 2021. Does putting a linguist in the loop improve NLU data collection? In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 4886-4901, Punta Cana, Dominican Republic. Association for Computational Linguistics. +Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9. +Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J Liu, et al. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(140):1-67. +Hannah Rashkin, David Reitter, Gaurav Singh Tomar, and Dipanjan Das. 2021. Increasing faithfulness in knowledge-grounded dialogue with controllable features. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics + +and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 704-718, Online. Association for Computational Linguistics. +Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. 2014. Dropout: A simple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 15(56):1929-1958. +Emma Strubell, Ananya Ganesh, and Andrew McCallum. 2019. Energy and policy considerations for deep learning in NLP. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3645-3650, Florence, Italy. Association for Computational Linguistics. +James Thorne, Andreas Vlachos, Christos Christodoulopoulos, and Arpit Mittal. 2018. FEVER: a large-scale dataset for fact extraction and VERIFICATION. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 809-819, New Orleans, Louisiana. Association for Computational Linguistics. +Prasetya Utama, Joshua Bambrick, Nafise Moosavi, and Iryna Gurevych. 2022. Falsesum: Generating document-level NLI examples for recognizing factual inconsistency in summarization. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2763-2776, Seattle, United States. Association for Computational Linguistics. +Alex Wang, Kyunghyun Cho, and Mike Lewis. 2020. Asking and answering questions to evaluate the factual consistency of summaries. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5008-5020, Online. Association for Computational Linguistics. +Alex Wang, Yada Pruksachatkun, Nikita Nangia, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. 2019. Superglue: A stickier benchmark for general-purpose language understanding systems. Advances in neural information processing systems, 32. +Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. 2018. GLUE: A multi-task benchmark and analysis platform for natural language understanding. In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 353-355, Brussels, Belgium. Association for Computational Linguistics. +Adina Williams, Nikita Nangia, and Samuel Bowman. 2018. A broad-coverage challenge corpus for sentence understanding through inference. In Proceedings of the 2018 Conference of the North American + +
Introductory Statements
Here is what I know:
yep. Also
Sure! Here is what I know:
Hedging
I am not sure, but
I am not sure but I do know that
I do not have information on this but
I think
I believe
Sentiment
I love that!
I like that!
+ +Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1112-1122, New Orleans, Louisiana. Association for Computational Linguistics. + +Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online. Association for Computational Linguistics. + +Yuan Zhang, Jason Baldridge, and Luheng He. 2019. PAWS: Paraphrase adversaries from word scrambling. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 1298-1308, Minneapolis, Minnesota. Association for Computational Linguistics. + +# A Augmentation Training Details + +# A.1 Augmentation Phrases + +Table 6 lists our manually curated list of phrases inserted during data augmentation. All phrases were derived via a small manual error analysis on the Base model. + +We broadly divide our phrases into three categories: introductory statements, hedging, and sentiment statements. For each instance in ANLI, one random phrase from the list is presupended to the hypothesis. We use all three rounds of ANLI annotations. This results in 162,865 augmented instances + +Table 6: Manually curated list of dialogue phrases + +
ParameterVal.
Warmup Ratio0.06
Weight Decay0.01
Effective Batch Size64
+ +Table 7: Hyperparameters + +which, together with the original ANLI instances, leads to a total of 325,730 training instances. + +# A.2 Hyperparameters + +Table 7 lists the hyperparameter settings for our model. We use the same optimizer hyperparameters as Laurer et al. (2022) except for an increased batch size and the learning rate. For the latter we tested three learning rates $(5e - 6, 5e - 2, 5e - 1)$ and select the one that provided the best loss on the augmented ANLI validation set. We initially ran models for 10,000 steps with a checkpoint every 1,000 steps and selected the checkpoint with the lowest loss on the augmented ANLI validation set. Later we reduced the number of training steps to 2,000 since we found we would usually select an early checkpoint as validation loss increased later in training, likely related to overfitting on the augmented data. + +# A.3 Training + +We use the DeBERTa implementation in the huggingface transformers library (Wolf et al., 2020) and trained our model on a single node using two RX6800 GPUs, with one training run taking about three hours. Later experiments with fewer steps cut that time by $80\%$ . + +# B Dataset Bias in BEGIN + +BEGIN is the only dialogue corpus on which first person pronoun occurrence shows a strong (negative) correlation with faithfulness (see Table 3). Since there is nothing in the annotation guidelines that would explain this correlation, we instead hypothesize that this is the consequence of a model induced bias in the data. Specifically, we hypothesize that one of the two models in BEGIN is (1) more likely to generate personal statements and (2) less likely to generate faithful responses. + +To avoid confusion in the remainder of this section, we highlight that there are two variants of BEGIN: + +BEGIN-v1 is the variant used in TRUE. It contains labeled generations by a fine-tuned GPT + +2 base (Radford et al., 2019) and a fine-tuned T5 base model (Raffel et al., 2020) on the Wizard of Wikipedia dataset (Dinan et al., 2019).6 + +BEGIN-v2 is a more recent variant of BEGIN that is not part of TRUE. In addition to new instances generated by T5 and GPT-2 it contains outputs from two additional models. It also has a revised annotation procedure. When we refer to BEGIN-v2, we exclusively mean the Wizard of Wikipedia subset. + +Unfortunately, BEGIN-v1 does not allow us to retrieve which model generated which instance. This makes it impossible to directly investigate for model bias. However, BEGIN-v2 includes outputs by the same two models, fine-tuned on the same data. Since we only need corpus level statistics to verify our assumptions, we conduct our analysis on the GPT-2 and T5 instances in BEGIN-v2. + +To verify (1), we compute the correlation between a binary variable indicating which model generated each instance (T5: 0, GPT-2: 1) and first-person pronoun occurrence. We find a positive correlation (Kendall's $\tau$ wrt. to $I$ -pronoun occurrence: $0.18, p < 0.001$ ), indicating that GPT-2 generates outputs including more first-person pronouns. + +To investigate whether GPT-2 is also more likely to be unfaithful, i.e. to verify (2), we compute the correlation between the binary model indicator variable and a faithfulness variable that is 1 when the output is labelled as Fully attributable and 0 otherwise. We find a negative correlation (Kendall's $\tau$ wrt. to Faithfulness: $-0.25$ , $p < 0.001$ ), supporting our hypothesis that GPT-2 is also overall less faithful. To ensure that this is not an effect of additional personal statements leading to more unfaithful generations, we conduct the same analysis only on instances where we identify no first-person pronouns. We find a similarly strong negative correlation of $-0.29$ ( $p < 0.001$ ). + +Our analysis shows that GPT-2 produces both overall less faithful outputs and more first-person pronouns than T5. Since BEGIN-v1 contains only outputs from T5 and GPT-2 this suggests that the root cause for the negative correlation between faithfulness label and first-person pronoun occurrence in BEGIN-v1 is model bias confounding faithfulness and first-person pronoun occurrence. + +
CorpusFaith.Non. FaithTotal
Frank223 (33.2%)448 (66.8%)671
MNBM255 (10.2%)2245 (89.8%)2500
SummEval1306 (81.6%)294 (18.4%)1600
QAGS-X116 (48.5%)123 (51.5%)239
QAGS-C113 (48.1%)122 (51.9%)235
BEGIN282 (33.7%)554 (66.3%)836
DialFact3341 (38.5%)5348 (61.5%)8689
Q2628 (57.7%)460 (42.3%)1088
PAWS3539 (44.2%)4461 (55.8%)8000
+ +Table 8: Dataset statistics for all constituent corpora in TRUE + +# B.1 Dataset Bias in BEGIN-v2 + +We conduct a preliminary study to investigate whether similar biases also exist in BEGIN-v2. + +We observe that while BEGIN-v2 uses data from four dialogue systems, a majority of faithful generations is produced by a single system called CTRL-DIALOG (Rashkin et al., 2021). CTRL-DIALOG is specifically trained to generate less subjective text, which we hypothesize might result in fewer first person pronouns. Since CTRL-DIALOG also produces more faithful texts, this would lead to a negative correlation between faithfulness and first person pronouns, similar to what we observe on BEGIN-v1. + +We verify this assumption by computing the correlation of a binary variable indicating an instance has been generated by CTRL-DIALOG with a) the faithfulness labels on BEGIN-v2 and b) first-person pronoun occurrence. We find that an instance being generated by CTRL-DIALOG is positively correlated with it having a faithful label (Kendall $\tau$ w.r.t. faithfulness: 0.48, $p < 0.001$ ) while being negatively correlated with the number of pronouns (Kendall $\tau$ w.r.t. I-pronoun occurrence: -0.34, $p < 0.001$ ). This suggests future evaluations on the BEGIN-v2 might run into similar bias issues. + +# C Dataset Statistics + +We report the number of instances, as well as the class distribution of TRUE in Table 8. + +A For every submission: + +A1. Did you describe the limitations of your work? 8 +A2. Did you discuss any potential risks of your work? 9 +A3. Do the abstract and introduction summarize the paper's main claims? +A4. Have you used AI writing assistants when working on this paper? Left blank. + +B Did you use or create scientific artifacts? + +1,3 + +B1. Did you cite the creators of artifacts you used? 1,3 +B2. Did you discuss the license or terms for use and / or distribution of any artifacts? 1,9 +B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? 9 +B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? + +Most data is machine generated and thus unlikely to reveal personal information. All data is also already publicly available and has been introduced in peer-reviewed publications, providing an additional safeguard. + +B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? We discuss the limitation to English in Section 9. +B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. + +Appendix C + +C Did you run computational experiments? + +3,4 + +C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? + +Appendix A + +C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Appendix A +C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? +C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? 5.2,Appendix A + +D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank. + +D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response. +D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response. +D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response. +D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response. +D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response. \ No newline at end of file diff --git a/2023/With a Little Push, NLI Models can Robustly and Efficiently Predict Faithfulness/images.zip b/2023/With a Little Push, NLI Models can Robustly and Efficiently Predict Faithfulness/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..60d2adff584ea8949d41a3cbddfdf086c0ec37ce --- /dev/null +++ b/2023/With a Little Push, NLI Models can Robustly and Efficiently Predict Faithfulness/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a5a3fd04ce5722ac8c6a89a8125a34e67dbe813d00ed6a0375d78e10cb698974 +size 354999 diff --git a/2023/With a Little Push, NLI Models can Robustly and Efficiently Predict Faithfulness/layout.json b/2023/With a Little Push, NLI Models can Robustly and Efficiently Predict Faithfulness/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..884085d155b0a2d2db6d9005b3beec71aecdce58 --- /dev/null +++ b/2023/With a Little Push, NLI Models can Robustly and Efficiently Predict Faithfulness/layout.json @@ -0,0 +1,8246 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 86, + 74, + 507, + 107 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 74, + 507, + 107 + ], + "spans": [ + { + "bbox": [ + 86, + 74, + 507, + 107 + ], + "type": "text", + "content": "With a Little Push, NLI Models can Robustly and Efficiently Predict Faithfulness" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 154, + 120, + 442, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 120, + 442, + 133 + ], + "spans": [ + { + "bbox": [ + 154, + 120, + 442, + 133 + ], + "type": "text", + "content": "Julius Steen Juri Opitz Anette Frank Katja Markert" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 197, + 135, + 399, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 135, + 399, + 148 + ], + "spans": [ + { + "bbox": [ + 197, + 135, + 399, + 148 + ], + "type": "text", + "content": "Department of Computational Linguistics" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 244, + 149, + 352, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 149, + 352, + 162 + ], + "spans": [ + { + "bbox": [ + 244, + 149, + 352, + 162 + ], + "type": "text", + "content": "Heidelberg University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 229, + 163, + 367, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 163, + 367, + 176 + ], + "spans": [ + { + "bbox": [ + 229, + 163, + 367, + 176 + ], + "type": "text", + "content": "69120 Heidelberg, Germany" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 154, + 177, + 443, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 177, + 443, + 190 + ], + "spans": [ + { + "bbox": [ + 154, + 177, + 443, + 190 + ], + "type": "text", + "content": "(steen|opitz|frank|markert)@cl.uni-heidelberg.de" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 155, + 212, + 204, + 226 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 212, + 204, + 226 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 204, + 226 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 84, + 238, + 274, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 238, + 274, + 417 + ], + "spans": [ + { + "bbox": [ + 84, + 238, + 274, + 417 + ], + "type": "text", + "content": "Conditional language models still generate unfaithful output that is not supported by their input. These unfaithful generations jeopardize trust in real-world applications such as summarization or human-machine interaction, motivating a need for automatic faithfulness metrics. To implement such metrics, NLI models seem attractive, since they solve a strongly related task that comes with a wealth of prior research and data. But recent research suggests that NLI models require costly additional machinery to perform reliably across datasets, e.g., by running inference on a cartesian product of input and generated sentences, or supporting them with a question-generation/answering step." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 84, + 423, + 274, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 423, + 274, + 602 + ], + "spans": [ + { + "bbox": [ + 84, + 423, + 274, + 602 + ], + "type": "text", + "content": "In this work we show that pure NLI models can outperform more complex metrics when combining task-adaptive data augmentation with robust inference procedures. We propose: (1) Augmenting NLI training data to adapt NL inferences to the specificities of faithfulness prediction in dialogue; (2) Making use of both entailment and contradiction probabilities in NLI, and (3) Using Monte-Carlo dropout during inference. Applied to the TRUE benchmark, which combines faithfulness datasets across diverse domains and tasks, our approach strongly improves a vanilla NLI model and significantly outperforms previous work, while showing favourable computational cost." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 68, + 614, + 155, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 614, + 155, + 628 + ], + "spans": [ + { + "bbox": [ + 68, + 614, + 155, + 628 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 638, + 292, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 638, + 292, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 638, + 292, + 773 + ], + "type": "text", + "content": "Conditional language models suffer from a tendency to hallucinate information (Maynez et al., 2020), resulting in generations that are not faithful to their input documents, which limits the trustworthiness of such models. This raises a need for automatic faithfulness metrics. In this context, models trained on natural language inference (NLI) (Bowman et al., 2015) are attractive since, intuitively, a generation being faithful implies it must be entailed by the source (Falke et al., 2019)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 213, + 527, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 213, + 527, + 416 + ], + "spans": [ + { + "bbox": [ + 302, + 213, + 527, + 416 + ], + "type": "text", + "content": "However, pure NLI models have seen mixed success in faithfulness evaluation (Falke et al., 2019; Kryscinski et al., 2020; Wang et al., 2020; Maynez et al., 2020). While in recent evaluation on the TRUE benchmark (Honovich et al., 2022), which contains datasets from knowledge-grounded dialogue, summarization and paraphrasing, NLI-derived metrics perform best overall, they require impractically large models, or costly additional machinery such as question generation and answering models at inference, while still showing robustness issues. Thus we ask: What is still needed for pure NLI models to perform robustly across faithfulness datasets – while remaining cheap enough to serve as a lean and practical evaluation tool?" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 417, + 525, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 417, + 525, + 444 + ], + "spans": [ + { + "bbox": [ + 302, + 417, + 525, + 444 + ], + "type": "text", + "content": "We enhance a relatively small NLI model to make it work robustly across tasks in three ways:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 446, + 527, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 446, + 527, + 608 + ], + "spans": [ + { + "bbox": [ + 302, + 446, + 527, + 608 + ], + "type": "text", + "content": "Task-Adaptive Data Augmentation. In NLI, a hypothesis must be fully entailed by its supporting premise. However, in faithfulness, not all parts of the generation always need to be grounded. We identify an instance of this phenomenon in dialogue where parts of a turn can fulfill communicative functions such as hedging or establishing emotional connection and are often disregarded in faithfulness annotation. Hence, when applying NLI models to complete dialogue turns that may include statements irrelevant for grounding, we run a risk of producing incorrect unfaithfulness predictions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 609, + 527, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 609, + 527, + 690 + ], + "spans": [ + { + "bbox": [ + 302, + 609, + 527, + 690 + ], + "type": "text", + "content": "To alleviate this issue, we propose a simple data augmentation method to adapt NLI models to genres where they need to be aware of statements that must be exempt from NLI-based faithfulness evaluation. Our approach is computationally attractive, as it avoids an increase of cost at inference time." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 692, + 527, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 692, + 527, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 692, + 527, + 773 + ], + "type": "text", + "content": "Integration of NLI Contradiction Scores. Existing NLI faithfulness metrics typically use the entailment score for their predictions (Honovich et al., 2022; Falke et al., 2019; Kryscinski et al., 2020). However, Chen and Eger (2022) show that subtracting the contradiction score from the entail" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "text", + "content": "914" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "spans": [ + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "type": "text", + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 224, + 806, + 369, + 817 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 806, + 369, + 817 + ], + "spans": [ + { + "bbox": [ + 224, + 806, + 369, + 817 + ], + "type": "text", + "content": "Volume 2: Short Papers, pages 914-924" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "spans": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "text", + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 290, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 290, + 166 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 290, + 166 + ], + "type": "text", + "content": "ment score (referred to as " + }, + { + "bbox": [ + 67, + 71, + 290, + 166 + ], + "type": "inline_equation", + "content": "e-c" + }, + { + "bbox": [ + 67, + 71, + 290, + 166 + ], + "type": "text", + "content": ") can improve NLI performance in certain evaluation tasks. We show that there also is a strong positive effect of " + }, + { + "bbox": [ + 67, + 71, + 290, + 166 + ], + "type": "inline_equation", + "content": "e-c" + }, + { + "bbox": [ + 67, + 71, + 290, + 166 + ], + "type": "text", + "content": " for faithfulness prediction, and demonstrate that this is due to a high contradiction probability being a more reliable predictor of unfaithfulness than low entailment probability." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 168, + 289, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 168, + 289, + 302 + ], + "spans": [ + { + "bbox": [ + 67, + 168, + 289, + 302 + ], + "type": "text", + "content": "Monte-Carlo Dropout Inference. Applying NLI models to faithfulness prediction involves a domain shift from largely human-written data to automatically generated text. To make NLI model scores more robust under this shift, we propose to use Monte-Carlo dropout during inference (Srivastava et al., 2014). This essentially creates a cheap ensemble and has been shown to deal better with noisy labels (Goel and Chen, 2021). This approach leads to consistent score improvements in our tasks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 304, + 289, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 304, + 289, + 359 + ], + "spans": [ + { + "bbox": [ + 67, + 304, + 289, + 359 + ], + "type": "text", + "content": "The combination of all modifications not only strongly improves over a baseline NLI model, but also outperforms all other metrics on TRUE, on average, while being cheaper and smaller." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 374, + 167, + 386 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 374, + 167, + 386 + ], + "spans": [ + { + "bbox": [ + 68, + 374, + 167, + 386 + ], + "type": "text", + "content": "2 Method Details" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 400, + 255, + 413 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 400, + 255, + 413 + ], + "spans": [ + { + "bbox": [ + 68, + 400, + 255, + 413 + ], + "type": "text", + "content": "2.1 Task-adaptive Data Augmentation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 421, + 290, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 421, + 290, + 474 + ], + "spans": [ + { + "bbox": [ + 67, + 421, + 290, + 474 + ], + "type": "text", + "content": "To illustrate that task requirements can be incompatible between faithfulness and NLI, consider the following instance from the Q2 dialogue corpus (Honovich et al., 2021) that is labelled as faithful:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 89, + 490, + 268, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 490, + 268, + 517 + ], + "spans": [ + { + "bbox": [ + 89, + 490, + 268, + 517 + ], + "type": "text", + "content": "Grounding: American pancakes are similar to Scotch pancakes or drop scones." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 89, + 518, + 269, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 518, + 269, + 544 + ], + "spans": [ + { + "bbox": [ + 89, + 518, + 269, + 544 + ], + "type": "text", + "content": "Generation: yes, i love american pancakes, they are like scotch pancakes" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 560, + 289, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 560, + 289, + 601 + ], + "spans": [ + { + "bbox": [ + 67, + 560, + 289, + 601 + ], + "type": "text", + "content": "From an NLI perspective, the generation is clearly not entailed, since the statement \"I love american pancakes\" is not supported by the input." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 602, + 290, + 737 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 602, + 290, + 737 + ], + "spans": [ + { + "bbox": [ + 67, + 602, + 290, + 737 + ], + "type": "text", + "content": "To better prepare an NLI system for such genre or task-specific cases, we manually curate a small list of statements that should not influence the faithfulness prediction. We augment NLI data from the ANLI corpus (Nie et al., 2020) by adding a randomly chosen phrase from this set to each instance, while preserving the label. We then train an already fine-tuned NLI model on a concatenation of these augmented samples and original ANLI data. For training details see Appendix A." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 71, + 433, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 433, + 84 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 433, + 84 + ], + "type": "text", + "content": "2.2 Monte-Carlo Dropout" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 89, + 525, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 89, + 525, + 157 + ], + "spans": [ + { + "bbox": [ + 302, + 89, + 525, + 157 + ], + "type": "text", + "content": "To compute scores under Monte-Carlo dropout, we randomly sample " + }, + { + "bbox": [ + 302, + 89, + 525, + 157 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 302, + 89, + 525, + 157 + ], + "type": "text", + "content": " dropout masks and compute the average of the model predictions. We set " + }, + { + "bbox": [ + 302, + 89, + 525, + 157 + ], + "type": "inline_equation", + "content": "k = 15" + }, + { + "bbox": [ + 302, + 89, + 525, + 157 + ], + "type": "text", + "content": ", since preliminary experiments showed that performance did not profit from additional samples." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 167, + 427, + 181 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 167, + 427, + 181 + ], + "spans": [ + { + "bbox": [ + 302, + 167, + 427, + 181 + ], + "type": "text", + "content": "3 Experimental Setup" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 188, + 525, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 188, + 525, + 323 + ], + "spans": [ + { + "bbox": [ + 302, + 188, + 525, + 323 + ], + "type": "text", + "content": "We run experiments on TRUE (Honovich et al., 2022), a benchmark that compiles a wide variety of faithfulness tasks in a standardized format. It contains summarization (Pagnoni et al., 2021; Maynez et al., 2020; Wang et al., 2020; Fabbri et al., 2021), knowledge-grounded dialog (Honovich et al., 2021; Gupta et al., 2022; Dziri et al., 2022) and paraphrasing (Zhang et al., 2019) datasets. Following recommendations in TRUE, we evaluate using Area under the ROC Curve (AUC)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 324, + 525, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 324, + 525, + 444 + ], + "spans": [ + { + "bbox": [ + 302, + 324, + 525, + 444 + ], + "type": "text", + "content": "As our BASE model, we use the DeBERTa-large (He et al., 2020) model of Laurer et al. (2022), trained on MultiNLI (Williams et al., 2018), FeverNLI (Thorne et al., 2018), ANLI (Nie et al., 2020), LingNLI (Parrish et al., 2021) and WANLI (Liu et al., 2022). The metric A11 uses all three of our proposed modifications to Base. We also investigate a variant without MC dropout inference " + }, + { + "bbox": [ + 302, + 324, + 525, + 444 + ], + "type": "inline_equation", + "content": "(-MC)" + }, + { + "bbox": [ + 302, + 324, + 525, + 444 + ], + "type": "text", + "content": " as a more cost efficient alternative." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 446, + 523, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 446, + 523, + 459 + ], + "spans": [ + { + "bbox": [ + 314, + 446, + 523, + 459 + ], + "type": "text", + "content": "We compare to the strongest models on TRUE:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 460, + 524, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 460, + 524, + 486 + ], + "spans": [ + { + "bbox": [ + 302, + 460, + 524, + 486 + ], + "type": "text", + "content": "T5 ANLI (Honovich et al., 2022) is a T5-11B (Raffel et al., 2020) model trained on ANLI.4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 302, + 487, + 525, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 487, + 525, + 540 + ], + "spans": [ + { + "bbox": [ + 302, + 487, + 525, + 540 + ], + "type": "text", + "content": "SummacZS (Laban et al., 2022) evaluates an NLI model on all pairs of input and generated sentences and then averages maximum entailment probabilities for each generated sentence." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 302, + 541, + 524, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 541, + 524, + 568 + ], + "spans": [ + { + "bbox": [ + 302, + 541, + 524, + 568 + ], + "type": "text", + "content": "Q2 (Honovich et al., 2021) combines a question generation/answering pipeline with an NLI score." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 302, + 569, + 524, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 569, + 524, + 623 + ], + "spans": [ + { + "bbox": [ + 302, + 569, + 524, + 623 + ], + "type": "text", + "content": "Finally, Honovich et al. (2022) introduce a strong ensemble of these 3 methods (Eorig). To further verify our approach, we construct a new ensemble (Eour) by replacing T5 with A11." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 302, + 633, + 361, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 633, + 361, + 645 + ], + "spans": [ + { + "bbox": [ + 302, + 633, + 361, + 645 + ], + "type": "text", + "content": "4 Results" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 302, + 655, + 525, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 655, + 525, + 682 + ], + "spans": [ + { + "bbox": [ + 302, + 655, + 525, + 682 + ], + "type": "text", + "content": "Table 1 shows the AUC scores for each metric. Our model A11 not only significantly improves over" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 689, + 524, + 710 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 689, + 524, + 710 + ], + "spans": [ + { + "bbox": [ + 302, + 689, + 524, + 710 + ], + "type": "text", + "content": "2TRUE uses an earlier variant of BEGIN that is described in https://arxiv.org/pdf/2105.00071v1.pdf" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 302, + 710, + 525, + 741 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 710, + 525, + 741 + ], + "spans": [ + { + "bbox": [ + 302, + 710, + 525, + 741 + ], + "type": "text", + "content": "3TRUE also has a fact-checking part, which was not included in average metric performance. We also exclude it here, as our base NLI model was trained on parts of it." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 302, + 741, + 524, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 741, + 524, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 524, + 772 + ], + "type": "text", + "content": "4The original T5 model is also pretrained on GLUE (Wang et al., 2018) and SuperGLUE (Wang et al., 2019) data, which contains additional NLI data." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 67, + 751, + 290, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 751, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 751, + 290, + 772 + ], + "type": "text", + "content": "1All code is available at https://github.com/julmaxi/ with_a_little.push" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "915" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 68, + 522, + 223 + ], + "blocks": [ + { + "bbox": [ + 70, + 68, + 522, + 223 + ], + "lines": [ + { + "bbox": [ + 70, + 68, + 522, + 223 + ], + "spans": [ + { + "bbox": [ + 70, + 68, + 522, + 223 + ], + "type": "table", + "html": "
MethodQ2SummacZST5 ANLIBase-MCA11EorigEour
Summarization
Frank85.487.890.086.789.191.187.389.491.283.185.688.084.286.6†88.985.587.7†89.889.491.293.089.791.593.2
MNBM65.668.717.768.671.374.175.577.980.271.774.677.470.173.576.671.374.577.474.076.679.473.676.479.2
SummEval75.978.881.479.481.783.978.080.583.069.672.875.872.375.2†78.173.276.1†78.880.482.985.480.383.085.3
QAGS-X65.570.976.273.178.182.979.583.888.276.981.686.577.782.286.876.381.185.480.484.888.979.483.888.0
QAGS-C79.183.587.976.380.985.277.582.186.768.774.179.373.078.4†82.973.278.0†82.983.587.791.383.186.790.3
Dialogue
BEGIN77.279.782.279.282.084.680.382.685.177.580.482.975.778.581.476.479.382.384.186.288.282.184.787.1
DialFact85.486.186.883.384.184.876.877.778.681.081.8*82.591.391.8*†x92.392.092.5*†x93.089.990.491.094.194.5x94.9
Q278.880.983.074.977.479.770.372.775.277.579.8*†82.087.288.9*†x90.387.889.4*†x90.980.882.884.986.888.5x90.1
Paraphrasing
PAWS89.189.790.387.588.288.785.786.487.187.287.8*†88.488.489.0*†89.689.490.0*†90.590.791.291.791.892.3x92.8
Avg79.780.781.780.481.482.380.681.582.478.879.880.881.782.7†83.682.283.2*†84.185.186.086.886.086.8x87.7
", + "image_path": "a810c5362656a8a5f192e99dc3bd0de08207345413b7c6f8cda9841998e9c07a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 289, + 291, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 289, + 291, + 329 + ], + "spans": [ + { + "bbox": [ + 67, + 289, + 291, + 329 + ], + "type": "text", + "content": "Base on six out of nine corpora, but also significantly outperforms all other competitors on average, while being more computationally efficient." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 332, + 291, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 332, + 291, + 493 + ], + "spans": [ + { + "bbox": [ + 67, + 332, + 291, + 493 + ], + "type": "text", + "content": "As expected, we find the biggest gains in dialogue, where the A11 model even outperforms Eorig on 2 out of 3 corpora. We do not improve on BEGIN, which is likely due to bias in the dataset construction, which we elaborate on in Section 5.1. On the summarization part, A11 improves significantly over Base on 3 out of 5 corpora, while not significantly harming performance on any corpus. However, it still falls short of the best models in TRUE. The strong showing of T5 on these corpora suggests that this might be alleviated with a stronger base model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 496, + 290, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 496, + 290, + 537 + ], + "spans": [ + { + "bbox": [ + 66, + 496, + 290, + 537 + ], + "type": "text", + "content": "Overall, a very similar behaviour is exhibited by -MC, presenting an attractive option when the added overhead of multiple samples is undesirable." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 539, + 290, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 539, + 290, + 580 + ], + "spans": [ + { + "bbox": [ + 67, + 539, + 290, + 580 + ], + "type": "text", + "content": "Eour is on par with Eorig, despite massively reduced costs; it even significantly outperforms it on two dialog and the paraphrasing corpora." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 581, + 291, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 581, + 291, + 650 + ], + "spans": [ + { + "bbox": [ + 67, + 581, + 291, + 650 + ], + "type": "text", + "content": "We also investigate the performance of each individual modification to our model (Table 2). They all improve average scores, while only leading to a notable decrease on BEGIN for both " + }, + { + "bbox": [ + 67, + 581, + 291, + 650 + ], + "type": "inline_equation", + "content": "e-c" + }, + { + "bbox": [ + 67, + 581, + 291, + 650 + ], + "type": "text", + "content": " and dialogue augmentations and on MNBM for " + }, + { + "bbox": [ + 67, + 581, + 291, + 650 + ], + "type": "inline_equation", + "content": "e-c" + }, + { + "bbox": [ + 67, + 581, + 291, + 650 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 651, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 651, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 651, + 291, + 773 + ], + "type": "text", + "content": "Outside of dialogue, we find that the augmentation methods have a positive impact on PAWS, as well as all summarization corpora that are at least partially based on summaries for the CNN/DM dataset (Hermann et al., 2015) (Frank, QAGS-C, and SummEval). While we do not have a definitive explanation for this phenomenon, we hypothesize that on these datasets our augmentations aid in making the model robust in the presence of noise" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 311, + 286, + 518, + 402 + ], + "blocks": [ + { + "bbox": [ + 67, + 231, + 526, + 268 + ], + "lines": [ + { + "bbox": [ + 67, + 231, + 526, + 268 + ], + "spans": [ + { + "bbox": [ + 67, + 231, + 526, + 268 + ], + "type": "text", + "content": "Table 1: AUC scores for all models on TRUE. Small numbers indicate " + }, + { + "bbox": [ + 67, + 231, + 526, + 268 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 67, + 231, + 526, + 268 + ], + "type": "text", + "content": " CIs computed via bootstrap. * indicates statistically significant improvement over T5; †: statistically sign. improvement over Base; " + }, + { + "bbox": [ + 67, + 231, + 526, + 268 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 67, + 231, + 526, + 268 + ], + "type": "text", + "content": ": statistically sign. improvement over Eorig (" + }, + { + "bbox": [ + 67, + 231, + 526, + 268 + ], + "type": "inline_equation", + "content": "p \\leq 0.05" + }, + { + "bbox": [ + 67, + 231, + 526, + 268 + ], + "type": "text", + "content": ", approximate randomization test). Best non-ensemble models in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 311, + 286, + 518, + 402 + ], + "lines": [ + { + "bbox": [ + 311, + 286, + 518, + 402 + ], + "spans": [ + { + "bbox": [ + 311, + 286, + 518, + 402 + ], + "type": "table", + "html": "
Corpus+e-c+MC+Aug.
Frank-0.0+0.3+0.5+0.1+0.9+1.8+0.3+1.0+1.7
MNBM-2.1-0.8+0.5+1.4+2.1+2.9-0.4+0.0+0.6
SummEval+0.7+1.0+1.3+0.1+1.2+2.3+0.6+1.6+2.6
QAGS-X-0.4+0.3+0.9-1.5-0.2+1.1-0.3+0.9+2.1
QAGS-C+0.5+1.2+2.0-1.6-0.1+1.5+2.2+3.5+5.0
BEGIN-3.0-1.1+0.6+0.0+0.6+1.3-1.6-1.0-0.5
DialFact+8.3+9.1+9.9+1.1+1.3+1.5+3.1+3.3+3.5
Q2+5.1+6.5+7.9-0.4-0.0+0.4+3.5+4.2+5.0
PAWS+0.3+0.4+0.5+1.1+1.3+1.4+0.8+0.9+1.0
Avg+1.6+1.9+2.2+0.5+0.8+1.1+1.4+1.6+1.9
", + "image_path": "8723d87d40c6c72534961f8eeb1eafa0627425e13c1a21b91273974cc78ebd85.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 411, + 525, + 435 + ], + "lines": [ + { + "bbox": [ + 302, + 411, + 525, + 435 + ], + "spans": [ + { + "bbox": [ + 302, + 411, + 525, + 435 + ], + "type": "text", + "content": "Table 2: AUC differences for individual modifications of Base. Small numbers: " + }, + { + "bbox": [ + 302, + 411, + 525, + 435 + ], + "type": "inline_equation", + "content": "{95}\\%" + }, + { + "bbox": [ + 302, + 411, + 525, + 435 + ], + "type": "text", + "content": " CIs (bootstrap resampling)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 302, + 458, + 525, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 458, + 525, + 498 + ], + "spans": [ + { + "bbox": [ + 302, + 458, + 525, + 498 + ], + "type": "text", + "content": "or irrelevant context since our augmentations are label-neutral and must similarly be 'ignored' during training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 510, + 368, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 510, + 368, + 523 + ], + "spans": [ + { + "bbox": [ + 302, + 510, + 368, + 523 + ], + "type": "text", + "content": "5 Analysis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 533, + 470, + 546 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 533, + 470, + 546 + ], + "spans": [ + { + "bbox": [ + 302, + 533, + 470, + 546 + ], + "type": "text", + "content": "5.1 Effect of Dialogue Adaptation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 551, + 525, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 551, + 525, + 592 + ], + "spans": [ + { + "bbox": [ + 302, + 551, + 525, + 592 + ], + "type": "text", + "content": "We investigate whether the improvements via our augmentation approach are indeed due to them improving the handling of personal statements." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 592, + 526, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 592, + 526, + 714 + ], + "spans": [ + { + "bbox": [ + 302, + 592, + 526, + 714 + ], + "type": "text", + "content": "We use the occurrences of the pronoun " + }, + { + "bbox": [ + 302, + 592, + 526, + 714 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 302, + 592, + 526, + 714 + ], + "type": "text", + "content": " in a generation as a proxy measure5 and compute its correlation with human labels and metrics (see Table 3). On both Q2 and Dialfact, our proxy measure, while uncorrelated with human labels, is strongly correlated with the scores of both Base and T5. This indicates these metrics indeed tend to incorrectly reject generations with personal statements. A11 on the other hand reduces this dependency." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 714, + 525, + 741 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 714, + 525, + 741 + ], + "spans": [ + { + "bbox": [ + 302, + 714, + 525, + 741 + ], + "type": "text", + "content": "Our results also help explain why A11 fails to improve on BEGIN, since BEGIN gold labels are" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 751, + 525, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 751, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 525, + 772 + ], + "type": "text", + "content": "5We use spacy (spacy.io) for POS tagging to identify pronouns." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "916" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 79, + 68, + 280, + 143 + ], + "blocks": [ + { + "bbox": [ + 79, + 68, + 280, + 143 + ], + "lines": [ + { + "bbox": [ + 79, + 68, + 280, + 143 + ], + "spans": [ + { + "bbox": [ + 79, + 68, + 280, + 143 + ], + "type": "table", + "html": "
Method(Begin)Q2DialFact
T5(-0.27)-0.40-0.13
Base(-0.28)-0.32-0.10
A11(-0.19)-0.190.04
Gold Label(-0.35)-0.030.05
", + "image_path": "c469ebc43607dcd002a83e393457f0b0ee8e910d2cba7126c683378bbfb53b7d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 67, + 208, + 290, + 317 + ], + "blocks": [ + { + "bbox": [ + 67, + 208, + 290, + 317 + ], + "lines": [ + { + "bbox": [ + 67, + 208, + 290, + 317 + ], + "spans": [ + { + "bbox": [ + 67, + 208, + 290, + 317 + ], + "type": "image", + "image_path": "b68b2418de9c7cb2e08cd7e712ac85c3d6d574b69e559527ee4fb02f463cd2d8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 323, + 290, + 348 + ], + "lines": [ + { + "bbox": [ + 67, + 323, + 290, + 348 + ], + "spans": [ + { + "bbox": [ + 67, + 323, + 290, + 348 + ], + "type": "text", + "content": "Figure 1: Histogram of the score distributions with and without " + }, + { + "bbox": [ + 67, + 323, + 290, + 348 + ], + "type": "inline_equation", + "content": "e-c" + }, + { + "bbox": [ + 67, + 323, + 290, + 348 + ], + "type": "text", + "content": " for faithful and non-faithful instances." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 369, + 291, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 369, + 291, + 451 + ], + "spans": [ + { + "bbox": [ + 67, + 369, + 291, + 451 + ], + "type": "text", + "content": "negatively correlated with first person pronouns. This is likely due to a bias in dataset construction: The BEGIN dataset used in TRUE has generations from two models, one of which is both more likely to generate pronouns and more likely to generate unfaithful output (see Appendix B)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 459, + 286, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 459, + 286, + 473 + ], + "spans": [ + { + "bbox": [ + 67, + 459, + 286, + 473 + ], + "type": "text", + "content": "5.2 Effect of integrating contradiction scores" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 476, + 291, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 476, + 291, + 678 + ], + "spans": [ + { + "bbox": [ + 67, + 476, + 291, + 678 + ], + "type": "text", + "content": "To isolate the effect of " + }, + { + "bbox": [ + 67, + 476, + 291, + 678 + ], + "type": "inline_equation", + "content": "e-c" + }, + { + "bbox": [ + 67, + 476, + 291, + 678 + ], + "type": "text", + "content": " we compare score distributions of Base and " + }, + { + "bbox": [ + 67, + 476, + 291, + 678 + ], + "type": "inline_equation", + "content": "\\text{Base} + e-c" + }, + { + "bbox": [ + 67, + 476, + 291, + 678 + ], + "type": "text", + "content": " in Figure 1. The left-hand side of the figure shows that in Base ca. 2700 faithful instances are predicted as non-entailed (i.e., " + }, + { + "bbox": [ + 67, + 476, + 291, + 678 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 67, + 476, + 291, + 678 + ], + "type": "text", + "content": "-score near 0), which implies they are labelled as contradictory or neutral. " + }, + { + "bbox": [ + 67, + 476, + 291, + 678 + ], + "type": "inline_equation", + "content": "e-c" + }, + { + "bbox": [ + 67, + 476, + 291, + 678 + ], + "type": "text", + "content": ", on the other hand, further differentiates these instances into instances with high contradiction (negative " + }, + { + "bbox": [ + 67, + 476, + 291, + 678 + ], + "type": "inline_equation", + "content": "e-c" + }, + { + "bbox": [ + 67, + 476, + 291, + 678 + ], + "type": "text", + "content": " score) and high neutral probability (" + }, + { + "bbox": [ + 67, + 476, + 291, + 678 + ], + "type": "inline_equation", + "content": "e-c" + }, + { + "bbox": [ + 67, + 476, + 291, + 678 + ], + "type": "text", + "content": " score near 0). We observe that almost all low-scoring faithful generations are classified as neutral, whereas nearly all instances that are classified as contradictory are indeed unfaithful. Where Base has no way to make use of this information, " + }, + { + "bbox": [ + 67, + 476, + 291, + 678 + ], + "type": "inline_equation", + "content": "e-c" + }, + { + "bbox": [ + 67, + 476, + 291, + 678 + ], + "type": "text", + "content": " allows to reliably label contradictory instances as unfaithful." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 688, + 270, + 702 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 688, + 270, + 702 + ], + "spans": [ + { + "bbox": [ + 67, + 688, + 270, + 702 + ], + "type": "text", + "content": "5.3 Cost comparison to other approaches" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 706, + 292, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 706, + 292, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 706, + 292, + 772 + ], + "type": "text", + "content": "There is increasing awareness of the resource-hungriness of deep learning (Strubell et al., 2019). Especially for faithfulness, cheap and reliable metrics are critical, given rising demands for NLG in research and industry. Table 4 shows that our model" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 305, + 68, + 524, + 139 + ], + "blocks": [ + { + "bbox": [ + 67, + 151, + 290, + 200 + ], + "lines": [ + { + "bbox": [ + 67, + 151, + 290, + 200 + ], + "spans": [ + { + "bbox": [ + 67, + 151, + 290, + 200 + ], + "type": "text", + "content": "Table 3: Kendall's " + }, + { + "bbox": [ + 67, + 151, + 290, + 200 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 67, + 151, + 290, + 200 + ], + "type": "text", + "content": " correlations of gold labels/system scores with first person pronoun occurrence. BEGIN shows a strong negative correlation which we attribute to model-induced dataset bias (see Appendix B)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 305, + 68, + 524, + 139 + ], + "lines": [ + { + "bbox": [ + 305, + 68, + 524, + 139 + ], + "spans": [ + { + "bbox": [ + 305, + 68, + 524, + 139 + ], + "type": "table", + "html": "
MethodAUC↑Param·106↓Model calls↓
SummacZS80.7355#snt×#snt
T5 ANLI81.511,0001
Q281.4220 + 355 + 355#Q × (Q1 + 2)
-MC82.73501
A1183.235015
", + "image_path": "3754aea18a0759d7fb7dbf67927c53a87c0803e369bca6412e2786b90d04c23f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 304, + 171, + 524, + 297 + ], + "blocks": [ + { + "bbox": [ + 336, + 148, + 492, + 160 + ], + "lines": [ + { + "bbox": [ + 336, + 148, + 492, + 160 + ], + "spans": [ + { + "bbox": [ + 336, + 148, + 492, + 160 + ], + "type": "text", + "content": "Table 4: Performance vs. cost analysis" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 304, + 171, + 524, + 297 + ], + "lines": [ + { + "bbox": [ + 304, + 171, + 524, + 297 + ], + "spans": [ + { + "bbox": [ + 304, + 171, + 524, + 297 + ], + "type": "table", + "html": "
Datasetw/ Five AugmentationsNo Aug. Avg.
Avg.Std.MinMax
Frank86.7-1.00.485.887.686.2
MBNM74.4-0.10.473.774.975.1
SummEval75.2-0.90.574.576.074.3
QAGS-X81.6+0.50.580.882.480.7
QAGS-C76.4-1.60.874.777.975.2
DialFact92.1-0.40.291.592.391.2
BEGIN79.6+0.30.579.080.680.9
Q288.8-0.60.388.189.286.3
PAWS89.7-0.30.189.590.089.3
Avg.82.7-0.50.282.382.982.1
", + "image_path": "35502bb19b402effb419aed66d1b150a7d8baad5213f8d83296bbe14c13d1a6f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 306, + 526, + 426 + ], + "lines": [ + { + "bbox": [ + 302, + 306, + 526, + 426 + ], + "spans": [ + { + "bbox": [ + 302, + 306, + 526, + 426 + ], + "type": "text", + "content": "Table 5: Results of our phrase selection robustness analysis. For each run, we sample five phrases, recreated our dataset and retrain our model. We repeat this process ten times and report the average, as well as the standard deviation, minimum and maximum scores of the runs. Small numbers indicate difference to the original scores. All results were computed using " + }, + { + "bbox": [ + 302, + 306, + 526, + 426 + ], + "type": "inline_equation", + "content": "e-c" + }, + { + "bbox": [ + 302, + 306, + 526, + 426 + ], + "type": "text", + "content": " and MC dropout. For better comparison, we also report the scores of a model without any augmentation (i.e. without any additional training) with " + }, + { + "bbox": [ + 302, + 306, + 526, + 426 + ], + "type": "inline_equation", + "content": "e-c" + }, + { + "bbox": [ + 302, + 306, + 526, + 426 + ], + "type": "text", + "content": " and MC dropout." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 302, + 448, + 525, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 448, + 525, + 609 + ], + "spans": [ + { + "bbox": [ + 302, + 448, + 525, + 609 + ], + "type": "text", + "content": "requires fewer parameters than any other metric, including a more than " + }, + { + "bbox": [ + 302, + 448, + 525, + 609 + ], + "type": "inline_equation", + "content": "30\\mathrm{x}" + }, + { + "bbox": [ + 302, + 448, + 525, + 609 + ], + "type": "text", + "content": " reduction compared to T5. During inference our model always requires a constant number of calls which can be reduced to a single call when ablating MC dropout. On the other hand, the number of calls in SummacZS scales with the number of input and output sentences. Q2 needs to generate questions by calling an auto-regressive QG model " + }, + { + "bbox": [ + 302, + 448, + 525, + 609 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 302, + 448, + 525, + 609 + ], + "type": "text", + "content": " times, where " + }, + { + "bbox": [ + 302, + 448, + 525, + 609 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 302, + 448, + 525, + 609 + ], + "type": "text", + "content": " factors in the amount and length of questions " + }, + { + "bbox": [ + 302, + 448, + 525, + 609 + ], + "type": "inline_equation", + "content": "(\\# \\mathbf{Q}\\times \\mathbf{Q}1)" + }, + { + "bbox": [ + 302, + 448, + 525, + 609 + ], + "type": "text", + "content": ", answer #Q questions with the QA model and finally check #Q answers with an NLI model " + }, + { + "bbox": [ + 302, + 448, + 525, + 609 + ], + "type": "inline_equation", + "content": "(\\# \\mathbf{Q}\\times 2)" + }, + { + "bbox": [ + 302, + 448, + 525, + 609 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 610, + 525, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 610, + 525, + 650 + ], + "spans": [ + { + "bbox": [ + 302, + 610, + 525, + 650 + ], + "type": "text", + "content": "In sum, our model compares favourably with other approaches, while also allowing for a performance/cost tradeoff by forgiving MC dropout." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 661, + 463, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 661, + 463, + 672 + ], + "spans": [ + { + "bbox": [ + 302, + 661, + 463, + 672 + ], + "type": "text", + "content": "5.4 Phrase Selection Robustness" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 678, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 678, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 678, + 526, + 772 + ], + "type": "text", + "content": "To ensure that our augmentation is robust and not overly reliant on any particular choice of phrases, we repeat our dataset augmentation process multiple times with five randomly chosen augmentation phrases out of the original ten. We sample ten such datasets and retrain our model for each. Table 5 shows the average score, minimum and maxi" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "917" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 290, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 290, + 166 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 290, + 166 + ], + "type": "text", + "content": "mum score, as well as the standard deviation of the scores. We also report results of a model with both MC dropout and " + }, + { + "bbox": [ + 67, + 71, + 290, + 166 + ], + "type": "inline_equation", + "content": "e-c" + }, + { + "bbox": [ + 67, + 71, + 290, + 166 + ], + "type": "text", + "content": " but without any additional training and augmentations to directly quantify whether the augmentations are still helpful in their reduced form. This corresponds to applying MC dropout and " + }, + { + "bbox": [ + 67, + 71, + 290, + 166 + ], + "type": "inline_equation", + "content": "e-c" + }, + { + "bbox": [ + 67, + 71, + 290, + 166 + ], + "type": "text", + "content": " to Base." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 167, + 290, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 167, + 290, + 260 + ], + "spans": [ + { + "bbox": [ + 67, + 167, + 290, + 260 + ], + "type": "text", + "content": "As expected, we find that reducing the variety of available phrases leads to a drop in performance across almost all datasets, compared to A11. The only exception is BEGIN, where we instead see a slight improvement. This is likely to be related to the construction of BEGIN (see the discussion in Section 5.1)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 263, + 291, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 263, + 291, + 384 + ], + "spans": [ + { + "bbox": [ + 67, + 263, + 291, + 384 + ], + "type": "text", + "content": "When comparing our limited augmentation models to the non-augmented model, we find that they still outperform the non-augmented model in almost all cases. In particular for Q2 and DialFact, for which we expect the strongest impact of our augmentations, we find that even the worst run still outperforms non-augmented model. This suggests that our augmentations can robustly adapt the model to the dialogue task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 386, + 291, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 386, + 291, + 465 + ], + "spans": [ + { + "bbox": [ + 67, + 386, + 291, + 465 + ], + "type": "text", + "content": "Finally, we observe a relatively large drop in scores for all datasets that are at (least partially) derived from CNN/DM (Frank, SummEval and QAGS-C). This mirrors our earlier observation in Section 4 that these datasets profit from our augmentation procedure." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 478, + 160, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 478, + 160, + 491 + ], + "spans": [ + { + "bbox": [ + 67, + 478, + 160, + 491 + ], + "type": "text", + "content": "6 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 502, + 291, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 502, + 291, + 663 + ], + "spans": [ + { + "bbox": [ + 67, + 502, + 291, + 663 + ], + "type": "text", + "content": "Previous work on the utility of NLI for faithfulness led to mixed conclusions. In summarization, Falke et al. (2019) and Kryscinski et al. (2020) find out-of-the-box models have only limited utility in a faithfulness setting. In Wang et al. (2020), an NLI model is outperformed by a question generation/answering (QA/QG)-based method. In contrast, Maynez et al. (2020) find that a similar NLI model vastly outperforms a QA/QG metric on their data. In knowledge-grounded dialogue, Dziri et al. (2022), Gupta et al. (2022) and Honovich et al. (2021) find out-of-the-box models underperform." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 666, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 666, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 666, + 291, + 773 + ], + "type": "text", + "content": "To improve NLI models for faithfulness in summarization, Kryscinski et al. (2020) propose FactCC, which is trained on artificially noised summaries. Utama et al. (2022) propose a controllable generation model to generate artificial faithfulness data. In knowledge-grounded dialogue, Dziri et al. (2022) and Gupta et al. (2022) combine noising techniques to generate additional training data for" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 71, + 526, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 193 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 193 + ], + "type": "text", + "content": "NLI-based faithfulness models. In contrast to our work, these approaches a) generate training data from external sources, instead of directly augmenting NLI data, and b) do not explicitly focus on reconciling differences between NLI and faithfulness with their augmentation. Outside of augmentation-based approaches, Goyal and Durrett (2020) propose to train NLI models to label faithfulness at the dependency arc level." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 303, + 202, + 381, + 215 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 202, + 381, + 215 + ], + "spans": [ + { + "bbox": [ + 303, + 202, + 381, + 215 + ], + "type": "text", + "content": "7 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 223, + 525, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 223, + 525, + 264 + ], + "spans": [ + { + "bbox": [ + 302, + 223, + 525, + 264 + ], + "type": "text", + "content": "We have demonstrated that with a small number of focused adaptations, even a relatively small NLI model can robustly predict faithfulness. We have:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 272, + 525, + 423 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 312, + 272, + 525, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 272, + 525, + 326 + ], + "spans": [ + { + "bbox": [ + 312, + 272, + 525, + 326 + ], + "type": "text", + "content": "1. Shown that NLI-based metrics can be incompatible with task-specific requirements and identified and fixed one such incompatibility in dialogue with an augmentation strategy." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 334, + 525, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 334, + 525, + 387 + ], + "spans": [ + { + "bbox": [ + 312, + 334, + 525, + 387 + ], + "type": "text", + "content": "2. Demonstrated the importance of contradiction probability for scoring and that the underlying mechanism is the high reliability of NLI contradiction scores for detecting unfaithfulness" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 396, + 525, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 396, + 525, + 423 + ], + "spans": [ + { + "bbox": [ + 312, + 396, + 525, + 423 + ], + "type": "text", + "content": "3. Shown that using Monte-Carlo dropout improves metric performance." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 302, + 431, + 525, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 431, + 525, + 485 + ], + "spans": [ + { + "bbox": [ + 302, + 431, + 525, + 485 + ], + "type": "text", + "content": "Our improved NLI model significantly improves over its baseline across many corpora and outperforms all competitors in average score on TRUE, while being much more efficient at inference." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 486, + 525, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 486, + 525, + 553 + ], + "spans": [ + { + "bbox": [ + 302, + 486, + 525, + 553 + ], + "type": "text", + "content": "Our work suggests that strong improvements are possible for NLI-based faithfulness metrics, by combining data augmentation with adapted NLI score computation. We hope this finding will spurn advances in cheap and robust NLI for faithfulness." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 562, + 383, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 562, + 383, + 576 + ], + "spans": [ + { + "bbox": [ + 302, + 562, + 383, + 576 + ], + "type": "text", + "content": "8 Limitations" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 302, + 584, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 584, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 584, + 526, + 772 + ], + "type": "text", + "content": "Some of the summarization datasets annotated for faithfulness are relatively small, which makes score estimates uncertain. Furthermore, many datasets contain only output from a limited number of generation systems, which makes it hard to properly account for potential biases towards certain generation systems that may confound scores (see Pagnoni et al. (2021)). These concerns are, however, alleviated to some extent since we study trends across many independently created datasets, which makes it less likely for a single bias to persist in all of them. Furthermore the availability of generation and thus annotated faithfulness data limits our experiments to English. Finally, it remains" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "918" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 291, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 291, + 126 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 291, + 126 + ], + "type": "text", + "content": "unclear whether our results would still provide advantages when applied to larger models such as T5-11B, whose parameter count makes experimentation infeasible on the hardware available to us." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 134, + 176, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 134, + 176, + 147 + ], + "spans": [ + { + "bbox": [ + 67, + 134, + 176, + 147 + ], + "type": "text", + "content": "9 Ethics Statement" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 155, + 292, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 155, + 292, + 235 + ], + "spans": [ + { + "bbox": [ + 67, + 155, + 292, + 235 + ], + "type": "text", + "content": "Faithfulness metrics help reduce the amount of incorrect information generated by NLG systems, reducing the risk associated which such generations. However, faulty or unreliable faithfulness metrics might cause harm by incorrectly classifying faithful content as unfaithful and vice versa." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 236, + 291, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 236, + 291, + 344 + ], + "spans": [ + { + "bbox": [ + 67, + 236, + 291, + 344 + ], + "type": "text", + "content": "We run all experiments on publicly available data that has been specifically constructed for faithfulness evaluation. The underlying publication has been published at a conference whose review process involved an ethics review. For a specific discussion of the human effort involved in creation of the datasets we refer the reader to the original publications." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 366, + 127, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 366, + 127, + 379 + ], + "spans": [ + { + "bbox": [ + 68, + 366, + 127, + 379 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 384, + 291, + 773 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 69, + 384, + 291, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 384, + 291, + 462 + ], + "spans": [ + { + "bbox": [ + 69, + 384, + 291, + 462 + ], + "type": "text", + "content": "Samuel R. Bowman, Gabor Angeli, Christopher Potts, and Christopher D. Manning. 2015. A large annotated corpus for learning natural language inference. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 632-642, Lisbon, Portugal. Association for Computational Linguistics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 468, + 290, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 468, + 290, + 502 + ], + "spans": [ + { + "bbox": [ + 69, + 468, + 290, + 502 + ], + "type": "text", + "content": "Yanran Chen and Steffen Eger. 2022. Menli: Robust evaluation metrics from natural language inference. arXiv preprint arXiv:2208.07316." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 508, + 290, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 508, + 290, + 564 + ], + "spans": [ + { + "bbox": [ + 69, + 508, + 290, + 564 + ], + "type": "text", + "content": "Emily Dinan, Stephen Roller, Kurt Shuster, Angela Fan, Michael Auli, and Jason Weston. 2019. Wizard of wikipedia: Knowledge-powered conversational agents. In International Conference on Learning Representations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 571, + 290, + 659 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 571, + 290, + 659 + ], + "spans": [ + { + "bbox": [ + 69, + 571, + 290, + 659 + ], + "type": "text", + "content": "Nouha Dziri, Hannah Rashkin, Tal Linzen, and David Reitter. 2022. Evaluating Attribution in Dialogue Systems: The BEGIN Benchmark. Transactions of the Association for Computational Linguistics, 10:1066-1083. Note: TRUE uses an earlier version of the BEGIN dataset. The version used in TRUE is described in an earlier preprint at https://arxiv.org/pdf/2105.00071v1.pdf." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 666, + 291, + 721 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 666, + 291, + 721 + ], + "spans": [ + { + "bbox": [ + 69, + 666, + 291, + 721 + ], + "type": "text", + "content": "Alexander R. Fabbri, Wojciech Krysciński, Bryan McCann, Caiming Xiong, Richard Socher, and Dragomir Radev. 2021. SummEval: Re-evaluating summarization evaluation. Transactions of the Association for Computational Linguistics, 9:391-409." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 728, + 291, + 773 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 728, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 69, + 728, + 291, + 773 + ], + "type": "text", + "content": "Tobias Falke, Leonardo F. R. Ribeiro, Prasetya Ajie Utama, Ido Dagan, and Iryna Gurevych. 2019. Ranking generated summaries by correctness: An interesting but challenging application for natural language" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 526, + 772 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 314, + 72, + 526, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 526, + 117 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 526, + 117 + ], + "type": "text", + "content": "inference. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2214-2220, Florence, Italy. Association for Computational Linguistics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 123, + 526, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 123, + 526, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 123, + 526, + 179 + ], + "type": "text", + "content": "Purvi Goel and Li Chen. 2021. On the robustness of monte carlo dropout trained with noisy labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 2219-2228." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 186, + 526, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 186, + 526, + 243 + ], + "spans": [ + { + "bbox": [ + 304, + 186, + 526, + 243 + ], + "type": "text", + "content": "Tanya Goyal and Greg Durrett. 2020. Evaluating factuality in generation with dependency-level entailment. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 3592-3603, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 249, + 526, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 249, + 526, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 249, + 526, + 327 + ], + "type": "text", + "content": "Prakhar Gupta, Chien-Sheng Wu, Wenhao Liu, and Caiming Xiong. 2022. DialFact: A benchmark for fact-checking in dialogue. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3785-3801, Dublin, Ireland. Association for Computational Linguistics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 334, + 525, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 334, + 525, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 334, + 525, + 380 + ], + "type": "text", + "content": "Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. 2020. Deberta: Decoding-enhanced bert with disentangled attention. In International Conference on Learning Representations." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 386, + 526, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 386, + 526, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 386, + 526, + 464 + ], + "type": "text", + "content": "Karl Moritz Hermann, Tomáš Kočisky, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa Suleyman, and Phil Blunsom. 2015. Teaching machines to read and comprehend. In Proceedings of the 28th International Conference on Neural Information Processing Systems - Volume 1, NIPS'15, page 1693-1701, Cambridge, MA, USA. MIT Press." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 470, + 525, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 470, + 525, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 470, + 525, + 581 + ], + "type": "text", + "content": "Or Honovich, Roee Aharoni, Jonathan Herzig, Hagai Taitelbaum, Doron Kukliansy, Vered Cohen, Thomas Scialom, Idan Szpektor, Avinatan Hassidim, and Yossi Matias. 2022. TRUE: Re-evaluating factual consistency evaluation. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3905-3920, Seattle, United States. Association for Computational Linguistics." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 588, + 526, + 687 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 588, + 526, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 588, + 526, + 687 + ], + "type": "text", + "content": "Or Honovich, Leshem Choshen, Roee Aharoni, Ella Neeman, Idan Szpektor, and Omri Abend. 2021. " + }, + { + "bbox": [ + 304, + 588, + 526, + 687 + ], + "type": "inline_equation", + "content": "q^2" + }, + { + "bbox": [ + 304, + 588, + 526, + 687 + ], + "type": "text", + "content": ": Evaluating factual consistency in knowledge-grounded dialogues via question generation and question answering. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7856-7870, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 694, + 526, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 694, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 694, + 526, + 772 + ], + "type": "text", + "content": "Wojciech Kryscinski, Bryan McCann, Caiming Xiong, and Richard Socher. 2020. Evaluating the factual consistency of abstractive text summarization. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 9332-9346, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "919" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 291, + 127 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 291, + 127 + ], + "type": "text", + "content": "Philippe Laban, Tobias Schnabel, Paul N. Bennett, and Marti A. Hearst. 2022. SummaC: Re-visiting NLIBased models for inconsistency detection in summarization. Transactions of the Association for Computational Linguistics, 10:163-177." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 135, + 290, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 135, + 290, + 190 + ], + "spans": [ + { + "bbox": [ + 69, + 135, + 290, + 190 + ], + "type": "text", + "content": "Moritz Laurer, W v Atteveldt, Andreu Casas, and Kasper Welbers. 2022. Less annotating, more classifying-addressing the data scarcity issue of supervised machine learning with deep transfer learning and bert-nli." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 198, + 289, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 198, + 289, + 243 + ], + "spans": [ + { + "bbox": [ + 69, + 198, + 289, + 243 + ], + "type": "text", + "content": "Alisa Liu, Swabha Swayamdipta, Noah A Smith, and Yejin Choi. 2022. Wanli: Worker and ai collaboration for natural language inference dataset creation. arXiv preprint arXiv:2201.05955." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 250, + 290, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 250, + 290, + 317 + ], + "spans": [ + { + "bbox": [ + 69, + 250, + 290, + 317 + ], + "type": "text", + "content": "Joshua Maynez, Shashi Narayan, Bernd Bohnet, and Ryan McDonald. 2020. On faithfulness and factuality in abstractive summarization. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1906-1919, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 324, + 290, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 324, + 290, + 401 + ], + "spans": [ + { + "bbox": [ + 69, + 324, + 290, + 401 + ], + "type": "text", + "content": "Yixin Nie, Adina Williams, Emily Dinan, Mohit Bansal, Jason Weston, and Douwe Kiela. 2020. Adversarial NLI: A new benchmark for natural language understanding. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4885-4901, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 409, + 290, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 409, + 290, + 497 + ], + "spans": [ + { + "bbox": [ + 69, + 409, + 290, + 497 + ], + "type": "text", + "content": "Artidoro Pagnoni, Vidhisha Balachandran, and Yulia Tsvetkov. 2021. Understanding factuality in abstractive summarization with FRANK: A benchmark for factuality metrics. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4812-4829, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 506, + 290, + 594 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 506, + 290, + 594 + ], + "spans": [ + { + "bbox": [ + 69, + 506, + 290, + 594 + ], + "type": "text", + "content": "Alicia Parrish, William Huang, Omar Agha, Soo-Hwan Lee, Nikita Nangia, Alexia Warstadt, Karmanya Aggarwal, Emily Allaway, Tal Linzen, and Samuel R. Bowman. 2021. Does putting a linguist in the loop improve NLU data collection? In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 4886-4901, Punta Cana, Dominican Republic. Association for Computational Linguistics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 602, + 290, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 602, + 290, + 645 + ], + "spans": [ + { + "bbox": [ + 69, + 602, + 290, + 645 + ], + "type": "text", + "content": "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 653, + 290, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 653, + 290, + 708 + ], + "spans": [ + { + "bbox": [ + 69, + 653, + 290, + 708 + ], + "type": "text", + "content": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J Liu, et al. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(140):1-67." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 716, + 290, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 716, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 716, + 290, + 772 + ], + "type": "text", + "content": "Hannah Rashkin, David Reitter, Gaurav Singh Tomar, and Dipanjan Das. 2021. Increasing faithfulness in knowledge-grounded dialogue with controllable features. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 525, + 772 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 314, + 72, + 525, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 525, + 116 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 525, + 116 + ], + "type": "text", + "content": "and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 704-718, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 125, + 525, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 125, + 525, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 125, + 525, + 179 + ], + "type": "text", + "content": "Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. 2014. Dropout: A simple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 15(56):1929-1958." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 190, + 525, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 190, + 525, + 256 + ], + "spans": [ + { + "bbox": [ + 304, + 190, + 525, + 256 + ], + "type": "text", + "content": "Emma Strubell, Ananya Ganesh, and Andrew McCallum. 2019. Energy and policy considerations for deep learning in NLP. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3645-3650, Florence, Italy. Association for Computational Linguistics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 264, + 525, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 264, + 525, + 364 + ], + "spans": [ + { + "bbox": [ + 304, + 264, + 525, + 364 + ], + "type": "text", + "content": "James Thorne, Andreas Vlachos, Christos Christodoulopoulos, and Arpit Mittal. 2018. FEVER: a large-scale dataset for fact extraction and VERIFICATION. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 809-819, New Orleans, Louisiana. Association for Computational Linguistics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 372, + 525, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 372, + 525, + 471 + ], + "spans": [ + { + "bbox": [ + 304, + 372, + 525, + 471 + ], + "type": "text", + "content": "Prasetya Utama, Joshua Bambrick, Nafise Moosavi, and Iryna Gurevych. 2022. Falsesum: Generating document-level NLI examples for recognizing factual inconsistency in summarization. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2763-2776, Seattle, United States. Association for Computational Linguistics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 481, + 525, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 481, + 525, + 546 + ], + "spans": [ + { + "bbox": [ + 304, + 481, + 525, + 546 + ], + "type": "text", + "content": "Alex Wang, Kyunghyun Cho, and Mike Lewis. 2020. Asking and answering questions to evaluate the factual consistency of summaries. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5008-5020, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 555, + 525, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 525, + 622 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 525, + 622 + ], + "type": "text", + "content": "Alex Wang, Yada Pruksachatkun, Nikita Nangia, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. 2019. Superglue: A stickier benchmark for general-purpose language understanding systems. Advances in neural information processing systems, 32." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 630, + 525, + 719 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 525, + 719 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 525, + 719 + ], + "type": "text", + "content": "Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. 2018. GLUE: A multi-task benchmark and analysis platform for natural language understanding. In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 353-355, Brussels, Belgium. Association for Computational Linguistics." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 728, + 525, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 728, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 728, + 525, + 772 + ], + "type": "text", + "content": "Adina Williams, Nikita Nangia, and Samuel Bowman. 2018. A broad-coverage challenge corpus for sentence understanding through inference. In Proceedings of the 2018 Conference of the North American" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 790 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 790 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 790 + ], + "type": "text", + "content": "920" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 92, + 68, + 267, + 257 + ], + "blocks": [ + { + "bbox": [ + 92, + 68, + 267, + 257 + ], + "lines": [ + { + "bbox": [ + 92, + 68, + 267, + 257 + ], + "spans": [ + { + "bbox": [ + 92, + 68, + 267, + 257 + ], + "type": "table", + "html": "
Introductory Statements
Here is what I know:
yep. Also
Sure! Here is what I know:
Hedging
I am not sure, but
I am not sure but I do know that
I do not have information on this but
I think
I believe
Sentiment
I love that!
I like that!
", + "image_path": "06ab63d2277531f27567d7eca39e01fd8f66b688b9ec31cb55590a15c6e7cba3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 78, + 298, + 291, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 298, + 291, + 354 + ], + "spans": [ + { + "bbox": [ + 78, + 298, + 291, + 354 + ], + "type": "text", + "content": "Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1112-1122, New Orleans, Louisiana. Association for Computational Linguistics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 361, + 291, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 361, + 291, + 495 + ], + "spans": [ + { + "bbox": [ + 69, + 361, + 291, + 495 + ], + "type": "text", + "content": "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 501, + 291, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 501, + 291, + 591 + ], + "spans": [ + { + "bbox": [ + 69, + 501, + 291, + 591 + ], + "type": "text", + "content": "Yuan Zhang, Jason Baldridge, and Luheng He. 2019. PAWS: Paraphrase adversaries from word scrambling. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 1298-1308, Minneapolis, Minnesota. Association for Computational Linguistics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 599, + 250, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 599, + 250, + 613 + ], + "spans": [ + { + "bbox": [ + 68, + 599, + 250, + 613 + ], + "type": "text", + "content": "A Augmentation Training Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 620, + 205, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 620, + 205, + 634 + ], + "spans": [ + { + "bbox": [ + 68, + 620, + 205, + 634 + ], + "type": "text", + "content": "A.1 Augmentation Phrases" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 638, + 290, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 638, + 290, + 691 + ], + "spans": [ + { + "bbox": [ + 67, + 638, + 290, + 691 + ], + "type": "text", + "content": "Table 6 lists our manually curated list of phrases inserted during data augmentation. All phrases were derived via a small manual error analysis on the Base model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 692, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 692, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 692, + 291, + 772 + ], + "type": "text", + "content": "We broadly divide our phrases into three categories: introductory statements, hedging, and sentiment statements. For each instance in ANLI, one random phrase from the list is presupended to the hypothesis. We use all three rounds of ANLI annotations. This results in 162,865 augmented instances" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 347, + 68, + 481, + 126 + ], + "blocks": [ + { + "bbox": [ + 77, + 264, + 280, + 278 + ], + "lines": [ + { + "bbox": [ + 77, + 264, + 280, + 278 + ], + "spans": [ + { + "bbox": [ + 77, + 264, + 280, + 278 + ], + "type": "text", + "content": "Table 6: Manually curated list of dialogue phrases" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 347, + 68, + 481, + 126 + ], + "lines": [ + { + "bbox": [ + 347, + 68, + 481, + 126 + ], + "spans": [ + { + "bbox": [ + 347, + 68, + 481, + 126 + ], + "type": "table", + "html": "
ParameterVal.
Warmup Ratio0.06
Weight Decay0.01
Effective Batch Size64
", + "image_path": "6f282655f4d8975ecd11d6244cc838b609584d1676fa0d44545c8bb9b26ea6cc.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 360, + 135, + 467, + 147 + ], + "lines": [ + { + "bbox": [ + 360, + 135, + 467, + 147 + ], + "spans": [ + { + "bbox": [ + 360, + 135, + 467, + 147 + ], + "type": "text", + "content": "Table 7: Hyperparameters" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 302, + 168, + 525, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 168, + 525, + 195 + ], + "spans": [ + { + "bbox": [ + 302, + 168, + 525, + 195 + ], + "type": "text", + "content": "which, together with the original ANLI instances, leads to a total of 325,730 training instances." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 303, + 203, + 417, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 203, + 417, + 216 + ], + "spans": [ + { + "bbox": [ + 303, + 203, + 417, + 216 + ], + "type": "text", + "content": "A.2 Hyperparameters" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 221, + 526, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 221, + 526, + 423 + ], + "spans": [ + { + "bbox": [ + 302, + 221, + 526, + 423 + ], + "type": "text", + "content": "Table 7 lists the hyperparameter settings for our model. We use the same optimizer hyperparameters as Laurer et al. (2022) except for an increased batch size and the learning rate. For the latter we tested three learning rates " + }, + { + "bbox": [ + 302, + 221, + 526, + 423 + ], + "type": "inline_equation", + "content": "(5e - 6, 5e - 2, 5e - 1)" + }, + { + "bbox": [ + 302, + 221, + 526, + 423 + ], + "type": "text", + "content": " and select the one that provided the best loss on the augmented ANLI validation set. We initially ran models for 10,000 steps with a checkpoint every 1,000 steps and selected the checkpoint with the lowest loss on the augmented ANLI validation set. Later we reduced the number of training steps to 2,000 since we found we would usually select an early checkpoint as validation loss increased later in training, likely related to overfitting on the augmented data." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 303, + 433, + 374, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 433, + 374, + 445 + ], + "spans": [ + { + "bbox": [ + 303, + 433, + 374, + 445 + ], + "type": "text", + "content": "A.3 Training" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 449, + 525, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 449, + 525, + 530 + ], + "spans": [ + { + "bbox": [ + 302, + 449, + 525, + 530 + ], + "type": "text", + "content": "We use the DeBERTa implementation in the huggingface transformers library (Wolf et al., 2020) and trained our model on a single node using two RX6800 GPUs, with one training run taking about three hours. Later experiments with fewer steps cut that time by " + }, + { + "bbox": [ + 302, + 449, + 525, + 530 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 302, + 449, + 525, + 530 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 303, + 541, + 444, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 541, + 444, + 553 + ], + "spans": [ + { + "bbox": [ + 303, + 541, + 444, + 553 + ], + "type": "text", + "content": "B Dataset Bias in BEGIN" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 302, + 561, + 525, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 561, + 525, + 697 + ], + "spans": [ + { + "bbox": [ + 302, + 561, + 525, + 697 + ], + "type": "text", + "content": "BEGIN is the only dialogue corpus on which first person pronoun occurrence shows a strong (negative) correlation with faithfulness (see Table 3). Since there is nothing in the annotation guidelines that would explain this correlation, we instead hypothesize that this is the consequence of a model induced bias in the data. Specifically, we hypothesize that one of the two models in BEGIN is (1) more likely to generate personal statements and (2) less likely to generate faithful responses." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 302, + 698, + 525, + 736 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 698, + 525, + 736 + ], + "spans": [ + { + "bbox": [ + 302, + 698, + 525, + 736 + ], + "type": "text", + "content": "To avoid confusion in the remainder of this section, we highlight that there are two variants of BEGIN:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 303, + 746, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 746, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 303, + 746, + 525, + 772 + ], + "type": "text", + "content": "BEGIN-v1 is the variant used in TRUE. It contains labeled generations by a fine-tuned GPT" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 305, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 305, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 305, + 791 + ], + "type": "text", + "content": "921" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 89, + 71, + 290, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 71, + 290, + 111 + ], + "spans": [ + { + "bbox": [ + 89, + 71, + 290, + 111 + ], + "type": "text", + "content": "2 base (Radford et al., 2019) and a fine-tuned T5 base model (Raffel et al., 2020) on the Wizard of Wikipedia dataset (Dinan et al., 2019).6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 120, + 291, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 120, + 291, + 215 + ], + "spans": [ + { + "bbox": [ + 69, + 120, + 291, + 215 + ], + "type": "text", + "content": "BEGIN-v2 is a more recent variant of BEGIN that is not part of TRUE. In addition to new instances generated by T5 and GPT-2 it contains outputs from two additional models. It also has a revised annotation procedure. When we refer to BEGIN-v2, we exclusively mean the Wizard of Wikipedia subset." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 224, + 290, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 224, + 290, + 331 + ], + "spans": [ + { + "bbox": [ + 67, + 224, + 290, + 331 + ], + "type": "text", + "content": "Unfortunately, BEGIN-v1 does not allow us to retrieve which model generated which instance. This makes it impossible to directly investigate for model bias. However, BEGIN-v2 includes outputs by the same two models, fine-tuned on the same data. Since we only need corpus level statistics to verify our assumptions, we conduct our analysis on the GPT-2 and T5 instances in BEGIN-v2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 332, + 290, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 332, + 290, + 427 + ], + "spans": [ + { + "bbox": [ + 67, + 332, + 290, + 427 + ], + "type": "text", + "content": "To verify (1), we compute the correlation between a binary variable indicating which model generated each instance (T5: 0, GPT-2: 1) and first-person pronoun occurrence. We find a positive correlation (Kendall's " + }, + { + "bbox": [ + 67, + 332, + 290, + 427 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 67, + 332, + 290, + 427 + ], + "type": "text", + "content": " wrt. to " + }, + { + "bbox": [ + 67, + 332, + 290, + 427 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 67, + 332, + 290, + 427 + ], + "type": "text", + "content": "-pronoun occurrence: " + }, + { + "bbox": [ + 67, + 332, + 290, + 427 + ], + "type": "inline_equation", + "content": "0.18, p < 0.001" + }, + { + "bbox": [ + 67, + 332, + 290, + 427 + ], + "type": "text", + "content": "), indicating that GPT-2 generates outputs including more first-person pronouns." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 427, + 290, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 427, + 290, + 615 + ], + "spans": [ + { + "bbox": [ + 67, + 427, + 290, + 615 + ], + "type": "text", + "content": "To investigate whether GPT-2 is also more likely to be unfaithful, i.e. to verify (2), we compute the correlation between the binary model indicator variable and a faithfulness variable that is 1 when the output is labelled as Fully attributable and 0 otherwise. We find a negative correlation (Kendall's " + }, + { + "bbox": [ + 67, + 427, + 290, + 615 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 67, + 427, + 290, + 615 + ], + "type": "text", + "content": " wrt. to Faithfulness: " + }, + { + "bbox": [ + 67, + 427, + 290, + 615 + ], + "type": "inline_equation", + "content": "-0.25" + }, + { + "bbox": [ + 67, + 427, + 290, + 615 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 67, + 427, + 290, + 615 + ], + "type": "inline_equation", + "content": "p < 0.001" + }, + { + "bbox": [ + 67, + 427, + 290, + 615 + ], + "type": "text", + "content": "), supporting our hypothesis that GPT-2 is also overall less faithful. To ensure that this is not an effect of additional personal statements leading to more unfaithful generations, we conduct the same analysis only on instances where we identify no first-person pronouns. We find a similarly strong negative correlation of " + }, + { + "bbox": [ + 67, + 427, + 290, + 615 + ], + "type": "inline_equation", + "content": "-0.29" + }, + { + "bbox": [ + 67, + 427, + 290, + 615 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 67, + 427, + 290, + 615 + ], + "type": "inline_equation", + "content": "p < 0.001" + }, + { + "bbox": [ + 67, + 427, + 290, + 615 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 616, + 290, + 724 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 616, + 290, + 724 + ], + "spans": [ + { + "bbox": [ + 67, + 616, + 290, + 724 + ], + "type": "text", + "content": "Our analysis shows that GPT-2 produces both overall less faithful outputs and more first-person pronouns than T5. Since BEGIN-v1 contains only outputs from T5 and GPT-2 this suggests that the root cause for the negative correlation between faithfulness label and first-person pronoun occurrence in BEGIN-v1 is model bias confounding faithfulness and first-person pronoun occurrence." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 311, + 68, + 518, + 172 + ], + "blocks": [ + { + "bbox": [ + 311, + 68, + 518, + 172 + ], + "lines": [ + { + "bbox": [ + 311, + 68, + 518, + 172 + ], + "spans": [ + { + "bbox": [ + 311, + 68, + 518, + 172 + ], + "type": "table", + "html": "
CorpusFaith.Non. FaithTotal
Frank223 (33.2%)448 (66.8%)671
MNBM255 (10.2%)2245 (89.8%)2500
SummEval1306 (81.6%)294 (18.4%)1600
QAGS-X116 (48.5%)123 (51.5%)239
QAGS-C113 (48.1%)122 (51.9%)235
BEGIN282 (33.7%)554 (66.3%)836
DialFact3341 (38.5%)5348 (61.5%)8689
Q2628 (57.7%)460 (42.3%)1088
PAWS3539 (44.2%)4461 (55.8%)8000
", + "image_path": "bbabbe41630dc254d59f33a0ce15224ff234b93320468df5f2729a95508bda9a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 180, + 524, + 204 + ], + "lines": [ + { + "bbox": [ + 302, + 180, + 524, + 204 + ], + "spans": [ + { + "bbox": [ + 302, + 180, + 524, + 204 + ], + "type": "text", + "content": "Table 8: Dataset statistics for all constituent corpora in TRUE" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 302, + 227, + 455, + 238 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 227, + 455, + 238 + ], + "spans": [ + { + "bbox": [ + 302, + 227, + 455, + 238 + ], + "type": "text", + "content": "B.1 Dataset Bias in BEGIN-v2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 244, + 524, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 244, + 524, + 269 + ], + "spans": [ + { + "bbox": [ + 302, + 244, + 524, + 269 + ], + "type": "text", + "content": "We conduct a preliminary study to investigate whether similar biases also exist in BEGIN-v2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 271, + 525, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 271, + 525, + 417 + ], + "spans": [ + { + "bbox": [ + 302, + 271, + 525, + 417 + ], + "type": "text", + "content": "We observe that while BEGIN-v2 uses data from four dialogue systems, a majority of faithful generations is produced by a single system called CTRL-DIALOG (Rashkin et al., 2021). CTRL-DIALOG is specifically trained to generate less subjective text, which we hypothesize might result in fewer first person pronouns. Since CTRL-DIALOG also produces more faithful texts, this would lead to a negative correlation between faithfulness and first person pronouns, similar to what we observe on BEGIN-v1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 420, + 525, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 420, + 525, + 594 + ], + "spans": [ + { + "bbox": [ + 302, + 420, + 525, + 594 + ], + "type": "text", + "content": "We verify this assumption by computing the correlation of a binary variable indicating an instance has been generated by CTRL-DIALOG with a) the faithfulness labels on BEGIN-v2 and b) first-person pronoun occurrence. We find that an instance being generated by CTRL-DIALOG is positively correlated with it having a faithful label (Kendall " + }, + { + "bbox": [ + 302, + 420, + 525, + 594 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 302, + 420, + 525, + 594 + ], + "type": "text", + "content": " w.r.t. faithfulness: 0.48, " + }, + { + "bbox": [ + 302, + 420, + 525, + 594 + ], + "type": "inline_equation", + "content": "p < 0.001" + }, + { + "bbox": [ + 302, + 420, + 525, + 594 + ], + "type": "text", + "content": ") while being negatively correlated with the number of pronouns (Kendall " + }, + { + "bbox": [ + 302, + 420, + 525, + 594 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 302, + 420, + 525, + 594 + ], + "type": "text", + "content": " w.r.t. I-pronoun occurrence: -0.34, " + }, + { + "bbox": [ + 302, + 420, + 525, + 594 + ], + "type": "inline_equation", + "content": "p < 0.001" + }, + { + "bbox": [ + 302, + 420, + 525, + 594 + ], + "type": "text", + "content": "). This suggests future evaluations on the BEGIN-v2 might run into similar bias issues." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 606, + 414, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 606, + 414, + 618 + ], + "spans": [ + { + "bbox": [ + 302, + 606, + 414, + 618 + ], + "type": "text", + "content": "C Dataset Statistics" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 627, + 524, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 627, + 524, + 653 + ], + "spans": [ + { + "bbox": [ + 302, + 627, + 524, + 653 + ], + "type": "text", + "content": "We report the number of instances, as well as the class distribution of TRUE in Table 8." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 731, + 290, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 731, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 731, + 290, + 772 + ], + "type": "text", + "content": "The relevant data can be found at https://raw. githubusercontent.com/google/BEGIN-dataset/ 5fa0cb0dde0e653d2016724a52a5ca27fe8b6a3f/dev_05_ 24_21.tsv" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "922" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "spans": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "type": "text", + "content": "A For every submission:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 106, + 414, + 240 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "spans": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "type": "text", + "content": "A1. Did you describe the limitations of your work? 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 142, + 329, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 142, + 329, + 168 + ], + "spans": [ + { + "bbox": [ + 77, + 142, + 329, + 168 + ], + "type": "text", + "content": "A2. Did you discuss any potential risks of your work? 9" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 178, + 414, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 178, + 414, + 203 + ], + "spans": [ + { + "bbox": [ + 77, + 178, + 414, + 203 + ], + "type": "text", + "content": "A3. Do the abstract and introduction summarize the paper's main claims?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 77, + 214, + 398, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 214, + 398, + 240 + ], + "spans": [ + { + "bbox": [ + 77, + 214, + 398, + 240 + ], + "type": "text", + "content": "A4. Have you used AI writing assistants when working on this paper? Left blank." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 69, + 249, + 290, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 249, + 290, + 264 + ], + "spans": [ + { + "bbox": [ + 69, + 249, + 290, + 264 + ], + "type": "text", + "content": "B Did you use or create scientific artifacts?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 80, + 269, + 95, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 269, + 95, + 280 + ], + "spans": [ + { + "bbox": [ + 80, + 269, + 95, + 280 + ], + "type": "text", + "content": "1,3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 77, + 289, + 524, + 477 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 289, + 315, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 289, + 315, + 315 + ], + "spans": [ + { + "bbox": [ + 77, + 289, + 315, + 315 + ], + "type": "text", + "content": "B1. Did you cite the creators of artifacts you used? 1,3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 77, + 324, + 463, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 324, + 463, + 350 + ], + "spans": [ + { + "bbox": [ + 77, + 324, + 463, + 350 + ], + "type": "text", + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? 1,9" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 77, + 360, + 524, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 360, + 524, + 426 + ], + "spans": [ + { + "bbox": [ + 77, + 360, + 524, + 426 + ], + "type": "text", + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? 9" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 77, + 438, + 524, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 438, + 524, + 477 + ], + "spans": [ + { + "bbox": [ + 77, + 438, + 524, + 477 + ], + "type": "text", + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 88, + 479, + 524, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 479, + 524, + 518 + ], + "spans": [ + { + "bbox": [ + 88, + 479, + 524, + 518 + ], + "type": "text", + "content": "Most data is machine generated and thus unlikely to reveal personal information. All data is also already publicly available and has been introduced in peer-reviewed publications, providing an additional safeguard." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 77, + 527, + 524, + 643 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 77, + 527, + 524, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 527, + 524, + 567 + ], + "spans": [ + { + "bbox": [ + 77, + 527, + 524, + 567 + ], + "type": "text", + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? We discuss the limitation to English in Section 9." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 77, + 576, + 524, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 576, + 524, + 643 + ], + "spans": [ + { + "bbox": [ + 77, + 576, + 524, + 643 + ], + "type": "text", + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 89, + 645, + 143, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 645, + 143, + 657 + ], + "spans": [ + { + "bbox": [ + 89, + 645, + 143, + 657 + ], + "type": "text", + "content": "Appendix C" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 69, + 665, + 293, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 665, + 293, + 679 + ], + "spans": [ + { + "bbox": [ + 69, + 665, + 293, + 679 + ], + "type": "text", + "content": "C Did you run computational experiments?" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 80, + 684, + 95, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 684, + 95, + 695 + ], + "spans": [ + { + "bbox": [ + 80, + 684, + 95, + 695 + ], + "type": "text", + "content": "3,4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 77, + 704, + 524, + 731 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 704, + 524, + 731 + ], + "spans": [ + { + "bbox": [ + 77, + 704, + 524, + 731 + ], + "type": "text", + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 89, + 733, + 142, + 745 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 733, + 142, + 745 + ], + "spans": [ + { + "bbox": [ + 89, + 733, + 142, + 745 + ], + "type": "text", + "content": "Appendix A" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "spans": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "text", + "content": "ACL 2023 Responsible NLP Checklist" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 751, + 522, + 771 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 751, + 522, + 771 + ], + "spans": [ + { + "bbox": [ + 67, + 751, + 522, + 771 + ], + "type": "text", + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "923" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 77, + 70, + 523, + 238 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 77, + 70, + 523, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 70, + 523, + 111 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 523, + 111 + ], + "type": "text", + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Appendix A" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 77, + 120, + 523, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 120, + 523, + 172 + ], + "spans": [ + { + "bbox": [ + 77, + 120, + 523, + 172 + ], + "type": "text", + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 183, + 523, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 183, + 523, + 238 + ], + "spans": [ + { + "bbox": [ + 77, + 183, + 523, + 238 + ], + "type": "text", + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? 5.2,Appendix A" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 67, + 247, + 522, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 247, + 522, + 278 + ], + "spans": [ + { + "bbox": [ + 67, + 247, + 522, + 278 + ], + "type": "text", + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 76, + 286, + 523, + 539 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 76, + 286, + 523, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 286, + 523, + 326 + ], + "spans": [ + { + "bbox": [ + 76, + 286, + 523, + 326 + ], + "type": "text", + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 76, + 336, + 523, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 336, + 523, + 390 + ], + "spans": [ + { + "bbox": [ + 76, + 336, + 523, + 390 + ], + "type": "text", + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 76, + 400, + 523, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 400, + 523, + 453 + ], + "spans": [ + { + "bbox": [ + 76, + 400, + 523, + 453 + ], + "type": "text", + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 76, + 462, + 519, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 462, + 519, + 489 + ], + "spans": [ + { + "bbox": [ + 76, + 462, + 519, + 489 + ], + "type": "text", + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 76, + 498, + 523, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 498, + 523, + 539 + ], + "spans": [ + { + "bbox": [ + 76, + 498, + 523, + 539 + ], + "type": "text", + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "924" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/XL-LEXEME_ WiC Pretrained Model for Cross-Lingual LEXical sEMantic changE/4754c7cf-ef79-4237-89fc-dec2c7680fce_content_list.json b/2023/XL-LEXEME_ WiC Pretrained Model for Cross-Lingual LEXical sEMantic changE/4754c7cf-ef79-4237-89fc-dec2c7680fce_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a80e109e4c03f4b8b43ffc59bb4d727e21ff693a --- /dev/null +++ b/2023/XL-LEXEME_ WiC Pretrained Model for Cross-Lingual LEXical sEMantic changE/4754c7cf-ef79-4237-89fc-dec2c7680fce_content_list.json @@ -0,0 +1,1294 @@ +[ + { + "type": "text", + "text": "XL-LEXEME: WiC Pretrained Model for Cross-Lingual LEXical sEMantic change", + "text_level": 1, + "bbox": [ + 280, + 89, + 715, + 130 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pierluigi Cassotti, Lucia Siciliani, Marco de Gemmis, Giovanni Semeraro and Pierpaolo Basile", + "bbox": [ + 186, + 149, + 815, + 181 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "University of Bari Aldo Moro", + "bbox": [ + 376, + 183, + 625, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{firstname.lastname} @uniba.it", + "bbox": [ + 376, + 199, + 628, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 252, + 339, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The recent introduction of large-scale datasets for the WiC (Word in Context) task enables the creation of more reliable and meaningful contextualized word embeddings. However, most of the approaches to the WiC task use cross-encoders, which prevent the possibility of deriving comparable word embeddings. In this work, we introduce XL-LEXEME, a Lexical Semantic Change Detection model. XL-LEXEME extends SBERT, highlighting the target word in the sentence. We evaluate XL-LEXEME on the multilingual benchmarks for SemEval-2020 Task 1 - Lexical Semantic Change (LSC) Detection and the RuShiftEval shared task involving five languages: English, German, Swedish, Latin, and Russian. XL-LEXEME outperforms the state-of-the-art in English, German and Swedish with statistically significant differences from the baseline results and obtains state-of-the-art performance in the RuShiftEval shared task.", + "bbox": [ + 144, + 278, + 460, + 575 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction and Motivation", + "text_level": 1, + "bbox": [ + 114, + 588, + 394, + 602 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lexical Semantic Change (LSC) Detection is the task of automatically identifying words that change their meaning over time. The LSC Detection task implicitly aims to disambiguate synchronic word sense occurrences and then find differences in the word sense frequencies in different periods. Word Sense Disambiguation (WSD) is a long-studied task in Natural Language Processing (Navigli, 2009), which consists of associating the correct sense to a word occurring in a specific context. WSD involves some crucial issues, such as relying on a fixed sense inventory. Fixed sense inventories ignore the diachronic aspect of language because they can miss older unused senses or be outdated and missing new senses.", + "bbox": [ + 115, + 613, + 487, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The Word in Context task (WiC) (Pilehvar and Camacho-Collados, 2019) aims to overcome these issues. In this work, we train a model on the WiC task and then use it to perform LSC Detection. In", + "bbox": [ + 112, + 854, + 485, + 917 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the WiC task, given the word $w$ and two different contexts $C1$ , $C2$ , the systems have to determine whether the meaning of $w$ is the same in the two contexts or not. Our approach is grounded on the assumption that models trained on the WiC tasks are robust enough to transfer the knowledge learned in a synchronic setting to a diachronic one. We summarise the main contribution of this work as follows: (i) We propose a pre-trained bi-encoder model, called XL-LEXEME, on a largescale dataset for the WiC task, which allows us to obtain comparable lexical-based representations; (ii) We assert the effectiveness of XL-LEXEME despite the computational limitation compared to the cross-encoder architecture for the LSC Detection task; (iii) Experiments on the LSC Detection task show that XL-LEXEME outperforms state-of-the-art LSC Detection models for English, German, Swedish, and Russian.", + "bbox": [ + 507, + 252, + 885, + 557 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 509, + 570, + 663, + 586 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "LSC Detection systems can be categorized based on the distributional embeddings used to tackle the LSC Detection task. One category is represented by those approaches that adopt type-base (i.e., static) embeddings. UWB (Prazák et al., 2020; Prazák et al., 2021) represents an example of this category of systems. First, it employs word2vec Skip-gram with Negative Sampling (Mikolov et al., 2013) to compute a semantic space for each corpus. It uses techniques like the Canonical Correlation Analysis (Hardoon et al., 2004) and the Orthogonal Transformation (Hamilton et al., 2016) to align the abovementioned spaces. Therefore, the cosine similarity between the vectors representing the word in two different spaces is used to detect the semantic shift.", + "bbox": [ + 507, + 596, + 882, + 851 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the increasing use of contextualized word embeddings, numerous approaches employing BERT-base models have been developed for LSC Detection (Montanelli and Periti, 2023; Laicher", + "bbox": [ + 507, + 854, + 882, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1577", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics", + "bbox": [ + 226, + 945, + 769, + 957 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Volume 2: Short Papers, pages 1577-1585", + "bbox": [ + 368, + 958, + 628, + 971 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "July 9-14, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 295, + 972, + 700, + 985 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "et al., 2021). In TempoBERT (Rosin et al., 2022), the authors exploit the concept of Masked Language Modeling (MLM), where the goal is to train a language model to predict a masked portion of text given the remaining part. In particular, they employ this technique to encode the concept of time into a BERT model. This is done by concatenating a specific token representing time to the text sequence. At inference time, TempoBERT can be used to predict the year of a sentence, masking the time reference, or to predict a masked token of the sentence conditioned by the time reference. In the same line of research, in Temporal Attention (Rosin and Radinsky, 2022), the authors investigate the effect of modifying the model instead of the input sentence like in TempoBERT. This is done by extending the model's attention mechanism to consider the time when computing the weight of each word. The time dimension is encoded using a different query embedding matrix for each timestamp.", + "bbox": [ + 110, + 84, + 492, + 420 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another kind of approach exploits the information coming from other tasks to perform LSC Detection. GlossReader represents an example (Rachinskiy and Arefyev, 2021), where a model based on XML-R (Conneau et al., 2020b) is first trained on English SemCor (Miller et al., 1994) with glosses from WordNet 3.0 (Miller, 1992) to perform WSD. Exploiting the zero-shot cross-lingual characteristics of XML-R, the authors used the same model to perform LSC Detection in the Russian language. With DeepMistake (Arefyev et al., 2021), the authors take advantage of the WiC task instead of WSD. They train a cross-encoder with XML-R as an underlying Language Model on the MCL-WiC training and development set and fine-tune on the RuSemShift dataset (Rodina and Kutuzov, 2020). DeepMistake, differently from XL-LEXEME, relies on the cross-encoder architecture and exploits only the MCL-WiC training dataset.", + "bbox": [ + 110, + 423, + 490, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3 XL-LEXEME", + "text_level": 1, + "bbox": [ + 112, + 760, + 272, + 778 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generally, for pairwise sentence similarity tasks, BERT models use a cross-encoder, in which the pairwise sequences are jointly encoded, and the overall vectors are used for the classification. However, in several tasks, the cross-encoder is not suitable since it cannot provide a distinct meaningful representation for each sentence. An approach to overcome this issue involves pooling the BERT out", + "bbox": [ + 110, + 790, + 490, + 919 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "put encoded vectors, which often results in worse performance. Sentence-BERT (SBERT) (Reimers and Gurevych, 2019) overcomes the limitation of cross-encoders using a Siamese Network, i.e., the weights of the underlying networks are shared. SBERT encodes the two sequences separately in the BERT model exploiting the Siamese architecture. The sequence-level representation is obtained by averaging the output encoded vectors, which are directly compared using similarity measures such as cosine similarity.", + "bbox": [ + 507, + 84, + 884, + 260 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Meanwhile, cross-encoders perform better since they are trained to profit from the attention over the whole input. In this work, we introduce XLLEXEME1 which mirrors models for pairwise sequence similarity tasks and adapts them to the WiC task, giving prominence to the target word, i.e. the word for which we want to detect the LSC. The model takes as input two sequences $s_1$ and $s_2$ . The sequences are tokenized using subwords tokenizer, such as Sentence Piece (Kudo and Richardson, 2018), and the special tokens $$ and $$ are used as target word delimiters (Xie et al., 2021):", + "bbox": [ + 507, + 261, + 885, + 455 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\ns _ {1} = w _ {1}, \\dots , < t >, w _ {i} ^ {t}, \\dots , w _ {i + k} ^ {t}, < / t >, \\dots , w _ {N} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 461, + 880, + 487 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\ns _ {2} = w _ {1}, \\dots , < \\mathsf {t} >, w _ {j} ^ {t}, \\dots , w _ {j + p} ^ {t}, < / \\mathsf {t} >, \\dots , w _ {M}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 482, + 845, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $N$ and $M$ represent the number of subwords of the sequence $s_1$ and $s_2$ respectively, while $w_i^t,\\ldots ,w_{i + k}^t$ and $w_j^t,\\ldots ,w_{j + p}^t$ are the subwords of the target words. In the following, we describe the baseline cross-encoder and XLLEXEME based on a bi-encoder. For the cross-encoder, the two input sequences are concatenated by the special token [SEP] in an overall sequence $s = [CLS] s_1[SEP] s_2[SEP]$ . If the length of $s$ , i.e. $N + M + 3$ , is greater than the maximum sequence length $\\lambda$ , then the sequence $s$ is cut such that the length of $s_1$ and $s_2$ is less than $\\lambda^{*} = \\frac{\\lambda - 3}{2}$ . To comply with the maximum length, the left and right contexts of the sequence are truncated. For instance, $s_1$ is truncated as follows:", + "bbox": [ + 507, + 507, + 885, + 747 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\ns _ {1} = w _ {n _ {0}}, \\dots , < \\mathrm {t} >, w _ {i} ^ {t}, \\dots , w _ {i + k} ^ {t}, < / \\mathrm {t} >, \\dots , w _ {n _ {1}} (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 757, + 880, + 775 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $n_0 = \\max (0,i - 1 - \\frac{\\lambda^* - k - 2}{2})$ and $n_1 = \\min (N,i + k + 1 + \\frac{\\lambda^* - k - 2}{2})$ . The truncated sequence has a length $\\gamma < \\lambda$ . The encoded representations of each subword $(v_{1},v_{2},\\ldots ,v_{\\gamma})$ are", + "bbox": [ + 507, + 782, + 884, + 850 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1The XL-LEXEME code is available on GitHub https://github.com/pierluigic/xl-lexeme. The XL-LEXEME model is available in the Hugging Face Model Hub https://huggingface.co/ pierluigic/xl-lexeme.", + "bbox": [ + 507, + 856, + 884, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1578", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "summed to get the encoded representation of the overall sequence, i.e. $s^{enc} = \\sum_{i}^{\\gamma} v_{i}$ . Finally, the vector $s^{enc}$ is used to compute the logits:", + "bbox": [ + 112, + 84, + 487, + 131 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {l o g i t} = \\log \\sigma (W s ^ {\\text {e n c}}) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 143, + 485, + 160 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $W\\in \\mathbb{R}^{1\\times d}$ . The model is trained to minimize the Binary Cross-entropy loss function.", + "bbox": [ + 112, + 170, + 487, + 203 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "XL-LEXEME is a bi-encoder that encodes the input sequences using a Siamese Network into two different vector representations. Each sequence is tokenized and truncated according to the maximum length $\\lambda^{*}$ , using Equation (2). We thus obtain the new lengths $\\gamma_{1},\\gamma_{2}$ . The vector representation is computed as the sum of the encoded subwords $(v_{1},v_{2},\\dots,v_{\\gamma})$ , i.e. $s_1^{enc} = \\sum_i^{\\gamma_1}v_i$ and $s_2^{enc} = \\sum_j^{\\gamma_2}v_j$ .", + "bbox": [ + 112, + 204, + 487, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "XL-LEXEME is trained to minimize the Contrastive loss (Hadsell et al., 2006):", + "bbox": [ + 112, + 349, + 489, + 380 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\ell = \\frac {1}{2} [ y \\cdot \\delta^ {2} + (1 - y) \\cdot \\max (0, m - \\delta) ^ {2} ] \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 388, + 487, + 419 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where we adopt a margin $m = 0.5$ . We use as default distance $\\delta$ the cosine distance between the encoded representations of $s_1$ and $s_2$ , i.e. $\\delta = \\cos(s_1^{enc}, s_2^{enc})$ . The main advantage of XL-LEXEME concerning models based on the cross-encoder architecture is efficiency. The time cost can be directly derived from the different architectures that exploit XL-LEXEME and the cross-encoder baseline. The self-attention time complexity $O(N^2 * d)$ depends on the vector dimension $d$ and the sequence length, which is $N$ for the cross-encoder and $\\frac{N}{2}$ for XL-LEXEME. For XL-LEXEME, the time complexity is reduced to $O((\\frac{N}{2})^2 * 2d)$ .", + "bbox": [ + 112, + 428, + 489, + 653 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4 Experimental setting", + "text_level": 1, + "bbox": [ + 112, + 664, + 331, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1 Lexical Semantic Change Detection", + "text_level": 1, + "bbox": [ + 112, + 689, + 438, + 703 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SemEval-2020 Task 1: Unsupervised Lexical Semantic Change Detection (Schlechtweg et al., 2020) is the first task on Unsupervised Lexical Semantic Change Detection in English, German, Swedish, and Latin languages. For each language, two corpora represent two different periods (T0, T1). Moreover, a set of target words, annotated using the DUREL framework (Schlechtweg et al., 2018), are provided. SemEval-2020 Task 1 involves two subtasks. The binary classification task requires assigning a label (changed/stable) to each target word. The ranking task sorts the target words according to their degree of semantic change. In", + "bbox": [ + 112, + 709, + 489, + 917 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "this work, we focus on Subtask 2, and for the sake of simplicity, we refer to SemEval-2020 Task 1 Subtask 2 as SemEval-2020 Task 1.", + "bbox": [ + 507, + 84, + 880, + 131 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "RuShiftEval, different from SemEval-2020 Task 1, involves three sub-corpora extracted from the Russian National Corpus spanning three periods. Models are evaluated on the resulting three test sets, namely RuShiftEval1 (pre-Soviet and Soviet), RuShiftEval2 (Soviet and post-Soviet), and RuShiftEval3 (pre-Soviet and post-Soviet). RuShiftEval provides participants with development data that can be used for tuning models. RuShiftEval aims to corroborate if training data can improve LSC Detection models. The development data rely on the RuSemShift dataset (Rodina and Kutuzov, 2020), which includes two sets of 70 target words for the pre-Soviet to Soviet period and Soviet to post-Soviet period, respectively. The dataset also includes annotated pairwise sentences, which can be used for training the models.", + "bbox": [ + 507, + 133, + 884, + 406 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.2 Training details", + "text_level": 1, + "bbox": [ + 507, + 420, + 680, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "XL-LEXEME and the cross-encoder are trained using XLM-RoBERTa (XLM-R) (Conneau et al., 2020a) large as the underlying Language Model $^2$ and using an NVIDIA GeForce RTX 3090. As for training data, the model uses the training data of MCL-WiC (Martelli et al., 2021), $\\mathrm{AM}^2\\mathrm{ICO}$ (Liu et al., 2021), and XL-WiC datasets (Raganato et al., 2020) merged with the randomly sampled $75\\%$ of the respective development data of each dataset. The remaining $25\\%$ of the development data is used to fine-tune hyper-parameters. Moreover, we augment training data for the cross-encoder by swapping the order of sentences in the training set (Martelli et al., 2021).", + "bbox": [ + 507, + 441, + 882, + 665 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We use AdamW optimizer and linear learning warm-up over the $10\\%$ of training data. We perform a grid search for the hyper-parameters optimization, tuning the learning rate in $\\{1\\mathrm{e} - 6,2\\mathrm{e} - 6,$ $5\\mathrm{e} - 6,1\\mathrm{e} - 5,2\\mathrm{e} - 5\\}$ and the weight decay $\\{0.0,0.01\\}$ . Table 3 (Appendix A) shows the selected hyperparameters. We sample 200 sentences containing the target word for each language and each period. The sampling is repeated ten times, and the results are averaged over the ten iterations. We use the same methodology of Rachinskiy and Arefyev (2021) for sampling sentences from the RuShiftEval corpora. We sample sentences in which we find the exact match with the target words with no pre", + "bbox": [ + 507, + 668, + 884, + 892 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "2The XLM-R model is fine-tuned during the training.", + "bbox": [ + 529, + 903, + 857, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1579", + "bbox": [ + 482, + 927, + 519, + 940 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "processing of the SemEval dataset. The LSC score is computed as the average distance between the vectors over the two different periods:", + "bbox": [ + 112, + 84, + 487, + 134 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {L S C} (s ^ {t _ {0}}, s ^ {t _ {1}}) = \\frac {1}{N \\cdot M} \\sum_ {i = 0} ^ {N} \\sum_ {j = 0} ^ {M} \\delta (s _ {i} ^ {t _ {0}}, s _ {j} ^ {t _ {1}}) \\quad (5)\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 145, + 487, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\delta$ is the distance measure, i.e. $\\delta = 1 - \\log \\sigma (W s^{enc})$ for the cross-encoder baseline and $\\delta = \\cos (s_1^{enc},s_2^{enc})$ for XL-LEXEME.", + "bbox": [ + 112, + 202, + 489, + 252 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5 Results", + "text_level": 1, + "bbox": [ + 112, + 263, + 213, + 279 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1 and Table 2 report the results on the SemEval-2020 Task 1 Subtask 2 and the results on the RuShiftEval test set. The results of the best systems are in bold. XL-LEXEME achieve the best score for English, German, Swedish, RuShiftEval1, RuShiftEval2, and RuShiftEval3. XL-LEXEME achieves a strong Spearman correlation for English and Swedish languages and a solid correlation on the German dataset, obtaining a significant correlation $(p < 0.001)$ . XL-LEXEME obtains no significant results in the Latin language since the predicted scores for the target words are not correlated with the test set. Latin is underrepresented in the training data of XLM-R, and there are no similar languages in the WiC dataset that we use for training XL-LEXEME. Moreover, the Latin dataset is more challenging as it involves the first corpus written in ancient Latin, which differs in many aspects from modern Latin. For this reason, XL-LEXEME could be ineffective in ancient languages and, in general, in languages that are not widely covered by the WiC dataset.", + "bbox": [ + 112, + 290, + 489, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We report the statistical significance of the difference between the performance of XL-LEXEME concerning the other models. The statistical significance of the difference is computed using Fisher's $z$ -transformation (Press, 2002). XL-LEXEME obtains stronger correlations than the cross-encoder, but the differences are not significant. The correlations obtained on the English and the German datasets are significantly different $(p < 0.05)$ for all the systems that participated in the SemEval2020 Task 1 but not for TempoBERT and Temporal Attention. On the other side, TempoBERT and Temporal Attention obtain a Spearman correlation on English and German that is not statistically different from the systems on the SemEval-2020 Task 1 leaderboard. In the Swedish language, XL-LEXEME is the only one obtaining a significantly", + "bbox": [ + 112, + 645, + 490, + 920 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "different correlation from the Count baseline results. XL-LEXEME showed its effectiveness also in Swedish, although the WiC dataset does not cover this language. Presumably, Swedish benefits from the presence of other languages descending from the Old Norse language, namely Danish and Norwegian.", + "bbox": [ + 507, + 84, + 884, + 197 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "XL-LEXEME obtains competitive results for the Russian language in the RuShiftEval leaderboard. Contrary to XL-LEXEME, Deep Mistake and Gloss Reader are fine-tuned on the RuSemShift dataset. The differences between XL-LEXEME and the best two systems in the leaderboard are not statically significant. Moreover, in Table 2, the results of XL-LEXEME fine-tuned on the RuSemShift are shown. Although the fine-tuned model achieves the best correlation scores in the three datasets, the difference between DeepMistake and GlossReader is not significant.", + "bbox": [ + 507, + 198, + 885, + 391 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 507, + 403, + 640, + 418 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this work, we introduced XL-LEXEME, a model for LSC Detection. XL-LEXEME is pre-trained on a large WiC dataset to mirror sentence-level encoders focusing on specific words in contexts. We evaluated our model on two Lexical Semantic Change Detection datasets: SemEval-2020 Task 1 and RuShiftEval. XL-LEXEME outperforms state-of-the-art models for LSC Detection in English, German, Swedish, and Russian datasets, with significant differences from the baselines. The XL-LEXEME effectiveness and efficiency make it reliable for LSC Detection on large diachronic corpora.", + "bbox": [ + 507, + 429, + 885, + 639 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "7 Limitations", + "text_level": 1, + "bbox": [ + 507, + 651, + 645, + 665 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While the vector representations obtained using XL-LEXEME for different languages are potentially comparable, lying on the same geometric space, the evaluation of cross-lingual semantic changes cannot be performed for lacking cross-lingual LSC Detection resources. SemEval 2020 Task 1 datasets consist of small sets of target words, i.e., the number of target words for English, German, Latin, and Swedish is 37, 48, 40, and 31, respectively. The example of the Latin language highlights that XL-LEXEME can perform poorly on languages that are underrepresented in the training set of XLM-R and not covered by the WiC dataset. Generally, at the moment is not possible to state precisely how and how much XL-LEXEME", + "bbox": [ + 507, + 677, + 884, + 919 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1580", + "bbox": [ + 482, + 927, + 521, + 940 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/963ca8017937b1e0e0fe220ddd28b5d73c7fc59709c60f69f742179a8b492a25.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SemEval-2020 Task 1 Subtask 2 LeaderboardTemporal BERTcross-encoderXL-LEXEME
Lang.UG_Student _InternJiaxin & Jinancs2020UWBCount baselineFreq. baselineTempoBERTTemporal Attention
EN0.4220.3250.3750.3670.022-0.2170.467†0.520†0.7520.757
DE0.7250.7170.7020.6970.2160.014-†0.763†0.8370.877
SV†0.547†0.588†0.536†0.604-0.022-0.150--†0.6800.754
LA0.4120.4400.3990.2540.359†0.0200.5120.565†0.016-0.056
Avg.0.5270.5180.5030.4810.144-0.083--0.5710.583
", + "bbox": [ + 115, + 80, + 884, + 200 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/48a7a89adc54cd27695d77cd0ec3e1a155e0fdab7c86034fddc91ea32cd000c0.jpg", + "table_caption": [ + "Table 1: Results (Spearman correlation) on the SemEval-2020 Task 1 Subtask 2 test set. The symbol $\\dagger$ indicates there is no statistical difference with the correlation obtained by XL-LEXEME." + ], + "table_footnote": [], + "table_body": "
RuShiftEval Leaderboardcross-encoderXL-LEXEMEXL-LEXEME (Fine-tuned)
DatasetGlossReaderDeepMistakeUWBBaseline
RuShiftEval1†0.781†0.7980.3620.314†0.7270.7750.799
RuShiftEval2†0.803†0.7730.3540.302†0.7530.8220.833
RuShiftEval3†0.822†0.8030.5330.381†0.7480.8090.842
Avg.0.8020.7910.4170.3320.7430.8020.825
", + "bbox": [ + 115, + 250, + 884, + 332 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2: Results (Spearman correlation) on the RuShiftEval test set. The symbol $\\dagger$ indicates there is no statistical difference with the correlation obtained by XL-LEXEME.", + "bbox": [ + 112, + 341, + 882, + 370 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "performance is affected by the language distribution in the XLM-R training set and the WiC dataset.", + "bbox": [ + 112, + 395, + 489, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 114, + 439, + 285, + 457 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We acknowledge the support of the PNRR project FAIR - Future AI Research (PE00000013), Spoke 6 - Symbiotic AI (CUP H97G22000210007) under the NRRP MUR program funded by the NextGenerationEU.", + "bbox": [ + 112, + 464, + 489, + 545 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This work has in part been funded by the research program Change is Key! supported by Riksbankens Jubileumsfond (under reference number M21-0021).", + "bbox": [ + 112, + 546, + 489, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 114, + 637, + 213, + 653 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Nikolay Arefyev, Daniil Homskiy, Maksim Fedoseev, Adis Davletov, Vitaly Protasov, and Alexander Panchenko. 2021. DeepMistake: Which Senses are Hard to Distinguish for a WordinContext Model. In Computational Linguistics and Intellectual Technologies - Papers from the Annual International Conference \"Dialogue\" 2021, volume 2021-June. Section: 20.", + "Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettle-moyer, and Veselin Stoyanov. 2020a. Unsupervised Cross-lingual Representation Learning at Scale. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 8440-8451. Association for Computational Linguistics.", + "Alexis Conneau, Kartikay Khandelwal, Naman Goyal," + ], + "bbox": [ + 115, + 659, + 489, + 919 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2020b. Unsupervised Cross-lingual Representation Learning at Scale. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 8440-8451. Association for Computational Linguistics.", + "Raia Hadsell, Sumit Chopra, and Yann LeCun. 2006. Dimensionality reduction by learning an invariant mapping. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2006), 17-22 June 2006, New York, NY, USA, pages 1735-1742. IEEE Computer Society.", + "William L. Hamilton, Jure Leskovec, and Dan Jurafsky. 2016. Diachronic Word Embeddings Reveal Statistical Laws of Semantic Change. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1489-1501, Berlin, Germany. Association for Computational Linguistics.", + "David R. Hardoon, Sandor Szedmak, and John Shawe-Taylor. 2004. Canonical Correlation Analysis: An Overview with Application to Learning Methods. Neural Computation, 16(12):2639-2664.", + "Taku Kudo and John Richardson. 2018. SentencePiece: A simple and language independent subword tokenizer and tokenizer for Neural Text Processing. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 66-71, Brussels, Belgium. Association for Computational Linguistics.", + "Severin Laicher, Sinan Kurtyigit, Dominik Schlechtweg, Jonas Kuhn, and Sabine Schulte im Walde. 2021. Explaining and improving BERT performance on lexical semantic change detection. In Proceedings of" + ], + "bbox": [ + 510, + 395, + 884, + 917 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1581", + "bbox": [ + 482, + 928, + 517, + 940 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the 16th Conference of the European Chapter of the Association for Computational Linguistics: Student Research Workshop, pages 192-202, Online. Association for Computational Linguistics.", + "Qianchu Liu, Edoardo Maria Ponti, Diana McCarthy, Ivan Vulic, and Anna Korhonen. 2021. AM2iCo: Evaluating Word Meaning in Context across Low-Resource Languages with Adversarial Examples. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, EMNLP 2021, Virtual Event / Punta Cana, Dominican Republic, 7-11 November, 2021, pages 7151-7162. Association for Computational Linguistics.", + "Federico Martelli, Najla Kalach, Gabriele Tola, and Roberto Navigli. 2021. SemEval-2021 Task 2: Multilingual and Cross-lingual Word-in-Context Disambiguation (MCL-WiC). In Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 24–36, Online. Association for Computational Linguistics.", + "Tomás Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. Efficient estimation of word representations in vector space. In 1st International Conference on Learning Representations, ICLR 2013, Workshop Track Proceedings.", + "George A. Miller. 1992. WordNet: A Lexical Database for English. In Speech and Natural Language: Proceedings of a Workshop Held at Harriman, New York, February 23-26, 1992.", + "George A. Miller, Martin Chodorow, Shari Landes, Claudia Leacock, and Robert G. Thomas. 1994. Using a Semantic Concordance for Sense Identification. In Human Language Technology, Proceedings of a Workshop held at Plainsboro, New Jerey, USA, March 8-11, 1994. Morgan Kaufmann.", + "Stefano Montanelli and Francesco Periti. 2023. A survey on contextualised semantic shift detection. arXiv preprint arXiv:2304.01666.", + "Roberto Navigli. 2009. Word Sense Disambiguation: A Survey. ACM Comput. Surv., 41(2).", + "Mohammad Taher Pilehvar and José Camacho-Collados. 2019. WiC: the Word-in-Context Dataset for Evaluating Context-Sensitive Meaning Representations. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 1267-1273. Association for Computational Linguistics.", + "Ondrej Prazák, Pavel Pribán, and Stephen Taylor. 2021. UWB@ RuShiftEval Measuring Semantic Difference as per-word Variation in Aligned Semantic Spaces. In Computational Linguistics and Intellectual Technologies - Papers from the Annual International Conference \"Dialogue\" 2021, volume 2021-June. Section: 20." + ], + "bbox": [ + 115, + 85, + 489, + 917 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Ondrej Prazák, Pavel Pribán, Stephen Taylor, and Jakub Sido. 2020. UWB at SemEval-2020 Task 1: Lexical Semantic Change Detection. In Proceedings of the Fourteenth Workshop on Semantic Evaluation, SemEval@COLING2020, pages 246–254. International Committee for Computational Linguistics.", + "William H. Press. 2002. Numerical recipes in $C++$ : the art of scientific computing, 2nd Edition (C++ ed., print. is corrected to software version 2.10). Cambridge University Press.", + "Maxim Rachinskiy and Nikolay Arefyev. 2021. Zeroshot Crosslingual Transfer of a Gloss Language Model for Semantic Change Detection. In Computational Linguistics and Intellectual Technologies - Papers from the Annual International Conference \"Dialogue\" 2021, volume 2021-June. Section: 20.", + "Alessandro Raganato, Tommaso Pasini, José Camacho-Collados, and Mohammad Taher Pilehvar. 2020. XL-WiC: A Multilingual Benchmark for Evaluating Semantic Contextualization. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online, November 16-20, 2020, pages 7193-7206. Association for Computational Linguistics.", + "Nils Reimers and Iryna Gurevych. 2019. SentenceBERT: Sentence Embeddings using Siamese BERT-Networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3982-3992, Hong Kong, China. Association for Computational Linguistics.", + "Julia Rodina and Andrey Kutuzov. 2020. RuSemShift: a dataset of historical lexical semantic change in Russian. In Proceedings of the 28th International Conference on Computational Linguistics, pages 1037-1047, Barcelona, Spain (Online). International Committee on Computational Linguistics.", + "Guy D. Rosin, Ido Guy, and Kira Radinsky. 2022. Time Masking for Temporal Language Models. In WSDM '22: The Fifteenth ACM International Conference on Web Search and Data Mining, Virtual Event / Tempe, AZ, USA, February 21 - 25, 2022, pages 833-841. ACM.", + "Guy D. Rosin and Kira Radinsky. 2022. Temporal Attention for Language Models. CoRR, abs/2202.02093.", + "Dominik Schlechtweg, Barbara McGillivray, Simon Hengchen, Haim Dubossarsky, and Nina Tahmasebi. 2020. SemEval-2020 Task 1: Unsupervised Lexical Semantic Change Detection. In Proceedings of the Fourteenth Workshop on Semantic Evaluation, SemEval@COLING2020, pages 1-23. International Committee for Computational Linguistics.", + "Dominik Schlechtweg, Sabine Schulte im Walde, and Stefanie Eckmann. 2018. Diachronic Usage Relatedness (DUREl): A Framework for the Annotation" + ], + "bbox": [ + 510, + 85, + 882, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1582", + "bbox": [ + 482, + 928, + 521, + 940 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "of Lexical Semantic Change. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 169-174, New Orleans, Louisiana. Association for Computational Linguistics.", + "bbox": [ + 131, + 85, + 489, + 165 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Shuyi Xie, Jian Ma, Haiqin Yang, Lianxin Jiang, Yang Mo, and Jianping Shen. 2021. PALI at SemEval-2021 Task 2: Fine-Tune XLM-RoBERTa for Word in Context Disambiguation. In Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 713-718, Online. Association for Computational Linguistics.", + "bbox": [ + 114, + 174, + 489, + 267 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A Hyper-parameters", + "text_level": 1, + "bbox": [ + 114, + 279, + 314, + 294 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/acbc1d7372c2c7034e3f5877211fa20c541a0b7ceb2df51fd84539470e35f834.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Hyper-parameterValue
hidden actgelu
hidden dropout prob0.1
hidden size1024
initializer range0.02
intermediate size4096
layer norm eps1e-05
max position embeddings514
num attention heads16
num hidden layers24
position embedding typeabsolute
vocab size250004
learning rate
cross-encoder1e-05
XL-LEXEME1e-05
weight decay
cross-encoder0.01
XL-LEXEME0.00
max sequence length
cross-encoderλ = 256
XL-LEXEMEλ* = 128
", + "bbox": [ + 147, + 303, + 453, + 659 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3: XL-LEXEME and cross-encoder hyperparameters.", + "bbox": [ + 112, + 670, + 489, + 699 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1583", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A For every submission:", + "bbox": [ + 115, + 107, + 322, + 122 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "A1. Did you describe the limitations of your work? Section 7", + "A2. Did you discuss any potential risks of your work? Not applicable. Left blank.", + "A3. Do the abstract and introduction summarize the paper's main claims? Section 1", + "A4. Have you used AI writing assistants when working on this paper? Left blank." + ], + "bbox": [ + 127, + 126, + 695, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "B Did you use or create scientific artifacts?", + "bbox": [ + 114, + 299, + 489, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Section 3 and Section 4", + "bbox": [ + 132, + 321, + 310, + 335 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "B1. Did you cite the creators of artifacts you used?", + "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Section 4 and References", + "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Section 3 and Section 4", + "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Not applicable. Left blank.", + "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Section 3 and Section 4", + "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Section 4" + ], + "bbox": [ + 127, + 346, + 880, + 752 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "C Did you run computational experiments?", + "bbox": [ + 114, + 764, + 492, + 781 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Section 3 and Section 4", + "bbox": [ + 132, + 785, + 309, + 800 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Section 4 and Appendix A", + "bbox": [ + 129, + 810, + 880, + 860 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.", + "bbox": [ + 112, + 868, + 877, + 892 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "ACL 2023 Responsible NLP Checklist", + "bbox": [ + 132, + 84, + 433, + 99 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1584", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?", + "bbox": [ + 129, + 83, + 878, + 115 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Section 4", + "bbox": [ + 149, + 117, + 223, + 130 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?", + "bbox": [ + 129, + 142, + 882, + 190 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Section 4", + "bbox": [ + 149, + 192, + 223, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?", + "bbox": [ + 129, + 217, + 882, + 265 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Section 3", + "bbox": [ + 149, + 267, + 223, + 280 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?", + "bbox": [ + 112, + 293, + 877, + 309 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 132, + 313, + 213, + 329 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?", + "bbox": [ + 127, + 340, + 882, + 372 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 374, + 248, + 388 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?", + "bbox": [ + 127, + 399, + 882, + 447 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 449, + 248, + 463 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?", + "bbox": [ + 127, + 475, + 882, + 521 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 524, + 248, + 539 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?", + "bbox": [ + 127, + 549, + 873, + 564 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 565, + 248, + 581 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?", + "bbox": [ + 127, + 592, + 880, + 623 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 626, + 248, + 640 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1585", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 8 + } +] \ No newline at end of file diff --git a/2023/XL-LEXEME_ WiC Pretrained Model for Cross-Lingual LEXical sEMantic changE/4754c7cf-ef79-4237-89fc-dec2c7680fce_model.json b/2023/XL-LEXEME_ WiC Pretrained Model for Cross-Lingual LEXical sEMantic changE/4754c7cf-ef79-4237-89fc-dec2c7680fce_model.json new file mode 100644 index 0000000000000000000000000000000000000000..891da8cf69d7ff5e0f8a2dc34c9493df73cfb555 --- /dev/null +++ b/2023/XL-LEXEME_ WiC Pretrained Model for Cross-Lingual LEXical sEMantic changE/4754c7cf-ef79-4237-89fc-dec2c7680fce_model.json @@ -0,0 +1,1670 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.281, + 0.09, + 0.717, + 0.131 + ], + "angle": 0, + "content": "XL-LEXEME: WiC Pretrained Model for Cross-Lingual LEXical sEMantic change" + }, + { + "type": "text", + "bbox": [ + 0.187, + 0.15, + 0.816, + 0.183 + ], + "angle": 0, + "content": "Pierluigi Cassotti, Lucia Siciliani, Marco de Gemmis, Giovanni Semeraro and Pierpaolo Basile" + }, + { + "type": "text", + "bbox": [ + 0.378, + 0.184, + 0.626, + 0.199 + ], + "angle": 0, + "content": "University of Bari Aldo Moro" + }, + { + "type": "text", + "bbox": [ + 0.378, + 0.2, + 0.629, + 0.216 + ], + "angle": 0, + "content": "{firstname.lastname} @uniba.it" + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.253, + 0.341, + 0.268 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.279, + 0.461, + 0.576 + ], + "angle": 0, + "content": "The recent introduction of large-scale datasets for the WiC (Word in Context) task enables the creation of more reliable and meaningful contextualized word embeddings. However, most of the approaches to the WiC task use cross-encoders, which prevent the possibility of deriving comparable word embeddings. In this work, we introduce XL-LEXEME, a Lexical Semantic Change Detection model. XL-LEXEME extends SBERT, highlighting the target word in the sentence. We evaluate XL-LEXEME on the multilingual benchmarks for SemEval-2020 Task 1 - Lexical Semantic Change (LSC) Detection and the RuShiftEval shared task involving five languages: English, German, Swedish, Latin, and Russian. XL-LEXEME outperforms the state-of-the-art in English, German and Swedish with statistically significant differences from the baseline results and obtains state-of-the-art performance in the RuShiftEval shared task." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.589, + 0.396, + 0.604 + ], + "angle": 0, + "content": "1 Introduction and Motivation" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.614, + 0.488, + 0.854 + ], + "angle": 0, + "content": "Lexical Semantic Change (LSC) Detection is the task of automatically identifying words that change their meaning over time. The LSC Detection task implicitly aims to disambiguate synchronic word sense occurrences and then find differences in the word sense frequencies in different periods. Word Sense Disambiguation (WSD) is a long-studied task in Natural Language Processing (Navigli, 2009), which consists of associating the correct sense to a word occurring in a specific context. WSD involves some crucial issues, such as relying on a fixed sense inventory. Fixed sense inventories ignore the diachronic aspect of language because they can miss older unused senses or be outdated and missing new senses." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.855, + 0.487, + 0.919 + ], + "angle": 0, + "content": "The Word in Context task (WiC) (Pilehvar and Camacho-Collados, 2019) aims to overcome these issues. In this work, we train a model on the WiC task and then use it to perform LSC Detection. In" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.253, + 0.886, + 0.558 + ], + "angle": 0, + "content": "the WiC task, given the word \\( w \\) and two different contexts \\( C1 \\), \\( C2 \\), the systems have to determine whether the meaning of \\( w \\) is the same in the two contexts or not. Our approach is grounded on the assumption that models trained on the WiC tasks are robust enough to transfer the knowledge learned in a synchronic setting to a diachronic one. We summarise the main contribution of this work as follows: (i) We propose a pre-trained bi-encoder model, called XL-LEXEME, on a largescale dataset for the WiC task, which allows us to obtain comparable lexical-based representations; (ii) We assert the effectiveness of XL-LEXEME despite the computational limitation compared to the cross-encoder architecture for the LSC Detection task; (iii) Experiments on the LSC Detection task show that XL-LEXEME outperforms state-of-the-art LSC Detection models for English, German, Swedish, and Russian." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.571, + 0.665, + 0.587 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.597, + 0.884, + 0.852 + ], + "angle": 0, + "content": "LSC Detection systems can be categorized based on the distributional embeddings used to tackle the LSC Detection task. One category is represented by those approaches that adopt type-base (i.e., static) embeddings. UWB (Prazák et al., 2020; Prazák et al., 2021) represents an example of this category of systems. First, it employs word2vec Skip-gram with Negative Sampling (Mikolov et al., 2013) to compute a semantic space for each corpus. It uses techniques like the Canonical Correlation Analysis (Hardoon et al., 2004) and the Orthogonal Transformation (Hamilton et al., 2016) to align the abovementioned spaces. Therefore, the cosine similarity between the vectors representing the word in two different spaces is used to detect the semantic shift." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.855, + 0.884, + 0.919 + ], + "angle": 0, + "content": "With the increasing use of contextualized word embeddings, numerous approaches employing BERT-base models have been developed for LSC Detection (Montanelli and Periti, 2023; Laicher" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1577" + }, + { + "type": "footer", + "bbox": [ + 0.227, + 0.946, + 0.771, + 0.958 + ], + "angle": 0, + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + }, + { + "type": "footer", + "bbox": [ + 0.369, + 0.959, + 0.63, + 0.972 + ], + "angle": 0, + "content": "Volume 2: Short Papers, pages 1577-1585" + }, + { + "type": "footer", + "bbox": [ + 0.297, + 0.973, + 0.702, + 0.986 + ], + "angle": 0, + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.112, + 0.085, + 0.493, + 0.422 + ], + "angle": 0, + "content": "et al., 2021). In TempoBERT (Rosin et al., 2022), the authors exploit the concept of Masked Language Modeling (MLM), where the goal is to train a language model to predict a masked portion of text given the remaining part. In particular, they employ this technique to encode the concept of time into a BERT model. This is done by concatenating a specific token representing time to the text sequence. At inference time, TempoBERT can be used to predict the year of a sentence, masking the time reference, or to predict a masked token of the sentence conditioned by the time reference. In the same line of research, in Temporal Attention (Rosin and Radinsky, 2022), the authors investigate the effect of modifying the model instead of the input sentence like in TempoBERT. This is done by extending the model's attention mechanism to consider the time when computing the weight of each word. The time dimension is encoded using a different query embedding matrix for each timestamp." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.424, + 0.492, + 0.745 + ], + "angle": 0, + "content": "Another kind of approach exploits the information coming from other tasks to perform LSC Detection. GlossReader represents an example (Rachinskiy and Arefyev, 2021), where a model based on XML-R (Conneau et al., 2020b) is first trained on English SemCor (Miller et al., 1994) with glosses from WordNet 3.0 (Miller, 1992) to perform WSD. Exploiting the zero-shot cross-lingual characteristics of XML-R, the authors used the same model to perform LSC Detection in the Russian language. With DeepMistake (Arefyev et al., 2021), the authors take advantage of the WiC task instead of WSD. They train a cross-encoder with XML-R as an underlying Language Model on the MCL-WiC training and development set and fine-tune on the RuSemShift dataset (Rodina and Kutuzov, 2020). DeepMistake, differently from XL-LEXEME, relies on the cross-encoder architecture and exploits only the MCL-WiC training dataset." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.762, + 0.273, + 0.779 + ], + "angle": 0, + "content": "3 XL-LEXEME" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.791, + 0.491, + 0.92 + ], + "angle": 0, + "content": "Generally, for pairwise sentence similarity tasks, BERT models use a cross-encoder, in which the pairwise sequences are jointly encoded, and the overall vectors are used for the classification. However, in several tasks, the cross-encoder is not suitable since it cannot provide a distinct meaningful representation for each sentence. An approach to overcome this issue involves pooling the BERT out" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.261 + ], + "angle": 0, + "content": "put encoded vectors, which often results in worse performance. Sentence-BERT (SBERT) (Reimers and Gurevych, 2019) overcomes the limitation of cross-encoders using a Siamese Network, i.e., the weights of the underlying networks are shared. SBERT encodes the two sequences separately in the BERT model exploiting the Siamese architecture. The sequence-level representation is obtained by averaging the output encoded vectors, which are directly compared using similarity measures such as cosine similarity." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.262, + 0.887, + 0.456 + ], + "angle": 0, + "content": "Meanwhile, cross-encoders perform better since they are trained to profit from the attention over the whole input. In this work, we introduce XLLEXEME1 which mirrors models for pairwise sequence similarity tasks and adapts them to the WiC task, giving prominence to the target word, i.e. the word for which we want to detect the LSC. The model takes as input two sequences \\( s_1 \\) and \\( s_2 \\). The sequences are tokenized using subwords tokenizer, such as Sentence Piece (Kudo and Richardson, 2018), and the special tokens \\( \\) and \\( \\) are used as target word delimiters (Xie et al., 2021):" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.462, + 0.882, + 0.488 + ], + "angle": 0, + "content": "\\[\ns _ {1} = w _ {1}, \\dots , < t >, w _ {i} ^ {t}, \\dots , w _ {i + k} ^ {t}, < / t >, \\dots , w _ {N} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.483, + 0.846, + 0.502 + ], + "angle": 0, + "content": "\\[\ns _ {2} = w _ {1}, \\dots , < \\mathsf {t} >, w _ {j} ^ {t}, \\dots , w _ {j + p} ^ {t}, < / \\mathsf {t} >, \\dots , w _ {M}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.508, + 0.886, + 0.749 + ], + "angle": 0, + "content": "where \\(N\\) and \\(M\\) represent the number of subwords of the sequence \\(s_1\\) and \\(s_2\\) respectively, while \\(w_i^t,\\ldots ,w_{i + k}^t\\) and \\(w_j^t,\\ldots ,w_{j + p}^t\\) are the subwords of the target words. In the following, we describe the baseline cross-encoder and XLLEXEME based on a bi-encoder. For the cross-encoder, the two input sequences are concatenated by the special token [SEP] in an overall sequence \\(s = [CLS] s_1[SEP] s_2[SEP]\\). If the length of \\(s\\), i.e. \\(N + M + 3\\), is greater than the maximum sequence length \\(\\lambda\\), then the sequence \\(s\\) is cut such that the length of \\(s_1\\) and \\(s_2\\) is less than \\(\\lambda^{*} = \\frac{\\lambda - 3}{2}\\). To comply with the maximum length, the left and right contexts of the sequence are truncated. For instance, \\(s_1\\) is truncated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.758, + 0.882, + 0.776 + ], + "angle": 0, + "content": "\\[\ns _ {1} = w _ {n _ {0}}, \\dots , < \\mathrm {t} >, w _ {i} ^ {t}, \\dots , w _ {i + k} ^ {t}, < / \\mathrm {t} >, \\dots , w _ {n _ {1}} (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.783, + 0.885, + 0.851 + ], + "angle": 0, + "content": "where \\(n_0 = \\max (0,i - 1 - \\frac{\\lambda^* - k - 2}{2})\\) and \\(n_1 = \\min (N,i + k + 1 + \\frac{\\lambda^* - k - 2}{2})\\). The truncated sequence has a length \\(\\gamma < \\lambda\\). The encoded representations of each subword \\((v_{1},v_{2},\\ldots ,v_{\\gamma})\\) are" + }, + { + "type": "page_footnote", + "bbox": [ + 0.508, + 0.857, + 0.885, + 0.919 + ], + "angle": 0, + "content": "1The XL-LEXEME code is available on GitHub https://github.com/pierluigic/xl-lexeme. The XL-LEXEME model is available in the Hugging Face Model Hub https://huggingface.co/ pierluigic/xl-lexeme." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1578" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.489, + 0.133 + ], + "angle": 0, + "content": "summed to get the encoded representation of the overall sequence, i.e. \\( s^{enc} = \\sum_{i}^{\\gamma} v_{i} \\). Finally, the vector \\( s^{enc} \\) is used to compute the logits:" + }, + { + "type": "equation", + "bbox": [ + 0.216, + 0.144, + 0.487, + 0.161 + ], + "angle": 0, + "content": "\\[\n\\operatorname {l o g i t} = \\log \\sigma (W s ^ {\\text {e n c}}) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.171, + 0.489, + 0.204 + ], + "angle": 0, + "content": "where \\(W\\in \\mathbb{R}^{1\\times d}\\). The model is trained to minimize the Binary Cross-entropy loss function." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.205, + 0.489, + 0.35 + ], + "angle": 0, + "content": "XL-LEXEME is a bi-encoder that encodes the input sequences using a Siamese Network into two different vector representations. Each sequence is tokenized and truncated according to the maximum length \\(\\lambda^{*}\\), using Equation (2). We thus obtain the new lengths \\(\\gamma_{1},\\gamma_{2}\\). The vector representation is computed as the sum of the encoded subwords \\((v_{1},v_{2},\\dots,v_{\\gamma})\\), i.e. \\(s_1^{enc} = \\sum_i^{\\gamma_1}v_i\\) and \\(s_2^{enc} = \\sum_j^{\\gamma_2}v_j\\)." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.35, + 0.49, + 0.381 + ], + "angle": 0, + "content": "XL-LEXEME is trained to minimize the Contrastive loss (Hadsell et al., 2006):" + }, + { + "type": "equation", + "bbox": [ + 0.129, + 0.39, + 0.488, + 0.42 + ], + "angle": 0, + "content": "\\[\n\\ell = \\frac {1}{2} [ y \\cdot \\delta^ {2} + (1 - y) \\cdot \\max (0, m - \\delta) ^ {2} ] \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.429, + 0.49, + 0.655 + ], + "angle": 0, + "content": "where we adopt a margin \\(m = 0.5\\). We use as default distance \\(\\delta\\) the cosine distance between the encoded representations of \\(s_1\\) and \\(s_2\\), i.e. \\(\\delta = \\cos(s_1^{enc}, s_2^{enc})\\). The main advantage of XL-LEXEME concerning models based on the cross-encoder architecture is efficiency. The time cost can be directly derived from the different architectures that exploit XL-LEXEME and the cross-encoder baseline. The self-attention time complexity \\(O(N^2 * d)\\) depends on the vector dimension \\(d\\) and the sequence length, which is \\(N\\) for the cross-encoder and \\(\\frac{N}{2}\\) for XL-LEXEME. For XL-LEXEME, the time complexity is reduced to \\(O((\\frac{N}{2})^2 * 2d)\\)." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.665, + 0.332, + 0.681 + ], + "angle": 0, + "content": "4 Experimental setting" + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.69, + 0.439, + 0.705 + ], + "angle": 0, + "content": "4.1 Lexical Semantic Change Detection" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.71, + 0.49, + 0.919 + ], + "angle": 0, + "content": "SemEval-2020 Task 1: Unsupervised Lexical Semantic Change Detection (Schlechtweg et al., 2020) is the first task on Unsupervised Lexical Semantic Change Detection in English, German, Swedish, and Latin languages. For each language, two corpora represent two different periods (T0, T1). Moreover, a set of target words, annotated using the DUREL framework (Schlechtweg et al., 2018), are provided. SemEval-2020 Task 1 involves two subtasks. The binary classification task requires assigning a label (changed/stable) to each target word. The ranking task sorts the target words according to their degree of semantic change. In" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.882, + 0.132 + ], + "angle": 0, + "content": "this work, we focus on Subtask 2, and for the sake of simplicity, we refer to SemEval-2020 Task 1 Subtask 2 as SemEval-2020 Task 1." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.134, + 0.885, + 0.407 + ], + "angle": 0, + "content": "RuShiftEval, different from SemEval-2020 Task 1, involves three sub-corpora extracted from the Russian National Corpus spanning three periods. Models are evaluated on the resulting three test sets, namely RuShiftEval1 (pre-Soviet and Soviet), RuShiftEval2 (Soviet and post-Soviet), and RuShiftEval3 (pre-Soviet and post-Soviet). RuShiftEval provides participants with development data that can be used for tuning models. RuShiftEval aims to corroborate if training data can improve LSC Detection models. The development data rely on the RuSemShift dataset (Rodina and Kutuzov, 2020), which includes two sets of 70 target words for the pre-Soviet to Soviet period and Soviet to post-Soviet period, respectively. The dataset also includes annotated pairwise sentences, which can be used for training the models." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.421, + 0.681, + 0.435 + ], + "angle": 0, + "content": "4.2 Training details" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.442, + 0.884, + 0.667 + ], + "angle": 0, + "content": "XL-LEXEME and the cross-encoder are trained using XLM-RoBERTa (XLM-R) (Conneau et al., 2020a) large as the underlying Language Model\\(^2\\) and using an NVIDIA GeForce RTX 3090. As for training data, the model uses the training data of MCL-WiC (Martelli et al., 2021), \\(\\mathrm{AM}^2\\mathrm{ICO}\\) (Liu et al., 2021), and XL-WiC datasets (Raganato et al., 2020) merged with the randomly sampled \\(75\\%\\) of the respective development data of each dataset. The remaining \\(25\\%\\) of the development data is used to fine-tune hyper-parameters. Moreover, we augment training data for the cross-encoder by swapping the order of sentences in the training set (Martelli et al., 2021)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.669, + 0.885, + 0.894 + ], + "angle": 0, + "content": "We use AdamW optimizer and linear learning warm-up over the \\(10\\%\\) of training data. We perform a grid search for the hyper-parameters optimization, tuning the learning rate in \\(\\{1\\mathrm{e} - 6,2\\mathrm{e} - 6,\\) \\(5\\mathrm{e} - 6,1\\mathrm{e} - 5,2\\mathrm{e} - 5\\}\\) and the weight decay \\(\\{0.0,0.01\\}\\). Table 3 (Appendix A) shows the selected hyperparameters. We sample 200 sentences containing the target word for each language and each period. The sampling is repeated ten times, and the results are averaged over the ten iterations. We use the same methodology of Rachinskiy and Arefyev (2021) for sampling sentences from the RuShiftEval corpora. We sample sentences in which we find the exact match with the target words with no pre" + }, + { + "type": "page_footnote", + "bbox": [ + 0.53, + 0.904, + 0.858, + 0.919 + ], + "angle": 0, + "content": "2The XLM-R model is fine-tuned during the training." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1579" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.488, + 0.135 + ], + "angle": 0, + "content": "processing of the SemEval dataset. The LSC score is computed as the average distance between the vectors over the two different periods:" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.146, + 0.488, + 0.192 + ], + "angle": 0, + "content": "\\[\n\\operatorname {L S C} (s ^ {t _ {0}}, s ^ {t _ {1}}) = \\frac {1}{N \\cdot M} \\sum_ {i = 0} ^ {N} \\sum_ {j = 0} ^ {M} \\delta (s _ {i} ^ {t _ {0}}, s _ {j} ^ {t _ {1}}) \\quad (5)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.203, + 0.49, + 0.253 + ], + "angle": 0, + "content": "where \\(\\delta\\) is the distance measure, i.e. \\(\\delta = 1 - \\log \\sigma (W s^{enc})\\) for the cross-encoder baseline and \\(\\delta = \\cos (s_1^{enc},s_2^{enc})\\) for XL-LEXEME." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.265, + 0.214, + 0.28 + ], + "angle": 0, + "content": "5 Results" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.291, + 0.49, + 0.644 + ], + "angle": 0, + "content": "Table 1 and Table 2 report the results on the SemEval-2020 Task 1 Subtask 2 and the results on the RuShiftEval test set. The results of the best systems are in bold. XL-LEXEME achieve the best score for English, German, Swedish, RuShiftEval1, RuShiftEval2, and RuShiftEval3. XL-LEXEME achieves a strong Spearman correlation for English and Swedish languages and a solid correlation on the German dataset, obtaining a significant correlation \\((p < 0.001)\\). XL-LEXEME obtains no significant results in the Latin language since the predicted scores for the target words are not correlated with the test set. Latin is underrepresented in the training data of XLM-R, and there are no similar languages in the WiC dataset that we use for training XL-LEXEME. Moreover, the Latin dataset is more challenging as it involves the first corpus written in ancient Latin, which differs in many aspects from modern Latin. For this reason, XL-LEXEME could be ineffective in ancient languages and, in general, in languages that are not widely covered by the WiC dataset." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.646, + 0.492, + 0.921 + ], + "angle": 0, + "content": "We report the statistical significance of the difference between the performance of XL-LEXEME concerning the other models. The statistical significance of the difference is computed using Fisher's \\(z\\)-transformation (Press, 2002). XL-LEXEME obtains stronger correlations than the cross-encoder, but the differences are not significant. The correlations obtained on the English and the German datasets are significantly different \\((p < 0.05)\\) for all the systems that participated in the SemEval2020 Task 1 but not for TempoBERT and Temporal Attention. On the other side, TempoBERT and Temporal Attention obtain a Spearman correlation on English and German that is not statistically different from the systems on the SemEval-2020 Task 1 leaderboard. In the Swedish language, XL-LEXEME is the only one obtaining a significantly" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.198 + ], + "angle": 0, + "content": "different correlation from the Count baseline results. XL-LEXEME showed its effectiveness also in Swedish, although the WiC dataset does not cover this language. Presumably, Swedish benefits from the presence of other languages descending from the Old Norse language, namely Danish and Norwegian." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.199, + 0.886, + 0.392 + ], + "angle": 0, + "content": "XL-LEXEME obtains competitive results for the Russian language in the RuShiftEval leaderboard. Contrary to XL-LEXEME, Deep Mistake and Gloss Reader are fine-tuned on the RuSemShift dataset. The differences between XL-LEXEME and the best two systems in the leaderboard are not statically significant. Moreover, in Table 2, the results of XL-LEXEME fine-tuned on the RuSemShift are shown. Although the fine-tuned model achieves the best correlation scores in the three datasets, the difference between DeepMistake and GlossReader is not significant." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.404, + 0.642, + 0.419 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.43, + 0.886, + 0.64 + ], + "angle": 0, + "content": "In this work, we introduced XL-LEXEME, a model for LSC Detection. XL-LEXEME is pre-trained on a large WiC dataset to mirror sentence-level encoders focusing on specific words in contexts. We evaluated our model on two Lexical Semantic Change Detection datasets: SemEval-2020 Task 1 and RuShiftEval. XL-LEXEME outperforms state-of-the-art models for LSC Detection in English, German, Swedish, and Russian datasets, with significant differences from the baselines. The XL-LEXEME effectiveness and efficiency make it reliable for LSC Detection on large diachronic corpora." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.652, + 0.646, + 0.667 + ], + "angle": 0, + "content": "7 Limitations" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.678, + 0.885, + 0.92 + ], + "angle": 0, + "content": "While the vector representations obtained using XL-LEXEME for different languages are potentially comparable, lying on the same geometric space, the evaluation of cross-lingual semantic changes cannot be performed for lacking cross-lingual LSC Detection resources. SemEval 2020 Task 1 datasets consist of small sets of target words, i.e., the number of target words for English, German, Latin, and Swedish is 37, 48, 40, and 31, respectively. The example of the Latin language highlights that XL-LEXEME can perform poorly on languages that are underrepresented in the training set of XLM-R and not covered by the WiC dataset. Generally, at the moment is not possible to state precisely how and how much XL-LEXEME" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.928, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1580" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.116, + 0.082, + 0.885, + 0.201 + ], + "angle": 0, + "content": "
SemEval-2020 Task 1 Subtask 2 LeaderboardTemporal BERTcross-encoderXL-LEXEME
Lang.UG_Student _InternJiaxin & Jinancs2020UWBCount baselineFreq. baselineTempoBERTTemporal Attention
EN0.4220.3250.3750.3670.022-0.2170.467†0.520†0.7520.757
DE0.7250.7170.7020.6970.2160.014-†0.763†0.8370.877
SV†0.547†0.588†0.536†0.604-0.022-0.150--†0.6800.754
LA0.4120.4400.3990.2540.359†0.0200.5120.565†0.016-0.056
Avg.0.5270.5180.5030.4810.144-0.083--0.5710.583
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.21, + 0.884, + 0.24 + ], + "angle": 0, + "content": "Table 1: Results (Spearman correlation) on the SemEval-2020 Task 1 Subtask 2 test set. The symbol \\(\\dagger\\) indicates there is no statistical difference with the correlation obtained by XL-LEXEME." + }, + { + "type": "table", + "bbox": [ + 0.116, + 0.252, + 0.885, + 0.334 + ], + "angle": 0, + "content": "
RuShiftEval Leaderboardcross-encoderXL-LEXEMEXL-LEXEME (Fine-tuned)
DatasetGlossReaderDeepMistakeUWBBaseline
RuShiftEval1†0.781†0.7980.3620.314†0.7270.7750.799
RuShiftEval2†0.803†0.7730.3540.302†0.7530.8220.833
RuShiftEval3†0.822†0.8030.5330.381†0.7480.8090.842
Avg.0.8020.7910.4170.3320.7430.8020.825
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.342, + 0.884, + 0.372 + ], + "angle": 0, + "content": "Table 2: Results (Spearman correlation) on the RuShiftEval test set. The symbol \\(\\dagger\\) indicates there is no statistical difference with the correlation obtained by XL-LEXEME." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.397, + 0.49, + 0.43 + ], + "angle": 0, + "content": "performance is affected by the language distribution in the XLM-R training set and the WiC dataset." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.441, + 0.287, + 0.458 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.466, + 0.49, + 0.546 + ], + "angle": 0, + "content": "We acknowledge the support of the PNRR project FAIR - Future AI Research (PE00000013), Spoke 6 - Symbiotic AI (CUP H97G22000210007) under the NRRP MUR program funded by the NextGenerationEU." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.547, + 0.49, + 0.61 + ], + "angle": 0, + "content": "This work has in part been funded by the research program Change is Key! supported by Riksbankens Jubileumsfond (under reference number M21-0021)." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.638, + 0.214, + 0.654 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.661, + 0.49, + 0.766 + ], + "angle": 0, + "content": "Nikolay Arefyev, Daniil Homskiy, Maksim Fedoseev, Adis Davletov, Vitaly Protasov, and Alexander Panchenko. 2021. DeepMistake: Which Senses are Hard to Distinguish for a WordinContext Model. In Computational Linguistics and Intellectual Technologies - Papers from the Annual International Conference \"Dialogue\" 2021, volume 2021-June. Section: 20." + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.776, + 0.49, + 0.895 + ], + "angle": 0, + "content": "Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettle-moyer, and Veselin Stoyanov. 2020a. Unsupervised Cross-lingual Representation Learning at Scale. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 8440-8451. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.904, + 0.49, + 0.92 + ], + "angle": 0, + "content": "Alexis Conneau, Kartikay Khandelwal, Naman Goyal," + }, + { + "type": "list", + "bbox": [ + 0.116, + 0.661, + 0.49, + 0.92 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.527, + 0.397, + 0.885, + 0.504 + ], + "angle": 0, + "content": "Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2020b. Unsupervised Cross-lingual Representation Learning at Scale. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 8440-8451. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.512, + 0.885, + 0.593 + ], + "angle": 0, + "content": "Raia Hadsell, Sumit Chopra, and Yann LeCun. 2006. Dimensionality reduction by learning an invariant mapping. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2006), 17-22 June 2006, New York, NY, USA, pages 1735-1742. IEEE Computer Society." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.6, + 0.885, + 0.693 + ], + "angle": 0, + "content": "William L. Hamilton, Jure Leskovec, and Dan Jurafsky. 2016. Diachronic Word Embeddings Reveal Statistical Laws of Semantic Change. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1489-1501, Berlin, Germany. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.702, + 0.885, + 0.755 + ], + "angle": 0, + "content": "David R. Hardoon, Sandor Szedmak, and John Shawe-Taylor. 2004. Canonical Correlation Analysis: An Overview with Application to Learning Methods. Neural Computation, 16(12):2639-2664." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.764, + 0.885, + 0.856 + ], + "angle": 0, + "content": "Taku Kudo and John Richardson. 2018. SentencePiece: A simple and language independent subword tokenizer and tokenizer for Neural Text Processing. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 66-71, Brussels, Belgium. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.866, + 0.885, + 0.919 + ], + "angle": 0, + "content": "Severin Laicher, Sinan Kurtyigit, Dominik Schlechtweg, Jonas Kuhn, and Sabine Schulte im Walde. 2021. Explaining and improving BERT performance on lexical semantic change detection. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.397, + 0.885, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.519, + 0.941 + ], + "angle": 0, + "content": "1581" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.135, + 0.086, + 0.489, + 0.14 + ], + "angle": 0, + "content": "the 16th Conference of the European Chapter of the Association for Computational Linguistics: Student Research Workshop, pages 192-202, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.149, + 0.49, + 0.268 + ], + "angle": 0, + "content": "Qianchu Liu, Edoardo Maria Ponti, Diana McCarthy, Ivan Vulic, and Anna Korhonen. 2021. AM2iCo: Evaluating Word Meaning in Context across Low-Resource Languages with Adversarial Examples. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, EMNLP 2021, Virtual Event / Punta Cana, Dominican Republic, 7-11 November, 2021, pages 7151-7162. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.278, + 0.489, + 0.37 + ], + "angle": 0, + "content": "Federico Martelli, Najla Kalach, Gabriele Tola, and Roberto Navigli. 2021. SemEval-2021 Task 2: Multilingual and Cross-lingual Word-in-Context Disambiguation (MCL-WiC). In Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 24–36, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.381, + 0.489, + 0.446 + ], + "angle": 0, + "content": "Tomás Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. Efficient estimation of word representations in vector space. In 1st International Conference on Learning Representations, ICLR 2013, Workshop Track Proceedings." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.457, + 0.489, + 0.51 + ], + "angle": 0, + "content": "George A. Miller. 1992. WordNet: A Lexical Database for English. In Speech and Natural Language: Proceedings of a Workshop Held at Harriman, New York, February 23-26, 1992." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.521, + 0.489, + 0.6 + ], + "angle": 0, + "content": "George A. Miller, Martin Chodorow, Shari Landes, Claudia Leacock, and Robert G. Thomas. 1994. Using a Semantic Concordance for Sense Identification. In Human Language Technology, Proceedings of a Workshop held at Plainsboro, New Jerey, USA, March 8-11, 1994. Morgan Kaufmann." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.61, + 0.489, + 0.651 + ], + "angle": 0, + "content": "Stefano Montanelli and Francesco Periti. 2023. A survey on contextualised semantic shift detection. arXiv preprint arXiv:2304.01666." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.661, + 0.489, + 0.688 + ], + "angle": 0, + "content": "Roberto Navigli. 2009. Word Sense Disambiguation: A Survey. ACM Comput. Surv., 41(2)." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.698, + 0.49, + 0.817 + ], + "angle": 0, + "content": "Mohammad Taher Pilehvar and José Camacho-Collados. 2019. WiC: the Word-in-Context Dataset for Evaluating Context-Sensitive Meaning Representations. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 1267-1273. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.826, + 0.49, + 0.918 + ], + "angle": 0, + "content": "Ondrej Prazák, Pavel Pribán, and Stephen Taylor. 2021. UWB@ RuShiftEval Measuring Semantic Difference as per-word Variation in Aligned Semantic Spaces. In Computational Linguistics and Intellectual Technologies - Papers from the Annual International Conference \"Dialogue\" 2021, volume 2021-June. Section: 20." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.086, + 0.883, + 0.166 + ], + "angle": 0, + "content": "Ondrej Prazák, Pavel Pribán, Stephen Taylor, and Jakub Sido. 2020. UWB at SemEval-2020 Task 1: Lexical Semantic Change Detection. In Proceedings of the Fourteenth Workshop on Semantic Evaluation, SemEval@COLING2020, pages 246–254. International Committee for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.174, + 0.883, + 0.228 + ], + "angle": 0, + "content": "William H. Press. 2002. Numerical recipes in \\( C++ \\): the art of scientific computing, 2nd Edition (C++ ed., print. is corrected to software version 2.10). Cambridge University Press." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.236, + 0.883, + 0.315 + ], + "angle": 0, + "content": "Maxim Rachinskiy and Nikolay Arefyev. 2021. Zeroshot Crosslingual Transfer of a Gloss Language Model for Semantic Change Detection. In Computational Linguistics and Intellectual Technologies - Papers from the Annual International Conference \"Dialogue\" 2021, volume 2021-June. Section: 20." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.324, + 0.883, + 0.43 + ], + "angle": 0, + "content": "Alessandro Raganato, Tommaso Pasini, José Camacho-Collados, and Mohammad Taher Pilehvar. 2020. XL-WiC: A Multilingual Benchmark for Evaluating Semantic Contextualization. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online, November 16-20, 2020, pages 7193-7206. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.438, + 0.883, + 0.543 + ], + "angle": 0, + "content": "Nils Reimers and Iryna Gurevych. 2019. SentenceBERT: Sentence Embeddings using Siamese BERT-Networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3982-3992, Hong Kong, China. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.552, + 0.883, + 0.632 + ], + "angle": 0, + "content": "Julia Rodina and Andrey Kutuzov. 2020. RuSemShift: a dataset of historical lexical semantic change in Russian. In Proceedings of the 28th International Conference on Computational Linguistics, pages 1037-1047, Barcelona, Spain (Online). International Committee on Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.641, + 0.883, + 0.718 + ], + "angle": 0, + "content": "Guy D. Rosin, Ido Guy, and Kira Radinsky. 2022. Time Masking for Temporal Language Models. In WSDM '22: The Fifteenth ACM International Conference on Web Search and Data Mining, Virtual Event / Tempe, AZ, USA, February 21 - 25, 2022, pages 833-841. ACM." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.728, + 0.883, + 0.768 + ], + "angle": 0, + "content": "Guy D. Rosin and Kira Radinsky. 2022. Temporal Attention for Language Models. CoRR, abs/2202.02093." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.778, + 0.883, + 0.87 + ], + "angle": 0, + "content": "Dominik Schlechtweg, Barbara McGillivray, Simon Hengchen, Haim Dubossarsky, and Nina Tahmasebi. 2020. SemEval-2020 Task 1: Unsupervised Lexical Semantic Change Detection. In Proceedings of the Fourteenth Workshop on Semantic Evaluation, SemEval@COLING2020, pages 1-23. International Committee for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.879, + 0.883, + 0.918 + ], + "angle": 0, + "content": "Dominik Schlechtweg, Sabine Schulte im Walde, and Stefanie Eckmann. 2018. Diachronic Usage Relatedness (DUREl): A Framework for the Annotation" + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.883, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1582" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.132, + 0.086, + 0.49, + 0.166 + ], + "angle": 0, + "content": "of Lexical Semantic Change. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 169-174, New Orleans, Louisiana. Association for Computational Linguistics." + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.175, + 0.49, + 0.268 + ], + "angle": 0, + "content": "Shuyi Xie, Jian Ma, Haiqin Yang, Lianxin Jiang, Yang Mo, and Jianping Shen. 2021. PALI at SemEval-2021 Task 2: Fine-Tune XLM-RoBERTa for Word in Context Disambiguation. In Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 713-718, Online. Association for Computational Linguistics." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.28, + 0.315, + 0.296 + ], + "angle": 0, + "content": "A Hyper-parameters" + }, + { + "type": "table", + "bbox": [ + 0.149, + 0.304, + 0.454, + 0.66 + ], + "angle": 0, + "content": "
Hyper-parameterValue
hidden actgelu
hidden dropout prob0.1
hidden size1024
initializer range0.02
intermediate size4096
layer norm eps1e-05
max position embeddings514
num attention heads16
num hidden layers24
position embedding typeabsolute
vocab size250004
learning rate
cross-encoder1e-05
XL-LEXEME1e-05
weight decay
cross-encoder0.01
XL-LEXEME0.00
max sequence length
cross-encoderλ = 256
XL-LEXEMEλ* = 128
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.671, + 0.49, + 0.7 + ], + "angle": 0, + "content": "Table 3: XL-LEXEME and cross-encoder hyperparameters." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1583" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.133, + 0.085, + 0.435, + 0.1 + ], + "angle": 0, + "content": "ACL 2023 Responsible NLP Checklist" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.108, + 0.323, + 0.123 + ], + "angle": 0, + "content": "A For every submission:" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.127, + 0.533, + 0.158 + ], + "angle": 0, + "content": "A1. Did you describe the limitations of your work? Section 7" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.171, + 0.553, + 0.202 + ], + "angle": 0, + "content": "A2. Did you discuss any potential risks of your work? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.212, + 0.697, + 0.244 + ], + "angle": 0, + "content": "A3. Do the abstract and introduction summarize the paper's main claims? Section 1" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.256, + 0.669, + 0.288 + ], + "angle": 0, + "content": "A4. Have you used AI writing assistants when working on this paper? Left blank." + }, + { + "type": "list", + "bbox": [ + 0.129, + 0.127, + 0.697, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.3, + 0.49, + 0.317 + ], + "angle": 0, + "content": "B Did you use or create scientific artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.322, + 0.311, + 0.336 + ], + "angle": 0, + "content": "Section 3 and Section 4" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.347, + 0.53, + 0.379 + ], + "angle": 0, + "content": "B1. Did you cite the creators of artifacts you used?" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.39, + 0.779, + 0.423 + ], + "angle": 0, + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Section 4 and References" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.433, + 0.882, + 0.513 + ], + "angle": 0, + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Section 3 and Section 4" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.525, + 0.882, + 0.589 + ], + "angle": 0, + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.599, + 0.882, + 0.647 + ], + "angle": 0, + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Section 3 and Section 4" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.658, + 0.882, + 0.753 + ], + "angle": 0, + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Section 4" + }, + { + "type": "list", + "bbox": [ + 0.129, + 0.347, + 0.882, + 0.753 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.765, + 0.494, + 0.782 + ], + "angle": 0, + "content": "C Did you run computational experiments?" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.787, + 0.31, + 0.801 + ], + "angle": 0, + "content": "Section 3 and Section 4" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.812, + 0.882, + 0.861 + ], + "angle": 0, + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Section 4 and Appendix A" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.869, + 0.878, + 0.893 + ], + "angle": 0, + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1584" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.13, + 0.084, + 0.88, + 0.116 + ], + "angle": 0, + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.118, + 0.225, + 0.131 + ], + "angle": 0, + "content": "Section 4" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.143, + 0.884, + 0.191 + ], + "angle": 0, + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.193, + 0.225, + 0.206 + ], + "angle": 0, + "content": "Section 4" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.218, + 0.884, + 0.266 + ], + "angle": 0, + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.268, + 0.225, + 0.281 + ], + "angle": 0, + "content": "Section 3" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.294, + 0.878, + 0.31 + ], + "angle": 0, + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.315, + 0.215, + 0.33 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.341, + 0.883, + 0.373 + ], + "angle": 0, + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.375, + 0.25, + 0.389 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.4, + 0.884, + 0.448 + ], + "angle": 0, + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.45, + 0.25, + 0.464 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.476, + 0.884, + 0.523 + ], + "angle": 0, + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.525, + 0.25, + 0.54 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.55, + 0.875, + 0.565 + ], + "angle": 0, + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.567, + 0.25, + 0.582 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.593, + 0.881, + 0.624 + ], + "angle": 0, + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.627, + 0.25, + 0.641 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1585" + } + ] +] \ No newline at end of file diff --git a/2023/XL-LEXEME_ WiC Pretrained Model for Cross-Lingual LEXical sEMantic changE/4754c7cf-ef79-4237-89fc-dec2c7680fce_origin.pdf b/2023/XL-LEXEME_ WiC Pretrained Model for Cross-Lingual LEXical sEMantic changE/4754c7cf-ef79-4237-89fc-dec2c7680fce_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8734c803e53184016163daea5de7dba74880b2a4 --- /dev/null +++ b/2023/XL-LEXEME_ WiC Pretrained Model for Cross-Lingual LEXical sEMantic changE/4754c7cf-ef79-4237-89fc-dec2c7680fce_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:821fad04033572dd35f6b9a90cbf377311f8c057d77bf95d5ccf7177745b9835 +size 269713 diff --git a/2023/XL-LEXEME_ WiC Pretrained Model for Cross-Lingual LEXical sEMantic changE/full.md b/2023/XL-LEXEME_ WiC Pretrained Model for Cross-Lingual LEXical sEMantic changE/full.md new file mode 100644 index 0000000000000000000000000000000000000000..34f853abad0c9c8d3fe54c0d47787d0888e67774 --- /dev/null +++ b/2023/XL-LEXEME_ WiC Pretrained Model for Cross-Lingual LEXical sEMantic changE/full.md @@ -0,0 +1,236 @@ +# XL-LEXEME: WiC Pretrained Model for Cross-Lingual LEXical sEMantic change + +Pierluigi Cassotti, Lucia Siciliani, Marco de Gemmis, Giovanni Semeraro and Pierpaolo Basile + +University of Bari Aldo Moro + +{firstname.lastname} @uniba.it + +# Abstract + +The recent introduction of large-scale datasets for the WiC (Word in Context) task enables the creation of more reliable and meaningful contextualized word embeddings. However, most of the approaches to the WiC task use cross-encoders, which prevent the possibility of deriving comparable word embeddings. In this work, we introduce XL-LEXEME, a Lexical Semantic Change Detection model. XL-LEXEME extends SBERT, highlighting the target word in the sentence. We evaluate XL-LEXEME on the multilingual benchmarks for SemEval-2020 Task 1 - Lexical Semantic Change (LSC) Detection and the RuShiftEval shared task involving five languages: English, German, Swedish, Latin, and Russian. XL-LEXEME outperforms the state-of-the-art in English, German and Swedish with statistically significant differences from the baseline results and obtains state-of-the-art performance in the RuShiftEval shared task. + +# 1 Introduction and Motivation + +Lexical Semantic Change (LSC) Detection is the task of automatically identifying words that change their meaning over time. The LSC Detection task implicitly aims to disambiguate synchronic word sense occurrences and then find differences in the word sense frequencies in different periods. Word Sense Disambiguation (WSD) is a long-studied task in Natural Language Processing (Navigli, 2009), which consists of associating the correct sense to a word occurring in a specific context. WSD involves some crucial issues, such as relying on a fixed sense inventory. Fixed sense inventories ignore the diachronic aspect of language because they can miss older unused senses or be outdated and missing new senses. + +The Word in Context task (WiC) (Pilehvar and Camacho-Collados, 2019) aims to overcome these issues. In this work, we train a model on the WiC task and then use it to perform LSC Detection. In + +the WiC task, given the word $w$ and two different contexts $C1$ , $C2$ , the systems have to determine whether the meaning of $w$ is the same in the two contexts or not. Our approach is grounded on the assumption that models trained on the WiC tasks are robust enough to transfer the knowledge learned in a synchronic setting to a diachronic one. We summarise the main contribution of this work as follows: (i) We propose a pre-trained bi-encoder model, called XL-LEXEME, on a largescale dataset for the WiC task, which allows us to obtain comparable lexical-based representations; (ii) We assert the effectiveness of XL-LEXEME despite the computational limitation compared to the cross-encoder architecture for the LSC Detection task; (iii) Experiments on the LSC Detection task show that XL-LEXEME outperforms state-of-the-art LSC Detection models for English, German, Swedish, and Russian. + +# 2 Related Work + +LSC Detection systems can be categorized based on the distributional embeddings used to tackle the LSC Detection task. One category is represented by those approaches that adopt type-base (i.e., static) embeddings. UWB (Prazák et al., 2020; Prazák et al., 2021) represents an example of this category of systems. First, it employs word2vec Skip-gram with Negative Sampling (Mikolov et al., 2013) to compute a semantic space for each corpus. It uses techniques like the Canonical Correlation Analysis (Hardoon et al., 2004) and the Orthogonal Transformation (Hamilton et al., 2016) to align the abovementioned spaces. Therefore, the cosine similarity between the vectors representing the word in two different spaces is used to detect the semantic shift. + +With the increasing use of contextualized word embeddings, numerous approaches employing BERT-base models have been developed for LSC Detection (Montanelli and Periti, 2023; Laicher + +et al., 2021). In TempoBERT (Rosin et al., 2022), the authors exploit the concept of Masked Language Modeling (MLM), where the goal is to train a language model to predict a masked portion of text given the remaining part. In particular, they employ this technique to encode the concept of time into a BERT model. This is done by concatenating a specific token representing time to the text sequence. At inference time, TempoBERT can be used to predict the year of a sentence, masking the time reference, or to predict a masked token of the sentence conditioned by the time reference. In the same line of research, in Temporal Attention (Rosin and Radinsky, 2022), the authors investigate the effect of modifying the model instead of the input sentence like in TempoBERT. This is done by extending the model's attention mechanism to consider the time when computing the weight of each word. The time dimension is encoded using a different query embedding matrix for each timestamp. + +Another kind of approach exploits the information coming from other tasks to perform LSC Detection. GlossReader represents an example (Rachinskiy and Arefyev, 2021), where a model based on XML-R (Conneau et al., 2020b) is first trained on English SemCor (Miller et al., 1994) with glosses from WordNet 3.0 (Miller, 1992) to perform WSD. Exploiting the zero-shot cross-lingual characteristics of XML-R, the authors used the same model to perform LSC Detection in the Russian language. With DeepMistake (Arefyev et al., 2021), the authors take advantage of the WiC task instead of WSD. They train a cross-encoder with XML-R as an underlying Language Model on the MCL-WiC training and development set and fine-tune on the RuSemShift dataset (Rodina and Kutuzov, 2020). DeepMistake, differently from XL-LEXEME, relies on the cross-encoder architecture and exploits only the MCL-WiC training dataset. + +# 3 XL-LEXEME + +Generally, for pairwise sentence similarity tasks, BERT models use a cross-encoder, in which the pairwise sequences are jointly encoded, and the overall vectors are used for the classification. However, in several tasks, the cross-encoder is not suitable since it cannot provide a distinct meaningful representation for each sentence. An approach to overcome this issue involves pooling the BERT out + +put encoded vectors, which often results in worse performance. Sentence-BERT (SBERT) (Reimers and Gurevych, 2019) overcomes the limitation of cross-encoders using a Siamese Network, i.e., the weights of the underlying networks are shared. SBERT encodes the two sequences separately in the BERT model exploiting the Siamese architecture. The sequence-level representation is obtained by averaging the output encoded vectors, which are directly compared using similarity measures such as cosine similarity. + +Meanwhile, cross-encoders perform better since they are trained to profit from the attention over the whole input. In this work, we introduce XLLEXEME1 which mirrors models for pairwise sequence similarity tasks and adapts them to the WiC task, giving prominence to the target word, i.e. the word for which we want to detect the LSC. The model takes as input two sequences $s_1$ and $s_2$ . The sequences are tokenized using subwords tokenizer, such as Sentence Piece (Kudo and Richardson, 2018), and the special tokens $$ and $$ are used as target word delimiters (Xie et al., 2021): + +$$ +s _ {1} = w _ {1}, \dots , < t >, w _ {i} ^ {t}, \dots , w _ {i + k} ^ {t}, < / t >, \dots , w _ {N} \tag {1} +$$ + +$$ +s _ {2} = w _ {1}, \dots , < \mathsf {t} >, w _ {j} ^ {t}, \dots , w _ {j + p} ^ {t}, < / \mathsf {t} >, \dots , w _ {M} +$$ + +where $N$ and $M$ represent the number of subwords of the sequence $s_1$ and $s_2$ respectively, while $w_i^t,\ldots ,w_{i + k}^t$ and $w_j^t,\ldots ,w_{j + p}^t$ are the subwords of the target words. In the following, we describe the baseline cross-encoder and XLLEXEME based on a bi-encoder. For the cross-encoder, the two input sequences are concatenated by the special token [SEP] in an overall sequence $s = [CLS] s_1[SEP] s_2[SEP]$ . If the length of $s$ , i.e. $N + M + 3$ , is greater than the maximum sequence length $\lambda$ , then the sequence $s$ is cut such that the length of $s_1$ and $s_2$ is less than $\lambda^{*} = \frac{\lambda - 3}{2}$ . To comply with the maximum length, the left and right contexts of the sequence are truncated. For instance, $s_1$ is truncated as follows: + +$$ +s _ {1} = w _ {n _ {0}}, \dots , < \mathrm {t} >, w _ {i} ^ {t}, \dots , w _ {i + k} ^ {t}, < / \mathrm {t} >, \dots , w _ {n _ {1}} (2) +$$ + +where $n_0 = \max (0,i - 1 - \frac{\lambda^* - k - 2}{2})$ and $n_1 = \min (N,i + k + 1 + \frac{\lambda^* - k - 2}{2})$ . The truncated sequence has a length $\gamma < \lambda$ . The encoded representations of each subword $(v_{1},v_{2},\ldots ,v_{\gamma})$ are + +summed to get the encoded representation of the overall sequence, i.e. $s^{enc} = \sum_{i}^{\gamma} v_{i}$ . Finally, the vector $s^{enc}$ is used to compute the logits: + +$$ +\operatorname {l o g i t} = \log \sigma (W s ^ {\text {e n c}}) \tag {3} +$$ + +where $W\in \mathbb{R}^{1\times d}$ . The model is trained to minimize the Binary Cross-entropy loss function. + +XL-LEXEME is a bi-encoder that encodes the input sequences using a Siamese Network into two different vector representations. Each sequence is tokenized and truncated according to the maximum length $\lambda^{*}$ , using Equation (2). We thus obtain the new lengths $\gamma_{1},\gamma_{2}$ . The vector representation is computed as the sum of the encoded subwords $(v_{1},v_{2},\dots,v_{\gamma})$ , i.e. $s_1^{enc} = \sum_i^{\gamma_1}v_i$ and $s_2^{enc} = \sum_j^{\gamma_2}v_j$ . + +XL-LEXEME is trained to minimize the Contrastive loss (Hadsell et al., 2006): + +$$ +\ell = \frac {1}{2} [ y \cdot \delta^ {2} + (1 - y) \cdot \max (0, m - \delta) ^ {2} ] \tag {4} +$$ + +where we adopt a margin $m = 0.5$ . We use as default distance $\delta$ the cosine distance between the encoded representations of $s_1$ and $s_2$ , i.e. $\delta = \cos(s_1^{enc}, s_2^{enc})$ . The main advantage of XL-LEXEME concerning models based on the cross-encoder architecture is efficiency. The time cost can be directly derived from the different architectures that exploit XL-LEXEME and the cross-encoder baseline. The self-attention time complexity $O(N^2 * d)$ depends on the vector dimension $d$ and the sequence length, which is $N$ for the cross-encoder and $\frac{N}{2}$ for XL-LEXEME. For XL-LEXEME, the time complexity is reduced to $O((\frac{N}{2})^2 * 2d)$ . + +# 4 Experimental setting + +# 4.1 Lexical Semantic Change Detection + +SemEval-2020 Task 1: Unsupervised Lexical Semantic Change Detection (Schlechtweg et al., 2020) is the first task on Unsupervised Lexical Semantic Change Detection in English, German, Swedish, and Latin languages. For each language, two corpora represent two different periods (T0, T1). Moreover, a set of target words, annotated using the DUREL framework (Schlechtweg et al., 2018), are provided. SemEval-2020 Task 1 involves two subtasks. The binary classification task requires assigning a label (changed/stable) to each target word. The ranking task sorts the target words according to their degree of semantic change. In + +this work, we focus on Subtask 2, and for the sake of simplicity, we refer to SemEval-2020 Task 1 Subtask 2 as SemEval-2020 Task 1. + +RuShiftEval, different from SemEval-2020 Task 1, involves three sub-corpora extracted from the Russian National Corpus spanning three periods. Models are evaluated on the resulting three test sets, namely RuShiftEval1 (pre-Soviet and Soviet), RuShiftEval2 (Soviet and post-Soviet), and RuShiftEval3 (pre-Soviet and post-Soviet). RuShiftEval provides participants with development data that can be used for tuning models. RuShiftEval aims to corroborate if training data can improve LSC Detection models. The development data rely on the RuSemShift dataset (Rodina and Kutuzov, 2020), which includes two sets of 70 target words for the pre-Soviet to Soviet period and Soviet to post-Soviet period, respectively. The dataset also includes annotated pairwise sentences, which can be used for training the models. + +# 4.2 Training details + +XL-LEXEME and the cross-encoder are trained using XLM-RoBERTa (XLM-R) (Conneau et al., 2020a) large as the underlying Language Model $^2$ and using an NVIDIA GeForce RTX 3090. As for training data, the model uses the training data of MCL-WiC (Martelli et al., 2021), $\mathrm{AM}^2\mathrm{ICO}$ (Liu et al., 2021), and XL-WiC datasets (Raganato et al., 2020) merged with the randomly sampled $75\%$ of the respective development data of each dataset. The remaining $25\%$ of the development data is used to fine-tune hyper-parameters. Moreover, we augment training data for the cross-encoder by swapping the order of sentences in the training set (Martelli et al., 2021). + +We use AdamW optimizer and linear learning warm-up over the $10\%$ of training data. We perform a grid search for the hyper-parameters optimization, tuning the learning rate in $\{1\mathrm{e} - 6,2\mathrm{e} - 6,$ $5\mathrm{e} - 6,1\mathrm{e} - 5,2\mathrm{e} - 5\}$ and the weight decay $\{0.0,0.01\}$ . Table 3 (Appendix A) shows the selected hyperparameters. We sample 200 sentences containing the target word for each language and each period. The sampling is repeated ten times, and the results are averaged over the ten iterations. We use the same methodology of Rachinskiy and Arefyev (2021) for sampling sentences from the RuShiftEval corpora. We sample sentences in which we find the exact match with the target words with no pre + +processing of the SemEval dataset. The LSC score is computed as the average distance between the vectors over the two different periods: + +$$ +\operatorname {L S C} (s ^ {t _ {0}}, s ^ {t _ {1}}) = \frac {1}{N \cdot M} \sum_ {i = 0} ^ {N} \sum_ {j = 0} ^ {M} \delta (s _ {i} ^ {t _ {0}}, s _ {j} ^ {t _ {1}}) \quad (5) +$$ + +where $\delta$ is the distance measure, i.e. $\delta = 1 - \log \sigma (W s^{enc})$ for the cross-encoder baseline and $\delta = \cos (s_1^{enc},s_2^{enc})$ for XL-LEXEME. + +# 5 Results + +Table 1 and Table 2 report the results on the SemEval-2020 Task 1 Subtask 2 and the results on the RuShiftEval test set. The results of the best systems are in bold. XL-LEXEME achieve the best score for English, German, Swedish, RuShiftEval1, RuShiftEval2, and RuShiftEval3. XL-LEXEME achieves a strong Spearman correlation for English and Swedish languages and a solid correlation on the German dataset, obtaining a significant correlation $(p < 0.001)$ . XL-LEXEME obtains no significant results in the Latin language since the predicted scores for the target words are not correlated with the test set. Latin is underrepresented in the training data of XLM-R, and there are no similar languages in the WiC dataset that we use for training XL-LEXEME. Moreover, the Latin dataset is more challenging as it involves the first corpus written in ancient Latin, which differs in many aspects from modern Latin. For this reason, XL-LEXEME could be ineffective in ancient languages and, in general, in languages that are not widely covered by the WiC dataset. + +We report the statistical significance of the difference between the performance of XL-LEXEME concerning the other models. The statistical significance of the difference is computed using Fisher's $z$ -transformation (Press, 2002). XL-LEXEME obtains stronger correlations than the cross-encoder, but the differences are not significant. The correlations obtained on the English and the German datasets are significantly different $(p < 0.05)$ for all the systems that participated in the SemEval2020 Task 1 but not for TempoBERT and Temporal Attention. On the other side, TempoBERT and Temporal Attention obtain a Spearman correlation on English and German that is not statistically different from the systems on the SemEval-2020 Task 1 leaderboard. In the Swedish language, XL-LEXEME is the only one obtaining a significantly + +different correlation from the Count baseline results. XL-LEXEME showed its effectiveness also in Swedish, although the WiC dataset does not cover this language. Presumably, Swedish benefits from the presence of other languages descending from the Old Norse language, namely Danish and Norwegian. + +XL-LEXEME obtains competitive results for the Russian language in the RuShiftEval leaderboard. Contrary to XL-LEXEME, Deep Mistake and Gloss Reader are fine-tuned on the RuSemShift dataset. The differences between XL-LEXEME and the best two systems in the leaderboard are not statically significant. Moreover, in Table 2, the results of XL-LEXEME fine-tuned on the RuSemShift are shown. Although the fine-tuned model achieves the best correlation scores in the three datasets, the difference between DeepMistake and GlossReader is not significant. + +# 6 Conclusion + +In this work, we introduced XL-LEXEME, a model for LSC Detection. XL-LEXEME is pre-trained on a large WiC dataset to mirror sentence-level encoders focusing on specific words in contexts. We evaluated our model on two Lexical Semantic Change Detection datasets: SemEval-2020 Task 1 and RuShiftEval. XL-LEXEME outperforms state-of-the-art models for LSC Detection in English, German, Swedish, and Russian datasets, with significant differences from the baselines. The XL-LEXEME effectiveness and efficiency make it reliable for LSC Detection on large diachronic corpora. + +# 7 Limitations + +While the vector representations obtained using XL-LEXEME for different languages are potentially comparable, lying on the same geometric space, the evaluation of cross-lingual semantic changes cannot be performed for lacking cross-lingual LSC Detection resources. SemEval 2020 Task 1 datasets consist of small sets of target words, i.e., the number of target words for English, German, Latin, and Swedish is 37, 48, 40, and 31, respectively. The example of the Latin language highlights that XL-LEXEME can perform poorly on languages that are underrepresented in the training set of XLM-R and not covered by the WiC dataset. Generally, at the moment is not possible to state precisely how and how much XL-LEXEME + +
SemEval-2020 Task 1 Subtask 2 LeaderboardTemporal BERTcross-encoderXL-LEXEME
Lang.UG_Student _InternJiaxin & Jinancs2020UWBCount baselineFreq. baselineTempoBERTTemporal Attention
EN0.4220.3250.3750.3670.022-0.2170.467†0.520†0.7520.757
DE0.7250.7170.7020.6970.2160.014-†0.763†0.8370.877
SV†0.547†0.588†0.536†0.604-0.022-0.150--†0.6800.754
LA0.4120.4400.3990.2540.359†0.0200.5120.565†0.016-0.056
Avg.0.5270.5180.5030.4810.144-0.083--0.5710.583
+ +Table 1: Results (Spearman correlation) on the SemEval-2020 Task 1 Subtask 2 test set. The symbol $\dagger$ indicates there is no statistical difference with the correlation obtained by XL-LEXEME. + +
RuShiftEval Leaderboardcross-encoderXL-LEXEMEXL-LEXEME (Fine-tuned)
DatasetGlossReaderDeepMistakeUWBBaseline
RuShiftEval1†0.781†0.7980.3620.314†0.7270.7750.799
RuShiftEval2†0.803†0.7730.3540.302†0.7530.8220.833
RuShiftEval3†0.822†0.8030.5330.381†0.7480.8090.842
Avg.0.8020.7910.4170.3320.7430.8020.825
+ +Table 2: Results (Spearman correlation) on the RuShiftEval test set. The symbol $\dagger$ indicates there is no statistical difference with the correlation obtained by XL-LEXEME. + +performance is affected by the language distribution in the XLM-R training set and the WiC dataset. + +# Acknowledgements + +We acknowledge the support of the PNRR project FAIR - Future AI Research (PE00000013), Spoke 6 - Symbiotic AI (CUP H97G22000210007) under the NRRP MUR program funded by the NextGenerationEU. + +This work has in part been funded by the research program Change is Key! supported by Riksbankens Jubileumsfond (under reference number M21-0021). + +# References + +Nikolay Arefyev, Daniil Homskiy, Maksim Fedoseev, Adis Davletov, Vitaly Protasov, and Alexander Panchenko. 2021. DeepMistake: Which Senses are Hard to Distinguish for a WordinContext Model. In Computational Linguistics and Intellectual Technologies - Papers from the Annual International Conference "Dialogue" 2021, volume 2021-June. Section: 20. +Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettle-moyer, and Veselin Stoyanov. 2020a. Unsupervised Cross-lingual Representation Learning at Scale. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 8440-8451. Association for Computational Linguistics. +Alexis Conneau, Kartikay Khandelwal, Naman Goyal, + +Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2020b. Unsupervised Cross-lingual Representation Learning at Scale. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 8440-8451. Association for Computational Linguistics. +Raia Hadsell, Sumit Chopra, and Yann LeCun. 2006. Dimensionality reduction by learning an invariant mapping. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2006), 17-22 June 2006, New York, NY, USA, pages 1735-1742. IEEE Computer Society. +William L. Hamilton, Jure Leskovec, and Dan Jurafsky. 2016. Diachronic Word Embeddings Reveal Statistical Laws of Semantic Change. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1489-1501, Berlin, Germany. Association for Computational Linguistics. +David R. Hardoon, Sandor Szedmak, and John Shawe-Taylor. 2004. Canonical Correlation Analysis: An Overview with Application to Learning Methods. Neural Computation, 16(12):2639-2664. +Taku Kudo and John Richardson. 2018. SentencePiece: A simple and language independent subword tokenizer and tokenizer for Neural Text Processing. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 66-71, Brussels, Belgium. Association for Computational Linguistics. +Severin Laicher, Sinan Kurtyigit, Dominik Schlechtweg, Jonas Kuhn, and Sabine Schulte im Walde. 2021. Explaining and improving BERT performance on lexical semantic change detection. In Proceedings of + +the 16th Conference of the European Chapter of the Association for Computational Linguistics: Student Research Workshop, pages 192-202, Online. Association for Computational Linguistics. +Qianchu Liu, Edoardo Maria Ponti, Diana McCarthy, Ivan Vulic, and Anna Korhonen. 2021. AM2iCo: Evaluating Word Meaning in Context across Low-Resource Languages with Adversarial Examples. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, EMNLP 2021, Virtual Event / Punta Cana, Dominican Republic, 7-11 November, 2021, pages 7151-7162. Association for Computational Linguistics. +Federico Martelli, Najla Kalach, Gabriele Tola, and Roberto Navigli. 2021. SemEval-2021 Task 2: Multilingual and Cross-lingual Word-in-Context Disambiguation (MCL-WiC). In Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 24–36, Online. Association for Computational Linguistics. +Tomás Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. Efficient estimation of word representations in vector space. In 1st International Conference on Learning Representations, ICLR 2013, Workshop Track Proceedings. +George A. Miller. 1992. WordNet: A Lexical Database for English. In Speech and Natural Language: Proceedings of a Workshop Held at Harriman, New York, February 23-26, 1992. +George A. Miller, Martin Chodorow, Shari Landes, Claudia Leacock, and Robert G. Thomas. 1994. Using a Semantic Concordance for Sense Identification. In Human Language Technology, Proceedings of a Workshop held at Plainsboro, New Jerey, USA, March 8-11, 1994. Morgan Kaufmann. +Stefano Montanelli and Francesco Periti. 2023. A survey on contextualised semantic shift detection. arXiv preprint arXiv:2304.01666. +Roberto Navigli. 2009. Word Sense Disambiguation: A Survey. ACM Comput. Surv., 41(2). +Mohammad Taher Pilehvar and José Camacho-Collados. 2019. WiC: the Word-in-Context Dataset for Evaluating Context-Sensitive Meaning Representations. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 1267-1273. Association for Computational Linguistics. +Ondrej Prazák, Pavel Pribán, and Stephen Taylor. 2021. UWB@ RuShiftEval Measuring Semantic Difference as per-word Variation in Aligned Semantic Spaces. In Computational Linguistics and Intellectual Technologies - Papers from the Annual International Conference "Dialogue" 2021, volume 2021-June. Section: 20. + +Ondrej Prazák, Pavel Pribán, Stephen Taylor, and Jakub Sido. 2020. UWB at SemEval-2020 Task 1: Lexical Semantic Change Detection. In Proceedings of the Fourteenth Workshop on Semantic Evaluation, SemEval@COLING2020, pages 246–254. International Committee for Computational Linguistics. +William H. Press. 2002. Numerical recipes in $C++$ : the art of scientific computing, 2nd Edition (C++ ed., print. is corrected to software version 2.10). Cambridge University Press. +Maxim Rachinskiy and Nikolay Arefyev. 2021. Zeroshot Crosslingual Transfer of a Gloss Language Model for Semantic Change Detection. In Computational Linguistics and Intellectual Technologies - Papers from the Annual International Conference "Dialogue" 2021, volume 2021-June. Section: 20. +Alessandro Raganato, Tommaso Pasini, José Camacho-Collados, and Mohammad Taher Pilehvar. 2020. XL-WiC: A Multilingual Benchmark for Evaluating Semantic Contextualization. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online, November 16-20, 2020, pages 7193-7206. Association for Computational Linguistics. +Nils Reimers and Iryna Gurevych. 2019. SentenceBERT: Sentence Embeddings using Siamese BERT-Networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3982-3992, Hong Kong, China. Association for Computational Linguistics. +Julia Rodina and Andrey Kutuzov. 2020. RuSemShift: a dataset of historical lexical semantic change in Russian. In Proceedings of the 28th International Conference on Computational Linguistics, pages 1037-1047, Barcelona, Spain (Online). International Committee on Computational Linguistics. +Guy D. Rosin, Ido Guy, and Kira Radinsky. 2022. Time Masking for Temporal Language Models. In WSDM '22: The Fifteenth ACM International Conference on Web Search and Data Mining, Virtual Event / Tempe, AZ, USA, February 21 - 25, 2022, pages 833-841. ACM. +Guy D. Rosin and Kira Radinsky. 2022. Temporal Attention for Language Models. CoRR, abs/2202.02093. +Dominik Schlechtweg, Barbara McGillivray, Simon Hengchen, Haim Dubossarsky, and Nina Tahmasebi. 2020. SemEval-2020 Task 1: Unsupervised Lexical Semantic Change Detection. In Proceedings of the Fourteenth Workshop on Semantic Evaluation, SemEval@COLING2020, pages 1-23. International Committee for Computational Linguistics. +Dominik Schlechtweg, Sabine Schulte im Walde, and Stefanie Eckmann. 2018. Diachronic Usage Relatedness (DUREl): A Framework for the Annotation + +of Lexical Semantic Change. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 169-174, New Orleans, Louisiana. Association for Computational Linguistics. + +Shuyi Xie, Jian Ma, Haiqin Yang, Lianxin Jiang, Yang Mo, and Jianping Shen. 2021. PALI at SemEval-2021 Task 2: Fine-Tune XLM-RoBERTa for Word in Context Disambiguation. In Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 713-718, Online. Association for Computational Linguistics. + +# A Hyper-parameters + +
Hyper-parameterValue
hidden actgelu
hidden dropout prob0.1
hidden size1024
initializer range0.02
intermediate size4096
layer norm eps1e-05
max position embeddings514
num attention heads16
num hidden layers24
position embedding typeabsolute
vocab size250004
learning rate
cross-encoder1e-05
XL-LEXEME1e-05
weight decay
cross-encoder0.01
XL-LEXEME0.00
max sequence length
cross-encoderλ = 256
XL-LEXEMEλ* = 128
+ +Table 3: XL-LEXEME and cross-encoder hyperparameters. + +A For every submission: + +A1. Did you describe the limitations of your work? Section 7 +A2. Did you discuss any potential risks of your work? Not applicable. Left blank. +A3. Do the abstract and introduction summarize the paper's main claims? Section 1 +A4. Have you used AI writing assistants when working on this paper? Left blank. + +B Did you use or create scientific artifacts? + +Section 3 and Section 4 + +B1. Did you cite the creators of artifacts you used? +B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Section 4 and References +B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Section 3 and Section 4 +B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Not applicable. Left blank. +B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Section 3 and Section 4 +B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Section 4 + +C Did you run computational experiments? + +Section 3 and Section 4 + +C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Section 4 and Appendix A + +The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance. + +C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? + +Section 4 + +C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? + +Section 4 + +C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? + +Section 3 + +D Did you use human annotators (e.g., crowdworkers) or research with human participants? + +Left blank. + +D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? + +No response. + +D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? + +No response. + +D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? + +No response. + +D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? + +No response. + +D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? + +No response. \ No newline at end of file diff --git a/2023/XL-LEXEME_ WiC Pretrained Model for Cross-Lingual LEXical sEMantic changE/images.zip b/2023/XL-LEXEME_ WiC Pretrained Model for Cross-Lingual LEXical sEMantic changE/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ab74b0ddd23f8fe3d1d6db4cb707dc195ce90ce7 --- /dev/null +++ b/2023/XL-LEXEME_ WiC Pretrained Model for Cross-Lingual LEXical sEMantic changE/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8d463b311b4738617ad028e1f6d749c5a6be19dea534da1b51ce7065697965da +size 200821 diff --git a/2023/XL-LEXEME_ WiC Pretrained Model for Cross-Lingual LEXical sEMantic changE/layout.json b/2023/XL-LEXEME_ WiC Pretrained Model for Cross-Lingual LEXical sEMantic changE/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..005f2f9a5390d18ed9794ce9f3174d8461234896 --- /dev/null +++ b/2023/XL-LEXEME_ WiC Pretrained Model for Cross-Lingual LEXical sEMantic changE/layout.json @@ -0,0 +1,6216 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 167, + 75, + 426, + 110 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 75, + 426, + 110 + ], + "spans": [ + { + "bbox": [ + 167, + 75, + 426, + 110 + ], + "type": "text", + "content": "XL-LEXEME: WiC Pretrained Model for Cross-Lingual LEXical sEMantic change" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 111, + 126, + 485, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 126, + 485, + 153 + ], + "spans": [ + { + "bbox": [ + 111, + 126, + 485, + 153 + ], + "type": "text", + "content": "Pierluigi Cassotti, Lucia Siciliani, Marco de Gemmis, Giovanni Semeraro and Pierpaolo Basile" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 224, + 154, + 372, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 154, + 372, + 167 + ], + "spans": [ + { + "bbox": [ + 224, + 154, + 372, + 167 + ], + "type": "text", + "content": "University of Bari Aldo Moro" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 224, + 168, + 374, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 168, + 374, + 181 + ], + "spans": [ + { + "bbox": [ + 224, + 168, + 374, + 181 + ], + "type": "text", + "content": "{firstname.lastname} @uniba.it" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 155, + 212, + 202, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 212, + 202, + 225 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 202, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 86, + 234, + 274, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 234, + 274, + 484 + ], + "spans": [ + { + "bbox": [ + 86, + 234, + 274, + 484 + ], + "type": "text", + "content": "The recent introduction of large-scale datasets for the WiC (Word in Context) task enables the creation of more reliable and meaningful contextualized word embeddings. However, most of the approaches to the WiC task use cross-encoders, which prevent the possibility of deriving comparable word embeddings. In this work, we introduce XL-LEXEME, a Lexical Semantic Change Detection model. XL-LEXEME extends SBERT, highlighting the target word in the sentence. We evaluate XL-LEXEME on the multilingual benchmarks for SemEval-2020 Task 1 - Lexical Semantic Change (LSC) Detection and the RuShiftEval shared task involving five languages: English, German, Swedish, Latin, and Russian. XL-LEXEME outperforms the state-of-the-art in English, German and Swedish with statistically significant differences from the baseline results and obtains state-of-the-art performance in the RuShiftEval shared task." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 495, + 235, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 495, + 235, + 507 + ], + "spans": [ + { + "bbox": [ + 68, + 495, + 235, + 507 + ], + "type": "text", + "content": "1 Introduction and Motivation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 516, + 290, + 718 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 516, + 290, + 718 + ], + "spans": [ + { + "bbox": [ + 69, + 516, + 290, + 718 + ], + "type": "text", + "content": "Lexical Semantic Change (LSC) Detection is the task of automatically identifying words that change their meaning over time. The LSC Detection task implicitly aims to disambiguate synchronic word sense occurrences and then find differences in the word sense frequencies in different periods. Word Sense Disambiguation (WSD) is a long-studied task in Natural Language Processing (Navigli, 2009), which consists of associating the correct sense to a word occurring in a specific context. WSD involves some crucial issues, such as relying on a fixed sense inventory. Fixed sense inventories ignore the diachronic aspect of language because they can miss older unused senses or be outdated and missing new senses." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 719, + 289, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 719, + 289, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 719, + 289, + 772 + ], + "type": "text", + "content": "The Word in Context task (WiC) (Pilehvar and Camacho-Collados, 2019) aims to overcome these issues. In this work, we train a model on the WiC task and then use it to perform LSC Detection. In" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 212, + 527, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 212, + 527, + 469 + ], + "spans": [ + { + "bbox": [ + 302, + 212, + 527, + 469 + ], + "type": "text", + "content": "the WiC task, given the word " + }, + { + "bbox": [ + 302, + 212, + 527, + 469 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 302, + 212, + 527, + 469 + ], + "type": "text", + "content": " and two different contexts " + }, + { + "bbox": [ + 302, + 212, + 527, + 469 + ], + "type": "inline_equation", + "content": "C1" + }, + { + "bbox": [ + 302, + 212, + 527, + 469 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 302, + 212, + 527, + 469 + ], + "type": "inline_equation", + "content": "C2" + }, + { + "bbox": [ + 302, + 212, + 527, + 469 + ], + "type": "text", + "content": ", the systems have to determine whether the meaning of " + }, + { + "bbox": [ + 302, + 212, + 527, + 469 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 302, + 212, + 527, + 469 + ], + "type": "text", + "content": " is the same in the two contexts or not. Our approach is grounded on the assumption that models trained on the WiC tasks are robust enough to transfer the knowledge learned in a synchronic setting to a diachronic one. We summarise the main contribution of this work as follows: (i) We propose a pre-trained bi-encoder model, called XL-LEXEME, on a largescale dataset for the WiC task, which allows us to obtain comparable lexical-based representations; (ii) We assert the effectiveness of XL-LEXEME despite the computational limitation compared to the cross-encoder architecture for the LSC Detection task; (iii) Experiments on the LSC Detection task show that XL-LEXEME outperforms state-of-the-art LSC Detection models for English, German, Swedish, and Russian." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 303, + 480, + 395, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 480, + 395, + 493 + ], + "spans": [ + { + "bbox": [ + 303, + 480, + 395, + 493 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 502, + 525, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 502, + 525, + 716 + ], + "spans": [ + { + "bbox": [ + 302, + 502, + 525, + 716 + ], + "type": "text", + "content": "LSC Detection systems can be categorized based on the distributional embeddings used to tackle the LSC Detection task. One category is represented by those approaches that adopt type-base (i.e., static) embeddings. UWB (Prazák et al., 2020; Prazák et al., 2021) represents an example of this category of systems. First, it employs word2vec Skip-gram with Negative Sampling (Mikolov et al., 2013) to compute a semantic space for each corpus. It uses techniques like the Canonical Correlation Analysis (Hardoon et al., 2004) and the Orthogonal Transformation (Hamilton et al., 2016) to align the abovementioned spaces. Therefore, the cosine similarity between the vectors representing the word in two different spaces is used to detect the semantic shift." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 719, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 719, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 719, + 525, + 772 + ], + "type": "text", + "content": "With the increasing use of contextualized word embeddings, numerous approaches employing BERT-base models have been developed for LSC Detection (Montanelli and Periti, 2023; Laicher" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "1577" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "spans": [ + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "type": "text", + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 219, + 806, + 374, + 817 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 806, + 374, + 817 + ], + "spans": [ + { + "bbox": [ + 219, + 806, + 374, + 817 + ], + "type": "text", + "content": "Volume 2: Short Papers, pages 1577-1585" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "spans": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "text", + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 66, + 71, + 293, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 71, + 293, + 354 + ], + "spans": [ + { + "bbox": [ + 66, + 71, + 293, + 354 + ], + "type": "text", + "content": "et al., 2021). In TempoBERT (Rosin et al., 2022), the authors exploit the concept of Masked Language Modeling (MLM), where the goal is to train a language model to predict a masked portion of text given the remaining part. In particular, they employ this technique to encode the concept of time into a BERT model. This is done by concatenating a specific token representing time to the text sequence. At inference time, TempoBERT can be used to predict the year of a sentence, masking the time reference, or to predict a masked token of the sentence conditioned by the time reference. In the same line of research, in Temporal Attention (Rosin and Radinsky, 2022), the authors investigate the effect of modifying the model instead of the input sentence like in TempoBERT. This is done by extending the model's attention mechanism to consider the time when computing the weight of each word. The time dimension is encoded using a different query embedding matrix for each timestamp." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 356, + 292, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 356, + 292, + 626 + ], + "spans": [ + { + "bbox": [ + 66, + 356, + 292, + 626 + ], + "type": "text", + "content": "Another kind of approach exploits the information coming from other tasks to perform LSC Detection. GlossReader represents an example (Rachinskiy and Arefyev, 2021), where a model based on XML-R (Conneau et al., 2020b) is first trained on English SemCor (Miller et al., 1994) with glosses from WordNet 3.0 (Miller, 1992) to perform WSD. Exploiting the zero-shot cross-lingual characteristics of XML-R, the authors used the same model to perform LSC Detection in the Russian language. With DeepMistake (Arefyev et al., 2021), the authors take advantage of the WiC task instead of WSD. They train a cross-encoder with XML-R as an underlying Language Model on the MCL-WiC training and development set and fine-tune on the RuSemShift dataset (Rodina and Kutuzov, 2020). DeepMistake, differently from XL-LEXEME, relies on the cross-encoder architecture and exploits only the MCL-WiC training dataset." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 640, + 162, + 655 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 640, + 162, + 655 + ], + "spans": [ + { + "bbox": [ + 67, + 640, + 162, + 655 + ], + "type": "text", + "content": "3 XL-LEXEME" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 665, + 292, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 665, + 292, + 773 + ], + "spans": [ + { + "bbox": [ + 66, + 665, + 292, + 773 + ], + "type": "text", + "content": "Generally, for pairwise sentence similarity tasks, BERT models use a cross-encoder, in which the pairwise sequences are jointly encoded, and the overall vectors are used for the classification. However, in several tasks, the cross-encoder is not suitable since it cannot provide a distinct meaningful representation for each sentence. An approach to overcome this issue involves pooling the BERT out" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 302, + 71, + 526, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 219 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 219 + ], + "type": "text", + "content": "put encoded vectors, which often results in worse performance. Sentence-BERT (SBERT) (Reimers and Gurevych, 2019) overcomes the limitation of cross-encoders using a Siamese Network, i.e., the weights of the underlying networks are shared. SBERT encodes the two sequences separately in the BERT model exploiting the Siamese architecture. The sequence-level representation is obtained by averaging the output encoded vectors, which are directly compared using similarity measures such as cosine similarity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 302, + 220, + 527, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 220, + 527, + 383 + ], + "spans": [ + { + "bbox": [ + 302, + 220, + 527, + 383 + ], + "type": "text", + "content": "Meanwhile, cross-encoders perform better since they are trained to profit from the attention over the whole input. In this work, we introduce XLLEXEME1 which mirrors models for pairwise sequence similarity tasks and adapts them to the WiC task, giving prominence to the target word, i.e. the word for which we want to detect the LSC. The model takes as input two sequences " + }, + { + "bbox": [ + 302, + 220, + 527, + 383 + ], + "type": "inline_equation", + "content": "s_1" + }, + { + "bbox": [ + 302, + 220, + 527, + 383 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 302, + 220, + 527, + 383 + ], + "type": "inline_equation", + "content": "s_2" + }, + { + "bbox": [ + 302, + 220, + 527, + 383 + ], + "type": "text", + "content": ". The sequences are tokenized using subwords tokenizer, such as Sentence Piece (Kudo and Richardson, 2018), and the special tokens " + }, + { + "bbox": [ + 302, + 220, + 527, + 383 + ], + "type": "inline_equation", + "content": "" + }, + { + "bbox": [ + 302, + 220, + 527, + 383 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 302, + 220, + 527, + 383 + ], + "type": "inline_equation", + "content": "" + }, + { + "bbox": [ + 302, + 220, + 527, + 383 + ], + "type": "text", + "content": " are used as target word delimiters (Xie et al., 2021):" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 312, + 388, + 524, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 388, + 524, + 410 + ], + "spans": [ + { + "bbox": [ + 312, + 388, + 524, + 410 + ], + "type": "interline_equation", + "content": "s _ {1} = w _ {1}, \\dots , < t >, w _ {i} ^ {t}, \\dots , w _ {i + k} ^ {t}, < / t >, \\dots , w _ {N} \\tag {1}", + "image_path": "678b4d98536a202225bc432b5cc346535e85f872afde95bce45483e2ffeae0bb.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 312, + 406, + 503, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 406, + 503, + 422 + ], + "spans": [ + { + "bbox": [ + 312, + 406, + 503, + 422 + ], + "type": "interline_equation", + "content": "s _ {2} = w _ {1}, \\dots , < \\mathsf {t} >, w _ {j} ^ {t}, \\dots , w _ {j + p} ^ {t}, < / \\mathsf {t} >, \\dots , w _ {M}", + "image_path": "3e45706e61f66316f7f262f2db83385bb98000449da0a076227e93439adc6bd1.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 427, + 527, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 427, + 527, + 629 + ], + "spans": [ + { + "bbox": [ + 302, + 427, + 527, + 629 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 302, + 427, + 527, + 629 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 302, + 427, + 527, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 302, + 427, + 527, + 629 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 302, + 427, + 527, + 629 + ], + "type": "text", + "content": " represent the number of subwords of the sequence " + }, + { + "bbox": [ + 302, + 427, + 527, + 629 + ], + "type": "inline_equation", + "content": "s_1" + }, + { + "bbox": [ + 302, + 427, + 527, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 302, + 427, + 527, + 629 + ], + "type": "inline_equation", + "content": "s_2" + }, + { + "bbox": [ + 302, + 427, + 527, + 629 + ], + "type": "text", + "content": " respectively, while " + }, + { + "bbox": [ + 302, + 427, + 527, + 629 + ], + "type": "inline_equation", + "content": "w_i^t,\\ldots ,w_{i + k}^t" + }, + { + "bbox": [ + 302, + 427, + 527, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 302, + 427, + 527, + 629 + ], + "type": "inline_equation", + "content": "w_j^t,\\ldots ,w_{j + p}^t" + }, + { + "bbox": [ + 302, + 427, + 527, + 629 + ], + "type": "text", + "content": " are the subwords of the target words. In the following, we describe the baseline cross-encoder and XLLEXEME based on a bi-encoder. For the cross-encoder, the two input sequences are concatenated by the special token [SEP] in an overall sequence " + }, + { + "bbox": [ + 302, + 427, + 527, + 629 + ], + "type": "inline_equation", + "content": "s = [CLS] s_1[SEP] s_2[SEP]" + }, + { + "bbox": [ + 302, + 427, + 527, + 629 + ], + "type": "text", + "content": ". If the length of " + }, + { + "bbox": [ + 302, + 427, + 527, + 629 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 302, + 427, + 527, + 629 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 302, + 427, + 527, + 629 + ], + "type": "inline_equation", + "content": "N + M + 3" + }, + { + "bbox": [ + 302, + 427, + 527, + 629 + ], + "type": "text", + "content": ", is greater than the maximum sequence length " + }, + { + "bbox": [ + 302, + 427, + 527, + 629 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 302, + 427, + 527, + 629 + ], + "type": "text", + "content": ", then the sequence " + }, + { + "bbox": [ + 302, + 427, + 527, + 629 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 302, + 427, + 527, + 629 + ], + "type": "text", + "content": " is cut such that the length of " + }, + { + "bbox": [ + 302, + 427, + 527, + 629 + ], + "type": "inline_equation", + "content": "s_1" + }, + { + "bbox": [ + 302, + 427, + 527, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 302, + 427, + 527, + 629 + ], + "type": "inline_equation", + "content": "s_2" + }, + { + "bbox": [ + 302, + 427, + 527, + 629 + ], + "type": "text", + "content": " is less than " + }, + { + "bbox": [ + 302, + 427, + 527, + 629 + ], + "type": "inline_equation", + "content": "\\lambda^{*} = \\frac{\\lambda - 3}{2}" + }, + { + "bbox": [ + 302, + 427, + 527, + 629 + ], + "type": "text", + "content": ". To comply with the maximum length, the left and right contexts of the sequence are truncated. For instance, " + }, + { + "bbox": [ + 302, + 427, + 527, + 629 + ], + "type": "inline_equation", + "content": "s_1" + }, + { + "bbox": [ + 302, + 427, + 527, + 629 + ], + "type": "text", + "content": " is truncated as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 309, + 637, + 524, + 652 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 637, + 524, + 652 + ], + "spans": [ + { + "bbox": [ + 309, + 637, + 524, + 652 + ], + "type": "interline_equation", + "content": "s _ {1} = w _ {n _ {0}}, \\dots , < \\mathrm {t} >, w _ {i} ^ {t}, \\dots , w _ {i + k} ^ {t}, < / \\mathrm {t} >, \\dots , w _ {n _ {1}} (2)", + "image_path": "691cb008099ed727de252c46305355f39fd7a9f718d43ab553150fe6f6adab17.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 658, + 526, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 658, + 526, + 715 + ], + "spans": [ + { + "bbox": [ + 302, + 658, + 526, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 302, + 658, + 526, + 715 + ], + "type": "inline_equation", + "content": "n_0 = \\max (0,i - 1 - \\frac{\\lambda^* - k - 2}{2})" + }, + { + "bbox": [ + 302, + 658, + 526, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 302, + 658, + 526, + 715 + ], + "type": "inline_equation", + "content": "n_1 = \\min (N,i + k + 1 + \\frac{\\lambda^* - k - 2}{2})" + }, + { + "bbox": [ + 302, + 658, + 526, + 715 + ], + "type": "text", + "content": ". The truncated sequence has a length " + }, + { + "bbox": [ + 302, + 658, + 526, + 715 + ], + "type": "inline_equation", + "content": "\\gamma < \\lambda" + }, + { + "bbox": [ + 302, + 658, + 526, + 715 + ], + "type": "text", + "content": ". The encoded representations of each subword " + }, + { + "bbox": [ + 302, + 658, + 526, + 715 + ], + "type": "inline_equation", + "content": "(v_{1},v_{2},\\ldots ,v_{\\gamma})" + }, + { + "bbox": [ + 302, + 658, + 526, + 715 + ], + "type": "text", + "content": " are" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 720, + 526, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 720, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 720, + 526, + 772 + ], + "type": "text", + "content": "1The XL-LEXEME code is available on GitHub https://github.com/pierluigic/xl-lexeme. The XL-LEXEME model is available in the Hugging Face Model Hub https://huggingface.co/ pierluigic/xl-lexeme." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "1578" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 290, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 290, + 111 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 290, + 111 + ], + "type": "text", + "content": "summed to get the encoded representation of the overall sequence, i.e. " + }, + { + "bbox": [ + 67, + 71, + 290, + 111 + ], + "type": "inline_equation", + "content": "s^{enc} = \\sum_{i}^{\\gamma} v_{i}" + }, + { + "bbox": [ + 67, + 71, + 290, + 111 + ], + "type": "text", + "content": ". Finally, the vector " + }, + { + "bbox": [ + 67, + 71, + 290, + 111 + ], + "type": "inline_equation", + "content": "s^{enc}" + }, + { + "bbox": [ + 67, + 71, + 290, + 111 + ], + "type": "text", + "content": " is used to compute the logits:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 128, + 121, + 289, + 135 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 121, + 289, + 135 + ], + "spans": [ + { + "bbox": [ + 128, + 121, + 289, + 135 + ], + "type": "interline_equation", + "content": "\\operatorname {l o g i t} = \\log \\sigma (W s ^ {\\text {e n c}}) \\tag {3}", + "image_path": "170483933b02a23286817a19d4af112261e7cdd81e529274b9a859e97598a353.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 143, + 290, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 143, + 290, + 171 + ], + "spans": [ + { + "bbox": [ + 67, + 143, + 290, + 171 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 67, + 143, + 290, + 171 + ], + "type": "inline_equation", + "content": "W\\in \\mathbb{R}^{1\\times d}" + }, + { + "bbox": [ + 67, + 143, + 290, + 171 + ], + "type": "text", + "content": ". The model is trained to minimize the Binary Cross-entropy loss function." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 172, + 290, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 172, + 290, + 294 + ], + "spans": [ + { + "bbox": [ + 67, + 172, + 290, + 294 + ], + "type": "text", + "content": "XL-LEXEME is a bi-encoder that encodes the input sequences using a Siamese Network into two different vector representations. Each sequence is tokenized and truncated according to the maximum length " + }, + { + "bbox": [ + 67, + 172, + 290, + 294 + ], + "type": "inline_equation", + "content": "\\lambda^{*}" + }, + { + "bbox": [ + 67, + 172, + 290, + 294 + ], + "type": "text", + "content": ", using Equation (2). We thus obtain the new lengths " + }, + { + "bbox": [ + 67, + 172, + 290, + 294 + ], + "type": "inline_equation", + "content": "\\gamma_{1},\\gamma_{2}" + }, + { + "bbox": [ + 67, + 172, + 290, + 294 + ], + "type": "text", + "content": ". The vector representation is computed as the sum of the encoded subwords " + }, + { + "bbox": [ + 67, + 172, + 290, + 294 + ], + "type": "inline_equation", + "content": "(v_{1},v_{2},\\dots,v_{\\gamma})" + }, + { + "bbox": [ + 67, + 172, + 290, + 294 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 67, + 172, + 290, + 294 + ], + "type": "inline_equation", + "content": "s_1^{enc} = \\sum_i^{\\gamma_1}v_i" + }, + { + "bbox": [ + 67, + 172, + 290, + 294 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 172, + 290, + 294 + ], + "type": "inline_equation", + "content": "s_2^{enc} = \\sum_j^{\\gamma_2}v_j" + }, + { + "bbox": [ + 67, + 172, + 290, + 294 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 294, + 291, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 294, + 291, + 320 + ], + "spans": [ + { + "bbox": [ + 67, + 294, + 291, + 320 + ], + "type": "text", + "content": "XL-LEXEME is trained to minimize the Contrastive loss (Hadsell et al., 2006):" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 76, + 327, + 290, + 353 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 327, + 290, + 353 + ], + "spans": [ + { + "bbox": [ + 76, + 327, + 290, + 353 + ], + "type": "interline_equation", + "content": "\\ell = \\frac {1}{2} [ y \\cdot \\delta^ {2} + (1 - y) \\cdot \\max (0, m - \\delta) ^ {2} ] \\tag {4}", + "image_path": "6212d23cdc591b2d49f35de936a65d650baa8f2313208d82ba1fe9632889170d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 360, + 291, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 360, + 291, + 550 + ], + "spans": [ + { + "bbox": [ + 67, + 360, + 291, + 550 + ], + "type": "text", + "content": "where we adopt a margin " + }, + { + "bbox": [ + 67, + 360, + 291, + 550 + ], + "type": "inline_equation", + "content": "m = 0.5" + }, + { + "bbox": [ + 67, + 360, + 291, + 550 + ], + "type": "text", + "content": ". We use as default distance " + }, + { + "bbox": [ + 67, + 360, + 291, + 550 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 67, + 360, + 291, + 550 + ], + "type": "text", + "content": " the cosine distance between the encoded representations of " + }, + { + "bbox": [ + 67, + 360, + 291, + 550 + ], + "type": "inline_equation", + "content": "s_1" + }, + { + "bbox": [ + 67, + 360, + 291, + 550 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 360, + 291, + 550 + ], + "type": "inline_equation", + "content": "s_2" + }, + { + "bbox": [ + 67, + 360, + 291, + 550 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 67, + 360, + 291, + 550 + ], + "type": "inline_equation", + "content": "\\delta = \\cos(s_1^{enc}, s_2^{enc})" + }, + { + "bbox": [ + 67, + 360, + 291, + 550 + ], + "type": "text", + "content": ". The main advantage of XL-LEXEME concerning models based on the cross-encoder architecture is efficiency. The time cost can be directly derived from the different architectures that exploit XL-LEXEME and the cross-encoder baseline. The self-attention time complexity " + }, + { + "bbox": [ + 67, + 360, + 291, + 550 + ], + "type": "inline_equation", + "content": "O(N^2 * d)" + }, + { + "bbox": [ + 67, + 360, + 291, + 550 + ], + "type": "text", + "content": " depends on the vector dimension " + }, + { + "bbox": [ + 67, + 360, + 291, + 550 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 67, + 360, + 291, + 550 + ], + "type": "text", + "content": " and the sequence length, which is " + }, + { + "bbox": [ + 67, + 360, + 291, + 550 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 67, + 360, + 291, + 550 + ], + "type": "text", + "content": " for the cross-encoder and " + }, + { + "bbox": [ + 67, + 360, + 291, + 550 + ], + "type": "inline_equation", + "content": "\\frac{N}{2}" + }, + { + "bbox": [ + 67, + 360, + 291, + 550 + ], + "type": "text", + "content": " for XL-LEXEME. For XL-LEXEME, the time complexity is reduced to " + }, + { + "bbox": [ + 67, + 360, + 291, + 550 + ], + "type": "inline_equation", + "content": "O((\\frac{N}{2})^2 * 2d)" + }, + { + "bbox": [ + 67, + 360, + 291, + 550 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 559, + 197, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 559, + 197, + 572 + ], + "spans": [ + { + "bbox": [ + 67, + 559, + 197, + 572 + ], + "type": "text", + "content": "4 Experimental setting" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 580, + 261, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 580, + 261, + 592 + ], + "spans": [ + { + "bbox": [ + 67, + 580, + 261, + 592 + ], + "type": "text", + "content": "4.1 Lexical Semantic Change Detection" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 597, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 597, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 597, + 291, + 772 + ], + "type": "text", + "content": "SemEval-2020 Task 1: Unsupervised Lexical Semantic Change Detection (Schlechtweg et al., 2020) is the first task on Unsupervised Lexical Semantic Change Detection in English, German, Swedish, and Latin languages. For each language, two corpora represent two different periods (T0, T1). Moreover, a set of target words, annotated using the DUREL framework (Schlechtweg et al., 2018), are provided. SemEval-2020 Task 1 involves two subtasks. The binary classification task requires assigning a label (changed/stable) to each target word. The ranking task sorts the target words according to their degree of semantic change. In" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 71, + 524, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 524, + 111 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 524, + 111 + ], + "type": "text", + "content": "this work, we focus on Subtask 2, and for the sake of simplicity, we refer to SemEval-2020 Task 1 Subtask 2 as SemEval-2020 Task 1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 112, + 526, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 112, + 526, + 342 + ], + "spans": [ + { + "bbox": [ + 302, + 112, + 526, + 342 + ], + "type": "text", + "content": "RuShiftEval, different from SemEval-2020 Task 1, involves three sub-corpora extracted from the Russian National Corpus spanning three periods. Models are evaluated on the resulting three test sets, namely RuShiftEval1 (pre-Soviet and Soviet), RuShiftEval2 (Soviet and post-Soviet), and RuShiftEval3 (pre-Soviet and post-Soviet). RuShiftEval provides participants with development data that can be used for tuning models. RuShiftEval aims to corroborate if training data can improve LSC Detection models. The development data rely on the RuSemShift dataset (Rodina and Kutuzov, 2020), which includes two sets of 70 target words for the pre-Soviet to Soviet period and Soviet to post-Soviet period, respectively. The dataset also includes annotated pairwise sentences, which can be used for training the models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 354, + 405, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 354, + 405, + 365 + ], + "spans": [ + { + "bbox": [ + 302, + 354, + 405, + 365 + ], + "type": "text", + "content": "4.2 Training details" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 371, + 525, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 371, + 525, + 560 + ], + "spans": [ + { + "bbox": [ + 302, + 371, + 525, + 560 + ], + "type": "text", + "content": "XL-LEXEME and the cross-encoder are trained using XLM-RoBERTa (XLM-R) (Conneau et al., 2020a) large as the underlying Language Model" + }, + { + "bbox": [ + 302, + 371, + 525, + 560 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 302, + 371, + 525, + 560 + ], + "type": "text", + "content": " and using an NVIDIA GeForce RTX 3090. As for training data, the model uses the training data of MCL-WiC (Martelli et al., 2021), " + }, + { + "bbox": [ + 302, + 371, + 525, + 560 + ], + "type": "inline_equation", + "content": "\\mathrm{AM}^2\\mathrm{ICO}" + }, + { + "bbox": [ + 302, + 371, + 525, + 560 + ], + "type": "text", + "content": " (Liu et al., 2021), and XL-WiC datasets (Raganato et al., 2020) merged with the randomly sampled " + }, + { + "bbox": [ + 302, + 371, + 525, + 560 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 302, + 371, + 525, + 560 + ], + "type": "text", + "content": " of the respective development data of each dataset. The remaining " + }, + { + "bbox": [ + 302, + 371, + 525, + 560 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 302, + 371, + 525, + 560 + ], + "type": "text", + "content": " of the development data is used to fine-tune hyper-parameters. Moreover, we augment training data for the cross-encoder by swapping the order of sentences in the training set (Martelli et al., 2021)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 562, + 526, + 751 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 562, + 526, + 751 + ], + "spans": [ + { + "bbox": [ + 302, + 562, + 526, + 751 + ], + "type": "text", + "content": "We use AdamW optimizer and linear learning warm-up over the " + }, + { + "bbox": [ + 302, + 562, + 526, + 751 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 302, + 562, + 526, + 751 + ], + "type": "text", + "content": " of training data. We perform a grid search for the hyper-parameters optimization, tuning the learning rate in " + }, + { + "bbox": [ + 302, + 562, + 526, + 751 + ], + "type": "inline_equation", + "content": "\\{1\\mathrm{e} - 6,2\\mathrm{e} - 6," + }, + { + "bbox": [ + 302, + 562, + 526, + 751 + ], + "type": "inline_equation", + "content": "5\\mathrm{e} - 6,1\\mathrm{e} - 5,2\\mathrm{e} - 5\\}" + }, + { + "bbox": [ + 302, + 562, + 526, + 751 + ], + "type": "text", + "content": " and the weight decay " + }, + { + "bbox": [ + 302, + 562, + 526, + 751 + ], + "type": "inline_equation", + "content": "\\{0.0,0.01\\}" + }, + { + "bbox": [ + 302, + 562, + 526, + 751 + ], + "type": "text", + "content": ". Table 3 (Appendix A) shows the selected hyperparameters. We sample 200 sentences containing the target word for each language and each period. The sampling is repeated ten times, and the results are averaged over the ten iterations. We use the same methodology of Rachinskiy and Arefyev (2021) for sampling sentences from the RuShiftEval corpora. We sample sentences in which we find the exact match with the target words with no pre" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 760, + 510, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 760, + 510, + 772 + ], + "spans": [ + { + "bbox": [ + 315, + 760, + 510, + 772 + ], + "type": "text", + "content": "2The XLM-R model is fine-tuned during the training." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 287, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 780, + 309, + 791 + ], + "type": "text", + "content": "1579" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 290, + 113 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 290, + 113 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 290, + 113 + ], + "type": "text", + "content": "processing of the SemEval dataset. The LSC score is computed as the average distance between the vectors over the two different periods:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 77, + 122, + 290, + 161 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 122, + 290, + 161 + ], + "spans": [ + { + "bbox": [ + 77, + 122, + 290, + 161 + ], + "type": "interline_equation", + "content": "\\operatorname {L S C} (s ^ {t _ {0}}, s ^ {t _ {1}}) = \\frac {1}{N \\cdot M} \\sum_ {i = 0} ^ {N} \\sum_ {j = 0} ^ {M} \\delta (s _ {i} ^ {t _ {0}}, s _ {j} ^ {t _ {1}}) \\quad (5)", + "image_path": "12621753f74d6163f898cae7641ef2fb39fd01efaf85c1294183691893c4ea90.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 170, + 291, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 170, + 291, + 212 + ], + "spans": [ + { + "bbox": [ + 67, + 170, + 291, + 212 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 67, + 170, + 291, + 212 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 67, + 170, + 291, + 212 + ], + "type": "text", + "content": " is the distance measure, i.e. " + }, + { + "bbox": [ + 67, + 170, + 291, + 212 + ], + "type": "inline_equation", + "content": "\\delta = 1 - \\log \\sigma (W s^{enc})" + }, + { + "bbox": [ + 67, + 170, + 291, + 212 + ], + "type": "text", + "content": " for the cross-encoder baseline and " + }, + { + "bbox": [ + 67, + 170, + 291, + 212 + ], + "type": "inline_equation", + "content": "\\delta = \\cos (s_1^{enc},s_2^{enc})" + }, + { + "bbox": [ + 67, + 170, + 291, + 212 + ], + "type": "text", + "content": " for XL-LEXEME." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 222, + 127, + 235 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 222, + 127, + 235 + ], + "spans": [ + { + "bbox": [ + 67, + 222, + 127, + 235 + ], + "type": "text", + "content": "5 Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 244, + 291, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 244, + 291, + 541 + ], + "spans": [ + { + "bbox": [ + 67, + 244, + 291, + 541 + ], + "type": "text", + "content": "Table 1 and Table 2 report the results on the SemEval-2020 Task 1 Subtask 2 and the results on the RuShiftEval test set. The results of the best systems are in bold. XL-LEXEME achieve the best score for English, German, Swedish, RuShiftEval1, RuShiftEval2, and RuShiftEval3. XL-LEXEME achieves a strong Spearman correlation for English and Swedish languages and a solid correlation on the German dataset, obtaining a significant correlation " + }, + { + "bbox": [ + 67, + 244, + 291, + 541 + ], + "type": "inline_equation", + "content": "(p < 0.001)" + }, + { + "bbox": [ + 67, + 244, + 291, + 541 + ], + "type": "text", + "content": ". XL-LEXEME obtains no significant results in the Latin language since the predicted scores for the target words are not correlated with the test set. Latin is underrepresented in the training data of XLM-R, and there are no similar languages in the WiC dataset that we use for training XL-LEXEME. Moreover, the Latin dataset is more challenging as it involves the first corpus written in ancient Latin, which differs in many aspects from modern Latin. For this reason, XL-LEXEME could be ineffective in ancient languages and, in general, in languages that are not widely covered by the WiC dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 543, + 292, + 774 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 543, + 292, + 774 + ], + "spans": [ + { + "bbox": [ + 67, + 543, + 292, + 774 + ], + "type": "text", + "content": "We report the statistical significance of the difference between the performance of XL-LEXEME concerning the other models. The statistical significance of the difference is computed using Fisher's " + }, + { + "bbox": [ + 67, + 543, + 292, + 774 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 67, + 543, + 292, + 774 + ], + "type": "text", + "content": "-transformation (Press, 2002). XL-LEXEME obtains stronger correlations than the cross-encoder, but the differences are not significant. The correlations obtained on the English and the German datasets are significantly different " + }, + { + "bbox": [ + 67, + 543, + 292, + 774 + ], + "type": "inline_equation", + "content": "(p < 0.05)" + }, + { + "bbox": [ + 67, + 543, + 292, + 774 + ], + "type": "text", + "content": " for all the systems that participated in the SemEval2020 Task 1 but not for TempoBERT and Temporal Attention. On the other side, TempoBERT and Temporal Attention obtain a Spearman correlation on English and German that is not statistically different from the systems on the SemEval-2020 Task 1 leaderboard. In the Swedish language, XL-LEXEME is the only one obtaining a significantly" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 71, + 526, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 166 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 166 + ], + "type": "text", + "content": "different correlation from the Count baseline results. XL-LEXEME showed its effectiveness also in Swedish, although the WiC dataset does not cover this language. Presumably, Swedish benefits from the presence of other languages descending from the Old Norse language, namely Danish and Norwegian." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 167, + 527, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 167, + 527, + 329 + ], + "spans": [ + { + "bbox": [ + 302, + 167, + 527, + 329 + ], + "type": "text", + "content": "XL-LEXEME obtains competitive results for the Russian language in the RuShiftEval leaderboard. Contrary to XL-LEXEME, Deep Mistake and Gloss Reader are fine-tuned on the RuSemShift dataset. The differences between XL-LEXEME and the best two systems in the leaderboard are not statically significant. Moreover, in Table 2, the results of XL-LEXEME fine-tuned on the RuSemShift are shown. Although the fine-tuned model achieves the best correlation scores in the three datasets, the difference between DeepMistake and GlossReader is not significant." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 339, + 381, + 352 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 339, + 381, + 352 + ], + "spans": [ + { + "bbox": [ + 302, + 339, + 381, + 352 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 361, + 527, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 361, + 527, + 538 + ], + "spans": [ + { + "bbox": [ + 302, + 361, + 527, + 538 + ], + "type": "text", + "content": "In this work, we introduced XL-LEXEME, a model for LSC Detection. XL-LEXEME is pre-trained on a large WiC dataset to mirror sentence-level encoders focusing on specific words in contexts. We evaluated our model on two Lexical Semantic Change Detection datasets: SemEval-2020 Task 1 and RuShiftEval. XL-LEXEME outperforms state-of-the-art models for LSC Detection in English, German, Swedish, and Russian datasets, with significant differences from the baselines. The XL-LEXEME effectiveness and efficiency make it reliable for LSC Detection on large diachronic corpora." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 548, + 384, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 548, + 384, + 560 + ], + "spans": [ + { + "bbox": [ + 302, + 548, + 384, + 560 + ], + "type": "text", + "content": "7 Limitations" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 570, + 526, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 570, + 526, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 570, + 526, + 773 + ], + "type": "text", + "content": "While the vector representations obtained using XL-LEXEME for different languages are potentially comparable, lying on the same geometric space, the evaluation of cross-lingual semantic changes cannot be performed for lacking cross-lingual LSC Detection resources. SemEval 2020 Task 1 datasets consist of small sets of target words, i.e., the number of target words for English, German, Latin, and Swedish is 37, 48, 40, and 31, respectively. The example of the Latin language highlights that XL-LEXEME can perform poorly on languages that are underrepresented in the training set of XLM-R and not covered by the WiC dataset. Generally, at the moment is not possible to state precisely how and how much XL-LEXEME" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 780, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 780, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 780, + 310, + 791 + ], + "type": "text", + "content": "1580" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 68, + 526, + 169 + ], + "blocks": [ + { + "bbox": [ + 69, + 68, + 526, + 169 + ], + "lines": [ + { + "bbox": [ + 69, + 68, + 526, + 169 + ], + "spans": [ + { + "bbox": [ + 69, + 68, + 526, + 169 + ], + "type": "table", + "html": "
SemEval-2020 Task 1 Subtask 2 LeaderboardTemporal BERTcross-encoderXL-LEXEME
Lang.UG_Student _InternJiaxin & Jinancs2020UWBCount baselineFreq. baselineTempoBERTTemporal Attention
EN0.4220.3250.3750.3670.022-0.2170.467†0.520†0.7520.757
DE0.7250.7170.7020.6970.2160.014-†0.763†0.8370.877
SV†0.547†0.588†0.536†0.604-0.022-0.150--†0.6800.754
LA0.4120.4400.3990.2540.359†0.0200.5120.565†0.016-0.056
Avg.0.5270.5180.5030.4810.144-0.083--0.5710.583
", + "image_path": "963ca8017937b1e0e0fe220ddd28b5d73c7fc59709c60f69f742179a8b492a25.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 69, + 211, + 526, + 280 + ], + "blocks": [ + { + "bbox": [ + 67, + 176, + 525, + 201 + ], + "lines": [ + { + "bbox": [ + 67, + 176, + 525, + 201 + ], + "spans": [ + { + "bbox": [ + 67, + 176, + 525, + 201 + ], + "type": "text", + "content": "Table 1: Results (Spearman correlation) on the SemEval-2020 Task 1 Subtask 2 test set. The symbol " + }, + { + "bbox": [ + 67, + 176, + 525, + 201 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 67, + 176, + 525, + 201 + ], + "type": "text", + "content": " indicates there is no statistical difference with the correlation obtained by XL-LEXEME." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 211, + 526, + 280 + ], + "lines": [ + { + "bbox": [ + 69, + 211, + 526, + 280 + ], + "spans": [ + { + "bbox": [ + 69, + 211, + 526, + 280 + ], + "type": "table", + "html": "
RuShiftEval Leaderboardcross-encoderXL-LEXEMEXL-LEXEME (Fine-tuned)
DatasetGlossReaderDeepMistakeUWBBaseline
RuShiftEval1†0.781†0.7980.3620.314†0.7270.7750.799
RuShiftEval2†0.803†0.7730.3540.302†0.7530.8220.833
RuShiftEval3†0.822†0.8030.5330.381†0.7480.8090.842
Avg.0.8020.7910.4170.3320.7430.8020.825
", + "image_path": "48a7a89adc54cd27695d77cd0ec3e1a155e0fdab7c86034fddc91ea32cd000c0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 287, + 525, + 312 + ], + "lines": [ + { + "bbox": [ + 67, + 287, + 525, + 312 + ], + "spans": [ + { + "bbox": [ + 67, + 287, + 525, + 312 + ], + "type": "text", + "content": "Table 2: Results (Spearman correlation) on the RuShiftEval test set. The symbol " + }, + { + "bbox": [ + 67, + 287, + 525, + 312 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 67, + 287, + 525, + 312 + ], + "type": "text", + "content": " indicates there is no statistical difference with the correlation obtained by XL-LEXEME." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 333, + 291, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 333, + 291, + 361 + ], + "spans": [ + { + "bbox": [ + 67, + 333, + 291, + 361 + ], + "type": "text", + "content": "performance is affected by the language distribution in the XLM-R training set and the WiC dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 370, + 170, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 370, + 170, + 385 + ], + "spans": [ + { + "bbox": [ + 68, + 370, + 170, + 385 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 391, + 291, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 391, + 291, + 459 + ], + "spans": [ + { + "bbox": [ + 67, + 391, + 291, + 459 + ], + "type": "text", + "content": "We acknowledge the support of the PNRR project FAIR - Future AI Research (PE00000013), Spoke 6 - Symbiotic AI (CUP H97G22000210007) under the NRRP MUR program funded by the NextGenerationEU." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 460, + 291, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 460, + 291, + 513 + ], + "spans": [ + { + "bbox": [ + 67, + 460, + 291, + 513 + ], + "type": "text", + "content": "This work has in part been funded by the research program Change is Key! supported by Riksbankens Jubileumsfond (under reference number M21-0021)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 68, + 536, + 127, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 536, + 127, + 550 + ], + "spans": [ + { + "bbox": [ + 68, + 536, + 127, + 550 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 555, + 291, + 773 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 69, + 555, + 291, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 555, + 291, + 644 + ], + "spans": [ + { + "bbox": [ + 69, + 555, + 291, + 644 + ], + "type": "text", + "content": "Nikolay Arefyev, Daniil Homskiy, Maksim Fedoseev, Adis Davletov, Vitaly Protasov, and Alexander Panchenko. 2021. DeepMistake: Which Senses are Hard to Distinguish for a WordinContext Model. In Computational Linguistics and Intellectual Technologies - Papers from the Annual International Conference \"Dialogue\" 2021, volume 2021-June. Section: 20." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 652, + 291, + 752 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 652, + 291, + 752 + ], + "spans": [ + { + "bbox": [ + 69, + 652, + 291, + 752 + ], + "type": "text", + "content": "Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettle-moyer, and Veselin Stoyanov. 2020a. Unsupervised Cross-lingual Representation Learning at Scale. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 8440-8451. Association for Computational Linguistics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 69, + 760, + 291, + 773 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 760, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 69, + 760, + 291, + 773 + ], + "type": "text", + "content": "Alexis Conneau, Kartikay Khandelwal, Naman Goyal," + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 333, + 526, + 772 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 313, + 333, + 526, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 333, + 526, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 333, + 526, + 423 + ], + "type": "text", + "content": "Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2020b. Unsupervised Cross-lingual Representation Learning at Scale. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 8440-8451. Association for Computational Linguistics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 430, + 526, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 430, + 526, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 526, + 498 + ], + "type": "text", + "content": "Raia Hadsell, Sumit Chopra, and Yann LeCun. 2006. Dimensionality reduction by learning an invariant mapping. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2006), 17-22 June 2006, New York, NY, USA, pages 1735-1742. IEEE Computer Society." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 504, + 526, + 582 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 504, + 526, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 504, + 526, + 582 + ], + "type": "text", + "content": "William L. Hamilton, Jure Leskovec, and Dan Jurafsky. 2016. Diachronic Word Embeddings Reveal Statistical Laws of Semantic Change. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1489-1501, Berlin, Germany. Association for Computational Linguistics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 590, + 526, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 590, + 526, + 634 + ], + "spans": [ + { + "bbox": [ + 304, + 590, + 526, + 634 + ], + "type": "text", + "content": "David R. Hardoon, Sandor Szedmak, and John Shawe-Taylor. 2004. Canonical Correlation Analysis: An Overview with Application to Learning Methods. Neural Computation, 16(12):2639-2664." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 642, + 526, + 719 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 526, + 719 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 526, + 719 + ], + "type": "text", + "content": "Taku Kudo and John Richardson. 2018. SentencePiece: A simple and language independent subword tokenizer and tokenizer for Neural Text Processing. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 66-71, Brussels, Belgium. Association for Computational Linguistics." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 728, + 526, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 728, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 728, + 526, + 772 + ], + "type": "text", + "content": "Severin Laicher, Sinan Kurtyigit, Dominik Schlechtweg, Jonas Kuhn, and Sabine Schulte im Walde. 2021. Explaining and improving BERT performance on lexical semantic change detection. In Proceedings of" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 308, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 308, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 308, + 791 + ], + "type": "text", + "content": "1581" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 80, + 72, + 290, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 72, + 290, + 117 + ], + "spans": [ + { + "bbox": [ + 80, + 72, + 290, + 117 + ], + "type": "text", + "content": "the 16th Conference of the European Chapter of the Association for Computational Linguistics: Student Research Workshop, pages 192-202, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 125, + 291, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 125, + 291, + 225 + ], + "spans": [ + { + "bbox": [ + 69, + 125, + 291, + 225 + ], + "type": "text", + "content": "Qianchu Liu, Edoardo Maria Ponti, Diana McCarthy, Ivan Vulic, and Anna Korhonen. 2021. AM2iCo: Evaluating Word Meaning in Context across Low-Resource Languages with Adversarial Examples. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, EMNLP 2021, Virtual Event / Punta Cana, Dominican Republic, 7-11 November, 2021, pages 7151-7162. Association for Computational Linguistics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 233, + 290, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 233, + 290, + 311 + ], + "spans": [ + { + "bbox": [ + 69, + 233, + 290, + 311 + ], + "type": "text", + "content": "Federico Martelli, Najla Kalach, Gabriele Tola, and Roberto Navigli. 2021. SemEval-2021 Task 2: Multilingual and Cross-lingual Word-in-Context Disambiguation (MCL-WiC). In Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 24–36, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 320, + 290, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 320, + 290, + 375 + ], + "spans": [ + { + "bbox": [ + 69, + 320, + 290, + 375 + ], + "type": "text", + "content": "Tomás Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. Efficient estimation of word representations in vector space. In 1st International Conference on Learning Representations, ICLR 2013, Workshop Track Proceedings." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 384, + 290, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 384, + 290, + 428 + ], + "spans": [ + { + "bbox": [ + 69, + 384, + 290, + 428 + ], + "type": "text", + "content": "George A. Miller. 1992. WordNet: A Lexical Database for English. In Speech and Natural Language: Proceedings of a Workshop Held at Harriman, New York, February 23-26, 1992." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 438, + 290, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 438, + 290, + 504 + ], + "spans": [ + { + "bbox": [ + 69, + 438, + 290, + 504 + ], + "type": "text", + "content": "George A. Miller, Martin Chodorow, Shari Landes, Claudia Leacock, and Robert G. Thomas. 1994. Using a Semantic Concordance for Sense Identification. In Human Language Technology, Proceedings of a Workshop held at Plainsboro, New Jerey, USA, March 8-11, 1994. Morgan Kaufmann." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 513, + 290, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 513, + 290, + 547 + ], + "spans": [ + { + "bbox": [ + 69, + 513, + 290, + 547 + ], + "type": "text", + "content": "Stefano Montanelli and Francesco Periti. 2023. A survey on contextualised semantic shift detection. arXiv preprint arXiv:2304.01666." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 555, + 290, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 555, + 290, + 578 + ], + "spans": [ + { + "bbox": [ + 69, + 555, + 290, + 578 + ], + "type": "text", + "content": "Roberto Navigli. 2009. Word Sense Disambiguation: A Survey. ACM Comput. Surv., 41(2)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 587, + 291, + 687 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 587, + 291, + 687 + ], + "spans": [ + { + "bbox": [ + 69, + 587, + 291, + 687 + ], + "type": "text", + "content": "Mohammad Taher Pilehvar and José Camacho-Collados. 2019. WiC: the Word-in-Context Dataset for Evaluating Context-Sensitive Meaning Representations. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 1267-1273. Association for Computational Linguistics." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 694, + 291, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 694, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 694, + 291, + 772 + ], + "type": "text", + "content": "Ondrej Prazák, Pavel Pribán, and Stephen Taylor. 2021. UWB@ RuShiftEval Measuring Semantic Difference as per-word Variation in Aligned Semantic Spaces. In Computational Linguistics and Intellectual Technologies - Papers from the Annual International Conference \"Dialogue\" 2021, volume 2021-June. Section: 20." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 525, + 772 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 304, + 72, + 525, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 525, + 139 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 525, + 139 + ], + "type": "text", + "content": "Ondrej Prazák, Pavel Pribán, Stephen Taylor, and Jakub Sido. 2020. UWB at SemEval-2020 Task 1: Lexical Semantic Change Detection. In Proceedings of the Fourteenth Workshop on Semantic Evaluation, SemEval@COLING2020, pages 246–254. International Committee for Computational Linguistics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 146, + 525, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 146, + 525, + 191 + ], + "spans": [ + { + "bbox": [ + 304, + 146, + 525, + 191 + ], + "type": "text", + "content": "William H. Press. 2002. Numerical recipes in " + }, + { + "bbox": [ + 304, + 146, + 525, + 191 + ], + "type": "inline_equation", + "content": "C++" + }, + { + "bbox": [ + 304, + 146, + 525, + 191 + ], + "type": "text", + "content": ": the art of scientific computing, 2nd Edition (C++ ed., print. is corrected to software version 2.10). Cambridge University Press." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 198, + 525, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 525, + 264 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 525, + 264 + ], + "type": "text", + "content": "Maxim Rachinskiy and Nikolay Arefyev. 2021. Zeroshot Crosslingual Transfer of a Gloss Language Model for Semantic Change Detection. In Computational Linguistics and Intellectual Technologies - Papers from the Annual International Conference \"Dialogue\" 2021, volume 2021-June. Section: 20." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 272, + 525, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 272, + 525, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 272, + 525, + 361 + ], + "type": "text", + "content": "Alessandro Raganato, Tommaso Pasini, José Camacho-Collados, and Mohammad Taher Pilehvar. 2020. XL-WiC: A Multilingual Benchmark for Evaluating Semantic Contextualization. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online, November 16-20, 2020, pages 7193-7206. Association for Computational Linguistics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 368, + 525, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 368, + 525, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 368, + 525, + 456 + ], + "type": "text", + "content": "Nils Reimers and Iryna Gurevych. 2019. SentenceBERT: Sentence Embeddings using Siamese BERT-Networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3982-3992, Hong Kong, China. Association for Computational Linguistics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 464, + 525, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 525, + 531 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 525, + 531 + ], + "type": "text", + "content": "Julia Rodina and Andrey Kutuzov. 2020. RuSemShift: a dataset of historical lexical semantic change in Russian. In Proceedings of the 28th International Conference on Computational Linguistics, pages 1037-1047, Barcelona, Spain (Online). International Committee on Computational Linguistics." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 539, + 525, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 539, + 525, + 603 + ], + "spans": [ + { + "bbox": [ + 304, + 539, + 525, + 603 + ], + "type": "text", + "content": "Guy D. Rosin, Ido Guy, and Kira Radinsky. 2022. Time Masking for Temporal Language Models. In WSDM '22: The Fifteenth ACM International Conference on Web Search and Data Mining, Virtual Event / Tempe, AZ, USA, February 21 - 25, 2022, pages 833-841. ACM." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 612, + 525, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 612, + 525, + 645 + ], + "spans": [ + { + "bbox": [ + 304, + 612, + 525, + 645 + ], + "type": "text", + "content": "Guy D. Rosin and Kira Radinsky. 2022. Temporal Attention for Language Models. CoRR, abs/2202.02093." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 654, + 525, + 731 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 525, + 731 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 525, + 731 + ], + "type": "text", + "content": "Dominik Schlechtweg, Barbara McGillivray, Simon Hengchen, Haim Dubossarsky, and Nina Tahmasebi. 2020. SemEval-2020 Task 1: Unsupervised Lexical Semantic Change Detection. In Proceedings of the Fourteenth Workshop on Semantic Evaluation, SemEval@COLING2020, pages 1-23. International Committee for Computational Linguistics." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 739, + 525, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 739, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 739, + 525, + 772 + ], + "type": "text", + "content": "Dominik Schlechtweg, Sabine Schulte im Walde, and Stefanie Eckmann. 2018. Diachronic Usage Relatedness (DUREl): A Framework for the Annotation" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "text", + "content": "1582" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 78, + 72, + 291, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 72, + 291, + 139 + ], + "spans": [ + { + "bbox": [ + 78, + 72, + 291, + 139 + ], + "type": "text", + "content": "of Lexical Semantic Change. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 169-174, New Orleans, Louisiana. Association for Computational Linguistics." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 147, + 291, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 147, + 291, + 225 + ], + "spans": [ + { + "bbox": [ + 68, + 147, + 291, + 225 + ], + "type": "text", + "content": "Shuyi Xie, Jian Ma, Haiqin Yang, Lianxin Jiang, Yang Mo, and Jianping Shen. 2021. PALI at SemEval-2021 Task 2: Fine-Tune XLM-RoBERTa for Word in Context Disambiguation. In Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 713-718, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 68, + 235, + 187, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 235, + 187, + 248 + ], + "spans": [ + { + "bbox": [ + 68, + 235, + 187, + 248 + ], + "type": "text", + "content": "A Hyper-parameters" + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 88, + 255, + 270, + 555 + ], + "blocks": [ + { + "bbox": [ + 88, + 255, + 270, + 555 + ], + "lines": [ + { + "bbox": [ + 88, + 255, + 270, + 555 + ], + "spans": [ + { + "bbox": [ + 88, + 255, + 270, + 555 + ], + "type": "table", + "html": "
Hyper-parameterValue
hidden actgelu
hidden dropout prob0.1
hidden size1024
initializer range0.02
intermediate size4096
layer norm eps1e-05
max position embeddings514
num attention heads16
num hidden layers24
position embedding typeabsolute
vocab size250004
learning rate
cross-encoder1e-05
XL-LEXEME1e-05
weight decay
cross-encoder0.01
XL-LEXEME0.00
max sequence length
cross-encoderλ = 256
XL-LEXEMEλ* = 128
", + "image_path": "acbc1d7372c2c7034e3f5877211fa20c541a0b7ceb2df51fd84539470e35f834.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 564, + 291, + 588 + ], + "lines": [ + { + "bbox": [ + 67, + 564, + 291, + 588 + ], + "spans": [ + { + "bbox": [ + 67, + 564, + 291, + 588 + ], + "type": "text", + "content": "Table 3: XL-LEXEME and cross-encoder hyperparameters." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1583" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "spans": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "type": "text", + "content": "A For every submission:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 76, + 106, + 414, + 242 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "spans": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "type": "text", + "content": "A1. Did you describe the limitations of your work? Section 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 76, + 143, + 329, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 143, + 329, + 169 + ], + "spans": [ + { + "bbox": [ + 76, + 143, + 329, + 169 + ], + "type": "text", + "content": "A2. Did you discuss any potential risks of your work? Not applicable. Left blank." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 178, + 414, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 178, + 414, + 205 + ], + "spans": [ + { + "bbox": [ + 77, + 178, + 414, + 205 + ], + "type": "text", + "content": "A3. Do the abstract and introduction summarize the paper's main claims? Section 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 77, + 215, + 398, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 215, + 398, + 242 + ], + "spans": [ + { + "bbox": [ + 77, + 215, + 398, + 242 + ], + "type": "text", + "content": "A4. Have you used AI writing assistants when working on this paper? Left blank." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 68, + 252, + 291, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 252, + 291, + 266 + ], + "spans": [ + { + "bbox": [ + 68, + 252, + 291, + 266 + ], + "type": "text", + "content": "B Did you use or create scientific artifacts?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 79, + 270, + 185, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 270, + 185, + 282 + ], + "spans": [ + { + "bbox": [ + 79, + 270, + 185, + 282 + ], + "type": "text", + "content": "Section 3 and Section 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 76, + 291, + 524, + 633 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 77, + 291, + 315, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 291, + 315, + 318 + ], + "spans": [ + { + "bbox": [ + 77, + 291, + 315, + 318 + ], + "type": "text", + "content": "B1. Did you cite the creators of artifacts you used?" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 77, + 327, + 463, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 327, + 463, + 355 + ], + "spans": [ + { + "bbox": [ + 77, + 327, + 463, + 355 + ], + "type": "text", + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Section 4 and References" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 77, + 364, + 524, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 364, + 524, + 431 + ], + "spans": [ + { + "bbox": [ + 77, + 364, + 524, + 431 + ], + "type": "text", + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Section 3 and Section 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 76, + 441, + 524, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 441, + 524, + 495 + ], + "spans": [ + { + "bbox": [ + 76, + 441, + 524, + 495 + ], + "type": "text", + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Not applicable. Left blank." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 77, + 503, + 524, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 503, + 524, + 544 + ], + "spans": [ + { + "bbox": [ + 77, + 503, + 524, + 544 + ], + "type": "text", + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Section 3 and Section 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 77, + 553, + 524, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 553, + 524, + 633 + ], + "spans": [ + { + "bbox": [ + 77, + 553, + 524, + 633 + ], + "type": "text", + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Section 4" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "spans": [ + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "type": "text", + "content": "C Did you run computational experiments?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 79, + 661, + 184, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 661, + 184, + 673 + ], + "spans": [ + { + "bbox": [ + 79, + 661, + 184, + 673 + ], + "type": "text", + "content": "Section 3 and Section 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 77, + 682, + 524, + 724 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 682, + 524, + 724 + ], + "spans": [ + { + "bbox": [ + 77, + 682, + 524, + 724 + ], + "type": "text", + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Section 4 and Appendix A" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "spans": [ + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "type": "text", + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "spans": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "text", + "content": "ACL 2023 Responsible NLP Checklist" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1584" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 77, + 70, + 523, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 70, + 523, + 97 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 523, + 97 + ], + "type": "text", + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 89, + 99, + 133, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 99, + 133, + 110 + ], + "spans": [ + { + "bbox": [ + 89, + 99, + 133, + 110 + ], + "type": "text", + "content": "Section 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 120, + 525, + 160 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 120, + 525, + 160 + ], + "spans": [ + { + "bbox": [ + 77, + 120, + 525, + 160 + ], + "type": "text", + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 162, + 133, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 162, + 133, + 173 + ], + "spans": [ + { + "bbox": [ + 89, + 162, + 133, + 173 + ], + "type": "text", + "content": "Section 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 183, + 525, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 183, + 525, + 223 + ], + "spans": [ + { + "bbox": [ + 77, + 183, + 525, + 223 + ], + "type": "text", + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 225, + 133, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 225, + 133, + 236 + ], + "spans": [ + { + "bbox": [ + 89, + 225, + 133, + 236 + ], + "type": "text", + "content": "Section 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 247, + 522, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 247, + 522, + 260 + ], + "spans": [ + { + "bbox": [ + 67, + 247, + 522, + 260 + ], + "type": "text", + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 79, + 264, + 127, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 264, + 127, + 277 + ], + "spans": [ + { + "bbox": [ + 79, + 264, + 127, + 277 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 76, + 286, + 525, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 286, + 525, + 313 + ], + "spans": [ + { + "bbox": [ + 76, + 286, + 525, + 313 + ], + "type": "text", + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 89, + 315, + 148, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 315, + 148, + 327 + ], + "spans": [ + { + "bbox": [ + 89, + 315, + 148, + 327 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 76, + 336, + 525, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 336, + 525, + 376 + ], + "spans": [ + { + "bbox": [ + 76, + 336, + 525, + 376 + ], + "type": "text", + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 89, + 378, + 148, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 378, + 148, + 390 + ], + "spans": [ + { + "bbox": [ + 89, + 378, + 148, + 390 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 76, + 400, + 525, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 400, + 525, + 439 + ], + "spans": [ + { + "bbox": [ + 76, + 400, + 525, + 439 + ], + "type": "text", + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 89, + 441, + 148, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 441, + 148, + 454 + ], + "spans": [ + { + "bbox": [ + 89, + 441, + 148, + 454 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 76, + 462, + 520, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 462, + 520, + 475 + ], + "spans": [ + { + "bbox": [ + 76, + 462, + 520, + 475 + ], + "type": "text", + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 89, + 476, + 148, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 476, + 148, + 489 + ], + "spans": [ + { + "bbox": [ + 89, + 476, + 148, + 489 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 76, + 498, + 524, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 498, + 524, + 524 + ], + "spans": [ + { + "bbox": [ + 76, + 498, + 524, + 524 + ], + "type": "text", + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 89, + 527, + 148, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 527, + 148, + 539 + ], + "spans": [ + { + "bbox": [ + 89, + 527, + 148, + 539 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1585" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Zero-Shot and Few-Shot Stance Detection on Varied Topics via Conditional Generation/6aade9ed-d045-4bed-80d3-d9ddb4ea3243_content_list.json b/2023/Zero-Shot and Few-Shot Stance Detection on Varied Topics via Conditional Generation/6aade9ed-d045-4bed-80d3-d9ddb4ea3243_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..70918f8cb8b07cf06018b0bbc95fbddca2217611 --- /dev/null +++ b/2023/Zero-Shot and Few-Shot Stance Detection on Varied Topics via Conditional Generation/6aade9ed-d045-4bed-80d3-d9ddb4ea3243_content_list.json @@ -0,0 +1,1562 @@ +[ + { + "type": "text", + "text": "Zero-Shot and Few-Shot Stance Detection on Varied Topics via Conditional Generation", + "text_level": 1, + "bbox": [ + 115, + 89, + 884, + 130 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haoyang Wen and Alexander G. Hauptmann", + "bbox": [ + 300, + 155, + 699, + 172 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Language Technologies Institute, Carnegie Mellon University", + "bbox": [ + 250, + 172, + 749, + 187 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{hwen3, alex}@cs.cmu.edu", + "bbox": [ + 379, + 189, + 621, + 204 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 252, + 339, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zero-shot and few-shot stance detection identify the polarity of text with regard to a certain target when we have only limited or no training resources for the target. Previous work generally formulates the problem into a classification setting, ignoring the potential use of label text. In this paper, we instead utilize a conditional generation framework and formulate the problem as denoising from partially-filled templates, which can better utilize the semantics among input, label, and target texts. We further propose to jointly train an auxiliary task, target prediction, and to incorporate manually constructed incorrect samples with unlikelihood training to improve the representations for both target and label texts. We also verify the effectiveness of target-related Wikipedia knowledge with the generation framework. Experiments show that our proposed method significantly outperforms several strong baselines on VAST, and achieves new state-of-the-art performance. $^{1}$", + "bbox": [ + 141, + 280, + 460, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 114, + 592, + 260, + 607 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Stance detection is an important task that identifies the polarity of text with regard to certain target (Somasundaran and Wiebe, 2010; Augenstein et al., 2016; Mohammad et al., 2016; Sobhani et al., 2017; Allaway and McKeown, 2020), as shown in Table 1. It is crucial for understanding opinionated information expressed in natural language, and it can facilitate downstream social science analyses and applications (Zhang et al., 2017; Hanselowski et al., 2018; Jang and Allan, 2018).", + "bbox": [ + 112, + 618, + 489, + 778 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Previous work on stance detection mostly focuses on in-domain or leave-out targets with only a few target choices (Mohtarami et al., 2018; Xu et al., 2018; Graells-Garrido et al., 2020; Zhang et al., 2020; Liang et al., 2021; Allaway et al., 2021;", + "bbox": [ + 112, + 778, + 489, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Input Text: Airports and the roads on east nor west coast can not handle the present volume adequately as is. I did ride the vast trains in Europe, Japan and China and found them very comfortable and providing much better connections and more efficient.", + "bbox": [ + 517, + 250, + 880, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Target: high-speed rail Stance Label: Supportive (Pro)", + "bbox": [ + 519, + 310, + 877, + 323 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Table 1: A stance detection example from VAST.", + "bbox": [ + 527, + 334, + 862, + 349 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiang et al., 2022). Although achieving promising performance, those models are limited to generalize to a wide variety of targets. Zero-shot and few-shot stance detection on varied topics (VAST; Allaway and McKeown, 2020), instead, provides a diverse set of targets for training and testing. Efforts on this direction include involving graph modeling (Lin et al., 2021), common sense (Liu et al., 2021) or Wikipedia knowledge (He et al., 2022), and contrastive learning (Liang et al., 2022a,b). These methods generally formulate the problem into a classification setting, which directly trains the label representation from scratch, and does not fully utilize the semantics from those label and target texts.", + "bbox": [ + 507, + 382, + 884, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, connections among text semantics from input text, target, and label can be beneficial for stance detection. In this paper, we propose a new model by formulating the problem as a denoising task from text templates via conditional generation. Compared to direct classification, we can further exploit the label and topic semantics via learning to decode a series of natural language text containing the predicted label. The denoising scheme can also take advantage of the pretrained language model with similar pretraining task formulation (Lewis et al., 2020). To improve the target representation, we propose to jointly train target prediction with stance detection, which gives the input text and desired stance label to output possible targets. We use unlikelihood training (Welleck et al., 2020) that suppress the likelihood of manually constructed incorrect samples to enhance label", + "bbox": [ + 507, + 629, + 884, + 919 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1The resource for reproducing this paper is available at https://github.com/wenhycs/ACL2023-Zero-Shot-and-Few-Shot-Stance-Detection-on-Varied-Topics via -Conditional-Generation.", + "bbox": [ + 110, + 868, + 487, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1491", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics", + "bbox": [ + 226, + 945, + 771, + 957 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Volume 2: Short Papers, pages 1491-1499", + "bbox": [ + 368, + 958, + 630, + 971 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "July 9-14, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 295, + 972, + 700, + 985 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4bf32ce8c1161fd2fc799794b9e3e48e68aa1363f1db6efca4b907ce41afa9d4.jpg", + "image_caption": [ + "Figure 1: Overall framework of BART-based generation framework for stance detection." + ], + "image_footnote": [], + "bbox": [ + 121, + 87, + 480, + 237 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "representations. Recently, He et al. (2022) show the effectiveness of target-related Wikipedia knowledge for classification-based stance detection. We also follow the idea and incorporate target-related Wikipedia knowledge for our generation model.", + "bbox": [ + 112, + 305, + 487, + 384 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate our method on VAST. Experimental results show that the conditional generation formulation can achieve better performance compared to classification, demonstrating the effectiveness of connecting input, target, and label semantics for stance detection. Further analysis illustrates the benefits of joint target prediction, unlikelihood training, and Wikipedia knowledge. Our model can achieve new state-of-the-art performance, outperforming several strong baselines from previous work.", + "bbox": [ + 112, + 385, + 487, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Approach", + "text_level": 1, + "bbox": [ + 114, + 575, + 235, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we will discuss our approach to zero-shot and few-shot stance detection. We will first introduce the problem formulation, and then discuss our generation-based framework.", + "bbox": [ + 112, + 601, + 487, + 665 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 Problem Formulation", + "text_level": 1, + "bbox": [ + 114, + 677, + 331, + 690 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Stance detection aims to identify the polarity of an input text with regard to a specific target. Formally, a sample instance can be considered as a triple $(x,t,y)$ , where $x$ and $t$ are two sequences of tokens, representing input text and target respectively. $y\\in \\{\\mathrm{supportive}(\\mathrm{pro}),\\mathrm{opposite}(\\mathrm{con}),\\mathrm{neutral}\\}$ represents then stance label.", + "bbox": [ + 112, + 697, + 487, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A stance-detection model is to infer the stance label $y$ given $\\pmb{x}$ and $\\pmb{t}$ with parameter $\\theta$ :", + "bbox": [ + 112, + 810, + 487, + 843 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nf \\left(\\boldsymbol {x}, \\boldsymbol {t}; \\theta\\right) = y.\n$$\n", + "text_format": "latex", + "bbox": [ + 240, + 857, + 359, + 873 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the zero-shot and few-shot stance detection dataset with varied targets (Allaway and McKe", + "bbox": [ + 112, + 887, + 487, + 917 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "own, 2020), many target tokens only occur zero or a few times in the training set.", + "bbox": [ + 507, + 84, + 882, + 116 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2 A Generation-Based Framework", + "text_level": 1, + "bbox": [ + 507, + 126, + 811, + 140 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generation-based frameworks have demonstrated their effectiveness for problems beyond traditional generation tasks (Lewis and Fan, 2019; Yan et al., 2021; Li et al., 2021; Raffel et al., 2022). We use a conditional generation model for this problem, where the condition is a partially-filled template with the input text. The template is two sentences describing the target and stance with a placeholder for stance detection. An example of the partially-filled template with input text and output is shown in Figure 1.", + "bbox": [ + 507, + 147, + 882, + 323 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our base model is BART (Lewis et al., 2020), an encoder-decoder language model pretrained with denoising objectives, which is similar to our generation-based formulation. The generation process can be considered as using the conditional probability to select a new token at each step given input and previously generated tokens:", + "bbox": [ + 507, + 324, + 882, + 436 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\np \\left(\\boldsymbol {o} \\mid g \\left(\\boldsymbol {x}, \\boldsymbol {t}\\right); \\theta\\right) = \\prod_ {i = 1} ^ {| \\boldsymbol {o} |} p \\left(o _ {i} \\mid \\boldsymbol {o} _ {< i}, g \\left(\\boldsymbol {x}, \\boldsymbol {t}\\right); \\theta\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 443, + 873, + 487 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $g(\\pmb {x},\\pmb {t})$ represents the transformation function that fills the target $\\pmb{t}$ into the template and forms the input sequence with the input text $\\pmb{x}$ . Specifically, $g(\\pmb {x},\\pmb {t})$ will generate a combination of input text and template with special tokens: “ template x ”. The template contains two sentences: “The target is . The stance is ”. We will fill in placeholder with the actual target and keep the placeholder for the decoder to generate.", + "bbox": [ + 507, + 493, + 882, + 653 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The generated output $o$ is a fully-filled template, where both target and stance placeholders are replaced by actual or predicted values. The model is trained by minimizing the log-likelihood over the whole generated sequence:", + "bbox": [ + 507, + 653, + 882, + 734 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {s} = - \\log p (\\boldsymbol {o} \\mid g (\\boldsymbol {x}, t); \\theta) \\\\ = - \\sum_ {i = 1} ^ {| O |} \\log p \\left(o _ {i} \\mid o _ {< i}, g (\\boldsymbol {x}, t); \\theta\\right). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 741, + 842, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The final predicted stance label is obtained with a post-processing function that tries to find the polarity word after the prompt for stance.", + "bbox": [ + 507, + 812, + 882, + 860 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2.1 Joint Target Prediction", + "text_level": 1, + "bbox": [ + 507, + 868, + 752, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another advantage of using generation-based architecture is that we can leverage auxiliary generative", + "bbox": [ + 507, + 887, + 882, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1492", + "bbox": [ + 482, + 928, + 521, + 940 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/1ec07c3827b57e28d25f80c266b64154a72d7e929923d3717aac39e2c547f619.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Stance Detection
InputTarget is high-speed rail. Stance is<stance>.
OutputTarget is high-speed rail. Stance issupportive.
Target Prediction
InputStance is supportive. Target is<target>.
OutputStance is supportive. Target ishigh-speed rail.
Unlikelihood Training
InputTarget is high-speed rail. Stance is<stance>.
OutputTarget is high-speed rail. Stance isopposite.
", + "bbox": [ + 114, + 80, + 489, + 209 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 2: Examples input and output templates for stance detection, target prediction, and unlikelihood training.", + "bbox": [ + 112, + 217, + 487, + 247 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tasks to help train stance detection. We use target prediction, which is to infer the target tokens $t$ given stance label $y$ and input text $x$ :", + "bbox": [ + 112, + 269, + 489, + 317 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {t} (\\boldsymbol {x}, y; \\theta) = \\boldsymbol {t}.\n$$\n", + "text_format": "latex", + "bbox": [ + 238, + 326, + 361, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Target prediction can provide the connection of stance to target in an opposite direction of stance detection. It can also enhance the representation of target tokens by learning to decode them.", + "bbox": [ + 112, + 353, + 487, + 416 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The input sequence of target prediction is similar to stance detection, consisting of a partially-filled template and input text. The template used for joint target prediction is slightly different than the one used for stance detection, where we switch the position of two sentences so that the stance information shows up first. We will fill in the actual stance text in the input sequence, and leave the placeholder for the decoder to generate.", + "bbox": [ + 112, + 418, + 487, + 562 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2.2 Unlikelihood Training", + "text_level": 1, + "bbox": [ + 112, + 570, + 349, + 586 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Log-likelihood objective optimizes the likelihood over the entire distribution. However, in our task, especially when generating the stance labels, we should specifically focus on several candidate tokens. Therefore, we introduce unlikelihood training (Welleck et al., 2020), where we use unlikely tokens, i.e. incorrect stance predictions, to replace the ground-truth sequence and optimize with the unlikelihood loss for the replaced tokens.", + "bbox": [ + 112, + 589, + 487, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, for an output sequence $\\pmb{o}$ , we assume $o_k$ is the stance label and replaced it with an incorrect stance prediction $o_k'$ while keeping other tokens to form incorrect sequence $o'$ . The combination of likelihood and unlikelihood will be:", + "bbox": [ + 112, + 734, + 489, + 814 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {u} = \\log p \\left(o _ {k} ^ {\\prime} \\mid \\boldsymbol {o} _ {< k} ^ {\\prime}, g (\\boldsymbol {x}, \\boldsymbol {t}); \\theta\\right) \\\\ - \\sum_ {i \\neq k} \\log p \\left(o _ {i} ^ {\\prime} \\mid o _ {< i} ^ {\\prime}, g (\\boldsymbol {x}, \\boldsymbol {t}); \\theta\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 824, + 458, + 878 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For each ground-truth sequence, we can construct two sequences for unlikelihood training with the", + "bbox": [ + 112, + 887, + 487, + 917 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "other two incorrect stance labels. Table 2 illustrates the examples for different input and output templates for stance prediction, target prediction, and unlikelihood training.", + "bbox": [ + 507, + 84, + 884, + 149 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2.3 Incorporating Wikipedia Knowledge", + "text_level": 1, + "bbox": [ + 507, + 156, + 858, + 173 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "He et al. (2022) collect relevant Wikipedia snippets for each target and propose to incorporate Wikipedia knowledge to enhance target representations for BERT-based (Devlin et al., 2019) classification, which demonstrates a significant improvement. We follow He et al. (2022) and incorporate Wikipedia knowledge into our generation-based method. Specifically, we append Wikipedia snippets to the end of our input sequence: “ $$ template $\\langle /s\\rangle < / s\\rangle x$ $\\langle /s\\rangle < / s\\rangle$ Wikipedia snippet $\\langle /s\\rangle$ ”. We use the new input sequence to perform both training and inference while the output sequences remain as the fully-filled templates.", + "bbox": [ + 507, + 175, + 884, + 385 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2.4 Training Objective", + "text_level": 1, + "bbox": [ + 507, + 394, + 717, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The final training objective is the combination of loss functions from stance detection, target prediction, and unlikelihood training:", + "bbox": [ + 507, + 412, + 882, + 460 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {s} + \\alpha_ {t} \\mathcal {L} _ {t} + \\alpha_ {u} \\mathcal {L} _ {u},\n$$\n", + "text_format": "latex", + "bbox": [ + 600, + 472, + 791, + 488 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_t$ represents the log-likelihood loss over the output template for target prediction, $\\alpha_{t},\\alpha_{u}$ are used to balance different loss functions.", + "bbox": [ + 507, + 499, + 882, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 Experiments", + "text_level": 1, + "bbox": [ + 507, + 558, + 655, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1 Data", + "text_level": 1, + "bbox": [ + 509, + 583, + 594, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "VAST contains 18,548 examples from New York Times \"Room for Debate\" section with 5,630 different targets for zero-shot and few-shot stance detection. The original examples of VAST are collected from Habernal et al. (2018) under Apache-2.0 license2. We use Wikipedia knowledge collected by He et al. (2022), which uses API to crawl Wikipedia pages for targets. Wikipedia content can be used under Creative Commons Attribution Share-Alike license (CC-BY-SA)3. We use the same training/development/test split as Allaway and McKeown (2020).", + "bbox": [ + 507, + 602, + 884, + 796 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2 Experimental Setup", + "text_level": 1, + "bbox": [ + 507, + 807, + 714, + 822 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We conduct our experiments on VAST (Allaway and McKeown, 2020). We compare our model", + "bbox": [ + 507, + 828, + 882, + 860 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "$^{2}$ https://github.com/UKPLab/argument-reasoning-comprehension-task/blob/master/License", + "bbox": [ + 507, + 866, + 882, + 891 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "3https://en.wikipedia.org/wiki/Wikipedia:Reusing_Wikipedia_content", + "bbox": [ + 507, + 892, + 880, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1493", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/1eb1129e945c1b9414ec5383703471570bb731e8ce94291ec684ab64bbacf7a5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelPrecisionRecallF1
BERT Classification72.672.072.1
BART w/ Template75.775.175.3
+ Topic Prediction76.075.675.7
+Unlikelihood76.475.975.9
+Wikipedia78.077.377.4
", + "bbox": [ + 132, + 80, + 470, + 170 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/eaa78d12b608f420824ef696f6cb77dcd38a4f963653b44fe45601b2aebe71d8.jpg", + "table_caption": [ + "Table 3: Performance of different model variants on the overall precision, recall and $\\mathrm{F}_1$ on the development set (\\%). Each of our model variants is on top of the variant from its previous row." + ], + "table_footnote": [], + "table_body": "
ModelZero-ShotFew-ShotOverall
TGA-Net66.666.366.5
BERT-GCN68.669.769.2
CKE-Net70.270.170.1
WS-BERT75.373.674.5
Our Model76.478.077.3
", + "bbox": [ + 134, + 265, + 467, + 355 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 4: Stance detection performance $(\\%)$ on VAST. Our model significantly outperforms previous work on all metrics. Our results are obtained from averaging performances over 5 random seeds. $p < 0.001$ on overall $\\mathrm{F_1}$ using Z-test with variance as the standard deviation over multiple runs.", + "bbox": [ + 112, + 363, + 489, + 451 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with several existing systems including 1) TGA-Net (Allaway and McKeown, 2020); 2) BERTGCN (Lin et al., 2021); 3) CKE-Net (Liu et al., 2021); 4) WS-BERT (He et al., 2022). Following their setup, we use macro-average $\\mathrm{F}_1$ as the evaluation metric, and we report performance on the subset of test set for zero-shot and few-shot, and the overall test set.", + "bbox": [ + 112, + 491, + 489, + 618 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use BART-base $^4$ as our base model, of which the number of parameters is roughly consistent with baselines on BERT-base $^5$ . Our best model is optimized with AdamW (Loshchilov and Hutter, 2019) for 30 epochs with a learning rate of 1e-5. We use a linear scheduler with a warmup proportion of 0.1 and the training batch size is 32. We use greedy search during inference. We reported performances on development set and test set using the averaged results from 5 different random seeds. Test results are reported based on the best overall $\\mathrm{F_1}$ performance on the development set. $\\alpha_{t}$ is set to 1 and $\\alpha_{u}$ is set to 0.5. Our final model takes about 5 hours for training on one Nvidia RTX 3090 GPU.", + "bbox": [ + 112, + 627, + 489, + 853 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f9cd9180d99fc33a12ecf4039e1b8b13e7d4e3208478af6de17156cdc97d73d4.jpg", + "image_caption": [ + "(a) Our model", + "Figure 2: The t-SNE visualization of intermediate representations from our model and BERT classification model. Color map: Supportive, Opposite, Neutral." + ], + "image_footnote": [], + "bbox": [ + 519, + 84, + 694, + 206 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ee292fcfd6026bd798f7e0c63ebd63806084f7a1d24e58e3c1468a4fea63059f.jpg", + "image_caption": [ + "(b) BERT classification" + ], + "image_footnote": [], + "bbox": [ + 700, + 84, + 873, + 206 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3 Results", + "text_level": 1, + "bbox": [ + 507, + 309, + 613, + 323 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1 Comparing with Model Variants", + "text_level": 1, + "bbox": [ + 507, + 332, + 823, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first conduct comparison of some of our model variants to illustrate the effectiveness of our proposed components. The results are shown in Table 3. From the comparison of BERT-based classification (BERT Classification) and BART-based denoising generation from templates (BART w/ Template), we can find that adopting the generation framework can significantly improve the model performance. Our proposed topic prediction and un-likelihood training can further boost performance. The final model with knowledge from Wikipedia, verifies the effectiveness of Wikipedia knowledge for stance detection with a generative framework.", + "bbox": [ + 507, + 356, + 884, + 565 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.2 Comparing with Existing Systems", + "text_level": 1, + "bbox": [ + 507, + 581, + 835, + 596 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our overall performance is shown in Table 4. Our method can significantly outperform those previous baselines, indicating the effectiveness of our proposed generation framework for zero-shot and few-shot stance detection with varies topics.", + "bbox": [ + 507, + 602, + 882, + 684 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4 Qualitative Analysis", + "text_level": 1, + "bbox": [ + 507, + 702, + 714, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 2 show the t-SNE (van der Maaten and Hinton, 2008) visualization of intermediate representations before the classification layer from our model and BERT classification model on the development set. We use random initialization with perplexity as 50 for visualization and we color each visualized instance with its corresponding stance label. The visualization of BERT classification shows small clusters with hybrid labels, While we can see that instances with our generation method are clustered with labels, where neutral labels are at the top and supportive labels are generally at the bottom.", + "bbox": [ + 507, + 726, + 882, + 919 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "4https://huggingface.co/facebook/bart-base", + "bbox": [ + 134, + 890, + 455, + 904 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "5https://huggingface.co/bert-base-uncased", + "bbox": [ + 136, + 904, + 448, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1494", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4 Related Work", + "text_level": 1, + "bbox": [ + 114, + 83, + 270, + 98 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Zero-shot and few-shot stance detection. Zero-shot and few-shot stance detection focus on detecting stances for unseen or low-resource targets. Allaway and McKeown (2020) construct a dataset with varied topics that can be used to test stance detection under zero-shot and few-shot settings. Previous efforts mostly focus on modeling targets, documents, or their connections. Allaway and McKeown (2020) obtain generalized topic representation through clustering. Liu et al. (2021) use commonsense knowledge graph to enhance the connection between target and document. Liang et al. (2022a,b) use contrastive learning to learn target features. He et al. (2022) incorporate Wikipedia knowledge to enhance target representations. While in our work, we use a conditional generation framework to build the connections between input, target, and label text semantics.", + "bbox": [ + 112, + 112, + 492, + 401 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Text processing via conditional generation. Our work is also motivated by the recent success of tackling text processing problems as conditional generation (Lewis et al., 2020; Raffel et al., 2022). In addition to the conventional text generation problems, conditional generation frameworks are effectively applied in information extraction (Li et al., 2021), question answering (Lewis and Fan, 2019; Raffel et al., 2022) and sentiment analysis (Yan et al., 2021). In our work, we further explore stance detection via conditional generation.", + "bbox": [ + 112, + 414, + 490, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 112, + 606, + 247, + 621 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this paper, we propose a generation-based framework for zero-shot and few-shot stance detection that generate stance label from pre-defined templates. We further propose an auxiliary task, joint target prediction that takes stance label and input text to generate targets, and unlikelihood training on manually constructed incorrect generation output. Combining with Wikipedia knowledge for target from He et al. (2022), our model can achieve new state-of-the-art performance on VAST.", + "bbox": [ + 112, + 634, + 489, + 796 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Limitations", + "text_level": 1, + "bbox": [ + 112, + 810, + 220, + 826 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Because of the nature of our framework design, our work requires a diverse set of targets during training, which is important for target prediction and therefore the stance detection method. It is difficult to be applied to other stance detection datasets", + "bbox": [ + 112, + 839, + 490, + 919 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "when there are limited training resources with regard to targets, such as Conforti et al. (2020) and Mohammad et al. (2016). Besides, the model is trained on news-related debate corpus, so it may need further domain adaptation if applying the model to other domains such as social media.", + "bbox": [ + 507, + 84, + 884, + 179 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We are using an auto-regressive generation framework, which will also require extra inference time to generate the whole output sequence compared to the classification model. We would encourage readers to compare it with classification methods for efficiency when it will be applied in a time-sensitive scenario.", + "bbox": [ + 507, + 181, + 885, + 293 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 510, + 319, + 608, + 334 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Emily Allaway and Kathleen McKeown. 2020. Zero-Shot Stance Detection: A Dataset and Model using Generalized Topic Representations. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 8913-8931, Online. Association for Computational Linguistics.", + "Emily Allaway, Malavika Srikanth, and Kathleen McKeown. 2021. Adversarial learning for zero-shot stance detection on social media. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4756-4767, Online. Association for Computational Linguistics.", + "Isabelle Augenstein, Tim Rocktäschel, Andreas Vlachos, and Kalina Bontcheva. 2016. Stance detection with bidirectional conditional encoding. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 876-885, Austin, Texas. Association for Computational Linguistics.", + "Costanza Conforti, Jakob Berndt, Mohammad Taher Pilehvar, Chryssi Giannitsarou, Flavio Toxvaerd, and Nigel Collier. 2020. Will-they-won't-they: A very large dataset for stance detection on Twitter. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1715-1724, Online. Association for Computational Linguistics.", + "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics.", + "Eduardo Graells-Garrido, Ricardo Baeza-Yates, and Mounia Lalmas. 2020. Representativeness of abortion legislation debate on twitter: A case study in" + ], + "bbox": [ + 509, + 340, + 885, + 920 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1495", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "argentina and chile. In Companion Proceedings of the Web Conference 2020, WWW '20, page 765-774, New York, NY, USA. Association for Computing Machinery.", + "Ivan Habernal, Henning Wachsmuth, Iryna Gurevych, and Benno Stein. 2018. The argument reasoning comprehension task: Identification and reconstruction of implicit warrants. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1930-1940, New Orleans, Louisiana. Association for Computational Linguistics.", + "Andreas Hanselowski, Avinesh PVS, Benjamin Schiller, Felix Caspelherr, Debanjan Chaudhuri, Christian M. Meyer, and Iryna Gurevych. 2018. A retrospective analysis of the fake news challenge stance-detection task. In Proceedings of the 27th International Conference on Computational Linguistics, pages 1859-1874, Santa Fe, New Mexico, USA. Association for Computational Linguistics.", + "Zihao He, Negar Mokhberian, and Kristina Lerman. 2022. Infusing knowledge from Wikipedia to enhance stance detection. In Proceedings of the 12th Workshop on Computational Approaches to Subjectivity, Sentiment & Social Media Analysis, pages 71-77, Dublin, Ireland. Association for Computational Linguistics.", + "Myungha Jang and James Allan. 2018. Explaining controversy on social media via stance summarization. In The 41st International ACM SIGIR Conference on Research & Development in Information Retrieval, SIGIR '18, page 1221-1224, New York, NY, USA. Association for Computing Machinery.", + "Yan Jiang, Jinhua Gao, Huawei Shen, and Xueqi Cheng. 2022. Few-shot stance detection via target-aware prompt distillation. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '22, page 837-847, New York, NY, USA. Association for Computing Machinery.", + "Mike Lewis and Angela Fan. 2019. Generative question answering: Learning to answer the whole question. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net.", + "Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871-7880, Online. Association for Computational Linguistics.", + "Sha Li, Heng Ji, and Jiawei Han. 2021. Document-level event argument extraction by conditional generation." + ], + "bbox": [ + 115, + 85, + 489, + 917 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 894-908, Online. Association for Computational Linguistics.", + "Bin Liang, Zixiao Chen, Lin Gui, Yulan He, Min Yang, and Ruifeng Xu. 2022a. Zero-shot stance detection via contrastive learning. In Proceedings of the ACM Web Conference 2022, WWW '22, page 2738-2747, New York, NY, USA. Association for Computing Machinery.", + "Bin Liang, Yonghao Fu, Lin Gui, Min Yang, Jiachen Du, Yulan He, and Ruifeng Xu. 2021. Target-adaptive graph for cross-target stance detection. In Proceedings of the Web Conference 2021, WWW '21, page 3453-3464, New York, NY, USA. Association for Computing Machinery.", + "Bin Liang, Qinglin Zhu, Xiang Li, Min Yang, Lin Gui, Yulan He, and Ruifeng Xu. 2022b. JointCL: A joint contrastive learning framework for zero-shot stance detection. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 81-91, Dublin, Ireland. Association for Computational Linguistics.", + "Yuxiao Lin, Yuxian Meng, Xiaofei Sun, Qinghong Han, Kun Kuang, Jiwei Li, and Fei Wu. 2021. BertGCN: Transductive text classification by combining GNN and BERT. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 1456-1462, Online. Association for Computational Linguistics.", + "Rui Liu, Zheng Lin, Yutong Tan, and Weiping Wang. 2021. Enhancing zero-shot and few-shot stance detection with commonsense knowledge graph. In *Findings of the Association for Computational Linguistics: ACL-IJCNLP* 2021, pages 3152-3157, Online. Association for Computational Linguistics.", + "Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In International Conference on Learning Representations.", + "Saif Mohammad, Svetlana Kiritchenko, Parinaz Sobhani, Xiaodan Zhu, and Colin Cherry. 2016. SemEval-2016 task 6: Detecting stance in tweets. In Proceedings of the 10th International Workshop on Semantic Evaluation (SemEval-2016), pages 31-41, San Diego, California. Association for Computational Linguistics.", + "Mitra Mohtarami, Ramy Baly, James Glass, Preslav Nakov, Lluis Márquez, and Alessandro Moschitti. 2018. Automatic stance detection using end-to-end memory networks. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 767-776, New Orleans, Louisiana. Association for Computational Linguistics." + ], + "bbox": [ + 510, + 85, + 882, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1496", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2022. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(1).", + "Parinaz Sobhani, Diana Inkpen, and Xiaodan Zhu. 2017. A dataset for multi-target stance detection. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers, pages 551-557, Valencia, Spain. Association for Computational Linguistics.", + "Swapna Somasundaran and Janyce Wiebe. 2010. Recognizing stances in ideological on-line debates. In Proceedings of the NAACL HLT 2010 Workshop on Computational Approaches to Analysis and Generation of Emotion in Text, pages 116-124, Los Angeles, CA. Association for Computational Linguistics.", + "Laurens van der Maaten and Geoffrey Hinton. 2008. Visualizing data using t-sne. Journal of Machine Learning Research, 9(86):2579-2605.", + "Sean Welleck, Ilia Kulikov, Stephen Roller, Emily Dinan, Kyunghyun Cho, and Jason Weston. 2020. Neural text generation with unlikelihood training. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net.", + "Chang Xu, Cécile Paris, Surya Nepal, and Ross Sparks. 2018. Cross-target stance classification with self-attention networks. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 778-783, Melbourne, Australia. Association for Computational Linguistics.", + "Hang Yan, Junqi Dai, Tuo Ji, Xipeng Qiu, and Zheng Zhang. 2021. A unified generative framework for aspect-based sentiment analysis. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 2416-2429, Online. Association for Computational Linguistics.", + "Rong Zhang, Qifei Zhou, Bo An, Weiping Li, Tong Mo, and Bo Wu. 2020. Enhancing neural models with vulnerability via adversarial attack. In Proceedings of the 28th International Conference on Computational Linguistics, pages 1133-1146, Barcelona, Spain (Online). International Committee on Computational Linguistics.", + "Shaodian Zhang, Lin Qiu, Frank Chen, Weinan Zhang, Yong Yu, and Noémie Elhadad. 2017. We make choices we think are going to save us: Debate and stance identification for online breast cancer cam discussions. In Proceedings of the 26th International Conference on World Wide Web Companion, WWW '17 Companion, page 1073-1081, Republic and Canton of Geneva, CHE. International World Wide Web Conferences Steering Committee." + ], + "bbox": [ + 115, + 85, + 489, + 910 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1497", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A For every submission:", + "bbox": [ + 114, + 107, + 322, + 122 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "A1. Did you describe the limitations of your work? Limitations", + "A2. Did you discuss any potential risks of your work? Limitations", + "A3. Do the abstract and introduction summarize the paper's main claims? Abstract, Introduction", + "A4. Have you used AI writing assistants when working on this paper? Left blank." + ], + "bbox": [ + 129, + 126, + 695, + 285 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "B Did you use or create scientific artifacts?", + "bbox": [ + 114, + 297, + 487, + 313 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Introduction, Section 3.1 Data", + "bbox": [ + 131, + 319, + 359, + 332 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "B1. Did you cite the creators of artifacts you used? Introduction, Section 3.1 Data", + "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Section 3.1 Data", + "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Section 3.1 Data, Section 3.2 Experimental Setup" + ], + "bbox": [ + 129, + 343, + 880, + 508 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? We use an existing resource and detail of the data is discussed and introduced in their own published paper.", + "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? We use an existing resource and detail of the data is discussed and introduced in their own published paper.", + "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Section 3.1 Data" + ], + "bbox": [ + 129, + 520, + 880, + 778 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "C Did you run computational experiments?", + "bbox": [ + 114, + 790, + 492, + 807 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 132, + 813, + 213, + 827 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Section 3.2 Experimental Setup", + "bbox": [ + 129, + 837, + 880, + 885 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "ACL 2023 Responsible NLP Checklist", + "bbox": [ + 132, + 84, + 433, + 99 + ], + "page_idx": 7 + }, + { + "type": "footer", + "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.", + "bbox": [ + 112, + 892, + 877, + 916 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1498", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?", + "bbox": [ + 129, + 83, + 878, + 115 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Section 3.2 Experimental Setup", + "bbox": [ + 149, + 117, + 384, + 133 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?", + "bbox": [ + 129, + 142, + 882, + 190 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Section 3.2 Experimental Setup, Table 1", + "bbox": [ + 149, + 191, + 448, + 206 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?", + "bbox": [ + 129, + 217, + 882, + 265 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Section 3.2 Experimental Setup", + "bbox": [ + 149, + 267, + 386, + 282 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?", + "bbox": [ + 112, + 293, + 877, + 309 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 132, + 313, + 213, + 329 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?", + "bbox": [ + 127, + 340, + 882, + 372 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 373, + 248, + 388 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?", + "bbox": [ + 127, + 399, + 882, + 447 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 449, + 248, + 464 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?", + "bbox": [ + 127, + 475, + 882, + 521 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 524, + 248, + 539 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?", + "bbox": [ + 127, + 549, + 873, + 565 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 565, + 248, + 581 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?", + "bbox": [ + 127, + 592, + 880, + 623 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 626, + 248, + 640 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1499", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 8 + } +] \ No newline at end of file diff --git a/2023/Zero-Shot and Few-Shot Stance Detection on Varied Topics via Conditional Generation/6aade9ed-d045-4bed-80d3-d9ddb4ea3243_model.json b/2023/Zero-Shot and Few-Shot Stance Detection on Varied Topics via Conditional Generation/6aade9ed-d045-4bed-80d3-d9ddb4ea3243_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2e613135fb08ed8a577817df25a199fb43ad5c31 --- /dev/null +++ b/2023/Zero-Shot and Few-Shot Stance Detection on Varied Topics via Conditional Generation/6aade9ed-d045-4bed-80d3-d9ddb4ea3243_model.json @@ -0,0 +1,2022 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.116, + 0.09, + 0.885, + 0.131 + ], + "angle": 0, + "content": "Zero-Shot and Few-Shot Stance Detection on Varied Topics via Conditional Generation" + }, + { + "type": "text", + "bbox": [ + 0.302, + 0.156, + 0.7, + 0.173 + ], + "angle": 0, + "content": "Haoyang Wen and Alexander G. Hauptmann" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.173, + 0.751, + 0.189 + ], + "angle": 0, + "content": "Language Technologies Institute, Carnegie Mellon University" + }, + { + "type": "text", + "bbox": [ + 0.38, + 0.19, + 0.623, + 0.205 + ], + "angle": 0, + "content": "{hwen3, alex}@cs.cmu.edu" + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.253, + 0.341, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.281, + 0.461, + 0.581 + ], + "angle": 0, + "content": "Zero-shot and few-shot stance detection identify the polarity of text with regard to a certain target when we have only limited or no training resources for the target. Previous work generally formulates the problem into a classification setting, ignoring the potential use of label text. In this paper, we instead utilize a conditional generation framework and formulate the problem as denoising from partially-filled templates, which can better utilize the semantics among input, label, and target texts. We further propose to jointly train an auxiliary task, target prediction, and to incorporate manually constructed incorrect samples with unlikelihood training to improve the representations for both target and label texts. We also verify the effectiveness of target-related Wikipedia knowledge with the generation framework. Experiments show that our proposed method significantly outperforms several strong baselines on VAST, and achieves new state-of-the-art performance.\\(^{1}\\)" + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.593, + 0.262, + 0.608 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.619, + 0.49, + 0.779 + ], + "angle": 0, + "content": "Stance detection is an important task that identifies the polarity of text with regard to certain target (Somasundaran and Wiebe, 2010; Augenstein et al., 2016; Mohammad et al., 2016; Sobhani et al., 2017; Allaway and McKeown, 2020), as shown in Table 1. It is crucial for understanding opinionated information expressed in natural language, and it can facilitate downstream social science analyses and applications (Zhang et al., 2017; Hanselowski et al., 2018; Jang and Allan, 2018)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.78, + 0.49, + 0.86 + ], + "angle": 0, + "content": "Previous work on stance detection mostly focuses on in-domain or leave-out targets with only a few target choices (Mohtarami et al., 2018; Xu et al., 2018; Graells-Garrido et al., 2020; Zhang et al., 2020; Liang et al., 2021; Allaway et al., 2021;" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.251, + 0.882, + 0.31 + ], + "angle": 0, + "content": "Input Text: Airports and the roads on east nor west coast can not handle the present volume adequately as is. I did ride the vast trains in Europe, Japan and China and found them very comfortable and providing much better connections and more efficient." + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.311, + 0.878, + 0.324 + ], + "angle": 0, + "content": "Target: high-speed rail Stance Label: Supportive (Pro)" + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.335, + 0.863, + 0.35 + ], + "angle": 0, + "content": "Table 1: A stance detection example from VAST." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.384, + 0.885, + 0.626 + ], + "angle": 0, + "content": "Jiang et al., 2022). Although achieving promising performance, those models are limited to generalize to a wide variety of targets. Zero-shot and few-shot stance detection on varied topics (VAST; Allaway and McKeown, 2020), instead, provides a diverse set of targets for training and testing. Efforts on this direction include involving graph modeling (Lin et al., 2021), common sense (Liu et al., 2021) or Wikipedia knowledge (He et al., 2022), and contrastive learning (Liang et al., 2022a,b). These methods generally formulate the problem into a classification setting, which directly trains the label representation from scratch, and does not fully utilize the semantics from those label and target texts." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.63, + 0.885, + 0.92 + ], + "angle": 0, + "content": "However, connections among text semantics from input text, target, and label can be beneficial for stance detection. In this paper, we propose a new model by formulating the problem as a denoising task from text templates via conditional generation. Compared to direct classification, we can further exploit the label and topic semantics via learning to decode a series of natural language text containing the predicted label. The denoising scheme can also take advantage of the pretrained language model with similar pretraining task formulation (Lewis et al., 2020). To improve the target representation, we propose to jointly train target prediction with stance detection, which gives the input text and desired stance label to output possible targets. We use unlikelihood training (Welleck et al., 2020) that suppress the likelihood of manually constructed incorrect samples to enhance label" + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.869, + 0.488, + 0.918 + ], + "angle": 0, + "content": "1The resource for reproducing this paper is available at https://github.com/wenhycs/ACL2023-Zero-Shot-and-Few-Shot-Stance-Detection-on-Varied-Topics via -Conditional-Generation." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.928, + 0.52, + 0.941 + ], + "angle": 0, + "content": "1491" + }, + { + "type": "footer", + "bbox": [ + 0.227, + 0.946, + 0.772, + 0.958 + ], + "angle": 0, + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + }, + { + "type": "footer", + "bbox": [ + 0.369, + 0.959, + 0.631, + 0.972 + ], + "angle": 0, + "content": "Volume 2: Short Papers, pages 1491-1499" + }, + { + "type": "footer", + "bbox": [ + 0.297, + 0.973, + 0.702, + 0.986 + ], + "angle": 0, + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.122, + 0.088, + 0.482, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.251, + 0.489, + 0.281 + ], + "angle": 0, + "content": "Figure 1: Overall framework of BART-based generation framework for stance detection." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.306, + 0.489, + 0.385 + ], + "angle": 0, + "content": "representations. Recently, He et al. (2022) show the effectiveness of target-related Wikipedia knowledge for classification-based stance detection. We also follow the idea and incorporate target-related Wikipedia knowledge for our generation model." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.386, + 0.489, + 0.562 + ], + "angle": 0, + "content": "We evaluate our method on VAST. Experimental results show that the conditional generation formulation can achieve better performance compared to classification, demonstrating the effectiveness of connecting input, target, and label semantics for stance detection. Further analysis illustrates the benefits of joint target prediction, unlikelihood training, and Wikipedia knowledge. Our model can achieve new state-of-the-art performance, outperforming several strong baselines from previous work." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.576, + 0.236, + 0.592 + ], + "angle": 0, + "content": "2 Approach" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.602, + 0.489, + 0.667 + ], + "angle": 0, + "content": "In this section, we will discuss our approach to zero-shot and few-shot stance detection. We will first introduce the problem formulation, and then discuss our generation-based framework." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.678, + 0.332, + 0.692 + ], + "angle": 0, + "content": "2.1 Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.699, + 0.489, + 0.81 + ], + "angle": 0, + "content": "Stance detection aims to identify the polarity of an input text with regard to a specific target. Formally, a sample instance can be considered as a triple \\((x,t,y)\\), where \\(x\\) and \\(t\\) are two sequences of tokens, representing input text and target respectively. \\(y\\in \\{\\mathrm{supportive}(\\mathrm{pro}),\\mathrm{opposite}(\\mathrm{con}),\\mathrm{neutral}\\}\\) represents then stance label." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.812, + 0.488, + 0.844 + ], + "angle": 0, + "content": "A stance-detection model is to infer the stance label \\(y\\) given \\(\\pmb{x}\\) and \\(\\pmb{t}\\) with parameter \\(\\theta\\):" + }, + { + "type": "equation", + "bbox": [ + 0.242, + 0.858, + 0.36, + 0.875 + ], + "angle": 0, + "content": "\\[\nf \\left(\\boldsymbol {x}, \\boldsymbol {t}; \\theta\\right) = y.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.888, + 0.489, + 0.919 + ], + "angle": 0, + "content": "In the zero-shot and few-shot stance detection dataset with varied targets (Allaway and McKe" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.883, + 0.117 + ], + "angle": 0, + "content": "own, 2020), many target tokens only occur zero or a few times in the training set." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.127, + 0.813, + 0.141 + ], + "angle": 0, + "content": "2.2 A Generation-Based Framework" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.148, + 0.884, + 0.324 + ], + "angle": 0, + "content": "Generation-based frameworks have demonstrated their effectiveness for problems beyond traditional generation tasks (Lewis and Fan, 2019; Yan et al., 2021; Li et al., 2021; Raffel et al., 2022). We use a conditional generation model for this problem, where the condition is a partially-filled template with the input text. The template is two sentences describing the target and stance with a placeholder for stance detection. An example of the partially-filled template with input text and output is shown in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.325, + 0.884, + 0.437 + ], + "angle": 0, + "content": "Our base model is BART (Lewis et al., 2020), an encoder-decoder language model pretrained with denoising objectives, which is similar to our generation-based formulation. The generation process can be considered as using the conditional probability to select a new token at each step given input and previously generated tokens:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.444, + 0.875, + 0.488 + ], + "angle": 0, + "content": "\\[\np \\left(\\boldsymbol {o} \\mid g \\left(\\boldsymbol {x}, \\boldsymbol {t}\\right); \\theta\\right) = \\prod_ {i = 1} ^ {| \\boldsymbol {o} |} p \\left(o _ {i} \\mid \\boldsymbol {o} _ {< i}, g \\left(\\boldsymbol {x}, \\boldsymbol {t}\\right); \\theta\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.494, + 0.884, + 0.655 + ], + "angle": 0, + "content": "where \\(g(\\pmb {x},\\pmb {t})\\) represents the transformation function that fills the target \\(\\pmb{t}\\) into the template and forms the input sequence with the input text \\(\\pmb{x}\\). Specifically, \\(g(\\pmb {x},\\pmb {t})\\) will generate a combination of input text and template with special tokens: “ template x ”. The template contains two sentences: “The target is . The stance is ”. We will fill in placeholder with the actual target and keep the placeholder for the decoder to generate." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.655, + 0.884, + 0.735 + ], + "angle": 0, + "content": "The generated output \\( o \\) is a fully-filled template, where both target and stance placeholders are replaced by actual or predicted values. The model is trained by minimizing the log-likelihood over the whole generated sequence:" + }, + { + "type": "equation", + "bbox": [ + 0.546, + 0.742, + 0.843, + 0.807 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {s} = - \\log p (\\boldsymbol {o} \\mid g (\\boldsymbol {x}, t); \\theta) \\\\ = - \\sum_ {i = 1} ^ {| O |} \\log p \\left(o _ {i} \\mid o _ {< i}, g (\\boldsymbol {x}, t); \\theta\\right). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.813, + 0.884, + 0.861 + ], + "angle": 0, + "content": "The final predicted stance label is obtained with a post-processing function that tries to find the polarity word after the prompt for stance." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.869, + 0.753, + 0.884 + ], + "angle": 0, + "content": "2.2.1 Joint Target Prediction" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.888, + 0.884, + 0.919 + ], + "angle": 0, + "content": "Another advantage of using generation-based architecture is that we can leverage auxiliary generative" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1492" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.115, + 0.082, + 0.49, + 0.21 + ], + "angle": 0, + "content": "
Stance Detection
InputTarget is high-speed rail. Stance is<stance>.
OutputTarget is high-speed rail. Stance issupportive.
Target Prediction
InputStance is supportive. Target is<target>.
OutputStance is supportive. Target ishigh-speed rail.
Unlikelihood Training
InputTarget is high-speed rail. Stance is<stance>.
OutputTarget is high-speed rail. Stance isopposite.
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.218, + 0.489, + 0.248 + ], + "angle": 0, + "content": "Table 2: Examples input and output templates for stance detection, target prediction, and unlikelihood training." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.27, + 0.49, + 0.318 + ], + "angle": 0, + "content": "tasks to help train stance detection. We use target prediction, which is to infer the target tokens \\( t \\) given stance label \\( y \\) and input text \\( x \\):" + }, + { + "type": "equation", + "bbox": [ + 0.24, + 0.328, + 0.362, + 0.344 + ], + "angle": 0, + "content": "\\[\nf _ {t} (\\boldsymbol {x}, y; \\theta) = \\boldsymbol {t}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.354, + 0.489, + 0.417 + ], + "angle": 0, + "content": "Target prediction can provide the connection of stance to target in an opposite direction of stance detection. It can also enhance the representation of target tokens by learning to decode them." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.419, + 0.489, + 0.563 + ], + "angle": 0, + "content": "The input sequence of target prediction is similar to stance detection, consisting of a partially-filled template and input text. The template used for joint target prediction is slightly different than the one used for stance detection, where we switch the position of two sentences so that the stance information shows up first. We will fill in the actual stance text in the input sequence, and leave the placeholder for the decoder to generate." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.571, + 0.35, + 0.587 + ], + "angle": 0, + "content": "2.2.2 Unlikelihood Training" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.59, + 0.489, + 0.734 + ], + "angle": 0, + "content": "Log-likelihood objective optimizes the likelihood over the entire distribution. However, in our task, especially when generating the stance labels, we should specifically focus on several candidate tokens. Therefore, we introduce unlikelihood training (Welleck et al., 2020), where we use unlikely tokens, i.e. incorrect stance predictions, to replace the ground-truth sequence and optimize with the unlikelihood loss for the replaced tokens." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.736, + 0.49, + 0.815 + ], + "angle": 0, + "content": "Specifically, for an output sequence \\( \\pmb{o} \\), we assume \\( o_k \\) is the stance label and replaced it with an incorrect stance prediction \\( o_k' \\) while keeping other tokens to form incorrect sequence \\( o' \\). The combination of likelihood and unlikelihood will be:" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.825, + 0.46, + 0.879 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {u} = \\log p \\left(o _ {k} ^ {\\prime} \\mid \\boldsymbol {o} _ {< k} ^ {\\prime}, g (\\boldsymbol {x}, \\boldsymbol {t}); \\theta\\right) \\\\ - \\sum_ {i \\neq k} \\log p \\left(o _ {i} ^ {\\prime} \\mid o _ {< i} ^ {\\prime}, g (\\boldsymbol {x}, \\boldsymbol {t}); \\theta\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.888, + 0.489, + 0.919 + ], + "angle": 0, + "content": "For each ground-truth sequence, we can construct two sequences for unlikelihood training with the" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.15 + ], + "angle": 0, + "content": "other two incorrect stance labels. Table 2 illustrates the examples for different input and output templates for stance prediction, target prediction, and unlikelihood training." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.158, + 0.859, + 0.174 + ], + "angle": 0, + "content": "2.2.3 Incorporating Wikipedia Knowledge" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.177, + 0.885, + 0.386 + ], + "angle": 0, + "content": "He et al. (2022) collect relevant Wikipedia snippets for each target and propose to incorporate Wikipedia knowledge to enhance target representations for BERT-based (Devlin et al., 2019) classification, which demonstrates a significant improvement. We follow He et al. (2022) and incorporate Wikipedia knowledge into our generation-based method. Specifically, we append Wikipedia snippets to the end of our input sequence: “\\(\\) template \\(\\langle /s\\rangle < / s\\rangle x\\) \\(\\langle /s\\rangle < / s\\rangle\\) Wikipedia snippet \\(\\langle /s\\rangle\\)”. We use the new input sequence to perform both training and inference while the output sequences remain as the fully-filled templates." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.395, + 0.719, + 0.41 + ], + "angle": 0, + "content": "2.2.4 Training Objective" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.413, + 0.884, + 0.461 + ], + "angle": 0, + "content": "The final training objective is the combination of loss functions from stance detection, target prediction, and unlikelihood training:" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.473, + 0.792, + 0.489 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {s} + \\alpha_ {t} \\mathcal {L} _ {t} + \\alpha_ {u} \\mathcal {L} _ {u},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.5, + 0.884, + 0.547 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_t\\) represents the log-likelihood loss over the output template for target prediction, \\(\\alpha_{t},\\alpha_{u}\\) are used to balance different loss functions." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.559, + 0.657, + 0.575 + ], + "angle": 0, + "content": "3 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.585, + 0.596, + 0.599 + ], + "angle": 0, + "content": "3.1 Data" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.604, + 0.885, + 0.797 + ], + "angle": 0, + "content": "VAST contains 18,548 examples from New York Times \"Room for Debate\" section with 5,630 different targets for zero-shot and few-shot stance detection. The original examples of VAST are collected from Habernal et al. (2018) under Apache-2.0 license2. We use Wikipedia knowledge collected by He et al. (2022), which uses API to crawl Wikipedia pages for targets. Wikipedia content can be used under Creative Commons Attribution Share-Alike license (CC-BY-SA)3. We use the same training/development/test split as Allaway and McKeown (2020)." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.808, + 0.715, + 0.824 + ], + "angle": 0, + "content": "3.2 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.829, + 0.883, + 0.861 + ], + "angle": 0, + "content": "We conduct our experiments on VAST (Allaway and McKeown, 2020). We compare our model" + }, + { + "type": "page_footnote", + "bbox": [ + 0.508, + 0.868, + 0.883, + 0.892 + ], + "angle": 0, + "content": "\\(^{2}\\)https://github.com/UKPLab/argument-reasoning-comprehension-task/blob/master/License" + }, + { + "type": "page_footnote", + "bbox": [ + 0.508, + 0.894, + 0.882, + 0.918 + ], + "angle": 0, + "content": "3https://en.wikipedia.org/wiki/Wikipedia:Reusing_Wikipedia_content" + }, + { + "type": "list", + "bbox": [ + 0.508, + 0.868, + 0.883, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1493" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.134, + 0.082, + 0.472, + 0.171 + ], + "angle": 0, + "content": "
ModelPrecisionRecallF1
BERT Classification72.672.072.1
BART w/ Template75.775.175.3
+ Topic Prediction76.075.675.7
+Unlikelihood76.475.975.9
+Wikipedia78.077.377.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.181, + 0.489, + 0.238 + ], + "angle": 0, + "content": "Table 3: Performance of different model variants on the overall precision, recall and \\( \\mathrm{F}_1 \\) on the development set (\\%). Each of our model variants is on top of the variant from its previous row." + }, + { + "type": "table", + "bbox": [ + 0.136, + 0.266, + 0.468, + 0.356 + ], + "angle": 0, + "content": "
ModelZero-ShotFew-ShotOverall
TGA-Net66.666.366.5
BERT-GCN68.669.769.2
CKE-Net70.270.170.1
WS-BERT75.373.674.5
Our Model76.478.077.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.365, + 0.49, + 0.452 + ], + "angle": 0, + "content": "Table 4: Stance detection performance \\((\\%)\\) on VAST. Our model significantly outperforms previous work on all metrics. Our results are obtained from averaging performances over 5 random seeds. \\(p < 0.001\\) on overall \\(\\mathrm{F_1}\\) using Z-test with variance as the standard deviation over multiple runs." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.492, + 0.49, + 0.619 + ], + "angle": 0, + "content": "with several existing systems including 1) TGA-Net (Allaway and McKeown, 2020); 2) BERTGCN (Lin et al., 2021); 3) CKE-Net (Liu et al., 2021); 4) WS-BERT (He et al., 2022). Following their setup, we use macro-average \\(\\mathrm{F}_1\\) as the evaluation metric, and we report performance on the subset of test set for zero-shot and few-shot, and the overall test set." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.628, + 0.49, + 0.854 + ], + "angle": 0, + "content": "We use BART-base\\(^4\\) as our base model, of which the number of parameters is roughly consistent with baselines on BERT-base\\(^5\\). Our best model is optimized with AdamW (Loshchilov and Hutter, 2019) for 30 epochs with a learning rate of 1e-5. We use a linear scheduler with a warmup proportion of 0.1 and the training batch size is 32. We use greedy search during inference. We reported performances on development set and test set using the averaged results from 5 different random seeds. Test results are reported based on the best overall \\(\\mathrm{F_1}\\) performance on the development set. \\(\\alpha_{t}\\) is set to 1 and \\(\\alpha_{u}\\) is set to 0.5. Our final model takes about 5 hours for training on one Nvidia RTX 3090 GPU." + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.085, + 0.695, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.214, + 0.652, + 0.226 + ], + "angle": 0, + "content": "(a) Our model" + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.085, + 0.874, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.715, + 0.214, + 0.859, + 0.226 + ], + "angle": 0, + "content": "(b) BERT classification" + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.238, + 0.884, + 0.281 + ], + "angle": 0, + "content": "Figure 2: The t-SNE visualization of intermediate representations from our model and BERT classification model. Color map: Supportive, Opposite, Neutral." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.31, + 0.614, + 0.324 + ], + "angle": 0, + "content": "3.3 Results" + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.334, + 0.825, + 0.35 + ], + "angle": 0, + "content": "3.3.1 Comparing with Model Variants" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.357, + 0.885, + 0.567 + ], + "angle": 0, + "content": "We first conduct comparison of some of our model variants to illustrate the effectiveness of our proposed components. The results are shown in Table 3. From the comparison of BERT-based classification (BERT Classification) and BART-based denoising generation from templates (BART w/ Template), we can find that adopting the generation framework can significantly improve the model performance. Our proposed topic prediction and un-likelihood training can further boost performance. The final model with knowledge from Wikipedia, verifies the effectiveness of Wikipedia knowledge for stance detection with a generative framework." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.582, + 0.836, + 0.598 + ], + "angle": 0, + "content": "3.3.2 Comparing with Existing Systems" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.604, + 0.884, + 0.685 + ], + "angle": 0, + "content": "Our overall performance is shown in Table 4. Our method can significantly outperform those previous baselines, indicating the effectiveness of our proposed generation framework for zero-shot and few-shot stance detection with varies topics." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.703, + 0.715, + 0.718 + ], + "angle": 0, + "content": "3.4 Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.727, + 0.884, + 0.92 + ], + "angle": 0, + "content": "Figure 2 show the t-SNE (van der Maaten and Hinton, 2008) visualization of intermediate representations before the classification layer from our model and BERT classification model on the development set. We use random initialization with perplexity as 50 for visualization and we color each visualized instance with its corresponding stance label. The visualization of BERT classification shows small clusters with hybrid labels, While we can see that instances with our generation method are clustered with labels, where neutral labels are at the top and supportive labels are generally at the bottom." + }, + { + "type": "page_footnote", + "bbox": [ + 0.136, + 0.891, + 0.456, + 0.905 + ], + "angle": 0, + "content": "4https://huggingface.co/facebook/bart-base" + }, + { + "type": "page_footnote", + "bbox": [ + 0.137, + 0.905, + 0.45, + 0.918 + ], + "angle": 0, + "content": "5https://huggingface.co/bert-base-uncased" + }, + { + "type": "list", + "bbox": [ + 0.136, + 0.891, + 0.456, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1494" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.115, + 0.084, + 0.271, + 0.099 + ], + "angle": 0, + "content": "4 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.113, + 0.493, + 0.403 + ], + "angle": 0, + "content": "Zero-shot and few-shot stance detection. Zero-shot and few-shot stance detection focus on detecting stances for unseen or low-resource targets. Allaway and McKeown (2020) construct a dataset with varied topics that can be used to test stance detection under zero-shot and few-shot settings. Previous efforts mostly focus on modeling targets, documents, or their connections. Allaway and McKeown (2020) obtain generalized topic representation through clustering. Liu et al. (2021) use commonsense knowledge graph to enhance the connection between target and document. Liang et al. (2022a,b) use contrastive learning to learn target features. He et al. (2022) incorporate Wikipedia knowledge to enhance target representations. While in our work, we use a conditional generation framework to build the connections between input, target, and label text semantics." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.415, + 0.492, + 0.593 + ], + "angle": 0, + "content": "Text processing via conditional generation. Our work is also motivated by the recent success of tackling text processing problems as conditional generation (Lewis et al., 2020; Raffel et al., 2022). In addition to the conventional text generation problems, conditional generation frameworks are effectively applied in information extraction (Li et al., 2021), question answering (Lewis and Fan, 2019; Raffel et al., 2022) and sentiment analysis (Yan et al., 2021). In our work, we further explore stance detection via conditional generation." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.607, + 0.248, + 0.622 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.636, + 0.49, + 0.797 + ], + "angle": 0, + "content": "In this paper, we propose a generation-based framework for zero-shot and few-shot stance detection that generate stance label from pre-defined templates. We further propose an auxiliary task, joint target prediction that takes stance label and input text to generate targets, and unlikelihood training on manually constructed incorrect generation output. Combining with Wikipedia knowledge for target from He et al. (2022), our model can achieve new state-of-the-art performance on VAST." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.811, + 0.221, + 0.827 + ], + "angle": 0, + "content": "Limitations" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.84, + 0.491, + 0.92 + ], + "angle": 0, + "content": "Because of the nature of our framework design, our work requires a diverse set of targets during training, which is important for target prediction and therefore the stance detection method. It is difficult to be applied to other stance detection datasets" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.18 + ], + "angle": 0, + "content": "when there are limited training resources with regard to targets, such as Conforti et al. (2020) and Mohammad et al. (2016). Besides, the model is trained on news-related debate corpus, so it may need further domain adaptation if applying the model to other domains such as social media." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.182, + 0.886, + 0.294 + ], + "angle": 0, + "content": "We are using an auto-regressive generation framework, which will also require extra inference time to generate the whole output sequence compared to the classification model. We would encourage readers to compare it with classification methods for efficiency when it will be applied in a time-sensitive scenario." + }, + { + "type": "title", + "bbox": [ + 0.511, + 0.32, + 0.61, + 0.335 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.341, + 0.886, + 0.435 + ], + "angle": 0, + "content": "Emily Allaway and Kathleen McKeown. 2020. Zero-Shot Stance Detection: A Dataset and Model using Generalized Topic Representations. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 8913-8931, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.441, + 0.887, + 0.535 + ], + "angle": 0, + "content": "Emily Allaway, Malavika Srikanth, and Kathleen McKeown. 2021. Adversarial learning for zero-shot stance detection on social media. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4756-4767, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.54, + 0.886, + 0.634 + ], + "angle": 0, + "content": "Isabelle Augenstein, Tim Rocktäschel, Andreas Vlachos, and Kalina Bontcheva. 2016. Stance detection with bidirectional conditional encoding. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 876-885, Austin, Texas. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.64, + 0.886, + 0.745 + ], + "angle": 0, + "content": "Costanza Conforti, Jakob Berndt, Mohammad Taher Pilehvar, Chryssi Giannitsarou, Flavio Toxvaerd, and Nigel Collier. 2020. Will-they-won't-they: A very large dataset for stance detection on Twitter. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1715-1724, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.753, + 0.887, + 0.873 + ], + "angle": 0, + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.878, + 0.887, + 0.921 + ], + "angle": 0, + "content": "Eduardo Graells-Garrido, Ricardo Baeza-Yates, and Mounia Lalmas. 2020. Representativeness of abortion legislation debate on twitter: A case study in" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.341, + 0.887, + 0.921 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1495" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.135, + 0.086, + 0.49, + 0.14 + ], + "angle": 0, + "content": "argentina and chile. In Companion Proceedings of the Web Conference 2020, WWW '20, page 765-774, New York, NY, USA. Association for Computing Machinery." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.149, + 0.49, + 0.268 + ], + "angle": 0, + "content": "Ivan Habernal, Henning Wachsmuth, Iryna Gurevych, and Benno Stein. 2018. The argument reasoning comprehension task: Identification and reconstruction of implicit warrants. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1930-1940, New Orleans, Louisiana. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.278, + 0.49, + 0.383 + ], + "angle": 0, + "content": "Andreas Hanselowski, Avinesh PVS, Benjamin Schiller, Felix Caspelherr, Debanjan Chaudhuri, Christian M. Meyer, and Iryna Gurevych. 2018. A retrospective analysis of the fake news challenge stance-detection task. In Proceedings of the 27th International Conference on Computational Linguistics, pages 1859-1874, Santa Fe, New Mexico, USA. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.393, + 0.49, + 0.485 + ], + "angle": 0, + "content": "Zihao He, Negar Mokhberian, and Kristina Lerman. 2022. Infusing knowledge from Wikipedia to enhance stance detection. In Proceedings of the 12th Workshop on Computational Approaches to Subjectivity, Sentiment & Social Media Analysis, pages 71-77, Dublin, Ireland. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.495, + 0.49, + 0.575 + ], + "angle": 0, + "content": "Myungha Jang and James Allan. 2018. Explaining controversy on social media via stance summarization. In The 41st International ACM SIGIR Conference on Research & Development in Information Retrieval, SIGIR '18, page 1221-1224, New York, NY, USA. Association for Computing Machinery." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.585, + 0.49, + 0.676 + ], + "angle": 0, + "content": "Yan Jiang, Jinhua Gao, Huawei Shen, and Xueqi Cheng. 2022. Few-shot stance detection via target-aware prompt distillation. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '22, page 837-847, New York, NY, USA. Association for Computing Machinery." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.687, + 0.49, + 0.753 + ], + "angle": 0, + "content": "Mike Lewis and Angela Fan. 2019. Generative question answering: Learning to answer the whole question. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.763, + 0.49, + 0.881 + ], + "angle": 0, + "content": "Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871-7880, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.891, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Sha Li, Heng Ji, and Jiawei Han. 2021. Document-level event argument extraction by conditional generation." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.529, + 0.086, + 0.884, + 0.152 + ], + "angle": 0, + "content": "In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 894-908, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.164, + 0.884, + 0.243 + ], + "angle": 0, + "content": "Bin Liang, Zixiao Chen, Lin Gui, Yulan He, Min Yang, and Ruifeng Xu. 2022a. Zero-shot stance detection via contrastive learning. In Proceedings of the ACM Web Conference 2022, WWW '22, page 2738-2747, New York, NY, USA. Association for Computing Machinery." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.255, + 0.884, + 0.334 + ], + "angle": 0, + "content": "Bin Liang, Yonghao Fu, Lin Gui, Min Yang, Jiachen Du, Yulan He, and Ruifeng Xu. 2021. Target-adaptive graph for cross-target stance detection. In Proceedings of the Web Conference 2021, WWW '21, page 3453-3464, New York, NY, USA. Association for Computing Machinery." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.346, + 0.884, + 0.438 + ], + "angle": 0, + "content": "Bin Liang, Qinglin Zhu, Xiang Li, Min Yang, Lin Gui, Yulan He, and Ruifeng Xu. 2022b. JointCL: A joint contrastive learning framework for zero-shot stance detection. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 81-91, Dublin, Ireland. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.449, + 0.884, + 0.542 + ], + "angle": 0, + "content": "Yuxiao Lin, Yuxian Meng, Xiaofei Sun, Qinghong Han, Kun Kuang, Jiwei Li, and Fei Wu. 2021. BertGCN: Transductive text classification by combining GNN and BERT. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 1456-1462, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.554, + 0.884, + 0.633 + ], + "angle": 0, + "content": "Rui Liu, Zheng Lin, Yutong Tan, and Weiping Wang. 2021. Enhancing zero-shot and few-shot stance detection with commonsense knowledge graph. In *Findings of the Association for Computational Linguistics: ACL-IJCNLP* 2021, pages 3152-3157, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.645, + 0.884, + 0.685 + ], + "angle": 0, + "content": "Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In International Conference on Learning Representations." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.696, + 0.884, + 0.788 + ], + "angle": 0, + "content": "Saif Mohammad, Svetlana Kiritchenko, Parinaz Sobhani, Xiaodan Zhu, and Colin Cherry. 2016. SemEval-2016 task 6: Detecting stance in tweets. In Proceedings of the 10th International Workshop on Semantic Evaluation (SemEval-2016), pages 31-41, San Diego, California. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.8, + 0.884, + 0.919 + ], + "angle": 0, + "content": "Mitra Mohtarami, Ramy Baly, James Glass, Preslav Nakov, Lluis Márquez, and Alessandro Moschitti. 2018. Automatic stance detection using end-to-end memory networks. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 767-776, New Orleans, Louisiana. Association for Computational Linguistics." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.884, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1496" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.152 + ], + "angle": 0, + "content": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2022. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(1)." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.16, + 0.49, + 0.252 + ], + "angle": 0, + "content": "Parinaz Sobhani, Diana Inkpen, and Xiaodan Zhu. 2017. A dataset for multi-target stance detection. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers, pages 551-557, Valencia, Spain. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.26, + 0.49, + 0.34 + ], + "angle": 0, + "content": "Swapna Somasundaran and Janyce Wiebe. 2010. Recognizing stances in ideological on-line debates. In Proceedings of the NAACL HLT 2010 Workshop on Computational Approaches to Analysis and Generation of Emotion in Text, pages 116-124, Los Angeles, CA. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.347, + 0.489, + 0.386 + ], + "angle": 0, + "content": "Laurens van der Maaten and Geoffrey Hinton. 2008. Visualizing data using t-sne. Journal of Machine Learning Research, 9(86):2579-2605." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.395, + 0.49, + 0.473 + ], + "angle": 0, + "content": "Sean Welleck, Ilia Kulikov, Stephen Roller, Emily Dinan, Kyunghyun Cho, and Jason Weston. 2020. Neural text generation with unlikelihood training. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.482, + 0.49, + 0.573 + ], + "angle": 0, + "content": "Chang Xu, Cécile Paris, Surya Nepal, and Ross Sparks. 2018. Cross-target stance classification with self-attention networks. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 778-783, Melbourne, Australia. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.581, + 0.49, + 0.687 + ], + "angle": 0, + "content": "Hang Yan, Junqi Dai, Tuo Ji, Xipeng Qiu, and Zheng Zhang. 2021. A unified generative framework for aspect-based sentiment analysis. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 2416-2429, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.694, + 0.49, + 0.786 + ], + "angle": 0, + "content": "Rong Zhang, Qifei Zhou, Bo An, Weiping Li, Tong Mo, and Bo Wu. 2020. Enhancing neural models with vulnerability via adversarial attack. In Proceedings of the 28th International Conference on Computational Linguistics, pages 1133-1146, Barcelona, Spain (Online). International Committee on Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.794, + 0.49, + 0.912 + ], + "angle": 0, + "content": "Shaodian Zhang, Lin Qiu, Frank Chen, Weinan Zhang, Yong Yu, and Noémie Elhadad. 2017. We make choices we think are going to save us: Debate and stance identification for online breast cancer cam discussions. In Proceedings of the 26th International Conference on World Wide Web Companion, WWW '17 Companion, page 1073-1081, Republic and Canton of Geneva, CHE. International World Wide Web Conferences Steering Committee." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.912 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1497" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.133, + 0.085, + 0.435, + 0.1 + ], + "angle": 0, + "content": "ACL 2023 Responsible NLP Checklist" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.108, + 0.323, + 0.123 + ], + "angle": 0, + "content": "A For every submission:" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.127, + 0.533, + 0.158 + ], + "angle": 0, + "content": "A1. Did you describe the limitations of your work? Limitations" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.169, + 0.553, + 0.2 + ], + "angle": 0, + "content": "A2. Did you discuss any potential risks of your work? Limitations" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.212, + 0.696, + 0.243 + ], + "angle": 0, + "content": "A3. Do the abstract and introduction summarize the paper's main claims? Abstract, Introduction" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.255, + 0.669, + 0.286 + ], + "angle": 0, + "content": "A4. Have you used AI writing assistants when working on this paper? Left blank." + }, + { + "type": "list", + "bbox": [ + 0.13, + 0.127, + 0.696, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.298, + 0.489, + 0.314 + ], + "angle": 0, + "content": "B Did you use or create scientific artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.132, + 0.32, + 0.361, + 0.333 + ], + "angle": 0, + "content": "Introduction, Section 3.1 Data" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.344, + 0.53, + 0.375 + ], + "angle": 0, + "content": "B1. Did you cite the creators of artifacts you used? Introduction, Section 3.1 Data" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.386, + 0.779, + 0.417 + ], + "angle": 0, + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Section 3.1 Data" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.429, + 0.881, + 0.509 + ], + "angle": 0, + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Section 3.1 Data, Section 3.2 Experimental Setup" + }, + { + "type": "list", + "bbox": [ + 0.131, + 0.344, + 0.881, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.521, + 0.881, + 0.601 + ], + "angle": 0, + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? We use an existing resource and detail of the data is discussed and introduced in their own published paper." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.612, + 0.881, + 0.676 + ], + "angle": 0, + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? We use an existing resource and detail of the data is discussed and introduced in their own published paper." + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.685, + 0.882, + 0.78 + ], + "angle": 0, + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Section 3.1 Data" + }, + { + "type": "list", + "bbox": [ + 0.13, + 0.521, + 0.882, + 0.78 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.791, + 0.494, + 0.808 + ], + "angle": 0, + "content": "C Did you run computational experiments?" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.814, + 0.215, + 0.828 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.838, + 0.881, + 0.887 + ], + "angle": 0, + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Section 3.2 Experimental Setup" + }, + { + "type": "footer", + "bbox": [ + 0.114, + 0.893, + 0.878, + 0.917 + ], + "angle": 0, + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1498" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.13, + 0.084, + 0.88, + 0.116 + ], + "angle": 0, + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.118, + 0.386, + 0.134 + ], + "angle": 0, + "content": "Section 3.2 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.143, + 0.884, + 0.191 + ], + "angle": 0, + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.192, + 0.449, + 0.208 + ], + "angle": 0, + "content": "Section 3.2 Experimental Setup, Table 1" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.218, + 0.884, + 0.266 + ], + "angle": 0, + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.268, + 0.387, + 0.283 + ], + "angle": 0, + "content": "Section 3.2 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.294, + 0.878, + 0.31 + ], + "angle": 0, + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.315, + 0.215, + 0.33 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.341, + 0.883, + 0.373 + ], + "angle": 0, + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.374, + 0.25, + 0.39 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.4, + 0.884, + 0.448 + ], + "angle": 0, + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.45, + 0.25, + 0.465 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.476, + 0.884, + 0.523 + ], + "angle": 0, + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.525, + 0.25, + 0.54 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.55, + 0.875, + 0.566 + ], + "angle": 0, + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.567, + 0.25, + 0.582 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.593, + 0.881, + 0.624 + ], + "angle": 0, + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.627, + 0.25, + 0.642 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1499" + } + ] +] \ No newline at end of file diff --git a/2023/Zero-Shot and Few-Shot Stance Detection on Varied Topics via Conditional Generation/6aade9ed-d045-4bed-80d3-d9ddb4ea3243_origin.pdf b/2023/Zero-Shot and Few-Shot Stance Detection on Varied Topics via Conditional Generation/6aade9ed-d045-4bed-80d3-d9ddb4ea3243_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..62ef38c9e07514de7e59c8682bfb5a064955bb87 --- /dev/null +++ b/2023/Zero-Shot and Few-Shot Stance Detection on Varied Topics via Conditional Generation/6aade9ed-d045-4bed-80d3-d9ddb4ea3243_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3a7aa676e45fc36320b79ec4f15e4ef188629e93b80373862ee2d4cd962d0174 +size 394537 diff --git a/2023/Zero-Shot and Few-Shot Stance Detection on Varied Topics via Conditional Generation/full.md b/2023/Zero-Shot and Few-Shot Stance Detection on Varied Topics via Conditional Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..59fa0d5f3780a7a2626e368604834630f9d2966b --- /dev/null +++ b/2023/Zero-Shot and Few-Shot Stance Detection on Varied Topics via Conditional Generation/full.md @@ -0,0 +1,279 @@ +# Zero-Shot and Few-Shot Stance Detection on Varied Topics via Conditional Generation + +Haoyang Wen and Alexander G. Hauptmann + +Language Technologies Institute, Carnegie Mellon University + +{hwen3, alex}@cs.cmu.edu + +# Abstract + +Zero-shot and few-shot stance detection identify the polarity of text with regard to a certain target when we have only limited or no training resources for the target. Previous work generally formulates the problem into a classification setting, ignoring the potential use of label text. In this paper, we instead utilize a conditional generation framework and formulate the problem as denoising from partially-filled templates, which can better utilize the semantics among input, label, and target texts. We further propose to jointly train an auxiliary task, target prediction, and to incorporate manually constructed incorrect samples with unlikelihood training to improve the representations for both target and label texts. We also verify the effectiveness of target-related Wikipedia knowledge with the generation framework. Experiments show that our proposed method significantly outperforms several strong baselines on VAST, and achieves new state-of-the-art performance. $^{1}$ + +# 1 Introduction + +Stance detection is an important task that identifies the polarity of text with regard to certain target (Somasundaran and Wiebe, 2010; Augenstein et al., 2016; Mohammad et al., 2016; Sobhani et al., 2017; Allaway and McKeown, 2020), as shown in Table 1. It is crucial for understanding opinionated information expressed in natural language, and it can facilitate downstream social science analyses and applications (Zhang et al., 2017; Hanselowski et al., 2018; Jang and Allan, 2018). + +Previous work on stance detection mostly focuses on in-domain or leave-out targets with only a few target choices (Mohtarami et al., 2018; Xu et al., 2018; Graells-Garrido et al., 2020; Zhang et al., 2020; Liang et al., 2021; Allaway et al., 2021; + +Input Text: Airports and the roads on east nor west coast can not handle the present volume adequately as is. I did ride the vast trains in Europe, Japan and China and found them very comfortable and providing much better connections and more efficient. + +Target: high-speed rail Stance Label: Supportive (Pro) + +Table 1: A stance detection example from VAST. + +Jiang et al., 2022). Although achieving promising performance, those models are limited to generalize to a wide variety of targets. Zero-shot and few-shot stance detection on varied topics (VAST; Allaway and McKeown, 2020), instead, provides a diverse set of targets for training and testing. Efforts on this direction include involving graph modeling (Lin et al., 2021), common sense (Liu et al., 2021) or Wikipedia knowledge (He et al., 2022), and contrastive learning (Liang et al., 2022a,b). These methods generally formulate the problem into a classification setting, which directly trains the label representation from scratch, and does not fully utilize the semantics from those label and target texts. + +However, connections among text semantics from input text, target, and label can be beneficial for stance detection. In this paper, we propose a new model by formulating the problem as a denoising task from text templates via conditional generation. Compared to direct classification, we can further exploit the label and topic semantics via learning to decode a series of natural language text containing the predicted label. The denoising scheme can also take advantage of the pretrained language model with similar pretraining task formulation (Lewis et al., 2020). To improve the target representation, we propose to jointly train target prediction with stance detection, which gives the input text and desired stance label to output possible targets. We use unlikelihood training (Welleck et al., 2020) that suppress the likelihood of manually constructed incorrect samples to enhance label + +![](images/4bf32ce8c1161fd2fc799794b9e3e48e68aa1363f1db6efca4b907ce41afa9d4.jpg) +Figure 1: Overall framework of BART-based generation framework for stance detection. + +representations. Recently, He et al. (2022) show the effectiveness of target-related Wikipedia knowledge for classification-based stance detection. We also follow the idea and incorporate target-related Wikipedia knowledge for our generation model. + +We evaluate our method on VAST. Experimental results show that the conditional generation formulation can achieve better performance compared to classification, demonstrating the effectiveness of connecting input, target, and label semantics for stance detection. Further analysis illustrates the benefits of joint target prediction, unlikelihood training, and Wikipedia knowledge. Our model can achieve new state-of-the-art performance, outperforming several strong baselines from previous work. + +# 2 Approach + +In this section, we will discuss our approach to zero-shot and few-shot stance detection. We will first introduce the problem formulation, and then discuss our generation-based framework. + +# 2.1 Problem Formulation + +Stance detection aims to identify the polarity of an input text with regard to a specific target. Formally, a sample instance can be considered as a triple $(x,t,y)$ , where $x$ and $t$ are two sequences of tokens, representing input text and target respectively. $y\in \{\mathrm{supportive}(\mathrm{pro}),\mathrm{opposite}(\mathrm{con}),\mathrm{neutral}\}$ represents then stance label. + +A stance-detection model is to infer the stance label $y$ given $\pmb{x}$ and $\pmb{t}$ with parameter $\theta$ : + +$$ +f \left(\boldsymbol {x}, \boldsymbol {t}; \theta\right) = y. +$$ + +In the zero-shot and few-shot stance detection dataset with varied targets (Allaway and McKe + +own, 2020), many target tokens only occur zero or a few times in the training set. + +# 2.2 A Generation-Based Framework + +Generation-based frameworks have demonstrated their effectiveness for problems beyond traditional generation tasks (Lewis and Fan, 2019; Yan et al., 2021; Li et al., 2021; Raffel et al., 2022). We use a conditional generation model for this problem, where the condition is a partially-filled template with the input text. The template is two sentences describing the target and stance with a placeholder for stance detection. An example of the partially-filled template with input text and output is shown in Figure 1. + +Our base model is BART (Lewis et al., 2020), an encoder-decoder language model pretrained with denoising objectives, which is similar to our generation-based formulation. The generation process can be considered as using the conditional probability to select a new token at each step given input and previously generated tokens: + +$$ +p \left(\boldsymbol {o} \mid g \left(\boldsymbol {x}, \boldsymbol {t}\right); \theta\right) = \prod_ {i = 1} ^ {| \boldsymbol {o} |} p \left(o _ {i} \mid \boldsymbol {o} _ {< i}, g \left(\boldsymbol {x}, \boldsymbol {t}\right); \theta\right), +$$ + +where $g(\pmb {x},\pmb {t})$ represents the transformation function that fills the target $\pmb{t}$ into the template and forms the input sequence with the input text $\pmb{x}$ . Specifically, $g(\pmb {x},\pmb {t})$ will generate a combination of input text and template with special tokens: “ template x ”. The template contains two sentences: “The target is . The stance is ”. We will fill in placeholder with the actual target and keep the placeholder for the decoder to generate. + +The generated output $o$ is a fully-filled template, where both target and stance placeholders are replaced by actual or predicted values. The model is trained by minimizing the log-likelihood over the whole generated sequence: + +$$ +\begin{array}{l} \mathcal {L} _ {s} = - \log p (\boldsymbol {o} \mid g (\boldsymbol {x}, t); \theta) \\ = - \sum_ {i = 1} ^ {| O |} \log p \left(o _ {i} \mid o _ {< i}, g (\boldsymbol {x}, t); \theta\right). \\ \end{array} +$$ + +The final predicted stance label is obtained with a post-processing function that tries to find the polarity word after the prompt for stance. + +# 2.2.1 Joint Target Prediction + +Another advantage of using generation-based architecture is that we can leverage auxiliary generative + +
Stance Detection
InputTarget is high-speed rail. Stance is<stance>.
OutputTarget is high-speed rail. Stance issupportive.
Target Prediction
InputStance is supportive. Target is<target>.
OutputStance is supportive. Target ishigh-speed rail.
Unlikelihood Training
InputTarget is high-speed rail. Stance is<stance>.
OutputTarget is high-speed rail. Stance isopposite.
+ +Table 2: Examples input and output templates for stance detection, target prediction, and unlikelihood training. + +tasks to help train stance detection. We use target prediction, which is to infer the target tokens $t$ given stance label $y$ and input text $x$ : + +$$ +f _ {t} (\boldsymbol {x}, y; \theta) = \boldsymbol {t}. +$$ + +Target prediction can provide the connection of stance to target in an opposite direction of stance detection. It can also enhance the representation of target tokens by learning to decode them. + +The input sequence of target prediction is similar to stance detection, consisting of a partially-filled template and input text. The template used for joint target prediction is slightly different than the one used for stance detection, where we switch the position of two sentences so that the stance information shows up first. We will fill in the actual stance text in the input sequence, and leave the placeholder for the decoder to generate. + +# 2.2.2 Unlikelihood Training + +Log-likelihood objective optimizes the likelihood over the entire distribution. However, in our task, especially when generating the stance labels, we should specifically focus on several candidate tokens. Therefore, we introduce unlikelihood training (Welleck et al., 2020), where we use unlikely tokens, i.e. incorrect stance predictions, to replace the ground-truth sequence and optimize with the unlikelihood loss for the replaced tokens. + +Specifically, for an output sequence $\pmb{o}$ , we assume $o_k$ is the stance label and replaced it with an incorrect stance prediction $o_k'$ while keeping other tokens to form incorrect sequence $o'$ . The combination of likelihood and unlikelihood will be: + +$$ +\begin{array}{l} \mathcal {L} _ {u} = \log p \left(o _ {k} ^ {\prime} \mid \boldsymbol {o} _ {< k} ^ {\prime}, g (\boldsymbol {x}, \boldsymbol {t}); \theta\right) \\ - \sum_ {i \neq k} \log p \left(o _ {i} ^ {\prime} \mid o _ {< i} ^ {\prime}, g (\boldsymbol {x}, \boldsymbol {t}); \theta\right), \\ \end{array} +$$ + +For each ground-truth sequence, we can construct two sequences for unlikelihood training with the + +other two incorrect stance labels. Table 2 illustrates the examples for different input and output templates for stance prediction, target prediction, and unlikelihood training. + +# 2.2.3 Incorporating Wikipedia Knowledge + +He et al. (2022) collect relevant Wikipedia snippets for each target and propose to incorporate Wikipedia knowledge to enhance target representations for BERT-based (Devlin et al., 2019) classification, which demonstrates a significant improvement. We follow He et al. (2022) and incorporate Wikipedia knowledge into our generation-based method. Specifically, we append Wikipedia snippets to the end of our input sequence: “ $$ template $\langle /s\rangle < / s\rangle x$ $\langle /s\rangle < / s\rangle$ Wikipedia snippet $\langle /s\rangle$ ”. We use the new input sequence to perform both training and inference while the output sequences remain as the fully-filled templates. + +# 2.2.4 Training Objective + +The final training objective is the combination of loss functions from stance detection, target prediction, and unlikelihood training: + +$$ +\mathcal {L} = \mathcal {L} _ {s} + \alpha_ {t} \mathcal {L} _ {t} + \alpha_ {u} \mathcal {L} _ {u}, +$$ + +where $\mathcal{L}_t$ represents the log-likelihood loss over the output template for target prediction, $\alpha_{t},\alpha_{u}$ are used to balance different loss functions. + +# 3 Experiments + +# 3.1 Data + +VAST contains 18,548 examples from New York Times "Room for Debate" section with 5,630 different targets for zero-shot and few-shot stance detection. The original examples of VAST are collected from Habernal et al. (2018) under Apache-2.0 license2. We use Wikipedia knowledge collected by He et al. (2022), which uses API to crawl Wikipedia pages for targets. Wikipedia content can be used under Creative Commons Attribution Share-Alike license (CC-BY-SA)3. We use the same training/development/test split as Allaway and McKeown (2020). + +# 3.2 Experimental Setup + +We conduct our experiments on VAST (Allaway and McKeown, 2020). We compare our model + +
ModelPrecisionRecallF1
BERT Classification72.672.072.1
BART w/ Template75.775.175.3
+ Topic Prediction76.075.675.7
+Unlikelihood76.475.975.9
+Wikipedia78.077.377.4
+ +Table 3: Performance of different model variants on the overall precision, recall and $\mathrm{F}_1$ on the development set (\%). Each of our model variants is on top of the variant from its previous row. + +
ModelZero-ShotFew-ShotOverall
TGA-Net66.666.366.5
BERT-GCN68.669.769.2
CKE-Net70.270.170.1
WS-BERT75.373.674.5
Our Model76.478.077.3
+ +Table 4: Stance detection performance $(\%)$ on VAST. Our model significantly outperforms previous work on all metrics. Our results are obtained from averaging performances over 5 random seeds. $p < 0.001$ on overall $\mathrm{F_1}$ using Z-test with variance as the standard deviation over multiple runs. + +with several existing systems including 1) TGA-Net (Allaway and McKeown, 2020); 2) BERTGCN (Lin et al., 2021); 3) CKE-Net (Liu et al., 2021); 4) WS-BERT (He et al., 2022). Following their setup, we use macro-average $\mathrm{F}_1$ as the evaluation metric, and we report performance on the subset of test set for zero-shot and few-shot, and the overall test set. + +We use BART-base $^4$ as our base model, of which the number of parameters is roughly consistent with baselines on BERT-base $^5$ . Our best model is optimized with AdamW (Loshchilov and Hutter, 2019) for 30 epochs with a learning rate of 1e-5. We use a linear scheduler with a warmup proportion of 0.1 and the training batch size is 32. We use greedy search during inference. We reported performances on development set and test set using the averaged results from 5 different random seeds. Test results are reported based on the best overall $\mathrm{F_1}$ performance on the development set. $\alpha_{t}$ is set to 1 and $\alpha_{u}$ is set to 0.5. Our final model takes about 5 hours for training on one Nvidia RTX 3090 GPU. + +![](images/f9cd9180d99fc33a12ecf4039e1b8b13e7d4e3208478af6de17156cdc97d73d4.jpg) +(a) Our model +Figure 2: The t-SNE visualization of intermediate representations from our model and BERT classification model. Color map: Supportive, Opposite, Neutral. + +![](images/ee292fcfd6026bd798f7e0c63ebd63806084f7a1d24e58e3c1468a4fea63059f.jpg) +(b) BERT classification + +# 3.3 Results + +# 3.3.1 Comparing with Model Variants + +We first conduct comparison of some of our model variants to illustrate the effectiveness of our proposed components. The results are shown in Table 3. From the comparison of BERT-based classification (BERT Classification) and BART-based denoising generation from templates (BART w/ Template), we can find that adopting the generation framework can significantly improve the model performance. Our proposed topic prediction and un-likelihood training can further boost performance. The final model with knowledge from Wikipedia, verifies the effectiveness of Wikipedia knowledge for stance detection with a generative framework. + +# 3.3.2 Comparing with Existing Systems + +Our overall performance is shown in Table 4. Our method can significantly outperform those previous baselines, indicating the effectiveness of our proposed generation framework for zero-shot and few-shot stance detection with varies topics. + +# 3.4 Qualitative Analysis + +Figure 2 show the t-SNE (van der Maaten and Hinton, 2008) visualization of intermediate representations before the classification layer from our model and BERT classification model on the development set. We use random initialization with perplexity as 50 for visualization and we color each visualized instance with its corresponding stance label. The visualization of BERT classification shows small clusters with hybrid labels, While we can see that instances with our generation method are clustered with labels, where neutral labels are at the top and supportive labels are generally at the bottom. + +# 4 Related Work + +Zero-shot and few-shot stance detection. Zero-shot and few-shot stance detection focus on detecting stances for unseen or low-resource targets. Allaway and McKeown (2020) construct a dataset with varied topics that can be used to test stance detection under zero-shot and few-shot settings. Previous efforts mostly focus on modeling targets, documents, or their connections. Allaway and McKeown (2020) obtain generalized topic representation through clustering. Liu et al. (2021) use commonsense knowledge graph to enhance the connection between target and document. Liang et al. (2022a,b) use contrastive learning to learn target features. He et al. (2022) incorporate Wikipedia knowledge to enhance target representations. While in our work, we use a conditional generation framework to build the connections between input, target, and label text semantics. + +Text processing via conditional generation. Our work is also motivated by the recent success of tackling text processing problems as conditional generation (Lewis et al., 2020; Raffel et al., 2022). In addition to the conventional text generation problems, conditional generation frameworks are effectively applied in information extraction (Li et al., 2021), question answering (Lewis and Fan, 2019; Raffel et al., 2022) and sentiment analysis (Yan et al., 2021). In our work, we further explore stance detection via conditional generation. + +# 5 Conclusion + +In this paper, we propose a generation-based framework for zero-shot and few-shot stance detection that generate stance label from pre-defined templates. We further propose an auxiliary task, joint target prediction that takes stance label and input text to generate targets, and unlikelihood training on manually constructed incorrect generation output. Combining with Wikipedia knowledge for target from He et al. (2022), our model can achieve new state-of-the-art performance on VAST. + +# Limitations + +Because of the nature of our framework design, our work requires a diverse set of targets during training, which is important for target prediction and therefore the stance detection method. It is difficult to be applied to other stance detection datasets + +when there are limited training resources with regard to targets, such as Conforti et al. (2020) and Mohammad et al. (2016). Besides, the model is trained on news-related debate corpus, so it may need further domain adaptation if applying the model to other domains such as social media. + +We are using an auto-regressive generation framework, which will also require extra inference time to generate the whole output sequence compared to the classification model. We would encourage readers to compare it with classification methods for efficiency when it will be applied in a time-sensitive scenario. + +# References + +Emily Allaway and Kathleen McKeown. 2020. Zero-Shot Stance Detection: A Dataset and Model using Generalized Topic Representations. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 8913-8931, Online. Association for Computational Linguistics. +Emily Allaway, Malavika Srikanth, and Kathleen McKeown. 2021. Adversarial learning for zero-shot stance detection on social media. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4756-4767, Online. Association for Computational Linguistics. +Isabelle Augenstein, Tim Rocktäschel, Andreas Vlachos, and Kalina Bontcheva. 2016. Stance detection with bidirectional conditional encoding. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 876-885, Austin, Texas. Association for Computational Linguistics. +Costanza Conforti, Jakob Berndt, Mohammad Taher Pilehvar, Chryssi Giannitsarou, Flavio Toxvaerd, and Nigel Collier. 2020. Will-they-won't-they: A very large dataset for stance detection on Twitter. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1715-1724, Online. Association for Computational Linguistics. +Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics. +Eduardo Graells-Garrido, Ricardo Baeza-Yates, and Mounia Lalmas. 2020. Representativeness of abortion legislation debate on twitter: A case study in + +argentina and chile. In Companion Proceedings of the Web Conference 2020, WWW '20, page 765-774, New York, NY, USA. Association for Computing Machinery. +Ivan Habernal, Henning Wachsmuth, Iryna Gurevych, and Benno Stein. 2018. The argument reasoning comprehension task: Identification and reconstruction of implicit warrants. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1930-1940, New Orleans, Louisiana. Association for Computational Linguistics. +Andreas Hanselowski, Avinesh PVS, Benjamin Schiller, Felix Caspelherr, Debanjan Chaudhuri, Christian M. Meyer, and Iryna Gurevych. 2018. A retrospective analysis of the fake news challenge stance-detection task. In Proceedings of the 27th International Conference on Computational Linguistics, pages 1859-1874, Santa Fe, New Mexico, USA. Association for Computational Linguistics. +Zihao He, Negar Mokhberian, and Kristina Lerman. 2022. Infusing knowledge from Wikipedia to enhance stance detection. In Proceedings of the 12th Workshop on Computational Approaches to Subjectivity, Sentiment & Social Media Analysis, pages 71-77, Dublin, Ireland. Association for Computational Linguistics. +Myungha Jang and James Allan. 2018. Explaining controversy on social media via stance summarization. In The 41st International ACM SIGIR Conference on Research & Development in Information Retrieval, SIGIR '18, page 1221-1224, New York, NY, USA. Association for Computing Machinery. +Yan Jiang, Jinhua Gao, Huawei Shen, and Xueqi Cheng. 2022. Few-shot stance detection via target-aware prompt distillation. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '22, page 837-847, New York, NY, USA. Association for Computing Machinery. +Mike Lewis and Angela Fan. 2019. Generative question answering: Learning to answer the whole question. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net. +Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871-7880, Online. Association for Computational Linguistics. +Sha Li, Heng Ji, and Jiawei Han. 2021. Document-level event argument extraction by conditional generation. + +In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 894-908, Online. Association for Computational Linguistics. +Bin Liang, Zixiao Chen, Lin Gui, Yulan He, Min Yang, and Ruifeng Xu. 2022a. Zero-shot stance detection via contrastive learning. In Proceedings of the ACM Web Conference 2022, WWW '22, page 2738-2747, New York, NY, USA. Association for Computing Machinery. +Bin Liang, Yonghao Fu, Lin Gui, Min Yang, Jiachen Du, Yulan He, and Ruifeng Xu. 2021. Target-adaptive graph for cross-target stance detection. In Proceedings of the Web Conference 2021, WWW '21, page 3453-3464, New York, NY, USA. Association for Computing Machinery. +Bin Liang, Qinglin Zhu, Xiang Li, Min Yang, Lin Gui, Yulan He, and Ruifeng Xu. 2022b. JointCL: A joint contrastive learning framework for zero-shot stance detection. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 81-91, Dublin, Ireland. Association for Computational Linguistics. +Yuxiao Lin, Yuxian Meng, Xiaofei Sun, Qinghong Han, Kun Kuang, Jiwei Li, and Fei Wu. 2021. BertGCN: Transductive text classification by combining GNN and BERT. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 1456-1462, Online. Association for Computational Linguistics. +Rui Liu, Zheng Lin, Yutong Tan, and Weiping Wang. 2021. Enhancing zero-shot and few-shot stance detection with commonsense knowledge graph. In *Findings of the Association for Computational Linguistics: ACL-IJCNLP* 2021, pages 3152-3157, Online. Association for Computational Linguistics. +Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In International Conference on Learning Representations. +Saif Mohammad, Svetlana Kiritchenko, Parinaz Sobhani, Xiaodan Zhu, and Colin Cherry. 2016. SemEval-2016 task 6: Detecting stance in tweets. In Proceedings of the 10th International Workshop on Semantic Evaluation (SemEval-2016), pages 31-41, San Diego, California. Association for Computational Linguistics. +Mitra Mohtarami, Ramy Baly, James Glass, Preslav Nakov, Lluis Márquez, and Alessandro Moschitti. 2018. Automatic stance detection using end-to-end memory networks. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 767-776, New Orleans, Louisiana. Association for Computational Linguistics. + +Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2022. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(1). +Parinaz Sobhani, Diana Inkpen, and Xiaodan Zhu. 2017. A dataset for multi-target stance detection. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers, pages 551-557, Valencia, Spain. Association for Computational Linguistics. +Swapna Somasundaran and Janyce Wiebe. 2010. Recognizing stances in ideological on-line debates. In Proceedings of the NAACL HLT 2010 Workshop on Computational Approaches to Analysis and Generation of Emotion in Text, pages 116-124, Los Angeles, CA. Association for Computational Linguistics. +Laurens van der Maaten and Geoffrey Hinton. 2008. Visualizing data using t-sne. Journal of Machine Learning Research, 9(86):2579-2605. +Sean Welleck, Ilia Kulikov, Stephen Roller, Emily Dinan, Kyunghyun Cho, and Jason Weston. 2020. Neural text generation with unlikelihood training. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net. +Chang Xu, Cécile Paris, Surya Nepal, and Ross Sparks. 2018. Cross-target stance classification with self-attention networks. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 778-783, Melbourne, Australia. Association for Computational Linguistics. +Hang Yan, Junqi Dai, Tuo Ji, Xipeng Qiu, and Zheng Zhang. 2021. A unified generative framework for aspect-based sentiment analysis. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 2416-2429, Online. Association for Computational Linguistics. +Rong Zhang, Qifei Zhou, Bo An, Weiping Li, Tong Mo, and Bo Wu. 2020. Enhancing neural models with vulnerability via adversarial attack. In Proceedings of the 28th International Conference on Computational Linguistics, pages 1133-1146, Barcelona, Spain (Online). International Committee on Computational Linguistics. +Shaodian Zhang, Lin Qiu, Frank Chen, Weinan Zhang, Yong Yu, and Noémie Elhadad. 2017. We make choices we think are going to save us: Debate and stance identification for online breast cancer cam discussions. In Proceedings of the 26th International Conference on World Wide Web Companion, WWW '17 Companion, page 1073-1081, Republic and Canton of Geneva, CHE. International World Wide Web Conferences Steering Committee. + +A For every submission: + +A1. Did you describe the limitations of your work? Limitations +A2. Did you discuss any potential risks of your work? Limitations +A3. Do the abstract and introduction summarize the paper's main claims? Abstract, Introduction +A4. Have you used AI writing assistants when working on this paper? Left blank. + +B Did you use or create scientific artifacts? + +Introduction, Section 3.1 Data + +B1. Did you cite the creators of artifacts you used? Introduction, Section 3.1 Data +B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Section 3.1 Data +B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Section 3.1 Data, Section 3.2 Experimental Setup + +B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? We use an existing resource and detail of the data is discussed and introduced in their own published paper. +B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? We use an existing resource and detail of the data is discussed and introduced in their own published paper. +B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Section 3.1 Data + +C Did you run computational experiments? + +Left blank. + +C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Section 3.2 Experimental Setup + +C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? + +Section 3.2 Experimental Setup + +C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? + +Section 3.2 Experimental Setup, Table 1 + +C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? + +Section 3.2 Experimental Setup + +D Did you use human annotators (e.g., crowdworkers) or research with human participants? + +Left blank. + +D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? + +No response. + +D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? + +No response. + +D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? + +No response. + +D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? + +No response. + +D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? + +No response. \ No newline at end of file diff --git a/2023/Zero-Shot and Few-Shot Stance Detection on Varied Topics via Conditional Generation/images.zip b/2023/Zero-Shot and Few-Shot Stance Detection on Varied Topics via Conditional Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1f0047e0b070ae0a6d6b9b67bffbf468b19fc23b --- /dev/null +++ b/2023/Zero-Shot and Few-Shot Stance Detection on Varied Topics via Conditional Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:54dc4e613bbab77b20ffe2df29bb14947b874b01eac471a65526d2a431bb39cf +size 161955 diff --git a/2023/Zero-Shot and Few-Shot Stance Detection on Varied Topics via Conditional Generation/layout.json b/2023/Zero-Shot and Few-Shot Stance Detection on Varied Topics via Conditional Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ae851bc8b140b3eed4dac348fd91e9b461747c4e --- /dev/null +++ b/2023/Zero-Shot and Few-Shot Stance Detection on Varied Topics via Conditional Generation/layout.json @@ -0,0 +1,6723 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 69, + 75, + 526, + 110 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 75, + 526, + 110 + ], + "spans": [ + { + "bbox": [ + 69, + 75, + 526, + 110 + ], + "type": "text", + "content": "Zero-Shot and Few-Shot Stance Detection on Varied Topics via Conditional Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 179, + 131, + 416, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 131, + 416, + 145 + ], + "spans": [ + { + "bbox": [ + 179, + 131, + 416, + 145 + ], + "type": "text", + "content": "Haoyang Wen and Alexander G. Hauptmann" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 149, + 145, + 446, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 145, + 446, + 158 + ], + "spans": [ + { + "bbox": [ + 149, + 145, + 446, + 158 + ], + "type": "text", + "content": "Language Technologies Institute, Carnegie Mellon University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 226, + 159, + 370, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 226, + 159, + 370, + 172 + ], + "spans": [ + { + "bbox": [ + 226, + 159, + 370, + 172 + ], + "type": "text", + "content": "{hwen3, alex}@cs.cmu.edu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 84, + 236, + 274, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 236, + 274, + 488 + ], + "spans": [ + { + "bbox": [ + 84, + 236, + 274, + 488 + ], + "type": "text", + "content": "Zero-shot and few-shot stance detection identify the polarity of text with regard to a certain target when we have only limited or no training resources for the target. Previous work generally formulates the problem into a classification setting, ignoring the potential use of label text. In this paper, we instead utilize a conditional generation framework and formulate the problem as denoising from partially-filled templates, which can better utilize the semantics among input, label, and target texts. We further propose to jointly train an auxiliary task, target prediction, and to incorporate manually constructed incorrect samples with unlikelihood training to improve the representations for both target and label texts. We also verify the effectiveness of target-related Wikipedia knowledge with the generation framework. Experiments show that our proposed method significantly outperforms several strong baselines on VAST, and achieves new state-of-the-art performance." + }, + { + "bbox": [ + 84, + 236, + 274, + 488 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 498, + 155, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 498, + 155, + 511 + ], + "spans": [ + { + "bbox": [ + 68, + 498, + 155, + 511 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 520, + 291, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 520, + 291, + 655 + ], + "spans": [ + { + "bbox": [ + 67, + 520, + 291, + 655 + ], + "type": "text", + "content": "Stance detection is an important task that identifies the polarity of text with regard to certain target (Somasundaran and Wiebe, 2010; Augenstein et al., 2016; Mohammad et al., 2016; Sobhani et al., 2017; Allaway and McKeown, 2020), as shown in Table 1. It is crucial for understanding opinionated information expressed in natural language, and it can facilitate downstream social science analyses and applications (Zhang et al., 2017; Hanselowski et al., 2018; Jang and Allan, 2018)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 655, + 291, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 655, + 291, + 723 + ], + "spans": [ + { + "bbox": [ + 67, + 655, + 291, + 723 + ], + "type": "text", + "content": "Previous work on stance detection mostly focuses on in-domain or leave-out targets with only a few target choices (Mohtarami et al., 2018; Xu et al., 2018; Graells-Garrido et al., 2020; Zhang et al., 2020; Liang et al., 2021; Allaway et al., 2021;" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 308, + 211, + 524, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 211, + 524, + 260 + ], + "spans": [ + { + "bbox": [ + 308, + 211, + 524, + 260 + ], + "type": "text", + "content": "Input Text: Airports and the roads on east nor west coast can not handle the present volume adequately as is. I did ride the vast trains in Europe, Japan and China and found them very comfortable and providing much better connections and more efficient." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 309, + 261, + 522, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 261, + 522, + 272 + ], + "spans": [ + { + "bbox": [ + 309, + 261, + 522, + 272 + ], + "type": "text", + "content": "Target: high-speed rail Stance Label: Supportive (Pro)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 281, + 513, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 281, + 513, + 294 + ], + "spans": [ + { + "bbox": [ + 314, + 281, + 513, + 294 + ], + "type": "text", + "content": "Table 1: A stance detection example from VAST." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 322, + 526, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 322, + 526, + 526 + ], + "spans": [ + { + "bbox": [ + 302, + 322, + 526, + 526 + ], + "type": "text", + "content": "Jiang et al., 2022). Although achieving promising performance, those models are limited to generalize to a wide variety of targets. Zero-shot and few-shot stance detection on varied topics (VAST; Allaway and McKeown, 2020), instead, provides a diverse set of targets for training and testing. Efforts on this direction include involving graph modeling (Lin et al., 2021), common sense (Liu et al., 2021) or Wikipedia knowledge (He et al., 2022), and contrastive learning (Liang et al., 2022a,b). These methods generally formulate the problem into a classification setting, which directly trains the label representation from scratch, and does not fully utilize the semantics from those label and target texts." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 529, + 526, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 529, + 526, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 529, + 526, + 773 + ], + "type": "text", + "content": "However, connections among text semantics from input text, target, and label can be beneficial for stance detection. In this paper, we propose a new model by formulating the problem as a denoising task from text templates via conditional generation. Compared to direct classification, we can further exploit the label and topic semantics via learning to decode a series of natural language text containing the predicted label. The denoising scheme can also take advantage of the pretrained language model with similar pretraining task formulation (Lewis et al., 2020). To improve the target representation, we propose to jointly train target prediction with stance detection, which gives the input text and desired stance label to output possible targets. We use unlikelihood training (Welleck et al., 2020) that suppress the likelihood of manually constructed incorrect samples to enhance label" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 730, + 290, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 730, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 66, + 730, + 290, + 772 + ], + "type": "text", + "content": "1The resource for reproducing this paper is available at https://github.com/wenhycs/ACL2023-Zero-Shot-and-Few-Shot-Stance-Detection-on-Varied-Topics via -Conditional-Generation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "1491" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 135, + 795, + 459, + 805 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 795, + 459, + 805 + ], + "spans": [ + { + "bbox": [ + 135, + 795, + 459, + 805 + ], + "type": "text", + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 219, + 806, + 375, + 817 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 806, + 375, + 817 + ], + "spans": [ + { + "bbox": [ + 219, + 806, + 375, + 817 + ], + "type": "text", + "content": "Volume 2: Short Papers, pages 1491-1499" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "spans": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "text", + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 74, + 286, + 200 + ], + "blocks": [ + { + "bbox": [ + 72, + 74, + 286, + 200 + ], + "lines": [ + { + "bbox": [ + 72, + 74, + 286, + 200 + ], + "spans": [ + { + "bbox": [ + 72, + 74, + 286, + 200 + ], + "type": "image", + "image_path": "4bf32ce8c1161fd2fc799794b9e3e48e68aa1363f1db6efca4b907ce41afa9d4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 211, + 290, + 236 + ], + "lines": [ + { + "bbox": [ + 67, + 211, + 290, + 236 + ], + "spans": [ + { + "bbox": [ + 67, + 211, + 290, + 236 + ], + "type": "text", + "content": "Figure 1: Overall framework of BART-based generation framework for stance detection." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 257, + 290, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 257, + 290, + 323 + ], + "spans": [ + { + "bbox": [ + 67, + 257, + 290, + 323 + ], + "type": "text", + "content": "representations. Recently, He et al. (2022) show the effectiveness of target-related Wikipedia knowledge for classification-based stance detection. We also follow the idea and incorporate target-related Wikipedia knowledge for our generation model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 324, + 290, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 324, + 290, + 472 + ], + "spans": [ + { + "bbox": [ + 67, + 324, + 290, + 472 + ], + "type": "text", + "content": "We evaluate our method on VAST. Experimental results show that the conditional generation formulation can achieve better performance compared to classification, demonstrating the effectiveness of connecting input, target, and label semantics for stance detection. Further analysis illustrates the benefits of joint target prediction, unlikelihood training, and Wikipedia knowledge. Our model can achieve new state-of-the-art performance, outperforming several strong baselines from previous work." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 484, + 140, + 497 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 484, + 140, + 497 + ], + "spans": [ + { + "bbox": [ + 68, + 484, + 140, + 497 + ], + "type": "text", + "content": "2 Approach" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 506, + 290, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 506, + 290, + 560 + ], + "spans": [ + { + "bbox": [ + 67, + 506, + 290, + 560 + ], + "type": "text", + "content": "In this section, we will discuss our approach to zero-shot and few-shot stance detection. We will first introduce the problem formulation, and then discuss our generation-based framework." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 570, + 197, + 581 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 570, + 197, + 581 + ], + "spans": [ + { + "bbox": [ + 68, + 570, + 197, + 581 + ], + "type": "text", + "content": "2.1 Problem Formulation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 587, + 290, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 587, + 290, + 681 + ], + "spans": [ + { + "bbox": [ + 67, + 587, + 290, + 681 + ], + "type": "text", + "content": "Stance detection aims to identify the polarity of an input text with regard to a specific target. Formally, a sample instance can be considered as a triple " + }, + { + "bbox": [ + 67, + 587, + 290, + 681 + ], + "type": "inline_equation", + "content": "(x,t,y)" + }, + { + "bbox": [ + 67, + 587, + 290, + 681 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 67, + 587, + 290, + 681 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 67, + 587, + 290, + 681 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 587, + 290, + 681 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 67, + 587, + 290, + 681 + ], + "type": "text", + "content": " are two sequences of tokens, representing input text and target respectively. " + }, + { + "bbox": [ + 67, + 587, + 290, + 681 + ], + "type": "inline_equation", + "content": "y\\in \\{\\mathrm{supportive}(\\mathrm{pro}),\\mathrm{opposite}(\\mathrm{con}),\\mathrm{neutral}\\}" + }, + { + "bbox": [ + 67, + 587, + 290, + 681 + ], + "type": "text", + "content": " represents then stance label." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 682, + 290, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 682, + 290, + 709 + ], + "spans": [ + { + "bbox": [ + 67, + 682, + 290, + 709 + ], + "type": "text", + "content": "A stance-detection model is to infer the stance label " + }, + { + "bbox": [ + 67, + 682, + 290, + 709 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 67, + 682, + 290, + 709 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 67, + 682, + 290, + 709 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 67, + 682, + 290, + 709 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 682, + 290, + 709 + ], + "type": "inline_equation", + "content": "\\pmb{t}" + }, + { + "bbox": [ + 67, + 682, + 290, + 709 + ], + "type": "text", + "content": " with parameter " + }, + { + "bbox": [ + 67, + 682, + 290, + 709 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 67, + 682, + 290, + 709 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 721, + 214, + 735 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 721, + 214, + 735 + ], + "spans": [ + { + "bbox": [ + 143, + 721, + 214, + 735 + ], + "type": "interline_equation", + "content": "f \\left(\\boldsymbol {x}, \\boldsymbol {t}; \\theta\\right) = y.", + "image_path": "d697d472a26c5cbcf1aa08cefedda4b688a2919a67c02ea4d9ea6da28a4039a1.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "type": "text", + "content": "In the zero-shot and few-shot stance detection dataset with varied targets (Allaway and McKe" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 71, + 525, + 98 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 525, + 98 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 525, + 98 + ], + "type": "text", + "content": "own, 2020), many target tokens only occur zero or a few times in the training set." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 106, + 483, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 106, + 483, + 118 + ], + "spans": [ + { + "bbox": [ + 302, + 106, + 483, + 118 + ], + "type": "text", + "content": "2.2 A Generation-Based Framework" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 124, + 525, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 124, + 525, + 272 + ], + "spans": [ + { + "bbox": [ + 302, + 124, + 525, + 272 + ], + "type": "text", + "content": "Generation-based frameworks have demonstrated their effectiveness for problems beyond traditional generation tasks (Lewis and Fan, 2019; Yan et al., 2021; Li et al., 2021; Raffel et al., 2022). We use a conditional generation model for this problem, where the condition is a partially-filled template with the input text. The template is two sentences describing the target and stance with a placeholder for stance detection. An example of the partially-filled template with input text and output is shown in Figure 1." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 273, + 525, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 273, + 525, + 367 + ], + "spans": [ + { + "bbox": [ + 302, + 273, + 525, + 367 + ], + "type": "text", + "content": "Our base model is BART (Lewis et al., 2020), an encoder-decoder language model pretrained with denoising objectives, which is similar to our generation-based formulation. The generation process can be considered as using the conditional probability to select a new token at each step given input and previously generated tokens:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 373, + 520, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 373, + 520, + 410 + ], + "spans": [ + { + "bbox": [ + 306, + 373, + 520, + 410 + ], + "type": "interline_equation", + "content": "p \\left(\\boldsymbol {o} \\mid g \\left(\\boldsymbol {x}, \\boldsymbol {t}\\right); \\theta\\right) = \\prod_ {i = 1} ^ {| \\boldsymbol {o} |} p \\left(o _ {i} \\mid \\boldsymbol {o} _ {< i}, g \\left(\\boldsymbol {x}, \\boldsymbol {t}\\right); \\theta\\right),", + "image_path": "5c1113d68ff894ab83d6775c782ba50aa5cd4d9f7bc42f72610a47197f99982a.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 415, + 525, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 415, + 525, + 550 + ], + "spans": [ + { + "bbox": [ + 302, + 415, + 525, + 550 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 302, + 415, + 525, + 550 + ], + "type": "inline_equation", + "content": "g(\\pmb {x},\\pmb {t})" + }, + { + "bbox": [ + 302, + 415, + 525, + 550 + ], + "type": "text", + "content": " represents the transformation function that fills the target " + }, + { + "bbox": [ + 302, + 415, + 525, + 550 + ], + "type": "inline_equation", + "content": "\\pmb{t}" + }, + { + "bbox": [ + 302, + 415, + 525, + 550 + ], + "type": "text", + "content": " into the template and forms the input sequence with the input text " + }, + { + "bbox": [ + 302, + 415, + 525, + 550 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 302, + 415, + 525, + 550 + ], + "type": "text", + "content": ". Specifically, " + }, + { + "bbox": [ + 302, + 415, + 525, + 550 + ], + "type": "inline_equation", + "content": "g(\\pmb {x},\\pmb {t})" + }, + { + "bbox": [ + 302, + 415, + 525, + 550 + ], + "type": "text", + "content": " will generate a combination of input text and template with special tokens: “ template x ”. The template contains two sentences: “The target is . The stance is ”. We will fill in placeholder with the actual target and keep the placeholder for the decoder to generate." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 302, + 550, + 525, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 550, + 525, + 618 + ], + "spans": [ + { + "bbox": [ + 302, + 550, + 525, + 618 + ], + "type": "text", + "content": "The generated output " + }, + { + "bbox": [ + 302, + 550, + 525, + 618 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 302, + 550, + 525, + 618 + ], + "type": "text", + "content": " is a fully-filled template, where both target and stance placeholders are replaced by actual or predicted values. The model is trained by minimizing the log-likelihood over the whole generated sequence:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 324, + 624, + 501, + 678 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 624, + 501, + 678 + ], + "spans": [ + { + "bbox": [ + 324, + 624, + 501, + 678 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {s} = - \\log p (\\boldsymbol {o} \\mid g (\\boldsymbol {x}, t); \\theta) \\\\ = - \\sum_ {i = 1} ^ {| O |} \\log p \\left(o _ {i} \\mid o _ {< i}, g (\\boldsymbol {x}, t); \\theta\\right). \\\\ \\end{array}", + "image_path": "ffbf6550fee224521ae0296690b8ab2338f9355216d7d1db8a6d626092ad09ae.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 302, + 683, + 525, + 724 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 683, + 525, + 724 + ], + "spans": [ + { + "bbox": [ + 302, + 683, + 525, + 724 + ], + "type": "text", + "content": "The final predicted stance label is obtained with a post-processing function that tries to find the polarity word after the prompt for stance." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 302, + 730, + 448, + 743 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 730, + 448, + 743 + ], + "spans": [ + { + "bbox": [ + 302, + 730, + 448, + 743 + ], + "type": "text", + "content": "2.2.1 Joint Target Prediction" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "type": "text", + "content": "Another advantage of using generation-based architecture is that we can leverage auxiliary generative" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "text", + "content": "1492" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 68, + 68, + 291, + 176 + ], + "blocks": [ + { + "bbox": [ + 68, + 68, + 291, + 176 + ], + "lines": [ + { + "bbox": [ + 68, + 68, + 291, + 176 + ], + "spans": [ + { + "bbox": [ + 68, + 68, + 291, + 176 + ], + "type": "table", + "html": "
Stance Detection
InputTarget is high-speed rail. Stance is<stance>.
OutputTarget is high-speed rail. Stance issupportive.
Target Prediction
InputStance is supportive. Target is<target>.
OutputStance is supportive. Target ishigh-speed rail.
Unlikelihood Training
InputTarget is high-speed rail. Stance is<stance>.
OutputTarget is high-speed rail. Stance isopposite.
", + "image_path": "1ec07c3827b57e28d25f80c266b64154a72d7e929923d3717aac39e2c547f619.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 183, + 290, + 208 + ], + "lines": [ + { + "bbox": [ + 67, + 183, + 290, + 208 + ], + "spans": [ + { + "bbox": [ + 67, + 183, + 290, + 208 + ], + "type": "text", + "content": "Table 2: Examples input and output templates for stance detection, target prediction, and unlikelihood training." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 227, + 291, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 227, + 291, + 267 + ], + "spans": [ + { + "bbox": [ + 67, + 227, + 291, + 267 + ], + "type": "text", + "content": "tasks to help train stance detection. We use target prediction, which is to infer the target tokens " + }, + { + "bbox": [ + 67, + 227, + 291, + 267 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 67, + 227, + 291, + 267 + ], + "type": "text", + "content": " given stance label " + }, + { + "bbox": [ + 67, + 227, + 291, + 267 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 67, + 227, + 291, + 267 + ], + "type": "text", + "content": " and input text " + }, + { + "bbox": [ + 67, + 227, + 291, + 267 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 67, + 227, + 291, + 267 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 142, + 275, + 215, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 275, + 215, + 289 + ], + "spans": [ + { + "bbox": [ + 142, + 275, + 215, + 289 + ], + "type": "interline_equation", + "content": "f _ {t} (\\boldsymbol {x}, y; \\theta) = \\boldsymbol {t}.", + "image_path": "092dde9df2bbb8e317680a69b67fde24aef4cd0363f4b7221f55f78aaa6c5ebf.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 297, + 290, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 297, + 290, + 350 + ], + "spans": [ + { + "bbox": [ + 67, + 297, + 290, + 350 + ], + "type": "text", + "content": "Target prediction can provide the connection of stance to target in an opposite direction of stance detection. It can also enhance the representation of target tokens by learning to decode them." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 352, + 290, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 352, + 290, + 473 + ], + "spans": [ + { + "bbox": [ + 67, + 352, + 290, + 473 + ], + "type": "text", + "content": "The input sequence of target prediction is similar to stance detection, consisting of a partially-filled template and input text. The template used for joint target prediction is slightly different than the one used for stance detection, where we switch the position of two sentences so that the stance information shows up first. We will fill in the actual stance text in the input sequence, and leave the placeholder for the decoder to generate." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 480, + 208, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 480, + 208, + 493 + ], + "spans": [ + { + "bbox": [ + 67, + 480, + 208, + 493 + ], + "type": "text", + "content": "2.2.2 Unlikelihood Training" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 496, + 290, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 496, + 290, + 617 + ], + "spans": [ + { + "bbox": [ + 67, + 496, + 290, + 617 + ], + "type": "text", + "content": "Log-likelihood objective optimizes the likelihood over the entire distribution. However, in our task, especially when generating the stance labels, we should specifically focus on several candidate tokens. Therefore, we introduce unlikelihood training (Welleck et al., 2020), where we use unlikely tokens, i.e. incorrect stance predictions, to replace the ground-truth sequence and optimize with the unlikelihood loss for the replaced tokens." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 618, + 291, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 618, + 291, + 685 + ], + "spans": [ + { + "bbox": [ + 67, + 618, + 291, + 685 + ], + "type": "text", + "content": "Specifically, for an output sequence " + }, + { + "bbox": [ + 67, + 618, + 291, + 685 + ], + "type": "inline_equation", + "content": "\\pmb{o}" + }, + { + "bbox": [ + 67, + 618, + 291, + 685 + ], + "type": "text", + "content": ", we assume " + }, + { + "bbox": [ + 67, + 618, + 291, + 685 + ], + "type": "inline_equation", + "content": "o_k" + }, + { + "bbox": [ + 67, + 618, + 291, + 685 + ], + "type": "text", + "content": " is the stance label and replaced it with an incorrect stance prediction " + }, + { + "bbox": [ + 67, + 618, + 291, + 685 + ], + "type": "inline_equation", + "content": "o_k'" + }, + { + "bbox": [ + 67, + 618, + 291, + 685 + ], + "type": "text", + "content": " while keeping other tokens to form incorrect sequence " + }, + { + "bbox": [ + 67, + 618, + 291, + 685 + ], + "type": "inline_equation", + "content": "o'" + }, + { + "bbox": [ + 67, + 618, + 291, + 685 + ], + "type": "text", + "content": ". The combination of likelihood and unlikelihood will be:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 85, + 693, + 273, + 739 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 693, + 273, + 739 + ], + "spans": [ + { + "bbox": [ + 85, + 693, + 273, + 739 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {u} = \\log p \\left(o _ {k} ^ {\\prime} \\mid \\boldsymbol {o} _ {< k} ^ {\\prime}, g (\\boldsymbol {x}, \\boldsymbol {t}); \\theta\\right) \\\\ - \\sum_ {i \\neq k} \\log p \\left(o _ {i} ^ {\\prime} \\mid o _ {< i} ^ {\\prime}, g (\\boldsymbol {x}, \\boldsymbol {t}); \\theta\\right), \\\\ \\end{array}", + "image_path": "b86e3924a0ecf19ebf70f7059648ace0e8bca17fab26e156dab507d5d1528a1a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "type": "text", + "content": "For each ground-truth sequence, we can construct two sequences for unlikelihood training with the" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 71, + 526, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 126 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 126 + ], + "type": "text", + "content": "other two incorrect stance labels. Table 2 illustrates the examples for different input and output templates for stance prediction, target prediction, and unlikelihood training." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 132, + 511, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 132, + 511, + 146 + ], + "spans": [ + { + "bbox": [ + 302, + 132, + 511, + 146 + ], + "type": "text", + "content": "2.2.3 Incorporating Wikipedia Knowledge" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 148, + 526, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 148, + 526, + 324 + ], + "spans": [ + { + "bbox": [ + 302, + 148, + 526, + 324 + ], + "type": "text", + "content": "He et al. (2022) collect relevant Wikipedia snippets for each target and propose to incorporate Wikipedia knowledge to enhance target representations for BERT-based (Devlin et al., 2019) classification, which demonstrates a significant improvement. We follow He et al. (2022) and incorporate Wikipedia knowledge into our generation-based method. Specifically, we append Wikipedia snippets to the end of our input sequence: “" + }, + { + "bbox": [ + 302, + 148, + 526, + 324 + ], + "type": "inline_equation", + "content": "" + }, + { + "bbox": [ + 302, + 148, + 526, + 324 + ], + "type": "text", + "content": " template " + }, + { + "bbox": [ + 302, + 148, + 526, + 324 + ], + "type": "inline_equation", + "content": "\\langle /s\\rangle < / s\\rangle x" + }, + { + "bbox": [ + 302, + 148, + 526, + 324 + ], + "type": "inline_equation", + "content": "\\langle /s\\rangle < / s\\rangle" + }, + { + "bbox": [ + 302, + 148, + 526, + 324 + ], + "type": "text", + "content": " Wikipedia snippet " + }, + { + "bbox": [ + 302, + 148, + 526, + 324 + ], + "type": "inline_equation", + "content": "\\langle /s\\rangle" + }, + { + "bbox": [ + 302, + 148, + 526, + 324 + ], + "type": "text", + "content": "”. We use the new input sequence to perform both training and inference while the output sequences remain as the fully-filled templates." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 332, + 427, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 332, + 427, + 344 + ], + "spans": [ + { + "bbox": [ + 302, + 332, + 427, + 344 + ], + "type": "text", + "content": "2.2.4 Training Objective" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 347, + 525, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 347, + 525, + 387 + ], + "spans": [ + { + "bbox": [ + 302, + 347, + 525, + 387 + ], + "type": "text", + "content": "The final training objective is the combination of loss functions from stance detection, target prediction, and unlikelihood training:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 357, + 397, + 471, + 411 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 397, + 471, + 411 + ], + "spans": [ + { + "bbox": [ + 357, + 397, + 471, + 411 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {s} + \\alpha_ {t} \\mathcal {L} _ {t} + \\alpha_ {u} \\mathcal {L} _ {u},", + "image_path": "c502c6ef73dd9610b296e04bab0de9908f695df91e783d4d6f5cf96e1a88045c.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 302, + 420, + 525, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 420, + 525, + 460 + ], + "spans": [ + { + "bbox": [ + 302, + 420, + 525, + 460 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 302, + 420, + 525, + 460 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_t" + }, + { + "bbox": [ + 302, + 420, + 525, + 460 + ], + "type": "text", + "content": " represents the log-likelihood loss over the output template for target prediction, " + }, + { + "bbox": [ + 302, + 420, + 525, + 460 + ], + "type": "inline_equation", + "content": "\\alpha_{t},\\alpha_{u}" + }, + { + "bbox": [ + 302, + 420, + 525, + 460 + ], + "type": "text", + "content": " are used to balance different loss functions." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 302, + 470, + 390, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 470, + 390, + 483 + ], + "spans": [ + { + "bbox": [ + 302, + 470, + 390, + 483 + ], + "type": "text", + "content": "3 Experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 303, + 491, + 354, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 491, + 354, + 503 + ], + "spans": [ + { + "bbox": [ + 303, + 491, + 354, + 503 + ], + "type": "text", + "content": "3.1 Data" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 302, + 507, + 526, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 507, + 526, + 670 + ], + "spans": [ + { + "bbox": [ + 302, + 507, + 526, + 670 + ], + "type": "text", + "content": "VAST contains 18,548 examples from New York Times \"Room for Debate\" section with 5,630 different targets for zero-shot and few-shot stance detection. The original examples of VAST are collected from Habernal et al. (2018) under Apache-2.0 license2. We use Wikipedia knowledge collected by He et al. (2022), which uses API to crawl Wikipedia pages for targets. Wikipedia content can be used under Creative Commons Attribution Share-Alike license (CC-BY-SA)3. We use the same training/development/test split as Allaway and McKeown (2020)." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 302, + 679, + 425, + 692 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 679, + 425, + 692 + ], + "spans": [ + { + "bbox": [ + 302, + 679, + 425, + 692 + ], + "type": "text", + "content": "3.2 Experimental Setup" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 302, + 697, + 525, + 724 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 697, + 525, + 724 + ], + "spans": [ + { + "bbox": [ + 302, + 697, + 525, + 724 + ], + "type": "text", + "content": "We conduct our experiments on VAST (Allaway and McKeown, 2020). We compare our model" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 729, + 525, + 750 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 729, + 525, + 750 + ], + "spans": [ + { + "bbox": [ + 302, + 729, + 525, + 750 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 302, + 729, + 525, + 750 + ], + "type": "text", + "content": "https://github.com/UKPLab/argument-reasoning-comprehension-task/blob/master/License" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 302, + 751, + 524, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 751, + 524, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 524, + 772 + ], + "type": "text", + "content": "3https://en.wikipedia.org/wiki/Wikipedia:Reusing_Wikipedia_content" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1493" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 79, + 68, + 280, + 143 + ], + "blocks": [ + { + "bbox": [ + 79, + 68, + 280, + 143 + ], + "lines": [ + { + "bbox": [ + 79, + 68, + 280, + 143 + ], + "spans": [ + { + "bbox": [ + 79, + 68, + 280, + 143 + ], + "type": "table", + "html": "
ModelPrecisionRecallF1
BERT Classification72.672.072.1
BART w/ Template75.775.175.3
+ Topic Prediction76.075.675.7
+Unlikelihood76.475.975.9
+Wikipedia78.077.377.4
", + "image_path": "1eb1129e945c1b9414ec5383703471570bb731e8ce94291ec684ab64bbacf7a5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 80, + 223, + 278, + 299 + ], + "blocks": [ + { + "bbox": [ + 67, + 152, + 290, + 200 + ], + "lines": [ + { + "bbox": [ + 67, + 152, + 290, + 200 + ], + "spans": [ + { + "bbox": [ + 67, + 152, + 290, + 200 + ], + "type": "text", + "content": "Table 3: Performance of different model variants on the overall precision, recall and " + }, + { + "bbox": [ + 67, + 152, + 290, + 200 + ], + "type": "inline_equation", + "content": "\\mathrm{F}_1" + }, + { + "bbox": [ + 67, + 152, + 290, + 200 + ], + "type": "text", + "content": " on the development set (\\%). Each of our model variants is on top of the variant from its previous row." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 80, + 223, + 278, + 299 + ], + "lines": [ + { + "bbox": [ + 80, + 223, + 278, + 299 + ], + "spans": [ + { + "bbox": [ + 80, + 223, + 278, + 299 + ], + "type": "table", + "html": "
ModelZero-ShotFew-ShotOverall
TGA-Net66.666.366.5
BERT-GCN68.669.769.2
CKE-Net70.270.170.1
WS-BERT75.373.674.5
Our Model76.478.077.3
", + "image_path": "eaa78d12b608f420824ef696f6cb77dcd38a4f963653b44fe45601b2aebe71d8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 306, + 291, + 380 + ], + "lines": [ + { + "bbox": [ + 67, + 306, + 291, + 380 + ], + "spans": [ + { + "bbox": [ + 67, + 306, + 291, + 380 + ], + "type": "text", + "content": "Table 4: Stance detection performance " + }, + { + "bbox": [ + 67, + 306, + 291, + 380 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 67, + 306, + 291, + 380 + ], + "type": "text", + "content": " on VAST. Our model significantly outperforms previous work on all metrics. Our results are obtained from averaging performances over 5 random seeds. " + }, + { + "bbox": [ + 67, + 306, + 291, + 380 + ], + "type": "inline_equation", + "content": "p < 0.001" + }, + { + "bbox": [ + 67, + 306, + 291, + 380 + ], + "type": "text", + "content": " on overall " + }, + { + "bbox": [ + 67, + 306, + 291, + 380 + ], + "type": "inline_equation", + "content": "\\mathrm{F_1}" + }, + { + "bbox": [ + 67, + 306, + 291, + 380 + ], + "type": "text", + "content": " using Z-test with variance as the standard deviation over multiple runs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 413, + 291, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 413, + 291, + 520 + ], + "spans": [ + { + "bbox": [ + 67, + 413, + 291, + 520 + ], + "type": "text", + "content": "with several existing systems including 1) TGA-Net (Allaway and McKeown, 2020); 2) BERTGCN (Lin et al., 2021); 3) CKE-Net (Liu et al., 2021); 4) WS-BERT (He et al., 2022). Following their setup, we use macro-average " + }, + { + "bbox": [ + 67, + 413, + 291, + 520 + ], + "type": "inline_equation", + "content": "\\mathrm{F}_1" + }, + { + "bbox": [ + 67, + 413, + 291, + 520 + ], + "type": "text", + "content": " as the evaluation metric, and we report performance on the subset of test set for zero-shot and few-shot, and the overall test set." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 528, + 291, + 718 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 528, + 291, + 718 + ], + "spans": [ + { + "bbox": [ + 67, + 528, + 291, + 718 + ], + "type": "text", + "content": "We use BART-base" + }, + { + "bbox": [ + 67, + 528, + 291, + 718 + ], + "type": "inline_equation", + "content": "^4" + }, + { + "bbox": [ + 67, + 528, + 291, + 718 + ], + "type": "text", + "content": " as our base model, of which the number of parameters is roughly consistent with baselines on BERT-base" + }, + { + "bbox": [ + 67, + 528, + 291, + 718 + ], + "type": "inline_equation", + "content": "^5" + }, + { + "bbox": [ + 67, + 528, + 291, + 718 + ], + "type": "text", + "content": ". Our best model is optimized with AdamW (Loshchilov and Hutter, 2019) for 30 epochs with a learning rate of 1e-5. We use a linear scheduler with a warmup proportion of 0.1 and the training batch size is 32. We use greedy search during inference. We reported performances on development set and test set using the averaged results from 5 different random seeds. Test results are reported based on the best overall " + }, + { + "bbox": [ + 67, + 528, + 291, + 718 + ], + "type": "inline_equation", + "content": "\\mathrm{F_1}" + }, + { + "bbox": [ + 67, + 528, + 291, + 718 + ], + "type": "text", + "content": " performance on the development set. " + }, + { + "bbox": [ + 67, + 528, + 291, + 718 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 67, + 528, + 291, + 718 + ], + "type": "text", + "content": " is set to 1 and " + }, + { + "bbox": [ + 67, + 528, + 291, + 718 + ], + "type": "inline_equation", + "content": "\\alpha_{u}" + }, + { + "bbox": [ + 67, + 528, + 291, + 718 + ], + "type": "text", + "content": " is set to 0.5. Our final model takes about 5 hours for training on one Nvidia RTX 3090 GPU." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 309, + 71, + 413, + 174 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 413, + 174 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 413, + 174 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 413, + 174 + ], + "type": "image", + "image_path": "f9cd9180d99fc33a12ecf4039e1b8b13e7d4e3208478af6de17156cdc97d73d4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 179, + 387, + 190 + ], + "lines": [ + { + "bbox": [ + 334, + 179, + 387, + 190 + ], + "spans": [ + { + "bbox": [ + 334, + 179, + 387, + 190 + ], + "type": "text", + "content": "(a) Our model" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 302, + 200, + 525, + 236 + ], + "lines": [ + { + "bbox": [ + 302, + 200, + 525, + 236 + ], + "spans": [ + { + "bbox": [ + 302, + 200, + 525, + 236 + ], + "type": "text", + "content": "Figure 2: The t-SNE visualization of intermediate representations from our model and BERT classification model. Color map: Supportive, Opposite, Neutral." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 417, + 71, + 520, + 174 + ], + "blocks": [ + { + "bbox": [ + 417, + 71, + 520, + 174 + ], + "lines": [ + { + "bbox": [ + 417, + 71, + 520, + 174 + ], + "spans": [ + { + "bbox": [ + 417, + 71, + 520, + 174 + ], + "type": "image", + "image_path": "ee292fcfd6026bd798f7e0c63ebd63806084f7a1d24e58e3c1468a4fea63059f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 425, + 179, + 511, + 190 + ], + "lines": [ + { + "bbox": [ + 425, + 179, + 511, + 190 + ], + "spans": [ + { + "bbox": [ + 425, + 179, + 511, + 190 + ], + "type": "text", + "content": "(b) BERT classification" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 260, + 365, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 260, + 365, + 272 + ], + "spans": [ + { + "bbox": [ + 302, + 260, + 365, + 272 + ], + "type": "text", + "content": "3.3 Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 280, + 490, + 294 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 280, + 490, + 294 + ], + "spans": [ + { + "bbox": [ + 302, + 280, + 490, + 294 + ], + "type": "text", + "content": "3.3.1 Comparing with Model Variants" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 300, + 526, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 300, + 526, + 476 + ], + "spans": [ + { + "bbox": [ + 302, + 300, + 526, + 476 + ], + "type": "text", + "content": "We first conduct comparison of some of our model variants to illustrate the effectiveness of our proposed components. The results are shown in Table 3. From the comparison of BERT-based classification (BERT Classification) and BART-based denoising generation from templates (BART w/ Template), we can find that adopting the generation framework can significantly improve the model performance. Our proposed topic prediction and un-likelihood training can further boost performance. The final model with knowledge from Wikipedia, verifies the effectiveness of Wikipedia knowledge for stance detection with a generative framework." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 489, + 497, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 489, + 497, + 502 + ], + "spans": [ + { + "bbox": [ + 302, + 489, + 497, + 502 + ], + "type": "text", + "content": "3.3.2 Comparing with Existing Systems" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 507, + 525, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 507, + 525, + 576 + ], + "spans": [ + { + "bbox": [ + 302, + 507, + 525, + 576 + ], + "type": "text", + "content": "Our overall performance is shown in Table 4. Our method can significantly outperform those previous baselines, indicating the effectiveness of our proposed generation framework for zero-shot and few-shot stance detection with varies topics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 591, + 425, + 603 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 591, + 425, + 603 + ], + "spans": [ + { + "bbox": [ + 302, + 591, + 425, + 603 + ], + "type": "text", + "content": "3.4 Qualitative Analysis" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 302, + 611, + 525, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 611, + 525, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 611, + 525, + 773 + ], + "type": "text", + "content": "Figure 2 show the t-SNE (van der Maaten and Hinton, 2008) visualization of intermediate representations before the classification layer from our model and BERT classification model on the development set. We use random initialization with perplexity as 50 for visualization and we color each visualized instance with its corresponding stance label. The visualization of BERT classification shows small clusters with hybrid labels, While we can see that instances with our generation method are clustered with labels, where neutral labels are at the top and supportive labels are generally at the bottom." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 80, + 749, + 271, + 761 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 749, + 271, + 761 + ], + "spans": [ + { + "bbox": [ + 80, + 749, + 271, + 761 + ], + "type": "text", + "content": "4https://huggingface.co/facebook/bart-base" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 81, + 761, + 267, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 761, + 267, + 772 + ], + "spans": [ + { + "bbox": [ + 81, + 761, + 267, + 772 + ], + "type": "text", + "content": "5https://huggingface.co/bert-base-uncased" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1494" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 70, + 161, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 70, + 161, + 83 + ], + "spans": [ + { + "bbox": [ + 68, + 70, + 161, + 83 + ], + "type": "text", + "content": "4 Related Work" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 95, + 293, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 95, + 293, + 338 + ], + "spans": [ + { + "bbox": [ + 67, + 95, + 293, + 338 + ], + "type": "text", + "content": "Zero-shot and few-shot stance detection. Zero-shot and few-shot stance detection focus on detecting stances for unseen or low-resource targets. Allaway and McKeown (2020) construct a dataset with varied topics that can be used to test stance detection under zero-shot and few-shot settings. Previous efforts mostly focus on modeling targets, documents, or their connections. Allaway and McKeown (2020) obtain generalized topic representation through clustering. Liu et al. (2021) use commonsense knowledge graph to enhance the connection between target and document. Liang et al. (2022a,b) use contrastive learning to learn target features. He et al. (2022) incorporate Wikipedia knowledge to enhance target representations. While in our work, we use a conditional generation framework to build the connections between input, target, and label text semantics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 349, + 292, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 349, + 292, + 498 + ], + "spans": [ + { + "bbox": [ + 67, + 349, + 292, + 498 + ], + "type": "text", + "content": "Text processing via conditional generation. Our work is also motivated by the recent success of tackling text processing problems as conditional generation (Lewis et al., 2020; Raffel et al., 2022). In addition to the conventional text generation problems, conditional generation frameworks are effectively applied in information extraction (Li et al., 2021), question answering (Lewis and Fan, 2019; Raffel et al., 2022) and sentiment analysis (Yan et al., 2021). In our work, we further explore stance detection via conditional generation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 510, + 147, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 510, + 147, + 523 + ], + "spans": [ + { + "bbox": [ + 67, + 510, + 147, + 523 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 534, + 291, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 534, + 291, + 670 + ], + "spans": [ + { + "bbox": [ + 67, + 534, + 291, + 670 + ], + "type": "text", + "content": "In this paper, we propose a generation-based framework for zero-shot and few-shot stance detection that generate stance label from pre-defined templates. We further propose an auxiliary task, joint target prediction that takes stance label and input text to generate targets, and unlikelihood training on manually constructed incorrect generation output. Combining with Wikipedia knowledge for target from He et al. (2022), our model can achieve new state-of-the-art performance on VAST." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 682, + 131, + 695 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 682, + 131, + 695 + ], + "spans": [ + { + "bbox": [ + 67, + 682, + 131, + 695 + ], + "type": "text", + "content": "Limitations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 706, + 292, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 706, + 292, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 706, + 292, + 773 + ], + "type": "text", + "content": "Because of the nature of our framework design, our work requires a diverse set of targets during training, which is important for target prediction and therefore the stance detection method. It is difficult to be applied to other stance detection datasets" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 71, + 526, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 151 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 151 + ], + "type": "text", + "content": "when there are limited training resources with regard to targets, such as Conforti et al. (2020) and Mohammad et al. (2016). Besides, the model is trained on news-related debate corpus, so it may need further domain adaptation if applying the model to other domains such as social media." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 153, + 527, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 153, + 527, + 247 + ], + "spans": [ + { + "bbox": [ + 302, + 153, + 527, + 247 + ], + "type": "text", + "content": "We are using an auto-regressive generation framework, which will also require extra inference time to generate the whole output sequence compared to the classification model. We would encourage readers to compare it with classification methods for efficiency when it will be applied in a time-sensitive scenario." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 269, + 362, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 269, + 362, + 281 + ], + "spans": [ + { + "bbox": [ + 304, + 269, + 362, + 281 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 303, + 286, + 527, + 774 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 304, + 286, + 527, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 286, + 527, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 286, + 527, + 365 + ], + "type": "text", + "content": "Emily Allaway and Kathleen McKeown. 2020. Zero-Shot Stance Detection: A Dataset and Model using Generalized Topic Representations. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 8913-8931, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 370, + 527, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 370, + 527, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 370, + 527, + 449 + ], + "type": "text", + "content": "Emily Allaway, Malavika Srikanth, and Kathleen McKeown. 2021. Adversarial learning for zero-shot stance detection on social media. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4756-4767, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 454, + 527, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 454, + 527, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 454, + 527, + 533 + ], + "type": "text", + "content": "Isabelle Augenstein, Tim Rocktäschel, Andreas Vlachos, and Kalina Bontcheva. 2016. Stance detection with bidirectional conditional encoding. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 876-885, Austin, Texas. Association for Computational Linguistics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 303, + 538, + 527, + 626 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 538, + 527, + 626 + ], + "spans": [ + { + "bbox": [ + 303, + 538, + 527, + 626 + ], + "type": "text", + "content": "Costanza Conforti, Jakob Berndt, Mohammad Taher Pilehvar, Chryssi Giannitsarou, Flavio Toxvaerd, and Nigel Collier. 2020. Will-they-won't-they: A very large dataset for stance detection on Twitter. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1715-1724, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 303, + 633, + 527, + 734 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 633, + 527, + 734 + ], + "spans": [ + { + "bbox": [ + 303, + 633, + 527, + 734 + ], + "type": "text", + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 303, + 738, + 527, + 774 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 738, + 527, + 774 + ], + "spans": [ + { + "bbox": [ + 303, + 738, + 527, + 774 + ], + "type": "text", + "content": "Eduardo Graells-Garrido, Ricardo Baeza-Yates, and Mounia Lalmas. 2020. Representativeness of abortion legislation debate on twitter: A case study in" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1495" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 80, + 72, + 291, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 72, + 291, + 117 + ], + "spans": [ + { + "bbox": [ + 80, + 72, + 291, + 117 + ], + "type": "text", + "content": "argentina and chile. In Companion Proceedings of the Web Conference 2020, WWW '20, page 765-774, New York, NY, USA. Association for Computing Machinery." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 125, + 291, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 125, + 291, + 225 + ], + "spans": [ + { + "bbox": [ + 69, + 125, + 291, + 225 + ], + "type": "text", + "content": "Ivan Habernal, Henning Wachsmuth, Iryna Gurevych, and Benno Stein. 2018. The argument reasoning comprehension task: Identification and reconstruction of implicit warrants. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1930-1940, New Orleans, Louisiana. Association for Computational Linguistics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 233, + 291, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 233, + 291, + 322 + ], + "spans": [ + { + "bbox": [ + 69, + 233, + 291, + 322 + ], + "type": "text", + "content": "Andreas Hanselowski, Avinesh PVS, Benjamin Schiller, Felix Caspelherr, Debanjan Chaudhuri, Christian M. Meyer, and Iryna Gurevych. 2018. A retrospective analysis of the fake news challenge stance-detection task. In Proceedings of the 27th International Conference on Computational Linguistics, pages 1859-1874, Santa Fe, New Mexico, USA. Association for Computational Linguistics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 330, + 291, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 330, + 291, + 407 + ], + "spans": [ + { + "bbox": [ + 69, + 330, + 291, + 407 + ], + "type": "text", + "content": "Zihao He, Negar Mokhberian, and Kristina Lerman. 2022. Infusing knowledge from Wikipedia to enhance stance detection. In Proceedings of the 12th Workshop on Computational Approaches to Subjectivity, Sentiment & Social Media Analysis, pages 71-77, Dublin, Ireland. Association for Computational Linguistics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 416, + 291, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 416, + 291, + 483 + ], + "spans": [ + { + "bbox": [ + 69, + 416, + 291, + 483 + ], + "type": "text", + "content": "Myungha Jang and James Allan. 2018. Explaining controversy on social media via stance summarization. In The 41st International ACM SIGIR Conference on Research & Development in Information Retrieval, SIGIR '18, page 1221-1224, New York, NY, USA. Association for Computing Machinery." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 491, + 291, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 491, + 291, + 568 + ], + "spans": [ + { + "bbox": [ + 69, + 491, + 291, + 568 + ], + "type": "text", + "content": "Yan Jiang, Jinhua Gao, Huawei Shen, and Xueqi Cheng. 2022. Few-shot stance detection via target-aware prompt distillation. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '22, page 837-847, New York, NY, USA. Association for Computing Machinery." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 577, + 291, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 577, + 291, + 633 + ], + "spans": [ + { + "bbox": [ + 69, + 577, + 291, + 633 + ], + "type": "text", + "content": "Mike Lewis and Angela Fan. 2019. Generative question answering: Learning to answer the whole question. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 641, + 291, + 740 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 641, + 291, + 740 + ], + "spans": [ + { + "bbox": [ + 69, + 641, + 291, + 740 + ], + "type": "text", + "content": "Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871-7880, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 749, + 291, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 749, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 749, + 291, + 772 + ], + "type": "text", + "content": "Sha Li, Heng Ji, and Jiawei Han. 2021. Document-level event argument extraction by conditional generation." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 525, + 772 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 314, + 72, + 525, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 525, + 127 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 525, + 127 + ], + "type": "text", + "content": "In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 894-908, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 137, + 525, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 137, + 525, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 137, + 525, + 204 + ], + "type": "text", + "content": "Bin Liang, Zixiao Chen, Lin Gui, Yulan He, Min Yang, and Ruifeng Xu. 2022a. Zero-shot stance detection via contrastive learning. In Proceedings of the ACM Web Conference 2022, WWW '22, page 2738-2747, New York, NY, USA. Association for Computing Machinery." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 214, + 525, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 214, + 525, + 280 + ], + "spans": [ + { + "bbox": [ + 304, + 214, + 525, + 280 + ], + "type": "text", + "content": "Bin Liang, Yonghao Fu, Lin Gui, Min Yang, Jiachen Du, Yulan He, and Ruifeng Xu. 2021. Target-adaptive graph for cross-target stance detection. In Proceedings of the Web Conference 2021, WWW '21, page 3453-3464, New York, NY, USA. Association for Computing Machinery." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 290, + 525, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 290, + 525, + 368 + ], + "spans": [ + { + "bbox": [ + 304, + 290, + 525, + 368 + ], + "type": "text", + "content": "Bin Liang, Qinglin Zhu, Xiang Li, Min Yang, Lin Gui, Yulan He, and Ruifeng Xu. 2022b. JointCL: A joint contrastive learning framework for zero-shot stance detection. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 81-91, Dublin, Ireland. Association for Computational Linguistics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 377, + 525, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 525, + 455 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 525, + 455 + ], + "type": "text", + "content": "Yuxiao Lin, Yuxian Meng, Xiaofei Sun, Qinghong Han, Kun Kuang, Jiwei Li, and Fei Wu. 2021. BertGCN: Transductive text classification by combining GNN and BERT. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 1456-1462, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 465, + 525, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 465, + 525, + 532 + ], + "spans": [ + { + "bbox": [ + 304, + 465, + 525, + 532 + ], + "type": "text", + "content": "Rui Liu, Zheng Lin, Yutong Tan, and Weiping Wang. 2021. Enhancing zero-shot and few-shot stance detection with commonsense knowledge graph. In *Findings of the Association for Computational Linguistics: ACL-IJCNLP* 2021, pages 3152-3157, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 542, + 525, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 542, + 525, + 576 + ], + "spans": [ + { + "bbox": [ + 304, + 542, + 525, + 576 + ], + "type": "text", + "content": "Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In International Conference on Learning Representations." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 585, + 525, + 662 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 585, + 525, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 585, + 525, + 662 + ], + "type": "text", + "content": "Saif Mohammad, Svetlana Kiritchenko, Parinaz Sobhani, Xiaodan Zhu, and Colin Cherry. 2016. SemEval-2016 task 6: Detecting stance in tweets. In Proceedings of the 10th International Workshop on Semantic Evaluation (SemEval-2016), pages 31-41, San Diego, California. Association for Computational Linguistics." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 672, + 525, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 672, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 672, + 525, + 772 + ], + "type": "text", + "content": "Mitra Mohtarami, Ramy Baly, James Glass, Preslav Nakov, Lluis Márquez, and Alessandro Moschitti. 2018. Automatic stance detection using end-to-end memory networks. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 767-776, New Orleans, Louisiana. Association for Computational Linguistics." + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1496" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 766 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 291, + 127 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 291, + 127 + ], + "type": "text", + "content": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2022. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(1)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 134, + 291, + 211 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 134, + 291, + 211 + ], + "spans": [ + { + "bbox": [ + 69, + 134, + 291, + 211 + ], + "type": "text", + "content": "Parinaz Sobhani, Diana Inkpen, and Xiaodan Zhu. 2017. A dataset for multi-target stance detection. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers, pages 551-557, Valencia, Spain. Association for Computational Linguistics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 218, + 291, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 218, + 291, + 285 + ], + "spans": [ + { + "bbox": [ + 69, + 218, + 291, + 285 + ], + "type": "text", + "content": "Swapna Somasundaran and Janyce Wiebe. 2010. Recognizing stances in ideological on-line debates. In Proceedings of the NAACL HLT 2010 Workshop on Computational Approaches to Analysis and Generation of Emotion in Text, pages 116-124, Los Angeles, CA. Association for Computational Linguistics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 291, + 290, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 291, + 290, + 324 + ], + "spans": [ + { + "bbox": [ + 69, + 291, + 290, + 324 + ], + "type": "text", + "content": "Laurens van der Maaten and Geoffrey Hinton. 2008. Visualizing data using t-sne. Journal of Machine Learning Research, 9(86):2579-2605." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 332, + 291, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 332, + 291, + 397 + ], + "spans": [ + { + "bbox": [ + 69, + 332, + 291, + 397 + ], + "type": "text", + "content": "Sean Welleck, Ilia Kulikov, Stephen Roller, Emily Dinan, Kyunghyun Cho, and Jason Weston. 2020. Neural text generation with unlikelihood training. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 405, + 291, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 405, + 291, + 481 + ], + "spans": [ + { + "bbox": [ + 69, + 405, + 291, + 481 + ], + "type": "text", + "content": "Chang Xu, Cécile Paris, Surya Nepal, and Ross Sparks. 2018. Cross-target stance classification with self-attention networks. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 778-783, Melbourne, Australia. Association for Computational Linguistics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 488, + 291, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 488, + 291, + 577 + ], + "spans": [ + { + "bbox": [ + 69, + 488, + 291, + 577 + ], + "type": "text", + "content": "Hang Yan, Junqi Dai, Tuo Ji, Xipeng Qiu, and Zheng Zhang. 2021. A unified generative framework for aspect-based sentiment analysis. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 2416-2429, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 583, + 291, + 661 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 583, + 291, + 661 + ], + "spans": [ + { + "bbox": [ + 69, + 583, + 291, + 661 + ], + "type": "text", + "content": "Rong Zhang, Qifei Zhou, Bo An, Weiping Li, Tong Mo, and Bo Wu. 2020. Enhancing neural models with vulnerability via adversarial attack. In Proceedings of the 28th International Conference on Computational Linguistics, pages 1133-1146, Barcelona, Spain (Online). International Committee on Computational Linguistics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 667, + 291, + 766 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 667, + 291, + 766 + ], + "spans": [ + { + "bbox": [ + 69, + 667, + 291, + 766 + ], + "type": "text", + "content": "Shaodian Zhang, Lin Qiu, Frank Chen, Weinan Zhang, Yong Yu, and Noémie Elhadad. 2017. We make choices we think are going to save us: Debate and stance identification for online breast cancer cam discussions. In Proceedings of the 26th International Conference on World Wide Web Companion, WWW '17 Companion, page 1073-1081, Republic and Canton of Geneva, CHE. International World Wide Web Conferences Steering Committee." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1497" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 90, + 192, + 103 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 90, + 192, + 103 + ], + "spans": [ + { + "bbox": [ + 68, + 90, + 192, + 103 + ], + "type": "text", + "content": "A For every submission:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 106, + 414, + 240 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "spans": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "type": "text", + "content": "A1. Did you describe the limitations of your work? Limitations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 142, + 329, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 142, + 329, + 168 + ], + "spans": [ + { + "bbox": [ + 77, + 142, + 329, + 168 + ], + "type": "text", + "content": "A2. Did you discuss any potential risks of your work? Limitations" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 178, + 414, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 178, + 414, + 204 + ], + "spans": [ + { + "bbox": [ + 77, + 178, + 414, + 204 + ], + "type": "text", + "content": "A3. Do the abstract and introduction summarize the paper's main claims? Abstract, Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 77, + 214, + 398, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 214, + 398, + 240 + ], + "spans": [ + { + "bbox": [ + 77, + 214, + 398, + 240 + ], + "type": "text", + "content": "A4. Have you used AI writing assistants when working on this paper? Left blank." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 68, + 250, + 290, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 250, + 290, + 264 + ], + "spans": [ + { + "bbox": [ + 68, + 250, + 290, + 264 + ], + "type": "text", + "content": "B Did you use or create scientific artifacts?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 78, + 269, + 214, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 269, + 214, + 280 + ], + "spans": [ + { + "bbox": [ + 78, + 269, + 214, + 280 + ], + "type": "text", + "content": "Introduction, Section 3.1 Data" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 77, + 289, + 524, + 428 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 77, + 289, + 315, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 289, + 315, + 315 + ], + "spans": [ + { + "bbox": [ + 77, + 289, + 315, + 315 + ], + "type": "text", + "content": "B1. Did you cite the creators of artifacts you used? Introduction, Section 3.1 Data" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 77, + 324, + 463, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 324, + 463, + 350 + ], + "spans": [ + { + "bbox": [ + 77, + 324, + 463, + 350 + ], + "type": "text", + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Section 3.1 Data" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 77, + 360, + 524, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 360, + 524, + 428 + ], + "spans": [ + { + "bbox": [ + 77, + 360, + 524, + 428 + ], + "type": "text", + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Section 3.1 Data, Section 3.2 Experimental Setup" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 77, + 438, + 524, + 655 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 77, + 438, + 524, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 438, + 524, + 505 + ], + "spans": [ + { + "bbox": [ + 77, + 438, + 524, + 505 + ], + "type": "text", + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? We use an existing resource and detail of the data is discussed and introduced in their own published paper." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 77, + 514, + 524, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 514, + 524, + 568 + ], + "spans": [ + { + "bbox": [ + 77, + 514, + 524, + 568 + ], + "type": "text", + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? We use an existing resource and detail of the data is discussed and introduced in their own published paper." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 77, + 576, + 524, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 576, + 524, + 655 + ], + "spans": [ + { + "bbox": [ + 77, + 576, + 524, + 655 + ], + "type": "text", + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Section 3.1 Data" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 68, + 665, + 293, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 665, + 293, + 679 + ], + "spans": [ + { + "bbox": [ + 68, + 665, + 293, + 679 + ], + "type": "text", + "content": "C Did you run computational experiments?" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 79, + 684, + 127, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 684, + 127, + 696 + ], + "spans": [ + { + "bbox": [ + 79, + 684, + 127, + 696 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 77, + 704, + 524, + 745 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 704, + 524, + 745 + ], + "spans": [ + { + "bbox": [ + 77, + 704, + 524, + 745 + ], + "type": "text", + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Section 3.2 Experimental Setup" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "spans": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "text", + "content": "ACL 2023 Responsible NLP Checklist" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 751, + 522, + 771 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 751, + 522, + 771 + ], + "spans": [ + { + "bbox": [ + 67, + 751, + 522, + 771 + ], + "type": "text", + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1498" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 77, + 70, + 523, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 70, + 523, + 97 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 523, + 97 + ], + "type": "text", + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 89, + 99, + 229, + 112 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 99, + 229, + 112 + ], + "spans": [ + { + "bbox": [ + 89, + 99, + 229, + 112 + ], + "type": "text", + "content": "Section 3.2 Experimental Setup" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 120, + 525, + 160 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 120, + 525, + 160 + ], + "spans": [ + { + "bbox": [ + 77, + 120, + 525, + 160 + ], + "type": "text", + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 161, + 267, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 161, + 267, + 174 + ], + "spans": [ + { + "bbox": [ + 89, + 161, + 267, + 174 + ], + "type": "text", + "content": "Section 3.2 Experimental Setup, Table 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 183, + 525, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 183, + 525, + 223 + ], + "spans": [ + { + "bbox": [ + 77, + 183, + 525, + 223 + ], + "type": "text", + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 225, + 230, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 225, + 230, + 238 + ], + "spans": [ + { + "bbox": [ + 89, + 225, + 230, + 238 + ], + "type": "text", + "content": "Section 3.2 Experimental Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 247, + 522, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 247, + 522, + 260 + ], + "spans": [ + { + "bbox": [ + 67, + 247, + 522, + 260 + ], + "type": "text", + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 79, + 264, + 127, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 264, + 127, + 277 + ], + "spans": [ + { + "bbox": [ + 79, + 264, + 127, + 277 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 76, + 286, + 525, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 286, + 525, + 313 + ], + "spans": [ + { + "bbox": [ + 76, + 286, + 525, + 313 + ], + "type": "text", + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 89, + 314, + 148, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 314, + 148, + 327 + ], + "spans": [ + { + "bbox": [ + 89, + 314, + 148, + 327 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 76, + 336, + 525, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 336, + 525, + 376 + ], + "spans": [ + { + "bbox": [ + 76, + 336, + 525, + 376 + ], + "type": "text", + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 89, + 378, + 148, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 378, + 148, + 391 + ], + "spans": [ + { + "bbox": [ + 89, + 378, + 148, + 391 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 76, + 400, + 525, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 400, + 525, + 439 + ], + "spans": [ + { + "bbox": [ + 76, + 400, + 525, + 439 + ], + "type": "text", + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 89, + 441, + 148, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 441, + 148, + 454 + ], + "spans": [ + { + "bbox": [ + 89, + 441, + 148, + 454 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 76, + 462, + 520, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 462, + 520, + 476 + ], + "spans": [ + { + "bbox": [ + 76, + 462, + 520, + 476 + ], + "type": "text", + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 89, + 476, + 148, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 476, + 148, + 489 + ], + "spans": [ + { + "bbox": [ + 89, + 476, + 148, + 489 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 76, + 498, + 524, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 498, + 524, + 524 + ], + "spans": [ + { + "bbox": [ + 76, + 498, + 524, + 524 + ], + "type": "text", + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 89, + 527, + 148, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 527, + 148, + 539 + ], + "spans": [ + { + "bbox": [ + 89, + 527, + 148, + 539 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1499" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Zero-shot Cross-lingual Transfer With Learned Projections Using Unlabeled Target-Language Data/ce957704-0301-4c9b-84de-73d565462619_content_list.json b/2023/Zero-shot Cross-lingual Transfer With Learned Projections Using Unlabeled Target-Language Data/ce957704-0301-4c9b-84de-73d565462619_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..31694b53bea47f0ba9cde04a70792059d161850b --- /dev/null +++ b/2023/Zero-shot Cross-lingual Transfer With Learned Projections Using Unlabeled Target-Language Data/ce957704-0301-4c9b-84de-73d565462619_content_list.json @@ -0,0 +1,1435 @@ +[ + { + "type": "text", + "text": "Zero-shot Cross-lingual Transfer With Learned Projections Using Unlabeled Target-Language Data", + "text_level": 1, + "bbox": [ + 191, + 89, + 805, + 130 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ujan Deb*", + "bbox": [ + 203, + 155, + 295, + 172 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "IIT Bhilai", + "bbox": [ + 208, + 173, + 290, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ujand@iitbhilai.ac.in", + "bbox": [ + 142, + 189, + 357, + 204 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ridayesh Parab*", + "bbox": [ + 423, + 155, + 569, + 171 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "IIT Bombay", + "bbox": [ + 447, + 172, + 547, + 187 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ridayesh@gmail.com", + "bbox": [ + 406, + 189, + 589, + 204 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Preethi Jyothi", + "bbox": [ + 687, + 155, + 811, + 171 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "IIT Bombay", + "bbox": [ + 697, + 172, + 801, + 187 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "pjyothi@cse.iitb.ac.in", + "bbox": [ + 638, + 189, + 860, + 204 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 252, + 339, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Adapters have emerged as a parameter-efficient Transformer-based framework for cross-lingual transfer by inserting lightweight language-specific modules (language adapters) and task-specific modules (task adapters) within pretrained multilingual models. Zero-shot transfer is enabled by pairing the language adapter in the target language with an appropriate task adapter in a source language. If our target languages are known apriori, we explore how zero-shot transfer can be further improved within the adapter framework by utilizing unlabeled text during task-specific finetuning. We construct language-specific subspaces using standard linear algebra constructs and selectively project source-language representations into the target language subspace during task-specific finetuning using two schemes. Our experiments on three cross-lingual tasks, Named Entity Recognition (NER), Question Answering (QA) and Natural Language Inference (NLI) yield consistent benefits compared to adapter baselines over a wide variety of target languages with up to $11\\%$ relative improvement in NER, $2\\%$ relative improvement in QA and $5\\%$ relative improvement in NLI.", + "bbox": [ + 141, + 284, + 460, + 653 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 114, + 670, + 258, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zero-shot cross-lingual transfer refers to the transfer of task-specific knowledge from a (high-resource) source language to a (zero-resource) target language that has no labeled task-specific data for training. A popular paradigm for cross-lingual transfer learning is to finetune pretrained multilingual models using labeled task-specific data in the source language and directly evaluate these finetuned models on target language test sets.", + "bbox": [ + 112, + 697, + 489, + 843 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A parameter-efficient alternative to full finetuning for cross-lingual transfer is MAD-X (Pfeiffer et al., 2020b), an adapter-based framework that", + "bbox": [ + 112, + 844, + 489, + 892 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "scaffolds on multilingual pretrained models to combine task-specific and language-specific modules in a plug-and-play manner. Adapters (Houlsby et al., 2019) are feedforward layer blocks inserted within each Transformer layer to selectively learn task-specific and language-specific capabilities via task adapters and language adapters, respectively. Language adapters are trained using self-supervised objectives like masked language modeling (MLM) and task adapters are trained using task-specific objectives. To enable task transfer to a target language, the relevant language and task adapters are combined at test-time.", + "bbox": [ + 507, + 253, + 884, + 461 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the zero-shot setting, we assume access to unlabeled text in the target languages. In MAD-X, this text is only used to train target language adapters and not further used during finetuning. Given knowledge of which languages we want to target, can we make effective use of unlabeled text in the target languages even during task-specific finetuning? This is the main question we tackle in this work.", + "bbox": [ + 507, + 464, + 884, + 608 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose a general adapter-based technique to inject target language bias into task-specific finetuning. Using the unlabeled text in each target language, we construct an affine subspace from contextualized representations for every Transformer layer in the multilingual model. These subspaces are defined using singular value decomposition (SVD) and only need to be computed once per target language. During task-specific finetuning using labeled data in the source language, we project the source representations onto the target language subspaces. This projection can be invoked randomly using a projection probability defined as a hyperparameter. Projections can also be triggered depending on whether the current source representations are closer to the mean embedding of the source language subspace compared to the mean embedding of the target language subspace. We investigate both these projection policies and find", + "bbox": [ + 507, + 613, + 884, + 919 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 136, + 904, + 258, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "449", + "bbox": [ + 485, + 927, + 515, + 940 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics", + "bbox": [ + 226, + 945, + 769, + 957 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Volume 2: Short Papers, pages 449-457", + "bbox": [ + 376, + 958, + 620, + 971 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "July 9-14, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 295, + 972, + 700, + 985 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "that they both improve performance across multiple tasks in multiple languages compared to state-of-the-art adapter baselines. We also release code1 to reproduce our experiments.", + "bbox": [ + 112, + 84, + 489, + 149 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Methodology", + "text_level": 1, + "bbox": [ + 112, + 159, + 263, + 177 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adapters and MAD-X. Adapters for language models (Houlsby et al., 2019) are bottleneck feedforward modules, typically inserted in each Transformer layer of a multilingual model before layer normalization. Instead of finetuning the entire model, only adapters are tuned for a specific task. Pfeiffer et al. (2020b) extended adapter-based fine tuning to support cross-lingual transfer. Their framework called MAD-X (Multiple Adapters for Cross-lingual transfer) comprises of language adapters and task adapters. Language adapters are pretrained using masked language modeling to learn language-specific features. Task adapters are stacked on top of language adapters during downstream task finetuning to learn task-specific information. To achieve zero-shot transfer, the model is trained with a frozen source-language language adapter and a task adapter. During test time, the source-language adapter is replaced with the target-language adapter and evaluated on test instances in the target language.", + "bbox": [ + 112, + 184, + 489, + 523 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Overview of our technique. We are interested in the setting where we have apriori knowledge of which languages we want to target at test time. We aim to bias cross-lingual transfer towards known target languages during task-specific finetuning. We start with MAD-X as our underlying framework and adopt the following 3-step approach:", + "bbox": [ + 112, + 531, + 489, + 644 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We construct layer-specific subspaces for each of the target languages. This is done by computing SVD on contextualized token representations extracted from each layer. See §2.1 for more details.", + "- During task-specific training, we selectively project output representations from the language adapter of a chosen layer onto the target language subspace. These projections are triggered based on two policies: Random projection (§2.2) and Mean Cosine Distance (§2.3). The projected representations are further passed through the task adapter that is trained using labeled data in the source language.", + "- Similar to MAD-X, we evaluate on the target language by simply swapping the source language" + ], + "bbox": [ + 112, + 644, + 489, + 885 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "adapter with the target language adapter while keeping the task adapter fixed. No projection is done during inference.", + "bbox": [ + 507, + 84, + 884, + 131 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 Language Subspaces and Projections", + "text_level": 1, + "bbox": [ + 507, + 143, + 845, + 158 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our objective is to bias the model towards the target language while fine-tuning for a task. For this, we need to extract language-specific information from model representations that jointly exhibit language-specific and language-independent properties. Language-specific subspaces have been typically used to analyze representations in multilingual language models. Choenni and Shutova (2020) showed that individual representations can be used to predict linguistic typological features after projecting onto language-sensitive subspaces. Chang et al. (2022) construct language subspaces with SVD using language-specific contextualized token embeddings. They analyze model performance and other properties after computing layerwise projections of representations to various language subspaces.", + "bbox": [ + 507, + 162, + 884, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We construct subspaces for each of the target languages using SVD and contextualized token representations for unlabeled text in the target language. Consider a pretrained model like XLMR (Conneau et al., 2020) that takes text sequences from the target language as its input. $d$ -dimensional embeddings from a particular layer for a given language $A$ can be grouped into a matrix $\\mathbf{M}_A \\in R^{n \\times d}$ . SVD of $\\mathbf{M}_A$ (after subtracting the mean representation for $A$ ) can be written as: $\\mathbf{M}_A = \\mathbf{U}_A \\Sigma \\mathbf{V}_A^T$ . The right singular matrix $\\mathbf{V}_A$ is considered to be the subspace for language $A$ . These subspaces only need to be computed once for each layer. Next, we look at when projections should be invoked.", + "bbox": [ + 507, + 437, + 884, + 663 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2 Random Projection", + "text_level": 1, + "bbox": [ + 507, + 673, + 710, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For a given target language, during finetuning using task-specific data in the source language, we project the source representations onto the target language subspace with a predetermined probability $p$ . This projection is invoked right before passing the representation through the task adapter, having already passed through the language adapter. To project onto a target subspace, we first shift the target language subspace so that it passes through the source language mean embedding and then take the projection onto the target subspace (Chang et al., 2022). Let $S$ be the source language and $Q$ be the target language. Let subspaces and means of representations from one of the Transformer layers", + "bbox": [ + 507, + 694, + 884, + 919 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1https://github.com/csalt-research/adapter-projections", + "bbox": [ + 112, + 892, + 403, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "450", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "for the source language be $\\mathbf{V}_S$ and $\\pmb{\\mu}_{S}$ , respectively. Projection of a representation $\\mathbf{x}$ on $S$ is given by:", + "bbox": [ + 112, + 84, + 489, + 116 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {P r o j e c t} _ {S} (\\mathbf {x}) = \\mathbf {V} _ {S} \\mathbf {V} _ {S} ^ {T} (\\mathbf {x} - \\boldsymbol {\\mu} _ {S}) + \\boldsymbol {\\mu} _ {S}\n$$\n", + "text_format": "latex", + "bbox": [ + 152, + 129, + 448, + 149 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The projection of $\\mathbf{x}$ onto the target language subspace, that is shifted onto the source subspace, can be computed as:", + "bbox": [ + 112, + 162, + 489, + 209 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {P r o j e c t} _ {Q, \\pmb {\\mu} _ {S}} (\\mathbf {x}) = \\mathbf {V} _ {Q} \\mathbf {V} _ {Q} ^ {T} (\\mathbf {x} - \\pmb {\\mu} _ {S}) + \\pmb {\\mu} _ {S}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 222, + 460, + 243 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The main intuition here is that by probabilistically projecting source representations onto the target language subspace during task-specific finetuning, the model can encode both source and target language information in its representations. The model cannot solely rely on source-language specific features during task-specific training.", + "bbox": [ + 112, + 255, + 490, + 367 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3 Mean Cosine Distance (MCD)", + "text_level": 1, + "bbox": [ + 112, + 379, + 394, + 395 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We suggest another projection scheme, Mean Cosine Distance (MCD), that is more informed than randomly projecting source representations based on a projection probability $p$ . Using MCD, we project those embeddings that are deemed as being further away from the target language subspace compared to the source language subspace. This is quantified using a cosine distance between an embedding from a layer and means of source and target language subspaces. If an embedding is closer to the source language mean compared to the target language mean, we project it onto the target language subspace so as to make it more similar to target language embeddings. However, if an embedding is closer to the target language mean, we can possibly omit projection since it already contains information relevant to the target language.", + "bbox": [ + 112, + 401, + 489, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consider a set of embeddings extracted from one of the Transformer layers. Let the means of all embeddings from this layer and the associated subspace be denoted by $\\pmb{\\mu}$ and $\\mathbf{V}$ , respectively. $\\pmb{\\mu}_{S}$ and $\\pmb{\\mu}_{Q}$ denote the means for the source and target language, respectively. Similarly, $\\mathbf{V}_{S}$ and $\\mathbf{V}_{Q}$ refer to the respective subspaces. Let $\\mathbf{x}$ denote a token embedding from the source language. The MCD policy can be written as:", + "bbox": [ + 112, + 675, + 487, + 820 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} = \\left\\{ \\begin{array}{l l} \\operatorname {P r o j e c t} _ {Q, \\boldsymbol {\\mu} _ {S}} (\\mathbf {x}) & \\mathrm {i f} \\operatorname {c} (\\mathbf {x}, \\boldsymbol {\\mu} _ {Q}) < \\operatorname {c} (\\mathbf {x}, \\boldsymbol {\\mu} _ {S}) \\\\ \\mathbf {x} & \\mathrm {o t h e r w i s e} \\end{array} \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 833, + 485, + 875 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\operatorname{Project}_{Q, \\mu_S}(\\mathbf{x})$ is defined in Section 2.2 as the projection of $\\mathbf{x}$ onto the target subspace $\\mathbf{V}_Q$", + "bbox": [ + 112, + 887, + 487, + 920 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/08256902dcdae9869176721f35b7e68e6a6d47a6545b279356582f90fe9922d2.jpg", + "image_caption": [ + "Figure 1: A single Transformer layer as modified by the MAD-X setup and our projection scheme. During training, the output from the source language adapter is projected onto the target language subspace with probability $p$ for random projection (or, if deemed necessary, by the MCD scheme). Dotted arrows refer to the inference time pathway when representations pass through the target language adapter and no projection is applied." + ], + "image_footnote": [], + "bbox": [ + 600, + 82, + 794, + 363 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and $\\mathrm{c}(\\mathbf{x},\\mathbf{y})$ refers to the cosine similarity between two embeddings $\\mathbf{x}$ and $\\mathbf{y}$ .", + "bbox": [ + 507, + 517, + 882, + 549 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 1 provides an illustration of our proposed technique within a single Transformer layer that includes language and task adapters (as in the MAD-X framework).", + "bbox": [ + 507, + 551, + 882, + 613 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 Experimental Setup", + "text_level": 1, + "bbox": [ + 509, + 632, + 717, + 649 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Subspace construction. To construct language specific subspaces, we adopt the settings used by Chang et al. (2022). Text sequences of length 512 are taken from the OSCAR dataset (Ortiz Su'arez et al., 2019) and passed through XLMR (Conneau et al., 2020) to produce layer-wise contextualized embeddings. We pick 262K contextualized representations and subtract the representation mean before computing SVD. For a low-dimensional subspace, we select the greatest $k$ singular values such that their sum of squares is greater than or equal to $90\\%$ of the total variance. (Total variance is given by the sum of the squared singular values produced.) Finally, in order to compute the language-specific subspaces, the corresponding right singular vectors are taken as the basis.", + "bbox": [ + 505, + 661, + 884, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "451", + "bbox": [ + 485, + 928, + 512, + 940 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/588e62121cef0fa180dafb91ff10ef55087494ce0010df4fad6e79131ec65357.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
NER
hivideidisiloswmyjvavg
MAD-X Adapters68.366.875.949.476.274.074.852.757.366.1
Random Projection68.969.077.553.876.879.876.557.661.269.0
MCD68.568.177.154.776.176.975.453.659.367.7
", + "bbox": [ + 196, + 80, + 801, + 167 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/56d820cf54253f16825943a3ce9a73e3afee00013c8886f77effbdab1ea18583.jpg", + "table_caption": [ + "Table 1: NER results (F1 scores) for nine languages." + ], + "table_footnote": [], + "table_body": "
XQuAD
hivideavg
MAD-X Adapters68.171.471.870.4
Random Projection68.272.272.270.9
MCD68.672.973.571.7
", + "bbox": [ + 144, + 204, + 470, + 288 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Datasets. We conduct cross-lingual transfer experiments on three tasks, Named Entity Recognition (NER), Question Answering (QA) and Natural Language Inference (NLI), where the source language is always English. For NER, we use the WikiANN dataset (Rahimi et al., 2019), and show results for nine languages Hindi, Vietnamese, German, Indonesian, Icelandic, Ilocano, Swahili, Burmese and Javanese with roughly 20K instances in the English train set and between 1K and 10K instances in the target dev and test sets. For QA, we use XQuAD (Artetxe et al., 2019), a multilingual extension of SQuAD (Rajpurkar et al., 2016) and we report results for Hindi, Vietnamese and German consisting of around 87K examples in the English SQuAD train set and 1190 instances in the three target dev sets. For NLI, we use the AmericasNLI dataset (Ebrahimmi et al., 2021) which is an extension of the XNLI dataset (Conneau et al., 2018) with low-resource American languages. We report results on Quechua and Guarani, consisting of 392k instances in the English train set and 2490 and 5010 instances in the dev and test sets, respectively for each target language.", + "bbox": [ + 112, + 338, + 489, + 725 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training setup. We use transformer models from the adapter-transformers $^2$ fork of the HuggingFace transformers library (Wolf et al., 2020). We use pre-trained language adapters from AdapterHub (Pfeiffer et al., 2020a) for our transfer experiments. XQuAD and NLI fine-tuning experiments were conducted on a single NVIDIA A100 80 GB gpu for 15 epochs and 10 epochs, with learning rate 1e-4 and batch size 16. NER experiments were run", + "bbox": [ + 112, + 736, + 487, + 881 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/c30cd14ae80d52b4edb56b4b8a6afeb37a256301cfc4c67d87f6f30eec346e87.jpg", + "table_caption": [ + "Table 2: Results (F1) for QA for three languages" + ], + "table_footnote": [], + "table_body": "
NLI
qugnavg
MAD-X Adapters48.236.042.1
Random Projection49.337.543.4
MCD48.137.842.9
", + "bbox": [ + 549, + 204, + 828, + 288 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 3: Results (F1) for NLI for two languages", + "bbox": [ + 522, + 299, + 853, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for 30 epochs on Nvidia 1080 Ti with 12 GB ram. Further details can be found in Appendix A.", + "bbox": [ + 507, + 338, + 882, + 370 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4 Results", + "text_level": 1, + "bbox": [ + 507, + 382, + 608, + 397 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "NER, XQuAD and NLI results are shown in Table 1, Table 2 and Table 3 respectively. All values correspond to F1 scores averaged over 3 different seeds. We use the target language validation set to choose the best hyperparameter values for all experiments. Both MCD and random projections show consistent improvement over the MAD-X baseline numbers. With MCD, we explicitly instruct the model when to project. This removes a hyperparameter from the setup, compared to random projections, while maintaining consistent performance gains over the baseline. To further analyze MCD,", + "bbox": [ + 505, + 407, + 884, + 601 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d0403a355587a651bb4a059529d5c1231a98c2247c43be65ab97d8009572c6f2.jpg", + "image_caption": [ + "Figure 2: Projection fraction vs layer across epochs." + ], + "image_footnote": [], + "bbox": [ + 547, + 630, + 830, + 781 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we consider the fraction of embeddings being projected onto the target language subspace for NER. Table 4 shows the fraction of embeddings projected during training (averaged across all layers) for each language. For languages dissimilar to en (such as hi and id), it makes sense that the projection fractions", + "bbox": [ + 507, + 822, + 882, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "2https://github.com/adapter-hub/adapter-transformers", + "bbox": [ + 112, + 891, + 381, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "452", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/74fef88abfe1cfc790eecef614b348bec241d378ed65cccb42006edffa3e689a.jpg", + "table_caption": [ + "Table 4: Projection percentages for NER." + ], + "table_footnote": [], + "table_body": "
hivideidis
Proj. Frac.0.650.570.570.630.55
", + "bbox": [ + 134, + 107, + 468, + 145 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "are high since the language subspace means are closer to the source language mean (Chang et al., 2022), compared to languages more similar to en like de and is. Figure 2 shows how projection fractions vary across layers averaged across training epochs. We see high projection rates in early and final layers across languages. This correlates with these layers encoding a lot of English-specific information (Rogers et al., 2020) via training on the task-specific English data, thus triggering projections via MCD often.", + "bbox": [ + 112, + 168, + 489, + 344 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5 Related Work", + "text_level": 1, + "bbox": [ + 112, + 360, + 270, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multilingual language models like mBERT (Devlin, 2018), XLM-R (Conneau et al., 2020) possess some zero-shot cross-lingual capabilities, even without any explicit finetuning on the languages of interest (Wu and Dredze, 2019; Pires et al., 2019). Such transfer without any finetuning could lead to degradation in performance across certain language pairs (Hu et al., 2020). Nevertheless, multilingual models are a good foundation to bootstrap and further develop cross-lingual generalization. While there is a rapidly growing body of work on cross-lingual transfer, very few approaches utilize language-specific subspaces for this purpose. Both Choenni and Shutova (2020) and Chang et al. (2022) construct language-specific subspaces in multilingual models for an exploratory analysis of the model's representations. Yang et al. (2021) use projections on language specific subspaces to remove language specific information from the representations. We note such removal of language bias did not perform well on cross-lingual transfer in our experiments. Parovic et al. (2022) train bilingual language adapters using both source and target language text before task adapter training. However, this requires training language adapters using both source and target language unlabelled text, for every language pair, in addition to training task adapters. In contrast, our setup is a simple architectural extension of MAD-X, requiring no additional training once the subspaces are computed for each language. To the best of our knowledge, ours is the first work to exploit language-specific subspaces for cross-lingual transfer.", + "bbox": [ + 112, + 387, + 489, + 917 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6 Conclusions", + "text_level": 1, + "bbox": [ + 509, + 84, + 650, + 99 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this work, we present a new adapter-based cross-lingual transfer technique for an apriori known set of target languages. We construct language subspaces using contextualized representations for source and target languages. Representations during task-specific training are projected onto the target subspace if they exceed a probability threshold or if they are closer to a mean source embedding. Both schemes consistently improve zero-shot transfer for three natural language understanding tasks across many languages.", + "bbox": [ + 507, + 109, + 885, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 509, + 299, + 682, + 315 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The first author (Ujan) was supported by the Uplink Internship Program of the India Chapter of ACM SIGKDD. The authors are thankful to the anonymous reviewers for their constructive suggestions that helped improve this submission.", + "bbox": [ + 507, + 324, + 885, + 405 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Limitations", + "text_level": 1, + "bbox": [ + 509, + 417, + 616, + 432 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While our proposed projection techniques often improve cross-lingual transfer, the choice of the projection layer and the projection probability in the case of random projection are hyperparameters that vary across tasks and languages. Our ongoing work involves identifying a mechanism via which we can parameterize these quantities, enabling the model to directly learn the optimal layer and probability values for projection.", + "bbox": [ + 507, + 443, + 885, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 510, + 615, + 608, + 630 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mikel Artetxe, Sebastian Ruder, and Dani Yogatama. 2019. On the cross-lingual transferability of monolingual representations. CoRR, abs/1910.11856.", + "Tyler A. Chang, Zhuowen Tu, and Benjamin K. Bergen. 2022. The geometry of multilingual language model representations. arXiv:2205.10964.", + "Rochelle Choenni and Ekaterina Shutova. 2020. What does it mean to be language-agnostic? probing multilingual sentence encoders for typological properties. arXiv:2009.12862.", + "Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettle-moyer, and Veselin Stoyanov. 2020. Unsupervised cross-lingual representation learning at scale. arXiv:1911.02116.", + "Alexis Conneau, Rudy Rinott, Guillaume Lample, Adina Williams, Samuel R. Bowman, Holger Schwenk," + ], + "bbox": [ + 510, + 637, + 885, + 917 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "453", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Veselin Stoyanov. 2018. Xnli: Evaluating crosslingual sentence representations. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics.", + "Jacob Devlin. 2018. Multilingual bert readme document.", + "Abteen Ebrahimi, Manuel Mager, Arturo Oncevay, Vishrav Chaudhary, Luis Chiruzzo, Angela Fan, John Ortega, Ricardo Ramos, Annette Rios, Ivan Vladimir, Gustavo A. Giménez-Lugo, Elisabeth Mager, Graham Neubig, Alexis Palmer, Rolando A. Coto Solano, Ngoc Thang Vu, and Katharina Kann. 2021. Americasnli: Evaluating zero-shot natural language understanding of pretrained multilingual models in truly low-resource languages. CoRR, abs/2104.08726.", + "Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin de Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. 2019. Parameter-efficient transfer learning for nlp. arXiv:1902.00751.", + "Junjie Hu, Sebastian Ruder, Aditya Siddhant, Graham Neubig, Orhan First, and Melvin Johnson. 2020. Xtreme: A massively multilingual multi-task benchmark for evaluating cross-lingual generalization. arXiv:2003.11080.", + "Quentin Lhoest, Albert Villanova del Moral, Yacine Jernite, Abhishek Thakur, Patrick von Platen, Suraj Patil, Julien Chaumond, Mariama Drame, Julien Plu, Lewis Tunstall, Joe Davison, Mario Šaško, Gunjan Chhablani, Bhavitvya Malik, Simon Brandeis, Teven Le Scao, Victor Sanh, Canwen Xu, and Nicolas Patry. 2020. Datasets: A community library for natural language processing. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 175-184, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.", + "Pedro Javier Ortiz Su'arez, Benoit Sagot, and Laurent Romary. 2019. Asynchronous pipelines for processing huge corpora on medium to low resource infrastructures. Proceedings of the Workshop on Challenges in the Management of Large Corpora (CMLC-7) 2019. Cardiff, 22nd July 2019, pages 9 - 16, Mannheim. Leibniz-Institut f\"ur Deutsche Sprache.", + "Marinela Parovic, Goran Glavaš, Ivan Vulić, and Anna Korhonen. 2022. Bad-x: Bilingual adapters improve zero-shot cross-lingual transfer. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1791–1799, Seattle, United States. Association for Computational Linguistics.", + "Jonas Pfeiffer, Andreas Rückle, Clifton Poth, Aishwarya Kamath, Ivan Vulic, Sebastian Ruder, Kyunghyun Cho, and Iryna Gurevych. 2020a. Adapterhub: A framework for adapting transformers. In Proceedings" + ], + "bbox": [ + 115, + 85, + 489, + 917 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 46-54, Online. Association for Computational Linguistics.", + "Jonas Pfeiffer, Ivan Vulic, Iryna Gurevych, and Sebastian Ruder. 2020b. Mad-x: An adapter-based framework for multi-task cross-lingual transfer. arXiv:2005.00052.", + "Telmo Pires, Eva Schlinger, and Dan Garrette. 2019. How multilingual is multilingual bert? Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics.", + "Afshin Rahimi, Yuan Li, and Trevor Cohn. 2019. Massively multilingual transfer for NER. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 151-164, Florence, Italy. Association for Computational Linguistics.", + "Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD: 100,000+ Questions for Machine Comprehension of Text. arXiv e-prints, page arXiv:1606.05250.", + "Anna Rogers, Olga Kovaleva, and Anna Rumshisky. 2020. A primer in bertology: What we know about how bert works. Transactions of the Association for Computational Linguistics, 8:842-866.", + "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Anthony Moi Clement Delangue, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, and Sylvain Gugger. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online. Association for Computational Linguistics.", + "Shijie Wu and Mark Dredze. 2019. Beto, bentz, becas: The surprising cross-lingual effectiveness of bert. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP).", + "Ziyi Yang, Yinfei Yang, Daniel Cer, and Eric Darve. 2021. A simple and effective method to eliminate the self language bias in multilingual representations. arXiv:2109.04727." + ], + "bbox": [ + 510, + 85, + 882, + 766 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A Implementation Details", + "text_level": 1, + "bbox": [ + 510, + 791, + 749, + 807 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use the xlm-roberta-base model from HuggingFace Transformers (Wolf et al., 2020) pretrained on 2.5 TB of CommonCrawl data3, for all our experiments. NLI and XQuAD experiments were conducted on a single NVIDIA A100 GPU (80 GB", + "bbox": [ + 510, + 816, + 882, + 896 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "3https://commoncrawl.org/", + "bbox": [ + 532, + 904, + 721, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "454", + "bbox": [ + 485, + 928, + 515, + 939 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c1546ba656b0badb67c5abe8437f7cf670c45e0ba31d7e3aed373fb1d8168e9f.jpg", + "table_caption": [ + "Table 5: For random projection, best-performing projection layers for different languages obtained via a grid search on validation sets." + ], + "table_footnote": [], + "table_body": "
NERXQuADNLI
hivideidisswilojvmyhividequgn
Random Projections56848668801291
MCD10284060079291111
", + "bbox": [ + 168, + 121, + 828, + 191 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ca7dfc5722eef60b418badf408860ecb3a472fc389bc64886c682dbee69006a9.jpg", + "table_caption": [ + "Table 6: Probability values (as determined by tuning on validation sets) for the layers in Table 5." + ], + "table_footnote": [], + "table_body": "
NERXQuADNLI
hivideidisswilojvmyhividequgn
Random Projections0.10.30.30.90.50.50.30.50.50.50.70.50.10.1
", + "bbox": [ + 149, + 228, + 848, + 281 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "RAM) and the NER experiments ran on a single Nvidia 1080Ti GPU (12 GB RAM). We used a learning rate of 1e-4 with a batch size of 16. The hyperparameter choices for layers and probabilities for our experiments are given in Tables 5 and 6, respectively.", + "bbox": [ + 112, + 305, + 487, + 400 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "All datasets used are taken from HuggingFace Datasets (Lhoest et al., 2020). For evaluating models, we use the HuggingFace Evaluate library as well as the seqval python package", + "bbox": [ + 112, + 401, + 489, + 464 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "4https://huggingface.co/docs/evaluate/index", + "bbox": [ + 134, + 890, + 462, + 904 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "5https://pypi.org/project/seqeval/", + "bbox": [ + 136, + 904, + 394, + 917 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "455", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A For every submission:", + "bbox": [ + 115, + 107, + 322, + 122 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "A1. Did you describe the limitations of your work? 7", + "A2. Did you discuss any potential risks of your work? No potential risks", + "A3. Do the abstract and introduction summarize the paper's main claims?", + "A4. Have you used AI writing assistants when working on this paper? Left blank." + ], + "bbox": [ + 129, + 126, + 695, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "B Did you use or create scientific artifacts?", + "bbox": [ + 114, + 300, + 489, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 132, + 321, + 215, + 336 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "B1. Did you cite the creators of artifacts you used? No response.", + "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? No response.", + "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? No response.", + "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? No response.", + "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? No response.", + "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. No response." + ], + "bbox": [ + 127, + 347, + 880, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "C Did you run computational experiments?", + "bbox": [ + 114, + 764, + 492, + 781 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3", + "bbox": [ + 134, + 787, + 146, + 799 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? 3, appendix", + "bbox": [ + 129, + 810, + 880, + 860 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.", + "bbox": [ + 112, + 868, + 877, + 892 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "ACL 2023 Responsible NLP Checklist", + "bbox": [ + 132, + 84, + 433, + 99 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "456", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?", + "bbox": [ + 129, + 83, + 878, + 115 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3, appendix", + "bbox": [ + 149, + 117, + 240, + 131 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?", + "bbox": [ + 129, + 142, + 882, + 190 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4", + "bbox": [ + 151, + 192, + 166, + 204 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?", + "bbox": [ + 129, + 217, + 882, + 265 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "appendix", + "bbox": [ + 149, + 267, + 223, + 282 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?", + "bbox": [ + 112, + 293, + 877, + 309 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 132, + 313, + 213, + 329 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?", + "bbox": [ + 127, + 340, + 882, + 372 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 373, + 248, + 388 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?", + "bbox": [ + 127, + 399, + 882, + 447 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 449, + 248, + 464 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?", + "bbox": [ + 127, + 475, + 882, + 521 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 524, + 248, + 539 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?", + "bbox": [ + 127, + 549, + 873, + 565 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 565, + 248, + 582 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?", + "bbox": [ + 127, + 592, + 880, + 623 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "No response.", + "bbox": [ + 149, + 626, + 248, + 640 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "457", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 8 + } +] \ No newline at end of file diff --git a/2023/Zero-shot Cross-lingual Transfer With Learned Projections Using Unlabeled Target-Language Data/ce957704-0301-4c9b-84de-73d565462619_model.json b/2023/Zero-shot Cross-lingual Transfer With Learned Projections Using Unlabeled Target-Language Data/ce957704-0301-4c9b-84de-73d565462619_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8a8242cab6c6e294f1224db02de8992a5f028a4a --- /dev/null +++ b/2023/Zero-shot Cross-lingual Transfer With Learned Projections Using Unlabeled Target-Language Data/ce957704-0301-4c9b-84de-73d565462619_model.json @@ -0,0 +1,1835 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.192, + 0.09, + 0.806, + 0.131 + ], + "angle": 0, + "content": "Zero-shot Cross-lingual Transfer With Learned Projections Using Unlabeled Target-Language Data" + }, + { + "type": "text", + "bbox": [ + 0.205, + 0.156, + 0.296, + 0.173 + ], + "angle": 0, + "content": "Ujan Deb*" + }, + { + "type": "text", + "bbox": [ + 0.209, + 0.174, + 0.292, + 0.187 + ], + "angle": 0, + "content": "IIT Bhilai" + }, + { + "type": "text", + "bbox": [ + 0.144, + 0.19, + 0.358, + 0.205 + ], + "angle": 0, + "content": "ujand@iitbhilai.ac.in" + }, + { + "type": "text", + "bbox": [ + 0.425, + 0.156, + 0.571, + 0.172 + ], + "angle": 0, + "content": "Ridayesh Parab*" + }, + { + "type": "text", + "bbox": [ + 0.448, + 0.173, + 0.549, + 0.189 + ], + "angle": 0, + "content": "IIT Bombay" + }, + { + "type": "text", + "bbox": [ + 0.408, + 0.19, + 0.591, + 0.205 + ], + "angle": 0, + "content": "ridayesh@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.688, + 0.156, + 0.812, + 0.172 + ], + "angle": 0, + "content": "Preethi Jyothi" + }, + { + "type": "text", + "bbox": [ + 0.699, + 0.173, + 0.802, + 0.189 + ], + "angle": 0, + "content": "IIT Bombay" + }, + { + "type": "text", + "bbox": [ + 0.639, + 0.19, + 0.862, + 0.205 + ], + "angle": 0, + "content": "pjyothi@cse.iitb.ac.in" + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.253, + 0.341, + 0.268 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.285, + 0.461, + 0.655 + ], + "angle": 0, + "content": "Adapters have emerged as a parameter-efficient Transformer-based framework for cross-lingual transfer by inserting lightweight language-specific modules (language adapters) and task-specific modules (task adapters) within pretrained multilingual models. Zero-shot transfer is enabled by pairing the language adapter in the target language with an appropriate task adapter in a source language. If our target languages are known apriori, we explore how zero-shot transfer can be further improved within the adapter framework by utilizing unlabeled text during task-specific finetuning. We construct language-specific subspaces using standard linear algebra constructs and selectively project source-language representations into the target language subspace during task-specific finetuning using two schemes. Our experiments on three cross-lingual tasks, Named Entity Recognition (NER), Question Answering (QA) and Natural Language Inference (NLI) yield consistent benefits compared to adapter baselines over a wide variety of target languages with up to \\(11\\%\\) relative improvement in NER, \\(2\\%\\) relative improvement in QA and \\(5\\%\\) relative improvement in NLI." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.671, + 0.26, + 0.685 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.699, + 0.49, + 0.844 + ], + "angle": 0, + "content": "Zero-shot cross-lingual transfer refers to the transfer of task-specific knowledge from a (high-resource) source language to a (zero-resource) target language that has no labeled task-specific data for training. A popular paradigm for cross-lingual transfer learning is to finetune pretrained multilingual models using labeled task-specific data in the source language and directly evaluate these finetuned models on target language test sets." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.845, + 0.49, + 0.893 + ], + "angle": 0, + "content": "A parameter-efficient alternative to full finetuning for cross-lingual transfer is MAD-X (Pfeiffer et al., 2020b), an adapter-based framework that" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.254, + 0.885, + 0.462 + ], + "angle": 0, + "content": "scaffolds on multilingual pretrained models to combine task-specific and language-specific modules in a plug-and-play manner. Adapters (Houlsby et al., 2019) are feedforward layer blocks inserted within each Transformer layer to selectively learn task-specific and language-specific capabilities via task adapters and language adapters, respectively. Language adapters are trained using self-supervised objectives like masked language modeling (MLM) and task adapters are trained using task-specific objectives. To enable task transfer to a target language, the relevant language and task adapters are combined at test-time." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.466, + 0.885, + 0.609 + ], + "angle": 0, + "content": "In the zero-shot setting, we assume access to unlabeled text in the target languages. In MAD-X, this text is only used to train target language adapters and not further used during finetuning. Given knowledge of which languages we want to target, can we make effective use of unlabeled text in the target languages even during task-specific finetuning? This is the main question we tackle in this work." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.614, + 0.885, + 0.92 + ], + "angle": 0, + "content": "We propose a general adapter-based technique to inject target language bias into task-specific finetuning. Using the unlabeled text in each target language, we construct an affine subspace from contextualized representations for every Transformer layer in the multilingual model. These subspaces are defined using singular value decomposition (SVD) and only need to be computed once per target language. During task-specific finetuning using labeled data in the source language, we project the source representations onto the target language subspaces. This projection can be invoked randomly using a projection probability defined as a hyperparameter. Projections can also be triggered depending on whether the current source representations are closer to the mean embedding of the source language subspace compared to the mean embedding of the target language subspace. We investigate both these projection policies and find" + }, + { + "type": "page_footnote", + "bbox": [ + 0.137, + 0.905, + 0.26, + 0.919 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.928, + 0.516, + 0.941 + ], + "angle": 0, + "content": "449" + }, + { + "type": "footer", + "bbox": [ + 0.227, + 0.946, + 0.77, + 0.958 + ], + "angle": 0, + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + }, + { + "type": "footer", + "bbox": [ + 0.377, + 0.959, + 0.621, + 0.972 + ], + "angle": 0, + "content": "Volume 2: Short Papers, pages 449-457" + }, + { + "type": "footer", + "bbox": [ + 0.297, + 0.973, + 0.702, + 0.986 + ], + "angle": 0, + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.49, + 0.15 + ], + "angle": 0, + "content": "that they both improve performance across multiple tasks in multiple languages compared to state-of-the-art adapter baselines. We also release code1 to reproduce our experiments." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.16, + 0.265, + 0.178 + ], + "angle": 0, + "content": "2 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.185, + 0.49, + 0.524 + ], + "angle": 0, + "content": "Adapters and MAD-X. Adapters for language models (Houlsby et al., 2019) are bottleneck feedforward modules, typically inserted in each Transformer layer of a multilingual model before layer normalization. Instead of finetuning the entire model, only adapters are tuned for a specific task. Pfeiffer et al. (2020b) extended adapter-based fine tuning to support cross-lingual transfer. Their framework called MAD-X (Multiple Adapters for Cross-lingual transfer) comprises of language adapters and task adapters. Language adapters are pretrained using masked language modeling to learn language-specific features. Task adapters are stacked on top of language adapters during downstream task finetuning to learn task-specific information. To achieve zero-shot transfer, the model is trained with a frozen source-language language adapter and a task adapter. During test time, the source-language adapter is replaced with the target-language adapter and evaluated on test instances in the target language." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.532, + 0.49, + 0.645 + ], + "angle": 0, + "content": "Overview of our technique. We are interested in the setting where we have apriori knowledge of which languages we want to target at test time. We aim to bias cross-lingual transfer towards known target languages during task-specific finetuning. We start with MAD-X as our underlying framework and adopt the following 3-step approach:" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.645, + 0.489, + 0.709 + ], + "angle": 0, + "content": "- We construct layer-specific subspaces for each of the target languages. This is done by computing SVD on contextualized token representations extracted from each layer. See §2.1 for more details." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.71, + 0.49, + 0.853 + ], + "angle": 0, + "content": "- During task-specific training, we selectively project output representations from the language adapter of a chosen layer onto the target language subspace. These projections are triggered based on two policies: Random projection (§2.2) and Mean Cosine Distance (§2.3). The projected representations are further passed through the task adapter that is trained using labeled data in the source language." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.854, + 0.489, + 0.887 + ], + "angle": 0, + "content": "- Similar to MAD-X, we evaluate on the target language by simply swapping the source language" + }, + { + "type": "list", + "bbox": [ + 0.113, + 0.645, + 0.49, + 0.887 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.893, + 0.405, + 0.919 + ], + "angle": 0, + "content": "1https://github.com/csalt-research/adapter-projections" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.085, + 0.885, + 0.133 + ], + "angle": 0, + "content": "adapter with the target language adapter while keeping the task adapter fixed. No projection is done during inference." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.144, + 0.847, + 0.159 + ], + "angle": 0, + "content": "2.1 Language Subspaces and Projections" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.164, + 0.885, + 0.437 + ], + "angle": 0, + "content": "Our objective is to bias the model towards the target language while fine-tuning for a task. For this, we need to extract language-specific information from model representations that jointly exhibit language-specific and language-independent properties. Language-specific subspaces have been typically used to analyze representations in multilingual language models. Choenni and Shutova (2020) showed that individual representations can be used to predict linguistic typological features after projecting onto language-sensitive subspaces. Chang et al. (2022) construct language subspaces with SVD using language-specific contextualized token embeddings. They analyze model performance and other properties after computing layerwise projections of representations to various language subspaces." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.438, + 0.885, + 0.664 + ], + "angle": 0, + "content": "We construct subspaces for each of the target languages using SVD and contextualized token representations for unlabeled text in the target language. Consider a pretrained model like XLMR (Conneau et al., 2020) that takes text sequences from the target language as its input. \\(d\\)-dimensional embeddings from a particular layer for a given language \\(A\\) can be grouped into a matrix \\(\\mathbf{M}_A \\in R^{n \\times d}\\). SVD of \\(\\mathbf{M}_A\\) (after subtracting the mean representation for \\(A\\)) can be written as: \\(\\mathbf{M}_A = \\mathbf{U}_A \\Sigma \\mathbf{V}_A^T\\). The right singular matrix \\(\\mathbf{V}_A\\) is considered to be the subspace for language \\(A\\). These subspaces only need to be computed once for each layer. Next, we look at when projections should be invoked." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.674, + 0.712, + 0.689 + ], + "angle": 0, + "content": "2.2 Random Projection" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.695, + 0.885, + 0.92 + ], + "angle": 0, + "content": "For a given target language, during finetuning using task-specific data in the source language, we project the source representations onto the target language subspace with a predetermined probability \\( p \\). This projection is invoked right before passing the representation through the task adapter, having already passed through the language adapter. To project onto a target subspace, we first shift the target language subspace so that it passes through the source language mean embedding and then take the projection onto the target subspace (Chang et al., 2022). Let \\( S \\) be the source language and \\( Q \\) be the target language. Let subspaces and means of representations from one of the Transformer layers" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.517, + 0.941 + ], + "angle": 0, + "content": "450" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.114, + 0.085, + 0.49, + 0.117 + ], + "angle": 0, + "content": "for the source language be \\(\\mathbf{V}_S\\) and \\(\\pmb{\\mu}_{S}\\), respectively. Projection of a representation \\(\\mathbf{x}\\) on \\(S\\) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.13, + 0.449, + 0.15 + ], + "angle": 0, + "content": "\\[\n\\mathrm {P r o j e c t} _ {S} (\\mathbf {x}) = \\mathbf {V} _ {S} \\mathbf {V} _ {S} ^ {T} (\\mathbf {x} - \\boldsymbol {\\mu} _ {S}) + \\boldsymbol {\\mu} _ {S}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.163, + 0.49, + 0.21 + ], + "angle": 0, + "content": "The projection of \\(\\mathbf{x}\\) onto the target language subspace, that is shifted onto the source subspace, can be computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.139, + 0.223, + 0.462, + 0.244 + ], + "angle": 0, + "content": "\\[\n\\mathrm {P r o j e c t} _ {Q, \\pmb {\\mu} _ {S}} (\\mathbf {x}) = \\mathbf {V} _ {Q} \\mathbf {V} _ {Q} ^ {T} (\\mathbf {x} - \\pmb {\\mu} _ {S}) + \\pmb {\\mu} _ {S}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.256, + 0.491, + 0.368 + ], + "angle": 0, + "content": "The main intuition here is that by probabilistically projecting source representations onto the target language subspace during task-specific finetuning, the model can encode both source and target language information in its representations. The model cannot solely rely on source-language specific features during task-specific training." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.38, + 0.396, + 0.396 + ], + "angle": 0, + "content": "2.3 Mean Cosine Distance (MCD)" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.402, + 0.49, + 0.675 + ], + "angle": 0, + "content": "We suggest another projection scheme, Mean Cosine Distance (MCD), that is more informed than randomly projecting source representations based on a projection probability \\( p \\). Using MCD, we project those embeddings that are deemed as being further away from the target language subspace compared to the source language subspace. This is quantified using a cosine distance between an embedding from a layer and means of source and target language subspaces. If an embedding is closer to the source language mean compared to the target language mean, we project it onto the target language subspace so as to make it more similar to target language embeddings. However, if an embedding is closer to the target language mean, we can possibly omit projection since it already contains information relevant to the target language." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.676, + 0.489, + 0.821 + ], + "angle": 0, + "content": "Consider a set of embeddings extracted from one of the Transformer layers. Let the means of all embeddings from this layer and the associated subspace be denoted by \\(\\pmb{\\mu}\\) and \\(\\mathbf{V}\\), respectively. \\(\\pmb{\\mu}_{S}\\) and \\(\\pmb{\\mu}_{Q}\\) denote the means for the source and target language, respectively. Similarly, \\(\\mathbf{V}_{S}\\) and \\(\\mathbf{V}_{Q}\\) refer to the respective subspaces. Let \\(\\mathbf{x}\\) denote a token embedding from the source language. The MCD policy can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.834, + 0.486, + 0.876 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} = \\left\\{ \\begin{array}{l l} \\operatorname {P r o j e c t} _ {Q, \\boldsymbol {\\mu} _ {S}} (\\mathbf {x}) & \\mathrm {i f} \\operatorname {c} (\\mathbf {x}, \\boldsymbol {\\mu} _ {Q}) < \\operatorname {c} (\\mathbf {x}, \\boldsymbol {\\mu} _ {S}) \\\\ \\mathbf {x} & \\mathrm {o t h e r w i s e} \\end{array} \\right.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.888, + 0.488, + 0.921 + ], + "angle": 0, + "content": "where \\(\\operatorname{Project}_{Q, \\mu_S}(\\mathbf{x})\\) is defined in Section 2.2 as the projection of \\(\\mathbf{x}\\) onto the target subspace \\(\\mathbf{V}_Q\\)" + }, + { + "type": "image", + "bbox": [ + 0.601, + 0.083, + 0.795, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.375, + 0.885, + 0.492 + ], + "angle": 0, + "content": "Figure 1: A single Transformer layer as modified by the MAD-X setup and our projection scheme. During training, the output from the source language adapter is projected onto the target language subspace with probability \\( p \\) for random projection (or, if deemed necessary, by the MCD scheme). Dotted arrows refer to the inference time pathway when representations pass through the target language adapter and no projection is applied." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.518, + 0.883, + 0.55 + ], + "angle": 0, + "content": "and \\( \\mathrm{c}(\\mathbf{x},\\mathbf{y}) \\) refers to the cosine similarity between two embeddings \\( \\mathbf{x} \\) and \\( \\mathbf{y} \\)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.552, + 0.884, + 0.614 + ], + "angle": 0, + "content": "Figure 1 provides an illustration of our proposed technique within a single Transformer layer that includes language and task adapters (as in the MAD-X framework)." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.633, + 0.719, + 0.65 + ], + "angle": 0, + "content": "3 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.662, + 0.885, + 0.919 + ], + "angle": 0, + "content": "Subspace construction. To construct language specific subspaces, we adopt the settings used by Chang et al. (2022). Text sequences of length 512 are taken from the OSCAR dataset (Ortiz Su'arez et al., 2019) and passed through XLMR (Conneau et al., 2020) to produce layer-wise contextualized embeddings. We pick 262K contextualized representations and subtract the representation mean before computing SVD. For a low-dimensional subspace, we select the greatest \\( k \\) singular values such that their sum of squares is greater than or equal to \\( 90\\% \\) of the total variance. (Total variance is given by the sum of the squared singular values produced.) Finally, in order to compute the language-specific subspaces, the corresponding right singular vectors are taken as the basis." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.514, + 0.941 + ], + "angle": 0, + "content": "451" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.197, + 0.082, + 0.803, + 0.168 + ], + "angle": 0, + "content": "
NER
hivideidisiloswmyjvavg
MAD-X Adapters68.366.875.949.476.274.074.852.757.366.1
Random Projection68.969.077.553.876.879.876.557.661.269.0
MCD68.568.177.154.776.176.975.453.659.367.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.317, + 0.177, + 0.679, + 0.192 + ], + "angle": 0, + "content": "Table 1: NER results (F1 scores) for nine languages." + }, + { + "type": "table", + "bbox": [ + 0.146, + 0.205, + 0.472, + 0.29 + ], + "angle": 0, + "content": "
XQuAD
hivideavg
MAD-X Adapters68.171.471.870.4
Random Projection68.272.272.270.9
MCD68.672.973.571.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.14, + 0.3, + 0.476, + 0.314 + ], + "angle": 0, + "content": "Table 2: Results (F1) for QA for three languages" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.34, + 0.49, + 0.726 + ], + "angle": 0, + "content": "Datasets. We conduct cross-lingual transfer experiments on three tasks, Named Entity Recognition (NER), Question Answering (QA) and Natural Language Inference (NLI), where the source language is always English. For NER, we use the WikiANN dataset (Rahimi et al., 2019), and show results for nine languages Hindi, Vietnamese, German, Indonesian, Icelandic, Ilocano, Swahili, Burmese and Javanese with roughly 20K instances in the English train set and between 1K and 10K instances in the target dev and test sets. For QA, we use XQuAD (Artetxe et al., 2019), a multilingual extension of SQuAD (Rajpurkar et al., 2016) and we report results for Hindi, Vietnamese and German consisting of around 87K examples in the English SQuAD train set and 1190 instances in the three target dev sets. For NLI, we use the AmericasNLI dataset (Ebrahimmi et al., 2021) which is an extension of the XNLI dataset (Conneau et al., 2018) with low-resource American languages. We report results on Quechua and Guarani, consisting of 392k instances in the English train set and 2490 and 5010 instances in the dev and test sets, respectively for each target language." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.737, + 0.489, + 0.882 + ], + "angle": 0, + "content": "Training setup. We use transformer models from the adapter-transformers\\(^2\\) fork of the HuggingFace transformers library (Wolf et al., 2020). We use pre-trained language adapters from AdapterHub (Pfeiffer et al., 2020a) for our transfer experiments. XQuAD and NLI fine-tuning experiments were conducted on a single NVIDIA A100 80 GB gpu for 15 epochs and 10 epochs, with learning rate 1e-4 and batch size 16. NER experiments were run" + }, + { + "type": "table", + "bbox": [ + 0.55, + 0.205, + 0.83, + 0.29 + ], + "angle": 0, + "content": "
NLI
qugnavg
MAD-X Adapters48.236.042.1
Random Projection49.337.543.4
MCD48.137.842.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.523, + 0.3, + 0.855, + 0.314 + ], + "angle": 0, + "content": "Table 3: Results (F1) for NLI for two languages" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.34, + 0.884, + 0.372 + ], + "angle": 0, + "content": "for 30 epochs on Nvidia 1080 Ti with 12 GB ram. Further details can be found in Appendix A." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.383, + 0.609, + 0.398 + ], + "angle": 0, + "content": "4 Results" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.408, + 0.885, + 0.602 + ], + "angle": 0, + "content": "NER, XQuAD and NLI results are shown in Table 1, Table 2 and Table 3 respectively. All values correspond to F1 scores averaged over 3 different seeds. We use the target language validation set to choose the best hyperparameter values for all experiments. Both MCD and random projections show consistent improvement over the MAD-X baseline numbers. With MCD, we explicitly instruct the model when to project. This removes a hyperparameter from the setup, compared to random projections, while maintaining consistent performance gains over the baseline. To further analyze MCD," + }, + { + "type": "image", + "bbox": [ + 0.548, + 0.631, + 0.831, + 0.782 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.793, + 0.874, + 0.809 + ], + "angle": 0, + "content": "Figure 2: Projection fraction vs layer across epochs." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.823, + 0.884, + 0.919 + ], + "angle": 0, + "content": "we consider the fraction of embeddings being projected onto the target language subspace for NER. Table 4 shows the fraction of embeddings projected during training (averaged across all layers) for each language. For languages dissimilar to en (such as hi and id), it makes sense that the projection fractions" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.892, + 0.382, + 0.918 + ], + "angle": 0, + "content": "2https://github.com/adapter-hub/adapter-transformers" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.517, + 0.941 + ], + "angle": 0, + "content": "452" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.158, + 0.083, + 0.444, + 0.098 + ], + "angle": 0, + "content": "Table 4: Projection percentages for NER." + }, + { + "type": "table", + "bbox": [ + 0.135, + 0.108, + 0.47, + 0.146 + ], + "angle": 0, + "content": "
hivideidis
Proj. Frac.0.650.570.570.630.55
" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.17, + 0.49, + 0.346 + ], + "angle": 0, + "content": "are high since the language subspace means are closer to the source language mean (Chang et al., 2022), compared to languages more similar to en like de and is. Figure 2 shows how projection fractions vary across layers averaged across training epochs. We see high projection rates in early and final layers across languages. This correlates with these layers encoding a lot of English-specific information (Rogers et al., 2020) via training on the task-specific English data, thus triggering projections via MCD often." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.361, + 0.271, + 0.377 + ], + "angle": 0, + "content": "5 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.388, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Multilingual language models like mBERT (Devlin, 2018), XLM-R (Conneau et al., 2020) possess some zero-shot cross-lingual capabilities, even without any explicit finetuning on the languages of interest (Wu and Dredze, 2019; Pires et al., 2019). Such transfer without any finetuning could lead to degradation in performance across certain language pairs (Hu et al., 2020). Nevertheless, multilingual models are a good foundation to bootstrap and further develop cross-lingual generalization. While there is a rapidly growing body of work on cross-lingual transfer, very few approaches utilize language-specific subspaces for this purpose. Both Choenni and Shutova (2020) and Chang et al. (2022) construct language-specific subspaces in multilingual models for an exploratory analysis of the model's representations. Yang et al. (2021) use projections on language specific subspaces to remove language specific information from the representations. We note such removal of language bias did not perform well on cross-lingual transfer in our experiments. Parovic et al. (2022) train bilingual language adapters using both source and target language text before task adapter training. However, this requires training language adapters using both source and target language unlabelled text, for every language pair, in addition to training task adapters. In contrast, our setup is a simple architectural extension of MAD-X, requiring no additional training once the subspaces are computed for each language. To the best of our knowledge, ours is the first work to exploit language-specific subspaces for cross-lingual transfer." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.085, + 0.651, + 0.1 + ], + "angle": 0, + "content": "6 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.11, + 0.886, + 0.288 + ], + "angle": 0, + "content": "In this work, we present a new adapter-based cross-lingual transfer technique for an apriori known set of target languages. We construct language subspaces using contextualized representations for source and target languages. Representations during task-specific training are projected onto the target subspace if they exceed a probability threshold or if they are closer to a mean source embedding. Both schemes consistently improve zero-shot transfer for three natural language understanding tasks across many languages." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.3, + 0.683, + 0.316 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.325, + 0.886, + 0.406 + ], + "angle": 0, + "content": "The first author (Ujan) was supported by the Uplink Internship Program of the India Chapter of ACM SIGKDD. The authors are thankful to the anonymous reviewers for their constructive suggestions that helped improve this submission." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.418, + 0.617, + 0.433 + ], + "angle": 0, + "content": "Limitations" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.444, + 0.886, + 0.589 + ], + "angle": 0, + "content": "While our proposed projection techniques often improve cross-lingual transfer, the choice of the projection layer and the projection probability in the case of random projection are hyperparameters that vary across tasks and languages. Our ongoing work involves identifying a mechanism via which we can parameterize these quantities, enabling the model to directly learn the optimal layer and probability values for projection." + }, + { + "type": "title", + "bbox": [ + 0.511, + 0.616, + 0.61, + 0.631 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.638, + 0.886, + 0.68 + ], + "angle": 0, + "content": "Mikel Artetxe, Sebastian Ruder, and Dani Yogatama. 2019. On the cross-lingual transferability of monolingual representations. CoRR, abs/1910.11856." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.688, + 0.886, + 0.73 + ], + "angle": 0, + "content": "Tyler A. Chang, Zhuowen Tu, and Benjamin K. Bergen. 2022. The geometry of multilingual language model representations. arXiv:2205.10964." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.739, + 0.886, + 0.791 + ], + "angle": 0, + "content": "Rochelle Choenni and Ekaterina Shutova. 2020. What does it mean to be language-agnostic? probing multilingual sentence encoders for typological properties. arXiv:2009.12862." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.802, + 0.886, + 0.881 + ], + "angle": 0, + "content": "Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettle-moyer, and Veselin Stoyanov. 2020. Unsupervised cross-lingual representation learning at scale. arXiv:1911.02116." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.892, + 0.886, + 0.919 + ], + "angle": 0, + "content": "Alexis Conneau, Rudy Rinott, Guillaume Lample, Adina Williams, Samuel R. Bowman, Holger Schwenk," + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.638, + 0.886, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.517, + 0.941 + ], + "angle": 0, + "content": "453" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.135, + 0.086, + 0.49, + 0.152 + ], + "angle": 0, + "content": "and Veselin Stoyanov. 2018. Xnli: Evaluating crosslingual sentence representations. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.163, + 0.489, + 0.188 + ], + "angle": 0, + "content": "Jacob Devlin. 2018. Multilingual bert readme document." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.2, + 0.489, + 0.317 + ], + "angle": 0, + "content": "Abteen Ebrahimi, Manuel Mager, Arturo Oncevay, Vishrav Chaudhary, Luis Chiruzzo, Angela Fan, John Ortega, Ricardo Ramos, Annette Rios, Ivan Vladimir, Gustavo A. Giménez-Lugo, Elisabeth Mager, Graham Neubig, Alexis Palmer, Rolando A. Coto Solano, Ngoc Thang Vu, and Katharina Kann. 2021. Americasnli: Evaluating zero-shot natural language understanding of pretrained multilingual models in truly low-resource languages. CoRR, abs/2104.08726." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.328, + 0.489, + 0.393 + ], + "angle": 0, + "content": "Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin de Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. 2019. Parameter-efficient transfer learning for nlp. arXiv:1902.00751." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.405, + 0.489, + 0.469 + ], + "angle": 0, + "content": "Junjie Hu, Sebastian Ruder, Aditya Siddhant, Graham Neubig, Orhan First, and Melvin Johnson. 2020. Xtreme: A massively multilingual multi-task benchmark for evaluating cross-lingual generalization. arXiv:2003.11080." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.481, + 0.489, + 0.638 + ], + "angle": 0, + "content": "Quentin Lhoest, Albert Villanova del Moral, Yacine Jernite, Abhishek Thakur, Patrick von Platen, Suraj Patil, Julien Chaumond, Mariama Drame, Julien Plu, Lewis Tunstall, Joe Davison, Mario Šaško, Gunjan Chhablani, Bhavitvya Malik, Simon Brandeis, Teven Le Scao, Victor Sanh, Canwen Xu, and Nicolas Patry. 2020. Datasets: A community library for natural language processing. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 175-184, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.648, + 0.489, + 0.74 + ], + "angle": 0, + "content": "Pedro Javier Ortiz Su'arez, Benoit Sagot, and Laurent Romary. 2019. Asynchronous pipelines for processing huge corpora on medium to low resource infrastructures. Proceedings of the Workshop on Challenges in the Management of Large Corpora (CMLC-7) 2019. Cardiff, 22nd July 2019, pages 9 - 16, Mannheim. Leibniz-Institut f\"ur Deutsche Sprache." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.75, + 0.489, + 0.856 + ], + "angle": 0, + "content": "Marinela Parovic, Goran Glavaš, Ivan Vulić, and Anna Korhonen. 2022. Bad-x: Bilingual adapters improve zero-shot cross-lingual transfer. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1791–1799, Seattle, United States. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.866, + 0.489, + 0.919 + ], + "angle": 0, + "content": "Jonas Pfeiffer, Andreas Rückle, Clifton Poth, Aishwarya Kamath, Ivan Vulic, Sebastian Ruder, Kyunghyun Cho, and Iryna Gurevych. 2020a. Adapterhub: A framework for adapting transformers. In Proceedings" + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.529, + 0.086, + 0.883, + 0.139 + ], + "angle": 0, + "content": "of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 46-54, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.149, + 0.883, + 0.2 + ], + "angle": 0, + "content": "Jonas Pfeiffer, Ivan Vulic, Iryna Gurevych, and Sebastian Ruder. 2020b. Mad-x: An adapter-based framework for multi-task cross-lingual transfer. arXiv:2005.00052." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.211, + 0.883, + 0.264 + ], + "angle": 0, + "content": "Telmo Pires, Eva Schlinger, and Dan Garrette. 2019. How multilingual is multilingual bert? Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.274, + 0.883, + 0.34 + ], + "angle": 0, + "content": "Afshin Rahimi, Yuan Li, and Trevor Cohn. 2019. Massively multilingual transfer for NER. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 151-164, Florence, Italy. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.349, + 0.883, + 0.402 + ], + "angle": 0, + "content": "Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD: 100,000+ Questions for Machine Comprehension of Text. arXiv e-prints, page arXiv:1606.05250." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.412, + 0.883, + 0.464 + ], + "angle": 0, + "content": "Anna Rogers, Olga Kovaleva, and Anna Rumshisky. 2020. A primer in bertology: What we know about how bert works. Transactions of the Association for Computational Linguistics, 8:842-866." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.474, + 0.883, + 0.618 + ], + "angle": 0, + "content": "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Anthony Moi Clement Delangue, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, and Sylvain Gugger. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.627, + 0.883, + 0.706 + ], + "angle": 0, + "content": "Shijie Wu and Mark Dredze. 2019. Beto, bentz, becas: The surprising cross-lingual effectiveness of bert. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.715, + 0.883, + 0.768 + ], + "angle": 0, + "content": "Ziyi Yang, Yinfei Yang, Daniel Cer, and Eric Darve. 2021. A simple and effective method to eliminate the self language bias in multilingual representations. arXiv:2109.04727." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.883, + 0.768 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.512, + 0.792, + 0.75, + 0.808 + ], + "angle": 0, + "content": "A Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.817, + 0.883, + 0.897 + ], + "angle": 0, + "content": "We use the xlm-roberta-base model from HuggingFace Transformers (Wolf et al., 2020) pretrained on 2.5 TB of CommonCrawl data3, for all our experiments. NLI and XQuAD experiments were conducted on a single NVIDIA A100 GPU (80 GB" + }, + { + "type": "page_footnote", + "bbox": [ + 0.533, + 0.905, + 0.722, + 0.918 + ], + "angle": 0, + "content": "3https://commoncrawl.org/" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.516, + 0.94 + ], + "angle": 0, + "content": "454" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.083, + 0.883, + 0.111 + ], + "angle": 0, + "content": "Table 5: For random projection, best-performing projection layers for different languages obtained via a grid search on validation sets." + }, + { + "type": "table", + "bbox": [ + 0.169, + 0.122, + 0.83, + 0.192 + ], + "angle": 0, + "content": "
NERXQuADNLI
hivideidisswilojvmyhividequgn
Random Projections56848668801291
MCD10284060079291111
" + }, + { + "type": "table_caption", + "bbox": [ + 0.17, + 0.203, + 0.825, + 0.218 + ], + "angle": 0, + "content": "Table 6: Probability values (as determined by tuning on validation sets) for the layers in Table 5." + }, + { + "type": "table", + "bbox": [ + 0.15, + 0.229, + 0.849, + 0.282 + ], + "angle": 0, + "content": "
NERXQuADNLI
hivideidisswilojvmyhividequgn
Random Projections0.10.30.30.90.50.50.30.50.50.50.70.50.10.1
" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.306, + 0.489, + 0.401 + ], + "angle": 0, + "content": "RAM) and the NER experiments ran on a single Nvidia 1080Ti GPU (12 GB RAM). We used a learning rate of 1e-4 with a batch size of 16. The hyperparameter choices for layers and probabilities for our experiments are given in Tables 5 and 6, respectively." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.402, + 0.49, + 0.466 + ], + "angle": 0, + "content": "All datasets used are taken from HuggingFace Datasets (Lhoest et al., 2020). For evaluating models, we use the HuggingFace Evaluate library as well as the seqval python package" + }, + { + "type": "page_footnote", + "bbox": [ + 0.136, + 0.891, + 0.463, + 0.905 + ], + "angle": 0, + "content": "4https://huggingface.co/docs/evaluate/index" + }, + { + "type": "page_footnote", + "bbox": [ + 0.137, + 0.905, + 0.395, + 0.919 + ], + "angle": 0, + "content": "5https://pypi.org/project/seqeval/" + }, + { + "type": "list", + "bbox": [ + 0.136, + 0.891, + 0.463, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.516, + 0.941 + ], + "angle": 0, + "content": "455" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.133, + 0.085, + 0.435, + 0.1 + ], + "angle": 0, + "content": "ACL 2023 Responsible NLP Checklist" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.108, + 0.323, + 0.123 + ], + "angle": 0, + "content": "A For every submission:" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.127, + 0.533, + 0.158 + ], + "angle": 0, + "content": "A1. Did you describe the limitations of your work? 7" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.171, + 0.553, + 0.202 + ], + "angle": 0, + "content": "A2. Did you discuss any potential risks of your work? No potential risks" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.212, + 0.696, + 0.244 + ], + "angle": 0, + "content": "A3. Do the abstract and introduction summarize the paper's main claims?" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.256, + 0.669, + 0.288 + ], + "angle": 0, + "content": "A4. Have you used AI writing assistants when working on this paper? Left blank." + }, + { + "type": "list", + "bbox": [ + 0.13, + 0.127, + 0.696, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.301, + 0.49, + 0.317 + ], + "angle": 0, + "content": "B Did you use or create scientific artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.322, + 0.216, + 0.337 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.348, + 0.531, + 0.38 + ], + "angle": 0, + "content": "B1. Did you cite the creators of artifacts you used? No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.391, + 0.779, + 0.423 + ], + "angle": 0, + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.434, + 0.881, + 0.514 + ], + "angle": 0, + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.525, + 0.881, + 0.588 + ], + "angle": 0, + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.6, + 0.881, + 0.648 + ], + "angle": 0, + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.659, + 0.881, + 0.755 + ], + "angle": 0, + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. No response." + }, + { + "type": "list", + "bbox": [ + 0.129, + 0.348, + 0.881, + 0.755 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.765, + 0.494, + 0.782 + ], + "angle": 0, + "content": "C Did you run computational experiments?" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.788, + 0.147, + 0.8 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.812, + 0.881, + 0.861 + ], + "angle": 0, + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? 3, appendix" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.869, + 0.878, + 0.893 + ], + "angle": 0, + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.517, + 0.941 + ], + "angle": 0, + "content": "456" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.13, + 0.084, + 0.88, + 0.116 + ], + "angle": 0, + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.118, + 0.242, + 0.133 + ], + "angle": 0, + "content": "3, appendix" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.143, + 0.884, + 0.191 + ], + "angle": 0, + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.193, + 0.167, + 0.205 + ], + "angle": 0, + "content": "4" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.218, + 0.884, + 0.266 + ], + "angle": 0, + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.268, + 0.225, + 0.283 + ], + "angle": 0, + "content": "appendix" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.294, + 0.878, + 0.31 + ], + "angle": 0, + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.315, + 0.215, + 0.33 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.341, + 0.883, + 0.373 + ], + "angle": 0, + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.374, + 0.25, + 0.39 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.4, + 0.884, + 0.448 + ], + "angle": 0, + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.45, + 0.25, + 0.465 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.476, + 0.884, + 0.523 + ], + "angle": 0, + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.525, + 0.25, + 0.54 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.55, + 0.875, + 0.566 + ], + "angle": 0, + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.567, + 0.25, + 0.583 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.593, + 0.881, + 0.624 + ], + "angle": 0, + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.627, + 0.25, + 0.642 + ], + "angle": 0, + "content": "No response." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.516, + 0.941 + ], + "angle": 0, + "content": "457" + } + ] +] \ No newline at end of file diff --git a/2023/Zero-shot Cross-lingual Transfer With Learned Projections Using Unlabeled Target-Language Data/ce957704-0301-4c9b-84de-73d565462619_origin.pdf b/2023/Zero-shot Cross-lingual Transfer With Learned Projections Using Unlabeled Target-Language Data/ce957704-0301-4c9b-84de-73d565462619_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d63279fb33a28d813aea2da711366aec8ec34cc0 --- /dev/null +++ b/2023/Zero-shot Cross-lingual Transfer With Learned Projections Using Unlabeled Target-Language Data/ce957704-0301-4c9b-84de-73d565462619_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4408f0dd39f5a2a05e9a6a3ea7b67362b68d43b976e824f3d24d5ad0045dbdcb +size 420310 diff --git a/2023/Zero-shot Cross-lingual Transfer With Learned Projections Using Unlabeled Target-Language Data/full.md b/2023/Zero-shot Cross-lingual Transfer With Learned Projections Using Unlabeled Target-Language Data/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dca240fe8f213ad6243ce255cf9654cf2810d0ee --- /dev/null +++ b/2023/Zero-shot Cross-lingual Transfer With Learned Projections Using Unlabeled Target-Language Data/full.md @@ -0,0 +1,251 @@ +# Zero-shot Cross-lingual Transfer With Learned Projections Using Unlabeled Target-Language Data + +Ujan Deb* + +IIT Bhilai + +ujand@iitbhilai.ac.in + +Ridayesh Parab* + +IIT Bombay + +ridayesh@gmail.com + +Preethi Jyothi + +IIT Bombay + +pjyothi@cse.iitb.ac.in + +# Abstract + +Adapters have emerged as a parameter-efficient Transformer-based framework for cross-lingual transfer by inserting lightweight language-specific modules (language adapters) and task-specific modules (task adapters) within pretrained multilingual models. Zero-shot transfer is enabled by pairing the language adapter in the target language with an appropriate task adapter in a source language. If our target languages are known apriori, we explore how zero-shot transfer can be further improved within the adapter framework by utilizing unlabeled text during task-specific finetuning. We construct language-specific subspaces using standard linear algebra constructs and selectively project source-language representations into the target language subspace during task-specific finetuning using two schemes. Our experiments on three cross-lingual tasks, Named Entity Recognition (NER), Question Answering (QA) and Natural Language Inference (NLI) yield consistent benefits compared to adapter baselines over a wide variety of target languages with up to $11\%$ relative improvement in NER, $2\%$ relative improvement in QA and $5\%$ relative improvement in NLI. + +# 1 Introduction + +Zero-shot cross-lingual transfer refers to the transfer of task-specific knowledge from a (high-resource) source language to a (zero-resource) target language that has no labeled task-specific data for training. A popular paradigm for cross-lingual transfer learning is to finetune pretrained multilingual models using labeled task-specific data in the source language and directly evaluate these finetuned models on target language test sets. + +A parameter-efficient alternative to full finetuning for cross-lingual transfer is MAD-X (Pfeiffer et al., 2020b), an adapter-based framework that + +scaffolds on multilingual pretrained models to combine task-specific and language-specific modules in a plug-and-play manner. Adapters (Houlsby et al., 2019) are feedforward layer blocks inserted within each Transformer layer to selectively learn task-specific and language-specific capabilities via task adapters and language adapters, respectively. Language adapters are trained using self-supervised objectives like masked language modeling (MLM) and task adapters are trained using task-specific objectives. To enable task transfer to a target language, the relevant language and task adapters are combined at test-time. + +In the zero-shot setting, we assume access to unlabeled text in the target languages. In MAD-X, this text is only used to train target language adapters and not further used during finetuning. Given knowledge of which languages we want to target, can we make effective use of unlabeled text in the target languages even during task-specific finetuning? This is the main question we tackle in this work. + +We propose a general adapter-based technique to inject target language bias into task-specific finetuning. Using the unlabeled text in each target language, we construct an affine subspace from contextualized representations for every Transformer layer in the multilingual model. These subspaces are defined using singular value decomposition (SVD) and only need to be computed once per target language. During task-specific finetuning using labeled data in the source language, we project the source representations onto the target language subspaces. This projection can be invoked randomly using a projection probability defined as a hyperparameter. Projections can also be triggered depending on whether the current source representations are closer to the mean embedding of the source language subspace compared to the mean embedding of the target language subspace. We investigate both these projection policies and find + +that they both improve performance across multiple tasks in multiple languages compared to state-of-the-art adapter baselines. We also release code1 to reproduce our experiments. + +# 2 Methodology + +Adapters and MAD-X. Adapters for language models (Houlsby et al., 2019) are bottleneck feedforward modules, typically inserted in each Transformer layer of a multilingual model before layer normalization. Instead of finetuning the entire model, only adapters are tuned for a specific task. Pfeiffer et al. (2020b) extended adapter-based fine tuning to support cross-lingual transfer. Their framework called MAD-X (Multiple Adapters for Cross-lingual transfer) comprises of language adapters and task adapters. Language adapters are pretrained using masked language modeling to learn language-specific features. Task adapters are stacked on top of language adapters during downstream task finetuning to learn task-specific information. To achieve zero-shot transfer, the model is trained with a frozen source-language language adapter and a task adapter. During test time, the source-language adapter is replaced with the target-language adapter and evaluated on test instances in the target language. + +Overview of our technique. We are interested in the setting where we have apriori knowledge of which languages we want to target at test time. We aim to bias cross-lingual transfer towards known target languages during task-specific finetuning. We start with MAD-X as our underlying framework and adopt the following 3-step approach: + +- We construct layer-specific subspaces for each of the target languages. This is done by computing SVD on contextualized token representations extracted from each layer. See §2.1 for more details. +- During task-specific training, we selectively project output representations from the language adapter of a chosen layer onto the target language subspace. These projections are triggered based on two policies: Random projection (§2.2) and Mean Cosine Distance (§2.3). The projected representations are further passed through the task adapter that is trained using labeled data in the source language. +- Similar to MAD-X, we evaluate on the target language by simply swapping the source language + +adapter with the target language adapter while keeping the task adapter fixed. No projection is done during inference. + +# 2.1 Language Subspaces and Projections + +Our objective is to bias the model towards the target language while fine-tuning for a task. For this, we need to extract language-specific information from model representations that jointly exhibit language-specific and language-independent properties. Language-specific subspaces have been typically used to analyze representations in multilingual language models. Choenni and Shutova (2020) showed that individual representations can be used to predict linguistic typological features after projecting onto language-sensitive subspaces. Chang et al. (2022) construct language subspaces with SVD using language-specific contextualized token embeddings. They analyze model performance and other properties after computing layerwise projections of representations to various language subspaces. + +We construct subspaces for each of the target languages using SVD and contextualized token representations for unlabeled text in the target language. Consider a pretrained model like XLMR (Conneau et al., 2020) that takes text sequences from the target language as its input. $d$ -dimensional embeddings from a particular layer for a given language $A$ can be grouped into a matrix $\mathbf{M}_A \in R^{n \times d}$ . SVD of $\mathbf{M}_A$ (after subtracting the mean representation for $A$ ) can be written as: $\mathbf{M}_A = \mathbf{U}_A \Sigma \mathbf{V}_A^T$ . The right singular matrix $\mathbf{V}_A$ is considered to be the subspace for language $A$ . These subspaces only need to be computed once for each layer. Next, we look at when projections should be invoked. + +# 2.2 Random Projection + +For a given target language, during finetuning using task-specific data in the source language, we project the source representations onto the target language subspace with a predetermined probability $p$ . This projection is invoked right before passing the representation through the task adapter, having already passed through the language adapter. To project onto a target subspace, we first shift the target language subspace so that it passes through the source language mean embedding and then take the projection onto the target subspace (Chang et al., 2022). Let $S$ be the source language and $Q$ be the target language. Let subspaces and means of representations from one of the Transformer layers + +for the source language be $\mathbf{V}_S$ and $\pmb{\mu}_{S}$ , respectively. Projection of a representation $\mathbf{x}$ on $S$ is given by: + +$$ +\mathrm {P r o j e c t} _ {S} (\mathbf {x}) = \mathbf {V} _ {S} \mathbf {V} _ {S} ^ {T} (\mathbf {x} - \boldsymbol {\mu} _ {S}) + \boldsymbol {\mu} _ {S} +$$ + +The projection of $\mathbf{x}$ onto the target language subspace, that is shifted onto the source subspace, can be computed as: + +$$ +\mathrm {P r o j e c t} _ {Q, \pmb {\mu} _ {S}} (\mathbf {x}) = \mathbf {V} _ {Q} \mathbf {V} _ {Q} ^ {T} (\mathbf {x} - \pmb {\mu} _ {S}) + \pmb {\mu} _ {S} +$$ + +The main intuition here is that by probabilistically projecting source representations onto the target language subspace during task-specific finetuning, the model can encode both source and target language information in its representations. The model cannot solely rely on source-language specific features during task-specific training. + +# 2.3 Mean Cosine Distance (MCD) + +We suggest another projection scheme, Mean Cosine Distance (MCD), that is more informed than randomly projecting source representations based on a projection probability $p$ . Using MCD, we project those embeddings that are deemed as being further away from the target language subspace compared to the source language subspace. This is quantified using a cosine distance between an embedding from a layer and means of source and target language subspaces. If an embedding is closer to the source language mean compared to the target language mean, we project it onto the target language subspace so as to make it more similar to target language embeddings. However, if an embedding is closer to the target language mean, we can possibly omit projection since it already contains information relevant to the target language. + +Consider a set of embeddings extracted from one of the Transformer layers. Let the means of all embeddings from this layer and the associated subspace be denoted by $\pmb{\mu}$ and $\mathbf{V}$ , respectively. $\pmb{\mu}_{S}$ and $\pmb{\mu}_{Q}$ denote the means for the source and target language, respectively. Similarly, $\mathbf{V}_{S}$ and $\mathbf{V}_{Q}$ refer to the respective subspaces. Let $\mathbf{x}$ denote a token embedding from the source language. The MCD policy can be written as: + +$$ +\mathbf {x} = \left\{ \begin{array}{l l} \operatorname {P r o j e c t} _ {Q, \boldsymbol {\mu} _ {S}} (\mathbf {x}) & \mathrm {i f} \operatorname {c} (\mathbf {x}, \boldsymbol {\mu} _ {Q}) < \operatorname {c} (\mathbf {x}, \boldsymbol {\mu} _ {S}) \\ \mathbf {x} & \mathrm {o t h e r w i s e} \end{array} \right. +$$ + +where $\operatorname{Project}_{Q, \mu_S}(\mathbf{x})$ is defined in Section 2.2 as the projection of $\mathbf{x}$ onto the target subspace $\mathbf{V}_Q$ + +![](images/08256902dcdae9869176721f35b7e68e6a6d47a6545b279356582f90fe9922d2.jpg) +Figure 1: A single Transformer layer as modified by the MAD-X setup and our projection scheme. During training, the output from the source language adapter is projected onto the target language subspace with probability $p$ for random projection (or, if deemed necessary, by the MCD scheme). Dotted arrows refer to the inference time pathway when representations pass through the target language adapter and no projection is applied. + +and $\mathrm{c}(\mathbf{x},\mathbf{y})$ refers to the cosine similarity between two embeddings $\mathbf{x}$ and $\mathbf{y}$ . + +Figure 1 provides an illustration of our proposed technique within a single Transformer layer that includes language and task adapters (as in the MAD-X framework). + +# 3 Experimental Setup + +Subspace construction. To construct language specific subspaces, we adopt the settings used by Chang et al. (2022). Text sequences of length 512 are taken from the OSCAR dataset (Ortiz Su'arez et al., 2019) and passed through XLMR (Conneau et al., 2020) to produce layer-wise contextualized embeddings. We pick 262K contextualized representations and subtract the representation mean before computing SVD. For a low-dimensional subspace, we select the greatest $k$ singular values such that their sum of squares is greater than or equal to $90\%$ of the total variance. (Total variance is given by the sum of the squared singular values produced.) Finally, in order to compute the language-specific subspaces, the corresponding right singular vectors are taken as the basis. + +
NER
hivideidisiloswmyjvavg
MAD-X Adapters68.366.875.949.476.274.074.852.757.366.1
Random Projection68.969.077.553.876.879.876.557.661.269.0
MCD68.568.177.154.776.176.975.453.659.367.7
+ +Table 1: NER results (F1 scores) for nine languages. + +
XQuAD
hivideavg
MAD-X Adapters68.171.471.870.4
Random Projection68.272.272.270.9
MCD68.672.973.571.7
+ +Datasets. We conduct cross-lingual transfer experiments on three tasks, Named Entity Recognition (NER), Question Answering (QA) and Natural Language Inference (NLI), where the source language is always English. For NER, we use the WikiANN dataset (Rahimi et al., 2019), and show results for nine languages Hindi, Vietnamese, German, Indonesian, Icelandic, Ilocano, Swahili, Burmese and Javanese with roughly 20K instances in the English train set and between 1K and 10K instances in the target dev and test sets. For QA, we use XQuAD (Artetxe et al., 2019), a multilingual extension of SQuAD (Rajpurkar et al., 2016) and we report results for Hindi, Vietnamese and German consisting of around 87K examples in the English SQuAD train set and 1190 instances in the three target dev sets. For NLI, we use the AmericasNLI dataset (Ebrahimmi et al., 2021) which is an extension of the XNLI dataset (Conneau et al., 2018) with low-resource American languages. We report results on Quechua and Guarani, consisting of 392k instances in the English train set and 2490 and 5010 instances in the dev and test sets, respectively for each target language. + +Training setup. We use transformer models from the adapter-transformers $^2$ fork of the HuggingFace transformers library (Wolf et al., 2020). We use pre-trained language adapters from AdapterHub (Pfeiffer et al., 2020a) for our transfer experiments. XQuAD and NLI fine-tuning experiments were conducted on a single NVIDIA A100 80 GB gpu for 15 epochs and 10 epochs, with learning rate 1e-4 and batch size 16. NER experiments were run + +Table 2: Results (F1) for QA for three languages + +
NLI
qugnavg
MAD-X Adapters48.236.042.1
Random Projection49.337.543.4
MCD48.137.842.9
+ +Table 3: Results (F1) for NLI for two languages + +for 30 epochs on Nvidia 1080 Ti with 12 GB ram. Further details can be found in Appendix A. + +# 4 Results + +NER, XQuAD and NLI results are shown in Table 1, Table 2 and Table 3 respectively. All values correspond to F1 scores averaged over 3 different seeds. We use the target language validation set to choose the best hyperparameter values for all experiments. Both MCD and random projections show consistent improvement over the MAD-X baseline numbers. With MCD, we explicitly instruct the model when to project. This removes a hyperparameter from the setup, compared to random projections, while maintaining consistent performance gains over the baseline. To further analyze MCD, + +![](images/d0403a355587a651bb4a059529d5c1231a98c2247c43be65ab97d8009572c6f2.jpg) +Figure 2: Projection fraction vs layer across epochs. + +we consider the fraction of embeddings being projected onto the target language subspace for NER. Table 4 shows the fraction of embeddings projected during training (averaged across all layers) for each language. For languages dissimilar to en (such as hi and id), it makes sense that the projection fractions + +Table 4: Projection percentages for NER. + +
hivideidis
Proj. Frac.0.650.570.570.630.55
+ +are high since the language subspace means are closer to the source language mean (Chang et al., 2022), compared to languages more similar to en like de and is. Figure 2 shows how projection fractions vary across layers averaged across training epochs. We see high projection rates in early and final layers across languages. This correlates with these layers encoding a lot of English-specific information (Rogers et al., 2020) via training on the task-specific English data, thus triggering projections via MCD often. + +# 5 Related Work + +Multilingual language models like mBERT (Devlin, 2018), XLM-R (Conneau et al., 2020) possess some zero-shot cross-lingual capabilities, even without any explicit finetuning on the languages of interest (Wu and Dredze, 2019; Pires et al., 2019). Such transfer without any finetuning could lead to degradation in performance across certain language pairs (Hu et al., 2020). Nevertheless, multilingual models are a good foundation to bootstrap and further develop cross-lingual generalization. While there is a rapidly growing body of work on cross-lingual transfer, very few approaches utilize language-specific subspaces for this purpose. Both Choenni and Shutova (2020) and Chang et al. (2022) construct language-specific subspaces in multilingual models for an exploratory analysis of the model's representations. Yang et al. (2021) use projections on language specific subspaces to remove language specific information from the representations. We note such removal of language bias did not perform well on cross-lingual transfer in our experiments. Parovic et al. (2022) train bilingual language adapters using both source and target language text before task adapter training. However, this requires training language adapters using both source and target language unlabelled text, for every language pair, in addition to training task adapters. In contrast, our setup is a simple architectural extension of MAD-X, requiring no additional training once the subspaces are computed for each language. To the best of our knowledge, ours is the first work to exploit language-specific subspaces for cross-lingual transfer. + +# 6 Conclusions + +In this work, we present a new adapter-based cross-lingual transfer technique for an apriori known set of target languages. We construct language subspaces using contextualized representations for source and target languages. Representations during task-specific training are projected onto the target subspace if they exceed a probability threshold or if they are closer to a mean source embedding. Both schemes consistently improve zero-shot transfer for three natural language understanding tasks across many languages. + +# Acknowledgements + +The first author (Ujan) was supported by the Uplink Internship Program of the India Chapter of ACM SIGKDD. The authors are thankful to the anonymous reviewers for their constructive suggestions that helped improve this submission. + +# Limitations + +While our proposed projection techniques often improve cross-lingual transfer, the choice of the projection layer and the projection probability in the case of random projection are hyperparameters that vary across tasks and languages. Our ongoing work involves identifying a mechanism via which we can parameterize these quantities, enabling the model to directly learn the optimal layer and probability values for projection. + +# References + +Mikel Artetxe, Sebastian Ruder, and Dani Yogatama. 2019. On the cross-lingual transferability of monolingual representations. CoRR, abs/1910.11856. +Tyler A. Chang, Zhuowen Tu, and Benjamin K. Bergen. 2022. The geometry of multilingual language model representations. arXiv:2205.10964. +Rochelle Choenni and Ekaterina Shutova. 2020. What does it mean to be language-agnostic? probing multilingual sentence encoders for typological properties. arXiv:2009.12862. +Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettle-moyer, and Veselin Stoyanov. 2020. Unsupervised cross-lingual representation learning at scale. arXiv:1911.02116. +Alexis Conneau, Rudy Rinott, Guillaume Lample, Adina Williams, Samuel R. Bowman, Holger Schwenk, + +and Veselin Stoyanov. 2018. Xnli: Evaluating crosslingual sentence representations. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics. +Jacob Devlin. 2018. Multilingual bert readme document. +Abteen Ebrahimi, Manuel Mager, Arturo Oncevay, Vishrav Chaudhary, Luis Chiruzzo, Angela Fan, John Ortega, Ricardo Ramos, Annette Rios, Ivan Vladimir, Gustavo A. Giménez-Lugo, Elisabeth Mager, Graham Neubig, Alexis Palmer, Rolando A. Coto Solano, Ngoc Thang Vu, and Katharina Kann. 2021. Americasnli: Evaluating zero-shot natural language understanding of pretrained multilingual models in truly low-resource languages. CoRR, abs/2104.08726. +Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin de Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. 2019. Parameter-efficient transfer learning for nlp. arXiv:1902.00751. +Junjie Hu, Sebastian Ruder, Aditya Siddhant, Graham Neubig, Orhan First, and Melvin Johnson. 2020. Xtreme: A massively multilingual multi-task benchmark for evaluating cross-lingual generalization. arXiv:2003.11080. +Quentin Lhoest, Albert Villanova del Moral, Yacine Jernite, Abhishek Thakur, Patrick von Platen, Suraj Patil, Julien Chaumond, Mariama Drame, Julien Plu, Lewis Tunstall, Joe Davison, Mario Šaško, Gunjan Chhablani, Bhavitvya Malik, Simon Brandeis, Teven Le Scao, Victor Sanh, Canwen Xu, and Nicolas Patry. 2020. Datasets: A community library for natural language processing. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 175-184, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics. +Pedro Javier Ortiz Su'arez, Benoit Sagot, and Laurent Romary. 2019. Asynchronous pipelines for processing huge corpora on medium to low resource infrastructures. Proceedings of the Workshop on Challenges in the Management of Large Corpora (CMLC-7) 2019. Cardiff, 22nd July 2019, pages 9 - 16, Mannheim. Leibniz-Institut f"ur Deutsche Sprache. +Marinela Parovic, Goran Glavaš, Ivan Vulić, and Anna Korhonen. 2022. Bad-x: Bilingual adapters improve zero-shot cross-lingual transfer. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1791–1799, Seattle, United States. Association for Computational Linguistics. +Jonas Pfeiffer, Andreas Rückle, Clifton Poth, Aishwarya Kamath, Ivan Vulic, Sebastian Ruder, Kyunghyun Cho, and Iryna Gurevych. 2020a. Adapterhub: A framework for adapting transformers. In Proceedings + +of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 46-54, Online. Association for Computational Linguistics. +Jonas Pfeiffer, Ivan Vulic, Iryna Gurevych, and Sebastian Ruder. 2020b. Mad-x: An adapter-based framework for multi-task cross-lingual transfer. arXiv:2005.00052. +Telmo Pires, Eva Schlinger, and Dan Garrette. 2019. How multilingual is multilingual bert? Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. +Afshin Rahimi, Yuan Li, and Trevor Cohn. 2019. Massively multilingual transfer for NER. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 151-164, Florence, Italy. Association for Computational Linguistics. +Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD: 100,000+ Questions for Machine Comprehension of Text. arXiv e-prints, page arXiv:1606.05250. +Anna Rogers, Olga Kovaleva, and Anna Rumshisky. 2020. A primer in bertology: What we know about how bert works. Transactions of the Association for Computational Linguistics, 8:842-866. +Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Anthony Moi Clement Delangue, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, and Sylvain Gugger. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online. Association for Computational Linguistics. +Shijie Wu and Mark Dredze. 2019. Beto, bentz, becas: The surprising cross-lingual effectiveness of bert. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). +Ziyi Yang, Yinfei Yang, Daniel Cer, and Eric Darve. 2021. A simple and effective method to eliminate the self language bias in multilingual representations. arXiv:2109.04727. + +# A Implementation Details + +We use the xlm-roberta-base model from HuggingFace Transformers (Wolf et al., 2020) pretrained on 2.5 TB of CommonCrawl data3, for all our experiments. NLI and XQuAD experiments were conducted on a single NVIDIA A100 GPU (80 GB + +Table 5: For random projection, best-performing projection layers for different languages obtained via a grid search on validation sets. + +
NERXQuADNLI
hivideidisswilojvmyhividequgn
Random Projections56848668801291
MCD10284060079291111
+ +Table 6: Probability values (as determined by tuning on validation sets) for the layers in Table 5. + +
NERXQuADNLI
hivideidisswilojvmyhividequgn
Random Projections0.10.30.30.90.50.50.30.50.50.50.70.50.10.1
+ +RAM) and the NER experiments ran on a single Nvidia 1080Ti GPU (12 GB RAM). We used a learning rate of 1e-4 with a batch size of 16. The hyperparameter choices for layers and probabilities for our experiments are given in Tables 5 and 6, respectively. + +All datasets used are taken from HuggingFace Datasets (Lhoest et al., 2020). For evaluating models, we use the HuggingFace Evaluate library as well as the seqval python package + +A For every submission: + +A1. Did you describe the limitations of your work? 7 +A2. Did you discuss any potential risks of your work? No potential risks +A3. Do the abstract and introduction summarize the paper's main claims? +A4. Have you used AI writing assistants when working on this paper? Left blank. + +B Did you use or create scientific artifacts? + +Left blank. + +B1. Did you cite the creators of artifacts you used? No response. +B2. Did you discuss the license or terms for use and / or distribution of any artifacts? No response. +B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? No response. +B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? No response. +B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? No response. +B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. No response. + +C Did you run computational experiments? + +3 + +C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? 3, appendix + +The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance. + +C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? + +3, appendix + +C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? + +4 + +C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? + +appendix + +D Did you use human annotators (e.g., crowdworkers) or research with human participants? + +Left blank. + +D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? + +No response. + +D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? + +No response. + +D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? + +No response. + +D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? + +No response. + +D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? + +No response. \ No newline at end of file diff --git a/2023/Zero-shot Cross-lingual Transfer With Learned Projections Using Unlabeled Target-Language Data/images.zip b/2023/Zero-shot Cross-lingual Transfer With Learned Projections Using Unlabeled Target-Language Data/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a0aff76f0a9efd51567e9a040fc5c6ab1fdfcd03 --- /dev/null +++ b/2023/Zero-shot Cross-lingual Transfer With Learned Projections Using Unlabeled Target-Language Data/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:83276bed1fbfdd6bcd45d9874f757d6346209ee8287520272227809e0958e89a +size 161777 diff --git a/2023/Zero-shot Cross-lingual Transfer With Learned Projections Using Unlabeled Target-Language Data/layout.json b/2023/Zero-shot Cross-lingual Transfer With Learned Projections Using Unlabeled Target-Language Data/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5880513b9b96b404435d22d3c740df040976984c --- /dev/null +++ b/2023/Zero-shot Cross-lingual Transfer With Learned Projections Using Unlabeled Target-Language Data/layout.json @@ -0,0 +1,6238 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 114, + 75, + 479, + 110 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 75, + 479, + 110 + ], + "spans": [ + { + "bbox": [ + 114, + 75, + 479, + 110 + ], + "type": "text", + "content": "Zero-shot Cross-lingual Transfer With Learned Projections Using Unlabeled Target-Language Data" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 121, + 131, + 176, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 131, + 176, + 145 + ], + "spans": [ + { + "bbox": [ + 121, + 131, + 176, + 145 + ], + "type": "text", + "content": "Ujan Deb*" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 124, + 146, + 173, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 146, + 173, + 157 + ], + "spans": [ + { + "bbox": [ + 124, + 146, + 173, + 157 + ], + "type": "text", + "content": "IIT Bhilai" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 159, + 213, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 159, + 213, + 172 + ], + "spans": [ + { + "bbox": [ + 85, + 159, + 213, + 172 + ], + "type": "text", + "content": "ujand@iitbhilai.ac.in" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 252, + 131, + 339, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 131, + 339, + 144 + ], + "spans": [ + { + "bbox": [ + 252, + 131, + 339, + 144 + ], + "type": "text", + "content": "Ridayesh Parab*" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 266, + 145, + 326, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 145, + 326, + 158 + ], + "spans": [ + { + "bbox": [ + 266, + 145, + 326, + 158 + ], + "type": "text", + "content": "IIT Bombay" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 242, + 159, + 351, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 159, + 351, + 172 + ], + "spans": [ + { + "bbox": [ + 242, + 159, + 351, + 172 + ], + "type": "text", + "content": "ridayesh@gmail.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 409, + 131, + 483, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 409, + 131, + 483, + 144 + ], + "spans": [ + { + "bbox": [ + 409, + 131, + 483, + 144 + ], + "type": "text", + "content": "Preethi Jyothi" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 415, + 145, + 477, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 145, + 477, + 158 + ], + "spans": [ + { + "bbox": [ + 415, + 145, + 477, + 158 + ], + "type": "text", + "content": "IIT Bombay" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 380, + 159, + 512, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 159, + 512, + 172 + ], + "spans": [ + { + "bbox": [ + 380, + 159, + 512, + 172 + ], + "type": "text", + "content": "pjyothi@cse.iitb.ac.in" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 155, + 212, + 202, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 212, + 202, + 225 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 202, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 84, + 239, + 274, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 239, + 274, + 550 + ], + "spans": [ + { + "bbox": [ + 84, + 239, + 274, + 550 + ], + "type": "text", + "content": "Adapters have emerged as a parameter-efficient Transformer-based framework for cross-lingual transfer by inserting lightweight language-specific modules (language adapters) and task-specific modules (task adapters) within pretrained multilingual models. Zero-shot transfer is enabled by pairing the language adapter in the target language with an appropriate task adapter in a source language. If our target languages are known apriori, we explore how zero-shot transfer can be further improved within the adapter framework by utilizing unlabeled text during task-specific finetuning. We construct language-specific subspaces using standard linear algebra constructs and selectively project source-language representations into the target language subspace during task-specific finetuning using two schemes. Our experiments on three cross-lingual tasks, Named Entity Recognition (NER), Question Answering (QA) and Natural Language Inference (NLI) yield consistent benefits compared to adapter baselines over a wide variety of target languages with up to " + }, + { + "bbox": [ + 84, + 239, + 274, + 550 + ], + "type": "inline_equation", + "content": "11\\%" + }, + { + "bbox": [ + 84, + 239, + 274, + 550 + ], + "type": "text", + "content": " relative improvement in NER, " + }, + { + "bbox": [ + 84, + 239, + 274, + 550 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 84, + 239, + 274, + 550 + ], + "type": "text", + "content": " relative improvement in QA and " + }, + { + "bbox": [ + 84, + 239, + 274, + 550 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 84, + 239, + 274, + 550 + ], + "type": "text", + "content": " relative improvement in NLI." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 68, + 564, + 154, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 564, + 154, + 576 + ], + "spans": [ + { + "bbox": [ + 68, + 564, + 154, + 576 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 67, + 587, + 291, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 587, + 291, + 709 + ], + "spans": [ + { + "bbox": [ + 67, + 587, + 291, + 709 + ], + "type": "text", + "content": "Zero-shot cross-lingual transfer refers to the transfer of task-specific knowledge from a (high-resource) source language to a (zero-resource) target language that has no labeled task-specific data for training. A popular paradigm for cross-lingual transfer learning is to finetune pretrained multilingual models using labeled task-specific data in the source language and directly evaluate these finetuned models on target language test sets." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 67, + 710, + 291, + 751 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 710, + 291, + 751 + ], + "spans": [ + { + "bbox": [ + 67, + 710, + 291, + 751 + ], + "type": "text", + "content": "A parameter-efficient alternative to full finetuning for cross-lingual transfer is MAD-X (Pfeiffer et al., 2020b), an adapter-based framework that" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 213, + 526, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 213, + 526, + 388 + ], + "spans": [ + { + "bbox": [ + 302, + 213, + 526, + 388 + ], + "type": "text", + "content": "scaffolds on multilingual pretrained models to combine task-specific and language-specific modules in a plug-and-play manner. Adapters (Houlsby et al., 2019) are feedforward layer blocks inserted within each Transformer layer to selectively learn task-specific and language-specific capabilities via task adapters and language adapters, respectively. Language adapters are trained using self-supervised objectives like masked language modeling (MLM) and task adapters are trained using task-specific objectives. To enable task transfer to a target language, the relevant language and task adapters are combined at test-time." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 391, + 526, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 391, + 526, + 512 + ], + "spans": [ + { + "bbox": [ + 302, + 391, + 526, + 512 + ], + "type": "text", + "content": "In the zero-shot setting, we assume access to unlabeled text in the target languages. In MAD-X, this text is only used to train target language adapters and not further used during finetuning. Given knowledge of which languages we want to target, can we make effective use of unlabeled text in the target languages even during task-specific finetuning? This is the main question we tackle in this work." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 302, + 516, + 526, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 516, + 526, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 516, + 526, + 773 + ], + "type": "text", + "content": "We propose a general adapter-based technique to inject target language bias into task-specific finetuning. Using the unlabeled text in each target language, we construct an affine subspace from contextualized representations for every Transformer layer in the multilingual model. These subspaces are defined using singular value decomposition (SVD) and only need to be computed once per target language. During task-specific finetuning using labeled data in the source language, we project the source representations onto the target language subspaces. This projection can be invoked randomly using a projection probability defined as a hyperparameter. Projections can also be triggered depending on whether the current source representations are closer to the mean embedding of the source language subspace compared to the mean embedding of the target language subspace. We investigate both these projection policies and find" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 81, + 761, + 154, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 761, + 154, + 772 + ], + "spans": [ + { + "bbox": [ + 81, + 761, + 154, + 772 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "text", + "content": "449" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "spans": [ + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "type": "text", + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 224, + 806, + 369, + 817 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 806, + 369, + 817 + ], + "spans": [ + { + "bbox": [ + 224, + 806, + 369, + 817 + ], + "type": "text", + "content": "Volume 2: Short Papers, pages 449-457" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "spans": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "text", + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 291, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 291, + 126 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 291, + 126 + ], + "type": "text", + "content": "that they both improve performance across multiple tasks in multiple languages compared to state-of-the-art adapter baselines. We also release code1 to reproduce our experiments." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 134, + 157, + 149 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 134, + 157, + 149 + ], + "spans": [ + { + "bbox": [ + 67, + 134, + 157, + 149 + ], + "type": "text", + "content": "2 Methodology" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 155, + 291, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 155, + 291, + 440 + ], + "spans": [ + { + "bbox": [ + 67, + 155, + 291, + 440 + ], + "type": "text", + "content": "Adapters and MAD-X. Adapters for language models (Houlsby et al., 2019) are bottleneck feedforward modules, typically inserted in each Transformer layer of a multilingual model before layer normalization. Instead of finetuning the entire model, only adapters are tuned for a specific task. Pfeiffer et al. (2020b) extended adapter-based fine tuning to support cross-lingual transfer. Their framework called MAD-X (Multiple Adapters for Cross-lingual transfer) comprises of language adapters and task adapters. Language adapters are pretrained using masked language modeling to learn language-specific features. Task adapters are stacked on top of language adapters during downstream task finetuning to learn task-specific information. To achieve zero-shot transfer, the model is trained with a frozen source-language language adapter and a task adapter. During test time, the source-language adapter is replaced with the target-language adapter and evaluated on test instances in the target language." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 447, + 291, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 447, + 291, + 542 + ], + "spans": [ + { + "bbox": [ + 67, + 447, + 291, + 542 + ], + "type": "text", + "content": "Overview of our technique. We are interested in the setting where we have apriori knowledge of which languages we want to target at test time. We aim to bias cross-lingual transfer towards known target languages during task-specific finetuning. We start with MAD-X as our underlying framework and adopt the following 3-step approach:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 542, + 291, + 745 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 67, + 542, + 290, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 542, + 290, + 596 + ], + "spans": [ + { + "bbox": [ + 67, + 542, + 290, + 596 + ], + "type": "text", + "content": "- We construct layer-specific subspaces for each of the target languages. This is done by computing SVD on contextualized token representations extracted from each layer. See §2.1 for more details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 597, + 291, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 597, + 291, + 717 + ], + "spans": [ + { + "bbox": [ + 67, + 597, + 291, + 717 + ], + "type": "text", + "content": "- During task-specific training, we selectively project output representations from the language adapter of a chosen layer onto the target language subspace. These projections are triggered based on two policies: Random projection (§2.2) and Mean Cosine Distance (§2.3). The projected representations are further passed through the task adapter that is trained using labeled data in the source language." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 718, + 290, + 745 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 718, + 290, + 745 + ], + "spans": [ + { + "bbox": [ + 67, + 718, + 290, + 745 + ], + "type": "text", + "content": "- Similar to MAD-X, we evaluate on the target language by simply swapping the source language" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 302, + 71, + 526, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 111 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 111 + ], + "type": "text", + "content": "adapter with the target language adapter while keeping the task adapter fixed. No projection is done during inference." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 121, + 503, + 133 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 121, + 503, + 133 + ], + "spans": [ + { + "bbox": [ + 302, + 121, + 503, + 133 + ], + "type": "text", + "content": "2.1 Language Subspaces and Projections" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 137, + 526, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 137, + 526, + 367 + ], + "spans": [ + { + "bbox": [ + 302, + 137, + 526, + 367 + ], + "type": "text", + "content": "Our objective is to bias the model towards the target language while fine-tuning for a task. For this, we need to extract language-specific information from model representations that jointly exhibit language-specific and language-independent properties. Language-specific subspaces have been typically used to analyze representations in multilingual language models. Choenni and Shutova (2020) showed that individual representations can be used to predict linguistic typological features after projecting onto language-sensitive subspaces. Chang et al. (2022) construct language subspaces with SVD using language-specific contextualized token embeddings. They analyze model performance and other properties after computing layerwise projections of representations to various language subspaces." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 368, + 526, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 368, + 526, + 558 + ], + "spans": [ + { + "bbox": [ + 302, + 368, + 526, + 558 + ], + "type": "text", + "content": "We construct subspaces for each of the target languages using SVD and contextualized token representations for unlabeled text in the target language. Consider a pretrained model like XLMR (Conneau et al., 2020) that takes text sequences from the target language as its input. " + }, + { + "bbox": [ + 302, + 368, + 526, + 558 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 302, + 368, + 526, + 558 + ], + "type": "text", + "content": "-dimensional embeddings from a particular layer for a given language " + }, + { + "bbox": [ + 302, + 368, + 526, + 558 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 302, + 368, + 526, + 558 + ], + "type": "text", + "content": " can be grouped into a matrix " + }, + { + "bbox": [ + 302, + 368, + 526, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_A \\in R^{n \\times d}" + }, + { + "bbox": [ + 302, + 368, + 526, + 558 + ], + "type": "text", + "content": ". SVD of " + }, + { + "bbox": [ + 302, + 368, + 526, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_A" + }, + { + "bbox": [ + 302, + 368, + 526, + 558 + ], + "type": "text", + "content": " (after subtracting the mean representation for " + }, + { + "bbox": [ + 302, + 368, + 526, + 558 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 302, + 368, + 526, + 558 + ], + "type": "text", + "content": ") can be written as: " + }, + { + "bbox": [ + 302, + 368, + 526, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_A = \\mathbf{U}_A \\Sigma \\mathbf{V}_A^T" + }, + { + "bbox": [ + 302, + 368, + 526, + 558 + ], + "type": "text", + "content": ". The right singular matrix " + }, + { + "bbox": [ + 302, + 368, + 526, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_A" + }, + { + "bbox": [ + 302, + 368, + 526, + 558 + ], + "type": "text", + "content": " is considered to be the subspace for language " + }, + { + "bbox": [ + 302, + 368, + 526, + 558 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 302, + 368, + 526, + 558 + ], + "type": "text", + "content": ". These subspaces only need to be computed once for each layer. Next, we look at when projections should be invoked." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 566, + 423, + 579 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 566, + 423, + 579 + ], + "spans": [ + { + "bbox": [ + 302, + 566, + 423, + 579 + ], + "type": "text", + "content": "2.2 Random Projection" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 584, + 526, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 584, + 526, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 584, + 526, + 773 + ], + "type": "text", + "content": "For a given target language, during finetuning using task-specific data in the source language, we project the source representations onto the target language subspace with a predetermined probability " + }, + { + "bbox": [ + 302, + 584, + 526, + 773 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 302, + 584, + 526, + 773 + ], + "type": "text", + "content": ". This projection is invoked right before passing the representation through the task adapter, having already passed through the language adapter. To project onto a target subspace, we first shift the target language subspace so that it passes through the source language mean embedding and then take the projection onto the target subspace (Chang et al., 2022). Let " + }, + { + "bbox": [ + 302, + 584, + 526, + 773 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 302, + 584, + 526, + 773 + ], + "type": "text", + "content": " be the source language and " + }, + { + "bbox": [ + 302, + 584, + 526, + 773 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 302, + 584, + 526, + 773 + ], + "type": "text", + "content": " be the target language. Let subspaces and means of representations from one of the Transformer layers" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 751, + 240, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 751, + 240, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 751, + 240, + 772 + ], + "type": "text", + "content": "1https://github.com/csalt-research/adapter-projections" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "450" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 291, + 98 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 291, + 98 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 291, + 98 + ], + "type": "text", + "content": "for the source language be " + }, + { + "bbox": [ + 67, + 71, + 291, + 98 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_S" + }, + { + "bbox": [ + 67, + 71, + 291, + 98 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 71, + 291, + 98 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_{S}" + }, + { + "bbox": [ + 67, + 71, + 291, + 98 + ], + "type": "text", + "content": ", respectively. Projection of a representation " + }, + { + "bbox": [ + 67, + 71, + 291, + 98 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 67, + 71, + 291, + 98 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 67, + 71, + 291, + 98 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 67, + 71, + 291, + 98 + ], + "type": "text", + "content": " is given by:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 91, + 109, + 267, + 126 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 109, + 267, + 126 + ], + "spans": [ + { + "bbox": [ + 91, + 109, + 267, + 126 + ], + "type": "interline_equation", + "content": "\\mathrm {P r o j e c t} _ {S} (\\mathbf {x}) = \\mathbf {V} _ {S} \\mathbf {V} _ {S} ^ {T} (\\mathbf {x} - \\boldsymbol {\\mu} _ {S}) + \\boldsymbol {\\mu} _ {S}", + "image_path": "aa35db0bc546ef72e33097954ec6acc6bb27dd3e5e18cde59c33c778075fed25.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 137, + 291, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 137, + 291, + 176 + ], + "spans": [ + { + "bbox": [ + 67, + 137, + 291, + 176 + ], + "type": "text", + "content": "The projection of " + }, + { + "bbox": [ + 67, + 137, + 291, + 176 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 67, + 137, + 291, + 176 + ], + "type": "text", + "content": " onto the target language subspace, that is shifted onto the source subspace, can be computed as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 82, + 187, + 274, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 187, + 274, + 205 + ], + "spans": [ + { + "bbox": [ + 82, + 187, + 274, + 205 + ], + "type": "interline_equation", + "content": "\\mathrm {P r o j e c t} _ {Q, \\pmb {\\mu} _ {S}} (\\mathbf {x}) = \\mathbf {V} _ {Q} \\mathbf {V} _ {Q} ^ {T} (\\mathbf {x} - \\pmb {\\mu} _ {S}) + \\pmb {\\mu} _ {S}", + "image_path": "2b209bca2e3d1e505425d9b4e0e06baed2a2904dba108a5933ead14e7a38fe9f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 215, + 292, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 215, + 292, + 309 + ], + "spans": [ + { + "bbox": [ + 67, + 215, + 292, + 309 + ], + "type": "text", + "content": "The main intuition here is that by probabilistically projecting source representations onto the target language subspace during task-specific finetuning, the model can encode both source and target language information in its representations. The model cannot solely rely on source-language specific features during task-specific training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 319, + 235, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 319, + 235, + 333 + ], + "spans": [ + { + "bbox": [ + 67, + 319, + 235, + 333 + ], + "type": "text", + "content": "2.3 Mean Cosine Distance (MCD)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 338, + 291, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 338, + 291, + 567 + ], + "spans": [ + { + "bbox": [ + 67, + 338, + 291, + 567 + ], + "type": "text", + "content": "We suggest another projection scheme, Mean Cosine Distance (MCD), that is more informed than randomly projecting source representations based on a projection probability " + }, + { + "bbox": [ + 67, + 338, + 291, + 567 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 67, + 338, + 291, + 567 + ], + "type": "text", + "content": ". Using MCD, we project those embeddings that are deemed as being further away from the target language subspace compared to the source language subspace. This is quantified using a cosine distance between an embedding from a layer and means of source and target language subspaces. If an embedding is closer to the source language mean compared to the target language mean, we project it onto the target language subspace so as to make it more similar to target language embeddings. However, if an embedding is closer to the target language mean, we can possibly omit projection since it already contains information relevant to the target language." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 568, + 290, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 568, + 290, + 690 + ], + "spans": [ + { + "bbox": [ + 67, + 568, + 290, + 690 + ], + "type": "text", + "content": "Consider a set of embeddings extracted from one of the Transformer layers. Let the means of all embeddings from this layer and the associated subspace be denoted by " + }, + { + "bbox": [ + 67, + 568, + 290, + 690 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}" + }, + { + "bbox": [ + 67, + 568, + 290, + 690 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 568, + 290, + 690 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 67, + 568, + 290, + 690 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 67, + 568, + 290, + 690 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_{S}" + }, + { + "bbox": [ + 67, + 568, + 290, + 690 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 568, + 290, + 690 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_{Q}" + }, + { + "bbox": [ + 67, + 568, + 290, + 690 + ], + "type": "text", + "content": " denote the means for the source and target language, respectively. Similarly, " + }, + { + "bbox": [ + 67, + 568, + 290, + 690 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{S}" + }, + { + "bbox": [ + 67, + 568, + 290, + 690 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 568, + 290, + 690 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{Q}" + }, + { + "bbox": [ + 67, + 568, + 290, + 690 + ], + "type": "text", + "content": " refer to the respective subspaces. Let " + }, + { + "bbox": [ + 67, + 568, + 290, + 690 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 67, + 568, + 290, + 690 + ], + "type": "text", + "content": " denote a token embedding from the source language. The MCD policy can be written as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 701, + 289, + 736 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 289, + 736 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 289, + 736 + ], + "type": "interline_equation", + "content": "\\mathbf {x} = \\left\\{ \\begin{array}{l l} \\operatorname {P r o j e c t} _ {Q, \\boldsymbol {\\mu} _ {S}} (\\mathbf {x}) & \\mathrm {i f} \\operatorname {c} (\\mathbf {x}, \\boldsymbol {\\mu} _ {Q}) < \\operatorname {c} (\\mathbf {x}, \\boldsymbol {\\mu} _ {S}) \\\\ \\mathbf {x} & \\mathrm {o t h e r w i s e} \\end{array} \\right.", + "image_path": "21821f901a76b43311aa266e3e423608a755852c759082d87b35e30f0d76b197.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 746, + 290, + 774 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 746, + 290, + 774 + ], + "spans": [ + { + "bbox": [ + 67, + 746, + 290, + 774 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 67, + 746, + 290, + 774 + ], + "type": "inline_equation", + "content": "\\operatorname{Project}_{Q, \\mu_S}(\\mathbf{x})" + }, + { + "bbox": [ + 67, + 746, + 290, + 774 + ], + "type": "text", + "content": " is defined in Section 2.2 as the projection of " + }, + { + "bbox": [ + 67, + 746, + 290, + 774 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 67, + 746, + 290, + 774 + ], + "type": "text", + "content": " onto the target subspace " + }, + { + "bbox": [ + 67, + 746, + 290, + 774 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_Q" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 357, + 69, + 473, + 306 + ], + "blocks": [ + { + "bbox": [ + 357, + 69, + 473, + 306 + ], + "lines": [ + { + "bbox": [ + 357, + 69, + 473, + 306 + ], + "spans": [ + { + "bbox": [ + 357, + 69, + 473, + 306 + ], + "type": "image", + "image_path": "08256902dcdae9869176721f35b7e68e6a6d47a6545b279356582f90fe9922d2.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 315, + 526, + 413 + ], + "lines": [ + { + "bbox": [ + 302, + 315, + 526, + 413 + ], + "spans": [ + { + "bbox": [ + 302, + 315, + 526, + 413 + ], + "type": "text", + "content": "Figure 1: A single Transformer layer as modified by the MAD-X setup and our projection scheme. During training, the output from the source language adapter is projected onto the target language subspace with probability " + }, + { + "bbox": [ + 302, + 315, + 526, + 413 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 302, + 315, + 526, + 413 + ], + "type": "text", + "content": " for random projection (or, if deemed necessary, by the MCD scheme). Dotted arrows refer to the inference time pathway when representations pass through the target language adapter and no projection is applied." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 435, + 525, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 435, + 525, + 462 + ], + "spans": [ + { + "bbox": [ + 302, + 435, + 525, + 462 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 302, + 435, + 525, + 462 + ], + "type": "inline_equation", + "content": "\\mathrm{c}(\\mathbf{x},\\mathbf{y})" + }, + { + "bbox": [ + 302, + 435, + 525, + 462 + ], + "type": "text", + "content": " refers to the cosine similarity between two embeddings " + }, + { + "bbox": [ + 302, + 435, + 525, + 462 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 302, + 435, + 525, + 462 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 302, + 435, + 525, + 462 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 302, + 435, + 525, + 462 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 464, + 525, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 464, + 525, + 516 + ], + "spans": [ + { + "bbox": [ + 302, + 464, + 525, + 516 + ], + "type": "text", + "content": "Figure 1 provides an illustration of our proposed technique within a single Transformer layer that includes language and task adapters (as in the MAD-X framework)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 303, + 532, + 427, + 546 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 532, + 427, + 546 + ], + "spans": [ + { + "bbox": [ + 303, + 532, + 427, + 546 + ], + "type": "text", + "content": "3 Experimental Setup" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 301, + 556, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 556, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 301, + 556, + 526, + 772 + ], + "type": "text", + "content": "Subspace construction. To construct language specific subspaces, we adopt the settings used by Chang et al. (2022). Text sequences of length 512 are taken from the OSCAR dataset (Ortiz Su'arez et al., 2019) and passed through XLMR (Conneau et al., 2020) to produce layer-wise contextualized embeddings. We pick 262K contextualized representations and subtract the representation mean before computing SVD. For a low-dimensional subspace, we select the greatest " + }, + { + "bbox": [ + 301, + 556, + 526, + 772 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 301, + 556, + 526, + 772 + ], + "type": "text", + "content": " singular values such that their sum of squares is greater than or equal to " + }, + { + "bbox": [ + 301, + 556, + 526, + 772 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 301, + 556, + 526, + 772 + ], + "type": "text", + "content": " of the total variance. (Total variance is given by the sum of the squared singular values produced.) Finally, in order to compute the language-specific subspaces, the corresponding right singular vectors are taken as the basis." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 305, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 305, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 305, + 791 + ], + "type": "text", + "content": "451" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 117, + 68, + 477, + 141 + ], + "blocks": [ + { + "bbox": [ + 117, + 68, + 477, + 141 + ], + "lines": [ + { + "bbox": [ + 117, + 68, + 477, + 141 + ], + "spans": [ + { + "bbox": [ + 117, + 68, + 477, + 141 + ], + "type": "table", + "html": "
NER
hivideidisiloswmyjvavg
MAD-X Adapters68.366.875.949.476.274.074.852.757.366.1
Random Projection68.969.077.553.876.879.876.557.661.269.0
MCD68.568.177.154.776.176.975.453.659.367.7
", + "image_path": "588e62121cef0fa180dafb91ff10ef55087494ce0010df4fad6e79131ec65357.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 86, + 172, + 280, + 243 + ], + "blocks": [ + { + "bbox": [ + 188, + 148, + 404, + 161 + ], + "lines": [ + { + "bbox": [ + 188, + 148, + 404, + 161 + ], + "spans": [ + { + "bbox": [ + 188, + 148, + 404, + 161 + ], + "type": "text", + "content": "Table 1: NER results (F1 scores) for nine languages." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 86, + 172, + 280, + 243 + ], + "lines": [ + { + "bbox": [ + 86, + 172, + 280, + 243 + ], + "spans": [ + { + "bbox": [ + 86, + 172, + 280, + 243 + ], + "type": "table", + "html": "
XQuAD
hivideavg
MAD-X Adapters68.171.471.870.4
Random Projection68.272.272.270.9
MCD68.672.973.571.7
", + "image_path": "56d820cf54253f16825943a3ce9a73e3afee00013c8886f77effbdab1ea18583.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 285, + 291, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 285, + 291, + 610 + ], + "spans": [ + { + "bbox": [ + 67, + 285, + 291, + 610 + ], + "type": "text", + "content": "Datasets. We conduct cross-lingual transfer experiments on three tasks, Named Entity Recognition (NER), Question Answering (QA) and Natural Language Inference (NLI), where the source language is always English. For NER, we use the WikiANN dataset (Rahimi et al., 2019), and show results for nine languages Hindi, Vietnamese, German, Indonesian, Icelandic, Ilocano, Swahili, Burmese and Javanese with roughly 20K instances in the English train set and between 1K and 10K instances in the target dev and test sets. For QA, we use XQuAD (Artetxe et al., 2019), a multilingual extension of SQuAD (Rajpurkar et al., 2016) and we report results for Hindi, Vietnamese and German consisting of around 87K examples in the English SQuAD train set and 1190 instances in the three target dev sets. For NLI, we use the AmericasNLI dataset (Ebrahimmi et al., 2021) which is an extension of the XNLI dataset (Conneau et al., 2018) with low-resource American languages. We report results on Quechua and Guarani, consisting of 392k instances in the English train set and 2490 and 5010 instances in the dev and test sets, respectively for each target language." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 619, + 290, + 741 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 619, + 290, + 741 + ], + "spans": [ + { + "bbox": [ + 67, + 619, + 290, + 741 + ], + "type": "text", + "content": "Training setup. We use transformer models from the adapter-transformers" + }, + { + "bbox": [ + 67, + 619, + 290, + 741 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 67, + 619, + 290, + 741 + ], + "type": "text", + "content": " fork of the HuggingFace transformers library (Wolf et al., 2020). We use pre-trained language adapters from AdapterHub (Pfeiffer et al., 2020a) for our transfer experiments. XQuAD and NLI fine-tuning experiments were conducted on a single NVIDIA A100 80 GB gpu for 15 epochs and 10 epochs, with learning rate 1e-4 and batch size 16. NER experiments were run" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 327, + 172, + 493, + 243 + ], + "blocks": [ + { + "bbox": [ + 83, + 252, + 283, + 264 + ], + "lines": [ + { + "bbox": [ + 83, + 252, + 283, + 264 + ], + "spans": [ + { + "bbox": [ + 83, + 252, + 283, + 264 + ], + "type": "text", + "content": "Table 2: Results (F1) for QA for three languages" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 327, + 172, + 493, + 243 + ], + "lines": [ + { + "bbox": [ + 327, + 172, + 493, + 243 + ], + "spans": [ + { + "bbox": [ + 327, + 172, + 493, + 243 + ], + "type": "table", + "html": "
NLI
qugnavg
MAD-X Adapters48.236.042.1
Random Projection49.337.543.4
MCD48.137.842.9
", + "image_path": "c30cd14ae80d52b4edb56b4b8a6afeb37a256301cfc4c67d87f6f30eec346e87.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 311, + 252, + 508, + 264 + ], + "lines": [ + { + "bbox": [ + 311, + 252, + 508, + 264 + ], + "spans": [ + { + "bbox": [ + 311, + 252, + 508, + 264 + ], + "type": "text", + "content": "Table 3: Results (F1) for NLI for two languages" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 302, + 285, + 525, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 285, + 525, + 312 + ], + "spans": [ + { + "bbox": [ + 302, + 285, + 525, + 312 + ], + "type": "text", + "content": "for 30 epochs on Nvidia 1080 Ti with 12 GB ram. Further details can be found in Appendix A." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 322, + 362, + 334 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 322, + 362, + 334 + ], + "spans": [ + { + "bbox": [ + 302, + 322, + 362, + 334 + ], + "type": "text", + "content": "4 Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 301, + 343, + 526, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 343, + 526, + 506 + ], + "spans": [ + { + "bbox": [ + 301, + 343, + 526, + 506 + ], + "type": "text", + "content": "NER, XQuAD and NLI results are shown in Table 1, Table 2 and Table 3 respectively. All values correspond to F1 scores averaged over 3 different seeds. We use the target language validation set to choose the best hyperparameter values for all experiments. Both MCD and random projections show consistent improvement over the MAD-X baseline numbers. With MCD, we explicitly instruct the model when to project. This removes a hyperparameter from the setup, compared to random projections, while maintaining consistent performance gains over the baseline. To further analyze MCD," + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 326, + 530, + 494, + 657 + ], + "blocks": [ + { + "bbox": [ + 326, + 530, + 494, + 657 + ], + "lines": [ + { + "bbox": [ + 326, + 530, + 494, + 657 + ], + "spans": [ + { + "bbox": [ + 326, + 530, + 494, + 657 + ], + "type": "image", + "image_path": "d0403a355587a651bb4a059529d5c1231a98c2247c43be65ab97d8009572c6f2.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 308, + 666, + 520, + 680 + ], + "lines": [ + { + "bbox": [ + 308, + 666, + 520, + 680 + ], + "spans": [ + { + "bbox": [ + 308, + 666, + 520, + 680 + ], + "type": "text", + "content": "Figure 2: Projection fraction vs layer across epochs." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 692, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 692, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 692, + 525, + 772 + ], + "type": "text", + "content": "we consider the fraction of embeddings being projected onto the target language subspace for NER. Table 4 shows the fraction of embeddings projected during training (averaged across all layers) for each language. For languages dissimilar to en (such as hi and id), it makes sense that the projection fractions" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 750, + 227, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 750, + 227, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 750, + 227, + 772 + ], + "type": "text", + "content": "2https://github.com/adapter-hub/adapter-transformers" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "452" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 80, + 90, + 279, + 122 + ], + "blocks": [ + { + "bbox": [ + 94, + 69, + 264, + 82 + ], + "lines": [ + { + "bbox": [ + 94, + 69, + 264, + 82 + ], + "spans": [ + { + "bbox": [ + 94, + 69, + 264, + 82 + ], + "type": "text", + "content": "Table 4: Projection percentages for NER." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 80, + 90, + 279, + 122 + ], + "lines": [ + { + "bbox": [ + 80, + 90, + 279, + 122 + ], + "spans": [ + { + "bbox": [ + 80, + 90, + 279, + 122 + ], + "type": "table", + "html": "
hivideidis
Proj. Frac.0.650.570.570.630.55
", + "image_path": "74fef88abfe1cfc790eecef614b348bec241d378ed65cccb42006edffa3e689a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 142, + 291, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 142, + 291, + 290 + ], + "spans": [ + { + "bbox": [ + 67, + 142, + 291, + 290 + ], + "type": "text", + "content": "are high since the language subspace means are closer to the source language mean (Chang et al., 2022), compared to languages more similar to en like de and is. Figure 2 shows how projection fractions vary across layers averaged across training epochs. We see high projection rates in early and final layers across languages. This correlates with these layers encoding a lot of English-specific information (Rogers et al., 2020) via training on the task-specific English data, thus triggering projections via MCD often." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 303, + 161, + 317 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 303, + 161, + 317 + ], + "spans": [ + { + "bbox": [ + 67, + 303, + 161, + 317 + ], + "type": "text", + "content": "5 Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 326, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 326, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 326, + 291, + 772 + ], + "type": "text", + "content": "Multilingual language models like mBERT (Devlin, 2018), XLM-R (Conneau et al., 2020) possess some zero-shot cross-lingual capabilities, even without any explicit finetuning on the languages of interest (Wu and Dredze, 2019; Pires et al., 2019). Such transfer without any finetuning could lead to degradation in performance across certain language pairs (Hu et al., 2020). Nevertheless, multilingual models are a good foundation to bootstrap and further develop cross-lingual generalization. While there is a rapidly growing body of work on cross-lingual transfer, very few approaches utilize language-specific subspaces for this purpose. Both Choenni and Shutova (2020) and Chang et al. (2022) construct language-specific subspaces in multilingual models for an exploratory analysis of the model's representations. Yang et al. (2021) use projections on language specific subspaces to remove language specific information from the representations. We note such removal of language bias did not perform well on cross-lingual transfer in our experiments. Parovic et al. (2022) train bilingual language adapters using both source and target language text before task adapter training. However, this requires training language adapters using both source and target language unlabelled text, for every language pair, in addition to training task adapters. In contrast, our setup is a simple architectural extension of MAD-X, requiring no additional training once the subspaces are computed for each language. To the best of our knowledge, ours is the first work to exploit language-specific subspaces for cross-lingual transfer." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 303, + 71, + 387, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 71, + 387, + 84 + ], + "spans": [ + { + "bbox": [ + 303, + 71, + 387, + 84 + ], + "type": "text", + "content": "6 Conclusions" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 92, + 527, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 92, + 527, + 242 + ], + "spans": [ + { + "bbox": [ + 302, + 92, + 527, + 242 + ], + "type": "text", + "content": "In this work, we present a new adapter-based cross-lingual transfer technique for an apriori known set of target languages. We construct language subspaces using contextualized representations for source and target languages. Representations during task-specific training are projected onto the target subspace if they exceed a probability threshold or if they are closer to a mean source embedding. Both schemes consistently improve zero-shot transfer for three natural language understanding tasks across many languages." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 303, + 252, + 406, + 265 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 252, + 406, + 265 + ], + "spans": [ + { + "bbox": [ + 303, + 252, + 406, + 265 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 273, + 527, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 273, + 527, + 341 + ], + "spans": [ + { + "bbox": [ + 302, + 273, + 527, + 341 + ], + "type": "text", + "content": "The first author (Ujan) was supported by the Uplink Internship Program of the India Chapter of ACM SIGKDD. The authors are thankful to the anonymous reviewers for their constructive suggestions that helped improve this submission." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 303, + 351, + 367, + 364 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 351, + 367, + 364 + ], + "spans": [ + { + "bbox": [ + 303, + 351, + 367, + 364 + ], + "type": "text", + "content": "Limitations" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 373, + 527, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 373, + 527, + 495 + ], + "spans": [ + { + "bbox": [ + 302, + 373, + 527, + 495 + ], + "type": "text", + "content": "While our proposed projection techniques often improve cross-lingual transfer, the choice of the projection layer and the projection probability in the case of random projection are hyperparameters that vary across tasks and languages. Our ongoing work involves identifying a mechanism via which we can parameterize these quantities, enabling the model to directly learn the optimal layer and probability values for projection." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 518, + 362, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 362, + 530 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 362, + 530 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 536, + 527, + 772 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 304, + 536, + 527, + 571 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 536, + 527, + 571 + ], + "spans": [ + { + "bbox": [ + 304, + 536, + 527, + 571 + ], + "type": "text", + "content": "Mikel Artetxe, Sebastian Ruder, and Dani Yogatama. 2019. On the cross-lingual transferability of monolingual representations. CoRR, abs/1910.11856." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 578, + 527, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 527, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 527, + 613 + ], + "type": "text", + "content": "Tyler A. Chang, Zhuowen Tu, and Benjamin K. Bergen. 2022. The geometry of multilingual language model representations. arXiv:2205.10964." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 621, + 527, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 621, + 527, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 621, + 527, + 665 + ], + "type": "text", + "content": "Rochelle Choenni and Ekaterina Shutova. 2020. What does it mean to be language-agnostic? probing multilingual sentence encoders for typological properties. arXiv:2009.12862." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 674, + 527, + 740 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 674, + 527, + 740 + ], + "spans": [ + { + "bbox": [ + 304, + 674, + 527, + 740 + ], + "type": "text", + "content": "Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettle-moyer, and Veselin Stoyanov. 2020. Unsupervised cross-lingual representation learning at scale. arXiv:1911.02116." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 750, + 527, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 750, + 527, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 750, + 527, + 772 + ], + "type": "text", + "content": "Alexis Conneau, Rudy Rinott, Guillaume Lample, Adina Williams, Samuel R. Bowman, Holger Schwenk," + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "453" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 80, + 72, + 291, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 72, + 291, + 127 + ], + "spans": [ + { + "bbox": [ + 80, + 72, + 291, + 127 + ], + "type": "text", + "content": "and Veselin Stoyanov. 2018. Xnli: Evaluating crosslingual sentence representations. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 137, + 290, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 137, + 290, + 158 + ], + "spans": [ + { + "bbox": [ + 69, + 137, + 290, + 158 + ], + "type": "text", + "content": "Jacob Devlin. 2018. Multilingual bert readme document." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 168, + 290, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 168, + 290, + 266 + ], + "spans": [ + { + "bbox": [ + 69, + 168, + 290, + 266 + ], + "type": "text", + "content": "Abteen Ebrahimi, Manuel Mager, Arturo Oncevay, Vishrav Chaudhary, Luis Chiruzzo, Angela Fan, John Ortega, Ricardo Ramos, Annette Rios, Ivan Vladimir, Gustavo A. Giménez-Lugo, Elisabeth Mager, Graham Neubig, Alexis Palmer, Rolando A. Coto Solano, Ngoc Thang Vu, and Katharina Kann. 2021. Americasnli: Evaluating zero-shot natural language understanding of pretrained multilingual models in truly low-resource languages. CoRR, abs/2104.08726." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 275, + 290, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 275, + 290, + 330 + ], + "spans": [ + { + "bbox": [ + 69, + 275, + 290, + 330 + ], + "type": "text", + "content": "Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin de Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. 2019. Parameter-efficient transfer learning for nlp. arXiv:1902.00751." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 340, + 290, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 340, + 290, + 394 + ], + "spans": [ + { + "bbox": [ + 69, + 340, + 290, + 394 + ], + "type": "text", + "content": "Junjie Hu, Sebastian Ruder, Aditya Siddhant, Graham Neubig, Orhan First, and Melvin Johnson. 2020. Xtreme: A massively multilingual multi-task benchmark for evaluating cross-lingual generalization. arXiv:2003.11080." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 404, + 290, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 404, + 290, + 536 + ], + "spans": [ + { + "bbox": [ + 69, + 404, + 290, + 536 + ], + "type": "text", + "content": "Quentin Lhoest, Albert Villanova del Moral, Yacine Jernite, Abhishek Thakur, Patrick von Platen, Suraj Patil, Julien Chaumond, Mariama Drame, Julien Plu, Lewis Tunstall, Joe Davison, Mario Šaško, Gunjan Chhablani, Bhavitvya Malik, Simon Brandeis, Teven Le Scao, Victor Sanh, Canwen Xu, and Nicolas Patry. 2020. Datasets: A community library for natural language processing. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 175-184, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 544, + 290, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 544, + 290, + 622 + ], + "spans": [ + { + "bbox": [ + 69, + 544, + 290, + 622 + ], + "type": "text", + "content": "Pedro Javier Ortiz Su'arez, Benoit Sagot, and Laurent Romary. 2019. Asynchronous pipelines for processing huge corpora on medium to low resource infrastructures. Proceedings of the Workshop on Challenges in the Management of Large Corpora (CMLC-7) 2019. Cardiff, 22nd July 2019, pages 9 - 16, Mannheim. Leibniz-Institut f\"ur Deutsche Sprache." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 630, + 290, + 719 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 630, + 290, + 719 + ], + "spans": [ + { + "bbox": [ + 69, + 630, + 290, + 719 + ], + "type": "text", + "content": "Marinela Parovic, Goran Glavaš, Ivan Vulić, and Anna Korhonen. 2022. Bad-x: Bilingual adapters improve zero-shot cross-lingual transfer. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1791–1799, Seattle, United States. Association for Computational Linguistics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 728, + 290, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 728, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 728, + 290, + 772 + ], + "type": "text", + "content": "Jonas Pfeiffer, Andreas Rückle, Clifton Poth, Aishwarya Kamath, Ivan Vulic, Sebastian Ruder, Kyunghyun Cho, and Iryna Gurevych. 2020a. Adapterhub: A framework for adapting transformers. In Proceedings" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 525, + 645 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 314, + 72, + 525, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 525, + 116 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 525, + 116 + ], + "type": "text", + "content": "of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 46-54, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 125, + 525, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 125, + 525, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 125, + 525, + 168 + ], + "type": "text", + "content": "Jonas Pfeiffer, Ivan Vulic, Iryna Gurevych, and Sebastian Ruder. 2020b. Mad-x: An adapter-based framework for multi-task cross-lingual transfer. arXiv:2005.00052." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 177, + 525, + 222 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 177, + 525, + 222 + ], + "spans": [ + { + "bbox": [ + 304, + 177, + 525, + 222 + ], + "type": "text", + "content": "Telmo Pires, Eva Schlinger, and Dan Garrette. 2019. How multilingual is multilingual bert? Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 230, + 525, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 230, + 525, + 285 + ], + "spans": [ + { + "bbox": [ + 304, + 230, + 525, + 285 + ], + "type": "text", + "content": "Afshin Rahimi, Yuan Li, and Trevor Cohn. 2019. Massively multilingual transfer for NER. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 151-164, Florence, Italy. Association for Computational Linguistics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 293, + 525, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 525, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 525, + 338 + ], + "type": "text", + "content": "Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD: 100,000+ Questions for Machine Comprehension of Text. arXiv e-prints, page arXiv:1606.05250." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 346, + 525, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 525, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 525, + 390 + ], + "type": "text", + "content": "Anna Rogers, Olga Kovaleva, and Anna Rumshisky. 2020. A primer in bertology: What we know about how bert works. Transactions of the Association for Computational Linguistics, 8:842-866." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 398, + 525, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 398, + 525, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 398, + 525, + 519 + ], + "type": "text", + "content": "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Anthony Moi Clement Delangue, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, and Sylvain Gugger. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 527, + 525, + 593 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 527, + 525, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 527, + 525, + 593 + ], + "type": "text", + "content": "Shijie Wu and Mark Dredze. 2019. Beto, bentz, becas: The surprising cross-lingual effectiveness of bert. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 601, + 525, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 601, + 525, + 645 + ], + "spans": [ + { + "bbox": [ + 304, + 601, + 525, + 645 + ], + "type": "text", + "content": "Ziyi Yang, Yinfei Yang, Daniel Cer, and Eric Darve. 2021. A simple and effective method to eliminate the self language bias in multilingual representations. arXiv:2109.04727." + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 666, + 446, + 679 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 446, + 679 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 446, + 679 + ], + "type": "text", + "content": "A Implementation Details" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 687, + 525, + 754 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 687, + 525, + 754 + ], + "spans": [ + { + "bbox": [ + 304, + 687, + 525, + 754 + ], + "type": "text", + "content": "We use the xlm-roberta-base model from HuggingFace Transformers (Wolf et al., 2020) pretrained on 2.5 TB of CommonCrawl data3, for all our experiments. NLI and XQuAD experiments were conducted on a single NVIDIA A100 GPU (80 GB" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 761, + 429, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 761, + 429, + 772 + ], + "spans": [ + { + "bbox": [ + 317, + 761, + 429, + 772 + ], + "type": "text", + "content": "3https://commoncrawl.org/" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 289, + 781, + 307, + 790 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 790 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 790 + ], + "type": "text", + "content": "454" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 100, + 102, + 493, + 161 + ], + "blocks": [ + { + "bbox": [ + 67, + 69, + 525, + 93 + ], + "lines": [ + { + "bbox": [ + 67, + 69, + 525, + 93 + ], + "spans": [ + { + "bbox": [ + 67, + 69, + 525, + 93 + ], + "type": "text", + "content": "Table 5: For random projection, best-performing projection layers for different languages obtained via a grid search on validation sets." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 100, + 102, + 493, + 161 + ], + "lines": [ + { + "bbox": [ + 100, + 102, + 493, + 161 + ], + "spans": [ + { + "bbox": [ + 100, + 102, + 493, + 161 + ], + "type": "table", + "html": "
NERXQuADNLI
hivideidisswilojvmyhividequgn
Random Projections56848668801291
MCD10284060079291111
", + "image_path": "c1546ba656b0badb67c5abe8437f7cf670c45e0ba31d7e3aed373fb1d8168e9f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 89, + 192, + 505, + 237 + ], + "blocks": [ + { + "bbox": [ + 101, + 170, + 490, + 183 + ], + "lines": [ + { + "bbox": [ + 101, + 170, + 490, + 183 + ], + "spans": [ + { + "bbox": [ + 101, + 170, + 490, + 183 + ], + "type": "text", + "content": "Table 6: Probability values (as determined by tuning on validation sets) for the layers in Table 5." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 89, + 192, + 505, + 237 + ], + "lines": [ + { + "bbox": [ + 89, + 192, + 505, + 237 + ], + "spans": [ + { + "bbox": [ + 89, + 192, + 505, + 237 + ], + "type": "table", + "html": "
NERXQuADNLI
hivideidisswilojvmyhividequgn
Random Projections0.10.30.30.90.50.50.30.50.50.50.70.50.10.1
", + "image_path": "ca7dfc5722eef60b418badf408860ecb3a472fc389bc64886c682dbee69006a9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 257, + 290, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 257, + 290, + 337 + ], + "spans": [ + { + "bbox": [ + 67, + 257, + 290, + 337 + ], + "type": "text", + "content": "RAM) and the NER experiments ran on a single Nvidia 1080Ti GPU (12 GB RAM). We used a learning rate of 1e-4 with a batch size of 16. The hyperparameter choices for layers and probabilities for our experiments are given in Tables 5 and 6, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 338, + 291, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 338, + 291, + 391 + ], + "spans": [ + { + "bbox": [ + 67, + 338, + 291, + 391 + ], + "type": "text", + "content": "All datasets used are taken from HuggingFace Datasets (Lhoest et al., 2020). For evaluating models, we use the HuggingFace Evaluate library as well as the seqval python package" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 80, + 749, + 275, + 761 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 749, + 275, + 761 + ], + "spans": [ + { + "bbox": [ + 80, + 749, + 275, + 761 + ], + "type": "text", + "content": "4https://huggingface.co/docs/evaluate/index" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 81, + 761, + 235, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 761, + 235, + 772 + ], + "spans": [ + { + "bbox": [ + 81, + 761, + 235, + 772 + ], + "type": "text", + "content": "5https://pypi.org/project/seqeval/" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "455" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "spans": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "type": "text", + "content": "A For every submission:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 106, + 414, + 242 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "spans": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "type": "text", + "content": "A1. Did you describe the limitations of your work? 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 143, + 329, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 143, + 329, + 169 + ], + "spans": [ + { + "bbox": [ + 77, + 143, + 329, + 169 + ], + "type": "text", + "content": "A2. Did you discuss any potential risks of your work? No potential risks" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 178, + 414, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 178, + 414, + 205 + ], + "spans": [ + { + "bbox": [ + 77, + 178, + 414, + 205 + ], + "type": "text", + "content": "A3. Do the abstract and introduction summarize the paper's main claims?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 77, + 215, + 398, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 215, + 398, + 242 + ], + "spans": [ + { + "bbox": [ + 77, + 215, + 398, + 242 + ], + "type": "text", + "content": "A4. Have you used AI writing assistants when working on this paper? Left blank." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 68, + 253, + 291, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 253, + 291, + 266 + ], + "spans": [ + { + "bbox": [ + 68, + 253, + 291, + 266 + ], + "type": "text", + "content": "B Did you use or create scientific artifacts?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 79, + 270, + 128, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 270, + 128, + 283 + ], + "spans": [ + { + "bbox": [ + 79, + 270, + 128, + 283 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 76, + 292, + 524, + 634 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 292, + 315, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 292, + 315, + 319 + ], + "spans": [ + { + "bbox": [ + 76, + 292, + 315, + 319 + ], + "type": "text", + "content": "B1. Did you cite the creators of artifacts you used? No response." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 76, + 328, + 463, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 328, + 463, + 355 + ], + "spans": [ + { + "bbox": [ + 76, + 328, + 463, + 355 + ], + "type": "text", + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? No response." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 76, + 364, + 524, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 364, + 524, + 432 + ], + "spans": [ + { + "bbox": [ + 76, + 364, + 524, + 432 + ], + "type": "text", + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? No response." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 76, + 441, + 524, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 441, + 524, + 494 + ], + "spans": [ + { + "bbox": [ + 76, + 441, + 524, + 494 + ], + "type": "text", + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? No response." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 76, + 504, + 524, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 504, + 524, + 544 + ], + "spans": [ + { + "bbox": [ + 76, + 504, + 524, + 544 + ], + "type": "text", + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? No response." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 76, + 554, + 524, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 554, + 524, + 634 + ], + "spans": [ + { + "bbox": [ + 76, + 554, + 524, + 634 + ], + "type": "text", + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. No response." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "spans": [ + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "type": "text", + "content": "C Did you run computational experiments?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 80, + 662, + 87, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 662, + 87, + 672 + ], + "spans": [ + { + "bbox": [ + 80, + 662, + 87, + 672 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 77, + 682, + 524, + 724 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 682, + 524, + 724 + ], + "spans": [ + { + "bbox": [ + 77, + 682, + 524, + 724 + ], + "type": "text", + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? 3, appendix" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "spans": [ + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "type": "text", + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "spans": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "text", + "content": "ACL 2023 Responsible NLP Checklist" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "456" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 77, + 70, + 523, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 70, + 523, + 97 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 523, + 97 + ], + "type": "text", + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 89, + 99, + 143, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 99, + 143, + 111 + ], + "spans": [ + { + "bbox": [ + 89, + 99, + 143, + 111 + ], + "type": "text", + "content": "3, appendix" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 120, + 525, + 160 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 120, + 525, + 160 + ], + "spans": [ + { + "bbox": [ + 77, + 120, + 525, + 160 + ], + "type": "text", + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 90, + 162, + 99, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 162, + 99, + 172 + ], + "spans": [ + { + "bbox": [ + 90, + 162, + 99, + 172 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 183, + 525, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 183, + 525, + 223 + ], + "spans": [ + { + "bbox": [ + 77, + 183, + 525, + 223 + ], + "type": "text", + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 225, + 133, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 225, + 133, + 238 + ], + "spans": [ + { + "bbox": [ + 89, + 225, + 133, + 238 + ], + "type": "text", + "content": "appendix" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 247, + 522, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 247, + 522, + 260 + ], + "spans": [ + { + "bbox": [ + 67, + 247, + 522, + 260 + ], + "type": "text", + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 79, + 264, + 127, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 264, + 127, + 277 + ], + "spans": [ + { + "bbox": [ + 79, + 264, + 127, + 277 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 76, + 286, + 525, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 286, + 525, + 313 + ], + "spans": [ + { + "bbox": [ + 76, + 286, + 525, + 313 + ], + "type": "text", + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 89, + 314, + 148, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 314, + 148, + 327 + ], + "spans": [ + { + "bbox": [ + 89, + 314, + 148, + 327 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 76, + 336, + 525, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 336, + 525, + 376 + ], + "spans": [ + { + "bbox": [ + 76, + 336, + 525, + 376 + ], + "type": "text", + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 89, + 378, + 148, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 378, + 148, + 391 + ], + "spans": [ + { + "bbox": [ + 89, + 378, + 148, + 391 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 76, + 400, + 525, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 400, + 525, + 439 + ], + "spans": [ + { + "bbox": [ + 76, + 400, + 525, + 439 + ], + "type": "text", + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 89, + 441, + 148, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 441, + 148, + 454 + ], + "spans": [ + { + "bbox": [ + 89, + 441, + 148, + 454 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 76, + 462, + 520, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 462, + 520, + 476 + ], + "spans": [ + { + "bbox": [ + 76, + 462, + 520, + 476 + ], + "type": "text", + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 89, + 476, + 148, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 476, + 148, + 490 + ], + "spans": [ + { + "bbox": [ + 89, + 476, + 148, + 490 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 76, + 498, + 524, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 498, + 524, + 524 + ], + "spans": [ + { + "bbox": [ + 76, + 498, + 524, + 524 + ], + "type": "text", + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 89, + 527, + 148, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 527, + 148, + 539 + ], + "spans": [ + { + "bbox": [ + 89, + 527, + 148, + 539 + ], + "type": "text", + "content": "No response." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "457" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/mOKB6_ A Multilingual Open Knowledge Base Completion Benchmark/70c57a7f-70d1-4b2e-a713-7ef674b2ac69_content_list.json b/2023/mOKB6_ A Multilingual Open Knowledge Base Completion Benchmark/70c57a7f-70d1-4b2e-a713-7ef674b2ac69_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e142c7f565ae983b358db409ad522c45ef380bd1 --- /dev/null +++ b/2023/mOKB6_ A Multilingual Open Knowledge Base Completion Benchmark/70c57a7f-70d1-4b2e-a713-7ef674b2ac69_content_list.json @@ -0,0 +1,1751 @@ +[ + { + "type": "text", + "text": "mOKB6: A Multilingual Open Knowledge Base Completion Benchmark", + "text_level": 1, + "bbox": [ + 124, + 89, + 872, + 111 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shubham Mittal $^{\\alpha \\dagger}$ Keshay Kolluru $^{\\beta \\dagger}$ Soumen Chakrabarti $^{\\gamma}$ Mausam $^{\\alpha}$", + "bbox": [ + 161, + 123, + 843, + 140 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{\\alpha}$ Indian Institute of Technology Delhi", + "bbox": [ + 344, + 142, + 657, + 156 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{\\beta}$ KnowDis AI, New Delhi", + "bbox": [ + 391, + 158, + 611, + 173 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$\\gamma$ Indian Institute of Technology Bombay", + "bbox": [ + 332, + 175, + 668, + 191 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "shubhamiitd18@gmail.com, keshav.kolluru@gmail.com", + "bbox": [ + 253, + 192, + 749, + 208 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "soumen@cse.iitb.ac.in, mausam@cse.iitd.ac.in", + "bbox": [ + 277, + 210, + 724, + 224 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 268, + 339, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Automated completion of open knowledge bases (Open KBs), which are constructed from triples of the form (subject phrase, relation phrase, object phrase), obtained via open information extraction (Open IE) system, are useful for discovering novel facts that may not be directly present in the text. However, research in Open KB completion (Open KBC) has so far been limited to resource-rich languages like English. Using the latest advances in multilingual Open IE, we construct the first multilingual Open KBC dataset, called mOKB6, containing facts from Wikipedia in six languages (including English). Improving the previous Open KB construction pipeline by doing multilingual coreference resolution and keeping only entity-linked triples, we create a dense Open KB. We experiment with several models for the task and observe a consistent benefit of combining languages with the help of shared embedding space as well as translations of facts. We also observe that current multilingual models struggle to remember facts seen in languages of different scripts.", + "bbox": [ + 141, + 300, + 460, + 640 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 114, + 659, + 258, + 674 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Open information extraction (Open IE) systems (Mausam, 2016) such as ReVerb (Etzioni et al., 2011) and OpenIE6 (Kolluru et al., 2020) can extract triples, or facts, of the form (subject phrase, relation phrase, object phrase), which can be denoted as $(s,r,o)$ , from text (e.g., Wikipedia articles) without using any pre-defined ontology. Open knowledge base (Open KB) is constructed using these Open IE triples where the subject phrases and object phrases are nodes and relation phrases are labels on edges connecting the nodes in the graph. Open knowledge base completion (Open KBC) is", + "bbox": [ + 112, + 687, + 489, + 881 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the task of discovering new links between nodes using the graph structure of the Open KB. Knowledge graph embedding (KGE) models are typically used for the Open KBC task, where they are asked to answer questions of the form $(s,r,?)$ and $(?,r,o)$ .", + "bbox": [ + 507, + 253, + 884, + 332 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Research in Open KBC has been restricted to English (Vashishth et al., 2018) due to lack of Open KBs in other languages. We aim to study multilingual Open KBC, with the motivation that the information available in high resource languages like English may help when inferring links in Open KBs that use low resource languages like Telugu. Moreover, intuitively, if all the information in different languages can be pooled together, then it may help the model learn better, and allow information flow across Open KBs in different languages.", + "bbox": [ + 507, + 335, + 884, + 512 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We design the first multilingual Open KB construction pipeline (shown in Figure 1) using a multilingual Open IE system, GEN2OIE (Kolluru et al., 2022). We find that coreference resolution is missing in existing Open KB construction (Gashteovski et al., 2019) but is important for increasing the coverage of facts (as described in Figure 4). We re-train a recent coref model (Dobrovolskii, 2021) using XLM-R (Conneau et al., 2020) as the underlying multilingual encoder and add it to our pipeline. For constructing a high quality test set, we use 988 manually verified facts in English. For extending to other languages, we automatically translate English facts. The dataset thus constructed, called mOKB6, contains 42K facts in six languages: English, Hindi, Telugu, Spanish, Portuguese, and Chinese.", + "bbox": [ + 507, + 514, + 884, + 771 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We report the first baselines for multilingual Open KBC task. We find that they are able to benefit from information in multiple languages when compared to using facts from a single language. Translations of Open KB facts also help the models. However, we notice that although the multilingual KGE models learn facts in a particular language, they struggle to remember the same fact, when queried in another language with different script.", + "bbox": [ + 507, + 774, + 884, + 919 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Major part of work done as students at IIT Delhi.", + "bbox": [ + 139, + 891, + 452, + 904 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Dataset and code released at github.com:dair-iitd/mokb6", + "bbox": [ + 137, + 904, + 485, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "201", + "bbox": [ + 485, + 927, + 512, + 940 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics", + "bbox": [ + 226, + 945, + 769, + 957 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Volume 2: Short Papers, pages 201-214", + "bbox": [ + 376, + 958, + 620, + 971 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "July 9-14, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 295, + 972, + 700, + 984 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 114, + 83, + 270, + 98 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multilingual Open KBC datasets are absent in literature to the best of our knowledge, although multiple English Open KBC datasets are available. OLPBench (Broscheit et al., 2020), derived from OPIEC (Gashteovski et al., 2019), is a large-scale Open KBC dataset that contains 30M triples and is constructed from English Wikipedia using MinIE system (Gashteovski et al., 2017). The evaluation data contains 10K triples randomly sampled from 1.25M linked triples. ReVerb45K (Vashisth et al., 2018) and ReVerb20K (Galarraga et al., 2014) are smaller Open KBC datasets constructed from Clueweb09 corpus $^2$ using ReVerb Open IE system (Fader et al., 2011). Both the datasets keep only those tuples in which both the subject phrase and object phrase link to a finite set of Freebase entities.", + "bbox": [ + 112, + 110, + 489, + 367 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multilingual Open IE (mOpenIE) systems like GEN2OIE (Kolluru et al., 2022) and Multi $^{2}$ OIE (Ro et al., 2020) enable extracting facts from multiple languages. We use the GEN2OIE model for constructing mOKB6 dataset as it is trained with language-specific facts transferred from English, while Multi $^{2}$ OIE relies on zero-shot transfer for languages other than English.", + "bbox": [ + 112, + 368, + 489, + 495 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Knowledge Graph Embedding (KGE) Models: Conventional KGE models like TransE (Bordes et al., 2013), ComplEx (Trouillon et al., 2016), ConvE (Dettmers et al., 2018), and TuckER (Balazevic et al., 2019) have been used for Open KBC task (Gupta et al., 2019; Broscheit et al., 2020; Chandrahas and Talukdar, 2021; Kocijan and Lukasiewicz, 2021). Given a triple $(s,r,o)$ , these models encode the subject phrase, relation phrase, and object phrase from free text, and pass the encodings to a triple-scoring function, which is optimized using binary cross entropy loss. ComplEx has also been used for multilingual closed KBC task (Chakrabarti et al., 2022).", + "bbox": [ + 112, + 507, + 489, + 732 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Pretrained language models like BERT (Devlin et al., 2019) have been used in KGE models for the KBC task (Lovelace and Rosé, 2022; Lv et al., 2022; Chandrahas and Talukdar, 2021; Kim et al., 2020). SimKGC (Wang et al., 2022) is the state of the art KGE model on closed KBC task. It computes the score of a triple $(s, r, o)$ as the cosine similarity of the embeddings of $(s; r)$ and $(o)$ , computed using two separate pretrained BERT models without any weight sharing.", + "bbox": [ + 112, + 733, + 489, + 892 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3 Dataset Curation", + "text_level": 1, + "bbox": [ + 509, + 83, + 692, + 98 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We aim to construct a dense multilingual Open KB that maximizes the information about a given real-world entity, which may be represented as multiple nodes across languages. Therefore, we consider those Wikipedia articles3 that are available in six languages: English, Hindi, Telugu, Spanish, Portuguese, and Chinese4. This will also help the model learn from facts in high resource language like English and answer queries in low resource language like Telugu. We work with 300 titles randomly sampled from the ones common among all six languages (found using MediaWiki-Langlinks (MediaWiki, 2021)). Thus, we extract facts from $6 \\times 300$ Wikipedia articles. We discuss the three stages of our pipeline below.", + "bbox": [ + 507, + 109, + 884, + 350 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Stage 1 We first process each Wikipedia article through a coreference resolution system. Although language-specific end-to-end neural coref models have been developed (Žabokrtský et al., 2022; Xia and Van Durme, 2021), multilingual models that work on all our languages of interest are absent in the literature. Therefore, we retrain wl-coref (Dobrovolskii, 2021) with XLM-R (Conneau et al., 2020) on the English training data (available in OntoNotes (Weischedel et al., 2013)) that can work zero-shot for other languages.", + "bbox": [ + 507, + 359, + 884, + 535 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Coref models detect and cluster mentions, but do not identify a canonical cluster name, which is needed for standardizing all the mentions in the cluster. To find cluster names, entity linking systems such as mGENRE (De Cao et al., 2022) or Wikipedia hyperlinks can be used. However, we found that they result in low recall, particularly for low resource languages. Thus, we employ a heuristic to find the cluster name and replace each of the coreferent mentions with it. The score for each mention is represented by a tuple, computed as: Score(mention phrase) = (#proper nouns, #nouns, #numerals, #adjectives, #pronouns, #verbs). The tuple is ordered according to the importance of each field (POS tags) for the cluster name, which is determined empirically. Two tuples are compared index-wise with higher priority given to lower indices to determine the best scoring mention that is chosen as the canonical name (Table 1).", + "bbox": [ + 507, + 537, + 884, + 843 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Stage 2 We use GEN2OIE to extract Open IE triples from the coreference resolved sentences.", + "bbox": [ + 507, + 851, + 882, + 882 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "3Wikidump of April 02, 2022", + "bbox": [ + 529, + 890, + 715, + 904 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "4languages are chosen to match availability of Gen2OIE", + "bbox": [ + 529, + 904, + 873, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "2http://www.lemurproject.org/clueweb09.php/", + "bbox": [ + 134, + 903, + 462, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "202", + "bbox": [ + 485, + 927, + 515, + 940 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d8cf19105aad3c532c70c6bfe2852b999abae3175fb081a4cd65e4a3f6442568.jpg", + "image_caption": [ + "Figure 1: Our three-staged multilingual Open KB construction pipeline for mOKB6. mCoref is multilingual coreference resolution system, having XLM-R (Conneau et al., 2020) encoder based wl-coref (Dobrovolskii, 2021), and mOpenIE is multilingual open information extraction system, consisting of GEN2OIE (Kolluru et al., 2022)." + ], + "image_footnote": [], + "bbox": [ + 115, + 82, + 884, + 134 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/9c8dc2061c47852e3421b2da94e3a16579cbafb6e40911a03b062af8932394db.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MentionsScoresCluster Name
Barack Obama(2,0,0,0,0,0)
Obama(1,0,0,0,0,0)Barack Obama
He(0,0,0,0,1,0)
", + "bbox": [ + 144, + 205, + 457, + 271 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Stage 3 Similar to Gashteovski et al. (2019), we apply various filters to remove noisy triples that have empty or very long arguments, or have less confidence than 0.3 (as assigned by GEN2OIE). We further only keep triples that have the article's title as either the subject phrase or object phrase, to avoid generic or specific triples, valid only in the particular context. Examples of contextual triples (Choi et al., 2021) are discussed in Appendix E. See Appendix A for further data curation details.", + "bbox": [ + 112, + 324, + 487, + 483 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "These automatically extracted triples form the train set of mOKB6. To form a high quality test set in six languages with limited access to experts in all languages, the test set is created in a semiautomatic way. We sample 1600 English triples from the train set (which are subsequently filtered) and manually remove noisy triples. We use inter-annotation agreement between two annotators to check if they both agree that the given triple is noisy or clean. With an agreement of $91\\%$ , we retain 988 English triples, which we automatically translate to the other five languages. As illustrated in Figure 2, to translate a triple, we convert it to a sentence after removing tags and use Google translate for translating the triple-converted sentence to the remaining five languages. We observed high quality of translated triples, with $88\\%$ satisfactory translations as determined by native-speakers of three languages on a set of 75 translated triples. To get the Open IE subject phrase, relation phrase and object phrase tags, we project the labels from the original English triple to the translated sentence using word alignments (Kolluru et al., 2022). Finally, we are left with 550 triples in each language after removing examples where some labels could", + "bbox": [ + 115, + 488, + 489, + 889 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "not be aligned. We use these $6 \\times 550$ triples as the test sets. The train and dev sets are created from the remaining triples in each language such that the dev set has 500 randomly sampled triples (Table 2).", + "bbox": [ + 507, + 208, + 884, + 274 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5ef6efdb8a343a6921e6766304a6e31bfc2b48d19ad192936565512a7cf2369b.jpg", + "image_caption": [ + "Figure 2: Method to translate Open IE triple using Google translate, and followed by label projection using word alignments (Kolluru et al., 2022)." + ], + "image_footnote": [], + "bbox": [ + 573, + 287, + 821, + 437 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We analyse the entity overlap across languages and find that on an average, a test entity (which is present in either the subject phrase or object phrase of a test tuple) is present 17.73 times in English, 0.94 times in Hindi, 0.47 times in Telugu, 2.33 times in Spanish, 1.69 times in Portuguese, and 1.45 times in Chinese train set.", + "bbox": [ + 507, + 513, + 882, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our construction pipeline improves over OPIEC in three ways: (1) we use a multilingual Open IE system, instead of an English-specific Open IE system like in OPIEC, enabling us to curate Open KBs in many languages, (2) we add a multilingual coreference resolution system in our pipeline, and (3) the English test triples are manually verified. Further, we manually evaluate and review the noise at each step of data curation in Section 4.", + "bbox": [ + 507, + 626, + 884, + 771 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/d615963f2978c0599795ca5812bb17b69c514af75cfdce81ff9c830d99989c63.jpg", + "table_caption": [ + "Table 1: Parts of speech tags are used to find the canonical name of the coreferent cluster of entity mentions." + ], + "table_footnote": [], + "table_body": "
EnHiTeEsPtZh
#entity2063746253972565153045037
#relation787021771907282326442325
#train2019527861992396635283420
", + "bbox": [ + 510, + 781, + 882, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 2: Statistics of individual Open KBs in mOKB6 in English (En), Hindi (Hi), Telugu (Te), Spanish (Es), Portuguese (Pt), and Chinese (Zh). The dev and test set for each Open KB contain 500 and 550 triples each.", + "bbox": [ + 507, + 857, + 882, + 914 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "5https://translate.google.co.in/", + "bbox": [ + 134, + 903, + 379, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "203", + "bbox": [ + 485, + 927, + 515, + 940 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4 Noise Evaluation", + "text_level": 1, + "bbox": [ + 112, + 84, + 295, + 98 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Curating an Open KB involves various stages and each stage induces its noise in the construction pipeline (Gashteovski et al., 2019). We manually evaluate the noise induced at each stage of our pipeline (Figure 1) and discuss the same in this section. We ask native speakers of four (out of six) languages - English, Hindi, Telugu, and Chinese to assess the output quality, or precision, of each stage as discussed below.", + "bbox": [ + 112, + 109, + 487, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the first stage, we assess the performance of the coreference resolution system over Wikipedia articles. We find a high precision of $95.5\\%$ in coref's mention clustering and $89.82\\%$ accuracy in finding canonical cluster name (using the heuristic illustrated in Table 1), computed over 40 randomly sampled coref clusters (10 in each language).", + "bbox": [ + 112, + 254, + 487, + 366 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For evaluating the Open IE system, GEN2OIE, in the second stage, we mark an extraction of a sentence as correct if it has syntactically correct arguments and it is coherent with the sentence. We get an average precision of $63.4\\%$ on 80 extractions (20 in each language).", + "bbox": [ + 112, + 367, + 487, + 462 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We evaluate the triples, or Open KB facts, at the last stage after passing through various noise-removing filters. Note that these triples also form the train set (and dev set) in mOKB6 dataset. We mark triples as correct when they contain real-world entities, and also, factual information about them. If the triple is very generic or contextual (see Appendix E), it is marked as incorrect. We find the train (and dev) set quality to be $69.3\\%$ , averaged over 80 triples in four languages.", + "bbox": [ + 112, + 464, + 487, + 624 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 112, + 636, + 260, + 653 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our experimental study on multilingual open KBC task investigates the following research questions:", + "bbox": [ + 112, + 661, + 485, + 693 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Does the KGE model benefit from facts in different languages? (Section 5.1)", + "2. Can translation help transfer among languages? (Section 5.2)", + "3. Does the KGE model remember facts seen across different languages? (Section 5.3)" + ], + "bbox": [ + 127, + 694, + 487, + 788 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use SimKGC model (Wang et al., 2022) with pretrained mBERT initialization to run our experiments, after comparing with recent KGE models (Appendix C). For evaluation, we use three metrics -hits at rank 1 (H@1), hits at rank 10 (H@10), and mean reciprocal rank (MRR). The formal definitions of them are provided in Appendix B. We discuss further model training details in Appendix D.", + "bbox": [ + 110, + 790, + 487, + 917 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.1 Training on Multilingual Facts", + "text_level": 1, + "bbox": [ + 507, + 84, + 796, + 99 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We train and compare monolingual model, called MONO, with multilingual models, UNION and UNION w/o En. In MONO, we train one model for each language using its respective Open KB, whereas in UNION, a single model is trained on six languages' Open KBs together. UNION outperforms MONO in all languages by an average of $4.6\\%$ H@10 and $2.8\\%$ MRR (see Table 3), which provides evidence of information flow across languages and the model benefits from it.", + "bbox": [ + 507, + 112, + 882, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To check the extent of flow from (high-resource) English to the other languages, we also train on the five languages except English, which we call UNION w/o En. We find UNION w/o En also outperforms MONO by $2.7\\%$ H@10 and $1.2\\%$ MRR over the five languages, hinting that interlingual transfer is more general and pervasive.", + "bbox": [ + 507, + 277, + 882, + 390 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.2 Open KB Facts Translation", + "text_level": 1, + "bbox": [ + 507, + 414, + 769, + 429 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Apart from relying only on multilingual transfer in the embedding space, we analyse the effect of using translated triples in the training of the KGE model. We translate the English training triples to the other five languages (Section 3) and train monolingual models using only the translated triples (TRANS). To leverage facts present in each language's Open KB, we make MONO+TRANS, where we add language-specific MONO data to the translated triples. Table 3 shows that MONO+TRANS is better than MONO by a large margin of $15.5\\%$ H@1, $29.2\\%$ H@10, and $20.0\\%$ MRR, averaged over five languages. Also, MONO+TRANS improves over TRANS by $2.1\\%$ H@10 and $2.0\\%$ MRR, showcasing the importance of facts in each language's Open KBs.", + "bbox": [ + 507, + 442, + 882, + 699 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To effectively gain from transfer in both the embedding space as well as translation, we introduce $\\text{UNION+TRANS}$ . We train one model for each language, on the combination of UNION triples and the translated train triples from English Open KB to that language. $\\text{UNION+TRANS}$ is better than UNION by $25.9\\%$ H@10 and $18.4\\%$ MRR. This suggests that the model is able to benefit from English facts when they are translated to the query language, unlike in UNION where the English facts are present only in English.", + "bbox": [ + 507, + 703, + 882, + 879 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "6English source achieved the best translation quality.", + "bbox": [ + 529, + 904, + 852, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "204", + "bbox": [ + 485, + 927, + 515, + 940 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/d3f48cb6c2a32bfdaa827b2e60cf500e66de206dbf122e4de14514f84dfa3ec1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
English (En)Hindi (Hi)Telugu (Te)Spanish (Es)Portuguese (Pt)Chinese (Zh)
H@1H@10MRRH@1H@10MRRH@1H@10MRRH@1H@10MRRH@1H@10MRRH@1H@10MRR
MONO14.838.722.83.014.87.21.58.13.96.423.712.36.321.711.42.413.16.2
UNION w/o En5.721.510.92.915.47.41.810.24.98.127.814.56.726.112.93.215.57.5
UNION16.740.824.83.616.68.11.59.34.510.632.217.69.729.316.64.018.88.9
TRANS---20.547.629.78.728.715.523.250.632.420.550.730.514.039.422.5
MONO+TRANS---20.245.428.414.338.522.223.551.532.921.448.930.717.943.226.6
UNION+TRANS---23.349.732.315.138.523.123.952.433.423.552.133.116.943.626.0
", + "bbox": [ + 119, + 80, + 885, + 197 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 3: Performance (%) of SimKGC model on mOKB6 dataset, comprising of Open KBs in six languages. MONO, TRANS, and MONO+TRANS are monolingual models trained only on facts of one language whereas UNION, UNION w/o En, and UNION+TRANS are multilingual models trained with facts from multiple languages. All reported numbers are an average of three runs using different seeds. Best scores are highlighted in bold.", + "bbox": [ + 112, + 205, + 885, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.3 Cross-lingual Memorization", + "text_level": 1, + "bbox": [ + 112, + 288, + 381, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pretrained multilingual language models such as mBERT have demonstrated strong cross-lingual transfer capabilities (Wu and Dredze, 2019). We investigate cross-lingual memorization of the KGE model by showing facts in one language and querying the same facts in other five languages. For each language, $L$ , we take the UNION model and train it further on the test set of that language's Open KB, which we call MEMORIZE $_L$ model. Then, we test each MEMORIZE $_L$ model on the six test sets. Since the test sets (in mOKB6 dataset) of the different languages contain the same facts, this experiment allows us to investigate cross-lingual memorization. We provide the H@10 scores of MEMORIZE models in Figure 3 and the performance on other metrics (H@1 and MRR) is reported in Table 7.", + "bbox": [ + 112, + 311, + 489, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The model achieves at least $97\\%$ H@10 when tested on the language used for training (diagonal). We observe that there is relatively good crosslingual memorization among languages that share the same script (Latin in English, Spanish, and Portuguese), but the model struggles to remember facts when seen in languages of different scripts. Many entities look similar in shared scripts, possibly leading to better information transfer. For example, the $\\mathsf{MEMORIZE}_{En}$ achieves H@10 of $50.7\\%$ in Spanish (Es) compared to $22.3\\%$ in Chinese (Zh) and $11\\%$ in Telugu (Te).", + "bbox": [ + 112, + 569, + 489, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6 Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 112, + 778, + 401, + 793 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We create and release the mOKB6 dataset, the first multilingual Open Knowledge Base Completion dataset with 42K facts in six languages: English, Hindi, Telugu, Spanish, Portuguese, and Chinese. Its construction uses multilingual coreference resolution, entity-mention cluster naming, multilingual open information extraction and various filtering", + "bbox": [ + 112, + 806, + 489, + 919 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a1dc8f82f5d3ed3389f4fadaa56129d4f2c29635a09c9311958a3e04644f4744.jpg", + "image_caption": [ + "Figure 3: Performance (H@10) of MEMORIZE models. Row $L$ shows the performance of $\\text{MEMORIZE}_L$ model across the test sets of all languages (columns). For example, the performance of $\\text{MEMORIZE}_{En}$ when tested on English (En) is $97.1\\%$ H@10, and $\\text{MEMORIZE}_{En}$ when tested on Spanish (Es) gives $50.7\\%$ H@10. We find relatively good cross-lingual transfer among languages that use same script (Latin in English, Spanish and Portuguese) compared to those using different scripts (English, Hindi, Telugu and Chinese)." + ], + "image_footnote": [], + "bbox": [ + 547, + 304, + 830, + 464 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "steps to improve the quality of the extracted facts. We also report the first baselines on the task using the existing state of the art KGE models trained with facts from different languages using various augmentation strategies.", + "bbox": [ + 507, + 653, + 884, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our work opens many important research questions: (1) Can we develop better strategies to combine facts in different languages? (2) Can we build models that achieve strong information transfer across unrelated languages with same or different scripts? (3) Can we train the neural model to ignore contextual triples (Appendix E), thus improving overall performance? and (4) Can tying the same entities across various languages help the model generalize better? We leave these questions to be addressed in future work.", + "bbox": [ + 507, + 741, + 885, + 917 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "205", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "7 Acknowledgements", + "text_level": 1, + "bbox": [ + 114, + 84, + 315, + 99 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Keshav was supported by TCS Research Fellowship during his PhD. Mausam is supported by grants from Huawei, Google, Verisk and IBM, and a Jai Gupta Chair Fellowship. He also acknowledges Google and Yardi School of AI travel grants. Soumen is partly supported by a Jagadish Bose Fellowship and a grant from Cisco. We thank IIT Delhi HPC facility for compute resources.", + "bbox": [ + 112, + 109, + 489, + 237 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "8 Limitations", + "text_level": 1, + "bbox": [ + 112, + 249, + 250, + 265 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Although multilingual, the constructed open KB is limited to the sampling of the chosen six languages. We do not know how well the system will generalize to various language families that have not been considered here. Further, even among the languages considered, the performance of even the best-performing systems, as measured through $\\mathrm{H@1}$ is still in the low 20's. Therefore the models are not yet ready to be deployed for real-world applications.", + "bbox": [ + 112, + 274, + 489, + 435 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 114, + 462, + 213, + 476 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Ivana Balazevic, Carl Allen, and Timothy Hospedales. 2019. TuckER: Tensor factorization for knowledge graph completion. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5185-5194, Hong Kong, China. Association for Computational Linguistics.", + "Antoine Bordes, Nicolas Usunier, Alberto Garcia-Duran, Jason Weston, and Oksana Yakhnenko. 2013. Translating embeddings for modeling multi-relational data. In Advances in Neural Information Processing Systems, volume 26. Curran Associates, Inc.", + "Samuel Broscheit, Kiril Gashteovski, Yanjie Wang, and Rainer Gemulla. 2020. Can we predict new facts with open knowledge graph embeddings? a benchmark for open link prediction. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2296-2308, Online. Association for Computational Linguistics.", + "Soumen Chakrabarti, Harkanwar Singh, Shubham Lohiya, Prachi Jain, and Mausam. 2022. Joint completion and alignment of multilingual knowledge graphs. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 11922-11938, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.", + ". Chandrahas and Partha Talukdar. 2021. OKGIT: Open knowledge graph link prediction with implicit" + ], + "bbox": [ + 115, + 483, + 489, + 917 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "types. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 2546-2559, Online. Association for Computational Linguistics.", + "Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. 2014. Learning phrase representations using RNN encoder-decoder for statistical machine translation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1724-1734, Doha, Qatar. Association for Computational Linguistics.", + "Eunsol Choi, Jennimaria Palomaki, Matthew Lamm, Tom Kwiatkowski, Dipanjan Das, and Michael Collins. 2021. Decontextualization: Making sentences stand-alone. Transactions of the Association for Computational Linguistics, 9:447-461.", + "Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzman, Edouard Grave, Myle Ott, Luke Zettle-moyer, and Veselin Stoyanov. 2020. Unsupervised cross-lingual representation learning at scale. In ACL Conference, pages 8440-8451.", + "Nicola De Cao, Ledell Wu, Kashyap Popat, Mikel Artetxe, Naman Goyal, Mikhail Plekhanov, Luke Zettlemoyer, Nicola Cancedda, Sebastian Riedel, and Fabio Petroni. 2022. Multilingual autoregressive entity linking. Transactions of the Association for Computational Linguistics, 10:274-290.", + "Tim Dettmers, Pasquale Minervini, Pontus Stenetorp, and Sebastian Riedel. 2018. Convolutional 2d knowledge graph embeddings. In Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence and Thirtieth Innovative Applications of Artificial Intelligence Conference and Eighth AAAI Symposium on Educational Advances in Artificial Intelligence, AAAI'18/IAAI'18/EAAI'18. AAAI Press.", + "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics.", + "Vladimir Dobrovolskii. 2021. Word-level coreference resolution. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7670-7675, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.", + "Oren Etzioni, Anthony Fader, Janara Christensen, Stephen Soderland, and Mausam. 2011. Open information extraction: The second generation. In *IJCAI*" + ], + "bbox": [ + 510, + 85, + 884, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "206", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "2011, Proceedings of the 22nd International Joint Conference on Artificial Intelligence, Barcelona, Catalonia, Spain, July 16-22, 2011, pages 3-10. IJ-CAI/AAAI.", + "Anthony Fader, Stephen Soderland, and Oren Etzioni. 2011. Identifying relations for open information extraction. In Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 1535-1545, Edinburgh, Scotland, UK. Association for Computational Linguistics.", + "Luis Galárraga, Geremy Heitz, Kevin Murphy, and Fabian M. Suchanek. 2014. Canonicalizing open knowledge bases. New York, NY, USA. Association for Computing Machinery.", + "Kiril Gashteovski, Rainer Gemulla, and Luciano del Corro. 2017. MinIE: Minimizing facts in open information extraction. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2630-2640, Copenhagen, Denmark. Association for Computational Linguistics.", + "Kiril Gashteovski, Sebastian Wanner, Sven Hertling, Samuel Broscheit, and Rainer Gemulla. 2019. Opiec: An open information extraction corpus. In Proceedings of the Conference on Automatic Knowledge Base Construction (AKBC).", + "Swapnil Gupta, Sreyash Kenkre, and Partha Talukdar. 2019. CaRe: Open knowledge graph embeddings. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 378-388, Hong Kong, China. Association for Computational Linguistics.", + "Bosung Kim, Taesuk Hong, Youngjoong Ko, and Jungyun Seo. 2020. Multi-task learning for knowledge graph completion with pre-trained language models. In Proceedings of the 28th International Conference on Computational Linguistics, pages 1737-1743, Barcelona, Spain (Online). International Committee on Computational Linguistics.", + "Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings.", + "Vid Kocijan and Thomas Lukasiewicz. 2021. Knowledge base completion meets transfer learning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP), Punta Cana, Dominican Republic. Association for Computational Linguistics.", + "Keshav Kolluru, Vaibhav Adlakha, Samarth Aggarwal, Mausam, and Soumen Chakrabarti. 2020. OpenIE6: Iterative Grid Labeling and Coordination Analysis for Open Information Extraction. In Proceedings of" + ], + "bbox": [ + 115, + 85, + 489, + 917 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3748-3761, Online. Association for Computational Linguistics.", + "Keshav Kolluru, Muqeeth Mohammed, Shubham Mittal, Soumen Chakrabarti, and Mausam. 2022. Alignment-augmented consistent translation for multilingual open information extraction. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2502-2517, Dublin, Ireland. Association for Computational Linguistics.", + "Justin Lovelace and Carolyn Rosé. 2022. A framework for adapting pre-trained language models to knowledge graph completion. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 5937-5955, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.", + "Xin Lv, Yankai Lin, Yixin Cao, Lei Hou, Juanzi Li, Zhiyuan Liu, Peng Li, and Jie Zhou. 2022. Do pre-trained models benefit knowledge graph completion? a reliable evaluation and a reasonable approach. In Findings of the Association for Computational Linguistics: ACL 2022, pages 3570-3581, Dublin, Ireland. Association for Computational Linguistics.", + "Mausam. 2016. Open information extraction systems and downstream applications. In International Joint Conference on Artificial Intelligence.", + "MediaWiki. 2021. Api:langlinks — mediawiki. [Online; accessed 02-April-2022].", + "Jeffrey Pennington, Richard Socher, and Christopher Manning. 2014. GloVe: Global vectors for word representation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1532-1543, Doha, Qatar. Association for Computational Linguistics.", + "Peng Qi, Yuhao Zhang, Yuhui Zhang, Jason Bolton, and Christopher D. Manning. 2020. Stanza: A Python natural language processing toolkit for many human languages. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations.", + "Youngbin Ro, Yukyung Lee, and Pilsung Kang. 2020. Multi^2OIE: Multilingual open information extraction based on multi-head attention with BERT. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 1107-1117, Online. Association for Computational Linguistics.", + "Théo Trouillon, Johannes Welbl, Sebastian Riedel, Eric Gaussier, and Guillaume Bouchard. 2016. Complex embeddings for simple link prediction. In Proceedings of The 33rd International Conference on Machine Learning, volume 48 of Proceedings of Machine Learning Research, pages 2071-2080, New York, New York, USA. PMLR." + ], + "bbox": [ + 510, + 85, + 882, + 917 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "207", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Shikhar Vashishth, Prince Jain, and Partha Talukdar. 2018. CESI: Canonicalizing open knowledge bases using embeddings and side information. In Proceedings of the 2018 World Wide Web Conference, WWW '18, pages 1317-1327, Republic and Canton of Geneva, Switzerland. International World Wide Web Conferences Steering Committee.", + "Liang Wang, Wei Zhao, Zhuoyu Wei, and Jingming Liu. 2022. SimKGC: Simple contrastive knowledge graph completion with pre-trained language models. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4281-4294, Dublin, Ireland. Association for Computational Linguistics.", + "Ralph Weischedel, Martha Palmer, Mitchell Marcus, Eduard Hovy, Sameer Pradhan, Lance Ramshaw, Nianwen Xue, Ann Taylor, Jeff Kaufman, and Michelle Franchini. 2013. Ontonotes release 5.0. In Linguistic Data Consortium, Philadelphia, PA.", + "Guillaume Wenzek, Marie-Anne Lachaux, Alexis Conneau, Vishrav Chaudhary, Francisco Guzmán, Armand Joulin, and Edouard Grave. 2020. CCNet: Extracting high quality monolingual datasets from web crawl data. In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 4003-4012, Marseille, France. European Language Resources Association.", + "Shijie Wu and Mark Dredze. 2019. Beto, bentz, becas: The surprising cross-lingual effectiveness of BERT. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 833–844, Hong Kong, China. Association for Computational Linguistics.", + "Patrick Xia and Benjamin Van Durme. 2021. Moving on from OntoNotes: Coreference resolution model transfer. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5241-5256, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.", + "Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. 2021. mT5: A massively multilingual pre-trained text-to-text transformer. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 483-498, Online. Association for Computational Linguistics.", + "Zdeněk Žabokrtský, Miloslav Konopík, Anna Nedoluzhko, Michal Novák, Maciej Ogrodniczuk, Martin Popel, Ondřej Pražák, Jakub Sido, Daniel Zeman, and Yilun Zhu. 2022. Findings of the shared task on multilingual coreference resolution. In Proceedings of the CRAC 2022 Shared Task on Multilingual Coreference Resolution, pages 1-17," + ], + "bbox": [ + 115, + 85, + 489, + 917 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Gyeongju, Republic of Korea. Association for Computational Linguistics.", + "bbox": [ + 527, + 85, + 880, + 112 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "208", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "mOKB6: A Multilingual Open Knowledge Base Completion Benchmark (Appendix)", + "text_level": 1, + "bbox": [ + 124, + 79, + 872, + 117 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "A Dataset Curation", + "text_level": 1, + "bbox": [ + 112, + 128, + 302, + 142 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As discussed in Section 3, we construct mOKB6 dataset in three stages after extracting the Wikipedia articles (using WikiExtractor7) from the Wikidump of April 02, 2022. We run our construction pipeline (as shown in Figure 1) for all six languages on a single V100 (32 GB) GPU, which required 14 hours of computation to create mOKB6 dataset.", + "bbox": [ + 112, + 155, + 487, + 282 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In the first stage, we keep the sentences containing at least 6 and at most 50 tokens since we find that most of the short sentences are headings or sub-headings present in Wikipedia articles, and very long sentences can't be input to GEN2OIE (in second stage) due to maximum sequence length constraint of 1024 in mT5 (Xue et al., 2021) based GEN2OIE. This filtering step discards $18.9\\%$ of sentences on an average in all six languages. We use Stanza (Qi et al., 2020) to perform sentence- and word-segmentation on Wikipedia articles in all six languages. After filtering the sentences, the articles are processed for coreference resolution using XLM-R (Conneau et al., 2020) encoder based wlcoref (Dobrovolskii, 2021), followed by replacing the coreferent cluster mentions with their canonical cluster name using the heuristic discussed in Section 3.", + "bbox": [ + 115, + 286, + 489, + 574 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In the second stage, the coreference resolved articles are passed through GEN2OIE to get the Open IE triples. The confidence scores for these triples are computed using label rescoring, for which we refer the readers to Kolluru et al. (2022) for more details.", + "bbox": [ + 112, + 577, + 487, + 671 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Finally, in the last stage, we apply various filters, adapted from Gashteovski et al. (2019), to remove triples that are of no interest to Open KBC task, like the triples: (1) having any of its argument or relation empty, (2) containing more than 10 tokens in any of its arguments or relation, (3) having confidence score less than 0.3, (4) containing pronouns (found using Stanza) in its arguments, (5) having same subject and object (i.e. self loops), and (6) that are duplicates. These filters keep $91.6\\%$ of the triples obtained from stage 2 in all six languages.", + "bbox": [ + 112, + 674, + 489, + 866 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Further in the last stage, in order to create a dense Open KB containing minimum noise and maximum facts about the entities, we keep the triples having the Wikipedia article's title as either the subject phrase or object phrase and discard the rest. We do this by finding all the coreference clusters (of entity mentions) that contain the titles, then get the entities, or cluster names, of those clusters using the heuristic discussed in section 3, and keep those triples that contain these cluster names. This filtering step retains $23.6\\%$ of the triples.", + "bbox": [ + 507, + 129, + 884, + 305 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "B Metrics", + "text_level": 1, + "bbox": [ + 509, + 319, + 613, + 335 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We follow the previous works (Wang et al., 2022) on the evaluation methodology of Open KBC task and apply it to the multilingual Open KBC task, containing facts in multiple languages. Given an Open KB, containing a finite set of entities and open relations, the KGE model answers forward and backward queries of the form $(s,r,?)$ and $(?,r,o)$ respectively. The model ranks all the entities based on their correctness with, say, $s$ and $r$ in the forward query. Further, the evaluation is in filtered setting, where the other known correct answers, apart from $o$ , are removed from rank list.", + "bbox": [ + 507, + 346, + 882, + 539 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The commonly used evaluation metrics are hits at rank N (H@N), where $N$ is a natural number, and mean reciprocal rank (MRR). Suppose, the model ranks $o$ at $R$ among all entities. Then, H@N measures how many times $R$ is less than or equal to $N$ . MRR is the average of reciprocal ranks $\\left( \\frac{1}{R} \\right)$ . Both, H@N and MRR, are computed as average over both forms of queries over the full test set.", + "bbox": [ + 507, + 541, + 882, + 668 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "C Knowledge Graph Embedding Models", + "text_level": 1, + "bbox": [ + 507, + 683, + 878, + 699 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "SimKGC (Wang et al., 2022) is a text-based KGE model that uses two unshared pretrained BERT models (Devlin et al., 2019) for encoding (subject phrase; relation phrase) and object phrase separately. GRU-ConvE (Kocijan and Lukasiewicz, 2021) encodes both the relation phrase and argument phrase from their surface forms using two unshared GRU (Cho et al., 2014). CaRe (Gupta et al., 2019) learns separate embeddings for each argument phrase and uses a bi-directional GRU to encode the relation phrase from its surface form. Both, GRU-ConvE and CaRe, are initialised with Glove embeddings (Pennington et al., 2014).", + "bbox": [ + 507, + 709, + 884, + 917 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "$^{7}$ https://github.com/samuelbroscheit/wikiextractor-wikimentions", + "bbox": [ + 112, + 891, + 410, + 917 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "209", + "bbox": [ + 485, + 927, + 515, + 940 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/5806f2ef64071ea69556004f4def11da50b67e67341ac2a61a4802cd3edd4777.jpg", + "image_caption": [ + "Figure 4: Previous Open KB construction pipelines like Gashteovski et al. (2019) (shown by green arrows) lack coreference resolution system, which result in filtering important facts like (Barack Obama; returned to Honolulu, Hawaii in; 1971). Our pipeline (shown by blue arrows) increases the coverage of facts due to mCoref system." + ], + "image_footnote": [], + "bbox": [ + 114, + 80, + 884, + 156 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To choose the best model for our experiments (Table 3, Figure 3), we train the recent knowledge graph embedding (KGE) models — CaRe., GRUConvE and SimKGC on the English Open KB in mOKB6. We report performance in Table 4 using the three metrics: hits at rank 1 (H@1), hits at 10 (H@10), and mean reciprocal rank (MRR). We find that SimKGC with BERT encoder outperforms the other two models.", + "bbox": [ + 110, + 233, + 490, + 376 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/de213892889db43fdde203a034d8a16a57326f1090933c812519f9b97ed00732.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
H@1H@10MRR
CaRe6.611.38.3
GRU-ConvE12.427.817.8
SimKGC (BERT)16.140.024.3
SimKGC (mBERT)14.838.722.8
SimKGC (XLM-R)13.835.821.3
", + "bbox": [ + 154, + 386, + 448, + 483 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Since BERT supports only English language, we replace BERT in SimKGC with multilingual pretrained language models like mBERT (Devlin et al., 2019) or XLM-R (Conneau et al., 2020), to extend SimKGC model to other languages. We find in Table 4 that SimKGC with mBERT is better than with XLM-R by $2.9\\%$ H@10 and $1.5\\%$ MRR, possibly because mBERT (and mOKB6) uses Wikipedia while XLM-R uses CommonCrawl (Wenzek et al., 2020) during pre-training. Thus, we use SimKGC with mBERT as the underlying encoder to run our experiments for all the languages.", + "bbox": [ + 112, + 556, + 489, + 750 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "D KGE Model Training Details", + "text_level": 1, + "bbox": [ + 112, + 763, + 403, + 780 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We use the code from official repositories of the KGE models — SimKGC (Wang et al., 2022), GRU-ConvE (Kocijan and Lukasiewicz, 2021), and CaRe (Gupta et al., 2019) for our experiments. The models are trained using Adam optimizer (Kingma and Ba, 2015) on a single A100 (40 GB) GPU with three different random seeds and we report the average of three evaluation runs.", + "bbox": [ + 112, + 790, + 490, + 917 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We do not perform hyperparameter search trials, except for batch size, and use the default hyperparameters from the respective codes of KGE models (see Table 5). We use early stopping to find the best model checkpoints based on HITS@1. The dev set is different for each baseline: MONO, TRANS, MONO+TRANS, and UNION+TRANS use individual language's dev set, whereas UNION w/o En and UNION use the English dev set. We report the performance of baseline models on the dev sets in Table 9 and Table 10.", + "bbox": [ + 507, + 233, + 884, + 409 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/d3718ead02b038fcf077c22e328c0419ffb65e3796554eecfaa92302a8b3197d.jpg", + "table_caption": [ + "Table 4: Performance $(\\%)$ of the KGE models on the English test set in mOKB6 dataset. The reported numbers are an average of three runs using different seeds." + ], + "table_footnote": [], + "table_body": "
HyperparameterSimKGCGRU-ConvECaRe
#epochs100500500
#patience epochs101010
learning rate3e-53e-41e-3
dropout0.10.30.5
batch size2561024128
additive margin0.02N/AN/A
", + "bbox": [ + 527, + 418, + 865, + 520 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We provide the number of trainable parameters of each KGE model in Table 6. Based on the batch size and model size, different experiments consume different GPU hours. To train on English Open KB (in mOKB6 dataset), CaRe and GRU-ConvE models took 2.5 hours and 0.5 hours, respectively, whereas SimKGC takes nearly 1 hour of GPU time.", + "bbox": [ + 507, + 562, + 882, + 675 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/566686bbdf909d805d01895739fe45773e1eb857ef3b1e52fe8bd092bd59995c.jpg", + "table_caption": [ + "Table 5: Hyperparameters of the KGE models." + ], + "table_footnote": [], + "table_body": "
KGE model#trainable parameters
CaRe12,971,423
GRU-ConvE12,085,523
SimKGC (BERT)216,620,545
SimKGC (mBERT)355,706,881
SimKGC (XLM-R)1,119,780,865
", + "bbox": [ + 557, + 684, + 833, + 775 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 6: Number of trainable parameters in the KGE models.", + "bbox": [ + 507, + 784, + 882, + 812 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "210", + "bbox": [ + 485, + 927, + 515, + 940 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/1ae8b58d54b659eadae84e2e9e7bed390e9eee1cd63c89c1084ca7c1d879daa6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
EnglishHindiTeluguSpanishPortugueseChinese
H@1H@10MRRH@1H@10MRRH@1H@10MRRH@1H@10MRRH@1H@10MRRH@1H@10MRR
English68.497.178.83.417.28.31.611517.850.728.61744.6265.422.311.1
Hindi1942.226.780.699.588.32.412.55.912.33619.912.333.919.75.321.910.8
Telugu19.542.227.24.318.79.474.499.584.210.935.418.910.73418.54.721.410.1
Spanish27.960.438.84.117.88.91.810.75.18410090.337.67450.16.524.912.8
Portuguese27.858.738.24.418.29.31.710.55.141.578.553.684.299.990.86.62613.2
Chinese22.148.430.63.518.58.81.812.25.414.842.824.215.741.624.181.699.889.2
", + "bbox": [ + 119, + 80, + 884, + 200 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 7: Performance (%) of the six MEMORIZE models, which have been trained on each language's test set and tested on all the test sets in mOKB6 dataset.", + "bbox": [ + 112, + 210, + 882, + 239 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "E Contextual Triples", + "text_level": 1, + "bbox": [ + 112, + 263, + 314, + 280 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Open IE triples are of various kinds and not all of them can be used for Open KBC task. Various filtering steps are used to remove some of these in data curation (Section 3). We define contextual triples as another kind of noisy triples, which are specific to, and are not interpretable out of, the context of text from which they are extracted.", + "bbox": [ + 112, + 288, + 487, + 401 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "(Max Born; continued; scientific work)", + "bbox": [ + 179, + 414, + 420, + 426 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "(Robb Gravett; won; the championship)", + "bbox": [ + 179, + 426, + 421, + 439 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "(George Herbert Walker Bush; was; out of touch)", + "bbox": [ + 151, + 438, + 450, + 450 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "(Christianity; is; dominant)", + "bbox": [ + 216, + 450, + 384, + 462 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 8: Examples of contextual triples.", + "bbox": [ + 164, + 473, + 436, + 488 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "From the first two triples in Table 8, it is unclear which scientific work Max Born continued, or which championship Robb Gravett has won. The last two triples are too specific to the context and contain no factual information.", + "bbox": [ + 112, + 504, + 489, + 583 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "211", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/d91a900a0365a74f15084007720fc5ea4342636b49c33865f7c6a244dedb405f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
English (En)Hindi (Hi)Telugu (Te)Spanish (Es)Portuguese (Pt)Chinese (Zh)
H@1H@10MRRH@1H@10MRRH@1H@10MRRH@1H@10MRRH@1H@10MRRH@1H@10MRR
MONO16.238.723.918.239.425.98.52012.517.336.623.717.639.625.310.831.917.8
TRANS---8.123.713.53.315.47.512.933.620.312.637.220.6520.810.3
MONO+TRANS---20.843.228.67.824.813.420.24628.82145.929.210.630.116.7
UNION19.939.626.414.538.222.45.92010.619.843.227.919.743.82811.23318.8
UNION w/o En5.819.510.615.439.323.36.320.511.119.441.626.416.942.925.911.33318.4
UNION+TRANS---20.844.928.87.327.11421.445.329.619.449.129.16.93115.1
", + "bbox": [ + 119, + 222, + 878, + 337 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/2cbc6adebcb647e5e42b0fff45ac1e43cb9b74a46ce50b15727bb6b109119dbb.jpg", + "table_caption": [ + "Table 9: Performance (%) of SimKGC on the dev sets (of mOKB6 dataset) in six languages." + ], + "table_footnote": [], + "table_body": "
H@1H@10MRR
CaRe7.111.18.5
GRU-ConvE16.831.522.1
SimKGC (BERT)20.340.127.1
SimKGC (mBERT)16.238.723.9
SimKGC (XLM-R)1736.623.2
", + "bbox": [ + 352, + 653, + 643, + 747 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 10: Performance (%) of the KGE models on dev set of English Open KB in mOKB6 dataset.", + "bbox": [ + 164, + 758, + 830, + 772 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "212", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A For every submission:", + "bbox": [ + 115, + 107, + 322, + 122 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "A1. Did you describe the limitations of your work? 8", + "A2. Did you discuss any potential risks of your work? There are no potential risks of our work to our knowledge.", + "A3. Do the abstract and introduction summarize the paper's main claims?", + "A4. Have you used AI writing assistants when working on this paper? Left blank." + ], + "bbox": [ + 129, + 126, + 695, + 288 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "B Did you use or create scientific artifacts?", + "bbox": [ + 114, + 298, + 489, + 316 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "3,4", + "bbox": [ + 134, + 321, + 161, + 334 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "B1. Did you cite the creators of artifacts you used? 3,4", + "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Abstract", + "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Not applicable. Left blank.", + "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Not applicable. Left blank.", + "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? 3", + "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. 3" + ], + "bbox": [ + 129, + 346, + 880, + 752 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "C Did you run computational experiments?", + "bbox": [ + 114, + 764, + 492, + 781 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4", + "bbox": [ + 134, + 788, + 146, + 799 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Appendix D", + "bbox": [ + 129, + 810, + 880, + 860 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "ACL 2023 Responsible NLP Checklist", + "bbox": [ + 132, + 84, + 433, + 99 + ], + "page_idx": 12 + }, + { + "type": "page_footnote", + "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.", + "bbox": [ + 112, + 868, + 877, + 892 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "213", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Appendix D", + "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? \nAppendix D", + "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Appendix A, D" + ], + "bbox": [ + 129, + 83, + 878, + 282 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? 3", + "bbox": [ + 112, + 292, + 877, + 328 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?", + "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Not applicable. Left blank.", + "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Not applicable. Left blank.", + "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Not applicable. Left blank.", + "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Not applicable. Left blank." + ], + "bbox": [ + 127, + 338, + 878, + 640 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "214", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 13 + } +] \ No newline at end of file diff --git a/2023/mOKB6_ A Multilingual Open Knowledge Base Completion Benchmark/70c57a7f-70d1-4b2e-a713-7ef674b2ac69_model.json b/2023/mOKB6_ A Multilingual Open Knowledge Base Completion Benchmark/70c57a7f-70d1-4b2e-a713-7ef674b2ac69_model.json new file mode 100644 index 0000000000000000000000000000000000000000..18c715fa264c0291e4f6897b1ba2ce17bfcfe409 --- /dev/null +++ b/2023/mOKB6_ A Multilingual Open Knowledge Base Completion Benchmark/70c57a7f-70d1-4b2e-a713-7ef674b2ac69_model.json @@ -0,0 +1,2428 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.125, + 0.09, + 0.873, + 0.112 + ], + "angle": 0, + "content": "mOKB6: A Multilingual Open Knowledge Base Completion Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.163, + 0.124, + 0.844, + 0.141 + ], + "angle": 0, + "content": "Shubham Mittal\\(^{\\alpha \\dagger}\\) Keshay Kolluru\\(^{\\beta \\dagger}\\) Soumen Chakrabarti\\(^{\\gamma}\\) Mausam\\(^{\\alpha}\\)" + }, + { + "type": "text", + "bbox": [ + 0.346, + 0.143, + 0.658, + 0.158 + ], + "angle": 0, + "content": "\\(^{\\alpha}\\) Indian Institute of Technology Delhi" + }, + { + "type": "text", + "bbox": [ + 0.392, + 0.159, + 0.612, + 0.174 + ], + "angle": 0, + "content": "\\(^{\\beta}\\) KnowDis AI, New Delhi" + }, + { + "type": "text", + "bbox": [ + 0.334, + 0.176, + 0.669, + 0.192 + ], + "angle": 0, + "content": "\\(\\gamma\\) Indian Institute of Technology Bombay" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.193, + 0.75, + 0.209 + ], + "angle": 0, + "content": "shubhamiitd18@gmail.com, keshav.kolluru@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.278, + 0.211, + 0.725, + 0.225 + ], + "angle": 0, + "content": "soumen@cse.iitb.ac.in, mausam@cse.iitd.ac.in" + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.269, + 0.341, + 0.284 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.301, + 0.461, + 0.642 + ], + "angle": 0, + "content": "Automated completion of open knowledge bases (Open KBs), which are constructed from triples of the form (subject phrase, relation phrase, object phrase), obtained via open information extraction (Open IE) system, are useful for discovering novel facts that may not be directly present in the text. However, research in Open KB completion (Open KBC) has so far been limited to resource-rich languages like English. Using the latest advances in multilingual Open IE, we construct the first multilingual Open KBC dataset, called mOKB6, containing facts from Wikipedia in six languages (including English). Improving the previous Open KB construction pipeline by doing multilingual coreference resolution and keeping only entity-linked triples, we create a dense Open KB. We experiment with several models for the task and observe a consistent benefit of combining languages with the help of shared embedding space as well as translations of facts. We also observe that current multilingual models struggle to remember facts seen in languages of different scripts." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.66, + 0.26, + 0.675 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.688, + 0.49, + 0.882 + ], + "angle": 0, + "content": "Open information extraction (Open IE) systems (Mausam, 2016) such as ReVerb (Etzioni et al., 2011) and OpenIE6 (Kolluru et al., 2020) can extract triples, or facts, of the form (subject phrase, relation phrase, object phrase), which can be denoted as \\((s,r,o)\\), from text (e.g., Wikipedia articles) without using any pre-defined ontology. Open knowledge base (Open KB) is constructed using these Open IE triples where the subject phrases and object phrases are nodes and relation phrases are labels on edges connecting the nodes in the graph. Open knowledge base completion (Open KBC) is" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.254, + 0.885, + 0.334 + ], + "angle": 0, + "content": "the task of discovering new links between nodes using the graph structure of the Open KB. Knowledge graph embedding (KGE) models are typically used for the Open KBC task, where they are asked to answer questions of the form \\((s,r,?)\\) and \\((?,r,o)\\)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.336, + 0.885, + 0.513 + ], + "angle": 0, + "content": "Research in Open KBC has been restricted to English (Vashishth et al., 2018) due to lack of Open KBs in other languages. We aim to study multilingual Open KBC, with the motivation that the information available in high resource languages like English may help when inferring links in Open KBs that use low resource languages like Telugu. Moreover, intuitively, if all the information in different languages can be pooled together, then it may help the model learn better, and allow information flow across Open KBs in different languages." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.515, + 0.885, + 0.772 + ], + "angle": 0, + "content": "We design the first multilingual Open KB construction pipeline (shown in Figure 1) using a multilingual Open IE system, GEN2OIE (Kolluru et al., 2022). We find that coreference resolution is missing in existing Open KB construction (Gashteovski et al., 2019) but is important for increasing the coverage of facts (as described in Figure 4). We re-train a recent coref model (Dobrovolskii, 2021) using XLM-R (Conneau et al., 2020) as the underlying multilingual encoder and add it to our pipeline. For constructing a high quality test set, we use 988 manually verified facts in English. For extending to other languages, we automatically translate English facts. The dataset thus constructed, called mOKB6, contains 42K facts in six languages: English, Hindi, Telugu, Spanish, Portuguese, and Chinese." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.775, + 0.885, + 0.92 + ], + "angle": 0, + "content": "We report the first baselines for multilingual Open KBC task. We find that they are able to benefit from information in multiple languages when compared to using facts from a single language. Translations of Open KB facts also help the models. However, we notice that although the multilingual KGE models learn facts in a particular language, they struggle to remember the same fact, when queried in another language with different script." + }, + { + "type": "page_footnote", + "bbox": [ + 0.14, + 0.892, + 0.453, + 0.905 + ], + "angle": 0, + "content": "Major part of work done as students at IIT Delhi." + }, + { + "type": "page_footnote", + "bbox": [ + 0.138, + 0.905, + 0.486, + 0.918 + ], + "angle": 0, + "content": "\\(^{1}\\)Dataset and code released at github.com:dair-iitd/mokb6" + }, + { + "type": "list", + "bbox": [ + 0.138, + 0.892, + 0.486, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.928, + 0.514, + 0.941 + ], + "angle": 0, + "content": "201" + }, + { + "type": "footer", + "bbox": [ + 0.227, + 0.946, + 0.77, + 0.958 + ], + "angle": 0, + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + }, + { + "type": "footer", + "bbox": [ + 0.377, + 0.959, + 0.621, + 0.972 + ], + "angle": 0, + "content": "Volume 2: Short Papers, pages 201-214" + }, + { + "type": "footer", + "bbox": [ + 0.297, + 0.973, + 0.702, + 0.985 + ], + "angle": 0, + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.115, + 0.084, + 0.271, + 0.099 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.111, + 0.49, + 0.368 + ], + "angle": 0, + "content": "Multilingual Open KBC datasets are absent in literature to the best of our knowledge, although multiple English Open KBC datasets are available. OLPBench (Broscheit et al., 2020), derived from OPIEC (Gashteovski et al., 2019), is a large-scale Open KBC dataset that contains 30M triples and is constructed from English Wikipedia using MinIE system (Gashteovski et al., 2017). The evaluation data contains 10K triples randomly sampled from 1.25M linked triples. ReVerb45K (Vashisth et al., 2018) and ReVerb20K (Galarraga et al., 2014) are smaller Open KBC datasets constructed from Clueweb09 corpus\\(^2\\) using ReVerb Open IE system (Fader et al., 2011). Both the datasets keep only those tuples in which both the subject phrase and object phrase link to a finite set of Freebase entities." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.369, + 0.49, + 0.497 + ], + "angle": 0, + "content": "Multilingual Open IE (mOpenIE) systems like GEN2OIE (Kolluru et al., 2022) and Multi\\(^{2}\\)OIE (Ro et al., 2020) enable extracting facts from multiple languages. We use the GEN2OIE model for constructing mOKB6 dataset as it is trained with language-specific facts transferred from English, while Multi\\(^{2}\\)OIE relies on zero-shot transfer for languages other than English." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.508, + 0.49, + 0.733 + ], + "angle": 0, + "content": "Knowledge Graph Embedding (KGE) Models: Conventional KGE models like TransE (Bordes et al., 2013), ComplEx (Trouillon et al., 2016), ConvE (Dettmers et al., 2018), and TuckER (Balazevic et al., 2019) have been used for Open KBC task (Gupta et al., 2019; Broscheit et al., 2020; Chandrahas and Talukdar, 2021; Kocijan and Lukasiewicz, 2021). Given a triple \\((s,r,o)\\), these models encode the subject phrase, relation phrase, and object phrase from free text, and pass the encodings to a triple-scoring function, which is optimized using binary cross entropy loss. ComplEx has also been used for multilingual closed KBC task (Chakrabarti et al., 2022)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.734, + 0.49, + 0.894 + ], + "angle": 0, + "content": "Pretrained language models like BERT (Devlin et al., 2019) have been used in KGE models for the KBC task (Lovelace and Rosé, 2022; Lv et al., 2022; Chandrahas and Talukdar, 2021; Kim et al., 2020). SimKGC (Wang et al., 2022) is the state of the art KGE model on closed KBC task. It computes the score of a triple \\((s, r, o)\\) as the cosine similarity of the embeddings of \\((s; r)\\) and \\((o)\\), computed using two separate pretrained BERT models without any weight sharing." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.084, + 0.694, + 0.099 + ], + "angle": 0, + "content": "3 Dataset Curation" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.11, + 0.885, + 0.351 + ], + "angle": 0, + "content": "We aim to construct a dense multilingual Open KB that maximizes the information about a given real-world entity, which may be represented as multiple nodes across languages. Therefore, we consider those Wikipedia articles3 that are available in six languages: English, Hindi, Telugu, Spanish, Portuguese, and Chinese4. This will also help the model learn from facts in high resource language like English and answer queries in low resource language like Telugu. We work with 300 titles randomly sampled from the ones common among all six languages (found using MediaWiki-Langlinks (MediaWiki, 2021)). Thus, we extract facts from \\(6 \\times 300\\) Wikipedia articles. We discuss the three stages of our pipeline below." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.36, + 0.885, + 0.536 + ], + "angle": 0, + "content": "Stage 1 We first process each Wikipedia article through a coreference resolution system. Although language-specific end-to-end neural coref models have been developed (Žabokrtský et al., 2022; Xia and Van Durme, 2021), multilingual models that work on all our languages of interest are absent in the literature. Therefore, we retrain wl-coref (Dobrovolskii, 2021) with XLM-R (Conneau et al., 2020) on the English training data (available in OntoNotes (Weischedel et al., 2013)) that can work zero-shot for other languages." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.538, + 0.885, + 0.844 + ], + "angle": 0, + "content": "Coref models detect and cluster mentions, but do not identify a canonical cluster name, which is needed for standardizing all the mentions in the cluster. To find cluster names, entity linking systems such as mGENRE (De Cao et al., 2022) or Wikipedia hyperlinks can be used. However, we found that they result in low recall, particularly for low resource languages. Thus, we employ a heuristic to find the cluster name and replace each of the coreferent mentions with it. The score for each mention is represented by a tuple, computed as: Score(mention phrase) = (#proper nouns, #nouns, #numerals, #adjectives, #pronouns, #verbs). The tuple is ordered according to the importance of each field (POS tags) for the cluster name, which is determined empirically. Two tuples are compared index-wise with higher priority given to lower indices to determine the best scoring mention that is chosen as the canonical name (Table 1)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.852, + 0.883, + 0.883 + ], + "angle": 0, + "content": "Stage 2 We use GEN2OIE to extract Open IE triples from the coreference resolved sentences." + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.891, + 0.717, + 0.905 + ], + "angle": 0, + "content": "3Wikidump of April 02, 2022" + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.905, + 0.875, + 0.919 + ], + "angle": 0, + "content": "4languages are chosen to match availability of Gen2OIE" + }, + { + "type": "list", + "bbox": [ + 0.531, + 0.891, + 0.875, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.136, + 0.904, + 0.463, + 0.919 + ], + "angle": 0, + "content": "2http://www.lemurproject.org/clueweb09.php/" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.928, + 0.517, + 0.941 + ], + "angle": 0, + "content": "202" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.116, + 0.083, + 0.885, + 0.135 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.147, + 0.884, + 0.191 + ], + "angle": 0, + "content": "Figure 1: Our three-staged multilingual Open KB construction pipeline for mOKB6. mCoref is multilingual coreference resolution system, having XLM-R (Conneau et al., 2020) encoder based wl-coref (Dobrovolskii, 2021), and mOpenIE is multilingual open information extraction system, consisting of GEN2OIE (Kolluru et al., 2022)." + }, + { + "type": "table", + "bbox": [ + 0.146, + 0.206, + 0.458, + 0.272 + ], + "angle": 0, + "content": "
MentionsScoresCluster Name
Barack Obama(2,0,0,0,0,0)
Obama(1,0,0,0,0,0)Barack Obama
He(0,0,0,0,1,0)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.282, + 0.49, + 0.311 + ], + "angle": 0, + "content": "Table 1: Parts of speech tags are used to find the canonical name of the coreferent cluster of entity mentions." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.325, + 0.489, + 0.485 + ], + "angle": 0, + "content": "Stage 3 Similar to Gashteovski et al. (2019), we apply various filters to remove noisy triples that have empty or very long arguments, or have less confidence than 0.3 (as assigned by GEN2OIE). We further only keep triples that have the article's title as either the subject phrase or object phrase, to avoid generic or specific triples, valid only in the particular context. Examples of contextual triples (Choi et al., 2021) are discussed in Appendix E. See Appendix A for further data curation details." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.489, + 0.49, + 0.89 + ], + "angle": 0, + "content": "These automatically extracted triples form the train set of mOKB6. To form a high quality test set in six languages with limited access to experts in all languages, the test set is created in a semiautomatic way. We sample 1600 English triples from the train set (which are subsequently filtered) and manually remove noisy triples. We use inter-annotation agreement between two annotators to check if they both agree that the given triple is noisy or clean. With an agreement of \\(91\\%\\), we retain 988 English triples, which we automatically translate to the other five languages. As illustrated in Figure 2, to translate a triple, we convert it to a sentence after removing tags and use Google translate for translating the triple-converted sentence to the remaining five languages. We observed high quality of translated triples, with \\(88\\%\\) satisfactory translations as determined by native-speakers of three languages on a set of 75 translated triples. To get the Open IE subject phrase, relation phrase and object phrase tags, we project the labels from the original English triple to the translated sentence using word alignments (Kolluru et al., 2022). Finally, we are left with 550 triples in each language after removing examples where some labels could" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.209, + 0.885, + 0.275 + ], + "angle": 0, + "content": "not be aligned. We use these \\(6 \\times 550\\) triples as the test sets. The train and dev sets are created from the remaining triples in each language such that the dev set has 500 randomly sampled triples (Table 2)." + }, + { + "type": "image", + "bbox": [ + 0.574, + 0.288, + 0.822, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.455, + 0.884, + 0.498 + ], + "angle": 0, + "content": "Figure 2: Method to translate Open IE triple using Google translate, and followed by label projection using word alignments (Kolluru et al., 2022)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.514, + 0.884, + 0.626 + ], + "angle": 0, + "content": "We analyse the entity overlap across languages and find that on an average, a test entity (which is present in either the subject phrase or object phrase of a test tuple) is present 17.73 times in English, 0.94 times in Hindi, 0.47 times in Telugu, 2.33 times in Spanish, 1.69 times in Portuguese, and 1.45 times in Chinese train set." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.627, + 0.885, + 0.772 + ], + "angle": 0, + "content": "Our construction pipeline improves over OPIEC in three ways: (1) we use a multilingual Open IE system, instead of an English-specific Open IE system like in OPIEC, enabling us to curate Open KBs in many languages, (2) we add a multilingual coreference resolution system in our pipeline, and (3) the English test triples are manually verified. Further, we manually evaluate and review the noise at each step of data curation in Section 4." + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.782, + 0.883, + 0.849 + ], + "angle": 0, + "content": "
EnHiTeEsPtZh
#entity2063746253972565153045037
#relation787021771907282326442325
#train2019527861992396635283420
" + }, + { + "type": "table_caption", + "bbox": [ + 0.508, + 0.858, + 0.884, + 0.915 + ], + "angle": 0, + "content": "Table 2: Statistics of individual Open KBs in mOKB6 in English (En), Hindi (Hi), Telugu (Te), Spanish (Es), Portuguese (Pt), and Chinese (Zh). The dev and test set for each Open KB contain 500 and 550 triples each." + }, + { + "type": "page_footnote", + "bbox": [ + 0.136, + 0.904, + 0.381, + 0.919 + ], + "angle": 0, + "content": "5https://translate.google.co.in/" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.928, + 0.516, + 0.941 + ], + "angle": 0, + "content": "203" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.114, + 0.085, + 0.296, + 0.099 + ], + "angle": 0, + "content": "4 Noise Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.11, + 0.488, + 0.254 + ], + "angle": 0, + "content": "Curating an Open KB involves various stages and each stage induces its noise in the construction pipeline (Gashteovski et al., 2019). We manually evaluate the noise induced at each stage of our pipeline (Figure 1) and discuss the same in this section. We ask native speakers of four (out of six) languages - English, Hindi, Telugu, and Chinese to assess the output quality, or precision, of each stage as discussed below." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.255, + 0.488, + 0.367 + ], + "angle": 0, + "content": "In the first stage, we assess the performance of the coreference resolution system over Wikipedia articles. We find a high precision of \\(95.5\\%\\) in coref's mention clustering and \\(89.82\\%\\) accuracy in finding canonical cluster name (using the heuristic illustrated in Table 1), computed over 40 randomly sampled coref clusters (10 in each language)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.368, + 0.488, + 0.463 + ], + "angle": 0, + "content": "For evaluating the Open IE system, GEN2OIE, in the second stage, we mark an extraction of a sentence as correct if it has syntactically correct arguments and it is coherent with the sentence. We get an average precision of \\(63.4\\%\\) on 80 extractions (20 in each language)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.465, + 0.489, + 0.625 + ], + "angle": 0, + "content": "We evaluate the triples, or Open KB facts, at the last stage after passing through various noise-removing filters. Note that these triples also form the train set (and dev set) in mOKB6 dataset. We mark triples as correct when they contain real-world entities, and also, factual information about them. If the triple is very generic or contextual (see Appendix E), it is marked as incorrect. We find the train (and dev) set quality to be \\(69.3\\%\\), averaged over 80 triples in four languages." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.637, + 0.262, + 0.654 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.662, + 0.486, + 0.694 + ], + "angle": 0, + "content": "Our experimental study on multilingual open KBC task investigates the following research questions:" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.695, + 0.486, + 0.725 + ], + "angle": 0, + "content": "1. Does the KGE model benefit from facts in different languages? (Section 5.1)" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.727, + 0.488, + 0.757 + ], + "angle": 0, + "content": "2. Can translation help transfer among languages? (Section 5.2)" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.759, + 0.486, + 0.789 + ], + "angle": 0, + "content": "3. Does the KGE model remember facts seen across different languages? (Section 5.3)" + }, + { + "type": "list", + "bbox": [ + 0.129, + 0.695, + 0.488, + 0.789 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.791, + 0.489, + 0.919 + ], + "angle": 0, + "content": "We use SimKGC model (Wang et al., 2022) with pretrained mBERT initialization to run our experiments, after comparing with recent KGE models (Appendix C). For evaluation, we use three metrics -hits at rank 1 (H@1), hits at rank 10 (H@10), and mean reciprocal rank (MRR). The formal definitions of them are provided in Appendix B. We discuss further model training details in Appendix D." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.085, + 0.798, + 0.101 + ], + "angle": 0, + "content": "5.1 Training on Multilingual Facts" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.113, + 0.884, + 0.273 + ], + "angle": 0, + "content": "We train and compare monolingual model, called MONO, with multilingual models, UNION and UNION w/o En. In MONO, we train one model for each language using its respective Open KB, whereas in UNION, a single model is trained on six languages' Open KBs together. UNION outperforms MONO in all languages by an average of \\(4.6\\%\\) H@10 and \\(2.8\\%\\) MRR (see Table 3), which provides evidence of information flow across languages and the model benefits from it." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.278, + 0.884, + 0.391 + ], + "angle": 0, + "content": "To check the extent of flow from (high-resource) English to the other languages, we also train on the five languages except English, which we call UNION w/o En. We find UNION w/o En also outperforms MONO by \\(2.7\\%\\) H@10 and \\(1.2\\%\\) MRR over the five languages, hinting that interlingual transfer is more general and pervasive." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.415, + 0.771, + 0.43 + ], + "angle": 0, + "content": "5.2 Open KB Facts Translation" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.443, + 0.884, + 0.7 + ], + "angle": 0, + "content": "Apart from relying only on multilingual transfer in the embedding space, we analyse the effect of using translated triples in the training of the KGE model. We translate the English training triples to the other five languages (Section 3) and train monolingual models using only the translated triples (TRANS). To leverage facts present in each language's Open KB, we make MONO+TRANS, where we add language-specific MONO data to the translated triples. Table 3 shows that MONO+TRANS is better than MONO by a large margin of \\(15.5\\%\\) H@1, \\(29.2\\%\\) H@10, and \\(20.0\\%\\) MRR, averaged over five languages. Also, MONO+TRANS improves over TRANS by \\(2.1\\%\\) H@10 and \\(2.0\\%\\) MRR, showcasing the importance of facts in each language's Open KBs." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.704, + 0.884, + 0.881 + ], + "angle": 0, + "content": "To effectively gain from transfer in both the embedding space as well as translation, we introduce \\( \\text{UNION+TRANS} \\). We train one model for each language, on the combination of UNION triples and the translated train triples from English Open KB to that language. \\( \\text{UNION+TRANS} \\) is better than UNION by \\( 25.9\\% \\) H@10 and \\( 18.4\\% \\) MRR. This suggests that the model is able to benefit from English facts when they are translated to the query language, unlike in UNION where the English facts are present only in English." + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.905, + 0.853, + 0.919 + ], + "angle": 0, + "content": "6English source achieved the best translation quality." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.928, + 0.517, + 0.941 + ], + "angle": 0, + "content": "204" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.12, + 0.081, + 0.886, + 0.198 + ], + "angle": 0, + "content": "
English (En)Hindi (Hi)Telugu (Te)Spanish (Es)Portuguese (Pt)Chinese (Zh)
H@1H@10MRRH@1H@10MRRH@1H@10MRRH@1H@10MRRH@1H@10MRRH@1H@10MRR
MONO14.838.722.83.014.87.21.58.13.96.423.712.36.321.711.42.413.16.2
UNION w/o En5.721.510.92.915.47.41.810.24.98.127.814.56.726.112.93.215.57.5
UNION16.740.824.83.616.68.11.59.34.510.632.217.69.729.316.64.018.88.9
TRANS---20.547.629.78.728.715.523.250.632.420.550.730.514.039.422.5
MONO+TRANS---20.245.428.414.338.522.223.551.532.921.448.930.717.943.226.6
UNION+TRANS---23.349.732.315.138.523.123.952.433.423.552.133.116.943.626.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.206, + 0.886, + 0.265 + ], + "angle": 0, + "content": "Table 3: Performance (%) of SimKGC model on mOKB6 dataset, comprising of Open KBs in six languages. MONO, TRANS, and MONO+TRANS are monolingual models trained only on facts of one language whereas UNION, UNION w/o En, and UNION+TRANS are multilingual models trained with facts from multiple languages. All reported numbers are an average of three runs using different seeds. Best scores are highlighted in bold." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.289, + 0.383, + 0.304 + ], + "angle": 0, + "content": "5.3 Cross-lingual Memorization" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.312, + 0.49, + 0.569 + ], + "angle": 0, + "content": "Pretrained multilingual language models such as mBERT have demonstrated strong cross-lingual transfer capabilities (Wu and Dredze, 2019). We investigate cross-lingual memorization of the KGE model by showing facts in one language and querying the same facts in other five languages. For each language, \\( L \\), we take the UNION model and train it further on the test set of that language's Open KB, which we call MEMORIZE\\(_L\\) model. Then, we test each MEMORIZE\\(_L\\) model on the six test sets. Since the test sets (in mOKB6 dataset) of the different languages contain the same facts, this experiment allows us to investigate cross-lingual memorization. We provide the H@10 scores of MEMORIZE models in Figure 3 and the performance on other metrics (H@1 and MRR) is reported in Table 7." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.57, + 0.49, + 0.764 + ], + "angle": 0, + "content": "The model achieves at least \\(97\\%\\) H@10 when tested on the language used for training (diagonal). We observe that there is relatively good crosslingual memorization among languages that share the same script (Latin in English, Spanish, and Portuguese), but the model struggles to remember facts when seen in languages of different scripts. Many entities look similar in shared scripts, possibly leading to better information transfer. For example, the \\(\\mathsf{MEMORIZE}_{En}\\) achieves H@10 of \\(50.7\\%\\) in Spanish (Es) compared to \\(22.3\\%\\) in Chinese (Zh) and \\(11\\%\\) in Telugu (Te)." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.779, + 0.403, + 0.794 + ], + "angle": 0, + "content": "6 Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.807, + 0.49, + 0.92 + ], + "angle": 0, + "content": "We create and release the mOKB6 dataset, the first multilingual Open Knowledge Base Completion dataset with 42K facts in six languages: English, Hindi, Telugu, Spanish, Portuguese, and Chinese. Its construction uses multilingual coreference resolution, entity-mention cluster naming, multilingual open information extraction and various filtering" + }, + { + "type": "image", + "bbox": [ + 0.548, + 0.305, + 0.831, + 0.465 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.483, + 0.886, + 0.626 + ], + "angle": 0, + "content": "Figure 3: Performance (H@10) of MEMORIZE models. Row \\( L \\) shows the performance of \\( \\text{MEMORIZE}_L \\) model across the test sets of all languages (columns). For example, the performance of \\( \\text{MEMORIZE}_{En} \\) when tested on English (En) is \\( 97.1\\% \\) H@10, and \\( \\text{MEMORIZE}_{En} \\) when tested on Spanish (Es) gives \\( 50.7\\% \\) H@10. We find relatively good cross-lingual transfer among languages that use same script (Latin in English, Spanish and Portuguese) compared to those using different scripts (English, Hindi, Telugu and Chinese)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.655, + 0.885, + 0.737 + ], + "angle": 0, + "content": "steps to improve the quality of the extracted facts. We also report the first baselines on the task using the existing state of the art KGE models trained with facts from different languages using various augmentation strategies." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.743, + 0.886, + 0.919 + ], + "angle": 0, + "content": "Our work opens many important research questions: (1) Can we develop better strategies to combine facts in different languages? (2) Can we build models that achieve strong information transfer across unrelated languages with same or different scripts? (3) Can we train the neural model to ignore contextual triples (Appendix E), thus improving overall performance? and (4) Can tying the same entities across various languages help the model generalize better? We leave these questions to be addressed in future work." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.516, + 0.941 + ], + "angle": 0, + "content": "205" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.115, + 0.085, + 0.317, + 0.1 + ], + "angle": 0, + "content": "7 Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.11, + 0.49, + 0.239 + ], + "angle": 0, + "content": "Keshav was supported by TCS Research Fellowship during his PhD. Mausam is supported by grants from Huawei, Google, Verisk and IBM, and a Jai Gupta Chair Fellowship. He also acknowledges Google and Yardi School of AI travel grants. Soumen is partly supported by a Jagadish Bose Fellowship and a grant from Cisco. We thank IIT Delhi HPC facility for compute resources." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.25, + 0.251, + 0.266 + ], + "angle": 0, + "content": "8 Limitations" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.275, + 0.49, + 0.436 + ], + "angle": 0, + "content": "Although multilingual, the constructed open KB is limited to the sampling of the chosen six languages. We do not know how well the system will generalize to various language families that have not been considered here. Further, even among the languages considered, the performance of even the best-performing systems, as measured through \\(\\mathrm{H@1}\\) is still in the low 20's. Therefore the models are not yet ready to be deployed for real-world applications." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.463, + 0.214, + 0.478 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.485, + 0.49, + 0.591 + ], + "angle": 0, + "content": "Ivana Balazevic, Carl Allen, and Timothy Hospedales. 2019. TuckER: Tensor factorization for knowledge graph completion. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5185-5194, Hong Kong, China. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.6, + 0.49, + 0.679 + ], + "angle": 0, + "content": "Antoine Bordes, Nicolas Usunier, Alberto Garcia-Duran, Jason Weston, and Oksana Yakhnenko. 2013. Translating embeddings for modeling multi-relational data. In Advances in Neural Information Processing Systems, volume 26. Curran Associates, Inc." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.688, + 0.49, + 0.782 + ], + "angle": 0, + "content": "Samuel Broscheit, Kiril Gashteovski, Yanjie Wang, and Rainer Gemulla. 2020. Can we predict new facts with open knowledge graph embeddings? a benchmark for open link prediction. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2296-2308, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.79, + 0.49, + 0.883 + ], + "angle": 0, + "content": "Soumen Chakrabarti, Harkanwar Singh, Shubham Lohiya, Prachi Jain, and Mausam. 2022. Joint completion and alignment of multilingual knowledge graphs. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 11922-11938, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.891, + 0.49, + 0.919 + ], + "angle": 0, + "content": ". Chandrahas and Partha Talukdar. 2021. OKGIT: Open knowledge graph link prediction with implicit" + }, + { + "type": "list", + "bbox": [ + 0.116, + 0.485, + 0.49, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.529, + 0.086, + 0.885, + 0.14 + ], + "angle": 0, + "content": "types. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 2546-2559, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.149, + 0.885, + 0.268 + ], + "angle": 0, + "content": "Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. 2014. Learning phrase representations using RNN encoder-decoder for statistical machine translation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1724-1734, Doha, Qatar. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.278, + 0.885, + 0.343 + ], + "angle": 0, + "content": "Eunsol Choi, Jennimaria Palomaki, Matthew Lamm, Tom Kwiatkowski, Dipanjan Das, and Michael Collins. 2021. Decontextualization: Making sentences stand-alone. Transactions of the Association for Computational Linguistics, 9:447-461." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.354, + 0.885, + 0.433 + ], + "angle": 0, + "content": "Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzman, Edouard Grave, Myle Ott, Luke Zettle-moyer, and Veselin Stoyanov. 2020. Unsupervised cross-lingual representation learning at scale. In ACL Conference, pages 8440-8451." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.444, + 0.885, + 0.522 + ], + "angle": 0, + "content": "Nicola De Cao, Ledell Wu, Kashyap Popat, Mikel Artetxe, Naman Goyal, Mikhail Plekhanov, Luke Zettlemoyer, Nicola Cancedda, Sebastian Riedel, and Fabio Petroni. 2022. Multilingual autoregressive entity linking. Transactions of the Association for Computational Linguistics, 10:274-290." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.533, + 0.885, + 0.65 + ], + "angle": 0, + "content": "Tim Dettmers, Pasquale Minervini, Pontus Stenetorp, and Sebastian Riedel. 2018. Convolutional 2d knowledge graph embeddings. In Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence and Thirtieth Innovative Applications of Artificial Intelligence Conference and Eighth AAAI Symposium on Educational Advances in Artificial Intelligence, AAAI'18/IAAI'18/EAAI'18. AAAI Press." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.661, + 0.885, + 0.78 + ], + "angle": 0, + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.789, + 0.885, + 0.869 + ], + "angle": 0, + "content": "Vladimir Dobrovolskii. 2021. Word-level coreference resolution. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7670-7675, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.879, + 0.885, + 0.919 + ], + "angle": 0, + "content": "Oren Etzioni, Anthony Fader, Janara Christensen, Stephen Soderland, and Mausam. 2011. Open information extraction: The second generation. In *IJCAI*" + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.885, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.517, + 0.941 + ], + "angle": 0, + "content": "206" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.135, + 0.086, + 0.49, + 0.139 + ], + "angle": 0, + "content": "2011, Proceedings of the 22nd International Joint Conference on Artificial Intelligence, Barcelona, Catalonia, Spain, July 16-22, 2011, pages 3-10. IJ-CAI/AAAI." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.15, + 0.49, + 0.229 + ], + "angle": 0, + "content": "Anthony Fader, Stephen Soderland, and Oren Etzioni. 2011. Identifying relations for open information extraction. In Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 1535-1545, Edinburgh, Scotland, UK. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.239, + 0.488, + 0.293 + ], + "angle": 0, + "content": "Luis Galárraga, Geremy Heitz, Kevin Murphy, and Fabian M. Suchanek. 2014. Canonicalizing open knowledge bases. New York, NY, USA. Association for Computing Machinery." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.303, + 0.488, + 0.393 + ], + "angle": 0, + "content": "Kiril Gashteovski, Rainer Gemulla, and Luciano del Corro. 2017. MinIE: Minimizing facts in open information extraction. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2630-2640, Copenhagen, Denmark. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.405, + 0.488, + 0.471 + ], + "angle": 0, + "content": "Kiril Gashteovski, Sebastian Wanner, Sven Hertling, Samuel Broscheit, and Rainer Gemulla. 2019. Opiec: An open information extraction corpus. In Proceedings of the Conference on Automatic Knowledge Base Construction (AKBC)." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.481, + 0.488, + 0.587 + ], + "angle": 0, + "content": "Swapnil Gupta, Sreyash Kenkre, and Partha Talukdar. 2019. CaRe: Open knowledge graph embeddings. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 378-388, Hong Kong, China. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.597, + 0.488, + 0.689 + ], + "angle": 0, + "content": "Bosung Kim, Taesuk Hong, Youngjoong Ko, and Jungyun Seo. 2020. Multi-task learning for knowledge graph completion with pre-trained language models. In Proceedings of the 28th International Conference on Computational Linguistics, pages 1737-1743, Barcelona, Spain (Online). International Committee on Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.7, + 0.488, + 0.766 + ], + "angle": 0, + "content": "Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.776, + 0.488, + 0.856 + ], + "angle": 0, + "content": "Vid Kocijan and Thomas Lukasiewicz. 2021. Knowledge base completion meets transfer learning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP), Punta Cana, Dominican Republic. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.866, + 0.488, + 0.919 + ], + "angle": 0, + "content": "Keshav Kolluru, Vaibhav Adlakha, Samarth Aggarwal, Mausam, and Soumen Chakrabarti. 2020. OpenIE6: Iterative Grid Labeling and Coordination Analysis for Open Information Extraction. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.53, + 0.086, + 0.884, + 0.139 + ], + "angle": 0, + "content": "the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3748-3761, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.148, + 0.884, + 0.254 + ], + "angle": 0, + "content": "Keshav Kolluru, Muqeeth Mohammed, Shubham Mittal, Soumen Chakrabarti, and Mausam. 2022. Alignment-augmented consistent translation for multilingual open information extraction. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2502-2517, Dublin, Ireland. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.262, + 0.884, + 0.354 + ], + "angle": 0, + "content": "Justin Lovelace and Carolyn Rosé. 2022. A framework for adapting pre-trained language models to knowledge graph completion. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 5937-5955, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.363, + 0.884, + 0.468 + ], + "angle": 0, + "content": "Xin Lv, Yankai Lin, Yixin Cao, Lei Hou, Juanzi Li, Zhiyuan Liu, Peng Li, and Jie Zhou. 2022. Do pre-trained models benefit knowledge graph completion? a reliable evaluation and a reasonable approach. In Findings of the Association for Computational Linguistics: ACL 2022, pages 3570-3581, Dublin, Ireland. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.478, + 0.883, + 0.518 + ], + "angle": 0, + "content": "Mausam. 2016. Open information extraction systems and downstream applications. In International Joint Conference on Artificial Intelligence." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.526, + 0.883, + 0.554 + ], + "angle": 0, + "content": "MediaWiki. 2021. Api:langlinks — mediawiki. [Online; accessed 02-April-2022]." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.562, + 0.883, + 0.642 + ], + "angle": 0, + "content": "Jeffrey Pennington, Richard Socher, and Christopher Manning. 2014. GloVe: Global vectors for word representation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1532-1543, Doha, Qatar. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.65, + 0.883, + 0.73 + ], + "angle": 0, + "content": "Peng Qi, Yuhao Zhang, Yuhui Zhang, Jason Bolton, and Christopher D. Manning. 2020. Stanza: A Python natural language processing toolkit for many human languages. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.738, + 0.883, + 0.818 + ], + "angle": 0, + "content": "Youngbin Ro, Yukyung Lee, and Pilsung Kang. 2020. Multi^2OIE: Multilingual open information extraction based on multi-head attention with BERT. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 1107-1117, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.826, + 0.883, + 0.918 + ], + "angle": 0, + "content": "Théo Trouillon, Johannes Welbl, Sebastian Riedel, Eric Gaussier, and Guillaume Bouchard. 2016. Complex embeddings for simple link prediction. In Proceedings of The 33rd International Conference on Machine Learning, volume 48 of Proceedings of Machine Learning Research, pages 2071-2080, New York, New York, USA. PMLR." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.884, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.516, + 0.941 + ], + "angle": 0, + "content": "207" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.179 + ], + "angle": 0, + "content": "Shikhar Vashishth, Prince Jain, and Partha Talukdar. 2018. CESI: Canonicalizing open knowledge bases using embeddings and side information. In Proceedings of the 2018 World Wide Web Conference, WWW '18, pages 1317-1327, Republic and Canton of Geneva, Switzerland. International World Wide Web Conferences Steering Committee." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.188, + 0.49, + 0.281 + ], + "angle": 0, + "content": "Liang Wang, Wei Zhao, Zhuoyu Wei, and Jingming Liu. 2022. SimKGC: Simple contrastive knowledge graph completion with pre-trained language models. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4281-4294, Dublin, Ireland. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.29, + 0.49, + 0.357 + ], + "angle": 0, + "content": "Ralph Weischedel, Martha Palmer, Mitchell Marcus, Eduard Hovy, Sameer Pradhan, Lance Ramshaw, Nianwen Xue, Ann Taylor, Jeff Kaufman, and Michelle Franchini. 2013. Ontonotes release 5.0. In Linguistic Data Consortium, Philadelphia, PA." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.366, + 0.49, + 0.47 + ], + "angle": 0, + "content": "Guillaume Wenzek, Marie-Anne Lachaux, Alexis Conneau, Vishrav Chaudhary, Francisco Guzmán, Armand Joulin, and Edouard Grave. 2020. CCNet: Extracting high quality monolingual datasets from web crawl data. In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 4003-4012, Marseille, France. European Language Resources Association." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.481, + 0.49, + 0.586 + ], + "angle": 0, + "content": "Shijie Wu and Mark Dredze. 2019. Beto, bentz, becas: The surprising cross-lingual effectiveness of BERT. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 833–844, Hong Kong, China. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.596, + 0.49, + 0.687 + ], + "angle": 0, + "content": "Patrick Xia and Benjamin Van Durme. 2021. Moving on from OntoNotes: Coreference resolution model transfer. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5241-5256, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.697, + 0.49, + 0.816 + ], + "angle": 0, + "content": "Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. 2021. mT5: A massively multilingual pre-trained text-to-text transformer. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 483-498, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.826, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Zdeněk Žabokrtský, Miloslav Konopík, Anna Nedoluzhko, Michal Novák, Maciej Ogrodniczuk, Martin Popel, Ondřej Pražák, Jakub Sido, Daniel Zeman, and Yilun Zhu. 2022. Findings of the shared task on multilingual coreference resolution. In Proceedings of the CRAC 2022 Shared Task on Multilingual Coreference Resolution, pages 1-17," + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.529, + 0.086, + 0.882, + 0.114 + ], + "angle": 0, + "content": "Gyeongju, Republic of Korea. Association for Computational Linguistics." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.516, + 0.941 + ], + "angle": 0, + "content": "208" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.125, + 0.08, + 0.873, + 0.118 + ], + "angle": 0, + "content": "mOKB6: A Multilingual Open Knowledge Base Completion Benchmark (Appendix)" + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.129, + 0.303, + 0.143 + ], + "angle": 0, + "content": "A Dataset Curation" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.156, + 0.488, + 0.284 + ], + "angle": 0, + "content": "As discussed in Section 3, we construct mOKB6 dataset in three stages after extracting the Wikipedia articles (using WikiExtractor7) from the Wikidump of April 02, 2022. We run our construction pipeline (as shown in Figure 1) for all six languages on a single V100 (32 GB) GPU, which required 14 hours of computation to create mOKB6 dataset." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.287, + 0.49, + 0.575 + ], + "angle": 0, + "content": "In the first stage, we keep the sentences containing at least 6 and at most 50 tokens since we find that most of the short sentences are headings or sub-headings present in Wikipedia articles, and very long sentences can't be input to GEN2OIE (in second stage) due to maximum sequence length constraint of 1024 in mT5 (Xue et al., 2021) based GEN2OIE. This filtering step discards \\(18.9\\%\\) of sentences on an average in all six languages. We use Stanza (Qi et al., 2020) to perform sentence- and word-segmentation on Wikipedia articles in all six languages. After filtering the sentences, the articles are processed for coreference resolution using XLM-R (Conneau et al., 2020) encoder based wlcoref (Dobrovolskii, 2021), followed by replacing the coreferent cluster mentions with their canonical cluster name using the heuristic discussed in Section 3." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.578, + 0.489, + 0.673 + ], + "angle": 0, + "content": "In the second stage, the coreference resolved articles are passed through GEN2OIE to get the Open IE triples. The confidence scores for these triples are computed using label rescoring, for which we refer the readers to Kolluru et al. (2022) for more details." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.675, + 0.49, + 0.868 + ], + "angle": 0, + "content": "Finally, in the last stage, we apply various filters, adapted from Gashteovski et al. (2019), to remove triples that are of no interest to Open KBC task, like the triples: (1) having any of its argument or relation empty, (2) containing more than 10 tokens in any of its arguments or relation, (3) having confidence score less than 0.3, (4) containing pronouns (found using Stanza) in its arguments, (5) having same subject and object (i.e. self loops), and (6) that are duplicates. These filters keep \\(91.6\\%\\) of the triples obtained from stage 2 in all six languages." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.13, + 0.885, + 0.306 + ], + "angle": 0, + "content": "Further in the last stage, in order to create a dense Open KB containing minimum noise and maximum facts about the entities, we keep the triples having the Wikipedia article's title as either the subject phrase or object phrase and discard the rest. We do this by finding all the coreference clusters (of entity mentions) that contain the titles, then get the entities, or cluster names, of those clusters using the heuristic discussed in section 3, and keep those triples that contain these cluster names. This filtering step retains \\(23.6\\%\\) of the triples." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.32, + 0.615, + 0.336 + ], + "angle": 0, + "content": "B Metrics" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.347, + 0.883, + 0.54 + ], + "angle": 0, + "content": "We follow the previous works (Wang et al., 2022) on the evaluation methodology of Open KBC task and apply it to the multilingual Open KBC task, containing facts in multiple languages. Given an Open KB, containing a finite set of entities and open relations, the KGE model answers forward and backward queries of the form \\((s,r,?)\\) and \\((?,r,o)\\) respectively. The model ranks all the entities based on their correctness with, say, \\(s\\) and \\(r\\) in the forward query. Further, the evaluation is in filtered setting, where the other known correct answers, apart from \\(o\\), are removed from rank list." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.542, + 0.884, + 0.669 + ], + "angle": 0, + "content": "The commonly used evaluation metrics are hits at rank N (H@N), where \\( N \\) is a natural number, and mean reciprocal rank (MRR). Suppose, the model ranks \\( o \\) at \\( R \\) among all entities. Then, H@N measures how many times \\( R \\) is less than or equal to \\( N \\). MRR is the average of reciprocal ranks \\( \\left( \\frac{1}{R} \\right) \\). Both, H@N and MRR, are computed as average over both forms of queries over the full test set." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.684, + 0.88, + 0.7 + ], + "angle": 0, + "content": "C Knowledge Graph Embedding Models" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.71, + 0.885, + 0.919 + ], + "angle": 0, + "content": "SimKGC (Wang et al., 2022) is a text-based KGE model that uses two unshared pretrained BERT models (Devlin et al., 2019) for encoding (subject phrase; relation phrase) and object phrase separately. GRU-ConvE (Kocijan and Lukasiewicz, 2021) encodes both the relation phrase and argument phrase from their surface forms using two unshared GRU (Cho et al., 2014). CaRe (Gupta et al., 2019) learns separate embeddings for each argument phrase and uses a bi-directional GRU to encode the relation phrase from its surface form. Both, GRU-ConvE and CaRe, are initialised with Glove embeddings (Pennington et al., 2014)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.113, + 0.892, + 0.411, + 0.918 + ], + "angle": 0, + "content": "\\(^{7}\\)https://github.com/samuelbroscheit/wikiextractor-wikimentions" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.928, + 0.517, + 0.941 + ], + "angle": 0, + "content": "209" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.081, + 0.885, + 0.157 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.112, + 0.165, + 0.884, + 0.21 + ], + "angle": 0, + "content": "Figure 4: Previous Open KB construction pipelines like Gashteovski et al. (2019) (shown by green arrows) lack coreference resolution system, which result in filtering important facts like (Barack Obama; returned to Honolulu, Hawaii in; 1971). Our pipeline (shown by blue arrows) increases the coverage of facts due to mCoref system." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.234, + 0.491, + 0.378 + ], + "angle": 0, + "content": "To choose the best model for our experiments (Table 3, Figure 3), we train the recent knowledge graph embedding (KGE) models — CaRe., GRUConvE and SimKGC on the English Open KB in mOKB6. We report performance in Table 4 using the three metrics: hits at rank 1 (H@1), hits at 10 (H@10), and mean reciprocal rank (MRR). We find that SimKGC with BERT encoder outperforms the other two models." + }, + { + "type": "table", + "bbox": [ + 0.155, + 0.387, + 0.449, + 0.484 + ], + "angle": 0, + "content": "
H@1H@10MRR
CaRe6.611.38.3
GRU-ConvE12.427.817.8
SimKGC (BERT)16.140.024.3
SimKGC (mBERT)14.838.722.8
SimKGC (XLM-R)13.835.821.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.493, + 0.49, + 0.538 + ], + "angle": 0, + "content": "Table 4: Performance \\((\\%)\\) of the KGE models on the English test set in mOKB6 dataset. The reported numbers are an average of three runs using different seeds." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.557, + 0.49, + 0.751 + ], + "angle": 0, + "content": "Since BERT supports only English language, we replace BERT in SimKGC with multilingual pretrained language models like mBERT (Devlin et al., 2019) or XLM-R (Conneau et al., 2020), to extend SimKGC model to other languages. We find in Table 4 that SimKGC with mBERT is better than with XLM-R by \\(2.9\\%\\) H@10 and \\(1.5\\%\\) MRR, possibly because mBERT (and mOKB6) uses Wikipedia while XLM-R uses CommonCrawl (Wenzek et al., 2020) during pre-training. Thus, we use SimKGC with mBERT as the underlying encoder to run our experiments for all the languages." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.764, + 0.405, + 0.781 + ], + "angle": 0, + "content": "D KGE Model Training Details" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.791, + 0.491, + 0.919 + ], + "angle": 0, + "content": "We use the code from official repositories of the KGE models — SimKGC (Wang et al., 2022), GRU-ConvE (Kocijan and Lukasiewicz, 2021), and CaRe (Gupta et al., 2019) for our experiments. The models are trained using Adam optimizer (Kingma and Ba, 2015) on a single A100 (40 GB) GPU with three different random seeds and we report the average of three evaluation runs." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.234, + 0.885, + 0.41 + ], + "angle": 0, + "content": "We do not perform hyperparameter search trials, except for batch size, and use the default hyperparameters from the respective codes of KGE models (see Table 5). We use early stopping to find the best model checkpoints based on HITS@1. The dev set is different for each baseline: MONO, TRANS, MONO+TRANS, and UNION+TRANS use individual language's dev set, whereas UNION w/o En and UNION use the English dev set. We report the performance of baseline models on the dev sets in Table 9 and Table 10." + }, + { + "type": "table", + "bbox": [ + 0.528, + 0.419, + 0.866, + 0.521 + ], + "angle": 0, + "content": "
HyperparameterSimKGCGRU-ConvECaRe
#epochs100500500
#patience epochs101010
learning rate3e-53e-41e-3
dropout0.10.30.5
batch size2561024128
additive margin0.02N/AN/A
" + }, + { + "type": "table_caption", + "bbox": [ + 0.536, + 0.53, + 0.855, + 0.545 + ], + "angle": 0, + "content": "Table 5: Hyperparameters of the KGE models." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.563, + 0.884, + 0.676 + ], + "angle": 0, + "content": "We provide the number of trainable parameters of each KGE model in Table 6. Based on the batch size and model size, different experiments consume different GPU hours. To train on English Open KB (in mOKB6 dataset), CaRe and GRU-ConvE models took 2.5 hours and 0.5 hours, respectively, whereas SimKGC takes nearly 1 hour of GPU time." + }, + { + "type": "table", + "bbox": [ + 0.558, + 0.686, + 0.835, + 0.776 + ], + "angle": 0, + "content": "
KGE model#trainable parameters
CaRe12,971,423
GRU-ConvE12,085,523
SimKGC (BERT)216,620,545
SimKGC (mBERT)355,706,881
SimKGC (XLM-R)1,119,780,865
" + }, + { + "type": "table_caption", + "bbox": [ + 0.509, + 0.785, + 0.883, + 0.813 + ], + "angle": 0, + "content": "Table 6: Number of trainable parameters in the KGE models." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.928, + 0.517, + 0.941 + ], + "angle": 0, + "content": "210" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.12, + 0.082, + 0.885, + 0.202 + ], + "angle": 0, + "content": "
EnglishHindiTeluguSpanishPortugueseChinese
H@1H@10MRRH@1H@10MRRH@1H@10MRRH@1H@10MRRH@1H@10MRRH@1H@10MRR
English68.497.178.83.417.28.31.611517.850.728.61744.6265.422.311.1
Hindi1942.226.780.699.588.32.412.55.912.33619.912.333.919.75.321.910.8
Telugu19.542.227.24.318.79.474.499.584.210.935.418.910.73418.54.721.410.1
Spanish27.960.438.84.117.88.91.810.75.18410090.337.67450.16.524.912.8
Portuguese27.858.738.24.418.29.31.710.55.141.578.553.684.299.990.86.62613.2
Chinese22.148.430.63.518.58.81.812.25.414.842.824.215.741.624.181.699.889.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.211, + 0.884, + 0.24 + ], + "angle": 0, + "content": "Table 7: Performance (%) of the six MEMORIZE models, which have been trained on each language's test set and tested on all the test sets in mOKB6 dataset." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.264, + 0.315, + 0.281 + ], + "angle": 0, + "content": "E Contextual Triples" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.29, + 0.489, + 0.402 + ], + "angle": 0, + "content": "Open IE triples are of various kinds and not all of them can be used for Open KBC task. Various filtering steps are used to remove some of these in data curation (Section 3). We define contextual triples as another kind of noisy triples, which are specific to, and are not interpretable out of, the context of text from which they are extracted." + }, + { + "type": "text", + "bbox": [ + 0.18, + 0.415, + 0.421, + 0.427 + ], + "angle": 0, + "content": "(Max Born; continued; scientific work)" + }, + { + "type": "text", + "bbox": [ + 0.18, + 0.427, + 0.422, + 0.44 + ], + "angle": 0, + "content": "(Robb Gravett; won; the championship)" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.439, + 0.451, + 0.451 + ], + "angle": 0, + "content": "(George Herbert Walker Bush; was; out of touch)" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.451, + 0.386, + 0.463 + ], + "angle": 0, + "content": "(Christianity; is; dominant)" + }, + { + "type": "table_caption", + "bbox": [ + 0.165, + 0.474, + 0.437, + 0.489 + ], + "angle": 0, + "content": "Table 8: Examples of contextual triples." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.505, + 0.49, + 0.585 + ], + "angle": 0, + "content": "From the first two triples in Table 8, it is unclear which scientific work Max Born continued, or which championship Robb Gravett has won. The last two triples are too specific to the context and contain no factual information." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.516, + 0.941 + ], + "angle": 0, + "content": "211" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.12, + 0.223, + 0.88, + 0.338 + ], + "angle": 0, + "content": "
English (En)Hindi (Hi)Telugu (Te)Spanish (Es)Portuguese (Pt)Chinese (Zh)
H@1H@10MRRH@1H@10MRRH@1H@10MRRH@1H@10MRRH@1H@10MRRH@1H@10MRR
MONO16.238.723.918.239.425.98.52012.517.336.623.717.639.625.310.831.917.8
TRANS---8.123.713.53.315.47.512.933.620.312.637.220.6520.810.3
MONO+TRANS---20.843.228.67.824.813.420.24628.82145.929.210.630.116.7
UNION19.939.626.414.538.222.45.92010.619.843.227.919.743.82811.23318.8
UNION w/o En5.819.510.615.439.323.36.320.511.119.441.626.416.942.925.911.33318.4
UNION+TRANS---20.844.928.87.327.11421.445.329.619.449.129.16.93115.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.186, + 0.348, + 0.809, + 0.364 + ], + "angle": 0, + "content": "Table 9: Performance (%) of SimKGC on the dev sets (of mOKB6 dataset) in six languages." + }, + { + "type": "table", + "bbox": [ + 0.354, + 0.654, + 0.645, + 0.749 + ], + "angle": 0, + "content": "
H@1H@10MRR
CaRe7.111.18.5
GRU-ConvE16.831.522.1
SimKGC (BERT)20.340.127.1
SimKGC (mBERT)16.238.723.9
SimKGC (XLM-R)1736.623.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.165, + 0.759, + 0.831, + 0.774 + ], + "angle": 0, + "content": "Table 10: Performance (%) of the KGE models on dev set of English Open KB in mOKB6 dataset." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.516, + 0.941 + ], + "angle": 0, + "content": "212" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.133, + 0.085, + 0.435, + 0.1 + ], + "angle": 0, + "content": "ACL 2023 Responsible NLP Checklist" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.108, + 0.323, + 0.123 + ], + "angle": 0, + "content": "A For every submission:" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.127, + 0.533, + 0.158 + ], + "angle": 0, + "content": "A1. Did you describe the limitations of your work? 8" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.171, + 0.583, + 0.203 + ], + "angle": 0, + "content": "A2. Did you discuss any potential risks of your work? There are no potential risks of our work to our knowledge." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.212, + 0.697, + 0.244 + ], + "angle": 0, + "content": "A3. Do the abstract and introduction summarize the paper's main claims?" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.256, + 0.67, + 0.289 + ], + "angle": 0, + "content": "A4. Have you used AI writing assistants when working on this paper? Left blank." + }, + { + "type": "list", + "bbox": [ + 0.13, + 0.127, + 0.697, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.299, + 0.49, + 0.317 + ], + "angle": 0, + "content": "B Did you use or create scientific artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.322, + 0.162, + 0.335 + ], + "angle": 0, + "content": "3,4" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.347, + 0.531, + 0.379 + ], + "angle": 0, + "content": "B1. Did you cite the creators of artifacts you used? 3,4" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.39, + 0.779, + 0.422 + ], + "angle": 0, + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Abstract" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.434, + 0.882, + 0.514 + ], + "angle": 0, + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.525, + 0.882, + 0.589 + ], + "angle": 0, + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.599, + 0.882, + 0.647 + ], + "angle": 0, + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? 3" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.658, + 0.882, + 0.753 + ], + "angle": 0, + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. 3" + }, + { + "type": "list", + "bbox": [ + 0.13, + 0.347, + 0.882, + 0.753 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.765, + 0.494, + 0.782 + ], + "angle": 0, + "content": "C Did you run computational experiments?" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.789, + 0.147, + 0.8 + ], + "angle": 0, + "content": "4" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.812, + 0.882, + 0.862 + ], + "angle": 0, + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Appendix D" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.869, + 0.878, + 0.894 + ], + "angle": 0, + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.516, + 0.941 + ], + "angle": 0, + "content": "213" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.13, + 0.084, + 0.88, + 0.133 + ], + "angle": 0, + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Appendix D" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.143, + 0.88, + 0.208 + ], + "angle": 0, + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? \nAppendix D" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.218, + 0.88, + 0.283 + ], + "angle": 0, + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Appendix A, D" + }, + { + "type": "list", + "bbox": [ + 0.13, + 0.084, + 0.88, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.293, + 0.878, + 0.329 + ], + "angle": 0, + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? 3" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.34, + 0.88, + 0.386 + ], + "angle": 0, + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.4, + 0.88, + 0.464 + ], + "angle": 0, + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.475, + 0.88, + 0.54 + ], + "angle": 0, + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.55, + 0.873, + 0.582 + ], + "angle": 0, + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.593, + 0.88, + 0.641 + ], + "angle": 0, + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Not applicable. Left blank." + }, + { + "type": "list", + "bbox": [ + 0.129, + 0.34, + 0.88, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.929, + 0.516, + 0.941 + ], + "angle": 0, + "content": "214" + } + ] +] \ No newline at end of file diff --git a/2023/mOKB6_ A Multilingual Open Knowledge Base Completion Benchmark/70c57a7f-70d1-4b2e-a713-7ef674b2ac69_origin.pdf b/2023/mOKB6_ A Multilingual Open Knowledge Base Completion Benchmark/70c57a7f-70d1-4b2e-a713-7ef674b2ac69_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d1de7efd909b7aa6044c9286d071fca710c7faf5 --- /dev/null +++ b/2023/mOKB6_ A Multilingual Open Knowledge Base Completion Benchmark/70c57a7f-70d1-4b2e-a713-7ef674b2ac69_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1bb61c6a3b059b48b982e44f34bf5da0698a98dd6056cbd8cb0583c87e59ad7c +size 616795 diff --git a/2023/mOKB6_ A Multilingual Open Knowledge Base Completion Benchmark/full.md b/2023/mOKB6_ A Multilingual Open Knowledge Base Completion Benchmark/full.md new file mode 100644 index 0000000000000000000000000000000000000000..074bd883dfdc5285a62693f9c9fd8560bb19740f --- /dev/null +++ b/2023/mOKB6_ A Multilingual Open Knowledge Base Completion Benchmark/full.md @@ -0,0 +1,301 @@ +# mOKB6: A Multilingual Open Knowledge Base Completion Benchmark + +Shubham Mittal $^{\alpha \dagger}$ Keshay Kolluru $^{\beta \dagger}$ Soumen Chakrabarti $^{\gamma}$ Mausam $^{\alpha}$ + +$^{\alpha}$ Indian Institute of Technology Delhi + +$^{\beta}$ KnowDis AI, New Delhi + +$\gamma$ Indian Institute of Technology Bombay + +shubhamiitd18@gmail.com, keshav.kolluru@gmail.com + +soumen@cse.iitb.ac.in, mausam@cse.iitd.ac.in + +# Abstract + +Automated completion of open knowledge bases (Open KBs), which are constructed from triples of the form (subject phrase, relation phrase, object phrase), obtained via open information extraction (Open IE) system, are useful for discovering novel facts that may not be directly present in the text. However, research in Open KB completion (Open KBC) has so far been limited to resource-rich languages like English. Using the latest advances in multilingual Open IE, we construct the first multilingual Open KBC dataset, called mOKB6, containing facts from Wikipedia in six languages (including English). Improving the previous Open KB construction pipeline by doing multilingual coreference resolution and keeping only entity-linked triples, we create a dense Open KB. We experiment with several models for the task and observe a consistent benefit of combining languages with the help of shared embedding space as well as translations of facts. We also observe that current multilingual models struggle to remember facts seen in languages of different scripts. + +# 1 Introduction + +Open information extraction (Open IE) systems (Mausam, 2016) such as ReVerb (Etzioni et al., 2011) and OpenIE6 (Kolluru et al., 2020) can extract triples, or facts, of the form (subject phrase, relation phrase, object phrase), which can be denoted as $(s,r,o)$ , from text (e.g., Wikipedia articles) without using any pre-defined ontology. Open knowledge base (Open KB) is constructed using these Open IE triples where the subject phrases and object phrases are nodes and relation phrases are labels on edges connecting the nodes in the graph. Open knowledge base completion (Open KBC) is + +the task of discovering new links between nodes using the graph structure of the Open KB. Knowledge graph embedding (KGE) models are typically used for the Open KBC task, where they are asked to answer questions of the form $(s,r,?)$ and $(?,r,o)$ . + +Research in Open KBC has been restricted to English (Vashishth et al., 2018) due to lack of Open KBs in other languages. We aim to study multilingual Open KBC, with the motivation that the information available in high resource languages like English may help when inferring links in Open KBs that use low resource languages like Telugu. Moreover, intuitively, if all the information in different languages can be pooled together, then it may help the model learn better, and allow information flow across Open KBs in different languages. + +We design the first multilingual Open KB construction pipeline (shown in Figure 1) using a multilingual Open IE system, GEN2OIE (Kolluru et al., 2022). We find that coreference resolution is missing in existing Open KB construction (Gashteovski et al., 2019) but is important for increasing the coverage of facts (as described in Figure 4). We re-train a recent coref model (Dobrovolskii, 2021) using XLM-R (Conneau et al., 2020) as the underlying multilingual encoder and add it to our pipeline. For constructing a high quality test set, we use 988 manually verified facts in English. For extending to other languages, we automatically translate English facts. The dataset thus constructed, called mOKB6, contains 42K facts in six languages: English, Hindi, Telugu, Spanish, Portuguese, and Chinese. + +We report the first baselines for multilingual Open KBC task. We find that they are able to benefit from information in multiple languages when compared to using facts from a single language. Translations of Open KB facts also help the models. However, we notice that although the multilingual KGE models learn facts in a particular language, they struggle to remember the same fact, when queried in another language with different script. + +# 2 Related Work + +Multilingual Open KBC datasets are absent in literature to the best of our knowledge, although multiple English Open KBC datasets are available. OLPBench (Broscheit et al., 2020), derived from OPIEC (Gashteovski et al., 2019), is a large-scale Open KBC dataset that contains 30M triples and is constructed from English Wikipedia using MinIE system (Gashteovski et al., 2017). The evaluation data contains 10K triples randomly sampled from 1.25M linked triples. ReVerb45K (Vashisth et al., 2018) and ReVerb20K (Galarraga et al., 2014) are smaller Open KBC datasets constructed from Clueweb09 corpus $^2$ using ReVerb Open IE system (Fader et al., 2011). Both the datasets keep only those tuples in which both the subject phrase and object phrase link to a finite set of Freebase entities. + +Multilingual Open IE (mOpenIE) systems like GEN2OIE (Kolluru et al., 2022) and Multi $^{2}$ OIE (Ro et al., 2020) enable extracting facts from multiple languages. We use the GEN2OIE model for constructing mOKB6 dataset as it is trained with language-specific facts transferred from English, while Multi $^{2}$ OIE relies on zero-shot transfer for languages other than English. + +Knowledge Graph Embedding (KGE) Models: Conventional KGE models like TransE (Bordes et al., 2013), ComplEx (Trouillon et al., 2016), ConvE (Dettmers et al., 2018), and TuckER (Balazevic et al., 2019) have been used for Open KBC task (Gupta et al., 2019; Broscheit et al., 2020; Chandrahas and Talukdar, 2021; Kocijan and Lukasiewicz, 2021). Given a triple $(s,r,o)$ , these models encode the subject phrase, relation phrase, and object phrase from free text, and pass the encodings to a triple-scoring function, which is optimized using binary cross entropy loss. ComplEx has also been used for multilingual closed KBC task (Chakrabarti et al., 2022). + +Pretrained language models like BERT (Devlin et al., 2019) have been used in KGE models for the KBC task (Lovelace and Rosé, 2022; Lv et al., 2022; Chandrahas and Talukdar, 2021; Kim et al., 2020). SimKGC (Wang et al., 2022) is the state of the art KGE model on closed KBC task. It computes the score of a triple $(s, r, o)$ as the cosine similarity of the embeddings of $(s; r)$ and $(o)$ , computed using two separate pretrained BERT models without any weight sharing. + +# 3 Dataset Curation + +We aim to construct a dense multilingual Open KB that maximizes the information about a given real-world entity, which may be represented as multiple nodes across languages. Therefore, we consider those Wikipedia articles3 that are available in six languages: English, Hindi, Telugu, Spanish, Portuguese, and Chinese4. This will also help the model learn from facts in high resource language like English and answer queries in low resource language like Telugu. We work with 300 titles randomly sampled from the ones common among all six languages (found using MediaWiki-Langlinks (MediaWiki, 2021)). Thus, we extract facts from $6 \times 300$ Wikipedia articles. We discuss the three stages of our pipeline below. + +Stage 1 We first process each Wikipedia article through a coreference resolution system. Although language-specific end-to-end neural coref models have been developed (Žabokrtský et al., 2022; Xia and Van Durme, 2021), multilingual models that work on all our languages of interest are absent in the literature. Therefore, we retrain wl-coref (Dobrovolskii, 2021) with XLM-R (Conneau et al., 2020) on the English training data (available in OntoNotes (Weischedel et al., 2013)) that can work zero-shot for other languages. + +Coref models detect and cluster mentions, but do not identify a canonical cluster name, which is needed for standardizing all the mentions in the cluster. To find cluster names, entity linking systems such as mGENRE (De Cao et al., 2022) or Wikipedia hyperlinks can be used. However, we found that they result in low recall, particularly for low resource languages. Thus, we employ a heuristic to find the cluster name and replace each of the coreferent mentions with it. The score for each mention is represented by a tuple, computed as: Score(mention phrase) = (#proper nouns, #nouns, #numerals, #adjectives, #pronouns, #verbs). The tuple is ordered according to the importance of each field (POS tags) for the cluster name, which is determined empirically. Two tuples are compared index-wise with higher priority given to lower indices to determine the best scoring mention that is chosen as the canonical name (Table 1). + +Stage 2 We use GEN2OIE to extract Open IE triples from the coreference resolved sentences. + +![](images/d8cf19105aad3c532c70c6bfe2852b999abae3175fb081a4cd65e4a3f6442568.jpg) +Figure 1: Our three-staged multilingual Open KB construction pipeline for mOKB6. mCoref is multilingual coreference resolution system, having XLM-R (Conneau et al., 2020) encoder based wl-coref (Dobrovolskii, 2021), and mOpenIE is multilingual open information extraction system, consisting of GEN2OIE (Kolluru et al., 2022). + +
MentionsScoresCluster Name
Barack Obama(2,0,0,0,0,0)
Obama(1,0,0,0,0,0)Barack Obama
He(0,0,0,0,1,0)
+ +Stage 3 Similar to Gashteovski et al. (2019), we apply various filters to remove noisy triples that have empty or very long arguments, or have less confidence than 0.3 (as assigned by GEN2OIE). We further only keep triples that have the article's title as either the subject phrase or object phrase, to avoid generic or specific triples, valid only in the particular context. Examples of contextual triples (Choi et al., 2021) are discussed in Appendix E. See Appendix A for further data curation details. + +These automatically extracted triples form the train set of mOKB6. To form a high quality test set in six languages with limited access to experts in all languages, the test set is created in a semiautomatic way. We sample 1600 English triples from the train set (which are subsequently filtered) and manually remove noisy triples. We use inter-annotation agreement between two annotators to check if they both agree that the given triple is noisy or clean. With an agreement of $91\%$ , we retain 988 English triples, which we automatically translate to the other five languages. As illustrated in Figure 2, to translate a triple, we convert it to a sentence after removing tags and use Google translate for translating the triple-converted sentence to the remaining five languages. We observed high quality of translated triples, with $88\%$ satisfactory translations as determined by native-speakers of three languages on a set of 75 translated triples. To get the Open IE subject phrase, relation phrase and object phrase tags, we project the labels from the original English triple to the translated sentence using word alignments (Kolluru et al., 2022). Finally, we are left with 550 triples in each language after removing examples where some labels could + +not be aligned. We use these $6 \times 550$ triples as the test sets. The train and dev sets are created from the remaining triples in each language such that the dev set has 500 randomly sampled triples (Table 2). + +![](images/5ef6efdb8a343a6921e6766304a6e31bfc2b48d19ad192936565512a7cf2369b.jpg) +Figure 2: Method to translate Open IE triple using Google translate, and followed by label projection using word alignments (Kolluru et al., 2022). + +We analyse the entity overlap across languages and find that on an average, a test entity (which is present in either the subject phrase or object phrase of a test tuple) is present 17.73 times in English, 0.94 times in Hindi, 0.47 times in Telugu, 2.33 times in Spanish, 1.69 times in Portuguese, and 1.45 times in Chinese train set. + +Our construction pipeline improves over OPIEC in three ways: (1) we use a multilingual Open IE system, instead of an English-specific Open IE system like in OPIEC, enabling us to curate Open KBs in many languages, (2) we add a multilingual coreference resolution system in our pipeline, and (3) the English test triples are manually verified. Further, we manually evaluate and review the noise at each step of data curation in Section 4. + +Table 1: Parts of speech tags are used to find the canonical name of the coreferent cluster of entity mentions. + +
EnHiTeEsPtZh
#entity2063746253972565153045037
#relation787021771907282326442325
#train2019527861992396635283420
+ +Table 2: Statistics of individual Open KBs in mOKB6 in English (En), Hindi (Hi), Telugu (Te), Spanish (Es), Portuguese (Pt), and Chinese (Zh). The dev and test set for each Open KB contain 500 and 550 triples each. + +# 4 Noise Evaluation + +Curating an Open KB involves various stages and each stage induces its noise in the construction pipeline (Gashteovski et al., 2019). We manually evaluate the noise induced at each stage of our pipeline (Figure 1) and discuss the same in this section. We ask native speakers of four (out of six) languages - English, Hindi, Telugu, and Chinese to assess the output quality, or precision, of each stage as discussed below. + +In the first stage, we assess the performance of the coreference resolution system over Wikipedia articles. We find a high precision of $95.5\%$ in coref's mention clustering and $89.82\%$ accuracy in finding canonical cluster name (using the heuristic illustrated in Table 1), computed over 40 randomly sampled coref clusters (10 in each language). + +For evaluating the Open IE system, GEN2OIE, in the second stage, we mark an extraction of a sentence as correct if it has syntactically correct arguments and it is coherent with the sentence. We get an average precision of $63.4\%$ on 80 extractions (20 in each language). + +We evaluate the triples, or Open KB facts, at the last stage after passing through various noise-removing filters. Note that these triples also form the train set (and dev set) in mOKB6 dataset. We mark triples as correct when they contain real-world entities, and also, factual information about them. If the triple is very generic or contextual (see Appendix E), it is marked as incorrect. We find the train (and dev) set quality to be $69.3\%$ , averaged over 80 triples in four languages. + +# 5 Experiments + +Our experimental study on multilingual open KBC task investigates the following research questions: + +1. Does the KGE model benefit from facts in different languages? (Section 5.1) +2. Can translation help transfer among languages? (Section 5.2) +3. Does the KGE model remember facts seen across different languages? (Section 5.3) + +We use SimKGC model (Wang et al., 2022) with pretrained mBERT initialization to run our experiments, after comparing with recent KGE models (Appendix C). For evaluation, we use three metrics -hits at rank 1 (H@1), hits at rank 10 (H@10), and mean reciprocal rank (MRR). The formal definitions of them are provided in Appendix B. We discuss further model training details in Appendix D. + +# 5.1 Training on Multilingual Facts + +We train and compare monolingual model, called MONO, with multilingual models, UNION and UNION w/o En. In MONO, we train one model for each language using its respective Open KB, whereas in UNION, a single model is trained on six languages' Open KBs together. UNION outperforms MONO in all languages by an average of $4.6\%$ H@10 and $2.8\%$ MRR (see Table 3), which provides evidence of information flow across languages and the model benefits from it. + +To check the extent of flow from (high-resource) English to the other languages, we also train on the five languages except English, which we call UNION w/o En. We find UNION w/o En also outperforms MONO by $2.7\%$ H@10 and $1.2\%$ MRR over the five languages, hinting that interlingual transfer is more general and pervasive. + +# 5.2 Open KB Facts Translation + +Apart from relying only on multilingual transfer in the embedding space, we analyse the effect of using translated triples in the training of the KGE model. We translate the English training triples to the other five languages (Section 3) and train monolingual models using only the translated triples (TRANS). To leverage facts present in each language's Open KB, we make MONO+TRANS, where we add language-specific MONO data to the translated triples. Table 3 shows that MONO+TRANS is better than MONO by a large margin of $15.5\%$ H@1, $29.2\%$ H@10, and $20.0\%$ MRR, averaged over five languages. Also, MONO+TRANS improves over TRANS by $2.1\%$ H@10 and $2.0\%$ MRR, showcasing the importance of facts in each language's Open KBs. + +To effectively gain from transfer in both the embedding space as well as translation, we introduce $\text{UNION+TRANS}$ . We train one model for each language, on the combination of UNION triples and the translated train triples from English Open KB to that language. $\text{UNION+TRANS}$ is better than UNION by $25.9\%$ H@10 and $18.4\%$ MRR. This suggests that the model is able to benefit from English facts when they are translated to the query language, unlike in UNION where the English facts are present only in English. + +
English (En)Hindi (Hi)Telugu (Te)Spanish (Es)Portuguese (Pt)Chinese (Zh)
H@1H@10MRRH@1H@10MRRH@1H@10MRRH@1H@10MRRH@1H@10MRRH@1H@10MRR
MONO14.838.722.83.014.87.21.58.13.96.423.712.36.321.711.42.413.16.2
UNION w/o En5.721.510.92.915.47.41.810.24.98.127.814.56.726.112.93.215.57.5
UNION16.740.824.83.616.68.11.59.34.510.632.217.69.729.316.64.018.88.9
TRANS---20.547.629.78.728.715.523.250.632.420.550.730.514.039.422.5
MONO+TRANS---20.245.428.414.338.522.223.551.532.921.448.930.717.943.226.6
UNION+TRANS---23.349.732.315.138.523.123.952.433.423.552.133.116.943.626.0
+ +Table 3: Performance (%) of SimKGC model on mOKB6 dataset, comprising of Open KBs in six languages. MONO, TRANS, and MONO+TRANS are monolingual models trained only on facts of one language whereas UNION, UNION w/o En, and UNION+TRANS are multilingual models trained with facts from multiple languages. All reported numbers are an average of three runs using different seeds. Best scores are highlighted in bold. + +# 5.3 Cross-lingual Memorization + +Pretrained multilingual language models such as mBERT have demonstrated strong cross-lingual transfer capabilities (Wu and Dredze, 2019). We investigate cross-lingual memorization of the KGE model by showing facts in one language and querying the same facts in other five languages. For each language, $L$ , we take the UNION model and train it further on the test set of that language's Open KB, which we call MEMORIZE $_L$ model. Then, we test each MEMORIZE $_L$ model on the six test sets. Since the test sets (in mOKB6 dataset) of the different languages contain the same facts, this experiment allows us to investigate cross-lingual memorization. We provide the H@10 scores of MEMORIZE models in Figure 3 and the performance on other metrics (H@1 and MRR) is reported in Table 7. + +The model achieves at least $97\%$ H@10 when tested on the language used for training (diagonal). We observe that there is relatively good crosslingual memorization among languages that share the same script (Latin in English, Spanish, and Portuguese), but the model struggles to remember facts when seen in languages of different scripts. Many entities look similar in shared scripts, possibly leading to better information transfer. For example, the $\mathsf{MEMORIZE}_{En}$ achieves H@10 of $50.7\%$ in Spanish (Es) compared to $22.3\%$ in Chinese (Zh) and $11\%$ in Telugu (Te). + +# 6 Conclusion and Future Work + +We create and release the mOKB6 dataset, the first multilingual Open Knowledge Base Completion dataset with 42K facts in six languages: English, Hindi, Telugu, Spanish, Portuguese, and Chinese. Its construction uses multilingual coreference resolution, entity-mention cluster naming, multilingual open information extraction and various filtering + +![](images/a1dc8f82f5d3ed3389f4fadaa56129d4f2c29635a09c9311958a3e04644f4744.jpg) +Figure 3: Performance (H@10) of MEMORIZE models. Row $L$ shows the performance of $\text{MEMORIZE}_L$ model across the test sets of all languages (columns). For example, the performance of $\text{MEMORIZE}_{En}$ when tested on English (En) is $97.1\%$ H@10, and $\text{MEMORIZE}_{En}$ when tested on Spanish (Es) gives $50.7\%$ H@10. We find relatively good cross-lingual transfer among languages that use same script (Latin in English, Spanish and Portuguese) compared to those using different scripts (English, Hindi, Telugu and Chinese). + +steps to improve the quality of the extracted facts. We also report the first baselines on the task using the existing state of the art KGE models trained with facts from different languages using various augmentation strategies. + +Our work opens many important research questions: (1) Can we develop better strategies to combine facts in different languages? (2) Can we build models that achieve strong information transfer across unrelated languages with same or different scripts? (3) Can we train the neural model to ignore contextual triples (Appendix E), thus improving overall performance? and (4) Can tying the same entities across various languages help the model generalize better? We leave these questions to be addressed in future work. + +# 7 Acknowledgements + +Keshav was supported by TCS Research Fellowship during his PhD. Mausam is supported by grants from Huawei, Google, Verisk and IBM, and a Jai Gupta Chair Fellowship. He also acknowledges Google and Yardi School of AI travel grants. Soumen is partly supported by a Jagadish Bose Fellowship and a grant from Cisco. We thank IIT Delhi HPC facility for compute resources. + +# 8 Limitations + +Although multilingual, the constructed open KB is limited to the sampling of the chosen six languages. We do not know how well the system will generalize to various language families that have not been considered here. Further, even among the languages considered, the performance of even the best-performing systems, as measured through $\mathrm{H@1}$ is still in the low 20's. Therefore the models are not yet ready to be deployed for real-world applications. + +# References + +Ivana Balazevic, Carl Allen, and Timothy Hospedales. 2019. TuckER: Tensor factorization for knowledge graph completion. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5185-5194, Hong Kong, China. Association for Computational Linguistics. +Antoine Bordes, Nicolas Usunier, Alberto Garcia-Duran, Jason Weston, and Oksana Yakhnenko. 2013. Translating embeddings for modeling multi-relational data. In Advances in Neural Information Processing Systems, volume 26. Curran Associates, Inc. +Samuel Broscheit, Kiril Gashteovski, Yanjie Wang, and Rainer Gemulla. 2020. Can we predict new facts with open knowledge graph embeddings? a benchmark for open link prediction. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2296-2308, Online. Association for Computational Linguistics. +Soumen Chakrabarti, Harkanwar Singh, Shubham Lohiya, Prachi Jain, and Mausam. 2022. Joint completion and alignment of multilingual knowledge graphs. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 11922-11938, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics. +. Chandrahas and Partha Talukdar. 2021. OKGIT: Open knowledge graph link prediction with implicit + +types. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 2546-2559, Online. Association for Computational Linguistics. +Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. 2014. Learning phrase representations using RNN encoder-decoder for statistical machine translation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1724-1734, Doha, Qatar. Association for Computational Linguistics. +Eunsol Choi, Jennimaria Palomaki, Matthew Lamm, Tom Kwiatkowski, Dipanjan Das, and Michael Collins. 2021. Decontextualization: Making sentences stand-alone. Transactions of the Association for Computational Linguistics, 9:447-461. +Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzman, Edouard Grave, Myle Ott, Luke Zettle-moyer, and Veselin Stoyanov. 2020. Unsupervised cross-lingual representation learning at scale. In ACL Conference, pages 8440-8451. +Nicola De Cao, Ledell Wu, Kashyap Popat, Mikel Artetxe, Naman Goyal, Mikhail Plekhanov, Luke Zettlemoyer, Nicola Cancedda, Sebastian Riedel, and Fabio Petroni. 2022. Multilingual autoregressive entity linking. Transactions of the Association for Computational Linguistics, 10:274-290. +Tim Dettmers, Pasquale Minervini, Pontus Stenetorp, and Sebastian Riedel. 2018. Convolutional 2d knowledge graph embeddings. In Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence and Thirtieth Innovative Applications of Artificial Intelligence Conference and Eighth AAAI Symposium on Educational Advances in Artificial Intelligence, AAAI'18/IAAI'18/EAAI'18. AAAI Press. +Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics. +Vladimir Dobrovolskii. 2021. Word-level coreference resolution. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7670-7675, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics. +Oren Etzioni, Anthony Fader, Janara Christensen, Stephen Soderland, and Mausam. 2011. Open information extraction: The second generation. In *IJCAI* + +2011, Proceedings of the 22nd International Joint Conference on Artificial Intelligence, Barcelona, Catalonia, Spain, July 16-22, 2011, pages 3-10. IJ-CAI/AAAI. +Anthony Fader, Stephen Soderland, and Oren Etzioni. 2011. Identifying relations for open information extraction. In Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 1535-1545, Edinburgh, Scotland, UK. Association for Computational Linguistics. +Luis Galárraga, Geremy Heitz, Kevin Murphy, and Fabian M. Suchanek. 2014. Canonicalizing open knowledge bases. New York, NY, USA. Association for Computing Machinery. +Kiril Gashteovski, Rainer Gemulla, and Luciano del Corro. 2017. MinIE: Minimizing facts in open information extraction. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2630-2640, Copenhagen, Denmark. Association for Computational Linguistics. +Kiril Gashteovski, Sebastian Wanner, Sven Hertling, Samuel Broscheit, and Rainer Gemulla. 2019. Opiec: An open information extraction corpus. In Proceedings of the Conference on Automatic Knowledge Base Construction (AKBC). +Swapnil Gupta, Sreyash Kenkre, and Partha Talukdar. 2019. CaRe: Open knowledge graph embeddings. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 378-388, Hong Kong, China. Association for Computational Linguistics. +Bosung Kim, Taesuk Hong, Youngjoong Ko, and Jungyun Seo. 2020. Multi-task learning for knowledge graph completion with pre-trained language models. In Proceedings of the 28th International Conference on Computational Linguistics, pages 1737-1743, Barcelona, Spain (Online). International Committee on Computational Linguistics. +Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings. +Vid Kocijan and Thomas Lukasiewicz. 2021. Knowledge base completion meets transfer learning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP), Punta Cana, Dominican Republic. Association for Computational Linguistics. +Keshav Kolluru, Vaibhav Adlakha, Samarth Aggarwal, Mausam, and Soumen Chakrabarti. 2020. OpenIE6: Iterative Grid Labeling and Coordination Analysis for Open Information Extraction. In Proceedings of + +the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3748-3761, Online. Association for Computational Linguistics. +Keshav Kolluru, Muqeeth Mohammed, Shubham Mittal, Soumen Chakrabarti, and Mausam. 2022. Alignment-augmented consistent translation for multilingual open information extraction. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2502-2517, Dublin, Ireland. Association for Computational Linguistics. +Justin Lovelace and Carolyn Rosé. 2022. A framework for adapting pre-trained language models to knowledge graph completion. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 5937-5955, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics. +Xin Lv, Yankai Lin, Yixin Cao, Lei Hou, Juanzi Li, Zhiyuan Liu, Peng Li, and Jie Zhou. 2022. Do pre-trained models benefit knowledge graph completion? a reliable evaluation and a reasonable approach. In Findings of the Association for Computational Linguistics: ACL 2022, pages 3570-3581, Dublin, Ireland. Association for Computational Linguistics. +Mausam. 2016. Open information extraction systems and downstream applications. In International Joint Conference on Artificial Intelligence. +MediaWiki. 2021. Api:langlinks — mediawiki. [Online; accessed 02-April-2022]. +Jeffrey Pennington, Richard Socher, and Christopher Manning. 2014. GloVe: Global vectors for word representation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1532-1543, Doha, Qatar. Association for Computational Linguistics. +Peng Qi, Yuhao Zhang, Yuhui Zhang, Jason Bolton, and Christopher D. Manning. 2020. Stanza: A Python natural language processing toolkit for many human languages. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations. +Youngbin Ro, Yukyung Lee, and Pilsung Kang. 2020. Multi^2OIE: Multilingual open information extraction based on multi-head attention with BERT. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 1107-1117, Online. Association for Computational Linguistics. +Théo Trouillon, Johannes Welbl, Sebastian Riedel, Eric Gaussier, and Guillaume Bouchard. 2016. Complex embeddings for simple link prediction. In Proceedings of The 33rd International Conference on Machine Learning, volume 48 of Proceedings of Machine Learning Research, pages 2071-2080, New York, New York, USA. PMLR. + +Shikhar Vashishth, Prince Jain, and Partha Talukdar. 2018. CESI: Canonicalizing open knowledge bases using embeddings and side information. In Proceedings of the 2018 World Wide Web Conference, WWW '18, pages 1317-1327, Republic and Canton of Geneva, Switzerland. International World Wide Web Conferences Steering Committee. +Liang Wang, Wei Zhao, Zhuoyu Wei, and Jingming Liu. 2022. SimKGC: Simple contrastive knowledge graph completion with pre-trained language models. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4281-4294, Dublin, Ireland. Association for Computational Linguistics. +Ralph Weischedel, Martha Palmer, Mitchell Marcus, Eduard Hovy, Sameer Pradhan, Lance Ramshaw, Nianwen Xue, Ann Taylor, Jeff Kaufman, and Michelle Franchini. 2013. Ontonotes release 5.0. In Linguistic Data Consortium, Philadelphia, PA. +Guillaume Wenzek, Marie-Anne Lachaux, Alexis Conneau, Vishrav Chaudhary, Francisco Guzmán, Armand Joulin, and Edouard Grave. 2020. CCNet: Extracting high quality monolingual datasets from web crawl data. In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 4003-4012, Marseille, France. European Language Resources Association. +Shijie Wu and Mark Dredze. 2019. Beto, bentz, becas: The surprising cross-lingual effectiveness of BERT. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 833–844, Hong Kong, China. Association for Computational Linguistics. +Patrick Xia and Benjamin Van Durme. 2021. Moving on from OntoNotes: Coreference resolution model transfer. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5241-5256, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics. +Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. 2021. mT5: A massively multilingual pre-trained text-to-text transformer. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 483-498, Online. Association for Computational Linguistics. +Zdeněk Žabokrtský, Miloslav Konopík, Anna Nedoluzhko, Michal Novák, Maciej Ogrodniczuk, Martin Popel, Ondřej Pražák, Jakub Sido, Daniel Zeman, and Yilun Zhu. 2022. Findings of the shared task on multilingual coreference resolution. In Proceedings of the CRAC 2022 Shared Task on Multilingual Coreference Resolution, pages 1-17, + +Gyeongju, Republic of Korea. Association for Computational Linguistics. + +# mOKB6: A Multilingual Open Knowledge Base Completion Benchmark (Appendix) + +# A Dataset Curation + +As discussed in Section 3, we construct mOKB6 dataset in three stages after extracting the Wikipedia articles (using WikiExtractor7) from the Wikidump of April 02, 2022. We run our construction pipeline (as shown in Figure 1) for all six languages on a single V100 (32 GB) GPU, which required 14 hours of computation to create mOKB6 dataset. + +In the first stage, we keep the sentences containing at least 6 and at most 50 tokens since we find that most of the short sentences are headings or sub-headings present in Wikipedia articles, and very long sentences can't be input to GEN2OIE (in second stage) due to maximum sequence length constraint of 1024 in mT5 (Xue et al., 2021) based GEN2OIE. This filtering step discards $18.9\%$ of sentences on an average in all six languages. We use Stanza (Qi et al., 2020) to perform sentence- and word-segmentation on Wikipedia articles in all six languages. After filtering the sentences, the articles are processed for coreference resolution using XLM-R (Conneau et al., 2020) encoder based wlcoref (Dobrovolskii, 2021), followed by replacing the coreferent cluster mentions with their canonical cluster name using the heuristic discussed in Section 3. + +In the second stage, the coreference resolved articles are passed through GEN2OIE to get the Open IE triples. The confidence scores for these triples are computed using label rescoring, for which we refer the readers to Kolluru et al. (2022) for more details. + +Finally, in the last stage, we apply various filters, adapted from Gashteovski et al. (2019), to remove triples that are of no interest to Open KBC task, like the triples: (1) having any of its argument or relation empty, (2) containing more than 10 tokens in any of its arguments or relation, (3) having confidence score less than 0.3, (4) containing pronouns (found using Stanza) in its arguments, (5) having same subject and object (i.e. self loops), and (6) that are duplicates. These filters keep $91.6\%$ of the triples obtained from stage 2 in all six languages. + +Further in the last stage, in order to create a dense Open KB containing minimum noise and maximum facts about the entities, we keep the triples having the Wikipedia article's title as either the subject phrase or object phrase and discard the rest. We do this by finding all the coreference clusters (of entity mentions) that contain the titles, then get the entities, or cluster names, of those clusters using the heuristic discussed in section 3, and keep those triples that contain these cluster names. This filtering step retains $23.6\%$ of the triples. + +# B Metrics + +We follow the previous works (Wang et al., 2022) on the evaluation methodology of Open KBC task and apply it to the multilingual Open KBC task, containing facts in multiple languages. Given an Open KB, containing a finite set of entities and open relations, the KGE model answers forward and backward queries of the form $(s,r,?)$ and $(?,r,o)$ respectively. The model ranks all the entities based on their correctness with, say, $s$ and $r$ in the forward query. Further, the evaluation is in filtered setting, where the other known correct answers, apart from $o$ , are removed from rank list. + +The commonly used evaluation metrics are hits at rank N (H@N), where $N$ is a natural number, and mean reciprocal rank (MRR). Suppose, the model ranks $o$ at $R$ among all entities. Then, H@N measures how many times $R$ is less than or equal to $N$ . MRR is the average of reciprocal ranks $\left( \frac{1}{R} \right)$ . Both, H@N and MRR, are computed as average over both forms of queries over the full test set. + +# C Knowledge Graph Embedding Models + +SimKGC (Wang et al., 2022) is a text-based KGE model that uses two unshared pretrained BERT models (Devlin et al., 2019) for encoding (subject phrase; relation phrase) and object phrase separately. GRU-ConvE (Kocijan and Lukasiewicz, 2021) encodes both the relation phrase and argument phrase from their surface forms using two unshared GRU (Cho et al., 2014). CaRe (Gupta et al., 2019) learns separate embeddings for each argument phrase and uses a bi-directional GRU to encode the relation phrase from its surface form. Both, GRU-ConvE and CaRe, are initialised with Glove embeddings (Pennington et al., 2014). + +![](images/5806f2ef64071ea69556004f4def11da50b67e67341ac2a61a4802cd3edd4777.jpg) +Figure 4: Previous Open KB construction pipelines like Gashteovski et al. (2019) (shown by green arrows) lack coreference resolution system, which result in filtering important facts like (Barack Obama; returned to Honolulu, Hawaii in; 1971). Our pipeline (shown by blue arrows) increases the coverage of facts due to mCoref system. + +To choose the best model for our experiments (Table 3, Figure 3), we train the recent knowledge graph embedding (KGE) models — CaRe., GRUConvE and SimKGC on the English Open KB in mOKB6. We report performance in Table 4 using the three metrics: hits at rank 1 (H@1), hits at 10 (H@10), and mean reciprocal rank (MRR). We find that SimKGC with BERT encoder outperforms the other two models. + +
H@1H@10MRR
CaRe6.611.38.3
GRU-ConvE12.427.817.8
SimKGC (BERT)16.140.024.3
SimKGC (mBERT)14.838.722.8
SimKGC (XLM-R)13.835.821.3
+ +Since BERT supports only English language, we replace BERT in SimKGC with multilingual pretrained language models like mBERT (Devlin et al., 2019) or XLM-R (Conneau et al., 2020), to extend SimKGC model to other languages. We find in Table 4 that SimKGC with mBERT is better than with XLM-R by $2.9\%$ H@10 and $1.5\%$ MRR, possibly because mBERT (and mOKB6) uses Wikipedia while XLM-R uses CommonCrawl (Wenzek et al., 2020) during pre-training. Thus, we use SimKGC with mBERT as the underlying encoder to run our experiments for all the languages. + +# D KGE Model Training Details + +We use the code from official repositories of the KGE models — SimKGC (Wang et al., 2022), GRU-ConvE (Kocijan and Lukasiewicz, 2021), and CaRe (Gupta et al., 2019) for our experiments. The models are trained using Adam optimizer (Kingma and Ba, 2015) on a single A100 (40 GB) GPU with three different random seeds and we report the average of three evaluation runs. + +We do not perform hyperparameter search trials, except for batch size, and use the default hyperparameters from the respective codes of KGE models (see Table 5). We use early stopping to find the best model checkpoints based on HITS@1. The dev set is different for each baseline: MONO, TRANS, MONO+TRANS, and UNION+TRANS use individual language's dev set, whereas UNION w/o En and UNION use the English dev set. We report the performance of baseline models on the dev sets in Table 9 and Table 10. + +Table 4: Performance $(\%)$ of the KGE models on the English test set in mOKB6 dataset. The reported numbers are an average of three runs using different seeds. + +
HyperparameterSimKGCGRU-ConvECaRe
#epochs100500500
#patience epochs101010
learning rate3e-53e-41e-3
dropout0.10.30.5
batch size2561024128
additive margin0.02N/AN/A
+ +We provide the number of trainable parameters of each KGE model in Table 6. Based on the batch size and model size, different experiments consume different GPU hours. To train on English Open KB (in mOKB6 dataset), CaRe and GRU-ConvE models took 2.5 hours and 0.5 hours, respectively, whereas SimKGC takes nearly 1 hour of GPU time. + +Table 5: Hyperparameters of the KGE models. + +
KGE model#trainable parameters
CaRe12,971,423
GRU-ConvE12,085,523
SimKGC (BERT)216,620,545
SimKGC (mBERT)355,706,881
SimKGC (XLM-R)1,119,780,865
+ +Table 6: Number of trainable parameters in the KGE models. + +
EnglishHindiTeluguSpanishPortugueseChinese
H@1H@10MRRH@1H@10MRRH@1H@10MRRH@1H@10MRRH@1H@10MRRH@1H@10MRR
English68.497.178.83.417.28.31.611517.850.728.61744.6265.422.311.1
Hindi1942.226.780.699.588.32.412.55.912.33619.912.333.919.75.321.910.8
Telugu19.542.227.24.318.79.474.499.584.210.935.418.910.73418.54.721.410.1
Spanish27.960.438.84.117.88.91.810.75.18410090.337.67450.16.524.912.8
Portuguese27.858.738.24.418.29.31.710.55.141.578.553.684.299.990.86.62613.2
Chinese22.148.430.63.518.58.81.812.25.414.842.824.215.741.624.181.699.889.2
+ +Table 7: Performance (%) of the six MEMORIZE models, which have been trained on each language's test set and tested on all the test sets in mOKB6 dataset. + +# E Contextual Triples + +Open IE triples are of various kinds and not all of them can be used for Open KBC task. Various filtering steps are used to remove some of these in data curation (Section 3). We define contextual triples as another kind of noisy triples, which are specific to, and are not interpretable out of, the context of text from which they are extracted. + +(Max Born; continued; scientific work) + +(Robb Gravett; won; the championship) + +(George Herbert Walker Bush; was; out of touch) + +(Christianity; is; dominant) + +Table 8: Examples of contextual triples. + +From the first two triples in Table 8, it is unclear which scientific work Max Born continued, or which championship Robb Gravett has won. The last two triples are too specific to the context and contain no factual information. + +
English (En)Hindi (Hi)Telugu (Te)Spanish (Es)Portuguese (Pt)Chinese (Zh)
H@1H@10MRRH@1H@10MRRH@1H@10MRRH@1H@10MRRH@1H@10MRRH@1H@10MRR
MONO16.238.723.918.239.425.98.52012.517.336.623.717.639.625.310.831.917.8
TRANS---8.123.713.53.315.47.512.933.620.312.637.220.6520.810.3
MONO+TRANS---20.843.228.67.824.813.420.24628.82145.929.210.630.116.7
UNION19.939.626.414.538.222.45.92010.619.843.227.919.743.82811.23318.8
UNION w/o En5.819.510.615.439.323.36.320.511.119.441.626.416.942.925.911.33318.4
UNION+TRANS---20.844.928.87.327.11421.445.329.619.449.129.16.93115.1
+ +Table 9: Performance (%) of SimKGC on the dev sets (of mOKB6 dataset) in six languages. + +
H@1H@10MRR
CaRe7.111.18.5
GRU-ConvE16.831.522.1
SimKGC (BERT)20.340.127.1
SimKGC (mBERT)16.238.723.9
SimKGC (XLM-R)1736.623.2
+ +Table 10: Performance (%) of the KGE models on dev set of English Open KB in mOKB6 dataset. + +A For every submission: + +A1. Did you describe the limitations of your work? 8 +A2. Did you discuss any potential risks of your work? There are no potential risks of our work to our knowledge. +A3. Do the abstract and introduction summarize the paper's main claims? +A4. Have you used AI writing assistants when working on this paper? Left blank. + +B Did you use or create scientific artifacts? + +3,4 + +B1. Did you cite the creators of artifacts you used? 3,4 +B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Abstract +B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Not applicable. Left blank. +B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Not applicable. Left blank. +B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? 3 +B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. 3 + +C Did you run computational experiments? + +4 + +C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Appendix D + +C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Appendix D +C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? +Appendix D +C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Appendix A, D + +D Did you use human annotators (e.g., crowdworkers) or research with human participants? 3 + +D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? +D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Not applicable. Left blank. +D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Not applicable. Left blank. +D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Not applicable. Left blank. +D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Not applicable. Left blank. \ No newline at end of file diff --git a/2023/mOKB6_ A Multilingual Open Knowledge Base Completion Benchmark/images.zip b/2023/mOKB6_ A Multilingual Open Knowledge Base Completion Benchmark/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a8f50b496dc64faaf84b039847db7a7b32c6c8e8 --- /dev/null +++ b/2023/mOKB6_ A Multilingual Open Knowledge Base Completion Benchmark/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0f06d60726cec12ef92220d6e083266bfb0c52bace05d6c0054272eab270906d +size 425162 diff --git a/2023/mOKB6_ A Multilingual Open Knowledge Base Completion Benchmark/layout.json b/2023/mOKB6_ A Multilingual Open Knowledge Base Completion Benchmark/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..785f42a0efbe954fbcbf4c296ba858ffc8c64413 --- /dev/null +++ b/2023/mOKB6_ A Multilingual Open Knowledge Base Completion Benchmark/layout.json @@ -0,0 +1,8643 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 74, + 75, + 519, + 94 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 75, + 519, + 94 + ], + "spans": [ + { + "bbox": [ + 74, + 75, + 519, + 94 + ], + "type": "text", + "content": "mOKB6: A Multilingual Open Knowledge Base Completion Benchmark" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 96, + 104, + 502, + 118 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 104, + 502, + 118 + ], + "spans": [ + { + "bbox": [ + 96, + 104, + 502, + 118 + ], + "type": "text", + "content": "Shubham Mittal" + }, + { + "bbox": [ + 96, + 104, + 502, + 118 + ], + "type": "inline_equation", + "content": "^{\\alpha \\dagger}" + }, + { + "bbox": [ + 96, + 104, + 502, + 118 + ], + "type": "text", + "content": " Keshay Kolluru" + }, + { + "bbox": [ + 96, + 104, + 502, + 118 + ], + "type": "inline_equation", + "content": "^{\\beta \\dagger}" + }, + { + "bbox": [ + 96, + 104, + 502, + 118 + ], + "type": "text", + "content": " Soumen Chakrabarti" + }, + { + "bbox": [ + 96, + 104, + 502, + 118 + ], + "type": "inline_equation", + "content": "^{\\gamma}" + }, + { + "bbox": [ + 96, + 104, + 502, + 118 + ], + "type": "text", + "content": " Mausam" + }, + { + "bbox": [ + 96, + 104, + 502, + 118 + ], + "type": "inline_equation", + "content": "^{\\alpha}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 205, + 120, + 391, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 120, + 391, + 132 + ], + "spans": [ + { + "bbox": [ + 205, + 120, + 391, + 132 + ], + "type": "inline_equation", + "content": "^{\\alpha}" + }, + { + "bbox": [ + 205, + 120, + 391, + 132 + ], + "type": "text", + "content": " Indian Institute of Technology Delhi" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 233, + 133, + 364, + 146 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 133, + 364, + 146 + ], + "spans": [ + { + "bbox": [ + 233, + 133, + 364, + 146 + ], + "type": "inline_equation", + "content": "^{\\beta}" + }, + { + "bbox": [ + 233, + 133, + 364, + 146 + ], + "type": "text", + "content": " KnowDis AI, New Delhi" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 198, + 148, + 398, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 148, + 398, + 161 + ], + "spans": [ + { + "bbox": [ + 198, + 148, + 398, + 161 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 198, + 148, + 398, + 161 + ], + "type": "text", + "content": " Indian Institute of Technology Bombay" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 162, + 446, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 162, + 446, + 175 + ], + "spans": [ + { + "bbox": [ + 151, + 162, + 446, + 175 + ], + "type": "text", + "content": "shubhamiitd18@gmail.com, keshav.kolluru@gmail.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 165, + 177, + 431, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 177, + 431, + 189 + ], + "spans": [ + { + "bbox": [ + 165, + 177, + 431, + 189 + ], + "type": "text", + "content": "soumen@cse.iitb.ac.in, mausam@cse.iitd.ac.in" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 155, + 226, + 202, + 238 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 226, + 202, + 238 + ], + "spans": [ + { + "bbox": [ + 155, + 226, + 202, + 238 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 84, + 253, + 274, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 253, + 274, + 539 + ], + "spans": [ + { + "bbox": [ + 84, + 253, + 274, + 539 + ], + "type": "text", + "content": "Automated completion of open knowledge bases (Open KBs), which are constructed from triples of the form (subject phrase, relation phrase, object phrase), obtained via open information extraction (Open IE) system, are useful for discovering novel facts that may not be directly present in the text. However, research in Open KB completion (Open KBC) has so far been limited to resource-rich languages like English. Using the latest advances in multilingual Open IE, we construct the first multilingual Open KBC dataset, called mOKB6, containing facts from Wikipedia in six languages (including English). Improving the previous Open KB construction pipeline by doing multilingual coreference resolution and keeping only entity-linked triples, we create a dense Open KB. We experiment with several models for the task and observe a consistent benefit of combining languages with the help of shared embedding space as well as translations of facts. We also observe that current multilingual models struggle to remember facts seen in languages of different scripts." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 68, + 555, + 154, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 555, + 154, + 567 + ], + "spans": [ + { + "bbox": [ + 68, + 555, + 154, + 567 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 578, + 291, + 741 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 578, + 291, + 741 + ], + "spans": [ + { + "bbox": [ + 67, + 578, + 291, + 741 + ], + "type": "text", + "content": "Open information extraction (Open IE) systems (Mausam, 2016) such as ReVerb (Etzioni et al., 2011) and OpenIE6 (Kolluru et al., 2020) can extract triples, or facts, of the form (subject phrase, relation phrase, object phrase), which can be denoted as " + }, + { + "bbox": [ + 67, + 578, + 291, + 741 + ], + "type": "inline_equation", + "content": "(s,r,o)" + }, + { + "bbox": [ + 67, + 578, + 291, + 741 + ], + "type": "text", + "content": ", from text (e.g., Wikipedia articles) without using any pre-defined ontology. Open knowledge base (Open KB) is constructed using these Open IE triples where the subject phrases and object phrases are nodes and relation phrases are labels on edges connecting the nodes in the graph. Open knowledge base completion (Open KBC) is" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 213, + 526, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 213, + 526, + 280 + ], + "spans": [ + { + "bbox": [ + 302, + 213, + 526, + 280 + ], + "type": "text", + "content": "the task of discovering new links between nodes using the graph structure of the Open KB. Knowledge graph embedding (KGE) models are typically used for the Open KBC task, where they are asked to answer questions of the form " + }, + { + "bbox": [ + 302, + 213, + 526, + 280 + ], + "type": "inline_equation", + "content": "(s,r,?)" + }, + { + "bbox": [ + 302, + 213, + 526, + 280 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 302, + 213, + 526, + 280 + ], + "type": "inline_equation", + "content": "(?,r,o)" + }, + { + "bbox": [ + 302, + 213, + 526, + 280 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 282, + 526, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 282, + 526, + 431 + ], + "spans": [ + { + "bbox": [ + 302, + 282, + 526, + 431 + ], + "type": "text", + "content": "Research in Open KBC has been restricted to English (Vashishth et al., 2018) due to lack of Open KBs in other languages. We aim to study multilingual Open KBC, with the motivation that the information available in high resource languages like English may help when inferring links in Open KBs that use low resource languages like Telugu. Moreover, intuitively, if all the information in different languages can be pooled together, then it may help the model learn better, and allow information flow across Open KBs in different languages." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 433, + 526, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 433, + 526, + 649 + ], + "spans": [ + { + "bbox": [ + 302, + 433, + 526, + 649 + ], + "type": "text", + "content": "We design the first multilingual Open KB construction pipeline (shown in Figure 1) using a multilingual Open IE system, GEN2OIE (Kolluru et al., 2022). We find that coreference resolution is missing in existing Open KB construction (Gashteovski et al., 2019) but is important for increasing the coverage of facts (as described in Figure 4). We re-train a recent coref model (Dobrovolskii, 2021) using XLM-R (Conneau et al., 2020) as the underlying multilingual encoder and add it to our pipeline. For constructing a high quality test set, we use 988 manually verified facts in English. For extending to other languages, we automatically translate English facts. The dataset thus constructed, called mOKB6, contains 42K facts in six languages: English, Hindi, Telugu, Spanish, Portuguese, and Chinese." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 651, + 526, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 651, + 526, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 651, + 526, + 773 + ], + "type": "text", + "content": "We report the first baselines for multilingual Open KBC task. We find that they are able to benefit from information in multiple languages when compared to using facts from a single language. Translations of Open KB facts also help the models. However, we notice that although the multilingual KGE models learn facts in a particular language, they struggle to remember the same fact, when queried in another language with different script." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 83, + 750, + 269, + 761 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 750, + 269, + 761 + ], + "spans": [ + { + "bbox": [ + 83, + 750, + 269, + 761 + ], + "type": "text", + "content": "Major part of work done as students at IIT Delhi." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 82, + 761, + 289, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 761, + 289, + 772 + ], + "spans": [ + { + "bbox": [ + 82, + 761, + 289, + 772 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 82, + 761, + 289, + 772 + ], + "type": "text", + "content": "Dataset and code released at github.com:dair-iitd/mokb6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 289, + 780, + 305, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 780, + 305, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 780, + 305, + 791 + ], + "type": "text", + "content": "201" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "spans": [ + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "type": "text", + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 224, + 806, + 369, + 817 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 806, + 369, + 817 + ], + "spans": [ + { + "bbox": [ + 224, + 806, + 369, + 817 + ], + "type": "text", + "content": "Volume 2: Short Papers, pages 201-214" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 176, + 818, + 417, + 828 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 818, + 417, + 828 + ], + "spans": [ + { + "bbox": [ + 176, + 818, + 417, + 828 + ], + "type": "text", + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 70, + 161, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 70, + 161, + 83 + ], + "spans": [ + { + "bbox": [ + 68, + 70, + 161, + 83 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 93, + 291, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 93, + 291, + 309 + ], + "spans": [ + { + "bbox": [ + 67, + 93, + 291, + 309 + ], + "type": "text", + "content": "Multilingual Open KBC datasets are absent in literature to the best of our knowledge, although multiple English Open KBC datasets are available. OLPBench (Broscheit et al., 2020), derived from OPIEC (Gashteovski et al., 2019), is a large-scale Open KBC dataset that contains 30M triples and is constructed from English Wikipedia using MinIE system (Gashteovski et al., 2017). The evaluation data contains 10K triples randomly sampled from 1.25M linked triples. ReVerb45K (Vashisth et al., 2018) and ReVerb20K (Galarraga et al., 2014) are smaller Open KBC datasets constructed from Clueweb09 corpus" + }, + { + "bbox": [ + 67, + 93, + 291, + 309 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 67, + 93, + 291, + 309 + ], + "type": "text", + "content": " using ReVerb Open IE system (Fader et al., 2011). Both the datasets keep only those tuples in which both the subject phrase and object phrase link to a finite set of Freebase entities." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 310, + 291, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 310, + 291, + 417 + ], + "spans": [ + { + "bbox": [ + 67, + 310, + 291, + 417 + ], + "type": "text", + "content": "Multilingual Open IE (mOpenIE) systems like GEN2OIE (Kolluru et al., 2022) and Multi" + }, + { + "bbox": [ + 67, + 310, + 291, + 417 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 310, + 291, + 417 + ], + "type": "text", + "content": "OIE (Ro et al., 2020) enable extracting facts from multiple languages. We use the GEN2OIE model for constructing mOKB6 dataset as it is trained with language-specific facts transferred from English, while Multi" + }, + { + "bbox": [ + 67, + 310, + 291, + 417 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 310, + 291, + 417 + ], + "type": "text", + "content": "OIE relies on zero-shot transfer for languages other than English." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 427, + 291, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 427, + 291, + 616 + ], + "spans": [ + { + "bbox": [ + 67, + 427, + 291, + 616 + ], + "type": "text", + "content": "Knowledge Graph Embedding (KGE) Models: Conventional KGE models like TransE (Bordes et al., 2013), ComplEx (Trouillon et al., 2016), ConvE (Dettmers et al., 2018), and TuckER (Balazevic et al., 2019) have been used for Open KBC task (Gupta et al., 2019; Broscheit et al., 2020; Chandrahas and Talukdar, 2021; Kocijan and Lukasiewicz, 2021). Given a triple " + }, + { + "bbox": [ + 67, + 427, + 291, + 616 + ], + "type": "inline_equation", + "content": "(s,r,o)" + }, + { + "bbox": [ + 67, + 427, + 291, + 616 + ], + "type": "text", + "content": ", these models encode the subject phrase, relation phrase, and object phrase from free text, and pass the encodings to a triple-scoring function, which is optimized using binary cross entropy loss. ComplEx has also been used for multilingual closed KBC task (Chakrabarti et al., 2022)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 617, + 291, + 751 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 617, + 291, + 751 + ], + "spans": [ + { + "bbox": [ + 67, + 617, + 291, + 751 + ], + "type": "text", + "content": "Pretrained language models like BERT (Devlin et al., 2019) have been used in KGE models for the KBC task (Lovelace and Rosé, 2022; Lv et al., 2022; Chandrahas and Talukdar, 2021; Kim et al., 2020). SimKGC (Wang et al., 2022) is the state of the art KGE model on closed KBC task. It computes the score of a triple " + }, + { + "bbox": [ + 67, + 617, + 291, + 751 + ], + "type": "inline_equation", + "content": "(s, r, o)" + }, + { + "bbox": [ + 67, + 617, + 291, + 751 + ], + "type": "text", + "content": " as the cosine similarity of the embeddings of " + }, + { + "bbox": [ + 67, + 617, + 291, + 751 + ], + "type": "inline_equation", + "content": "(s; r)" + }, + { + "bbox": [ + 67, + 617, + 291, + 751 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 617, + 291, + 751 + ], + "type": "inline_equation", + "content": "(o)" + }, + { + "bbox": [ + 67, + 617, + 291, + 751 + ], + "type": "text", + "content": ", computed using two separate pretrained BERT models without any weight sharing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 303, + 70, + 412, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 70, + 412, + 83 + ], + "spans": [ + { + "bbox": [ + 303, + 70, + 412, + 83 + ], + "type": "text", + "content": "3 Dataset Curation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 92, + 526, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 92, + 526, + 295 + ], + "spans": [ + { + "bbox": [ + 302, + 92, + 526, + 295 + ], + "type": "text", + "content": "We aim to construct a dense multilingual Open KB that maximizes the information about a given real-world entity, which may be represented as multiple nodes across languages. Therefore, we consider those Wikipedia articles3 that are available in six languages: English, Hindi, Telugu, Spanish, Portuguese, and Chinese4. This will also help the model learn from facts in high resource language like English and answer queries in low resource language like Telugu. We work with 300 titles randomly sampled from the ones common among all six languages (found using MediaWiki-Langlinks (MediaWiki, 2021)). Thus, we extract facts from " + }, + { + "bbox": [ + 302, + 92, + 526, + 295 + ], + "type": "inline_equation", + "content": "6 \\times 300" + }, + { + "bbox": [ + 302, + 92, + 526, + 295 + ], + "type": "text", + "content": " Wikipedia articles. We discuss the three stages of our pipeline below." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 302, + 526, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 302, + 526, + 450 + ], + "spans": [ + { + "bbox": [ + 302, + 302, + 526, + 450 + ], + "type": "text", + "content": "Stage 1 We first process each Wikipedia article through a coreference resolution system. Although language-specific end-to-end neural coref models have been developed (Žabokrtský et al., 2022; Xia and Van Durme, 2021), multilingual models that work on all our languages of interest are absent in the literature. Therefore, we retrain wl-coref (Dobrovolskii, 2021) with XLM-R (Conneau et al., 2020) on the English training data (available in OntoNotes (Weischedel et al., 2013)) that can work zero-shot for other languages." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 452, + 526, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 452, + 526, + 709 + ], + "spans": [ + { + "bbox": [ + 302, + 452, + 526, + 709 + ], + "type": "text", + "content": "Coref models detect and cluster mentions, but do not identify a canonical cluster name, which is needed for standardizing all the mentions in the cluster. To find cluster names, entity linking systems such as mGENRE (De Cao et al., 2022) or Wikipedia hyperlinks can be used. However, we found that they result in low recall, particularly for low resource languages. Thus, we employ a heuristic to find the cluster name and replace each of the coreferent mentions with it. The score for each mention is represented by a tuple, computed as: Score(mention phrase) = (#proper nouns, #nouns, #numerals, #adjectives, #pronouns, #verbs). The tuple is ordered according to the importance of each field (POS tags) for the cluster name, which is determined empirically. Two tuples are compared index-wise with higher priority given to lower indices to determine the best scoring mention that is chosen as the canonical name (Table 1)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 716, + 525, + 742 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 716, + 525, + 742 + ], + "spans": [ + { + "bbox": [ + 302, + 716, + 525, + 742 + ], + "type": "text", + "content": "Stage 2 We use GEN2OIE to extract Open IE triples from the coreference resolved sentences." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 749, + 426, + 761 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 749, + 426, + 761 + ], + "spans": [ + { + "bbox": [ + 315, + 749, + 426, + 761 + ], + "type": "text", + "content": "3Wikidump of April 02, 2022" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 761, + 520, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 761, + 520, + 772 + ], + "spans": [ + { + "bbox": [ + 315, + 761, + 520, + 772 + ], + "type": "text", + "content": "4languages are chosen to match availability of Gen2OIE" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 80, + 760, + 275, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 760, + 275, + 772 + ], + "spans": [ + { + "bbox": [ + 80, + 760, + 275, + 772 + ], + "type": "text", + "content": "2http://www.lemurproject.org/clueweb09.php/" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "text", + "content": "202" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 69, + 526, + 113 + ], + "blocks": [ + { + "bbox": [ + 69, + 69, + 526, + 113 + ], + "lines": [ + { + "bbox": [ + 69, + 69, + 526, + 113 + ], + "spans": [ + { + "bbox": [ + 69, + 69, + 526, + 113 + ], + "type": "image", + "image_path": "d8cf19105aad3c532c70c6bfe2852b999abae3175fb081a4cd65e4a3f6442568.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 123, + 525, + 160 + ], + "lines": [ + { + "bbox": [ + 67, + 123, + 525, + 160 + ], + "spans": [ + { + "bbox": [ + 67, + 123, + 525, + 160 + ], + "type": "text", + "content": "Figure 1: Our three-staged multilingual Open KB construction pipeline for mOKB6. mCoref is multilingual coreference resolution system, having XLM-R (Conneau et al., 2020) encoder based wl-coref (Dobrovolskii, 2021), and mOpenIE is multilingual open information extraction system, consisting of GEN2OIE (Kolluru et al., 2022)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 86, + 173, + 272, + 228 + ], + "blocks": [ + { + "bbox": [ + 86, + 173, + 272, + 228 + ], + "lines": [ + { + "bbox": [ + 86, + 173, + 272, + 228 + ], + "spans": [ + { + "bbox": [ + 86, + 173, + 272, + 228 + ], + "type": "table", + "html": "
MentionsScoresCluster Name
Barack Obama(2,0,0,0,0,0)
Obama(1,0,0,0,0,0)Barack Obama
He(0,0,0,0,1,0)
", + "image_path": "9c8dc2061c47852e3421b2da94e3a16579cbafb6e40911a03b062af8932394db.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 273, + 290, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 273, + 290, + 407 + ], + "spans": [ + { + "bbox": [ + 67, + 273, + 290, + 407 + ], + "type": "text", + "content": "Stage 3 Similar to Gashteovski et al. (2019), we apply various filters to remove noisy triples that have empty or very long arguments, or have less confidence than 0.3 (as assigned by GEN2OIE). We further only keep triples that have the article's title as either the subject phrase or object phrase, to avoid generic or specific triples, valid only in the particular context. Examples of contextual triples (Choi et al., 2021) are discussed in Appendix E. See Appendix A for further data curation details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 411, + 291, + 748 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 411, + 291, + 748 + ], + "spans": [ + { + "bbox": [ + 69, + 411, + 291, + 748 + ], + "type": "text", + "content": "These automatically extracted triples form the train set of mOKB6. To form a high quality test set in six languages with limited access to experts in all languages, the test set is created in a semiautomatic way. We sample 1600 English triples from the train set (which are subsequently filtered) and manually remove noisy triples. We use inter-annotation agreement between two annotators to check if they both agree that the given triple is noisy or clean. With an agreement of " + }, + { + "bbox": [ + 69, + 411, + 291, + 748 + ], + "type": "inline_equation", + "content": "91\\%" + }, + { + "bbox": [ + 69, + 411, + 291, + 748 + ], + "type": "text", + "content": ", we retain 988 English triples, which we automatically translate to the other five languages. As illustrated in Figure 2, to translate a triple, we convert it to a sentence after removing tags and use Google translate for translating the triple-converted sentence to the remaining five languages. We observed high quality of translated triples, with " + }, + { + "bbox": [ + 69, + 411, + 291, + 748 + ], + "type": "inline_equation", + "content": "88\\%" + }, + { + "bbox": [ + 69, + 411, + 291, + 748 + ], + "type": "text", + "content": " satisfactory translations as determined by native-speakers of three languages on a set of 75 translated triples. To get the Open IE subject phrase, relation phrase and object phrase tags, we project the labels from the original English triple to the translated sentence using word alignments (Kolluru et al., 2022). Finally, we are left with 550 triples in each language after removing examples where some labels could" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 175, + 526, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 175, + 526, + 231 + ], + "spans": [ + { + "bbox": [ + 302, + 175, + 526, + 231 + ], + "type": "text", + "content": "not be aligned. We use these " + }, + { + "bbox": [ + 302, + 175, + 526, + 231 + ], + "type": "inline_equation", + "content": "6 \\times 550" + }, + { + "bbox": [ + 302, + 175, + 526, + 231 + ], + "type": "text", + "content": " triples as the test sets. The train and dev sets are created from the remaining triples in each language such that the dev set has 500 randomly sampled triples (Table 2)." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 341, + 242, + 489, + 368 + ], + "blocks": [ + { + "bbox": [ + 341, + 242, + 489, + 368 + ], + "lines": [ + { + "bbox": [ + 341, + 242, + 489, + 368 + ], + "spans": [ + { + "bbox": [ + 341, + 242, + 489, + 368 + ], + "type": "image", + "image_path": "5ef6efdb8a343a6921e6766304a6e31bfc2b48d19ad192936565512a7cf2369b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 382, + 525, + 418 + ], + "lines": [ + { + "bbox": [ + 302, + 382, + 525, + 418 + ], + "spans": [ + { + "bbox": [ + 302, + 382, + 525, + 418 + ], + "type": "text", + "content": "Figure 2: Method to translate Open IE triple using Google translate, and followed by label projection using word alignments (Kolluru et al., 2022)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 432, + 525, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 432, + 525, + 526 + ], + "spans": [ + { + "bbox": [ + 302, + 432, + 525, + 526 + ], + "type": "text", + "content": "We analyse the entity overlap across languages and find that on an average, a test entity (which is present in either the subject phrase or object phrase of a test tuple) is present 17.73 times in English, 0.94 times in Hindi, 0.47 times in Telugu, 2.33 times in Spanish, 1.69 times in Portuguese, and 1.45 times in Chinese train set." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 527, + 526, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 527, + 526, + 649 + ], + "spans": [ + { + "bbox": [ + 302, + 527, + 526, + 649 + ], + "type": "text", + "content": "Our construction pipeline improves over OPIEC in three ways: (1) we use a multilingual Open IE system, instead of an English-specific Open IE system like in OPIEC, enabling us to curate Open KBs in many languages, (2) we add a multilingual coreference resolution system in our pipeline, and (3) the English test triples are manually verified. Further, we manually evaluate and review the noise at each step of data curation in Section 4." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 304, + 657, + 525, + 714 + ], + "blocks": [ + { + "bbox": [ + 67, + 237, + 291, + 261 + ], + "lines": [ + { + "bbox": [ + 67, + 237, + 291, + 261 + ], + "spans": [ + { + "bbox": [ + 67, + 237, + 291, + 261 + ], + "type": "text", + "content": "Table 1: Parts of speech tags are used to find the canonical name of the coreferent cluster of entity mentions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 304, + 657, + 525, + 714 + ], + "lines": [ + { + "bbox": [ + 304, + 657, + 525, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 657, + 525, + 714 + ], + "type": "table", + "html": "
EnHiTeEsPtZh
#entity2063746253972565153045037
#relation787021771907282326442325
#train2019527861992396635283420
", + "image_path": "d615963f2978c0599795ca5812bb17b69c514af75cfdce81ff9c830d99989c63.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 721, + 525, + 769 + ], + "lines": [ + { + "bbox": [ + 302, + 721, + 525, + 769 + ], + "spans": [ + { + "bbox": [ + 302, + 721, + 525, + 769 + ], + "type": "text", + "content": "Table 2: Statistics of individual Open KBs in mOKB6 in English (En), Hindi (Hi), Telugu (Te), Spanish (Es), Portuguese (Pt), and Chinese (Zh). The dev and test set for each Open KB contain 500 and 550 triples each." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 80, + 760, + 226, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 760, + 226, + 772 + ], + "spans": [ + { + "bbox": [ + 80, + 760, + 226, + 772 + ], + "type": "text", + "content": "5https://translate.google.co.in/" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "text", + "content": "203" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 176, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 176, + 83 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 176, + 83 + ], + "type": "text", + "content": "4 Noise Evaluation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 92, + 290, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 92, + 290, + 213 + ], + "spans": [ + { + "bbox": [ + 67, + 92, + 290, + 213 + ], + "type": "text", + "content": "Curating an Open KB involves various stages and each stage induces its noise in the construction pipeline (Gashteovski et al., 2019). We manually evaluate the noise induced at each stage of our pipeline (Figure 1) and discuss the same in this section. We ask native speakers of four (out of six) languages - English, Hindi, Telugu, and Chinese to assess the output quality, or precision, of each stage as discussed below." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 214, + 290, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 214, + 290, + 308 + ], + "spans": [ + { + "bbox": [ + 67, + 214, + 290, + 308 + ], + "type": "text", + "content": "In the first stage, we assess the performance of the coreference resolution system over Wikipedia articles. We find a high precision of " + }, + { + "bbox": [ + 67, + 214, + 290, + 308 + ], + "type": "inline_equation", + "content": "95.5\\%" + }, + { + "bbox": [ + 67, + 214, + 290, + 308 + ], + "type": "text", + "content": " in coref's mention clustering and " + }, + { + "bbox": [ + 67, + 214, + 290, + 308 + ], + "type": "inline_equation", + "content": "89.82\\%" + }, + { + "bbox": [ + 67, + 214, + 290, + 308 + ], + "type": "text", + "content": " accuracy in finding canonical cluster name (using the heuristic illustrated in Table 1), computed over 40 randomly sampled coref clusters (10 in each language)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 309, + 290, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 309, + 290, + 389 + ], + "spans": [ + { + "bbox": [ + 67, + 309, + 290, + 389 + ], + "type": "text", + "content": "For evaluating the Open IE system, GEN2OIE, in the second stage, we mark an extraction of a sentence as correct if it has syntactically correct arguments and it is coherent with the sentence. We get an average precision of " + }, + { + "bbox": [ + 67, + 309, + 290, + 389 + ], + "type": "inline_equation", + "content": "63.4\\%" + }, + { + "bbox": [ + 67, + 309, + 290, + 389 + ], + "type": "text", + "content": " on 80 extractions (20 in each language)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 391, + 290, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 391, + 290, + 525 + ], + "spans": [ + { + "bbox": [ + 67, + 391, + 290, + 525 + ], + "type": "text", + "content": "We evaluate the triples, or Open KB facts, at the last stage after passing through various noise-removing filters. Note that these triples also form the train set (and dev set) in mOKB6 dataset. We mark triples as correct when they contain real-world entities, and also, factual information about them. If the triple is very generic or contextual (see Appendix E), it is marked as incorrect. We find the train (and dev) set quality to be " + }, + { + "bbox": [ + 67, + 391, + 290, + 525 + ], + "type": "inline_equation", + "content": "69.3\\%" + }, + { + "bbox": [ + 67, + 391, + 290, + 525 + ], + "type": "text", + "content": ", averaged over 80 triples in four languages." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 535, + 155, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 535, + 155, + 550 + ], + "spans": [ + { + "bbox": [ + 67, + 535, + 155, + 550 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 556, + 289, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 556, + 289, + 583 + ], + "spans": [ + { + "bbox": [ + 67, + 556, + 289, + 583 + ], + "type": "text", + "content": "Our experimental study on multilingual open KBC task investigates the following research questions:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 76, + 584, + 290, + 663 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 77, + 584, + 289, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 584, + 289, + 609 + ], + "spans": [ + { + "bbox": [ + 77, + 584, + 289, + 609 + ], + "type": "text", + "content": "1. Does the KGE model benefit from facts in different languages? (Section 5.1)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 76, + 611, + 290, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 611, + 290, + 636 + ], + "spans": [ + { + "bbox": [ + 76, + 611, + 290, + 636 + ], + "type": "text", + "content": "2. Can translation help transfer among languages? (Section 5.2)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 76, + 638, + 289, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 638, + 289, + 663 + ], + "spans": [ + { + "bbox": [ + 76, + 638, + 289, + 663 + ], + "type": "text", + "content": "3. Does the KGE model remember facts seen across different languages? (Section 5.3)" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 66, + 665, + 290, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 665, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 66, + 665, + 290, + 772 + ], + "type": "text", + "content": "We use SimKGC model (Wang et al., 2022) with pretrained mBERT initialization to run our experiments, after comparing with recent KGE models (Appendix C). For evaluation, we use three metrics -hits at rank 1 (H@1), hits at rank 10 (H@10), and mean reciprocal rank (MRR). The formal definitions of them are provided in Appendix B. We discuss further model training details in Appendix D." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 71, + 474, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 474, + 84 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 474, + 84 + ], + "type": "text", + "content": "5.1 Training on Multilingual Facts" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 95, + 525, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 95, + 525, + 229 + ], + "spans": [ + { + "bbox": [ + 302, + 95, + 525, + 229 + ], + "type": "text", + "content": "We train and compare monolingual model, called MONO, with multilingual models, UNION and UNION w/o En. In MONO, we train one model for each language using its respective Open KB, whereas in UNION, a single model is trained on six languages' Open KBs together. UNION outperforms MONO in all languages by an average of " + }, + { + "bbox": [ + 302, + 95, + 525, + 229 + ], + "type": "inline_equation", + "content": "4.6\\%" + }, + { + "bbox": [ + 302, + 95, + 525, + 229 + ], + "type": "text", + "content": " H@10 and " + }, + { + "bbox": [ + 302, + 95, + 525, + 229 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 302, + 95, + 525, + 229 + ], + "type": "text", + "content": " MRR (see Table 3), which provides evidence of information flow across languages and the model benefits from it." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 233, + 525, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 233, + 525, + 328 + ], + "spans": [ + { + "bbox": [ + 302, + 233, + 525, + 328 + ], + "type": "text", + "content": "To check the extent of flow from (high-resource) English to the other languages, we also train on the five languages except English, which we call UNION w/o En. We find UNION w/o En also outperforms MONO by " + }, + { + "bbox": [ + 302, + 233, + 525, + 328 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 302, + 233, + 525, + 328 + ], + "type": "text", + "content": " H@10 and " + }, + { + "bbox": [ + 302, + 233, + 525, + 328 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 302, + 233, + 525, + 328 + ], + "type": "text", + "content": " MRR over the five languages, hinting that interlingual transfer is more general and pervasive." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 349, + 458, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 349, + 458, + 361 + ], + "spans": [ + { + "bbox": [ + 302, + 349, + 458, + 361 + ], + "type": "text", + "content": "5.2 Open KB Facts Translation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 372, + 525, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 372, + 525, + 588 + ], + "spans": [ + { + "bbox": [ + 302, + 372, + 525, + 588 + ], + "type": "text", + "content": "Apart from relying only on multilingual transfer in the embedding space, we analyse the effect of using translated triples in the training of the KGE model. We translate the English training triples to the other five languages (Section 3) and train monolingual models using only the translated triples (TRANS). To leverage facts present in each language's Open KB, we make MONO+TRANS, where we add language-specific MONO data to the translated triples. Table 3 shows that MONO+TRANS is better than MONO by a large margin of " + }, + { + "bbox": [ + 302, + 372, + 525, + 588 + ], + "type": "inline_equation", + "content": "15.5\\%" + }, + { + "bbox": [ + 302, + 372, + 525, + 588 + ], + "type": "text", + "content": " H@1, " + }, + { + "bbox": [ + 302, + 372, + 525, + 588 + ], + "type": "inline_equation", + "content": "29.2\\%" + }, + { + "bbox": [ + 302, + 372, + 525, + 588 + ], + "type": "text", + "content": " H@10, and " + }, + { + "bbox": [ + 302, + 372, + 525, + 588 + ], + "type": "inline_equation", + "content": "20.0\\%" + }, + { + "bbox": [ + 302, + 372, + 525, + 588 + ], + "type": "text", + "content": " MRR, averaged over five languages. Also, MONO+TRANS improves over TRANS by " + }, + { + "bbox": [ + 302, + 372, + 525, + 588 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 302, + 372, + 525, + 588 + ], + "type": "text", + "content": " H@10 and " + }, + { + "bbox": [ + 302, + 372, + 525, + 588 + ], + "type": "inline_equation", + "content": "2.0\\%" + }, + { + "bbox": [ + 302, + 372, + 525, + 588 + ], + "type": "text", + "content": " MRR, showcasing the importance of facts in each language's Open KBs." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 302, + 592, + 525, + 740 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 592, + 525, + 740 + ], + "spans": [ + { + "bbox": [ + 302, + 592, + 525, + 740 + ], + "type": "text", + "content": "To effectively gain from transfer in both the embedding space as well as translation, we introduce " + }, + { + "bbox": [ + 302, + 592, + 525, + 740 + ], + "type": "inline_equation", + "content": "\\text{UNION+TRANS}" + }, + { + "bbox": [ + 302, + 592, + 525, + 740 + ], + "type": "text", + "content": ". We train one model for each language, on the combination of UNION triples and the translated train triples from English Open KB to that language. " + }, + { + "bbox": [ + 302, + 592, + 525, + 740 + ], + "type": "inline_equation", + "content": "\\text{UNION+TRANS}" + }, + { + "bbox": [ + 302, + 592, + 525, + 740 + ], + "type": "text", + "content": " is better than UNION by " + }, + { + "bbox": [ + 302, + 592, + 525, + 740 + ], + "type": "inline_equation", + "content": "25.9\\%" + }, + { + "bbox": [ + 302, + 592, + 525, + 740 + ], + "type": "text", + "content": " H@10 and " + }, + { + "bbox": [ + 302, + 592, + 525, + 740 + ], + "type": "inline_equation", + "content": "18.4\\%" + }, + { + "bbox": [ + 302, + 592, + 525, + 740 + ], + "type": "text", + "content": " MRR. This suggests that the model is able to benefit from English facts when they are translated to the query language, unlike in UNION where the English facts are present only in English." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 761, + 507, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 761, + 507, + 772 + ], + "spans": [ + { + "bbox": [ + 315, + 761, + 507, + 772 + ], + "type": "text", + "content": "6English source achieved the best translation quality." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "text", + "content": "204" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 71, + 68, + 527, + 166 + ], + "blocks": [ + { + "bbox": [ + 71, + 68, + 527, + 166 + ], + "lines": [ + { + "bbox": [ + 71, + 68, + 527, + 166 + ], + "spans": [ + { + "bbox": [ + 71, + 68, + 527, + 166 + ], + "type": "table", + "html": "
English (En)Hindi (Hi)Telugu (Te)Spanish (Es)Portuguese (Pt)Chinese (Zh)
H@1H@10MRRH@1H@10MRRH@1H@10MRRH@1H@10MRRH@1H@10MRRH@1H@10MRR
MONO14.838.722.83.014.87.21.58.13.96.423.712.36.321.711.42.413.16.2
UNION w/o En5.721.510.92.915.47.41.810.24.98.127.814.56.726.112.93.215.57.5
UNION16.740.824.83.616.68.11.59.34.510.632.217.69.729.316.64.018.88.9
TRANS---20.547.629.78.728.715.523.250.632.420.550.730.514.039.422.5
MONO+TRANS---20.245.428.414.338.522.223.551.532.921.448.930.717.943.226.6
UNION+TRANS---23.349.732.315.138.523.123.952.433.423.552.133.116.943.626.0
", + "image_path": "d3f48cb6c2a32bfdaa827b2e60cf500e66de206dbf122e4de14514f84dfa3ec1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 173, + 527, + 222 + ], + "lines": [ + { + "bbox": [ + 67, + 173, + 527, + 222 + ], + "spans": [ + { + "bbox": [ + 67, + 173, + 527, + 222 + ], + "type": "text", + "content": "Table 3: Performance (%) of SimKGC model on mOKB6 dataset, comprising of Open KBs in six languages. MONO, TRANS, and MONO+TRANS are monolingual models trained only on facts of one language whereas UNION, UNION w/o En, and UNION+TRANS are multilingual models trained with facts from multiple languages. All reported numbers are an average of three runs using different seeds. Best scores are highlighted in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 243, + 227, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 243, + 227, + 255 + ], + "spans": [ + { + "bbox": [ + 67, + 243, + 227, + 255 + ], + "type": "text", + "content": "5.3 Cross-lingual Memorization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 262, + 291, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 262, + 291, + 478 + ], + "spans": [ + { + "bbox": [ + 67, + 262, + 291, + 478 + ], + "type": "text", + "content": "Pretrained multilingual language models such as mBERT have demonstrated strong cross-lingual transfer capabilities (Wu and Dredze, 2019). We investigate cross-lingual memorization of the KGE model by showing facts in one language and querying the same facts in other five languages. For each language, " + }, + { + "bbox": [ + 67, + 262, + 291, + 478 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 67, + 262, + 291, + 478 + ], + "type": "text", + "content": ", we take the UNION model and train it further on the test set of that language's Open KB, which we call MEMORIZE" + }, + { + "bbox": [ + 67, + 262, + 291, + 478 + ], + "type": "inline_equation", + "content": "_L" + }, + { + "bbox": [ + 67, + 262, + 291, + 478 + ], + "type": "text", + "content": " model. Then, we test each MEMORIZE" + }, + { + "bbox": [ + 67, + 262, + 291, + 478 + ], + "type": "inline_equation", + "content": "_L" + }, + { + "bbox": [ + 67, + 262, + 291, + 478 + ], + "type": "text", + "content": " model on the six test sets. Since the test sets (in mOKB6 dataset) of the different languages contain the same facts, this experiment allows us to investigate cross-lingual memorization. We provide the H@10 scores of MEMORIZE models in Figure 3 and the performance on other metrics (H@1 and MRR) is reported in Table 7." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 479, + 291, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 479, + 291, + 642 + ], + "spans": [ + { + "bbox": [ + 67, + 479, + 291, + 642 + ], + "type": "text", + "content": "The model achieves at least " + }, + { + "bbox": [ + 67, + 479, + 291, + 642 + ], + "type": "inline_equation", + "content": "97\\%" + }, + { + "bbox": [ + 67, + 479, + 291, + 642 + ], + "type": "text", + "content": " H@10 when tested on the language used for training (diagonal). We observe that there is relatively good crosslingual memorization among languages that share the same script (Latin in English, Spanish, and Portuguese), but the model struggles to remember facts when seen in languages of different scripts. Many entities look similar in shared scripts, possibly leading to better information transfer. For example, the " + }, + { + "bbox": [ + 67, + 479, + 291, + 642 + ], + "type": "inline_equation", + "content": "\\mathsf{MEMORIZE}_{En}" + }, + { + "bbox": [ + 67, + 479, + 291, + 642 + ], + "type": "text", + "content": " achieves H@10 of " + }, + { + "bbox": [ + 67, + 479, + 291, + 642 + ], + "type": "inline_equation", + "content": "50.7\\%" + }, + { + "bbox": [ + 67, + 479, + 291, + 642 + ], + "type": "text", + "content": " in Spanish (Es) compared to " + }, + { + "bbox": [ + 67, + 479, + 291, + 642 + ], + "type": "inline_equation", + "content": "22.3\\%" + }, + { + "bbox": [ + 67, + 479, + 291, + 642 + ], + "type": "text", + "content": " in Chinese (Zh) and " + }, + { + "bbox": [ + 67, + 479, + 291, + 642 + ], + "type": "inline_equation", + "content": "11\\%" + }, + { + "bbox": [ + 67, + 479, + 291, + 642 + ], + "type": "text", + "content": " in Telugu (Te)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 655, + 239, + 667 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 655, + 239, + 667 + ], + "spans": [ + { + "bbox": [ + 67, + 655, + 239, + 667 + ], + "type": "text", + "content": "6 Conclusion and Future Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 678, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 678, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 678, + 291, + 773 + ], + "type": "text", + "content": "We create and release the mOKB6 dataset, the first multilingual Open Knowledge Base Completion dataset with 42K facts in six languages: English, Hindi, Telugu, Spanish, Portuguese, and Chinese. Its construction uses multilingual coreference resolution, entity-mention cluster naming, multilingual open information extraction and various filtering" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 326, + 256, + 494, + 391 + ], + "blocks": [ + { + "bbox": [ + 326, + 256, + 494, + 391 + ], + "lines": [ + { + "bbox": [ + 326, + 256, + 494, + 391 + ], + "spans": [ + { + "bbox": [ + 326, + 256, + 494, + 391 + ], + "type": "image", + "image_path": "a1dc8f82f5d3ed3389f4fadaa56129d4f2c29635a09c9311958a3e04644f4744.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 406, + 527, + 526 + ], + "lines": [ + { + "bbox": [ + 302, + 406, + 527, + 526 + ], + "spans": [ + { + "bbox": [ + 302, + 406, + 527, + 526 + ], + "type": "text", + "content": "Figure 3: Performance (H@10) of MEMORIZE models. Row " + }, + { + "bbox": [ + 302, + 406, + 527, + 526 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 302, + 406, + 527, + 526 + ], + "type": "text", + "content": " shows the performance of " + }, + { + "bbox": [ + 302, + 406, + 527, + 526 + ], + "type": "inline_equation", + "content": "\\text{MEMORIZE}_L" + }, + { + "bbox": [ + 302, + 406, + 527, + 526 + ], + "type": "text", + "content": " model across the test sets of all languages (columns). For example, the performance of " + }, + { + "bbox": [ + 302, + 406, + 527, + 526 + ], + "type": "inline_equation", + "content": "\\text{MEMORIZE}_{En}" + }, + { + "bbox": [ + 302, + 406, + 527, + 526 + ], + "type": "text", + "content": " when tested on English (En) is " + }, + { + "bbox": [ + 302, + 406, + 527, + 526 + ], + "type": "inline_equation", + "content": "97.1\\%" + }, + { + "bbox": [ + 302, + 406, + 527, + 526 + ], + "type": "text", + "content": " H@10, and " + }, + { + "bbox": [ + 302, + 406, + 527, + 526 + ], + "type": "inline_equation", + "content": "\\text{MEMORIZE}_{En}" + }, + { + "bbox": [ + 302, + 406, + 527, + 526 + ], + "type": "text", + "content": " when tested on Spanish (Es) gives " + }, + { + "bbox": [ + 302, + 406, + 527, + 526 + ], + "type": "inline_equation", + "content": "50.7\\%" + }, + { + "bbox": [ + 302, + 406, + 527, + 526 + ], + "type": "text", + "content": " H@10. We find relatively good cross-lingual transfer among languages that use same script (Latin in English, Spanish and Portuguese) compared to those using different scripts (English, Hindi, Telugu and Chinese)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 550, + 526, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 550, + 526, + 619 + ], + "spans": [ + { + "bbox": [ + 302, + 550, + 526, + 619 + ], + "type": "text", + "content": "steps to improve the quality of the extracted facts. We also report the first baselines on the task using the existing state of the art KGE models trained with facts from different languages using various augmentation strategies." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 624, + 527, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 624, + 527, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 624, + 527, + 772 + ], + "type": "text", + "content": "Our work opens many important research questions: (1) Can we develop better strategies to combine facts in different languages? (2) Can we build models that achieve strong information transfer across unrelated languages with same or different scripts? (3) Can we train the neural model to ignore contextual triples (Appendix E), thus improving overall performance? and (4) Can tying the same entities across various languages help the model generalize better? We leave these questions to be addressed in future work." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "205" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 71, + 188, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 71, + 188, + 84 + ], + "spans": [ + { + "bbox": [ + 68, + 71, + 188, + 84 + ], + "type": "text", + "content": "7 Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 92, + 291, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 92, + 291, + 200 + ], + "spans": [ + { + "bbox": [ + 67, + 92, + 291, + 200 + ], + "type": "text", + "content": "Keshav was supported by TCS Research Fellowship during his PhD. Mausam is supported by grants from Huawei, Google, Verisk and IBM, and a Jai Gupta Chair Fellowship. He also acknowledges Google and Yardi School of AI travel grants. Soumen is partly supported by a Jagadish Bose Fellowship and a grant from Cisco. We thank IIT Delhi HPC facility for compute resources." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 210, + 149, + 223 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 210, + 149, + 223 + ], + "spans": [ + { + "bbox": [ + 67, + 210, + 149, + 223 + ], + "type": "text", + "content": "8 Limitations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 231, + 291, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 231, + 291, + 366 + ], + "spans": [ + { + "bbox": [ + 67, + 231, + 291, + 366 + ], + "type": "text", + "content": "Although multilingual, the constructed open KB is limited to the sampling of the chosen six languages. We do not know how well the system will generalize to various language families that have not been considered here. Further, even among the languages considered, the performance of even the best-performing systems, as measured through " + }, + { + "bbox": [ + 67, + 231, + 291, + 366 + ], + "type": "inline_equation", + "content": "\\mathrm{H@1}" + }, + { + "bbox": [ + 67, + 231, + 291, + 366 + ], + "type": "text", + "content": " is still in the low 20's. Therefore the models are not yet ready to be deployed for real-world applications." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 389, + 127, + 401 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 389, + 127, + 401 + ], + "spans": [ + { + "bbox": [ + 68, + 389, + 127, + 401 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 407, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 69, + 407, + 291, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 407, + 291, + 497 + ], + "spans": [ + { + "bbox": [ + 69, + 407, + 291, + 497 + ], + "type": "text", + "content": "Ivana Balazevic, Carl Allen, and Timothy Hospedales. 2019. TuckER: Tensor factorization for knowledge graph completion. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5185-5194, Hong Kong, China. Association for Computational Linguistics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 504, + 291, + 571 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 504, + 291, + 571 + ], + "spans": [ + { + "bbox": [ + 69, + 504, + 291, + 571 + ], + "type": "text", + "content": "Antoine Bordes, Nicolas Usunier, Alberto Garcia-Duran, Jason Weston, and Oksana Yakhnenko. 2013. Translating embeddings for modeling multi-relational data. In Advances in Neural Information Processing Systems, volume 26. Curran Associates, Inc." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 578, + 291, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 578, + 291, + 657 + ], + "spans": [ + { + "bbox": [ + 69, + 578, + 291, + 657 + ], + "type": "text", + "content": "Samuel Broscheit, Kiril Gashteovski, Yanjie Wang, and Rainer Gemulla. 2020. Can we predict new facts with open knowledge graph embeddings? a benchmark for open link prediction. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2296-2308, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 664, + 291, + 742 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 664, + 291, + 742 + ], + "spans": [ + { + "bbox": [ + 69, + 664, + 291, + 742 + ], + "type": "text", + "content": "Soumen Chakrabarti, Harkanwar Singh, Shubham Lohiya, Prachi Jain, and Mausam. 2022. Joint completion and alignment of multilingual knowledge graphs. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 11922-11938, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 749, + 291, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 749, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 749, + 291, + 772 + ], + "type": "text", + "content": ". Chandrahas and Partha Talukdar. 2021. OKGIT: Open knowledge graph link prediction with implicit" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 526, + 772 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 314, + 72, + 526, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 526, + 117 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 526, + 117 + ], + "type": "text", + "content": "types. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 2546-2559, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 125, + 526, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 125, + 526, + 225 + ], + "spans": [ + { + "bbox": [ + 304, + 125, + 526, + 225 + ], + "type": "text", + "content": "Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. 2014. Learning phrase representations using RNN encoder-decoder for statistical machine translation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1724-1734, Doha, Qatar. Association for Computational Linguistics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 233, + 526, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 233, + 526, + 288 + ], + "spans": [ + { + "bbox": [ + 304, + 233, + 526, + 288 + ], + "type": "text", + "content": "Eunsol Choi, Jennimaria Palomaki, Matthew Lamm, Tom Kwiatkowski, Dipanjan Das, and Michael Collins. 2021. Decontextualization: Making sentences stand-alone. Transactions of the Association for Computational Linguistics, 9:447-461." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 297, + 526, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 297, + 526, + 364 + ], + "spans": [ + { + "bbox": [ + 304, + 297, + 526, + 364 + ], + "type": "text", + "content": "Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzman, Edouard Grave, Myle Ott, Luke Zettle-moyer, and Veselin Stoyanov. 2020. Unsupervised cross-lingual representation learning at scale. In ACL Conference, pages 8440-8451." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 373, + 526, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 526, + 439 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 526, + 439 + ], + "type": "text", + "content": "Nicola De Cao, Ledell Wu, Kashyap Popat, Mikel Artetxe, Naman Goyal, Mikhail Plekhanov, Luke Zettlemoyer, Nicola Cancedda, Sebastian Riedel, and Fabio Petroni. 2022. Multilingual autoregressive entity linking. Transactions of the Association for Computational Linguistics, 10:274-290." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 448, + 526, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 448, + 526, + 546 + ], + "spans": [ + { + "bbox": [ + 304, + 448, + 526, + 546 + ], + "type": "text", + "content": "Tim Dettmers, Pasquale Minervini, Pontus Stenetorp, and Sebastian Riedel. 2018. Convolutional 2d knowledge graph embeddings. In Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence and Thirtieth Innovative Applications of Artificial Intelligence Conference and Eighth AAAI Symposium on Educational Advances in Artificial Intelligence, AAAI'18/IAAI'18/EAAI'18. AAAI Press." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 555, + 526, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 526, + 655 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 526, + 655 + ], + "type": "text", + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 663, + 526, + 730 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 663, + 526, + 730 + ], + "spans": [ + { + "bbox": [ + 304, + 663, + 526, + 730 + ], + "type": "text", + "content": "Vladimir Dobrovolskii. 2021. Word-level coreference resolution. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7670-7675, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 739, + 526, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 739, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 739, + 526, + 772 + ], + "type": "text", + "content": "Oren Etzioni, Anthony Fader, Janara Christensen, Stephen Soderland, and Mausam. 2011. Open information extraction: The second generation. In *IJCAI*" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "206" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 80, + 72, + 291, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 72, + 291, + 116 + ], + "spans": [ + { + "bbox": [ + 80, + 72, + 291, + 116 + ], + "type": "text", + "content": "2011, Proceedings of the 22nd International Joint Conference on Artificial Intelligence, Barcelona, Catalonia, Spain, July 16-22, 2011, pages 3-10. IJ-CAI/AAAI." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 126, + 291, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 126, + 291, + 192 + ], + "spans": [ + { + "bbox": [ + 69, + 126, + 291, + 192 + ], + "type": "text", + "content": "Anthony Fader, Stephen Soderland, and Oren Etzioni. 2011. Identifying relations for open information extraction. In Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 1535-1545, Edinburgh, Scotland, UK. Association for Computational Linguistics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 200, + 290, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 200, + 290, + 246 + ], + "spans": [ + { + "bbox": [ + 69, + 200, + 290, + 246 + ], + "type": "text", + "content": "Luis Galárraga, Geremy Heitz, Kevin Murphy, and Fabian M. Suchanek. 2014. Canonicalizing open knowledge bases. New York, NY, USA. Association for Computing Machinery." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 254, + 290, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 254, + 290, + 330 + ], + "spans": [ + { + "bbox": [ + 69, + 254, + 290, + 330 + ], + "type": "text", + "content": "Kiril Gashteovski, Rainer Gemulla, and Luciano del Corro. 2017. MinIE: Minimizing facts in open information extraction. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2630-2640, Copenhagen, Denmark. Association for Computational Linguistics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 340, + 290, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 340, + 290, + 396 + ], + "spans": [ + { + "bbox": [ + 69, + 340, + 290, + 396 + ], + "type": "text", + "content": "Kiril Gashteovski, Sebastian Wanner, Sven Hertling, Samuel Broscheit, and Rainer Gemulla. 2019. Opiec: An open information extraction corpus. In Proceedings of the Conference on Automatic Knowledge Base Construction (AKBC)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 404, + 290, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 404, + 290, + 493 + ], + "spans": [ + { + "bbox": [ + 69, + 404, + 290, + 493 + ], + "type": "text", + "content": "Swapnil Gupta, Sreyash Kenkre, and Partha Talukdar. 2019. CaRe: Open knowledge graph embeddings. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 378-388, Hong Kong, China. Association for Computational Linguistics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 502, + 290, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 502, + 290, + 579 + ], + "spans": [ + { + "bbox": [ + 69, + 502, + 290, + 579 + ], + "type": "text", + "content": "Bosung Kim, Taesuk Hong, Youngjoong Ko, and Jungyun Seo. 2020. Multi-task learning for knowledge graph completion with pre-trained language models. In Proceedings of the 28th International Conference on Computational Linguistics, pages 1737-1743, Barcelona, Spain (Online). International Committee on Computational Linguistics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 588, + 290, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 588, + 290, + 644 + ], + "spans": [ + { + "bbox": [ + 69, + 588, + 290, + 644 + ], + "type": "text", + "content": "Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 652, + 290, + 719 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 652, + 290, + 719 + ], + "spans": [ + { + "bbox": [ + 69, + 652, + 290, + 719 + ], + "type": "text", + "content": "Vid Kocijan and Thomas Lukasiewicz. 2021. Knowledge base completion meets transfer learning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP), Punta Cana, Dominican Republic. Association for Computational Linguistics." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 728, + 290, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 728, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 728, + 290, + 772 + ], + "type": "text", + "content": "Keshav Kolluru, Vaibhav Adlakha, Samarth Aggarwal, Mausam, and Soumen Chakrabarti. 2020. OpenIE6: Iterative Grid Labeling and Coordination Analysis for Open Information Extraction. In Proceedings of" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 525, + 772 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 315, + 72, + 525, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 525, + 116 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 525, + 116 + ], + "type": "text", + "content": "the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3748-3761, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 124, + 525, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 124, + 525, + 213 + ], + "spans": [ + { + "bbox": [ + 304, + 124, + 525, + 213 + ], + "type": "text", + "content": "Keshav Kolluru, Muqeeth Mohammed, Shubham Mittal, Soumen Chakrabarti, and Mausam. 2022. Alignment-augmented consistent translation for multilingual open information extraction. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2502-2517, Dublin, Ireland. Association for Computational Linguistics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 220, + 525, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 220, + 525, + 297 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 525, + 297 + ], + "type": "text", + "content": "Justin Lovelace and Carolyn Rosé. 2022. A framework for adapting pre-trained language models to knowledge graph completion. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 5937-5955, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 305, + 525, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 305, + 525, + 393 + ], + "spans": [ + { + "bbox": [ + 304, + 305, + 525, + 393 + ], + "type": "text", + "content": "Xin Lv, Yankai Lin, Yixin Cao, Lei Hou, Juanzi Li, Zhiyuan Liu, Peng Li, and Jie Zhou. 2022. Do pre-trained models benefit knowledge graph completion? a reliable evaluation and a reasonable approach. In Findings of the Association for Computational Linguistics: ACL 2022, pages 3570-3581, Dublin, Ireland. Association for Computational Linguistics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 401, + 525, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 525, + 435 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 525, + 435 + ], + "type": "text", + "content": "Mausam. 2016. Open information extraction systems and downstream applications. In International Joint Conference on Artificial Intelligence." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 442, + 525, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 442, + 525, + 465 + ], + "spans": [ + { + "bbox": [ + 304, + 442, + 525, + 465 + ], + "type": "text", + "content": "MediaWiki. 2021. Api:langlinks — mediawiki. [Online; accessed 02-April-2022]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 472, + 525, + 539 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 525, + 539 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 525, + 539 + ], + "type": "text", + "content": "Jeffrey Pennington, Richard Socher, and Christopher Manning. 2014. GloVe: Global vectors for word representation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1532-1543, Doha, Qatar. Association for Computational Linguistics." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 546, + 525, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 525, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 525, + 613 + ], + "type": "text", + "content": "Peng Qi, Yuhao Zhang, Yuhui Zhang, Jason Bolton, and Christopher D. Manning. 2020. Stanza: A Python natural language processing toolkit for many human languages. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 620, + 525, + 687 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 620, + 525, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 620, + 525, + 687 + ], + "type": "text", + "content": "Youngbin Ro, Yukyung Lee, and Pilsung Kang. 2020. Multi^2OIE: Multilingual open information extraction based on multi-head attention with BERT. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 1107-1117, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 694, + 525, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 694, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 694, + 525, + 772 + ], + "type": "text", + "content": "Théo Trouillon, Johannes Welbl, Sebastian Riedel, Eric Gaussier, and Guillaume Bouchard. 2016. Complex embeddings for simple link prediction. In Proceedings of The 33rd International Conference on Machine Learning, volume 48 of Proceedings of Machine Learning Research, pages 2071-2080, New York, New York, USA. PMLR." + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "207" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 291, + 150 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 291, + 150 + ], + "type": "text", + "content": "Shikhar Vashishth, Prince Jain, and Partha Talukdar. 2018. CESI: Canonicalizing open knowledge bases using embeddings and side information. In Proceedings of the 2018 World Wide Web Conference, WWW '18, pages 1317-1327, Republic and Canton of Geneva, Switzerland. International World Wide Web Conferences Steering Committee." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 158, + 291, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 158, + 291, + 236 + ], + "spans": [ + { + "bbox": [ + 69, + 158, + 291, + 236 + ], + "type": "text", + "content": "Liang Wang, Wei Zhao, Zhuoyu Wei, and Jingming Liu. 2022. SimKGC: Simple contrastive knowledge graph completion with pre-trained language models. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4281-4294, Dublin, Ireland. Association for Computational Linguistics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 243, + 291, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 243, + 291, + 300 + ], + "spans": [ + { + "bbox": [ + 69, + 243, + 291, + 300 + ], + "type": "text", + "content": "Ralph Weischedel, Martha Palmer, Mitchell Marcus, Eduard Hovy, Sameer Pradhan, Lance Ramshaw, Nianwen Xue, Ann Taylor, Jeff Kaufman, and Michelle Franchini. 2013. Ontonotes release 5.0. In Linguistic Data Consortium, Philadelphia, PA." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 307, + 291, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 307, + 291, + 395 + ], + "spans": [ + { + "bbox": [ + 69, + 307, + 291, + 395 + ], + "type": "text", + "content": "Guillaume Wenzek, Marie-Anne Lachaux, Alexis Conneau, Vishrav Chaudhary, Francisco Guzmán, Armand Joulin, and Edouard Grave. 2020. CCNet: Extracting high quality monolingual datasets from web crawl data. In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 4003-4012, Marseille, France. European Language Resources Association." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 404, + 291, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 404, + 291, + 492 + ], + "spans": [ + { + "bbox": [ + 69, + 404, + 291, + 492 + ], + "type": "text", + "content": "Shijie Wu and Mark Dredze. 2019. Beto, bentz, becas: The surprising cross-lingual effectiveness of BERT. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 833–844, Hong Kong, China. Association for Computational Linguistics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 501, + 291, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 501, + 291, + 577 + ], + "spans": [ + { + "bbox": [ + 69, + 501, + 291, + 577 + ], + "type": "text", + "content": "Patrick Xia and Benjamin Van Durme. 2021. Moving on from OntoNotes: Coreference resolution model transfer. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5241-5256, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 586, + 291, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 586, + 291, + 686 + ], + "spans": [ + { + "bbox": [ + 69, + 586, + 291, + 686 + ], + "type": "text", + "content": "Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. 2021. mT5: A massively multilingual pre-trained text-to-text transformer. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 483-498, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 694, + 291, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 694, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 694, + 291, + 772 + ], + "type": "text", + "content": "Zdeněk Žabokrtský, Miloslav Konopík, Anna Nedoluzhko, Michal Novák, Maciej Ogrodniczuk, Martin Popel, Ondřej Pražák, Jakub Sido, Daniel Zeman, and Yilun Zhu. 2022. Findings of the shared task on multilingual coreference resolution. In Proceedings of the CRAC 2022 Shared Task on Multilingual Coreference Resolution, pages 1-17," + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 314, + 72, + 524, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 524, + 95 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 524, + 95 + ], + "type": "text", + "content": "Gyeongju, Republic of Korea. Association for Computational Linguistics." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "208" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 74, + 67, + 519, + 99 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 67, + 519, + 99 + ], + "spans": [ + { + "bbox": [ + 74, + 67, + 519, + 99 + ], + "type": "text", + "content": "mOKB6: A Multilingual Open Knowledge Base Completion Benchmark (Appendix)" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 108, + 180, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 108, + 180, + 120 + ], + "spans": [ + { + "bbox": [ + 67, + 108, + 180, + 120 + ], + "type": "text", + "content": "A Dataset Curation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 131, + 290, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 131, + 290, + 238 + ], + "spans": [ + { + "bbox": [ + 67, + 131, + 290, + 238 + ], + "type": "text", + "content": "As discussed in Section 3, we construct mOKB6 dataset in three stages after extracting the Wikipedia articles (using WikiExtractor7) from the Wikidump of April 02, 2022. We run our construction pipeline (as shown in Figure 1) for all six languages on a single V100 (32 GB) GPU, which required 14 hours of computation to create mOKB6 dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 241, + 291, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 241, + 291, + 483 + ], + "spans": [ + { + "bbox": [ + 69, + 241, + 291, + 483 + ], + "type": "text", + "content": "In the first stage, we keep the sentences containing at least 6 and at most 50 tokens since we find that most of the short sentences are headings or sub-headings present in Wikipedia articles, and very long sentences can't be input to GEN2OIE (in second stage) due to maximum sequence length constraint of 1024 in mT5 (Xue et al., 2021) based GEN2OIE. This filtering step discards " + }, + { + "bbox": [ + 69, + 241, + 291, + 483 + ], + "type": "inline_equation", + "content": "18.9\\%" + }, + { + "bbox": [ + 69, + 241, + 291, + 483 + ], + "type": "text", + "content": " of sentences on an average in all six languages. We use Stanza (Qi et al., 2020) to perform sentence- and word-segmentation on Wikipedia articles in all six languages. After filtering the sentences, the articles are processed for coreference resolution using XLM-R (Conneau et al., 2020) encoder based wlcoref (Dobrovolskii, 2021), followed by replacing the coreferent cluster mentions with their canonical cluster name using the heuristic discussed in Section 3." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 486, + 290, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 486, + 290, + 565 + ], + "spans": [ + { + "bbox": [ + 67, + 486, + 290, + 565 + ], + "type": "text", + "content": "In the second stage, the coreference resolved articles are passed through GEN2OIE to get the Open IE triples. The confidence scores for these triples are computed using label rescoring, for which we refer the readers to Kolluru et al. (2022) for more details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 567, + 291, + 729 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 567, + 291, + 729 + ], + "spans": [ + { + "bbox": [ + 67, + 567, + 291, + 729 + ], + "type": "text", + "content": "Finally, in the last stage, we apply various filters, adapted from Gashteovski et al. (2019), to remove triples that are of no interest to Open KBC task, like the triples: (1) having any of its argument or relation empty, (2) containing more than 10 tokens in any of its arguments or relation, (3) having confidence score less than 0.3, (4) containing pronouns (found using Stanza) in its arguments, (5) having same subject and object (i.e. self loops), and (6) that are duplicates. These filters keep " + }, + { + "bbox": [ + 67, + 567, + 291, + 729 + ], + "type": "inline_equation", + "content": "91.6\\%" + }, + { + "bbox": [ + 67, + 567, + 291, + 729 + ], + "type": "text", + "content": " of the triples obtained from stage 2 in all six languages." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 109, + 526, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 109, + 526, + 257 + ], + "spans": [ + { + "bbox": [ + 302, + 109, + 526, + 257 + ], + "type": "text", + "content": "Further in the last stage, in order to create a dense Open KB containing minimum noise and maximum facts about the entities, we keep the triples having the Wikipedia article's title as either the subject phrase or object phrase and discard the rest. We do this by finding all the coreference clusters (of entity mentions) that contain the titles, then get the entities, or cluster names, of those clusters using the heuristic discussed in section 3, and keep those triples that contain these cluster names. This filtering step retains " + }, + { + "bbox": [ + 302, + 109, + 526, + 257 + ], + "type": "inline_equation", + "content": "23.6\\%" + }, + { + "bbox": [ + 302, + 109, + 526, + 257 + ], + "type": "text", + "content": " of the triples." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 303, + 269, + 365, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 269, + 365, + 282 + ], + "spans": [ + { + "bbox": [ + 303, + 269, + 365, + 282 + ], + "type": "text", + "content": "B Metrics" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 291, + 525, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 291, + 525, + 454 + ], + "spans": [ + { + "bbox": [ + 302, + 291, + 525, + 454 + ], + "type": "text", + "content": "We follow the previous works (Wang et al., 2022) on the evaluation methodology of Open KBC task and apply it to the multilingual Open KBC task, containing facts in multiple languages. Given an Open KB, containing a finite set of entities and open relations, the KGE model answers forward and backward queries of the form " + }, + { + "bbox": [ + 302, + 291, + 525, + 454 + ], + "type": "inline_equation", + "content": "(s,r,?)" + }, + { + "bbox": [ + 302, + 291, + 525, + 454 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 302, + 291, + 525, + 454 + ], + "type": "inline_equation", + "content": "(?,r,o)" + }, + { + "bbox": [ + 302, + 291, + 525, + 454 + ], + "type": "text", + "content": " respectively. The model ranks all the entities based on their correctness with, say, " + }, + { + "bbox": [ + 302, + 291, + 525, + 454 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 302, + 291, + 525, + 454 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 302, + 291, + 525, + 454 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 302, + 291, + 525, + 454 + ], + "type": "text", + "content": " in the forward query. Further, the evaluation is in filtered setting, where the other known correct answers, apart from " + }, + { + "bbox": [ + 302, + 291, + 525, + 454 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 302, + 291, + 525, + 454 + ], + "type": "text", + "content": ", are removed from rank list." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 455, + 525, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 455, + 525, + 562 + ], + "spans": [ + { + "bbox": [ + 302, + 455, + 525, + 562 + ], + "type": "text", + "content": "The commonly used evaluation metrics are hits at rank N (H@N), where " + }, + { + "bbox": [ + 302, + 455, + 525, + 562 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 302, + 455, + 525, + 562 + ], + "type": "text", + "content": " is a natural number, and mean reciprocal rank (MRR). Suppose, the model ranks " + }, + { + "bbox": [ + 302, + 455, + 525, + 562 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 302, + 455, + 525, + 562 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 302, + 455, + 525, + 562 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 302, + 455, + 525, + 562 + ], + "type": "text", + "content": " among all entities. Then, H@N measures how many times " + }, + { + "bbox": [ + 302, + 455, + 525, + 562 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 302, + 455, + 525, + 562 + ], + "type": "text", + "content": " is less than or equal to " + }, + { + "bbox": [ + 302, + 455, + 525, + 562 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 302, + 455, + 525, + 562 + ], + "type": "text", + "content": ". MRR is the average of reciprocal ranks " + }, + { + "bbox": [ + 302, + 455, + 525, + 562 + ], + "type": "inline_equation", + "content": "\\left( \\frac{1}{R} \\right)" + }, + { + "bbox": [ + 302, + 455, + 525, + 562 + ], + "type": "text", + "content": ". Both, H@N and MRR, are computed as average over both forms of queries over the full test set." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 575, + 523, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 575, + 523, + 588 + ], + "spans": [ + { + "bbox": [ + 302, + 575, + 523, + 588 + ], + "type": "text", + "content": "C Knowledge Graph Embedding Models" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 597, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 597, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 597, + 526, + 772 + ], + "type": "text", + "content": "SimKGC (Wang et al., 2022) is a text-based KGE model that uses two unshared pretrained BERT models (Devlin et al., 2019) for encoding (subject phrase; relation phrase) and object phrase separately. GRU-ConvE (Kocijan and Lukasiewicz, 2021) encodes both the relation phrase and argument phrase from their surface forms using two unshared GRU (Cho et al., 2014). CaRe (Gupta et al., 2019) learns separate embeddings for each argument phrase and uses a bi-directional GRU to encode the relation phrase from its surface form. Both, GRU-ConvE and CaRe, are initialised with Glove embeddings (Pennington et al., 2014)." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 750, + 244, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 750, + 244, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 750, + 244, + 772 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 67, + 750, + 244, + 772 + ], + "type": "text", + "content": "https://github.com/samuelbroscheit/wikiextractor-wikimentions" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "text", + "content": "209" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 68, + 68, + 526, + 132 + ], + "blocks": [ + { + "bbox": [ + 68, + 68, + 526, + 132 + ], + "lines": [ + { + "bbox": [ + 68, + 68, + 526, + 132 + ], + "spans": [ + { + "bbox": [ + 68, + 68, + 526, + 132 + ], + "type": "image", + "image_path": "5806f2ef64071ea69556004f4def11da50b67e67341ac2a61a4802cd3edd4777.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 138, + 525, + 176 + ], + "lines": [ + { + "bbox": [ + 66, + 138, + 525, + 176 + ], + "spans": [ + { + "bbox": [ + 66, + 138, + 525, + 176 + ], + "type": "text", + "content": "Figure 4: Previous Open KB construction pipelines like Gashteovski et al. (2019) (shown by green arrows) lack coreference resolution system, which result in filtering important facts like (Barack Obama; returned to Honolulu, Hawaii in; 1971). Our pipeline (shown by blue arrows) increases the coverage of facts due to mCoref system." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 196, + 292, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 196, + 292, + 317 + ], + "spans": [ + { + "bbox": [ + 66, + 196, + 292, + 317 + ], + "type": "text", + "content": "To choose the best model for our experiments (Table 3, Figure 3), we train the recent knowledge graph embedding (KGE) models — CaRe., GRUConvE and SimKGC on the English Open KB in mOKB6. We report performance in Table 4 using the three metrics: hits at rank 1 (H@1), hits at 10 (H@10), and mean reciprocal rank (MRR). We find that SimKGC with BERT encoder outperforms the other two models." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 92, + 325, + 267, + 407 + ], + "blocks": [ + { + "bbox": [ + 92, + 325, + 267, + 407 + ], + "lines": [ + { + "bbox": [ + 92, + 325, + 267, + 407 + ], + "spans": [ + { + "bbox": [ + 92, + 325, + 267, + 407 + ], + "type": "table", + "html": "
H@1H@10MRR
CaRe6.611.38.3
GRU-ConvE12.427.817.8
SimKGC (BERT)16.140.024.3
SimKGC (mBERT)14.838.722.8
SimKGC (XLM-R)13.835.821.3
", + "image_path": "de213892889db43fdde203a034d8a16a57326f1090933c812519f9b97ed00732.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 468, + 291, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 468, + 291, + 631 + ], + "spans": [ + { + "bbox": [ + 67, + 468, + 291, + 631 + ], + "type": "text", + "content": "Since BERT supports only English language, we replace BERT in SimKGC with multilingual pretrained language models like mBERT (Devlin et al., 2019) or XLM-R (Conneau et al., 2020), to extend SimKGC model to other languages. We find in Table 4 that SimKGC with mBERT is better than with XLM-R by " + }, + { + "bbox": [ + 67, + 468, + 291, + 631 + ], + "type": "inline_equation", + "content": "2.9\\%" + }, + { + "bbox": [ + 67, + 468, + 291, + 631 + ], + "type": "text", + "content": " H@10 and " + }, + { + "bbox": [ + 67, + 468, + 291, + 631 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 67, + 468, + 291, + 631 + ], + "type": "text", + "content": " MRR, possibly because mBERT (and mOKB6) uses Wikipedia while XLM-R uses CommonCrawl (Wenzek et al., 2020) during pre-training. Thus, we use SimKGC with mBERT as the underlying encoder to run our experiments for all the languages." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 642, + 240, + 656 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 642, + 240, + 656 + ], + "spans": [ + { + "bbox": [ + 67, + 642, + 240, + 656 + ], + "type": "text", + "content": "D KGE Model Training Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 665, + 292, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 665, + 292, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 665, + 292, + 772 + ], + "type": "text", + "content": "We use the code from official repositories of the KGE models — SimKGC (Wang et al., 2022), GRU-ConvE (Kocijan and Lukasiewicz, 2021), and CaRe (Gupta et al., 2019) for our experiments. The models are trained using Adam optimizer (Kingma and Ba, 2015) on a single A100 (40 GB) GPU with three different random seeds and we report the average of three evaluation runs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 196, + 526, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 196, + 526, + 344 + ], + "spans": [ + { + "bbox": [ + 302, + 196, + 526, + 344 + ], + "type": "text", + "content": "We do not perform hyperparameter search trials, except for batch size, and use the default hyperparameters from the respective codes of KGE models (see Table 5). We use early stopping to find the best model checkpoints based on HITS@1. The dev set is different for each baseline: MONO, TRANS, MONO+TRANS, and UNION+TRANS use individual language's dev set, whereas UNION w/o En and UNION use the English dev set. We report the performance of baseline models on the dev sets in Table 9 and Table 10." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 314, + 352, + 515, + 438 + ], + "blocks": [ + { + "bbox": [ + 67, + 414, + 291, + 452 + ], + "lines": [ + { + "bbox": [ + 67, + 414, + 291, + 452 + ], + "spans": [ + { + "bbox": [ + 67, + 414, + 291, + 452 + ], + "type": "text", + "content": "Table 4: Performance " + }, + { + "bbox": [ + 67, + 414, + 291, + 452 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 67, + 414, + 291, + 452 + ], + "type": "text", + "content": " of the KGE models on the English test set in mOKB6 dataset. The reported numbers are an average of three runs using different seeds." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 352, + 515, + 438 + ], + "lines": [ + { + "bbox": [ + 314, + 352, + 515, + 438 + ], + "spans": [ + { + "bbox": [ + 314, + 352, + 515, + 438 + ], + "type": "table", + "html": "
HyperparameterSimKGCGRU-ConvECaRe
#epochs100500500
#patience epochs101010
learning rate3e-53e-41e-3
dropout0.10.30.5
batch size2561024128
additive margin0.02N/AN/A
", + "image_path": "d3718ead02b038fcf077c22e328c0419ffb65e3796554eecfaa92302a8b3197d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 473, + 525, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 473, + 525, + 568 + ], + "spans": [ + { + "bbox": [ + 302, + 473, + 525, + 568 + ], + "type": "text", + "content": "We provide the number of trainable parameters of each KGE model in Table 6. Based on the batch size and model size, different experiments consume different GPU hours. To train on English Open KB (in mOKB6 dataset), CaRe and GRU-ConvE models took 2.5 hours and 0.5 hours, respectively, whereas SimKGC takes nearly 1 hour of GPU time." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 332, + 576, + 496, + 652 + ], + "blocks": [ + { + "bbox": [ + 318, + 445, + 508, + 458 + ], + "lines": [ + { + "bbox": [ + 318, + 445, + 508, + 458 + ], + "spans": [ + { + "bbox": [ + 318, + 445, + 508, + 458 + ], + "type": "text", + "content": "Table 5: Hyperparameters of the KGE models." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 332, + 576, + 496, + 652 + ], + "lines": [ + { + "bbox": [ + 332, + 576, + 496, + 652 + ], + "spans": [ + { + "bbox": [ + 332, + 576, + 496, + 652 + ], + "type": "table", + "html": "
KGE model#trainable parameters
CaRe12,971,423
GRU-ConvE12,085,523
SimKGC (BERT)216,620,545
SimKGC (mBERT)355,706,881
SimKGC (XLM-R)1,119,780,865
", + "image_path": "566686bbdf909d805d01895739fe45773e1eb857ef3b1e52fe8bd092bd59995c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 660, + 525, + 683 + ], + "lines": [ + { + "bbox": [ + 302, + 660, + 525, + 683 + ], + "spans": [ + { + "bbox": [ + 302, + 660, + 525, + 683 + ], + "type": "text", + "content": "Table 6: Number of trainable parameters in the KGE models." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 780, + 307, + 791 + ], + "type": "text", + "content": "210" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 71, + 68, + 526, + 169 + ], + "blocks": [ + { + "bbox": [ + 71, + 68, + 526, + 169 + ], + "lines": [ + { + "bbox": [ + 71, + 68, + 526, + 169 + ], + "spans": [ + { + "bbox": [ + 71, + 68, + 526, + 169 + ], + "type": "table", + "html": "
EnglishHindiTeluguSpanishPortugueseChinese
H@1H@10MRRH@1H@10MRRH@1H@10MRRH@1H@10MRRH@1H@10MRRH@1H@10MRR
English68.497.178.83.417.28.31.611517.850.728.61744.6265.422.311.1
Hindi1942.226.780.699.588.32.412.55.912.33619.912.333.919.75.321.910.8
Telugu19.542.227.24.318.79.474.499.584.210.935.418.910.73418.54.721.410.1
Spanish27.960.438.84.117.88.91.810.75.18410090.337.67450.16.524.912.8
Portuguese27.858.738.24.418.29.31.710.55.141.578.553.684.299.990.86.62613.2
Chinese22.148.430.63.518.58.81.812.25.414.842.824.215.741.624.181.699.889.2
", + "image_path": "1ae8b58d54b659eadae84e2e9e7bed390e9eee1cd63c89c1084ca7c1d879daa6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 177, + 525, + 201 + ], + "lines": [ + { + "bbox": [ + 67, + 177, + 525, + 201 + ], + "spans": [ + { + "bbox": [ + 67, + 177, + 525, + 201 + ], + "type": "text", + "content": "Table 7: Performance (%) of the six MEMORIZE models, which have been trained on each language's test set and tested on all the test sets in mOKB6 dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 222, + 187, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 222, + 187, + 236 + ], + "spans": [ + { + "bbox": [ + 67, + 222, + 187, + 236 + ], + "type": "text", + "content": "E Contextual Triples" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 243, + 290, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 243, + 290, + 338 + ], + "spans": [ + { + "bbox": [ + 67, + 243, + 290, + 338 + ], + "type": "text", + "content": "Open IE triples are of various kinds and not all of them can be used for Open KBC task. Various filtering steps are used to remove some of these in data curation (Section 3). We define contextual triples as another kind of noisy triples, which are specific to, and are not interpretable out of, the context of text from which they are extracted." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 349, + 250, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 349, + 250, + 359 + ], + "spans": [ + { + "bbox": [ + 107, + 349, + 250, + 359 + ], + "type": "text", + "content": "(Max Born; continued; scientific work)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 359, + 251, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 359, + 251, + 370 + ], + "spans": [ + { + "bbox": [ + 107, + 359, + 251, + 370 + ], + "type": "text", + "content": "(Robb Gravett; won; the championship)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 90, + 369, + 268, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 369, + 268, + 379 + ], + "spans": [ + { + "bbox": [ + 90, + 369, + 268, + 379 + ], + "type": "text", + "content": "(George Herbert Walker Bush; was; out of touch)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 129, + 379, + 229, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 379, + 229, + 389 + ], + "spans": [ + { + "bbox": [ + 129, + 379, + 229, + 389 + ], + "type": "text", + "content": "(Christianity; is; dominant)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 98, + 398, + 260, + 411 + ], + "lines": [ + { + "bbox": [ + 98, + 398, + 260, + 411 + ], + "spans": [ + { + "bbox": [ + 98, + 398, + 260, + 411 + ], + "type": "text", + "content": "Table 8: Examples of contextual triples." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 424, + 291, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 424, + 291, + 491 + ], + "spans": [ + { + "bbox": [ + 67, + 424, + 291, + 491 + ], + "type": "text", + "content": "From the first two triples in Table 8, it is unclear which scientific work Max Born continued, or which championship Robb Gravett has won. The last two triples are too specific to the context and contain no factual information." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "211" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 71, + 187, + 523, + 284 + ], + "blocks": [ + { + "bbox": [ + 71, + 187, + 523, + 284 + ], + "lines": [ + { + "bbox": [ + 71, + 187, + 523, + 284 + ], + "spans": [ + { + "bbox": [ + 71, + 187, + 523, + 284 + ], + "type": "table", + "html": "
English (En)Hindi (Hi)Telugu (Te)Spanish (Es)Portuguese (Pt)Chinese (Zh)
H@1H@10MRRH@1H@10MRRH@1H@10MRRH@1H@10MRRH@1H@10MRRH@1H@10MRR
MONO16.238.723.918.239.425.98.52012.517.336.623.717.639.625.310.831.917.8
TRANS---8.123.713.53.315.47.512.933.620.312.637.220.6520.810.3
MONO+TRANS---20.843.228.67.824.813.420.24628.82145.929.210.630.116.7
UNION19.939.626.414.538.222.45.92010.619.843.227.919.743.82811.23318.8
UNION w/o En5.819.510.615.439.323.36.320.511.119.441.626.416.942.925.911.33318.4
UNION+TRANS---20.844.928.87.327.11421.445.329.619.449.129.16.93115.1
", + "image_path": "d91a900a0365a74f15084007720fc5ea4342636b49c33865f7c6a244dedb405f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 210, + 550, + 383, + 629 + ], + "blocks": [ + { + "bbox": [ + 110, + 292, + 481, + 306 + ], + "lines": [ + { + "bbox": [ + 110, + 292, + 481, + 306 + ], + "spans": [ + { + "bbox": [ + 110, + 292, + 481, + 306 + ], + "type": "text", + "content": "Table 9: Performance (%) of SimKGC on the dev sets (of mOKB6 dataset) in six languages." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 210, + 550, + 383, + 629 + ], + "lines": [ + { + "bbox": [ + 210, + 550, + 383, + 629 + ], + "spans": [ + { + "bbox": [ + 210, + 550, + 383, + 629 + ], + "type": "table", + "html": "
H@1H@10MRR
CaRe7.111.18.5
GRU-ConvE16.831.522.1
SimKGC (BERT)20.340.127.1
SimKGC (mBERT)16.238.723.9
SimKGC (XLM-R)1736.623.2
", + "image_path": "2cbc6adebcb647e5e42b0fff45ac1e43cb9b74a46ce50b15727bb6b109119dbb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 98, + 638, + 494, + 650 + ], + "lines": [ + { + "bbox": [ + 98, + 638, + 494, + 650 + ], + "spans": [ + { + "bbox": [ + 98, + 638, + 494, + 650 + ], + "type": "text", + "content": "Table 10: Performance (%) of the KGE models on dev set of English Open KB in mOKB6 dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "212" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "spans": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "type": "text", + "content": "A For every submission:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 106, + 414, + 243 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "spans": [ + { + "bbox": [ + 77, + 106, + 317, + 132 + ], + "type": "text", + "content": "A1. Did you describe the limitations of your work? 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 143, + 346, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 143, + 346, + 170 + ], + "spans": [ + { + "bbox": [ + 77, + 143, + 346, + 170 + ], + "type": "text", + "content": "A2. Did you discuss any potential risks of your work? There are no potential risks of our work to our knowledge." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 178, + 414, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 178, + 414, + 205 + ], + "spans": [ + { + "bbox": [ + 77, + 178, + 414, + 205 + ], + "type": "text", + "content": "A3. Do the abstract and introduction summarize the paper's main claims?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 77, + 215, + 398, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 215, + 398, + 243 + ], + "spans": [ + { + "bbox": [ + 77, + 215, + 398, + 243 + ], + "type": "text", + "content": "A4. Have you used AI writing assistants when working on this paper? Left blank." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 68, + 251, + 291, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 251, + 291, + 266 + ], + "spans": [ + { + "bbox": [ + 68, + 251, + 291, + 266 + ], + "type": "text", + "content": "B Did you use or create scientific artifacts?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 80, + 270, + 96, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 270, + 96, + 281 + ], + "spans": [ + { + "bbox": [ + 80, + 270, + 96, + 281 + ], + "type": "text", + "content": "3,4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 77, + 291, + 524, + 633 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 77, + 291, + 315, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 291, + 315, + 318 + ], + "spans": [ + { + "bbox": [ + 77, + 291, + 315, + 318 + ], + "type": "text", + "content": "B1. Did you cite the creators of artifacts you used? 3,4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 77, + 327, + 463, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 327, + 463, + 354 + ], + "spans": [ + { + "bbox": [ + 77, + 327, + 463, + 354 + ], + "type": "text", + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 77, + 364, + 524, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 364, + 524, + 432 + ], + "spans": [ + { + "bbox": [ + 77, + 364, + 524, + 432 + ], + "type": "text", + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Not applicable. Left blank." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 77, + 441, + 524, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 441, + 524, + 495 + ], + "spans": [ + { + "bbox": [ + 77, + 441, + 524, + 495 + ], + "type": "text", + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Not applicable. Left blank." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 77, + 503, + 524, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 503, + 524, + 544 + ], + "spans": [ + { + "bbox": [ + 77, + 503, + 524, + 544 + ], + "type": "text", + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 77, + 553, + 524, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 553, + 524, + 633 + ], + "spans": [ + { + "bbox": [ + 77, + 553, + 524, + 633 + ], + "type": "text", + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "spans": [ + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "type": "text", + "content": "C Did you run computational experiments?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 80, + 663, + 87, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 663, + 87, + 672 + ], + "spans": [ + { + "bbox": [ + 80, + 663, + 87, + 672 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 77, + 682, + 524, + 724 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 682, + 524, + 724 + ], + "spans": [ + { + "bbox": [ + 77, + 682, + 524, + 724 + ], + "type": "text", + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Appendix D" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "spans": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "text", + "content": "ACL 2023 Responsible NLP Checklist" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "spans": [ + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "type": "text", + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "213" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 77, + 70, + 523, + 238 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 77, + 70, + 523, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 70, + 523, + 111 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 523, + 111 + ], + "type": "text", + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Appendix D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 77, + 120, + 523, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 120, + 523, + 174 + ], + "spans": [ + { + "bbox": [ + 77, + 120, + 523, + 174 + ], + "type": "text", + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? \nAppendix D" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 183, + 523, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 183, + 523, + 238 + ], + "spans": [ + { + "bbox": [ + 77, + 183, + 523, + 238 + ], + "type": "text", + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Appendix A, D" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 67, + 246, + 522, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 246, + 522, + 276 + ], + "spans": [ + { + "bbox": [ + 67, + 246, + 522, + 276 + ], + "type": "text", + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 76, + 285, + 523, + 539 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 77, + 285, + 523, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 285, + 523, + 324 + ], + "spans": [ + { + "bbox": [ + 77, + 285, + 523, + 324 + ], + "type": "text", + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 76, + 336, + 523, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 336, + 523, + 390 + ], + "spans": [ + { + "bbox": [ + 76, + 336, + 523, + 390 + ], + "type": "text", + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Not applicable. Left blank." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 76, + 399, + 523, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 399, + 523, + 454 + ], + "spans": [ + { + "bbox": [ + 76, + 399, + 523, + 454 + ], + "type": "text", + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Not applicable. Left blank." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 76, + 462, + 519, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 462, + 519, + 489 + ], + "spans": [ + { + "bbox": [ + 76, + 462, + 519, + 489 + ], + "type": "text", + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Not applicable. Left blank." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 76, + 498, + 523, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 498, + 523, + 539 + ], + "spans": [ + { + "bbox": [ + 76, + 498, + 523, + 539 + ], + "type": "text", + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Not applicable. Left blank." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "214" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 13 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/mPMR_ A Multilingual Pre-trained Machine Reader at Scale/24809581-970f-402d-a256-c7d8e2f40ff3_content_list.json b/2023/mPMR_ A Multilingual Pre-trained Machine Reader at Scale/24809581-970f-402d-a256-c7d8e2f40ff3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c5523eae8288a428842fb4b98b557e0798cbee58 --- /dev/null +++ b/2023/mPMR_ A Multilingual Pre-trained Machine Reader at Scale/24809581-970f-402d-a256-c7d8e2f40ff3_content_list.json @@ -0,0 +1,1684 @@ +[ + { + "type": "text", + "text": "mPMR: A Multilingual Pre-trained Machine Reader at Scale*", + "text_level": 1, + "bbox": [ + 181, + 89, + 823, + 111 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Weiwen Xu $^{12,\\dagger}$ Xin Li $^{2,\\ddagger}$ Wai Lam $^{1}$ Lidong Bing $^{2}$", + "bbox": [ + 262, + 134, + 742, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1The Chinese University of Hong Kong", + "bbox": [ + 341, + 153, + 660, + 170 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ DAMO Academy, Alibaba Group", + "bbox": [ + 357, + 170, + 642, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{wxxu,wlam}@se.cuhk.edu.hk {xinting.lx,l.bing}@alibaba-inc.com", + "bbox": [ + 179, + 187, + 821, + 204 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 252, + 339, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present multilingual Pre-trained Machine Reader (mPMR), a novel method for multilingual machine reading comprehension (MRC)-style pre-training. mPMR aims to guide multilingual pre-trained language models (mPLMs) to perform natural language understanding (NLU) including both sequence classification and span extraction in multiple languages. To achieve cross-lingual generalization when only source-language fine-tuning data is available, existing mPLMs solely transfer NLU capability from a source language to target languages. In contrast, mPMR allows the direct inheritance of multilingual NLU capability from the MRC-style pre-training to downstream tasks. Therefore, mPMR acquires better NLU capability for target languages. mPMR also provides a unified solver for tackling cross-lingual span extraction and sequence classification, thereby enabling the extraction of rationales to explain the sentence-pair classification process. $^{1}$", + "bbox": [ + 141, + 275, + 460, + 575 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 114, + 585, + 258, + 600 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multilingual pre-trained language models, acronymed as mPLMs, have demonstrated strong Natural language understanding (NLU) capability in a wide range of languages (Xue et al., 2021; Cai et al., 2021, 2022; Conneau et al., 2020a; Ding et al., 2022; Li et al., 2020a). In particular, mPLMs can maintain exceptional cross-lingual language understanding (XLU) capability on unseen target languages though mPLMs are only fine-tuned on resource-rich source languages like English.", + "bbox": [ + 112, + 609, + 487, + 769 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "It has been proved that optimizing cross-lingual representations of mPLMs can improve XLU ca", + "bbox": [ + 112, + 771, + 487, + 802 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/56b276514635dad5b83103bb1b86a7c6db5594d38a67f412d88b687c12d37b95.jpg", + "image_caption": [ + "Figure 1: Pre-training and fine-tuning of mPMR." + ], + "image_footnote": [], + "bbox": [ + 517, + 252, + 873, + 451 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "pability. For example, cross-lingual supervisions, such as parallel sentences (Conneau and Lample, 2019) or bilingual dictionaries (Conneau et al., 2020b) could enhance cross-lingual representations with better language alignment. XLM-R (Conneau et al., 2020a) and mT5 (Xue et al., 2021) showed that appropriately incorporating more languages during pre-training leads to better cross-lingual representations. A few works enriched the cross-lingual representations with factual knowledge through the utilization of multilingual mentions of entities (Calixto et al., 2021; Ri et al., 2022) and relations (Liu et al., 2022; Jiang et al., 2022) annotated in knowledge graphs. Despite their differences, the above methods essentially constructed more diverse multilingual corpora for pre-training mPLMs. These mPLMs would presumably meet their saturation points and are known to suffer from curse of multilinguality (Conneau et al., 2020a; Pfeiffer et al., 2022; Berend, 2022). Under this situation, introducing more training data from either existing (Pfeiffer et al., 2022) or unseen (Conneau et al., 2020a) languages for enhancing mPLMs may not bring further improvement or even be detrimental to their cross-lingual representations.", + "bbox": [ + 507, + 517, + 885, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* This work was supported by Alibaba Group through Alibaba Research Intern Program. The work described in this paper was also partially supported by a grant from the Research Grant Council of the Hong Kong Special Administrative Region, China (Project Code: 14200719).† This work was done when Weiwen Xu was an intern at Alibaba DAMO Academy.‡ Xin Li is the corresponding author.", + "bbox": [ + 112, + 808, + 487, + 892 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1The code, data, and checkpoints are released at https: //github.com/DAMO-NLP-SG/PMR", + "bbox": [ + 112, + 892, + 487, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1533", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics", + "bbox": [ + 226, + 945, + 769, + 957 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Volume 2: Short Papers, pages 1533-1546", + "bbox": [ + 368, + 958, + 630, + 971 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "July 9-14, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 295, + 972, + 700, + 985 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the paper, instead of training a new mPLM with better cross-lingual representations, we propose multilingual Pre-trained Machine Reader (mPMR) to directly guide existing mPLMs to perform NLU in various languages. As shown in Figure 1, mPMR resembles PMR (Xu et al., 2022) for constructing multilingual machine reading comprehension (MRC)-style data with Wikipedia hyperlinks. These data are used to retrofit an mPLM into an mPMR through an MRC-style continual pre-training. During retrofitting process (i.e., pretraining), mPMR jointly learns the general sequence classification and span extraction capability for multiple languages. In XLU fine-tuning, mPLMs solely rely on cross-lingual representations to transfer NLU capability from a source language to target languages. By contrast, mPMR enables the direct inheritance of multilingual NLU capability from the MRC-style pre-training to downstream tasks in a unified MRC formulation, which alleviates the discrepancies between source-language fine-tuning and target-language inference (Zhou et al., 2022a,b, 2023). Therefore, mPMR shows greater potential in XLU than mPLMs.", + "bbox": [ + 115, + 84, + 490, + 469 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To improve the scalability of mPMR across multiple languages, we further propose Unified Q/C Construction and Stochastic answer position strategies for refining the curation of MRC data. With these two strategies, mPMR can better generalize to low-resource languages and becomes more robust to position bias (Ko et al., 2020).", + "bbox": [ + 112, + 470, + 489, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The experimental results show that mPMR obtains clear improvements over XLM-R (Conneau et al., 2020a) on span extraction, with an average improvement of up to 12.6 F1 on TyDiQA, and 8.7 F1 on WikiAnn respectively. The analysis reveals that mPMR benefits from more multilingual MRC data for pre-training. We also found that mPMR converges faster in downstream tasks and is capable of using its strong extraction capability for explaining the sequence classification process.", + "bbox": [ + 112, + 585, + 489, + 747 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 mPMR", + "text_level": 1, + "bbox": [ + 112, + 760, + 213, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We present the MRC model and training data of mPMR. We closely follow PMR (Xu et al., 2022) and introduce the modifications for enabling multilingual MRC-style pre-training.", + "bbox": [ + 112, + 787, + 490, + 853 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 Model Pre-training", + "text_level": 1, + "bbox": [ + 112, + 864, + 314, + 881 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our mPMR follows the same MRC architecture of Xu et al. (2022, 2023) with an encoder and an", + "bbox": [ + 112, + 887, + 487, + 917 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "extractor. The encoder maps input tokens $X$ , the concatenation of the query $Q$ , the context $C$ , and special markers (i.e., [CLS] and [SEP]), into hidden representations $H$ . For any two tokens $X_{i}$ and $X_{j}$ ( $i < j$ ), the extractor receives their contextualized representations $H_{i}$ and $H_{j}$ and predicts the probability score $S_{i,j}$ indicating the probability of the token span $X_{i:j}$ being the answer to the query $Q$ .", + "bbox": [ + 507, + 84, + 884, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "mPMR is guided with the Wiki Anchor Extraction (WAE) objective to train both the encoder and the extractor. WAE checks if the answer to the query exists in the context. If so, WAE would first regard the query and the context to be relevant and extracts the [CLS] token as a sequence-level relevance indicator. WAE would then extract all corresponding answers from the context.", + "bbox": [ + 507, + 214, + 884, + 342 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2 Multilingual MRC Data", + "text_level": 1, + "bbox": [ + 507, + 353, + 744, + 369 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Training mPMR requires the existence of labeled (query, context, answer) triplets. To obtain such data, we collected Wikipedia articles with anchor annotations for 24 languages, which are the most widely used and cover a reasonable number of languages used in XLU tasks (Ri et al., 2022).", + "bbox": [ + 507, + 374, + 882, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As shown in Figure 1, we utilized a Wikipedia anchor to obtain a pair of correlated articles. One side of the pair is the article that provides in-depth descriptions of the anchor entity, which we defined as the definition article. The other side of the pair is named as the mention article, which mentions the specific anchor text2. We composed an answerable MRC example in which the anchor is the answer, the surrounding text of the anchor in the mention article is the context, and the definition of the anchor entity in the definition article is the query. Additionally, we can generate an unanswerable MRC example by pairing a query with an irrelevant context without anchor association.", + "bbox": [ + 507, + 472, + 882, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unified Q/C Construction. PMR constructed the MRC query and context as valid sentences so as to keep the text coherent. However, sentence segmentation tools are usually not available for low-resource languages. To remedy this, we did not apply sentence segmentation but only preprocess Wikipedia articles with word tokenization in mPMR. For each anchor, the MRC query comprises the first $Q$ words in the definition article. To prevent information leakage during pre-training, similar to PMR, we anonymized the anchor entity", + "bbox": [ + 507, + 706, + 884, + 883 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "2definition/mention article refers to home/reference article of Xu et al. (2022).", + "bbox": [ + 507, + 891, + 882, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1534", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/a80c889344bf6a4cd7fabf2a39d922a5a6cbb745b7121e0224804c61cc784565.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Model#ParamsEQANERABSASentence PairAvg.
XQuADMLQATyDiQAWikiAnnCoNLLSemEval16PAWS-XXNLI
MetricsF1 / EMF1 / EMF1 / EMF1F1F1Acc.Acc.
XLM-R550M76.6 / 60.871.6 / 53.265.1 / 45.065.482.066.9‡86.479.274.2
mT5580M67.0 / 49.064.6 / 45.057.2 / 41.255.771.0‡62.5‡86.475.467.5
VECO550M77.3 / 61.871.7 / 53.267.6 / 49.165.781.3‡63.0‡88.779.974.4
mLUKE-W561M79.6 / -72.7 / -65.2 / 48.5‡67.7‡83.061.2‡88.2‡79.4‡74.6
Wiki-CL550M72.1 / 56.970.8 / 50.573.2 / 57.364.7--88.479.2-
KMLM550M77.3 / 61.772.1 / 53.767.9 / 50.466.7‡83.266.1‡88.079.275.1
Our MRC Formulation
XLM-Rbase270M70.8 / 56.964.4 / 47.950.8 / 38.257.979.260.085.073.367.7
mPMRbase270M74.0 / 59.565.3 / 48.763.4 / 49.066.681.762.186.173.671.6
XLM-R550M77.1 / 61.371.5 / 53.967.4 / 51.663.681.466.186.978.674.1
mPMR550M79.2 / 64.473.1 / 55.474.7 / 58.370.784.168.288.079.377.2
", + "bbox": [ + 126, + 80, + 870, + 296 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1: The results of all XLU tasks. We report the average results of all languages for each dataset. We also compute the overall average score among all datasets in the Avg. column. We reproduce the missing results with the $\\ddagger$ label. Some results of Wiki-CL are left blank because they do not release their model checkpoint.", + "bbox": [ + 112, + 307, + 882, + 350 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "in the query to the [MASK] token. The MRC context consists of $C$ words surrounding the anchor.", + "bbox": [ + 112, + 376, + 485, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Stochastic Answer Position. As mentioned by Ko et al. (2020), the model is prone to overfitting to the position shortcut if the answer in the context exhibits a fixed position pattern. In our case, suppose that the MRC context consists of $C / 2$ words on both the left and right sides of the anchor, the model may learn the shortcut that the middle part of the context is likely to be the answer. To prevent such position bias, we propose a stochastic answer position method, which allows the answer to be presented in any position within the context. Specifically, given an anchor in a Wikipedia article, the context comprises $\\xi$ words preceding the anchor and the $C - \\xi$ words following the anchor, where $\\xi$ is a random integer ranging from 0 to $C$ and varies across different contexts. In accordance with PMR, we treated all text spans identical to the anchor in the current context as valid answers.", + "bbox": [ + 115, + 418, + 489, + 706 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 Experimental Setup", + "text_level": 1, + "bbox": [ + 112, + 721, + 321, + 737 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Implementation Details. In mPMR, the encoder is loaded from XLM-R (Conneau et al., 2020a) and the extractor is randomly initialized. Both components are then continually pre-trained using the multilingual MRC data that we constructed. More hyper-parameters can be found in Appendix A.1.", + "bbox": [ + 112, + 747, + 487, + 843 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Downstream XLU Tasks. We evaluated mPMR on a series of span extraction tasks, including Extractive Question Answering (EQA), Named Entity Recognition (NER), and Aspect-Based Sentiment", + "bbox": [ + 112, + 854, + 487, + 917 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Analysis (ABSA). We also evaluated our mPMR on two sequence classification tasks. We followed Xu et al. (2022) to convert all tasks into MRC formulation to effectively leverage the knowledge that is acquired during MRC-style pre-training. For EQA, we used XQuAD (Artetxe et al., 2020), MLQA (Lewis et al., 2020), and TyDiQA (Clark et al., 2020). For NER, we used WikiAnn (Pan et al., 2017) and CoNLL (Tjong Kim Sang, 2002; Tjong Kim Sang and De Meulder, 2003). SemEval16 (Pontiki et al., 2016) was used for ABSA task. Regarding the sequence classification, we used XNLI (Conneau et al., 2018) and PAWS-X (Yang et al., 2019). Additional dataset information and concrete examples are provided in Appendix A.2", + "bbox": [ + 507, + 376, + 884, + 617 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Baselines. We compared mPMR with recent methods on improving cross-lingual representations, including 1) models pre-trained on a large number of languages: XLM-R (Conneau et al., 2020a), mT5 (Xue et al., 2021), and VECO (Luo et al., 2021); 2) models that exploited multilingual entity information: Wiki-CL (Calixto et al., 2021), and mLUKE-W (Ri et al., 2022); and 3) Model that utilized multilingual relation information: KMLM (Liu et al., 2022). For a fair comparison, all models have approximately the same parameter size.", + "bbox": [ + 507, + 625, + 882, + 802 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4 Results and Analyses", + "text_level": 1, + "bbox": [ + 507, + 814, + 726, + 829 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "XLU Performance. Table 1 shows the results on a variety of XLU tasks. mPMR outperforms all previous methods with an absolute improvement of 2.1 F1 over the best baseline (i.e. KMLM). mPMR shows greater improvements over previ", + "bbox": [ + 507, + 839, + 882, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1535", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/148e11b05851dacace3c001a12604c9de3997c40138ff8dcee66c90068b62d97.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
IndexModel#LangPAWS-XXQuADWikiAnnAvg.
#1XLM-Rbase085.070.857.971.2
#2#1 + MRC data in English185.2 (0.2↑)71.0 (0.2↑)59.5 (1.6↑)71.9 (0.7↑)
#3#2 + Stochastic Answer Position185.5 (0.3↑)73.0 (2.0↑)60.0 (0.5↑)72.8 (0.9↑)
#4#3 + MRC data in more languages1085.9 (0.4↑)73.5 (0.5↑)64.7 (4.7↑)74.7 (1.9↑)
#5#4 + MRC data in even more languages (mPMRbase)2486.1 (0.2↑)74.0 (0.5↑)66.6 (1.9↑)75.6 (0.9↑)
", + "bbox": [ + 115, + 82, + 885, + 170 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/9556b4fbac5433b04c4c78b23b9cd720fc6199f7aa29a32cc877240a09373244.jpg", + "table_caption": [ + "Table 2: The process of retrofitting XLM-R into mPMR using multilingual MRC data (English→10 languages→24 languages) and our Stochastic Answer Position method. Each row accumulates modifications from all rows above." + ], + "table_footnote": [], + "table_body": "
LabelSentence 1Sentence 2
EntailmentRami Nieminen ( born February 25 , 1966 ) is a Finnish footballer.Rami Nieminen ( born 25 February 1966 ) is a Finnish former footballer.
ContradictionIn 1938 he became the Government Anthropologist of the Egyptian-Anglo Sudan and conducted fieldwork with the Nuba.In 1938 he became the government anthropologist of the anglo-Egyptian Sudan and led fieldwork with the Nuba.
EntailmentStipsits 出生于科尔新堡，并在维也纳施塔莫斯多夫度过了他的童年。什蒂普西奇出生于德国科恩堡，在维也纳斯塔莫斯多夫度过了他的童年。
Contradiction纳舒厄白银骑士团队加入了夏季大学联盟，是本市的现役球队。Nashua Silver Knights 队是当前夏季联赛的一部分，也是该市的大学体育队。
Entailmentごれらの見方は、福音主義的、清教徒的、プロデ斯特兰トの動態が出現すると必に、しはしだは表明くださいます。ごれらの見解は多くの场合、新教徒、清教徒、福音主義者が出現する:NOか表示お願いいたします。
Contradiction1954年にスリーナムに戸った後、弁護士とでラマリポに定住したこと。1954年、バラマリポに戸ると、彼はスリーナムで弁護士とで定住,No理由。
", + "bbox": [ + 115, + 221, + 882, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 3: Case study on PAWS-X. mPMR can extract rationales to explain the sequence-pair classification in multiple languages.", + "bbox": [ + 112, + 442, + 882, + 472 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ous methods on span extraction tasks. In particular, mPMR achieves up to 7.3 and 7.1 F1 improvements over XLM-R on TyDiQA and WikiAnn respectively. Such significant improvements probably come from the following two facts: (1) WikiAnn comprises a larger number of target languages (i.e. 40). Therefore, existing methods may struggle to align these low-resource languages with English due to a lack of language-specific data. (2) TyDiQA is a more challenging cross-lingual EQA task with $2\\mathrm{x}$ less lexical overlap between the query and the answer than MLQA and XQuAD (Hu et al., 2020). Our mPMR, which acquires target-language span extraction capability from both MRC-style pretraining and English-only QA fine-tuning, achieves larger performance gains on more challenging task.", + "bbox": [ + 112, + 495, + 489, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "mPMR Pre-training. To reflect the impact of our MRC-style data and Stochastic Answer Position method on pre-training, we present a step-by-step analysis of the retrofitting process starting from XLM-R in Table 2. Our findings suggest that the significant improvements observed are largely due to the inclusion of multilingual MRC data. Introducing English MRC data (model #2) gives marginal improvements because model #2", + "bbox": [ + 112, + 774, + 489, + 919 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fde14688ed094287da5e02c2f3852fb01624809fa4b9d24f6b0d5328fe8edd71.jpg", + "image_caption": [ + "Figure 2: Convergence speed (Test set F1 and the training loss) of $\\mathrm{mPMR_{base}}$ and XLM- $\\mathbf{R}_{\\mathrm{base}}$ on WikiAnn." + ], + "image_footnote": [], + "bbox": [ + 521, + 494, + 870, + 618 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "can only rely on cross-lingual representations to transfer the knowledge acquired during MRC-style pre-training. When using MRC data on more languages (model #4 and #5), we can observe significant improvements on XLU tasks. This can be attributed to the NLU capability directly inherited from MRC-style pre-training in target languages. Additionally, with our Stochastic Answer Position method (model #3), mPMR becomes more robust to position bias and thus improves XLU tasks.", + "bbox": [ + 507, + 682, + 884, + 843 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Explainable Sentence-pair Classification. Inspired by PMR (Xu et al., 2022), we investigated if the extraction capability of mPMR can be leveraged to explain sentence-pair classification. Note", + "bbox": [ + 507, + 854, + 884, + 919 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1536", + "bbox": [ + 482, + 928, + 521, + 940 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/088e81490d4be3fc5e3553fa86604e19088ecab031db99dbd74c610ab80eb61f.jpg", + "image_caption": [ + "Figure 3: Convergence speed (Test set F1 and the training loss) of $\\mathrm{mPMR_{base}}$ and XLM- $\\mathbf{R}_{\\mathrm{base}}$ on XQuAD." + ], + "image_footnote": [], + "bbox": [ + 117, + 80, + 482, + 211 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "that sentence-pair classification focuses on the inference between the two sentences. If we construct the query with only the task label as PMR does, such query does not solely correspond to any meaningful span in the context, and thus is hard to guide the span extraction. Therefore, we leveraged another template \"[CLS] label Sen-1 [SEP] Sen-2 [SEP]\", where the two sentences are represented separately in the query and the context. In this template, we can extract the exact span from Sen-2 that leads to a contraction or entailment relation (i.e., the task label) with Sen-1. Specifically, we passed the sentence pair to the model twice, with each sentence of the pair being designated as the Sen-2 respectively, and extract the context span with the highest probability score from both sentences.", + "bbox": [ + 112, + 291, + 489, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Table 3, the extracted spans are indeed important rationales that determine the relationship between two sentences. Such a finding confirms that the extraction capability of mPMR can be appropriately used for explaining the sentence-pair classification process. While the extraction capability may affect the learning of sequence classification during fine-tuning, resulting in a 0.4 Acc. decrease on XNLI.", + "bbox": [ + 112, + 558, + 489, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "mPMR Fine-tuning. We investigated the effects of mPMR on XLU fine-tuning. Figure 2 shows that mPMR converges faster than XLM-R on WikiAnn with an extremely low loss value even fine-tuned for 500 steps. In terms of test set performance, mPMR outperforms XLM-R comprehensively and exhibits greater stability. As a result, mPMR provides a better starting point for addressing XLU tasks compared to XLM-R. More examples from XQuAD and PAWS-X are provided in Figure 3 and 4.", + "bbox": [ + 112, + 741, + 489, + 917 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/19c3ea4971b85aae02e844457d45899a7c9aa439a0f7c0812810227e18894eb3.jpg", + "image_caption": [ + "Figure 4: Convergence speed (Test set F1 and the training loss) of $\\mathrm{mPMR_{base}}$ and XLM- $\\mathbf{R}_{\\mathrm{base}}$ on PAWS-X." + ], + "image_footnote": [], + "bbox": [ + 512, + 80, + 865, + 209 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5 Conclusions", + "text_level": 1, + "bbox": [ + 509, + 275, + 650, + 290 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This paper presents a novel multilingual MRC-style pre-training method, namely mPMR. mPMR provides a unified solver for cross-lingual span extraction and sequence classification and enables direct transfer of NLU capability from pre-training to downstream tasks. mPMR clearly improves the previous baselines and provides a possible solution to explain the sentence-pair classification process.", + "bbox": [ + 507, + 302, + 884, + 432 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Limitations", + "text_level": 1, + "bbox": [ + 509, + 445, + 616, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We identify the following two limitations of our work:", + "bbox": [ + 507, + 473, + 882, + 505 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Different from raw text, constructing MRC-style data from Wikipedia requires the existence of hyperlinks. This idea works well for resource-rich languages, such as English and Chinese. While such an idea is less effective for languages with few hyperlink annotations in Wikipedia because a small amount of MRC-style training data is difficult to guide the learning of NLU capability in those languages. A possible solution is to explore other data resources to automatically construct large-scale MRC data for pre-training.", + "- As observed in Table 1, the improvements of sequence classification tasks are less significant than those of span extraction tasks. We suggest that the existence of anchors is not a strong relevance indicator between our constructed query and context. Such a finding is also observed in Chang et al. (2020). Therefore, constructing more relevant query-context pairs for sequence classification pre-training can possibly remedy this issue." + ], + "bbox": [ + 531, + 521, + 885, + 890 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1537", + "bbox": [ + 482, + 927, + 521, + 940 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 115, + 84, + 213, + 98 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mikel Artetxe, Sebastian Ruder, and Dani Yogatama. 2020. On the cross-lingual transferability of monolingual representations. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics.", + "Giuseppe Attardi. 2015. Wikiextractor. https://github.com/attardi/wikiextractor.", + "Gábor Berend. 2022. Combating the curse of multilinguality in cross-lingual WSD by aligning sparse contextualized word representations. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies.", + "Deng Cai, Xin Li, Jackie Chun-Sing Ho, Lidong Bing, and Wai Lam. 2021. Multilingual AMR parsing with noisy knowledge distillation. In Findings of the Association for Computational Linguistics: EMNLP 2021.", + "Deng Cai, Xin Li, Jackie Chun-Sing Ho, Lidong Bing, and Wai Lam. 2022. Retrofitting multilingual sentence embeddings with Abstract Meaning Representation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing.", + "Iacer Calixto, Alessandro Raganato, and Tommaso Pasini. 2021. Wikipedia entities as rendezvous across languages: Grounding multilingual language models by predicting Wikipedia hyperlinks. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies.", + "Wei-Cheng Chang, Felix X. Yu, Yin-Wen Chang, Yiming Yang, and Sanjiv Kumar. 2020. Pre-training tasks for embedding-based large-scale retrieval. In International Conference on Learning Representations.", + "Jonathan H. Clark, Eunsol Choi, Michael Collins, Dan Garrette, Tom Kwiatkowski, Vitaly Nikolaev, and Jennimaria Palomaki. 2020. TyDi QA: A benchmark for information-seeking question answering in typologically diverse languages. Transactions of the Association for Computational Linguistics.", + "Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2020a. Unsupervised cross-lingual representation learning at scale. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics.", + "Alexis Conneau and Guillaume Lample. 2019. Crosslingual language model pretraining. In Advances in Neural Information Processing Systems.", + "Alexis Conneau, Rudy Rinott, Guillaume Lample, Adina Williams, Samuel Bowman, Holger Schwenk, and Veselin Stoyanov. 2018. XNLI: Evaluating crosslingual sentence representations. In Proceedings of" + ], + "bbox": [ + 115, + 107, + 487, + 917 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the 2018 Conference on Empirical Methods in Natural Language Processing.", + "Alexis Conneau, Shijie Wu, Haoran Li, Luke Zettlemoyer, and Veselin Stoyanov. 2020b. Emerging cross-lingual structure in pretrained language models. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics.", + "Bosheng Ding, Junjie Hu, Lidong Bing, Mahani Aljunied, Shafiq Joty, Luo Si, and Chunyan Miao. 2022. GlobalWoZ: Globalizing MultiWoZ to develop multilingual task-oriented dialogue systems. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).", + "Junjie Hu, Sebastian Ruder, Aditya Siddhant, Graham Neubig, Orhan First, and Melvin Johnson. 2020. Xtreme: A massively multilingual multi-task benchmark for evaluating cross-lingual generalisation. In International Conference on Machine Learning.", + "Xiaoze Jiang, Yaobo Liang, Weizhu Chen, and Nan Duan. 2022. Xlm-k: Improving cross-lingual language model pre-training with multilingual knowledge. In Proceedings of the AAAI Conference on Artificial Intelligence.", + "Miyoung Ko, Jinhyuk Lee, Hyunjae Kim, Gangwoo Kim, and Jaewoo Kang. 2020. Look at the first sentence: Position bias in question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP).", + "Patrick Lewis, Barlas Oguz, Rudy Rinott, Sebastian Riedel, and Holger Schwenk. 2020. MLQA: Evaluating cross-lingual extractive question answering. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics.", + "Juntao Li, Ruidan He, Hai Ye, Hwee Tou Ng, Lidong Bing, and Rui Yan. 2020a. Unsupervised domain adaptation of a pretrained cross-lingual language model. In Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, IJCAI 2020.", + "Xin Li, Lidong Bing, Wenxuan Zhang, Zheng Li, and Wai Lam. 2020b. Unsupervised cross-lingual adaptation for sequence tagging and beyond. arXiv preprint arXiv:2010.12405.", + "Linlin Liu, Xin Li, Ruidan He, Lidong Bing, Shafiq Joty, and Luo Si. 2022. Enhancing multilingual language model with massive multilingual knowledge triples. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing.", + "Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In International Conference on Learning Representations.", + "Fuli Luo, Wei Wang, Jiahao Liu, Yijia Liu, Bin Bi, Songfang Huang, Fei Huang, and Luo Si. 2021. VECO: Variable and flexible cross-lingual pre-training for" + ], + "bbox": [ + 510, + 85, + 882, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1538", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "language understanding and generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers).", + "Xiaoman Pan, Boliang Zhang, Jonathan May, Joel Nothman, Kevin Knight, and Heng Ji. 2017. Cross-lingual name tagging and linking for 282 languages. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).", + "Jonas Pfeiffer, Naman Goyal, Xi Lin, Xian Li, James Cross, Sebastian Riedel, and Mikel Artetxe. 2022. Lifting the curse of multilinguality by pre-training modular transformers. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies.", + "Maria Pontiki, Dimitris Galanis, Haris Papageorgiou, Ion Androutsopoulos, Suresh Manandhar, Mohammad AL-Smadi, Mahmoud Al-Ayyoub, Yanyan Zhao, Bing Qin, Orphée De Clercq, Véronique Hoste, Marianna Apidianaki, Xavier Tannier, Natalia Loukachevitch, Evgeniy Kotelnikov, Nuria Bel, Salud María Jiménez-Zafra, and Gülşen Eryigit. 2016. SemEval-2016 task 5: Aspect based sentiment analysis. In Proceedings of the 10th International Workshop on Semantic Evaluation (SemEval-2016).", + "Ryokan Ri, Ikuya Yamada, and Yoshimasa Tsuruoka. 2022. mLUKE: The power of entity representations in multilingual pretrained language models. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).", + "Erik F. Tjong Kim Sang. 2002. Introduction to the CoNLL-2002 shared task: Language-independent named entity recognition. In COLING-02: The 6th Conference on Natural Language Learning 2002 (CoNLL-2002).", + "Erik F. Tjong Kim Sang and Fien De Meulder. 2003. Introduction to the CoNLL-2003 shared task: Language-independent named entity recognition. In Proceedings of the Seventh Conference on Natural Language Learning at HLT-NAACL 2003.", + "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations.", + "Weiwen Xu, Xin Li, Yang Deng, Wai Lam, and Lidong Bing. 2023. Peerda: Data augmentation via modeling" + ], + "bbox": [ + 115, + 85, + 487, + 917 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "peer relation for span identification tasks. In The 61th Annual Meeting of the Association for Computational Linguistics.", + "Weiwen Xu, Xin Li, Wenxuan Zhang, Meng Zhou, Lidong Bing, Wai Lam, and Luo Si. 2022. From clozing to comprehending: Retrofitting pre-trained language model to pre-trained machine reader. arXiv preprint arXiv:2212.04755.", + "Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. 2021. mT5: A massively multilingual pre-trained text-to-text transformer. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies.", + "Yinfei Yang, Yuan Zhang, Chris Tar, and Jason Baldridge. 2019. PAWS-X: A cross-lingual adversarial dataset for paraphrase identification. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP).", + "Wenxuan Zhang, Ruidan He, Haiyun Peng, Lidong Bing, and Wai Lam. 2021. Cross-lingual aspect-based sentiment analysis with aspect term code-switching. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing.", + "Meng Zhou, Xin Li, Yue Jiang, and Lidong Bing. 2022a. Enhancing cross-lingual prompting with mask token augmentation. arXiv preprint arXiv:2202.07255.", + "Ran Zhou, Xin Li, Lidong Bing, Erik Cambria, and Chunyan Miao. 2023. Improving self-training for cross-lingual named entity recognition with contrastive and prototype learning. In *The 61th Annual Meeting of the Association for Computational Linguistics*.", + "Ran Zhou, Xin Li, Lidong Bing, Erik Cambria, Luo Si, and Chunyan Miao. 2022b. ConNER: Consistency training for cross-lingual named entity recognition. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing." + ], + "bbox": [ + 510, + 85, + 882, + 695 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1539", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A Appendix", + "text_level": 1, + "bbox": [ + 114, + 84, + 238, + 99 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A.1 More Implementation Details", + "text_level": 1, + "bbox": [ + 114, + 110, + 394, + 124 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We collect the 2022-08-01 dump3 of Wikipedia articles for the 24 languages in consideration. The statistics of each language can be found in Table 4. Then for each article, we extract the plain text with anchors via WikiExtractor (Attardi, 2015). Word tokenization is performed using spaCy4 if the language is supported, otherwise, we utilize PyThaiNLP5 for Thai and Sacremoses6 for remaining languages. For each anchor entity, we construct 10 answerable MRC examples and 10 unanswerable MRC examples as described in Sec. 2.2. Anchor entities with low frequency (below 10 occurrences for English entities and 5 occurrences for entities in other languages) were excluded.", + "bbox": [ + 112, + 130, + 489, + 355 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In mPMR, we use Huggingface's implementations of XLM-R (Wolf et al., 2020). During the pre-training stage, the query length $Q$ is set to 50 words, and the context length $C$ is set to 200 words. Both are computed before the subword segmentation. We follow the default learning rate schedule and dropout settings used in XLM-R. We use AdamW (Loshchilov and Hutter, 2019) as our optimizer. We train both $\\mathrm{mPMR_{base}}$ and mPMR on 4 A100 GPU. The learning rate is set to 1e-5, and the effective batch size for each step is set to 256 and 80 for $\\mathrm{mPMR_{base}}$ and mPMR respectively in order to maximize the usage of the GPU memory. We use the average scores of XQuAD, CoNLL, and PAWS-X to select the best mPMR checkpoint. In fact, we continually pre-train $\\mathrm{mPMR_{base}}$ and mPMR for 250,000 and 100,000 steps. The training speed is around 6250 steps per hour. The hyper-parameters of $\\mathrm{mPMR_{large}}$ on downstream XLU tasks can be found in Table 5.", + "bbox": [ + 115, + 357, + 489, + 677 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A.2 Downstream XLU Tasks", + "text_level": 1, + "bbox": [ + 114, + 690, + 356, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate mPMR on XLU tasks including both span extraction (EQA, NER, and ABSA) and sequence classification (sentence pair classification). We follow (Xu et al., 2022) to convert all tasks into MRC formulation and tackle them accordingly. We show concrete examples for each task in Table 6. Specifically, we evaluate the performance of EQA on three benchmarks: XQuAD (Artetxe et al., 2020), MLQA (Lewis et al., 2020), and Ty", + "bbox": [ + 112, + 711, + 489, + 856 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "DiQA (Clark et al., 2020) covering 11, 7, and 9 languages respectively. For NER evaluation, we use the WikiAnn dataset (Pan et al., 2017) restricted to the 40 languages from XTREME (Hu et al., 2020), as well as the CoNLL dataset with 4 languages (Tjong Kim Sang, 2002; Tjong Kim Sang and De Meulder, 2003); We also evaluate the XLU performance of SemEval16 ABSA on 6 languages (Pontiki et al., 2016), where we collect the data from Li et al. (2020b); Zhang et al. (2021). Regarding the sequence classification task, we evaluate XNLI (Conneau et al., 2018) and PAWS-X (Yang et al., 2019) with 15 and 7 languages respectively.", + "bbox": [ + 507, + 84, + 884, + 294 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A.3 mPMR Performance per Language", + "text_level": 1, + "bbox": [ + 507, + 304, + 835, + 319 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We show the detailed results for each language in each task in Table 7 (XQuAD), Table 8 (MLQA), Table 9 (TyDiQA), Table 10 (WikiAnn), Table 11 (CoNLL), Table 12 (SemEval16), Table 13 (PAWS-X), and Table 14 (XNLI).", + "bbox": [ + 507, + 325, + 884, + 404 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "3https://dumps.wikimedia.org/enwiki/latest", + "bbox": [ + 134, + 865, + 400, + 879 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "4https://github.com/explosion/spaCy", + "bbox": [ + 136, + 878, + 359, + 892 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "5https://github.com/PyThaiNLP/pythainlp", + "bbox": [ + 136, + 891, + 391, + 904 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "$^{6}$ https://github.com/ Alvations/sacremoses", + "bbox": [ + 136, + 904, + 386, + 917 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1540", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/661dbfbb6e843fe62e3662662a55c377abc972b6a939cb660e1f8aac153b5086.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Language# Entities# MRC examplesLanguage# Entities# MRC examples
ar118,2922,020,502ko94,6161,597,076
bn25,081410,634nl251,3234,185,913
de864,74614,795,826pl283,9254,765,015
el56,383946,114pt216,6953,648,603
en966,19719,303,940ru432,4377,342,472
es412,4767,044,972sv169,0302,808,214
fi113,1181,960,636sw4,85765,724
fr595,87910,164,216te11,005170,664
hi15,350242,078th31,676522,434
id70,9601,164,662tr71,2941,175,276
it376,4176,421,850vi68,6651,147,772
ja423,8847,338,308zh259,7854,438,004
Total5,934,091103,680,905
", + "bbox": [ + 163, + 186, + 835, + 437 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/e95d9666cd872126df72706db3a769093eb6d05345671c9cb46bb11694796c53.jpg", + "table_caption": [ + "Table 4: Data statistics of mPMR pre-training data. The statistics is computed after removing the low-frequency entities. The number of MRC examples includes both answerable and unanswerable examples." + ], + "table_footnote": [], + "table_body": "
DatasetXQuADMLQATyDiQAWikiAnnCoNLLSemEval16PAWS-XXNLI
Query Length6464643232326464
Input Length384384384192192192192192
Batch Size8881616321632
Learning Rate3e-53e-52e-51e-51e-52e-55e-53e-5
Epoch3310101020103
", + "bbox": [ + 178, + 695, + 818, + 785 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 5: Hyper-parameters settings in fine-tuning XLU tasks.", + "bbox": [ + 289, + 793, + 705, + 809 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1541", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/529d9483dbe31ace10a0db411bc0a58a7b4d5ae548b131db3858aa13e8bbc42e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TaskExample InputExample Output
EQA(XSQuAD)Ori.Question: Who lost to the Broncos in the divisional round?Context: The Broncos defeated the Pittsburgh Steelers in the divi-sional round, 23–16, by scoring 11 points in the final three minutes of the game.Answer: "Pittsburgh Steelers"
PMR[CLS] Who lost to the Broncos in the divisional round ? [SEP] [SEP]The Broncos defeated the Pittsburgh Steelers in the divisional round, 23–16 , by scoring 11 points in the final three minutes of the game .[SEP](17,18) - "Pittsburgh Steelers"
NER(CoNLL)Ori.Two goals in the last six minutes gave holders Japan an uninspiring 2-1 Asian Cup victory over Syria on Friday.("Japan", LOC);("Syria", LOC);("Asian Cup", MISC)
PMR[CLS] "ORG". Organization entities are limited to named corporate,governmental, or other organizational entities. [SEP] [SEP] Two goals in the last six minutes gave holders Japan an uninspiring 2-1 Asian Cup victory over Syria on Friday . [SEP]∅
[CLS] "PER". Person entities are named persons or family . [SEP] [SEP] Two goals in the last six minutes gave holders Japan an uninspiring 2-1 Asian Cup victory over Syria on Friday . [SEP]∅
[CLS] "LOC". Location entities are the name of politically or geo-graphically defined locations such as cities , countries . [SEP] [SEP] Two goals in the last six minutes gave holders Japan an uninspiring 2-1 Asian Cup victory over Syria on Friday . [SEP](32,32) - "Japan";(40,40) - "Syria"
[CLS] "MISC". Examples of miscellaneous entities include events ,nationalities , products and works of art . [SEP] [SEP] Two goals in the last six minutes gave holders Japan an uninspiring 2-1 Asian Cup victory over Syria on Friday . [SEP](34,35) - "Asian Cup"
ABSA(SemEval16)Ori.Nice ambience, but highly overrated place.("ambience", POS);("place", NEG)
PMR[CLS] "POS". For aspect terms of positive sentiment . [SEP] [SEP] Nice ambience , but highly overrated place . [SEP](13,13) - "ambience"
[CLS] "NEG". For aspect terms of negative sentiment . [SEP] [SEP] Nice ambience , but highly overrated place . [SEP](18,18) - "place"
[CLS] "NEU". For aspect terms of neutral sentiment . [SEP] [SEP] Nice ambience , but highly overrated place . [SEP]∅
Sen. Pair Classification(PAWS-X)Ori.Hypothesis: The Tabaci River is a tributary of the River Leurda in Romania.Premise: The Leurda River is a tributary of the River Tabaci in Romania.Contradiction
PMR[CLS] Contradiction . The hypothesis is a sentence with a contradic-tory meaning to the premise . [SEP] [SEP] Hypothesis : The Tabaci River is a tributary of the River Leurda in Romania . Premise : The Leurda River is a tributary of the River Tabaci in Romania . [SEP](0,0) - "[CLS]"
[CLS] Entailment . The hypothesis is a sentence with a similar meaning as the premise . [SEP] [SEP] Hypothesis : The Tabaci River is a tributary of the River Leurda in Romania . Premise : The Leurda River is a tributary of the River Tabaci in Romania . [SEP]∅
", + "bbox": [ + 121, + 87, + 877, + 765 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/5553417b6d76c3c93b8475680099ce0d91d4dbfe3af4241c388707bac41aee2e.jpg", + "table_caption": [ + "Table 6: MRC examples of XLU tasks. We use English examples here for demonstration purposes. Ori. indicates the original data format of these tasks." + ], + "table_footnote": [], + "table_body": "
ModelenardeeleshiruthtrvizhAvg.
XLM-Rbase82.2 / 72.065.5 / 49.973.9 / 59.771.2 / 56.376.3 / 59.466.4 / 52.073.7 / 58.964.7 / 54.667.0 / 52.873.3 / 54.765.0 / 55.970.8 / 56.9
mPMRbase84.4 / 73.469.6 / 53.276.4 / 61.574.9 / 58.477.4 / 60.269.2 / 54.575.2 / 58.869.2 / 57.670.4 / 55.874.8 / 55.871.8 / 65.574.0 / 59.5
XLM-R86.5 / 75.672.4 / 54.879.3 / 63.079.2 / 61.682.0 / 62.976.1 / 59.179.0 / 62.972.2 / 59.875.4 / 60.879.7 / 60.868.2 / 58.277.3 / 61.7
mPMR87.6 / 76.575.9 / 60.081.5 / 65.080.8 / 63.982.8 / 65.176.5 / 60.380.9 / 65.375.5 / 65.576.7 / 61.381.5 / 62.271.5 / 63.479.2 / 64.4
", + "bbox": [ + 119, + 825, + 882, + 883 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 7: XQuAD results (F1 / EM) for each language.", + "bbox": [ + 314, + 892, + 678, + 908 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1542", + "bbox": [ + 482, + 928, + 521, + 940 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/8e6f20f0077c1a8a39e3ffdb3cfe824932d9d769058ee6587eff3dec4c024ed8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelenardeeshivizhAvg.
XLM-Rbase79.3 / 67.255.4 / 38.162.0 / 49.166.8 / 50.259.4 / 44.866.1 / 46.761.8 / 39.564.4 / 47.9
mPMRbase81.1 / 68.958.5 / 41.063.6 / 50.568.5 / 52.160.3 / 46.468.3 / 49.256.6 / 32.965.3 / 48.7
XLM-R83.4 / 71.064.9 / 45.869.6 / 54.874.1 / 56.870.7 / 53.473.3 / 53.064.4 / 42.471.5 / 53.9
mPMR84.0 / 71.466.4 / 47.070.3 / 56.274.5 / 57.171.4 / 54.174.7 / 54.470.5 / 47.373.1 / 55.4
", + "bbox": [ + 119, + 87, + 878, + 170 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/1b5ba5fe807c4a6f675f2589e3a00658e6068ac4dc90166fbd3297fd790dac4b.jpg", + "table_caption": [ + "Table 8: MLQA results (F1 / EM) for each language." + ], + "table_footnote": [], + "table_body": "
ModelenarbnfiidkoruswteAvg.
XLM-Rbase66.8 / 57.355.7 / 42.031.5 / 20.452.6 / 40.369.1 / 55.636.3 / 27.954.8 / 36.553.0 / 34.737.4 / 28.850.8 / 38.2
mPMRbase71.1 / 61.666.3 / 52.656.5 / 41.665.5 / 53.173.9 / 63.750.4 / 38.864.4 / 37.957.4 / 41.165.3 / 50.463.4 / 49.0
XLM-R71.3 / 60.769.3 / 52.366.2 / 53.164.3 / 51.376.5 / 62.558.3 / 46.764.7 / 43.468.6 / 53.167.3 / 41.167.4 / 51.6
mPMR76.4 / 65.276.0 / 58.072.3 / 55.874.4 / 56.584.1 / 71.362.2 / 50.772.5 / 43.276.5 / 63.177.7 / 60.874.7 / 58.3
", + "bbox": [ + 119, + 215, + 875, + 282 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/2b924d373700d88f3d63ddbff2aad515ae9917aa0e0817abe6eb8f1fde306eb0.jpg", + "table_caption": [ + "Table 9: TyDiQA-GoldP results (F1 / EM) for each language." + ], + "table_footnote": [], + "table_body": "
Modelenafarbgbndeeleseteufafifrhehihuiditjajv
XLM-Rbase84.275.347.379.066.377.575.378.069.656.038.170.481.450.867.972.451.079.619.663.9
mPMRbase85.180.757.680.271.981.277.679.579.171.349.680.482.465.271.782.258.683.543.272.0
XLM-R85.481.153.984.073.882.382.880.468.854.864.275.981.459.372.976.459.384.613.271.2
mPMR86.081.756.185.979.682.382.375.582.769.675.284.182.066.575.984.059.986.149.172.4
kakkkomlmrmsmynlptruswtatethtltrurviyozh
XLM-Rbase58.740.634.350.846.063.840.681.580.065.476.143.046.44.271.968.745.770.91.523.0
mPMRbase72.245.152.962.459.468.157.483.781.571.877.350.557.43.074.280.355.775.231.649.9
XLM-R59.941.741.356.858.276.729.686.185.272.277.652.351.67.178.870.964.080.027.222.4
mPMR77.346.857.970.668.173.857.886.083.672.879.862.658.13.883.080.376.283.636.154.4
", + "bbox": [ + 119, + 329, + 878, + 456 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/e39a5e9ac46de813adf7f5f2534e44b0c25b28019994c38c02fafc74419fd2ab.jpg", + "table_caption": [ + "Table 10: WikiAnn results (F1 Score) for each language." + ], + "table_footnote": [], + "table_body": "
ModelendeesnlAvg.
XLM-Rbase91.371.078.775.779.2
mPMRbase91.974.380.879.781.7
XLM-R92.873.781.677.781.4
mPMR93.575.085.083.184.1
", + "bbox": [ + 309, + 500, + 685, + 598 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/ff06593b4aad1ba656695b4b3b8ff3d54ea268de4b5e57efb95b4afc91919e8f.jpg", + "table_caption": [ + "Table 11: CoNLL results (F1 Score) for each language." + ], + "table_footnote": [], + "table_body": "
ModelenesfrnlrutrAvg.
XLM-Rbase76.565.455.661.256.145.460.0
mPMRbase77.668.656.462.259.548.462.1
XLM-R82.471.360.367.461.249.166.1
mPMR82.871.964.767.466.955.768.2
", + "bbox": [ + 257, + 643, + 739, + 741 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/715bade8a7ae6d5e3f69176701c3b2b8e53d8ee58fca914c606c1bdf513c16f8.jpg", + "table_caption": [ + "Table 12: SemEval16 results (F1 Score) for each language." + ], + "table_footnote": [], + "table_body": "
ModelendeesfrjakozhAvg.
XLM-Rbase94.387.789.188.777.076.681.385.0
mPMRbase94.388.490.188.979.079.482.486.1
XLM-R95.289.391.090.979.679.982.586.9
mPMR95.290.690.391.381.282.984.688.0
", + "bbox": [ + 231, + 785, + 764, + 883 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 13: PAWS-X accuracy scores (Acc.) for each language.", + "bbox": [ + 287, + 892, + 705, + 908 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1543", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/d55ac649e3c8d6574665651add12dec735a7988f8e06f35f210203d46bd94124.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelenarbgdeelesfrhiruswthtrurvizhAvg.
XLM-Rbase84.671.076.875.674.977.976.968.974.164.471.172.465.273.273.073.3
mPMRbase84.271.577.275.575.578.676.969.574.762.571.471.665.574.374.073.6
XLM-R88.277.081.781.281.284.281.774.978.970.875.777.470.678.077.778.6
mPMR88.377.982.982.281.083.582.275.279.871.276.178.971.678.979.079.3
", + "bbox": [ + 119, + 445, + 880, + 525 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 14: XNLI accuracy scores (Acc.) for each language.", + "bbox": [ + 299, + 535, + 695, + 550 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "1544", + "bbox": [ + 482, + 928, + 521, + 940 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A For every submission:", + "bbox": [ + 114, + 107, + 322, + 122 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A1. Did you describe the limitations of your work?", + "bbox": [ + 129, + 126, + 532, + 143 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Section Limitations", + "bbox": [ + 149, + 145, + 297, + 156 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A2. Did you discuss any potential risks of your work?", + "bbox": [ + 127, + 170, + 552, + 186 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Not applicable. Left blank.", + "bbox": [ + 149, + 187, + 349, + 200 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A3. Do the abstract and introduction summarize the paper's main claims?", + "bbox": [ + 129, + 212, + 695, + 228 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Abstract, Section 1", + "bbox": [ + 149, + 230, + 292, + 243 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A4. Have you used AI writing assistants when working on this paper?", + "bbox": [ + 129, + 255, + 668, + 272 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 149, + 273, + 231, + 287 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "B Did you use or create scientific artifacts?", + "bbox": [ + 114, + 299, + 487, + 316 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Section 3", + "bbox": [ + 132, + 321, + 205, + 335 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "B1. Did you cite the creators of artifacts you used?", + "bbox": [ + 129, + 346, + 529, + 363 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Section 3", + "bbox": [ + 152, + 363, + 221, + 376 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?", + "bbox": [ + 127, + 390, + 778, + 406 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Not applicable. Left blank.", + "bbox": [ + 149, + 407, + 349, + 422 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?", + "bbox": [ + 129, + 432, + 880, + 495 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Section 3, Appendix A.1", + "bbox": [ + 149, + 498, + 329, + 512 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?", + "bbox": [ + 127, + 524, + 880, + 571 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Not applicable. Left blank.", + "bbox": [ + 149, + 573, + 349, + 588 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?", + "bbox": [ + 129, + 598, + 880, + 631 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Appendix A.1, Appendix A.2", + "bbox": [ + 149, + 633, + 359, + 646 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be.", + "bbox": [ + 129, + 657, + 880, + 739 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Appendix A.1", + "bbox": [ + 149, + 740, + 252, + 753 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "C Did you run computational experiments?", + "bbox": [ + 114, + 764, + 492, + 781 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Section 4", + "bbox": [ + 132, + 787, + 205, + 800 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?", + "bbox": [ + 129, + 812, + 880, + 845 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Section 3, Appendix A.1", + "bbox": [ + 149, + 846, + 329, + 860 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.", + "bbox": [ + 112, + 868, + 877, + 892 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "ACL 2023 Responsible NLP Checklist", + "bbox": [ + 132, + 84, + 433, + 99 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "1545", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Appendix A.1", + "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Section 4", + "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Section 3, Appendix A.1" + ], + "bbox": [ + 129, + 83, + 878, + 282 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank.", + "bbox": [ + 112, + 293, + 877, + 330 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response.", + "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response.", + "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response.", + "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response.", + "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response." + ], + "bbox": [ + 127, + 340, + 878, + 640 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "1546", + "bbox": [ + 482, + 928, + 521, + 940 + ], + "page_idx": 13 + } +] \ No newline at end of file diff --git a/2023/mPMR_ A Multilingual Pre-trained Machine Reader at Scale/24809581-970f-402d-a256-c7d8e2f40ff3_model.json b/2023/mPMR_ A Multilingual Pre-trained Machine Reader at Scale/24809581-970f-402d-a256-c7d8e2f40ff3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8e865626fec3a51b28a2eb117fd9be02d7c93e26 --- /dev/null +++ b/2023/mPMR_ A Multilingual Pre-trained Machine Reader at Scale/24809581-970f-402d-a256-c7d8e2f40ff3_model.json @@ -0,0 +1,2263 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.182, + 0.09, + 0.825, + 0.112 + ], + "angle": 0, + "content": "mPMR: A Multilingual Pre-trained Machine Reader at Scale*" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.135, + 0.744, + 0.153 + ], + "angle": 0, + "content": "Weiwen Xu\\(^{12,\\dagger}\\) Xin Li\\(^{2,\\ddagger}\\) Wai Lam\\(^{1}\\) Lidong Bing\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.154, + 0.662, + 0.171 + ], + "angle": 0, + "content": "1The Chinese University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.359, + 0.171, + 0.643, + 0.187 + ], + "angle": 0, + "content": "\\(^{2}\\)DAMO Academy, Alibaba Group" + }, + { + "type": "text", + "bbox": [ + 0.18, + 0.188, + 0.823, + 0.205 + ], + "angle": 0, + "content": "{wxxu,wlam}@se.cuhk.edu.hk {xinting.lx,l.bing}@alibaba-inc.com" + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.253, + 0.341, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.277, + 0.461, + 0.576 + ], + "angle": 0, + "content": "We present multilingual Pre-trained Machine Reader (mPMR), a novel method for multilingual machine reading comprehension (MRC)-style pre-training. mPMR aims to guide multilingual pre-trained language models (mPLMs) to perform natural language understanding (NLU) including both sequence classification and span extraction in multiple languages. To achieve cross-lingual generalization when only source-language fine-tuning data is available, existing mPLMs solely transfer NLU capability from a source language to target languages. In contrast, mPMR allows the direct inheritance of multilingual NLU capability from the MRC-style pre-training to downstream tasks. Therefore, mPMR acquires better NLU capability for target languages. mPMR also provides a unified solver for tackling cross-lingual span extraction and sequence classification, thereby enabling the extraction of rationales to explain the sentence-pair classification process.\\(^{1}\\)" + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.586, + 0.26, + 0.601 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.611, + 0.489, + 0.77 + ], + "angle": 0, + "content": "Multilingual pre-trained language models, acronymed as mPLMs, have demonstrated strong Natural language understanding (NLU) capability in a wide range of languages (Xue et al., 2021; Cai et al., 2021, 2022; Conneau et al., 2020a; Ding et al., 2022; Li et al., 2020a). In particular, mPLMs can maintain exceptional cross-lingual language understanding (XLU) capability on unseen target languages though mPLMs are only fine-tuned on resource-rich source languages like English." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.772, + 0.489, + 0.803 + ], + "angle": 0, + "content": "It has been proved that optimizing cross-lingual representations of mPLMs can improve XLU ca" + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.253, + 0.875, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.462, + 0.862, + 0.477 + ], + "angle": 0, + "content": "Figure 1: Pre-training and fine-tuning of mPMR." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.518, + 0.886, + 0.919 + ], + "angle": 0, + "content": "pability. For example, cross-lingual supervisions, such as parallel sentences (Conneau and Lample, 2019) or bilingual dictionaries (Conneau et al., 2020b) could enhance cross-lingual representations with better language alignment. XLM-R (Conneau et al., 2020a) and mT5 (Xue et al., 2021) showed that appropriately incorporating more languages during pre-training leads to better cross-lingual representations. A few works enriched the cross-lingual representations with factual knowledge through the utilization of multilingual mentions of entities (Calixto et al., 2021; Ri et al., 2022) and relations (Liu et al., 2022; Jiang et al., 2022) annotated in knowledge graphs. Despite their differences, the above methods essentially constructed more diverse multilingual corpora for pre-training mPLMs. These mPLMs would presumably meet their saturation points and are known to suffer from curse of multilinguality (Conneau et al., 2020a; Pfeiffer et al., 2022; Berend, 2022). Under this situation, introducing more training data from either existing (Pfeiffer et al., 2022) or unseen (Conneau et al., 2020a) languages for enhancing mPLMs may not bring further improvement or even be detrimental to their cross-lingual representations." + }, + { + "type": "page_footnote", + "bbox": [ + 0.113, + 0.809, + 0.488, + 0.894 + ], + "angle": 0, + "content": "* This work was supported by Alibaba Group through Alibaba Research Intern Program. The work described in this paper was also partially supported by a grant from the Research Grant Council of the Hong Kong Special Administrative Region, China (Project Code: 14200719).† This work was done when Weiwen Xu was an intern at Alibaba DAMO Academy.‡ Xin Li is the corresponding author." + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.894, + 0.488, + 0.918 + ], + "angle": 0, + "content": "1The code, data, and checkpoints are released at https: //github.com/DAMO-NLP-SG/PMR" + }, + { + "type": "list", + "bbox": [ + 0.113, + 0.809, + 0.488, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1533" + }, + { + "type": "footer", + "bbox": [ + 0.227, + 0.946, + 0.771, + 0.958 + ], + "angle": 0, + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + }, + { + "type": "footer", + "bbox": [ + 0.369, + 0.959, + 0.631, + 0.972 + ], + "angle": 0, + "content": "Volume 2: Short Papers, pages 1533-1546" + }, + { + "type": "footer", + "bbox": [ + 0.297, + 0.973, + 0.702, + 0.986 + ], + "angle": 0, + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.117, + 0.085, + 0.491, + 0.47 + ], + "angle": 0, + "content": "In the paper, instead of training a new mPLM with better cross-lingual representations, we propose multilingual Pre-trained Machine Reader (mPMR) to directly guide existing mPLMs to perform NLU in various languages. As shown in Figure 1, mPMR resembles PMR (Xu et al., 2022) for constructing multilingual machine reading comprehension (MRC)-style data with Wikipedia hyperlinks. These data are used to retrofit an mPLM into an mPMR through an MRC-style continual pre-training. During retrofitting process (i.e., pretraining), mPMR jointly learns the general sequence classification and span extraction capability for multiple languages. In XLU fine-tuning, mPLMs solely rely on cross-lingual representations to transfer NLU capability from a source language to target languages. By contrast, mPMR enables the direct inheritance of multilingual NLU capability from the MRC-style pre-training to downstream tasks in a unified MRC formulation, which alleviates the discrepancies between source-language fine-tuning and target-language inference (Zhou et al., 2022a,b, 2023). Therefore, mPMR shows greater potential in XLU than mPLMs." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.472, + 0.49, + 0.584 + ], + "angle": 0, + "content": "To improve the scalability of mPMR across multiple languages, we further propose Unified Q/C Construction and Stochastic answer position strategies for refining the curation of MRC data. With these two strategies, mPMR can better generalize to low-resource languages and becomes more robust to position bias (Ko et al., 2020)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.586, + 0.49, + 0.749 + ], + "angle": 0, + "content": "The experimental results show that mPMR obtains clear improvements over XLM-R (Conneau et al., 2020a) on span extraction, with an average improvement of up to 12.6 F1 on TyDiQA, and 8.7 F1 on WikiAnn respectively. The analysis reveals that mPMR benefits from more multilingual MRC data for pre-training. We also found that mPMR converges faster in downstream tasks and is capable of using its strong extraction capability for explaining the sequence classification process." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.761, + 0.215, + 0.777 + ], + "angle": 0, + "content": "2 mPMR" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.788, + 0.491, + 0.854 + ], + "angle": 0, + "content": "We present the MRC model and training data of mPMR. We closely follow PMR (Xu et al., 2022) and introduce the modifications for enabling multilingual MRC-style pre-training." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.865, + 0.315, + 0.882 + ], + "angle": 0, + "content": "2.1 Model Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.888, + 0.489, + 0.919 + ], + "angle": 0, + "content": "Our mPMR follows the same MRC architecture of Xu et al. (2022, 2023) with an encoder and an" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.214 + ], + "angle": 0, + "content": "extractor. The encoder maps input tokens \\( X \\), the concatenation of the query \\( Q \\), the context \\( C \\), and special markers (i.e., [CLS] and [SEP]), into hidden representations \\( H \\). For any two tokens \\( X_{i} \\) and \\( X_{j} \\) (\\( i < j \\)), the extractor receives their contextualized representations \\( H_{i} \\) and \\( H_{j} \\) and predicts the probability score \\( S_{i,j} \\) indicating the probability of the token span \\( X_{i:j} \\) being the answer to the query \\( Q \\)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.215, + 0.885, + 0.343 + ], + "angle": 0, + "content": "mPMR is guided with the Wiki Anchor Extraction (WAE) objective to train both the encoder and the extractor. WAE checks if the answer to the query exists in the context. If so, WAE would first regard the query and the context to be relevant and extracts the [CLS] token as a sequence-level relevance indicator. WAE would then extract all corresponding answers from the context." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.354, + 0.746, + 0.37 + ], + "angle": 0, + "content": "2.2 Multilingual MRC Data" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.375, + 0.884, + 0.472 + ], + "angle": 0, + "content": "Training mPMR requires the existence of labeled (query, context, answer) triplets. To obtain such data, we collected Wikipedia articles with anchor annotations for 24 languages, which are the most widely used and cover a reasonable number of languages used in XLU tasks (Ri et al., 2022)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.473, + 0.884, + 0.697 + ], + "angle": 0, + "content": "As shown in Figure 1, we utilized a Wikipedia anchor to obtain a pair of correlated articles. One side of the pair is the article that provides in-depth descriptions of the anchor entity, which we defined as the definition article. The other side of the pair is named as the mention article, which mentions the specific anchor text2. We composed an answerable MRC example in which the anchor is the answer, the surrounding text of the anchor in the mention article is the context, and the definition of the anchor entity in the definition article is the query. Additionally, we can generate an unanswerable MRC example by pairing a query with an irrelevant context without anchor association." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.707, + 0.885, + 0.884 + ], + "angle": 0, + "content": "Unified Q/C Construction. PMR constructed the MRC query and context as valid sentences so as to keep the text coherent. However, sentence segmentation tools are usually not available for low-resource languages. To remedy this, we did not apply sentence segmentation but only preprocess Wikipedia articles with word tokenization in mPMR. For each anchor, the MRC query comprises the first \\(Q\\) words in the definition article. To prevent information leakage during pre-training, similar to PMR, we anonymized the anchor entity" + }, + { + "type": "page_footnote", + "bbox": [ + 0.509, + 0.892, + 0.883, + 0.919 + ], + "angle": 0, + "content": "2definition/mention article refers to home/reference article of Xu et al. (2022)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1534" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.127, + 0.082, + 0.871, + 0.297 + ], + "angle": 0, + "content": "
Model#ParamsEQANERABSASentence PairAvg.
XQuADMLQATyDiQAWikiAnnCoNLLSemEval16PAWS-XXNLI
MetricsF1 / EMF1 / EMF1 / EMF1F1F1Acc.Acc.
XLM-R550M76.6 / 60.871.6 / 53.265.1 / 45.065.482.066.9‡86.479.274.2
mT5580M67.0 / 49.064.6 / 45.057.2 / 41.255.771.0‡62.5‡86.475.467.5
VECO550M77.3 / 61.871.7 / 53.267.6 / 49.165.781.3‡63.0‡88.779.974.4
mLUKE-W561M79.6 / -72.7 / -65.2 / 48.5‡67.7‡83.061.2‡88.2‡79.4‡74.6
Wiki-CL550M72.1 / 56.970.8 / 50.573.2 / 57.364.7--88.479.2-
KMLM550M77.3 / 61.772.1 / 53.767.9 / 50.466.7‡83.266.1‡88.079.275.1
Our MRC Formulation
XLM-Rbase270M70.8 / 56.964.4 / 47.950.8 / 38.257.979.260.085.073.367.7
mPMRbase270M74.0 / 59.565.3 / 48.763.4 / 49.066.681.762.186.173.671.6
XLM-R550M77.1 / 61.371.5 / 53.967.4 / 51.663.681.466.186.978.674.1
mPMR550M79.2 / 64.473.1 / 55.474.7 / 58.370.784.168.288.079.377.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.308, + 0.883, + 0.351 + ], + "angle": 0, + "content": "Table 1: The results of all XLU tasks. We report the average results of all languages for each dataset. We also compute the overall average score among all datasets in the Avg. column. We reproduce the missing results with the \\(\\ddagger\\) label. Some results of Wiki-CL are left blank because they do not release their model checkpoint." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.377, + 0.486, + 0.408 + ], + "angle": 0, + "content": "in the query to the [MASK] token. The MRC context consists of \\( C \\) words surrounding the anchor." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.419, + 0.49, + 0.707 + ], + "angle": 0, + "content": "Stochastic Answer Position. As mentioned by Ko et al. (2020), the model is prone to overfitting to the position shortcut if the answer in the context exhibits a fixed position pattern. In our case, suppose that the MRC context consists of \\( C / 2 \\) words on both the left and right sides of the anchor, the model may learn the shortcut that the middle part of the context is likely to be the answer. To prevent such position bias, we propose a stochastic answer position method, which allows the answer to be presented in any position within the context. Specifically, given an anchor in a Wikipedia article, the context comprises \\( \\xi \\) words preceding the anchor and the \\( C - \\xi \\) words following the anchor, where \\( \\xi \\) is a random integer ranging from 0 to \\( C \\) and varies across different contexts. In accordance with PMR, we treated all text spans identical to the anchor in the current context as valid answers." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.722, + 0.322, + 0.738 + ], + "angle": 0, + "content": "3 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.749, + 0.488, + 0.844 + ], + "angle": 0, + "content": "Implementation Details. In mPMR, the encoder is loaded from XLM-R (Conneau et al., 2020a) and the extractor is randomly initialized. Both components are then continually pre-trained using the multilingual MRC data that we constructed. More hyper-parameters can be found in Appendix A.1." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.855, + 0.489, + 0.919 + ], + "angle": 0, + "content": "Downstream XLU Tasks. We evaluated mPMR on a series of span extraction tasks, including Extractive Question Answering (EQA), Named Entity Recognition (NER), and Aspect-Based Sentiment" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.377, + 0.885, + 0.618 + ], + "angle": 0, + "content": "Analysis (ABSA). We also evaluated our mPMR on two sequence classification tasks. We followed Xu et al. (2022) to convert all tasks into MRC formulation to effectively leverage the knowledge that is acquired during MRC-style pre-training. For EQA, we used XQuAD (Artetxe et al., 2020), MLQA (Lewis et al., 2020), and TyDiQA (Clark et al., 2020). For NER, we used WikiAnn (Pan et al., 2017) and CoNLL (Tjong Kim Sang, 2002; Tjong Kim Sang and De Meulder, 2003). SemEval16 (Pontiki et al., 2016) was used for ABSA task. Regarding the sequence classification, we used XNLI (Conneau et al., 2018) and PAWS-X (Yang et al., 2019). Additional dataset information and concrete examples are provided in Appendix A.2" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.626, + 0.884, + 0.803 + ], + "angle": 0, + "content": "Baselines. We compared mPMR with recent methods on improving cross-lingual representations, including 1) models pre-trained on a large number of languages: XLM-R (Conneau et al., 2020a), mT5 (Xue et al., 2021), and VECO (Luo et al., 2021); 2) models that exploited multilingual entity information: Wiki-CL (Calixto et al., 2021), and mLUKE-W (Ri et al., 2022); and 3) Model that utilized multilingual relation information: KMLM (Liu et al., 2022). For a fair comparison, all models have approximately the same parameter size." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.815, + 0.727, + 0.831 + ], + "angle": 0, + "content": "4 Results and Analyses" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.84, + 0.884, + 0.919 + ], + "angle": 0, + "content": "XLU Performance. Table 1 shows the results on a variety of XLU tasks. mPMR outperforms all previous methods with an absolute improvement of 2.1 F1 over the best baseline (i.e. KMLM). mPMR shows greater improvements over previ" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1535" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.116, + 0.083, + 0.887, + 0.171 + ], + "angle": 0, + "content": "
IndexModel#LangPAWS-XXQuADWikiAnnAvg.
#1XLM-Rbase085.070.857.971.2
#2#1 + MRC data in English185.2 (0.2↑)71.0 (0.2↑)59.5 (1.6↑)71.9 (0.7↑)
#3#2 + Stochastic Answer Position185.5 (0.3↑)73.0 (2.0↑)60.0 (0.5↑)72.8 (0.9↑)
#4#3 + MRC data in more languages1085.9 (0.4↑)73.5 (0.5↑)64.7 (4.7↑)74.7 (1.9↑)
#5#4 + MRC data in even more languages (mPMRbase)2486.1 (0.2↑)74.0 (0.5↑)66.6 (1.9↑)75.6 (0.9↑)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.18, + 0.884, + 0.21 + ], + "angle": 0, + "content": "Table 2: The process of retrofitting XLM-R into mPMR using multilingual MRC data (English→10 languages→24 languages) and our Stochastic Answer Position method. Each row accumulates modifications from all rows above." + }, + { + "type": "table", + "bbox": [ + 0.116, + 0.222, + 0.884, + 0.435 + ], + "angle": 0, + "content": "
LabelSentence 1Sentence 2
EntailmentRami Nieminen ( born February 25 , 1966 ) is a Finnish footballer.Rami Nieminen ( born 25 February 1966 ) is a Finnish former footballer.
ContradictionIn 1938 he became the Government Anthropologist of the Egyptian-Anglo Sudan and conducted fieldwork with the Nuba.In 1938 he became the government anthropologist of the anglo-Egyptian Sudan and led fieldwork with the Nuba.
EntailmentStipsits 出生于科尔新堡，并在维也纳施塔莫斯多夫度过了他的童年。什蒂普西奇出生于德国科恩堡，在维也纳斯塔莫斯多夫度过了他的童年。
Contradiction纳舒厄白银骑士团队加入了夏季大学联盟，是本市的现役球队。Nashua Silver Knights 队是当前夏季联赛的一部分，也是该市的大学体育队。
Entailmentごれらの見方は、福音主義的、清教徒的、プロデ斯特兰トの動態が出現すると必に、しはしだは表明くださいます。ごれらの見解は多くの场合、新教徒、清教徒、福音主義者が出現する:NOか表示お願いいたします。
Contradiction1954年にスリーナムに戸った後、弁護士とでラマリポに定住したこと。1954年、バラマリポに戸ると、彼はスリーナムで弁護士とで定住,No理由。
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.443, + 0.884, + 0.473 + ], + "angle": 0, + "content": "Table 3: Case study on PAWS-X. mPMR can extract rationales to explain the sequence-pair classification in multiple languages." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.497, + 0.49, + 0.756 + ], + "angle": 0, + "content": "ous methods on span extraction tasks. In particular, mPMR achieves up to 7.3 and 7.1 F1 improvements over XLM-R on TyDiQA and WikiAnn respectively. Such significant improvements probably come from the following two facts: (1) WikiAnn comprises a larger number of target languages (i.e. 40). Therefore, existing methods may struggle to align these low-resource languages with English due to a lack of language-specific data. (2) TyDiQA is a more challenging cross-lingual EQA task with \\(2\\mathrm{x}\\) less lexical overlap between the query and the answer than MLQA and XQuAD (Hu et al., 2020). Our mPMR, which acquires target-language span extraction capability from both MRC-style pretraining and English-only QA fine-tuning, achieves larger performance gains on more challenging task." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.775, + 0.49, + 0.92 + ], + "angle": 0, + "content": "mPMR Pre-training. To reflect the impact of our MRC-style data and Stochastic Answer Position method on pre-training, we present a step-by-step analysis of the retrofitting process starting from XLM-R in Table 2. Our findings suggest that the significant improvements observed are largely due to the inclusion of multilingual MRC data. Introducing English MRC data (model #2) gives marginal improvements because model #2" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.495, + 0.871, + 0.619 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.627, + 0.885, + 0.657 + ], + "angle": 0, + "content": "Figure 2: Convergence speed (Test set F1 and the training loss) of \\(\\mathrm{mPMR_{base}}\\) and XLM-\\(\\mathbf{R}_{\\mathrm{base}}\\) on WikiAnn." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.683, + 0.885, + 0.844 + ], + "angle": 0, + "content": "can only rely on cross-lingual representations to transfer the knowledge acquired during MRC-style pre-training. When using MRC data on more languages (model #4 and #5), we can observe significant improvements on XLU tasks. This can be attributed to the NLU capability directly inherited from MRC-style pre-training in target languages. Additionally, with our Stochastic Answer Position method (model #3), mPMR becomes more robust to position bias and thus improves XLU tasks." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.855, + 0.885, + 0.92 + ], + "angle": 0, + "content": "Explainable Sentence-pair Classification. Inspired by PMR (Xu et al., 2022), we investigated if the extraction capability of mPMR can be leveraged to explain sentence-pair classification. Note" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1536" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.081, + 0.484, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.22, + 0.49, + 0.25 + ], + "angle": 0, + "content": "Figure 3: Convergence speed (Test set F1 and the training loss) of \\(\\mathrm{mPMR_{base}}\\) and XLM-\\(\\mathbf{R}_{\\mathrm{base}}\\) on XQuAD." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.292, + 0.49, + 0.55 + ], + "angle": 0, + "content": "that sentence-pair classification focuses on the inference between the two sentences. If we construct the query with only the task label as PMR does, such query does not solely correspond to any meaningful span in the context, and thus is hard to guide the span extraction. Therefore, we leveraged another template \"[CLS] label Sen-1 [SEP] Sen-2 [SEP]\", where the two sentences are represented separately in the query and the context. In this template, we can extract the exact span from Sen-2 that leads to a contraction or entailment relation (i.e., the task label) with Sen-1. Specifically, we passed the sentence pair to the model twice, with each sentence of the pair being designated as the Sen-2 respectively, and extract the context span with the highest probability score from both sentences." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.559, + 0.49, + 0.704 + ], + "angle": 0, + "content": "As shown in Table 3, the extracted spans are indeed important rationales that determine the relationship between two sentences. Such a finding confirms that the extraction capability of mPMR can be appropriately used for explaining the sentence-pair classification process. While the extraction capability may affect the learning of sequence classification during fine-tuning, resulting in a 0.4 Acc. decrease on XNLI." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.743, + 0.49, + 0.918 + ], + "angle": 0, + "content": "mPMR Fine-tuning. We investigated the effects of mPMR on XLU fine-tuning. Figure 2 shows that mPMR converges faster than XLM-R on WikiAnn with an extremely low loss value even fine-tuned for 500 steps. In terms of test set performance, mPMR outperforms XLM-R comprehensively and exhibits greater stability. As a result, mPMR provides a better starting point for addressing XLU tasks compared to XLM-R. More examples from XQuAD and PAWS-X are provided in Figure 3 and 4." + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.082, + 0.867, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.22, + 0.885, + 0.25 + ], + "angle": 0, + "content": "Figure 4: Convergence speed (Test set F1 and the training loss) of \\(\\mathrm{mPMR_{base}}\\) and XLM-\\(\\mathbf{R}_{\\mathrm{base}}\\) on PAWS-X." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.276, + 0.651, + 0.291 + ], + "angle": 0, + "content": "5 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.303, + 0.885, + 0.433 + ], + "angle": 0, + "content": "This paper presents a novel multilingual MRC-style pre-training method, namely mPMR. mPMR provides a unified solver for cross-lingual span extraction and sequence classification and enables direct transfer of NLU capability from pre-training to downstream tasks. mPMR clearly improves the previous baselines and provides a possible solution to explain the sentence-pair classification process." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.447, + 0.617, + 0.463 + ], + "angle": 0, + "content": "Limitations" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.474, + 0.884, + 0.506 + ], + "angle": 0, + "content": "We identify the following two limitations of our work:" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.522, + 0.885, + 0.716 + ], + "angle": 0, + "content": "- Different from raw text, constructing MRC-style data from Wikipedia requires the existence of hyperlinks. This idea works well for resource-rich languages, such as English and Chinese. While such an idea is less effective for languages with few hyperlink annotations in Wikipedia because a small amount of MRC-style training data is difficult to guide the learning of NLU capability in those languages. A possible solution is to explore other data resources to automatically construct large-scale MRC data for pre-training." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.729, + 0.887, + 0.891 + ], + "angle": 0, + "content": "- As observed in Table 1, the improvements of sequence classification tasks are less significant than those of span extraction tasks. We suggest that the existence of anchors is not a strong relevance indicator between our constructed query and context. Such a finding is also observed in Chang et al. (2020). Therefore, constructing more relevant query-context pairs for sequence classification pre-training can possibly remedy this issue." + }, + { + "type": "list", + "bbox": [ + 0.532, + 0.522, + 0.887, + 0.891 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.928, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1537" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.116, + 0.085, + 0.214, + 0.099 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.108, + 0.489, + 0.173 + ], + "angle": 0, + "content": "Mikel Artetxe, Sebastian Ruder, and Dani Yogatama. 2020. On the cross-lingual transferability of monolingual representations. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.183, + 0.487, + 0.21 + ], + "angle": 0, + "content": "Giuseppe Attardi. 2015. Wikiextractor. https://github.com/attardi/wikiextractor." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.22, + 0.487, + 0.299 + ], + "angle": 0, + "content": "Gábor Berend. 2022. Combating the curse of multilinguality in cross-lingual WSD by aligning sparse contextualized word representations. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.309, + 0.487, + 0.374 + ], + "angle": 0, + "content": "Deng Cai, Xin Li, Jackie Chun-Sing Ho, Lidong Bing, and Wai Lam. 2021. Multilingual AMR parsing with noisy knowledge distillation. In Findings of the Association for Computational Linguistics: EMNLP 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.385, + 0.487, + 0.451 + ], + "angle": 0, + "content": "Deng Cai, Xin Li, Jackie Chun-Sing Ho, Lidong Bing, and Wai Lam. 2022. Retrofitting multilingual sentence embeddings with Abstract Meaning Representation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.461, + 0.487, + 0.552 + ], + "angle": 0, + "content": "Iacer Calixto, Alessandro Raganato, and Tommaso Pasini. 2021. Wikipedia entities as rendezvous across languages: Grounding multilingual language models by predicting Wikipedia hyperlinks. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.563, + 0.487, + 0.615 + ], + "angle": 0, + "content": "Wei-Cheng Chang, Felix X. Yu, Yin-Wen Chang, Yiming Yang, and Sanjiv Kumar. 2020. Pre-training tasks for embedding-based large-scale retrieval. In International Conference on Learning Representations." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.625, + 0.487, + 0.704 + ], + "angle": 0, + "content": "Jonathan H. Clark, Eunsol Choi, Michael Collins, Dan Garrette, Tom Kwiatkowski, Vitaly Nikolaev, and Jennimaria Palomaki. 2020. TyDi QA: A benchmark for information-seeking question answering in typologically diverse languages. Transactions of the Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.714, + 0.487, + 0.806 + ], + "angle": 0, + "content": "Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2020a. Unsupervised cross-lingual representation learning at scale. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.816, + 0.487, + 0.856 + ], + "angle": 0, + "content": "Alexis Conneau and Guillaume Lample. 2019. Crosslingual language model pretraining. In Advances in Neural Information Processing Systems." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.866, + 0.487, + 0.919 + ], + "angle": 0, + "content": "Alexis Conneau, Rudy Rinott, Guillaume Lample, Adina Williams, Samuel Bowman, Holger Schwenk, and Veselin Stoyanov. 2018. XNLI: Evaluating crosslingual sentence representations. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.108, + 0.489, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.53, + 0.086, + 0.883, + 0.113 + ], + "angle": 0, + "content": "the 2018 Conference on Empirical Methods in Natural Language Processing." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.124, + 0.883, + 0.19 + ], + "angle": 0, + "content": "Alexis Conneau, Shijie Wu, Haoran Li, Luke Zettlemoyer, and Veselin Stoyanov. 2020b. Emerging cross-lingual structure in pretrained language models. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.201, + 0.883, + 0.28 + ], + "angle": 0, + "content": "Bosheng Ding, Junjie Hu, Lidong Bing, Mahani Aljunied, Shafiq Joty, Luo Si, and Chunyan Miao. 2022. GlobalWoZ: Globalizing MultiWoZ to develop multilingual task-oriented dialogue systems. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.291, + 0.883, + 0.357 + ], + "angle": 0, + "content": "Junjie Hu, Sebastian Ruder, Aditya Siddhant, Graham Neubig, Orhan First, and Melvin Johnson. 2020. Xtreme: A massively multilingual multi-task benchmark for evaluating cross-lingual generalisation. In International Conference on Machine Learning." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.367, + 0.883, + 0.433 + ], + "angle": 0, + "content": "Xiaoze Jiang, Yaobo Liang, Weizhu Chen, and Nan Duan. 2022. Xlm-k: Improving cross-lingual language model pre-training with multilingual knowledge. In Proceedings of the AAAI Conference on Artificial Intelligence." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.444, + 0.883, + 0.51 + ], + "angle": 0, + "content": "Miyoung Ko, Jinhyuk Lee, Hyunjae Kim, Gangwoo Kim, and Jaewoo Kang. 2020. Look at the first sentence: Position bias in question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.521, + 0.883, + 0.587 + ], + "angle": 0, + "content": "Patrick Lewis, Barlas Oguz, Rudy Rinott, Sebastian Riedel, and Holger Schwenk. 2020. MLQA: Evaluating cross-lingual extractive question answering. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.598, + 0.883, + 0.675 + ], + "angle": 0, + "content": "Juntao Li, Ruidan He, Hai Ye, Hwee Tou Ng, Lidong Bing, and Rui Yan. 2020a. Unsupervised domain adaptation of a pretrained cross-lingual language model. In Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, IJCAI 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.687, + 0.883, + 0.74 + ], + "angle": 0, + "content": "Xin Li, Lidong Bing, Wenxuan Zhang, Zheng Li, and Wai Lam. 2020b. Unsupervised cross-lingual adaptation for sequence tagging and beyond. arXiv preprint arXiv:2010.12405." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.751, + 0.883, + 0.817 + ], + "angle": 0, + "content": "Linlin Liu, Xin Li, Ruidan He, Lidong Bing, Shafiq Joty, and Luo Si. 2022. Enhancing multilingual language model with massive multilingual knowledge triples. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.828, + 0.883, + 0.868 + ], + "angle": 0, + "content": "Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In International Conference on Learning Representations." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.878, + 0.883, + 0.919 + ], + "angle": 0, + "content": "Fuli Luo, Wei Wang, Jiahao Liu, Yijia Liu, Bin Bi, Songfang Huang, Fei Huang, and Luo Si. 2021. VECO: Variable and flexible cross-lingual pre-training for" + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.883, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1538" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.135, + 0.086, + 0.489, + 0.153 + ], + "angle": 0, + "content": "language understanding and generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers)." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.163, + 0.489, + 0.241 + ], + "angle": 0, + "content": "Xiaoman Pan, Boliang Zhang, Jonathan May, Joel Nothman, Kevin Knight, and Heng Ji. 2017. Cross-lingual name tagging and linking for 282 languages. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.252, + 0.488, + 0.343 + ], + "angle": 0, + "content": "Jonas Pfeiffer, Naman Goyal, Xi Lin, Xian Li, James Cross, Sebastian Riedel, and Mikel Artetxe. 2022. Lifting the curse of multilinguality by pre-training modular transformers. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.354, + 0.488, + 0.485 + ], + "angle": 0, + "content": "Maria Pontiki, Dimitris Galanis, Haris Papageorgiou, Ion Androutsopoulos, Suresh Manandhar, Mohammad AL-Smadi, Mahmoud Al-Ayyoub, Yanyan Zhao, Bing Qin, Orphée De Clercq, Véronique Hoste, Marianna Apidianaki, Xavier Tannier, Natalia Loukachevitch, Evgeniy Kotelnikov, Nuria Bel, Salud María Jiménez-Zafra, and Gülşen Eryigit. 2016. SemEval-2016 task 5: Aspect based sentiment analysis. In Proceedings of the 10th International Workshop on Semantic Evaluation (SemEval-2016)." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.496, + 0.488, + 0.574 + ], + "angle": 0, + "content": "Ryokan Ri, Ikuya Yamada, and Yoshimasa Tsuruoka. 2022. mLUKE: The power of entity representations in multilingual pretrained language models. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.585, + 0.488, + 0.65 + ], + "angle": 0, + "content": "Erik F. Tjong Kim Sang. 2002. Introduction to the CoNLL-2002 shared task: Language-independent named entity recognition. In COLING-02: The 6th Conference on Natural Language Learning 2002 (CoNLL-2002)." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.661, + 0.488, + 0.727 + ], + "angle": 0, + "content": "Erik F. Tjong Kim Sang and Fien De Meulder. 2003. Introduction to the CoNLL-2003 shared task: Language-independent named entity recognition. In Proceedings of the Seventh Conference on Natural Language Learning at HLT-NAACL 2003." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.737, + 0.488, + 0.88 + ], + "angle": 0, + "content": "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.892, + 0.488, + 0.919 + ], + "angle": 0, + "content": "Weiwen Xu, Xin Li, Yang Deng, Wai Lam, and Lidong Bing. 2023. Peerda: Data augmentation via modeling" + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.489, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.529, + 0.086, + 0.882, + 0.126 + ], + "angle": 0, + "content": "peer relation for span identification tasks. In The 61th Annual Meeting of the Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.136, + 0.883, + 0.202 + ], + "angle": 0, + "content": "Weiwen Xu, Xin Li, Wenxuan Zhang, Meng Zhou, Lidong Bing, Wai Lam, and Luo Si. 2022. From clozing to comprehending: Retrofitting pre-trained language model to pre-trained machine reader. arXiv preprint arXiv:2212.04755." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.212, + 0.883, + 0.304 + ], + "angle": 0, + "content": "Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. 2021. mT5: A massively multilingual pre-trained text-to-text transformer. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.313, + 0.883, + 0.405 + ], + "angle": 0, + "content": "Yinfei Yang, Yuan Zhang, Chris Tar, and Jason Baldridge. 2019. PAWS-X: A cross-lingual adversarial dataset for paraphrase identification. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.415, + 0.883, + 0.482 + ], + "angle": 0, + "content": "Wenxuan Zhang, Ruidan He, Haiyun Peng, Lidong Bing, and Wai Lam. 2021. Cross-lingual aspect-based sentiment analysis with aspect term code-switching. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.491, + 0.883, + 0.531 + ], + "angle": 0, + "content": "Meng Zhou, Xin Li, Yue Jiang, and Lidong Bing. 2022a. Enhancing cross-lingual prompting with mask token augmentation. arXiv preprint arXiv:2202.07255." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.54, + 0.883, + 0.619 + ], + "angle": 0, + "content": "Ran Zhou, Xin Li, Lidong Bing, Erik Cambria, and Chunyan Miao. 2023. Improving self-training for cross-lingual named entity recognition with contrastive and prototype learning. In *The 61th Annual Meeting of the Association for Computational Linguistics*." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.629, + 0.883, + 0.696 + ], + "angle": 0, + "content": "Ran Zhou, Xin Li, Lidong Bing, Erik Cambria, Luo Si, and Chunyan Miao. 2022b. ConNER: Consistency training for cross-lingual named entity recognition. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.883, + 0.696 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1539" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.115, + 0.085, + 0.239, + 0.1 + ], + "angle": 0, + "content": "A Appendix" + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.111, + 0.396, + 0.126 + ], + "angle": 0, + "content": "A.1 More Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.131, + 0.49, + 0.356 + ], + "angle": 0, + "content": "We collect the 2022-08-01 dump3 of Wikipedia articles for the 24 languages in consideration. The statistics of each language can be found in Table 4. Then for each article, we extract the plain text with anchors via WikiExtractor (Attardi, 2015). Word tokenization is performed using spaCy4 if the language is supported, otherwise, we utilize PyThaiNLP5 for Thai and Sacremoses6 for remaining languages. For each anchor entity, we construct 10 answerable MRC examples and 10 unanswerable MRC examples as described in Sec. 2.2. Anchor entities with low frequency (below 10 occurrences for English entities and 5 occurrences for entities in other languages) were excluded." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.358, + 0.49, + 0.678 + ], + "angle": 0, + "content": "In mPMR, we use Huggingface's implementations of XLM-R (Wolf et al., 2020). During the pre-training stage, the query length \\( Q \\) is set to 50 words, and the context length \\( C \\) is set to 200 words. Both are computed before the subword segmentation. We follow the default learning rate schedule and dropout settings used in XLM-R. We use AdamW (Loshchilov and Hutter, 2019) as our optimizer. We train both \\( \\mathrm{mPMR_{base}} \\) and mPMR on 4 A100 GPU. The learning rate is set to 1e-5, and the effective batch size for each step is set to 256 and 80 for \\( \\mathrm{mPMR_{base}} \\) and mPMR respectively in order to maximize the usage of the GPU memory. We use the average scores of XQuAD, CoNLL, and PAWS-X to select the best mPMR checkpoint. In fact, we continually pre-train \\( \\mathrm{mPMR_{base}} \\) and mPMR for 250,000 and 100,000 steps. The training speed is around 6250 steps per hour. The hyper-parameters of \\( \\mathrm{mPMR_{large}} \\) on downstream XLU tasks can be found in Table 5." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.691, + 0.357, + 0.705 + ], + "angle": 0, + "content": "A.2 Downstream XLU Tasks" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.712, + 0.49, + 0.857 + ], + "angle": 0, + "content": "We evaluate mPMR on XLU tasks including both span extraction (EQA, NER, and ABSA) and sequence classification (sentence pair classification). We follow (Xu et al., 2022) to convert all tasks into MRC formulation and tackle them accordingly. We show concrete examples for each task in Table 6. Specifically, we evaluate the performance of EQA on three benchmarks: XQuAD (Artetxe et al., 2020), MLQA (Lewis et al., 2020), and Ty" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.295 + ], + "angle": 0, + "content": "DiQA (Clark et al., 2020) covering 11, 7, and 9 languages respectively. For NER evaluation, we use the WikiAnn dataset (Pan et al., 2017) restricted to the 40 languages from XTREME (Hu et al., 2020), as well as the CoNLL dataset with 4 languages (Tjong Kim Sang, 2002; Tjong Kim Sang and De Meulder, 2003); We also evaluate the XLU performance of SemEval16 ABSA on 6 languages (Pontiki et al., 2016), where we collect the data from Li et al. (2020b); Zhang et al. (2021). Regarding the sequence classification task, we evaluate XNLI (Conneau et al., 2018) and PAWS-X (Yang et al., 2019) with 15 and 7 languages respectively." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.305, + 0.836, + 0.321 + ], + "angle": 0, + "content": "A.3 mPMR Performance per Language" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.326, + 0.885, + 0.405 + ], + "angle": 0, + "content": "We show the detailed results for each language in each task in Table 7 (XQuAD), Table 8 (MLQA), Table 9 (TyDiQA), Table 10 (WikiAnn), Table 11 (CoNLL), Table 12 (SemEval16), Table 13 (PAWS-X), and Table 14 (XNLI)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.136, + 0.866, + 0.401, + 0.88 + ], + "angle": 0, + "content": "3https://dumps.wikimedia.org/enwiki/latest" + }, + { + "type": "page_footnote", + "bbox": [ + 0.137, + 0.879, + 0.361, + 0.893 + ], + "angle": 0, + "content": "4https://github.com/explosion/spaCy" + }, + { + "type": "page_footnote", + "bbox": [ + 0.137, + 0.892, + 0.393, + 0.906 + ], + "angle": 0, + "content": "5https://github.com/PyThaiNLP/pythainlp" + }, + { + "type": "page_footnote", + "bbox": [ + 0.137, + 0.905, + 0.387, + 0.918 + ], + "angle": 0, + "content": "\\(^{6}\\)https://github.com/ Alvations/sacremoses" + }, + { + "type": "list", + "bbox": [ + 0.136, + 0.866, + 0.401, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1540" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.164, + 0.187, + 0.836, + 0.438 + ], + "angle": 0, + "content": "
Language# Entities# MRC examplesLanguage# Entities# MRC examples
ar118,2922,020,502ko94,6161,597,076
bn25,081410,634nl251,3234,185,913
de864,74614,795,826pl283,9254,765,015
el56,383946,114pt216,6953,648,603
en966,19719,303,940ru432,4377,342,472
es412,4767,044,972sv169,0302,808,214
fi113,1181,960,636sw4,85765,724
fr595,87910,164,216te11,005170,664
hi15,350242,078th31,676522,434
id70,9601,164,662tr71,2941,175,276
it376,4176,421,850vi68,6651,147,772
ja423,8847,338,308zh259,7854,438,004
Total5,934,091103,680,905
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.447, + 0.884, + 0.476 + ], + "angle": 0, + "content": "Table 4: Data statistics of mPMR pre-training data. The statistics is computed after removing the low-frequency entities. The number of MRC examples includes both answerable and unanswerable examples." + }, + { + "type": "table", + "bbox": [ + 0.179, + 0.696, + 0.82, + 0.786 + ], + "angle": 0, + "content": "
DatasetXQuADMLQATyDiQAWikiAnnCoNLLSemEval16PAWS-XXNLI
Query Length6464643232326464
Input Length384384384192192192192192
Batch Size8881616321632
Learning Rate3e-53e-52e-51e-51e-52e-55e-53e-5
Epoch3310101020103
" + }, + { + "type": "table_caption", + "bbox": [ + 0.29, + 0.794, + 0.707, + 0.81 + ], + "angle": 0, + "content": "Table 5: Hyper-parameters settings in fine-tuning XLU tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.52, + 0.941 + ], + "angle": 0, + "content": "1541" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.122, + 0.088, + 0.878, + 0.766 + ], + "angle": 0, + "content": "
TaskExample InputExample Output
EQA(XSQuAD)Ori.Question: Who lost to the Broncos in the divisional round?Context: The Broncos defeated the Pittsburgh Steelers in the divi-sional round, 23–16, by scoring 11 points in the final three minutes of the game.Answer: "Pittsburgh Steelers"
PMR[CLS] Who lost to the Broncos in the divisional round ? [SEP] [SEP]The Broncos defeated the Pittsburgh Steelers in the divisional round, 23–16 , by scoring 11 points in the final three minutes of the game .[SEP](17,18) - "Pittsburgh Steelers"
NER(CoNLL)Ori.Two goals in the last six minutes gave holders Japan an uninspiring 2-1 Asian Cup victory over Syria on Friday.("Japan", LOC);("Syria", LOC);("Asian Cup", MISC)
PMR[CLS] "ORG". Organization entities are limited to named corporate,governmental, or other organizational entities. [SEP] [SEP] Two goals in the last six minutes gave holders Japan an uninspiring 2-1 Asian Cup victory over Syria on Friday . [SEP]∅
[CLS] "PER". Person entities are named persons or family . [SEP] [SEP] Two goals in the last six minutes gave holders Japan an uninspiring 2-1 Asian Cup victory over Syria on Friday . [SEP]∅
[CLS] "LOC". Location entities are the name of politically or geo-graphically defined locations such as cities , countries . [SEP] [SEP] Two goals in the last six minutes gave holders Japan an uninspiring 2-1 Asian Cup victory over Syria on Friday . [SEP](32,32) - "Japan";(40,40) - "Syria"
[CLS] "MISC". Examples of miscellaneous entities include events ,nationalities , products and works of art . [SEP] [SEP] Two goals in the last six minutes gave holders Japan an uninspiring 2-1 Asian Cup victory over Syria on Friday . [SEP](34,35) - "Asian Cup"
ABSA(SemEval16)Ori.Nice ambience, but highly overrated place.("ambience", POS);("place", NEG)
PMR[CLS] "POS". For aspect terms of positive sentiment . [SEP] [SEP] Nice ambience , but highly overrated place . [SEP](13,13) - "ambience"
[CLS] "NEG". For aspect terms of negative sentiment . [SEP] [SEP] Nice ambience , but highly overrated place . [SEP](18,18) - "place"
[CLS] "NEU". For aspect terms of neutral sentiment . [SEP] [SEP] Nice ambience , but highly overrated place . [SEP]∅
Sen. Pair Classification(PAWS-X)Ori.Hypothesis: The Tabaci River is a tributary of the River Leurda in Romania.Premise: The Leurda River is a tributary of the River Tabaci in Romania.Contradiction
PMR[CLS] Contradiction . The hypothesis is a sentence with a contradic-tory meaning to the premise . [SEP] [SEP] Hypothesis : The Tabaci River is a tributary of the River Leurda in Romania . Premise : The Leurda River is a tributary of the River Tabaci in Romania . [SEP](0,0) - "[CLS]"
[CLS] Entailment . The hypothesis is a sentence with a similar meaning as the premise . [SEP] [SEP] Hypothesis : The Tabaci River is a tributary of the River Leurda in Romania . Premise : The Leurda River is a tributary of the River Tabaci in Romania . [SEP]∅
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.774, + 0.883, + 0.804 + ], + "angle": 0, + "content": "Table 6: MRC examples of XLU tasks. We use English examples here for demonstration purposes. Ori. indicates the original data format of these tasks." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.826, + 0.883, + 0.884 + ], + "angle": 0, + "content": "
ModelenardeeleshiruthtrvizhAvg.
XLM-Rbase82.2 / 72.065.5 / 49.973.9 / 59.771.2 / 56.376.3 / 59.466.4 / 52.073.7 / 58.964.7 / 54.667.0 / 52.873.3 / 54.765.0 / 55.970.8 / 56.9
mPMRbase84.4 / 73.469.6 / 53.276.4 / 61.574.9 / 58.477.4 / 60.269.2 / 54.575.2 / 58.869.2 / 57.670.4 / 55.874.8 / 55.871.8 / 65.574.0 / 59.5
XLM-R86.5 / 75.672.4 / 54.879.3 / 63.079.2 / 61.682.0 / 62.976.1 / 59.179.0 / 62.972.2 / 59.875.4 / 60.879.7 / 60.868.2 / 58.277.3 / 61.7
mPMR87.6 / 76.575.9 / 60.081.5 / 65.080.8 / 63.982.8 / 65.176.5 / 60.380.9 / 65.375.5 / 65.576.7 / 61.381.5 / 62.271.5 / 63.479.2 / 64.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.315, + 0.894, + 0.68, + 0.909 + ], + "angle": 0, + "content": "Table 7: XQuAD results (F1 / EM) for each language." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1542" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.12, + 0.089, + 0.88, + 0.171 + ], + "angle": 0, + "content": "
ModelenardeeshivizhAvg.
XLM-Rbase79.3 / 67.255.4 / 38.162.0 / 49.166.8 / 50.259.4 / 44.866.1 / 46.761.8 / 39.564.4 / 47.9
mPMRbase81.1 / 68.958.5 / 41.063.6 / 50.568.5 / 52.160.3 / 46.468.3 / 49.256.6 / 32.965.3 / 48.7
XLM-R83.4 / 71.064.9 / 45.869.6 / 54.874.1 / 56.870.7 / 53.473.3 / 53.064.4 / 42.471.5 / 53.9
mPMR84.0 / 71.466.4 / 47.070.3 / 56.274.5 / 57.171.4 / 54.174.7 / 54.470.5 / 47.373.1 / 55.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.318, + 0.181, + 0.678, + 0.196 + ], + "angle": 0, + "content": "Table 8: MLQA results (F1 / EM) for each language." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.216, + 0.876, + 0.284 + ], + "angle": 0, + "content": "
ModelenarbnfiidkoruswteAvg.
XLM-Rbase66.8 / 57.355.7 / 42.031.5 / 20.452.6 / 40.369.1 / 55.636.3 / 27.954.8 / 36.553.0 / 34.737.4 / 28.850.8 / 38.2
mPMRbase71.1 / 61.666.3 / 52.656.5 / 41.665.5 / 53.173.9 / 63.750.4 / 38.864.4 / 37.957.4 / 41.165.3 / 50.463.4 / 49.0
XLM-R71.3 / 60.769.3 / 52.366.2 / 53.164.3 / 51.376.5 / 62.558.3 / 46.764.7 / 43.468.6 / 53.167.3 / 41.167.4 / 51.6
mPMR76.4 / 65.276.0 / 58.072.3 / 55.874.4 / 56.584.1 / 71.362.2 / 50.772.5 / 43.276.5 / 63.177.7 / 60.874.7 / 58.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.29, + 0.295, + 0.706, + 0.309 + ], + "angle": 0, + "content": "Table 9: TyDiQA-GoldP results (F1 / EM) for each language." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.33, + 0.88, + 0.457 + ], + "angle": 0, + "content": "
Modelenafarbgbndeeleseteufafifrhehihuiditjajv
XLM-Rbase84.275.347.379.066.377.575.378.069.656.038.170.481.450.867.972.451.079.619.663.9
mPMRbase85.180.757.680.271.981.277.679.579.171.349.680.482.465.271.782.258.683.543.272.0
XLM-R85.481.153.984.073.882.382.880.468.854.864.275.981.459.372.976.459.384.613.271.2
mPMR86.081.756.185.979.682.382.375.582.769.675.284.182.066.575.984.059.986.149.172.4
kakkkomlmrmsmynlptruswtatethtltrurviyozh
XLM-Rbase58.740.634.350.846.063.840.681.580.065.476.143.046.44.271.968.745.770.91.523.0
mPMRbase72.245.152.962.459.468.157.483.781.571.877.350.557.43.074.280.355.775.231.649.9
XLM-R59.941.741.356.858.276.729.686.185.272.277.652.351.67.178.870.964.080.027.222.4
mPMR77.346.857.970.668.173.857.886.083.672.879.862.658.13.883.080.376.283.636.154.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.306, + 0.467, + 0.69, + 0.481 + ], + "angle": 0, + "content": "Table 10: WikiAnn results (F1 Score) for each language." + }, + { + "type": "table", + "bbox": [ + 0.31, + 0.501, + 0.687, + 0.599 + ], + "angle": 0, + "content": "
ModelendeesnlAvg.
XLM-Rbase91.371.078.775.779.2
mPMRbase91.974.380.879.781.7
XLM-R92.873.781.677.781.4
mPMR93.575.085.083.184.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.31, + 0.609, + 0.685, + 0.624 + ], + "angle": 0, + "content": "Table 11: CoNLL results (F1 Score) for each language." + }, + { + "type": "table", + "bbox": [ + 0.258, + 0.644, + 0.74, + 0.742 + ], + "angle": 0, + "content": "
ModelenesfrnlrutrAvg.
XLM-Rbase76.565.455.661.256.145.460.0
mPMRbase77.668.656.462.259.548.462.1
XLM-R82.471.360.367.461.249.166.1
mPMR82.871.964.767.466.955.768.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.298, + 0.752, + 0.698, + 0.767 + ], + "angle": 0, + "content": "Table 12: SemEval16 results (F1 Score) for each language." + }, + { + "type": "table", + "bbox": [ + 0.232, + 0.787, + 0.766, + 0.884 + ], + "angle": 0, + "content": "
ModelendeesfrjakozhAvg.
XLM-Rbase94.387.789.188.777.076.681.385.0
mPMRbase94.388.490.188.979.079.482.486.1
XLM-R95.289.391.090.979.679.982.586.9
mPMR95.290.690.391.381.282.984.688.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.289, + 0.894, + 0.707, + 0.909 + ], + "angle": 0, + "content": "Table 13: PAWS-X accuracy scores (Acc.) for each language." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1543" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.12, + 0.447, + 0.882, + 0.526 + ], + "angle": 0, + "content": "
ModelenarbgdeelesfrhiruswthtrurvizhAvg.
XLM-Rbase84.671.076.875.674.977.976.968.974.164.471.172.465.273.273.073.3
mPMRbase84.271.577.275.575.578.676.969.574.762.571.471.665.574.374.073.6
XLM-R88.277.081.781.281.284.281.774.978.970.875.777.470.678.077.778.6
mPMR88.377.982.982.281.083.582.275.279.871.276.178.971.678.979.079.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.3, + 0.536, + 0.696, + 0.551 + ], + "angle": 0, + "content": "Table 14: XNLI accuracy scores (Acc.) for each language." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1544" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.133, + 0.085, + 0.435, + 0.1 + ], + "angle": 0, + "content": "ACL 2023 Responsible NLP Checklist" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.108, + 0.323, + 0.123 + ], + "angle": 0, + "content": "A For every submission:" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.127, + 0.533, + 0.145 + ], + "angle": 0, + "content": "A1. Did you describe the limitations of your work?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.146, + 0.298, + 0.158 + ], + "angle": 0, + "content": "Section Limitations" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.171, + 0.553, + 0.187 + ], + "angle": 0, + "content": "A2. Did you discuss any potential risks of your work?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.188, + 0.351, + 0.202 + ], + "angle": 0, + "content": "Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.213, + 0.696, + 0.229 + ], + "angle": 0, + "content": "A3. Do the abstract and introduction summarize the paper's main claims?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.231, + 0.294, + 0.244 + ], + "angle": 0, + "content": "Abstract, Section 1" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.256, + 0.669, + 0.273 + ], + "angle": 0, + "content": "A4. Have you used AI writing assistants when working on this paper?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.274, + 0.233, + 0.288 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.3, + 0.489, + 0.317 + ], + "angle": 0, + "content": "B Did you use or create scientific artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.322, + 0.206, + 0.336 + ], + "angle": 0, + "content": "Section 3" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.347, + 0.53, + 0.364 + ], + "angle": 0, + "content": "B1. Did you cite the creators of artifacts you used?" + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.365, + 0.223, + 0.378 + ], + "angle": 0, + "content": "Section 3" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.391, + 0.779, + 0.407 + ], + "angle": 0, + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.408, + 0.351, + 0.423 + ], + "angle": 0, + "content": "Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.433, + 0.882, + 0.497 + ], + "angle": 0, + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.499, + 0.331, + 0.513 + ], + "angle": 0, + "content": "Section 3, Appendix A.1" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.525, + 0.882, + 0.573 + ], + "angle": 0, + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.574, + 0.351, + 0.589 + ], + "angle": 0, + "content": "Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.599, + 0.882, + 0.632 + ], + "angle": 0, + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.634, + 0.361, + 0.648 + ], + "angle": 0, + "content": "Appendix A.1, Appendix A.2" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.658, + 0.882, + 0.74 + ], + "angle": 0, + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be." + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.741, + 0.253, + 0.755 + ], + "angle": 0, + "content": "Appendix A.1" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.765, + 0.494, + 0.782 + ], + "angle": 0, + "content": "C Did you run computational experiments?" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.788, + 0.206, + 0.801 + ], + "angle": 0, + "content": "Section 4" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.813, + 0.882, + 0.846 + ], + "angle": 0, + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.847, + 0.331, + 0.861 + ], + "angle": 0, + "content": "Section 3, Appendix A.1" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.869, + 0.878, + 0.893 + ], + "angle": 0, + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "1545" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.13, + 0.084, + 0.88, + 0.133 + ], + "angle": 0, + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Appendix A.1" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.143, + 0.88, + 0.207 + ], + "angle": 0, + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Section 4" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.218, + 0.88, + 0.283 + ], + "angle": 0, + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Section 3, Appendix A.1" + }, + { + "type": "list", + "bbox": [ + 0.13, + 0.084, + 0.88, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.294, + 0.878, + 0.331 + ], + "angle": 0, + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.341, + 0.88, + 0.388 + ], + "angle": 0, + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.4, + 0.88, + 0.464 + ], + "angle": 0, + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.476, + 0.88, + 0.539 + ], + "angle": 0, + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.55, + 0.873, + 0.582 + ], + "angle": 0, + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response." + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.593, + 0.88, + 0.641 + ], + "angle": 0, + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response." + }, + { + "type": "list", + "bbox": [ + 0.128, + 0.341, + 0.88, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.522, + 0.941 + ], + "angle": 0, + "content": "1546" + } + ] +] \ No newline at end of file diff --git a/2023/mPMR_ A Multilingual Pre-trained Machine Reader at Scale/24809581-970f-402d-a256-c7d8e2f40ff3_origin.pdf b/2023/mPMR_ A Multilingual Pre-trained Machine Reader at Scale/24809581-970f-402d-a256-c7d8e2f40ff3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0d5c379d4f80f50cf3f4ebd1b9b2f0364a57c112 --- /dev/null +++ b/2023/mPMR_ A Multilingual Pre-trained Machine Reader at Scale/24809581-970f-402d-a256-c7d8e2f40ff3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:030b8d820c90ab252f2d3f5b6a743b79615a90e63ea506e6a4b133d58c1d8df3 +size 1215654 diff --git a/2023/mPMR_ A Multilingual Pre-trained Machine Reader at Scale/full.md b/2023/mPMR_ A Multilingual Pre-trained Machine Reader at Scale/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b87fa2f1571973388ed14f8be2d3dcfb5b48f2f3 --- /dev/null +++ b/2023/mPMR_ A Multilingual Pre-trained Machine Reader at Scale/full.md @@ -0,0 +1,290 @@ +# mPMR: A Multilingual Pre-trained Machine Reader at Scale* + +Weiwen Xu $^{12,\dagger}$ Xin Li $^{2,\ddagger}$ Wai Lam $^{1}$ Lidong Bing $^{2}$ + +1The Chinese University of Hong Kong + +$^{2}$ DAMO Academy, Alibaba Group + +{wxxu,wlam}@se.cuhk.edu.hk {xinting.lx,l.bing}@alibaba-inc.com + +# Abstract + +We present multilingual Pre-trained Machine Reader (mPMR), a novel method for multilingual machine reading comprehension (MRC)-style pre-training. mPMR aims to guide multilingual pre-trained language models (mPLMs) to perform natural language understanding (NLU) including both sequence classification and span extraction in multiple languages. To achieve cross-lingual generalization when only source-language fine-tuning data is available, existing mPLMs solely transfer NLU capability from a source language to target languages. In contrast, mPMR allows the direct inheritance of multilingual NLU capability from the MRC-style pre-training to downstream tasks. Therefore, mPMR acquires better NLU capability for target languages. mPMR also provides a unified solver for tackling cross-lingual span extraction and sequence classification, thereby enabling the extraction of rationales to explain the sentence-pair classification process. $^{1}$ + +# 1 Introduction + +Multilingual pre-trained language models, acronymed as mPLMs, have demonstrated strong Natural language understanding (NLU) capability in a wide range of languages (Xue et al., 2021; Cai et al., 2021, 2022; Conneau et al., 2020a; Ding et al., 2022; Li et al., 2020a). In particular, mPLMs can maintain exceptional cross-lingual language understanding (XLU) capability on unseen target languages though mPLMs are only fine-tuned on resource-rich source languages like English. + +It has been proved that optimizing cross-lingual representations of mPLMs can improve XLU ca + +![](images/56b276514635dad5b83103bb1b86a7c6db5594d38a67f412d88b687c12d37b95.jpg) +Figure 1: Pre-training and fine-tuning of mPMR. + +pability. For example, cross-lingual supervisions, such as parallel sentences (Conneau and Lample, 2019) or bilingual dictionaries (Conneau et al., 2020b) could enhance cross-lingual representations with better language alignment. XLM-R (Conneau et al., 2020a) and mT5 (Xue et al., 2021) showed that appropriately incorporating more languages during pre-training leads to better cross-lingual representations. A few works enriched the cross-lingual representations with factual knowledge through the utilization of multilingual mentions of entities (Calixto et al., 2021; Ri et al., 2022) and relations (Liu et al., 2022; Jiang et al., 2022) annotated in knowledge graphs. Despite their differences, the above methods essentially constructed more diverse multilingual corpora for pre-training mPLMs. These mPLMs would presumably meet their saturation points and are known to suffer from curse of multilinguality (Conneau et al., 2020a; Pfeiffer et al., 2022; Berend, 2022). Under this situation, introducing more training data from either existing (Pfeiffer et al., 2022) or unseen (Conneau et al., 2020a) languages for enhancing mPLMs may not bring further improvement or even be detrimental to their cross-lingual representations. + +In the paper, instead of training a new mPLM with better cross-lingual representations, we propose multilingual Pre-trained Machine Reader (mPMR) to directly guide existing mPLMs to perform NLU in various languages. As shown in Figure 1, mPMR resembles PMR (Xu et al., 2022) for constructing multilingual machine reading comprehension (MRC)-style data with Wikipedia hyperlinks. These data are used to retrofit an mPLM into an mPMR through an MRC-style continual pre-training. During retrofitting process (i.e., pretraining), mPMR jointly learns the general sequence classification and span extraction capability for multiple languages. In XLU fine-tuning, mPLMs solely rely on cross-lingual representations to transfer NLU capability from a source language to target languages. By contrast, mPMR enables the direct inheritance of multilingual NLU capability from the MRC-style pre-training to downstream tasks in a unified MRC formulation, which alleviates the discrepancies between source-language fine-tuning and target-language inference (Zhou et al., 2022a,b, 2023). Therefore, mPMR shows greater potential in XLU than mPLMs. + +To improve the scalability of mPMR across multiple languages, we further propose Unified Q/C Construction and Stochastic answer position strategies for refining the curation of MRC data. With these two strategies, mPMR can better generalize to low-resource languages and becomes more robust to position bias (Ko et al., 2020). + +The experimental results show that mPMR obtains clear improvements over XLM-R (Conneau et al., 2020a) on span extraction, with an average improvement of up to 12.6 F1 on TyDiQA, and 8.7 F1 on WikiAnn respectively. The analysis reveals that mPMR benefits from more multilingual MRC data for pre-training. We also found that mPMR converges faster in downstream tasks and is capable of using its strong extraction capability for explaining the sequence classification process. + +# 2 mPMR + +We present the MRC model and training data of mPMR. We closely follow PMR (Xu et al., 2022) and introduce the modifications for enabling multilingual MRC-style pre-training. + +# 2.1 Model Pre-training + +Our mPMR follows the same MRC architecture of Xu et al. (2022, 2023) with an encoder and an + +extractor. The encoder maps input tokens $X$ , the concatenation of the query $Q$ , the context $C$ , and special markers (i.e., [CLS] and [SEP]), into hidden representations $H$ . For any two tokens $X_{i}$ and $X_{j}$ ( $i < j$ ), the extractor receives their contextualized representations $H_{i}$ and $H_{j}$ and predicts the probability score $S_{i,j}$ indicating the probability of the token span $X_{i:j}$ being the answer to the query $Q$ . + +mPMR is guided with the Wiki Anchor Extraction (WAE) objective to train both the encoder and the extractor. WAE checks if the answer to the query exists in the context. If so, WAE would first regard the query and the context to be relevant and extracts the [CLS] token as a sequence-level relevance indicator. WAE would then extract all corresponding answers from the context. + +# 2.2 Multilingual MRC Data + +Training mPMR requires the existence of labeled (query, context, answer) triplets. To obtain such data, we collected Wikipedia articles with anchor annotations for 24 languages, which are the most widely used and cover a reasonable number of languages used in XLU tasks (Ri et al., 2022). + +As shown in Figure 1, we utilized a Wikipedia anchor to obtain a pair of correlated articles. One side of the pair is the article that provides in-depth descriptions of the anchor entity, which we defined as the definition article. The other side of the pair is named as the mention article, which mentions the specific anchor text2. We composed an answerable MRC example in which the anchor is the answer, the surrounding text of the anchor in the mention article is the context, and the definition of the anchor entity in the definition article is the query. Additionally, we can generate an unanswerable MRC example by pairing a query with an irrelevant context without anchor association. + +Unified Q/C Construction. PMR constructed the MRC query and context as valid sentences so as to keep the text coherent. However, sentence segmentation tools are usually not available for low-resource languages. To remedy this, we did not apply sentence segmentation but only preprocess Wikipedia articles with word tokenization in mPMR. For each anchor, the MRC query comprises the first $Q$ words in the definition article. To prevent information leakage during pre-training, similar to PMR, we anonymized the anchor entity + +
Model#ParamsEQANERABSASentence PairAvg.
XQuADMLQATyDiQAWikiAnnCoNLLSemEval16PAWS-XXNLI
MetricsF1 / EMF1 / EMF1 / EMF1F1F1Acc.Acc.
XLM-R550M76.6 / 60.871.6 / 53.265.1 / 45.065.482.066.9‡86.479.274.2
mT5580M67.0 / 49.064.6 / 45.057.2 / 41.255.771.0‡62.5‡86.475.467.5
VECO550M77.3 / 61.871.7 / 53.267.6 / 49.165.781.3‡63.0‡88.779.974.4
mLUKE-W561M79.6 / -72.7 / -65.2 / 48.5‡67.7‡83.061.2‡88.2‡79.4‡74.6
Wiki-CL550M72.1 / 56.970.8 / 50.573.2 / 57.364.7--88.479.2-
KMLM550M77.3 / 61.772.1 / 53.767.9 / 50.466.7‡83.266.1‡88.079.275.1
Our MRC Formulation
XLM-Rbase270M70.8 / 56.964.4 / 47.950.8 / 38.257.979.260.085.073.367.7
mPMRbase270M74.0 / 59.565.3 / 48.763.4 / 49.066.681.762.186.173.671.6
XLM-R550M77.1 / 61.371.5 / 53.967.4 / 51.663.681.466.186.978.674.1
mPMR550M79.2 / 64.473.1 / 55.474.7 / 58.370.784.168.288.079.377.2
+ +Table 1: The results of all XLU tasks. We report the average results of all languages for each dataset. We also compute the overall average score among all datasets in the Avg. column. We reproduce the missing results with the $\ddagger$ label. Some results of Wiki-CL are left blank because they do not release their model checkpoint. + +in the query to the [MASK] token. The MRC context consists of $C$ words surrounding the anchor. + +Stochastic Answer Position. As mentioned by Ko et al. (2020), the model is prone to overfitting to the position shortcut if the answer in the context exhibits a fixed position pattern. In our case, suppose that the MRC context consists of $C / 2$ words on both the left and right sides of the anchor, the model may learn the shortcut that the middle part of the context is likely to be the answer. To prevent such position bias, we propose a stochastic answer position method, which allows the answer to be presented in any position within the context. Specifically, given an anchor in a Wikipedia article, the context comprises $\xi$ words preceding the anchor and the $C - \xi$ words following the anchor, where $\xi$ is a random integer ranging from 0 to $C$ and varies across different contexts. In accordance with PMR, we treated all text spans identical to the anchor in the current context as valid answers. + +# 3 Experimental Setup + +Implementation Details. In mPMR, the encoder is loaded from XLM-R (Conneau et al., 2020a) and the extractor is randomly initialized. Both components are then continually pre-trained using the multilingual MRC data that we constructed. More hyper-parameters can be found in Appendix A.1. + +Downstream XLU Tasks. We evaluated mPMR on a series of span extraction tasks, including Extractive Question Answering (EQA), Named Entity Recognition (NER), and Aspect-Based Sentiment + +Analysis (ABSA). We also evaluated our mPMR on two sequence classification tasks. We followed Xu et al. (2022) to convert all tasks into MRC formulation to effectively leverage the knowledge that is acquired during MRC-style pre-training. For EQA, we used XQuAD (Artetxe et al., 2020), MLQA (Lewis et al., 2020), and TyDiQA (Clark et al., 2020). For NER, we used WikiAnn (Pan et al., 2017) and CoNLL (Tjong Kim Sang, 2002; Tjong Kim Sang and De Meulder, 2003). SemEval16 (Pontiki et al., 2016) was used for ABSA task. Regarding the sequence classification, we used XNLI (Conneau et al., 2018) and PAWS-X (Yang et al., 2019). Additional dataset information and concrete examples are provided in Appendix A.2 + +Baselines. We compared mPMR with recent methods on improving cross-lingual representations, including 1) models pre-trained on a large number of languages: XLM-R (Conneau et al., 2020a), mT5 (Xue et al., 2021), and VECO (Luo et al., 2021); 2) models that exploited multilingual entity information: Wiki-CL (Calixto et al., 2021), and mLUKE-W (Ri et al., 2022); and 3) Model that utilized multilingual relation information: KMLM (Liu et al., 2022). For a fair comparison, all models have approximately the same parameter size. + +# 4 Results and Analyses + +XLU Performance. Table 1 shows the results on a variety of XLU tasks. mPMR outperforms all previous methods with an absolute improvement of 2.1 F1 over the best baseline (i.e. KMLM). mPMR shows greater improvements over previ + +
IndexModel#LangPAWS-XXQuADWikiAnnAvg.
#1XLM-Rbase085.070.857.971.2
#2#1 + MRC data in English185.2 (0.2↑)71.0 (0.2↑)59.5 (1.6↑)71.9 (0.7↑)
#3#2 + Stochastic Answer Position185.5 (0.3↑)73.0 (2.0↑)60.0 (0.5↑)72.8 (0.9↑)
#4#3 + MRC data in more languages1085.9 (0.4↑)73.5 (0.5↑)64.7 (4.7↑)74.7 (1.9↑)
#5#4 + MRC data in even more languages (mPMRbase)2486.1 (0.2↑)74.0 (0.5↑)66.6 (1.9↑)75.6 (0.9↑)
+ +Table 2: The process of retrofitting XLM-R into mPMR using multilingual MRC data (English→10 languages→24 languages) and our Stochastic Answer Position method. Each row accumulates modifications from all rows above. + +
LabelSentence 1Sentence 2
EntailmentRami Nieminen ( born February 25 , 1966 ) is a Finnish footballer.Rami Nieminen ( born 25 February 1966 ) is a Finnish former footballer.
ContradictionIn 1938 he became the Government Anthropologist of the Egyptian-Anglo Sudan and conducted fieldwork with the Nuba.In 1938 he became the government anthropologist of the anglo-Egyptian Sudan and led fieldwork with the Nuba.
EntailmentStipsits 出生于科尔新堡，并在维也纳施塔莫斯多夫度过了他的童年。什蒂普西奇出生于德国科恩堡，在维也纳斯塔莫斯多夫度过了他的童年。
Contradiction纳舒厄白银骑士团队加入了夏季大学联盟，是本市的现役球队。Nashua Silver Knights 队是当前夏季联赛的一部分，也是该市的大学体育队。
Entailmentごれらの見方は、福音主義的、清教徒的、プロデ斯特兰トの動態が出現すると必に、しはしだは表明くださいます。ごれらの見解は多くの场合、新教徒、清教徒、福音主義者が出現する:NOか表示お願いいたします。
Contradiction1954年にスリーナムに戸った後、弁護士とでラマリポに定住したこと。1954年、バラマリポに戸ると、彼はスリーナムで弁護士とで定住,No理由。
+ +Table 3: Case study on PAWS-X. mPMR can extract rationales to explain the sequence-pair classification in multiple languages. + +ous methods on span extraction tasks. In particular, mPMR achieves up to 7.3 and 7.1 F1 improvements over XLM-R on TyDiQA and WikiAnn respectively. Such significant improvements probably come from the following two facts: (1) WikiAnn comprises a larger number of target languages (i.e. 40). Therefore, existing methods may struggle to align these low-resource languages with English due to a lack of language-specific data. (2) TyDiQA is a more challenging cross-lingual EQA task with $2\mathrm{x}$ less lexical overlap between the query and the answer than MLQA and XQuAD (Hu et al., 2020). Our mPMR, which acquires target-language span extraction capability from both MRC-style pretraining and English-only QA fine-tuning, achieves larger performance gains on more challenging task. + +mPMR Pre-training. To reflect the impact of our MRC-style data and Stochastic Answer Position method on pre-training, we present a step-by-step analysis of the retrofitting process starting from XLM-R in Table 2. Our findings suggest that the significant improvements observed are largely due to the inclusion of multilingual MRC data. Introducing English MRC data (model #2) gives marginal improvements because model #2 + +![](images/fde14688ed094287da5e02c2f3852fb01624809fa4b9d24f6b0d5328fe8edd71.jpg) +Figure 2: Convergence speed (Test set F1 and the training loss) of $\mathrm{mPMR_{base}}$ and XLM- $\mathbf{R}_{\mathrm{base}}$ on WikiAnn. + +can only rely on cross-lingual representations to transfer the knowledge acquired during MRC-style pre-training. When using MRC data on more languages (model #4 and #5), we can observe significant improvements on XLU tasks. This can be attributed to the NLU capability directly inherited from MRC-style pre-training in target languages. Additionally, with our Stochastic Answer Position method (model #3), mPMR becomes more robust to position bias and thus improves XLU tasks. + +Explainable Sentence-pair Classification. Inspired by PMR (Xu et al., 2022), we investigated if the extraction capability of mPMR can be leveraged to explain sentence-pair classification. Note + +![](images/088e81490d4be3fc5e3553fa86604e19088ecab031db99dbd74c610ab80eb61f.jpg) +Figure 3: Convergence speed (Test set F1 and the training loss) of $\mathrm{mPMR_{base}}$ and XLM- $\mathbf{R}_{\mathrm{base}}$ on XQuAD. + +that sentence-pair classification focuses on the inference between the two sentences. If we construct the query with only the task label as PMR does, such query does not solely correspond to any meaningful span in the context, and thus is hard to guide the span extraction. Therefore, we leveraged another template "[CLS] label Sen-1 [SEP] Sen-2 [SEP]", where the two sentences are represented separately in the query and the context. In this template, we can extract the exact span from Sen-2 that leads to a contraction or entailment relation (i.e., the task label) with Sen-1. Specifically, we passed the sentence pair to the model twice, with each sentence of the pair being designated as the Sen-2 respectively, and extract the context span with the highest probability score from both sentences. + +As shown in Table 3, the extracted spans are indeed important rationales that determine the relationship between two sentences. Such a finding confirms that the extraction capability of mPMR can be appropriately used for explaining the sentence-pair classification process. While the extraction capability may affect the learning of sequence classification during fine-tuning, resulting in a 0.4 Acc. decrease on XNLI. + +mPMR Fine-tuning. We investigated the effects of mPMR on XLU fine-tuning. Figure 2 shows that mPMR converges faster than XLM-R on WikiAnn with an extremely low loss value even fine-tuned for 500 steps. In terms of test set performance, mPMR outperforms XLM-R comprehensively and exhibits greater stability. As a result, mPMR provides a better starting point for addressing XLU tasks compared to XLM-R. More examples from XQuAD and PAWS-X are provided in Figure 3 and 4. + +![](images/19c3ea4971b85aae02e844457d45899a7c9aa439a0f7c0812810227e18894eb3.jpg) +Figure 4: Convergence speed (Test set F1 and the training loss) of $\mathrm{mPMR_{base}}$ and XLM- $\mathbf{R}_{\mathrm{base}}$ on PAWS-X. + +# 5 Conclusions + +This paper presents a novel multilingual MRC-style pre-training method, namely mPMR. mPMR provides a unified solver for cross-lingual span extraction and sequence classification and enables direct transfer of NLU capability from pre-training to downstream tasks. mPMR clearly improves the previous baselines and provides a possible solution to explain the sentence-pair classification process. + +# Limitations + +We identify the following two limitations of our work: + +- Different from raw text, constructing MRC-style data from Wikipedia requires the existence of hyperlinks. This idea works well for resource-rich languages, such as English and Chinese. While such an idea is less effective for languages with few hyperlink annotations in Wikipedia because a small amount of MRC-style training data is difficult to guide the learning of NLU capability in those languages. A possible solution is to explore other data resources to automatically construct large-scale MRC data for pre-training. +- As observed in Table 1, the improvements of sequence classification tasks are less significant than those of span extraction tasks. We suggest that the existence of anchors is not a strong relevance indicator between our constructed query and context. Such a finding is also observed in Chang et al. (2020). Therefore, constructing more relevant query-context pairs for sequence classification pre-training can possibly remedy this issue. + +# References + +Mikel Artetxe, Sebastian Ruder, and Dani Yogatama. 2020. On the cross-lingual transferability of monolingual representations. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. +Giuseppe Attardi. 2015. Wikiextractor. https://github.com/attardi/wikiextractor. +Gábor Berend. 2022. Combating the curse of multilinguality in cross-lingual WSD by aligning sparse contextualized word representations. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. +Deng Cai, Xin Li, Jackie Chun-Sing Ho, Lidong Bing, and Wai Lam. 2021. Multilingual AMR parsing with noisy knowledge distillation. In Findings of the Association for Computational Linguistics: EMNLP 2021. +Deng Cai, Xin Li, Jackie Chun-Sing Ho, Lidong Bing, and Wai Lam. 2022. Retrofitting multilingual sentence embeddings with Abstract Meaning Representation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. +Iacer Calixto, Alessandro Raganato, and Tommaso Pasini. 2021. Wikipedia entities as rendezvous across languages: Grounding multilingual language models by predicting Wikipedia hyperlinks. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. +Wei-Cheng Chang, Felix X. Yu, Yin-Wen Chang, Yiming Yang, and Sanjiv Kumar. 2020. Pre-training tasks for embedding-based large-scale retrieval. In International Conference on Learning Representations. +Jonathan H. Clark, Eunsol Choi, Michael Collins, Dan Garrette, Tom Kwiatkowski, Vitaly Nikolaev, and Jennimaria Palomaki. 2020. TyDi QA: A benchmark for information-seeking question answering in typologically diverse languages. Transactions of the Association for Computational Linguistics. +Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2020a. Unsupervised cross-lingual representation learning at scale. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. +Alexis Conneau and Guillaume Lample. 2019. Crosslingual language model pretraining. In Advances in Neural Information Processing Systems. +Alexis Conneau, Rudy Rinott, Guillaume Lample, Adina Williams, Samuel Bowman, Holger Schwenk, and Veselin Stoyanov. 2018. XNLI: Evaluating crosslingual sentence representations. In Proceedings of + +the 2018 Conference on Empirical Methods in Natural Language Processing. +Alexis Conneau, Shijie Wu, Haoran Li, Luke Zettlemoyer, and Veselin Stoyanov. 2020b. Emerging cross-lingual structure in pretrained language models. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. +Bosheng Ding, Junjie Hu, Lidong Bing, Mahani Aljunied, Shafiq Joty, Luo Si, and Chunyan Miao. 2022. GlobalWoZ: Globalizing MultiWoZ to develop multilingual task-oriented dialogue systems. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). +Junjie Hu, Sebastian Ruder, Aditya Siddhant, Graham Neubig, Orhan First, and Melvin Johnson. 2020. Xtreme: A massively multilingual multi-task benchmark for evaluating cross-lingual generalisation. In International Conference on Machine Learning. +Xiaoze Jiang, Yaobo Liang, Weizhu Chen, and Nan Duan. 2022. Xlm-k: Improving cross-lingual language model pre-training with multilingual knowledge. In Proceedings of the AAAI Conference on Artificial Intelligence. +Miyoung Ko, Jinhyuk Lee, Hyunjae Kim, Gangwoo Kim, and Jaewoo Kang. 2020. Look at the first sentence: Position bias in question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). +Patrick Lewis, Barlas Oguz, Rudy Rinott, Sebastian Riedel, and Holger Schwenk. 2020. MLQA: Evaluating cross-lingual extractive question answering. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. +Juntao Li, Ruidan He, Hai Ye, Hwee Tou Ng, Lidong Bing, and Rui Yan. 2020a. Unsupervised domain adaptation of a pretrained cross-lingual language model. In Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, IJCAI 2020. +Xin Li, Lidong Bing, Wenxuan Zhang, Zheng Li, and Wai Lam. 2020b. Unsupervised cross-lingual adaptation for sequence tagging and beyond. arXiv preprint arXiv:2010.12405. +Linlin Liu, Xin Li, Ruidan He, Lidong Bing, Shafiq Joty, and Luo Si. 2022. Enhancing multilingual language model with massive multilingual knowledge triples. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. +Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In International Conference on Learning Representations. +Fuli Luo, Wei Wang, Jiahao Liu, Yijia Liu, Bin Bi, Songfang Huang, Fei Huang, and Luo Si. 2021. VECO: Variable and flexible cross-lingual pre-training for + +language understanding and generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). +Xiaoman Pan, Boliang Zhang, Jonathan May, Joel Nothman, Kevin Knight, and Heng Ji. 2017. Cross-lingual name tagging and linking for 282 languages. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). +Jonas Pfeiffer, Naman Goyal, Xi Lin, Xian Li, James Cross, Sebastian Riedel, and Mikel Artetxe. 2022. Lifting the curse of multilinguality by pre-training modular transformers. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. +Maria Pontiki, Dimitris Galanis, Haris Papageorgiou, Ion Androutsopoulos, Suresh Manandhar, Mohammad AL-Smadi, Mahmoud Al-Ayyoub, Yanyan Zhao, Bing Qin, Orphée De Clercq, Véronique Hoste, Marianna Apidianaki, Xavier Tannier, Natalia Loukachevitch, Evgeniy Kotelnikov, Nuria Bel, Salud María Jiménez-Zafra, and Gülşen Eryigit. 2016. SemEval-2016 task 5: Aspect based sentiment analysis. In Proceedings of the 10th International Workshop on Semantic Evaluation (SemEval-2016). +Ryokan Ri, Ikuya Yamada, and Yoshimasa Tsuruoka. 2022. mLUKE: The power of entity representations in multilingual pretrained language models. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). +Erik F. Tjong Kim Sang. 2002. Introduction to the CoNLL-2002 shared task: Language-independent named entity recognition. In COLING-02: The 6th Conference on Natural Language Learning 2002 (CoNLL-2002). +Erik F. Tjong Kim Sang and Fien De Meulder. 2003. Introduction to the CoNLL-2003 shared task: Language-independent named entity recognition. In Proceedings of the Seventh Conference on Natural Language Learning at HLT-NAACL 2003. +Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations. +Weiwen Xu, Xin Li, Yang Deng, Wai Lam, and Lidong Bing. 2023. Peerda: Data augmentation via modeling + +peer relation for span identification tasks. In The 61th Annual Meeting of the Association for Computational Linguistics. +Weiwen Xu, Xin Li, Wenxuan Zhang, Meng Zhou, Lidong Bing, Wai Lam, and Luo Si. 2022. From clozing to comprehending: Retrofitting pre-trained language model to pre-trained machine reader. arXiv preprint arXiv:2212.04755. +Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. 2021. mT5: A massively multilingual pre-trained text-to-text transformer. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. +Yinfei Yang, Yuan Zhang, Chris Tar, and Jason Baldridge. 2019. PAWS-X: A cross-lingual adversarial dataset for paraphrase identification. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). +Wenxuan Zhang, Ruidan He, Haiyun Peng, Lidong Bing, and Wai Lam. 2021. Cross-lingual aspect-based sentiment analysis with aspect term code-switching. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. +Meng Zhou, Xin Li, Yue Jiang, and Lidong Bing. 2022a. Enhancing cross-lingual prompting with mask token augmentation. arXiv preprint arXiv:2202.07255. +Ran Zhou, Xin Li, Lidong Bing, Erik Cambria, and Chunyan Miao. 2023. Improving self-training for cross-lingual named entity recognition with contrastive and prototype learning. In *The 61th Annual Meeting of the Association for Computational Linguistics*. +Ran Zhou, Xin Li, Lidong Bing, Erik Cambria, Luo Si, and Chunyan Miao. 2022b. ConNER: Consistency training for cross-lingual named entity recognition. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. + +# A Appendix + +# A.1 More Implementation Details + +We collect the 2022-08-01 dump3 of Wikipedia articles for the 24 languages in consideration. The statistics of each language can be found in Table 4. Then for each article, we extract the plain text with anchors via WikiExtractor (Attardi, 2015). Word tokenization is performed using spaCy4 if the language is supported, otherwise, we utilize PyThaiNLP5 for Thai and Sacremoses6 for remaining languages. For each anchor entity, we construct 10 answerable MRC examples and 10 unanswerable MRC examples as described in Sec. 2.2. Anchor entities with low frequency (below 10 occurrences for English entities and 5 occurrences for entities in other languages) were excluded. + +In mPMR, we use Huggingface's implementations of XLM-R (Wolf et al., 2020). During the pre-training stage, the query length $Q$ is set to 50 words, and the context length $C$ is set to 200 words. Both are computed before the subword segmentation. We follow the default learning rate schedule and dropout settings used in XLM-R. We use AdamW (Loshchilov and Hutter, 2019) as our optimizer. We train both $\mathrm{mPMR_{base}}$ and mPMR on 4 A100 GPU. The learning rate is set to 1e-5, and the effective batch size for each step is set to 256 and 80 for $\mathrm{mPMR_{base}}$ and mPMR respectively in order to maximize the usage of the GPU memory. We use the average scores of XQuAD, CoNLL, and PAWS-X to select the best mPMR checkpoint. In fact, we continually pre-train $\mathrm{mPMR_{base}}$ and mPMR for 250,000 and 100,000 steps. The training speed is around 6250 steps per hour. The hyper-parameters of $\mathrm{mPMR_{large}}$ on downstream XLU tasks can be found in Table 5. + +# A.2 Downstream XLU Tasks + +We evaluate mPMR on XLU tasks including both span extraction (EQA, NER, and ABSA) and sequence classification (sentence pair classification). We follow (Xu et al., 2022) to convert all tasks into MRC formulation and tackle them accordingly. We show concrete examples for each task in Table 6. Specifically, we evaluate the performance of EQA on three benchmarks: XQuAD (Artetxe et al., 2020), MLQA (Lewis et al., 2020), and Ty + +DiQA (Clark et al., 2020) covering 11, 7, and 9 languages respectively. For NER evaluation, we use the WikiAnn dataset (Pan et al., 2017) restricted to the 40 languages from XTREME (Hu et al., 2020), as well as the CoNLL dataset with 4 languages (Tjong Kim Sang, 2002; Tjong Kim Sang and De Meulder, 2003); We also evaluate the XLU performance of SemEval16 ABSA on 6 languages (Pontiki et al., 2016), where we collect the data from Li et al. (2020b); Zhang et al. (2021). Regarding the sequence classification task, we evaluate XNLI (Conneau et al., 2018) and PAWS-X (Yang et al., 2019) with 15 and 7 languages respectively. + +# A.3 mPMR Performance per Language + +We show the detailed results for each language in each task in Table 7 (XQuAD), Table 8 (MLQA), Table 9 (TyDiQA), Table 10 (WikiAnn), Table 11 (CoNLL), Table 12 (SemEval16), Table 13 (PAWS-X), and Table 14 (XNLI). + +
Language# Entities# MRC examplesLanguage# Entities# MRC examples
ar118,2922,020,502ko94,6161,597,076
bn25,081410,634nl251,3234,185,913
de864,74614,795,826pl283,9254,765,015
el56,383946,114pt216,6953,648,603
en966,19719,303,940ru432,4377,342,472
es412,4767,044,972sv169,0302,808,214
fi113,1181,960,636sw4,85765,724
fr595,87910,164,216te11,005170,664
hi15,350242,078th31,676522,434
id70,9601,164,662tr71,2941,175,276
it376,4176,421,850vi68,6651,147,772
ja423,8847,338,308zh259,7854,438,004
Total5,934,091103,680,905
+ +Table 4: Data statistics of mPMR pre-training data. The statistics is computed after removing the low-frequency entities. The number of MRC examples includes both answerable and unanswerable examples. + +
DatasetXQuADMLQATyDiQAWikiAnnCoNLLSemEval16PAWS-XXNLI
Query Length6464643232326464
Input Length384384384192192192192192
Batch Size8881616321632
Learning Rate3e-53e-52e-51e-51e-52e-55e-53e-5
Epoch3310101020103
+ +Table 5: Hyper-parameters settings in fine-tuning XLU tasks. + +
TaskExample InputExample Output
EQA(XSQuAD)Ori.Question: Who lost to the Broncos in the divisional round?Context: The Broncos defeated the Pittsburgh Steelers in the divi-sional round, 23–16, by scoring 11 points in the final three minutes of the game.Answer: "Pittsburgh Steelers"
PMR[CLS] Who lost to the Broncos in the divisional round ? [SEP] [SEP]The Broncos defeated the Pittsburgh Steelers in the divisional round, 23–16 , by scoring 11 points in the final three minutes of the game .[SEP](17,18) - "Pittsburgh Steelers"
NER(CoNLL)Ori.Two goals in the last six minutes gave holders Japan an uninspiring 2-1 Asian Cup victory over Syria on Friday.("Japan", LOC);("Syria", LOC);("Asian Cup", MISC)
PMR[CLS] "ORG". Organization entities are limited to named corporate,governmental, or other organizational entities. [SEP] [SEP] Two goals in the last six minutes gave holders Japan an uninspiring 2-1 Asian Cup victory over Syria on Friday . [SEP]∅
[CLS] "PER". Person entities are named persons or family . [SEP] [SEP] Two goals in the last six minutes gave holders Japan an uninspiring 2-1 Asian Cup victory over Syria on Friday . [SEP]∅
[CLS] "LOC". Location entities are the name of politically or geo-graphically defined locations such as cities , countries . [SEP] [SEP] Two goals in the last six minutes gave holders Japan an uninspiring 2-1 Asian Cup victory over Syria on Friday . [SEP](32,32) - "Japan";(40,40) - "Syria"
[CLS] "MISC". Examples of miscellaneous entities include events ,nationalities , products and works of art . [SEP] [SEP] Two goals in the last six minutes gave holders Japan an uninspiring 2-1 Asian Cup victory over Syria on Friday . [SEP](34,35) - "Asian Cup"
ABSA(SemEval16)Ori.Nice ambience, but highly overrated place.("ambience", POS);("place", NEG)
PMR[CLS] "POS". For aspect terms of positive sentiment . [SEP] [SEP] Nice ambience , but highly overrated place . [SEP](13,13) - "ambience"
[CLS] "NEG". For aspect terms of negative sentiment . [SEP] [SEP] Nice ambience , but highly overrated place . [SEP](18,18) - "place"
[CLS] "NEU". For aspect terms of neutral sentiment . [SEP] [SEP] Nice ambience , but highly overrated place . [SEP]∅
Sen. Pair Classification(PAWS-X)Ori.Hypothesis: The Tabaci River is a tributary of the River Leurda in Romania.Premise: The Leurda River is a tributary of the River Tabaci in Romania.Contradiction
PMR[CLS] Contradiction . The hypothesis is a sentence with a contradic-tory meaning to the premise . [SEP] [SEP] Hypothesis : The Tabaci River is a tributary of the River Leurda in Romania . Premise : The Leurda River is a tributary of the River Tabaci in Romania . [SEP](0,0) - "[CLS]"
[CLS] Entailment . The hypothesis is a sentence with a similar meaning as the premise . [SEP] [SEP] Hypothesis : The Tabaci River is a tributary of the River Leurda in Romania . Premise : The Leurda River is a tributary of the River Tabaci in Romania . [SEP]∅
+ +Table 6: MRC examples of XLU tasks. We use English examples here for demonstration purposes. Ori. indicates the original data format of these tasks. + +
ModelenardeeleshiruthtrvizhAvg.
XLM-Rbase82.2 / 72.065.5 / 49.973.9 / 59.771.2 / 56.376.3 / 59.466.4 / 52.073.7 / 58.964.7 / 54.667.0 / 52.873.3 / 54.765.0 / 55.970.8 / 56.9
mPMRbase84.4 / 73.469.6 / 53.276.4 / 61.574.9 / 58.477.4 / 60.269.2 / 54.575.2 / 58.869.2 / 57.670.4 / 55.874.8 / 55.871.8 / 65.574.0 / 59.5
XLM-R86.5 / 75.672.4 / 54.879.3 / 63.079.2 / 61.682.0 / 62.976.1 / 59.179.0 / 62.972.2 / 59.875.4 / 60.879.7 / 60.868.2 / 58.277.3 / 61.7
mPMR87.6 / 76.575.9 / 60.081.5 / 65.080.8 / 63.982.8 / 65.176.5 / 60.380.9 / 65.375.5 / 65.576.7 / 61.381.5 / 62.271.5 / 63.479.2 / 64.4
+ +Table 7: XQuAD results (F1 / EM) for each language. + +
ModelenardeeshivizhAvg.
XLM-Rbase79.3 / 67.255.4 / 38.162.0 / 49.166.8 / 50.259.4 / 44.866.1 / 46.761.8 / 39.564.4 / 47.9
mPMRbase81.1 / 68.958.5 / 41.063.6 / 50.568.5 / 52.160.3 / 46.468.3 / 49.256.6 / 32.965.3 / 48.7
XLM-R83.4 / 71.064.9 / 45.869.6 / 54.874.1 / 56.870.7 / 53.473.3 / 53.064.4 / 42.471.5 / 53.9
mPMR84.0 / 71.466.4 / 47.070.3 / 56.274.5 / 57.171.4 / 54.174.7 / 54.470.5 / 47.373.1 / 55.4
+ +Table 8: MLQA results (F1 / EM) for each language. + +
ModelenarbnfiidkoruswteAvg.
XLM-Rbase66.8 / 57.355.7 / 42.031.5 / 20.452.6 / 40.369.1 / 55.636.3 / 27.954.8 / 36.553.0 / 34.737.4 / 28.850.8 / 38.2
mPMRbase71.1 / 61.666.3 / 52.656.5 / 41.665.5 / 53.173.9 / 63.750.4 / 38.864.4 / 37.957.4 / 41.165.3 / 50.463.4 / 49.0
XLM-R71.3 / 60.769.3 / 52.366.2 / 53.164.3 / 51.376.5 / 62.558.3 / 46.764.7 / 43.468.6 / 53.167.3 / 41.167.4 / 51.6
mPMR76.4 / 65.276.0 / 58.072.3 / 55.874.4 / 56.584.1 / 71.362.2 / 50.772.5 / 43.276.5 / 63.177.7 / 60.874.7 / 58.3
+ +Table 9: TyDiQA-GoldP results (F1 / EM) for each language. + +
Modelenafarbgbndeeleseteufafifrhehihuiditjajv
XLM-Rbase84.275.347.379.066.377.575.378.069.656.038.170.481.450.867.972.451.079.619.663.9
mPMRbase85.180.757.680.271.981.277.679.579.171.349.680.482.465.271.782.258.683.543.272.0
XLM-R85.481.153.984.073.882.382.880.468.854.864.275.981.459.372.976.459.384.613.271.2
mPMR86.081.756.185.979.682.382.375.582.769.675.284.182.066.575.984.059.986.149.172.4
kakkkomlmrmsmynlptruswtatethtltrurviyozh
XLM-Rbase58.740.634.350.846.063.840.681.580.065.476.143.046.44.271.968.745.770.91.523.0
mPMRbase72.245.152.962.459.468.157.483.781.571.877.350.557.43.074.280.355.775.231.649.9
XLM-R59.941.741.356.858.276.729.686.185.272.277.652.351.67.178.870.964.080.027.222.4
mPMR77.346.857.970.668.173.857.886.083.672.879.862.658.13.883.080.376.283.636.154.4
+ +Table 10: WikiAnn results (F1 Score) for each language. + +
ModelendeesnlAvg.
XLM-Rbase91.371.078.775.779.2
mPMRbase91.974.380.879.781.7
XLM-R92.873.781.677.781.4
mPMR93.575.085.083.184.1
+ +Table 11: CoNLL results (F1 Score) for each language. + +
ModelenesfrnlrutrAvg.
XLM-Rbase76.565.455.661.256.145.460.0
mPMRbase77.668.656.462.259.548.462.1
XLM-R82.471.360.367.461.249.166.1
mPMR82.871.964.767.466.955.768.2
+ +Table 12: SemEval16 results (F1 Score) for each language. + +
ModelendeesfrjakozhAvg.
XLM-Rbase94.387.789.188.777.076.681.385.0
mPMRbase94.388.490.188.979.079.482.486.1
XLM-R95.289.391.090.979.679.982.586.9
mPMR95.290.690.391.381.282.984.688.0
+ +Table 13: PAWS-X accuracy scores (Acc.) for each language. + +
ModelenarbgdeelesfrhiruswthtrurvizhAvg.
XLM-Rbase84.671.076.875.674.977.976.968.974.164.471.172.465.273.273.073.3
mPMRbase84.271.577.275.575.578.676.969.574.762.571.471.665.574.374.073.6
XLM-R88.277.081.781.281.284.281.774.978.970.875.777.470.678.077.778.6
mPMR88.377.982.982.281.083.582.275.279.871.276.178.971.678.979.079.3
+ +Table 14: XNLI accuracy scores (Acc.) for each language. + +A For every submission: + +A1. Did you describe the limitations of your work? + +Section Limitations + +A2. Did you discuss any potential risks of your work? + +Not applicable. Left blank. + +A3. Do the abstract and introduction summarize the paper's main claims? + +Abstract, Section 1 + +A4. Have you used AI writing assistants when working on this paper? + +Left blank. + +B Did you use or create scientific artifacts? + +Section 3 + +B1. Did you cite the creators of artifacts you used? + +Section 3 + +B2. Did you discuss the license or terms for use and / or distribution of any artifacts? + +Not applicable. Left blank. + +B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? + +Section 3, Appendix A.1 + +B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? + +Not applicable. Left blank. + +B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? + +Appendix A.1, Appendix A.2 + +B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. + +Appendix A.1 + +C Did you run computational experiments? + +Section 4 + +C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? + +Section 3, Appendix A.1 + +The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance. + +C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Appendix A.1 +C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Section 4 +C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Section 3, Appendix A.1 + +D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank. + +D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response. +D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response. +D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response. +D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response. +D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response. \ No newline at end of file diff --git a/2023/mPMR_ A Multilingual Pre-trained Machine Reader at Scale/images.zip b/2023/mPMR_ A Multilingual Pre-trained Machine Reader at Scale/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..65dab33c6e2140b3abe41cbd1f17ca073eed4545 --- /dev/null +++ b/2023/mPMR_ A Multilingual Pre-trained Machine Reader at Scale/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:db008972f45b335e304a77a2ffe67f4460899a1112f24d2874b9facf5d08db41 +size 1144389 diff --git a/2023/mPMR_ A Multilingual Pre-trained Machine Reader at Scale/layout.json b/2023/mPMR_ A Multilingual Pre-trained Machine Reader at Scale/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..81e0609bdddb8ec7aef28f09b96c27a5d5c837b6 --- /dev/null +++ b/2023/mPMR_ A Multilingual Pre-trained Machine Reader at Scale/layout.json @@ -0,0 +1,7650 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 108, + 75, + 490, + 94 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 75, + 490, + 94 + ], + "spans": [ + { + "bbox": [ + 108, + 75, + 490, + 94 + ], + "type": "text", + "content": "mPMR: A Multilingual Pre-trained Machine Reader at Scale*" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 156, + 113, + 442, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 113, + 442, + 128 + ], + "spans": [ + { + "bbox": [ + 156, + 113, + 442, + 128 + ], + "type": "text", + "content": "Weiwen Xu" + }, + { + "bbox": [ + 156, + 113, + 442, + 128 + ], + "type": "inline_equation", + "content": "^{12,\\dagger}" + }, + { + "bbox": [ + 156, + 113, + 442, + 128 + ], + "type": "text", + "content": " Xin Li" + }, + { + "bbox": [ + 156, + 113, + 442, + 128 + ], + "type": "inline_equation", + "content": "^{2,\\ddagger}" + }, + { + "bbox": [ + 156, + 113, + 442, + 128 + ], + "type": "text", + "content": " Wai Lam" + }, + { + "bbox": [ + 156, + 113, + 442, + 128 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 156, + 113, + 442, + 128 + ], + "type": "text", + "content": " Lidong Bing" + }, + { + "bbox": [ + 156, + 113, + 442, + 128 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 203, + 129, + 393, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 129, + 393, + 143 + ], + "spans": [ + { + "bbox": [ + 203, + 129, + 393, + 143 + ], + "type": "text", + "content": "1The Chinese University of Hong Kong" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 213, + 143, + 382, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 143, + 382, + 157 + ], + "spans": [ + { + "bbox": [ + 213, + 143, + 382, + 157 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 213, + 143, + 382, + 157 + ], + "type": "text", + "content": "DAMO Academy, Alibaba Group" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 158, + 489, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 158, + 489, + 172 + ], + "spans": [ + { + "bbox": [ + 107, + 158, + 489, + 172 + ], + "type": "text", + "content": "{wxxu,wlam}@se.cuhk.edu.hk {xinting.lx,l.bing}@alibaba-inc.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 84, + 232, + 274, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 232, + 274, + 484 + ], + "spans": [ + { + "bbox": [ + 84, + 232, + 274, + 484 + ], + "type": "text", + "content": "We present multilingual Pre-trained Machine Reader (mPMR), a novel method for multilingual machine reading comprehension (MRC)-style pre-training. mPMR aims to guide multilingual pre-trained language models (mPLMs) to perform natural language understanding (NLU) including both sequence classification and span extraction in multiple languages. To achieve cross-lingual generalization when only source-language fine-tuning data is available, existing mPLMs solely transfer NLU capability from a source language to target languages. In contrast, mPMR allows the direct inheritance of multilingual NLU capability from the MRC-style pre-training to downstream tasks. Therefore, mPMR acquires better NLU capability for target languages. mPMR also provides a unified solver for tackling cross-lingual span extraction and sequence classification, thereby enabling the extraction of rationales to explain the sentence-pair classification process." + }, + { + "bbox": [ + 84, + 232, + 274, + 484 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 68, + 492, + 154, + 505 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 492, + 154, + 505 + ], + "spans": [ + { + "bbox": [ + 68, + 492, + 154, + 505 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 513, + 290, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 513, + 290, + 647 + ], + "spans": [ + { + "bbox": [ + 67, + 513, + 290, + 647 + ], + "type": "text", + "content": "Multilingual pre-trained language models, acronymed as mPLMs, have demonstrated strong Natural language understanding (NLU) capability in a wide range of languages (Xue et al., 2021; Cai et al., 2021, 2022; Conneau et al., 2020a; Ding et al., 2022; Li et al., 2020a). In particular, mPLMs can maintain exceptional cross-lingual language understanding (XLU) capability on unseen target languages though mPLMs are only fine-tuned on resource-rich source languages like English." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 649, + 290, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 649, + 290, + 675 + ], + "spans": [ + { + "bbox": [ + 67, + 649, + 290, + 675 + ], + "type": "text", + "content": "It has been proved that optimizing cross-lingual representations of mPLMs can improve XLU ca" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 308, + 212, + 520, + 380 + ], + "blocks": [ + { + "bbox": [ + 308, + 212, + 520, + 380 + ], + "lines": [ + { + "bbox": [ + 308, + 212, + 520, + 380 + ], + "spans": [ + { + "bbox": [ + 308, + 212, + 520, + 380 + ], + "type": "image", + "image_path": "56b276514635dad5b83103bb1b86a7c6db5594d38a67f412d88b687c12d37b95.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 314, + 388, + 512, + 401 + ], + "lines": [ + { + "bbox": [ + 314, + 388, + 512, + 401 + ], + "spans": [ + { + "bbox": [ + 314, + 388, + 512, + 401 + ], + "type": "text", + "content": "Figure 1: Pre-training and fine-tuning of mPMR." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 435, + 527, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 435, + 527, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 435, + 527, + 772 + ], + "type": "text", + "content": "pability. For example, cross-lingual supervisions, such as parallel sentences (Conneau and Lample, 2019) or bilingual dictionaries (Conneau et al., 2020b) could enhance cross-lingual representations with better language alignment. XLM-R (Conneau et al., 2020a) and mT5 (Xue et al., 2021) showed that appropriately incorporating more languages during pre-training leads to better cross-lingual representations. A few works enriched the cross-lingual representations with factual knowledge through the utilization of multilingual mentions of entities (Calixto et al., 2021; Ri et al., 2022) and relations (Liu et al., 2022; Jiang et al., 2022) annotated in knowledge graphs. Despite their differences, the above methods essentially constructed more diverse multilingual corpora for pre-training mPLMs. These mPLMs would presumably meet their saturation points and are known to suffer from curse of multilinguality (Conneau et al., 2020a; Pfeiffer et al., 2022; Berend, 2022). Under this situation, introducing more training data from either existing (Pfeiffer et al., 2022) or unseen (Conneau et al., 2020a) languages for enhancing mPLMs may not bring further improvement or even be detrimental to their cross-lingual representations." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 680, + 290, + 751 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 680, + 290, + 751 + ], + "spans": [ + { + "bbox": [ + 67, + 680, + 290, + 751 + ], + "type": "text", + "content": "* This work was supported by Alibaba Group through Alibaba Research Intern Program. The work described in this paper was also partially supported by a grant from the Research Grant Council of the Hong Kong Special Administrative Region, China (Project Code: 14200719).† This work was done when Weiwen Xu was an intern at Alibaba DAMO Academy.‡ Xin Li is the corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 67, + 751, + 290, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 751, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 751, + 290, + 772 + ], + "type": "text", + "content": "1The code, data, and checkpoints are released at https: //github.com/DAMO-NLP-SG/PMR" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "1533" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "spans": [ + { + "bbox": [ + 135, + 795, + 458, + 805 + ], + "type": "text", + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 219, + 806, + 375, + 817 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 806, + 375, + 817 + ], + "spans": [ + { + "bbox": [ + 219, + 806, + 375, + 817 + ], + "type": "text", + "content": "Volume 2: Short Papers, pages 1533-1546" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "spans": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "text", + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 71, + 292, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 71, + 292, + 395 + ], + "spans": [ + { + "bbox": [ + 69, + 71, + 292, + 395 + ], + "type": "text", + "content": "In the paper, instead of training a new mPLM with better cross-lingual representations, we propose multilingual Pre-trained Machine Reader (mPMR) to directly guide existing mPLMs to perform NLU in various languages. As shown in Figure 1, mPMR resembles PMR (Xu et al., 2022) for constructing multilingual machine reading comprehension (MRC)-style data with Wikipedia hyperlinks. These data are used to retrofit an mPLM into an mPMR through an MRC-style continual pre-training. During retrofitting process (i.e., pretraining), mPMR jointly learns the general sequence classification and span extraction capability for multiple languages. In XLU fine-tuning, mPLMs solely rely on cross-lingual representations to transfer NLU capability from a source language to target languages. By contrast, mPMR enables the direct inheritance of multilingual NLU capability from the MRC-style pre-training to downstream tasks in a unified MRC formulation, which alleviates the discrepancies between source-language fine-tuning and target-language inference (Zhou et al., 2022a,b, 2023). Therefore, mPMR shows greater potential in XLU than mPLMs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 396, + 291, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 396, + 291, + 491 + ], + "spans": [ + { + "bbox": [ + 67, + 396, + 291, + 491 + ], + "type": "text", + "content": "To improve the scalability of mPMR across multiple languages, we further propose Unified Q/C Construction and Stochastic answer position strategies for refining the curation of MRC data. With these two strategies, mPMR can better generalize to low-resource languages and becomes more robust to position bias (Ko et al., 2020)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 492, + 291, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 492, + 291, + 629 + ], + "spans": [ + { + "bbox": [ + 67, + 492, + 291, + 629 + ], + "type": "text", + "content": "The experimental results show that mPMR obtains clear improvements over XLM-R (Conneau et al., 2020a) on span extraction, with an average improvement of up to 12.6 F1 on TyDiQA, and 8.7 F1 on WikiAnn respectively. The analysis reveals that mPMR benefits from more multilingual MRC data for pre-training. We also found that mPMR converges faster in downstream tasks and is capable of using its strong extraction capability for explaining the sequence classification process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 640, + 127, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 640, + 127, + 653 + ], + "spans": [ + { + "bbox": [ + 67, + 640, + 127, + 653 + ], + "type": "text", + "content": "2 mPMR" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 662, + 292, + 718 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 662, + 292, + 718 + ], + "spans": [ + { + "bbox": [ + 67, + 662, + 292, + 718 + ], + "type": "text", + "content": "We present the MRC model and training data of mPMR. We closely follow PMR (Xu et al., 2022) and introduce the modifications for enabling multilingual MRC-style pre-training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 727, + 187, + 741 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 727, + 187, + 741 + ], + "spans": [ + { + "bbox": [ + 67, + 727, + 187, + 741 + ], + "type": "text", + "content": "2.1 Model Pre-training" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "type": "text", + "content": "Our mPMR follows the same MRC architecture of Xu et al. (2022, 2023) with an encoder and an" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 71, + 526, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 179 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 179 + ], + "type": "text", + "content": "extractor. The encoder maps input tokens " + }, + { + "bbox": [ + 302, + 71, + 526, + 179 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 302, + 71, + 526, + 179 + ], + "type": "text", + "content": ", the concatenation of the query " + }, + { + "bbox": [ + 302, + 71, + 526, + 179 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 302, + 71, + 526, + 179 + ], + "type": "text", + "content": ", the context " + }, + { + "bbox": [ + 302, + 71, + 526, + 179 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 302, + 71, + 526, + 179 + ], + "type": "text", + "content": ", and special markers (i.e., [CLS] and [SEP]), into hidden representations " + }, + { + "bbox": [ + 302, + 71, + 526, + 179 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 302, + 71, + 526, + 179 + ], + "type": "text", + "content": ". For any two tokens " + }, + { + "bbox": [ + 302, + 71, + 526, + 179 + ], + "type": "inline_equation", + "content": "X_{i}" + }, + { + "bbox": [ + 302, + 71, + 526, + 179 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 302, + 71, + 526, + 179 + ], + "type": "inline_equation", + "content": "X_{j}" + }, + { + "bbox": [ + 302, + 71, + 526, + 179 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 302, + 71, + 526, + 179 + ], + "type": "inline_equation", + "content": "i < j" + }, + { + "bbox": [ + 302, + 71, + 526, + 179 + ], + "type": "text", + "content": "), the extractor receives their contextualized representations " + }, + { + "bbox": [ + 302, + 71, + 526, + 179 + ], + "type": "inline_equation", + "content": "H_{i}" + }, + { + "bbox": [ + 302, + 71, + 526, + 179 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 302, + 71, + 526, + 179 + ], + "type": "inline_equation", + "content": "H_{j}" + }, + { + "bbox": [ + 302, + 71, + 526, + 179 + ], + "type": "text", + "content": " and predicts the probability score " + }, + { + "bbox": [ + 302, + 71, + 526, + 179 + ], + "type": "inline_equation", + "content": "S_{i,j}" + }, + { + "bbox": [ + 302, + 71, + 526, + 179 + ], + "type": "text", + "content": " indicating the probability of the token span " + }, + { + "bbox": [ + 302, + 71, + 526, + 179 + ], + "type": "inline_equation", + "content": "X_{i:j}" + }, + { + "bbox": [ + 302, + 71, + 526, + 179 + ], + "type": "text", + "content": " being the answer to the query " + }, + { + "bbox": [ + 302, + 71, + 526, + 179 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 302, + 71, + 526, + 179 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 180, + 526, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 180, + 526, + 288 + ], + "spans": [ + { + "bbox": [ + 302, + 180, + 526, + 288 + ], + "type": "text", + "content": "mPMR is guided with the Wiki Anchor Extraction (WAE) objective to train both the encoder and the extractor. WAE checks if the answer to the query exists in the context. If so, WAE would first regard the query and the context to be relevant and extracts the [CLS] token as a sequence-level relevance indicator. WAE would then extract all corresponding answers from the context." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 297, + 443, + 311 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 297, + 443, + 311 + ], + "spans": [ + { + "bbox": [ + 302, + 297, + 443, + 311 + ], + "type": "text", + "content": "2.2 Multilingual MRC Data" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 315, + 525, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 315, + 525, + 396 + ], + "spans": [ + { + "bbox": [ + 302, + 315, + 525, + 396 + ], + "type": "text", + "content": "Training mPMR requires the existence of labeled (query, context, answer) triplets. To obtain such data, we collected Wikipedia articles with anchor annotations for 24 languages, which are the most widely used and cover a reasonable number of languages used in XLU tasks (Ri et al., 2022)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 397, + 525, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 397, + 525, + 586 + ], + "spans": [ + { + "bbox": [ + 302, + 397, + 525, + 586 + ], + "type": "text", + "content": "As shown in Figure 1, we utilized a Wikipedia anchor to obtain a pair of correlated articles. One side of the pair is the article that provides in-depth descriptions of the anchor entity, which we defined as the definition article. The other side of the pair is named as the mention article, which mentions the specific anchor text2. We composed an answerable MRC example in which the anchor is the answer, the surrounding text of the anchor in the mention article is the context, and the definition of the anchor entity in the definition article is the query. Additionally, we can generate an unanswerable MRC example by pairing a query with an irrelevant context without anchor association." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 594, + 526, + 743 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 594, + 526, + 743 + ], + "spans": [ + { + "bbox": [ + 302, + 594, + 526, + 743 + ], + "type": "text", + "content": "Unified Q/C Construction. PMR constructed the MRC query and context as valid sentences so as to keep the text coherent. However, sentence segmentation tools are usually not available for low-resource languages. To remedy this, we did not apply sentence segmentation but only preprocess Wikipedia articles with word tokenization in mPMR. For each anchor, the MRC query comprises the first " + }, + { + "bbox": [ + 302, + 594, + 526, + 743 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 302, + 594, + 526, + 743 + ], + "type": "text", + "content": " words in the definition article. To prevent information leakage during pre-training, similar to PMR, we anonymized the anchor entity" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 750, + 525, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 750, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 525, + 772 + ], + "type": "text", + "content": "2definition/mention article refers to home/reference article of Xu et al. (2022)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1534" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 68, + 518, + 249 + ], + "blocks": [ + { + "bbox": [ + 75, + 68, + 518, + 249 + ], + "lines": [ + { + "bbox": [ + 75, + 68, + 518, + 249 + ], + "spans": [ + { + "bbox": [ + 75, + 68, + 518, + 249 + ], + "type": "table", + "html": "
Model#ParamsEQANERABSASentence PairAvg.
XQuADMLQATyDiQAWikiAnnCoNLLSemEval16PAWS-XXNLI
MetricsF1 / EMF1 / EMF1 / EMF1F1F1Acc.Acc.
XLM-R550M76.6 / 60.871.6 / 53.265.1 / 45.065.482.066.9‡86.479.274.2
mT5580M67.0 / 49.064.6 / 45.057.2 / 41.255.771.0‡62.5‡86.475.467.5
VECO550M77.3 / 61.871.7 / 53.267.6 / 49.165.781.3‡63.0‡88.779.974.4
mLUKE-W561M79.6 / -72.7 / -65.2 / 48.5‡67.7‡83.061.2‡88.2‡79.4‡74.6
Wiki-CL550M72.1 / 56.970.8 / 50.573.2 / 57.364.7--88.479.2-
KMLM550M77.3 / 61.772.1 / 53.767.9 / 50.466.7‡83.266.1‡88.079.275.1
Our MRC Formulation
XLM-Rbase270M70.8 / 56.964.4 / 47.950.8 / 38.257.979.260.085.073.367.7
mPMRbase270M74.0 / 59.565.3 / 48.763.4 / 49.066.681.762.186.173.671.6
XLM-R550M77.1 / 61.371.5 / 53.967.4 / 51.663.681.466.186.978.674.1
mPMR550M79.2 / 64.473.1 / 55.474.7 / 58.370.784.168.288.079.377.2
", + "image_path": "a80c889344bf6a4cd7fabf2a39d922a5a6cbb745b7121e0224804c61cc784565.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 259, + 525, + 295 + ], + "lines": [ + { + "bbox": [ + 67, + 259, + 525, + 295 + ], + "spans": [ + { + "bbox": [ + 67, + 259, + 525, + 295 + ], + "type": "text", + "content": "Table 1: The results of all XLU tasks. We report the average results of all languages for each dataset. We also compute the overall average score among all datasets in the Avg. column. We reproduce the missing results with the " + }, + { + "bbox": [ + 67, + 259, + 525, + 295 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 67, + 259, + 525, + 295 + ], + "type": "text", + "content": " label. Some results of Wiki-CL are left blank because they do not release their model checkpoint." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 317, + 289, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 317, + 289, + 343 + ], + "spans": [ + { + "bbox": [ + 67, + 317, + 289, + 343 + ], + "type": "text", + "content": "in the query to the [MASK] token. The MRC context consists of " + }, + { + "bbox": [ + 67, + 317, + 289, + 343 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 67, + 317, + 289, + 343 + ], + "type": "text", + "content": " words surrounding the anchor." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 352, + 291, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 352, + 291, + 594 + ], + "spans": [ + { + "bbox": [ + 69, + 352, + 291, + 594 + ], + "type": "text", + "content": "Stochastic Answer Position. As mentioned by Ko et al. (2020), the model is prone to overfitting to the position shortcut if the answer in the context exhibits a fixed position pattern. In our case, suppose that the MRC context consists of " + }, + { + "bbox": [ + 69, + 352, + 291, + 594 + ], + "type": "inline_equation", + "content": "C / 2" + }, + { + "bbox": [ + 69, + 352, + 291, + 594 + ], + "type": "text", + "content": " words on both the left and right sides of the anchor, the model may learn the shortcut that the middle part of the context is likely to be the answer. To prevent such position bias, we propose a stochastic answer position method, which allows the answer to be presented in any position within the context. Specifically, given an anchor in a Wikipedia article, the context comprises " + }, + { + "bbox": [ + 69, + 352, + 291, + 594 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 69, + 352, + 291, + 594 + ], + "type": "text", + "content": " words preceding the anchor and the " + }, + { + "bbox": [ + 69, + 352, + 291, + 594 + ], + "type": "inline_equation", + "content": "C - \\xi" + }, + { + "bbox": [ + 69, + 352, + 291, + 594 + ], + "type": "text", + "content": " words following the anchor, where " + }, + { + "bbox": [ + 69, + 352, + 291, + 594 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 69, + 352, + 291, + 594 + ], + "type": "text", + "content": " is a random integer ranging from 0 to " + }, + { + "bbox": [ + 69, + 352, + 291, + 594 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 69, + 352, + 291, + 594 + ], + "type": "text", + "content": " and varies across different contexts. In accordance with PMR, we treated all text spans identical to the anchor in the current context as valid answers." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 607, + 191, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 607, + 191, + 620 + ], + "spans": [ + { + "bbox": [ + 67, + 607, + 191, + 620 + ], + "type": "text", + "content": "3 Experimental Setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 629, + 290, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 629, + 290, + 709 + ], + "spans": [ + { + "bbox": [ + 67, + 629, + 290, + 709 + ], + "type": "text", + "content": "Implementation Details. In mPMR, the encoder is loaded from XLM-R (Conneau et al., 2020a) and the extractor is randomly initialized. Both components are then continually pre-trained using the multilingual MRC data that we constructed. More hyper-parameters can be found in Appendix A.1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 719, + 290, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 719, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 719, + 290, + 772 + ], + "type": "text", + "content": "Downstream XLU Tasks. We evaluated mPMR on a series of span extraction tasks, including Extractive Question Answering (EQA), Named Entity Recognition (NER), and Aspect-Based Sentiment" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 317, + 526, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 317, + 526, + 519 + ], + "spans": [ + { + "bbox": [ + 302, + 317, + 526, + 519 + ], + "type": "text", + "content": "Analysis (ABSA). We also evaluated our mPMR on two sequence classification tasks. We followed Xu et al. (2022) to convert all tasks into MRC formulation to effectively leverage the knowledge that is acquired during MRC-style pre-training. For EQA, we used XQuAD (Artetxe et al., 2020), MLQA (Lewis et al., 2020), and TyDiQA (Clark et al., 2020). For NER, we used WikiAnn (Pan et al., 2017) and CoNLL (Tjong Kim Sang, 2002; Tjong Kim Sang and De Meulder, 2003). SemEval16 (Pontiki et al., 2016) was used for ABSA task. Regarding the sequence classification, we used XNLI (Conneau et al., 2018) and PAWS-X (Yang et al., 2019). Additional dataset information and concrete examples are provided in Appendix A.2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 526, + 525, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 526, + 525, + 675 + ], + "spans": [ + { + "bbox": [ + 302, + 526, + 525, + 675 + ], + "type": "text", + "content": "Baselines. We compared mPMR with recent methods on improving cross-lingual representations, including 1) models pre-trained on a large number of languages: XLM-R (Conneau et al., 2020a), mT5 (Xue et al., 2021), and VECO (Luo et al., 2021); 2) models that exploited multilingual entity information: Wiki-CL (Calixto et al., 2021), and mLUKE-W (Ri et al., 2022); and 3) Model that utilized multilingual relation information: KMLM (Liu et al., 2022). For a fair comparison, all models have approximately the same parameter size." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 685, + 432, + 698 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 685, + 432, + 698 + ], + "spans": [ + { + "bbox": [ + 302, + 685, + 432, + 698 + ], + "type": "text", + "content": "4 Results and Analyses" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 706, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 706, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 706, + 525, + 772 + ], + "type": "text", + "content": "XLU Performance. Table 1 shows the results on a variety of XLU tasks. mPMR outperforms all previous methods with an absolute improvement of 2.1 F1 over the best baseline (i.e. KMLM). mPMR shows greater improvements over previ" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1535" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 69, + 527, + 143 + ], + "blocks": [ + { + "bbox": [ + 69, + 69, + 527, + 143 + ], + "lines": [ + { + "bbox": [ + 69, + 69, + 527, + 143 + ], + "spans": [ + { + "bbox": [ + 69, + 69, + 527, + 143 + ], + "type": "table", + "html": "
IndexModel#LangPAWS-XXQuADWikiAnnAvg.
#1XLM-Rbase085.070.857.971.2
#2#1 + MRC data in English185.2 (0.2↑)71.0 (0.2↑)59.5 (1.6↑)71.9 (0.7↑)
#3#2 + Stochastic Answer Position185.5 (0.3↑)73.0 (2.0↑)60.0 (0.5↑)72.8 (0.9↑)
#4#3 + MRC data in more languages1085.9 (0.4↑)73.5 (0.5↑)64.7 (4.7↑)74.7 (1.9↑)
#5#4 + MRC data in even more languages (mPMRbase)2486.1 (0.2↑)74.0 (0.5↑)66.6 (1.9↑)75.6 (0.9↑)
", + "image_path": "148e11b05851dacace3c001a12604c9de3997c40138ff8dcee66c90068b62d97.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 69, + 186, + 525, + 365 + ], + "blocks": [ + { + "bbox": [ + 67, + 151, + 525, + 176 + ], + "lines": [ + { + "bbox": [ + 67, + 151, + 525, + 176 + ], + "spans": [ + { + "bbox": [ + 67, + 151, + 525, + 176 + ], + "type": "text", + "content": "Table 2: The process of retrofitting XLM-R into mPMR using multilingual MRC data (English→10 languages→24 languages) and our Stochastic Answer Position method. Each row accumulates modifications from all rows above." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 186, + 525, + 365 + ], + "lines": [ + { + "bbox": [ + 69, + 186, + 525, + 365 + ], + "spans": [ + { + "bbox": [ + 69, + 186, + 525, + 365 + ], + "type": "table", + "html": "
LabelSentence 1Sentence 2
EntailmentRami Nieminen ( born February 25 , 1966 ) is a Finnish footballer.Rami Nieminen ( born 25 February 1966 ) is a Finnish former footballer.
ContradictionIn 1938 he became the Government Anthropologist of the Egyptian-Anglo Sudan and conducted fieldwork with the Nuba.In 1938 he became the government anthropologist of the anglo-Egyptian Sudan and led fieldwork with the Nuba.
EntailmentStipsits 出生于科尔新堡，并在维也纳施塔莫斯多夫度过了他的童年。什蒂普西奇出生于德国科恩堡，在维也纳斯塔莫斯多夫度过了他的童年。
Contradiction纳舒厄白银骑士团队加入了夏季大学联盟，是本市的现役球队。Nashua Silver Knights 队是当前夏季联赛的一部分，也是该市的大学体育队。
Entailmentごれらの見方は、福音主義的、清教徒的、プロデ斯特兰トの動態が出現すると必に、しはしだは表明くださいます。ごれらの見解は多くの场合、新教徒、清教徒、福音主義者が出現する:NOか表示お願いいたします。
Contradiction1954年にスリーナムに戸った後、弁護士とでラマリポに定住したこと。1954年、バラマリポに戸ると、彼はスリーナムで弁護士とで定住,No理由。
", + "image_path": "9556b4fbac5433b04c4c78b23b9cd720fc6199f7aa29a32cc877240a09373244.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 372, + 525, + 397 + ], + "lines": [ + { + "bbox": [ + 67, + 372, + 525, + 397 + ], + "spans": [ + { + "bbox": [ + 67, + 372, + 525, + 397 + ], + "type": "text", + "content": "Table 3: Case study on PAWS-X. mPMR can extract rationales to explain the sequence-pair classification in multiple languages." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 417, + 291, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 417, + 291, + 635 + ], + "spans": [ + { + "bbox": [ + 67, + 417, + 291, + 635 + ], + "type": "text", + "content": "ous methods on span extraction tasks. In particular, mPMR achieves up to 7.3 and 7.1 F1 improvements over XLM-R on TyDiQA and WikiAnn respectively. Such significant improvements probably come from the following two facts: (1) WikiAnn comprises a larger number of target languages (i.e. 40). Therefore, existing methods may struggle to align these low-resource languages with English due to a lack of language-specific data. (2) TyDiQA is a more challenging cross-lingual EQA task with " + }, + { + "bbox": [ + 67, + 417, + 291, + 635 + ], + "type": "inline_equation", + "content": "2\\mathrm{x}" + }, + { + "bbox": [ + 67, + 417, + 291, + 635 + ], + "type": "text", + "content": " less lexical overlap between the query and the answer than MLQA and XQuAD (Hu et al., 2020). Our mPMR, which acquires target-language span extraction capability from both MRC-style pretraining and English-only QA fine-tuning, achieves larger performance gains on more challenging task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 651, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 651, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 651, + 291, + 773 + ], + "type": "text", + "content": "mPMR Pre-training. To reflect the impact of our MRC-style data and Stochastic Answer Position method on pre-training, we present a step-by-step analysis of the retrofitting process starting from XLM-R in Table 2. Our findings suggest that the significant improvements observed are largely due to the inclusion of multilingual MRC data. Introducing English MRC data (model #2) gives marginal improvements because model #2" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 310, + 416, + 518, + 520 + ], + "blocks": [ + { + "bbox": [ + 310, + 416, + 518, + 520 + ], + "lines": [ + { + "bbox": [ + 310, + 416, + 518, + 520 + ], + "spans": [ + { + "bbox": [ + 310, + 416, + 518, + 520 + ], + "type": "image", + "image_path": "fde14688ed094287da5e02c2f3852fb01624809fa4b9d24f6b0d5328fe8edd71.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 527, + 526, + 552 + ], + "lines": [ + { + "bbox": [ + 302, + 527, + 526, + 552 + ], + "spans": [ + { + "bbox": [ + 302, + 527, + 526, + 552 + ], + "type": "text", + "content": "Figure 2: Convergence speed (Test set F1 and the training loss) of " + }, + { + "bbox": [ + 302, + 527, + 526, + 552 + ], + "type": "inline_equation", + "content": "\\mathrm{mPMR_{base}}" + }, + { + "bbox": [ + 302, + 527, + 526, + 552 + ], + "type": "text", + "content": " and XLM-" + }, + { + "bbox": [ + 302, + 527, + 526, + 552 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{\\mathrm{base}}" + }, + { + "bbox": [ + 302, + 527, + 526, + 552 + ], + "type": "text", + "content": " on WikiAnn." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 574, + 526, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 574, + 526, + 709 + ], + "spans": [ + { + "bbox": [ + 302, + 574, + 526, + 709 + ], + "type": "text", + "content": "can only rely on cross-lingual representations to transfer the knowledge acquired during MRC-style pre-training. When using MRC data on more languages (model #4 and #5), we can observe significant improvements on XLU tasks. This can be attributed to the NLU capability directly inherited from MRC-style pre-training in target languages. Additionally, with our Stochastic Answer Position method (model #3), mPMR becomes more robust to position bias and thus improves XLU tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 719, + 526, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 719, + 526, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 719, + 526, + 773 + ], + "type": "text", + "content": "Explainable Sentence-pair Classification. Inspired by PMR (Xu et al., 2022), we investigated if the extraction capability of mPMR can be leveraged to explain sentence-pair classification. Note" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "text", + "content": "1536" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 68, + 287, + 178 + ], + "blocks": [ + { + "bbox": [ + 70, + 68, + 287, + 178 + ], + "lines": [ + { + "bbox": [ + 70, + 68, + 287, + 178 + ], + "spans": [ + { + "bbox": [ + 70, + 68, + 287, + 178 + ], + "type": "image", + "image_path": "088e81490d4be3fc5e3553fa86604e19088ecab031db99dbd74c610ab80eb61f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 185, + 291, + 210 + ], + "lines": [ + { + "bbox": [ + 67, + 185, + 291, + 210 + ], + "spans": [ + { + "bbox": [ + 67, + 185, + 291, + 210 + ], + "type": "text", + "content": "Figure 3: Convergence speed (Test set F1 and the training loss) of " + }, + { + "bbox": [ + 67, + 185, + 291, + 210 + ], + "type": "inline_equation", + "content": "\\mathrm{mPMR_{base}}" + }, + { + "bbox": [ + 67, + 185, + 291, + 210 + ], + "type": "text", + "content": " and XLM-" + }, + { + "bbox": [ + 67, + 185, + 291, + 210 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{\\mathrm{base}}" + }, + { + "bbox": [ + 67, + 185, + 291, + 210 + ], + "type": "text", + "content": " on XQuAD." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 245, + 291, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 245, + 291, + 462 + ], + "spans": [ + { + "bbox": [ + 67, + 245, + 291, + 462 + ], + "type": "text", + "content": "that sentence-pair classification focuses on the inference between the two sentences. If we construct the query with only the task label as PMR does, such query does not solely correspond to any meaningful span in the context, and thus is hard to guide the span extraction. Therefore, we leveraged another template \"[CLS] label Sen-1 [SEP] Sen-2 [SEP]\", where the two sentences are represented separately in the query and the context. In this template, we can extract the exact span from Sen-2 that leads to a contraction or entailment relation (i.e., the task label) with Sen-1. Specifically, we passed the sentence pair to the model twice, with each sentence of the pair being designated as the Sen-2 respectively, and extract the context span with the highest probability score from both sentences." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 470, + 291, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 470, + 291, + 592 + ], + "spans": [ + { + "bbox": [ + 67, + 470, + 291, + 592 + ], + "type": "text", + "content": "As shown in Table 3, the extracted spans are indeed important rationales that determine the relationship between two sentences. Such a finding confirms that the extraction capability of mPMR can be appropriately used for explaining the sentence-pair classification process. While the extraction capability may affect the learning of sequence classification during fine-tuning, resulting in a 0.4 Acc. decrease on XNLI." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 624, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 624, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 624, + 291, + 772 + ], + "type": "text", + "content": "mPMR Fine-tuning. We investigated the effects of mPMR on XLU fine-tuning. Figure 2 shows that mPMR converges faster than XLM-R on WikiAnn with an extremely low loss value even fine-tuned for 500 steps. In terms of test set performance, mPMR outperforms XLM-R comprehensively and exhibits greater stability. As a result, mPMR provides a better starting point for addressing XLU tasks compared to XLM-R. More examples from XQuAD and PAWS-X are provided in Figure 3 and 4." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 305, + 68, + 515, + 176 + ], + "blocks": [ + { + "bbox": [ + 305, + 68, + 515, + 176 + ], + "lines": [ + { + "bbox": [ + 305, + 68, + 515, + 176 + ], + "spans": [ + { + "bbox": [ + 305, + 68, + 515, + 176 + ], + "type": "image", + "image_path": "19c3ea4971b85aae02e844457d45899a7c9aa439a0f7c0812810227e18894eb3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 185, + 526, + 210 + ], + "lines": [ + { + "bbox": [ + 302, + 185, + 526, + 210 + ], + "spans": [ + { + "bbox": [ + 302, + 185, + 526, + 210 + ], + "type": "text", + "content": "Figure 4: Convergence speed (Test set F1 and the training loss) of " + }, + { + "bbox": [ + 302, + 185, + 526, + 210 + ], + "type": "inline_equation", + "content": "\\mathrm{mPMR_{base}}" + }, + { + "bbox": [ + 302, + 185, + 526, + 210 + ], + "type": "text", + "content": " and XLM-" + }, + { + "bbox": [ + 302, + 185, + 526, + 210 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{\\mathrm{base}}" + }, + { + "bbox": [ + 302, + 185, + 526, + 210 + ], + "type": "text", + "content": " on PAWS-X." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 303, + 232, + 387, + 244 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 232, + 387, + 244 + ], + "spans": [ + { + "bbox": [ + 303, + 232, + 387, + 244 + ], + "type": "text", + "content": "5 Conclusions" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 254, + 526, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 254, + 526, + 364 + ], + "spans": [ + { + "bbox": [ + 302, + 254, + 526, + 364 + ], + "type": "text", + "content": "This paper presents a novel multilingual MRC-style pre-training method, namely mPMR. mPMR provides a unified solver for cross-lingual span extraction and sequence classification and enables direct transfer of NLU capability from pre-training to downstream tasks. mPMR clearly improves the previous baselines and provides a possible solution to explain the sentence-pair classification process." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 303, + 375, + 367, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 375, + 367, + 389 + ], + "spans": [ + { + "bbox": [ + 303, + 375, + 367, + 389 + ], + "type": "text", + "content": "Limitations" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 398, + 525, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 398, + 525, + 425 + ], + "spans": [ + { + "bbox": [ + 302, + 398, + 525, + 425 + ], + "type": "text", + "content": "We identify the following two limitations of our work:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 316, + 439, + 527, + 749 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 316, + 439, + 526, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 439, + 526, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 439, + 526, + 602 + ], + "type": "text", + "content": "- Different from raw text, constructing MRC-style data from Wikipedia requires the existence of hyperlinks. This idea works well for resource-rich languages, such as English and Chinese. While such an idea is less effective for languages with few hyperlink annotations in Wikipedia because a small amount of MRC-style training data is difficult to guide the learning of NLU capability in those languages. A possible solution is to explore other data resources to automatically construct large-scale MRC data for pre-training." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 613, + 527, + 749 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 527, + 749 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 527, + 749 + ], + "type": "text", + "content": "- As observed in Table 1, the improvements of sequence classification tasks are less significant than those of span extraction tasks. We suggest that the existence of anchors is not a strong relevance indicator between our constructed query and context. Such a finding is also observed in Chang et al. (2020). Therefore, constructing more relevant query-context pairs for sequence classification pre-training can possibly remedy this issue." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 780, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 780, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 780, + 310, + 791 + ], + "type": "text", + "content": "1537" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 71, + 127, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 71, + 127, + 83 + ], + "spans": [ + { + "bbox": [ + 69, + 71, + 127, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 90, + 290, + 772 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 69, + 90, + 290, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 90, + 290, + 145 + ], + "spans": [ + { + "bbox": [ + 69, + 90, + 290, + 145 + ], + "type": "text", + "content": "Mikel Artetxe, Sebastian Ruder, and Dani Yogatama. 2020. On the cross-lingual transferability of monolingual representations. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 153, + 289, + 176 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 153, + 289, + 176 + ], + "spans": [ + { + "bbox": [ + 69, + 153, + 289, + 176 + ], + "type": "text", + "content": "Giuseppe Attardi. 2015. Wikiextractor. https://github.com/attardi/wikiextractor." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 185, + 289, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 185, + 289, + 251 + ], + "spans": [ + { + "bbox": [ + 69, + 185, + 289, + 251 + ], + "type": "text", + "content": "Gábor Berend. 2022. Combating the curse of multilinguality in cross-lingual WSD by aligning sparse contextualized word representations. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 259, + 289, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 259, + 289, + 314 + ], + "spans": [ + { + "bbox": [ + 69, + 259, + 289, + 314 + ], + "type": "text", + "content": "Deng Cai, Xin Li, Jackie Chun-Sing Ho, Lidong Bing, and Wai Lam. 2021. Multilingual AMR parsing with noisy knowledge distillation. In Findings of the Association for Computational Linguistics: EMNLP 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 323, + 289, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 323, + 289, + 379 + ], + "spans": [ + { + "bbox": [ + 69, + 323, + 289, + 379 + ], + "type": "text", + "content": "Deng Cai, Xin Li, Jackie Chun-Sing Ho, Lidong Bing, and Wai Lam. 2022. Retrofitting multilingual sentence embeddings with Abstract Meaning Representation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 387, + 289, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 387, + 289, + 464 + ], + "spans": [ + { + "bbox": [ + 69, + 387, + 289, + 464 + ], + "type": "text", + "content": "Iacer Calixto, Alessandro Raganato, and Tommaso Pasini. 2021. Wikipedia entities as rendezvous across languages: Grounding multilingual language models by predicting Wikipedia hyperlinks. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 473, + 289, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 473, + 289, + 517 + ], + "spans": [ + { + "bbox": [ + 69, + 473, + 289, + 517 + ], + "type": "text", + "content": "Wei-Cheng Chang, Felix X. Yu, Yin-Wen Chang, Yiming Yang, and Sanjiv Kumar. 2020. Pre-training tasks for embedding-based large-scale retrieval. In International Conference on Learning Representations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 525, + 289, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 525, + 289, + 592 + ], + "spans": [ + { + "bbox": [ + 69, + 525, + 289, + 592 + ], + "type": "text", + "content": "Jonathan H. Clark, Eunsol Choi, Michael Collins, Dan Garrette, Tom Kwiatkowski, Vitaly Nikolaev, and Jennimaria Palomaki. 2020. TyDi QA: A benchmark for information-seeking question answering in typologically diverse languages. Transactions of the Association for Computational Linguistics." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 600, + 289, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 600, + 289, + 677 + ], + "spans": [ + { + "bbox": [ + 69, + 600, + 289, + 677 + ], + "type": "text", + "content": "Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2020a. Unsupervised cross-lingual representation learning at scale. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 686, + 289, + 719 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 686, + 289, + 719 + ], + "spans": [ + { + "bbox": [ + 69, + 686, + 289, + 719 + ], + "type": "text", + "content": "Alexis Conneau and Guillaume Lample. 2019. Crosslingual language model pretraining. In Advances in Neural Information Processing Systems." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 69, + 728, + 289, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 728, + 289, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 728, + 289, + 772 + ], + "type": "text", + "content": "Alexis Conneau, Rudy Rinott, Guillaume Lample, Adina Williams, Samuel Bowman, Holger Schwenk, and Veselin Stoyanov. 2018. XNLI: Evaluating crosslingual sentence representations. In Proceedings of" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 525, + 772 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 315, + 72, + 525, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 525, + 95 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 525, + 95 + ], + "type": "text", + "content": "the 2018 Conference on Empirical Methods in Natural Language Processing." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 104, + 525, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 104, + 525, + 159 + ], + "spans": [ + { + "bbox": [ + 304, + 104, + 525, + 159 + ], + "type": "text", + "content": "Alexis Conneau, Shijie Wu, Haoran Li, Luke Zettlemoyer, and Veselin Stoyanov. 2020b. Emerging cross-lingual structure in pretrained language models. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 169, + 525, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 169, + 525, + 235 + ], + "spans": [ + { + "bbox": [ + 304, + 169, + 525, + 235 + ], + "type": "text", + "content": "Bosheng Ding, Junjie Hu, Lidong Bing, Mahani Aljunied, Shafiq Joty, Luo Si, and Chunyan Miao. 2022. GlobalWoZ: Globalizing MultiWoZ to develop multilingual task-oriented dialogue systems. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 244, + 525, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 244, + 525, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 244, + 525, + 300 + ], + "type": "text", + "content": "Junjie Hu, Sebastian Ruder, Aditya Siddhant, Graham Neubig, Orhan First, and Melvin Johnson. 2020. Xtreme: A massively multilingual multi-task benchmark for evaluating cross-lingual generalisation. In International Conference on Machine Learning." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 308, + 525, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 525, + 364 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 525, + 364 + ], + "type": "text", + "content": "Xiaoze Jiang, Yaobo Liang, Weizhu Chen, and Nan Duan. 2022. Xlm-k: Improving cross-lingual language model pre-training with multilingual knowledge. In Proceedings of the AAAI Conference on Artificial Intelligence." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 373, + 525, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 525, + 428 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 525, + 428 + ], + "type": "text", + "content": "Miyoung Ko, Jinhyuk Lee, Hyunjae Kim, Gangwoo Kim, and Jaewoo Kang. 2020. Look at the first sentence: Position bias in question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 438, + 525, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 525, + 493 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 525, + 493 + ], + "type": "text", + "content": "Patrick Lewis, Barlas Oguz, Rudy Rinott, Sebastian Riedel, and Holger Schwenk. 2020. MLQA: Evaluating cross-lingual extractive question answering. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 502, + 525, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 502, + 525, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 502, + 525, + 567 + ], + "type": "text", + "content": "Juntao Li, Ruidan He, Hai Ye, Hwee Tou Ng, Lidong Bing, and Rui Yan. 2020a. Unsupervised domain adaptation of a pretrained cross-lingual language model. In Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, IJCAI 2020." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 577, + 525, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 577, + 525, + 622 + ], + "spans": [ + { + "bbox": [ + 304, + 577, + 525, + 622 + ], + "type": "text", + "content": "Xin Li, Lidong Bing, Wenxuan Zhang, Zheng Li, and Wai Lam. 2020b. Unsupervised cross-lingual adaptation for sequence tagging and beyond. arXiv preprint arXiv:2010.12405." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 631, + 525, + 687 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 631, + 525, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 631, + 525, + 687 + ], + "type": "text", + "content": "Linlin Liu, Xin Li, Ruidan He, Lidong Bing, Shafiq Joty, and Luo Si. 2022. Enhancing multilingual language model with massive multilingual knowledge triples. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 696, + 525, + 729 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 696, + 525, + 729 + ], + "spans": [ + { + "bbox": [ + 304, + 696, + 525, + 729 + ], + "type": "text", + "content": "Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In International Conference on Learning Representations." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 738, + 525, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 738, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 738, + 525, + 772 + ], + "type": "text", + "content": "Fuli Luo, Wei Wang, Jiahao Liu, Yijia Liu, Bin Bi, Songfang Huang, Fei Huang, and Luo Si. 2021. VECO: Variable and flexible cross-lingual pre-training for" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1538" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 290, + 772 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 80, + 72, + 290, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 72, + 290, + 128 + ], + "spans": [ + { + "bbox": [ + 80, + 72, + 290, + 128 + ], + "type": "text", + "content": "language understanding and generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 137, + 290, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 137, + 290, + 202 + ], + "spans": [ + { + "bbox": [ + 69, + 137, + 290, + 202 + ], + "type": "text", + "content": "Xiaoman Pan, Boliang Zhang, Jonathan May, Joel Nothman, Kevin Knight, and Heng Ji. 2017. Cross-lingual name tagging and linking for 282 languages. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 211, + 290, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 211, + 290, + 288 + ], + "spans": [ + { + "bbox": [ + 69, + 211, + 290, + 288 + ], + "type": "text", + "content": "Jonas Pfeiffer, Naman Goyal, Xi Lin, Xian Li, James Cross, Sebastian Riedel, and Mikel Artetxe. 2022. Lifting the curse of multilinguality by pre-training modular transformers. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 297, + 290, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 297, + 290, + 407 + ], + "spans": [ + { + "bbox": [ + 69, + 297, + 290, + 407 + ], + "type": "text", + "content": "Maria Pontiki, Dimitris Galanis, Haris Papageorgiou, Ion Androutsopoulos, Suresh Manandhar, Mohammad AL-Smadi, Mahmoud Al-Ayyoub, Yanyan Zhao, Bing Qin, Orphée De Clercq, Véronique Hoste, Marianna Apidianaki, Xavier Tannier, Natalia Loukachevitch, Evgeniy Kotelnikov, Nuria Bel, Salud María Jiménez-Zafra, and Gülşen Eryigit. 2016. SemEval-2016 task 5: Aspect based sentiment analysis. In Proceedings of the 10th International Workshop on Semantic Evaluation (SemEval-2016)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 417, + 290, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 417, + 290, + 482 + ], + "spans": [ + { + "bbox": [ + 69, + 417, + 290, + 482 + ], + "type": "text", + "content": "Ryokan Ri, Ikuya Yamada, and Yoshimasa Tsuruoka. 2022. mLUKE: The power of entity representations in multilingual pretrained language models. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 491, + 290, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 491, + 290, + 546 + ], + "spans": [ + { + "bbox": [ + 69, + 491, + 290, + 546 + ], + "type": "text", + "content": "Erik F. Tjong Kim Sang. 2002. Introduction to the CoNLL-2002 shared task: Language-independent named entity recognition. In COLING-02: The 6th Conference on Natural Language Learning 2002 (CoNLL-2002)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 555, + 290, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 555, + 290, + 611 + ], + "spans": [ + { + "bbox": [ + 69, + 555, + 290, + 611 + ], + "type": "text", + "content": "Erik F. Tjong Kim Sang and Fien De Meulder. 2003. Introduction to the CoNLL-2003 shared task: Language-independent named entity recognition. In Proceedings of the Seventh Conference on Natural Language Learning at HLT-NAACL 2003." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 619, + 290, + 740 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 619, + 290, + 740 + ], + "spans": [ + { + "bbox": [ + 69, + 619, + 290, + 740 + ], + "type": "text", + "content": "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 750, + 290, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 750, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 750, + 290, + 772 + ], + "type": "text", + "content": "Weiwen Xu, Xin Li, Yang Deng, Wai Lam, and Lidong Bing. 2023. Peerda: Data augmentation via modeling" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 525, + 585 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 314, + 72, + 524, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 524, + 105 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 524, + 105 + ], + "type": "text", + "content": "peer relation for span identification tasks. In The 61th Annual Meeting of the Association for Computational Linguistics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 114, + 525, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 114, + 525, + 169 + ], + "spans": [ + { + "bbox": [ + 304, + 114, + 525, + 169 + ], + "type": "text", + "content": "Weiwen Xu, Xin Li, Wenxuan Zhang, Meng Zhou, Lidong Bing, Wai Lam, and Luo Si. 2022. From clozing to comprehending: Retrofitting pre-trained language model to pre-trained machine reader. arXiv preprint arXiv:2212.04755." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 178, + 525, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 178, + 525, + 255 + ], + "spans": [ + { + "bbox": [ + 304, + 178, + 525, + 255 + ], + "type": "text", + "content": "Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. 2021. mT5: A massively multilingual pre-trained text-to-text transformer. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 263, + 525, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 263, + 525, + 340 + ], + "spans": [ + { + "bbox": [ + 304, + 263, + 525, + 340 + ], + "type": "text", + "content": "Yinfei Yang, Yuan Zhang, Chris Tar, and Jason Baldridge. 2019. PAWS-X: A cross-lingual adversarial dataset for paraphrase identification. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 349, + 525, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 349, + 525, + 405 + ], + "spans": [ + { + "bbox": [ + 304, + 349, + 525, + 405 + ], + "type": "text", + "content": "Wenxuan Zhang, Ruidan He, Haiyun Peng, Lidong Bing, and Wai Lam. 2021. Cross-lingual aspect-based sentiment analysis with aspect term code-switching. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 412, + 525, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 412, + 525, + 446 + ], + "spans": [ + { + "bbox": [ + 304, + 412, + 525, + 446 + ], + "type": "text", + "content": "Meng Zhou, Xin Li, Yue Jiang, and Lidong Bing. 2022a. Enhancing cross-lingual prompting with mask token augmentation. arXiv preprint arXiv:2202.07255." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 454, + 525, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 454, + 525, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 454, + 525, + 520 + ], + "type": "text", + "content": "Ran Zhou, Xin Li, Lidong Bing, Erik Cambria, and Chunyan Miao. 2023. Improving self-training for cross-lingual named entity recognition with contrastive and prototype learning. In *The 61th Annual Meeting of the Association for Computational Linguistics*." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 528, + 525, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 528, + 525, + 585 + ], + "spans": [ + { + "bbox": [ + 304, + 528, + 525, + 585 + ], + "type": "text", + "content": "Ran Zhou, Xin Li, Lidong Bing, Erik Cambria, Luo Si, and Chunyan Miao. 2022b. ConNER: Consistency training for cross-lingual named entity recognition. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1539" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 71, + 142, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 71, + 142, + 84 + ], + "spans": [ + { + "bbox": [ + 68, + 71, + 142, + 84 + ], + "type": "text", + "content": "A Appendix" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 93, + 235, + 105 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 93, + 235, + 105 + ], + "spans": [ + { + "bbox": [ + 68, + 93, + 235, + 105 + ], + "type": "text", + "content": "A.1 More Implementation Details" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 110, + 291, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 110, + 291, + 299 + ], + "spans": [ + { + "bbox": [ + 67, + 110, + 291, + 299 + ], + "type": "text", + "content": "We collect the 2022-08-01 dump3 of Wikipedia articles for the 24 languages in consideration. The statistics of each language can be found in Table 4. Then for each article, we extract the plain text with anchors via WikiExtractor (Attardi, 2015). Word tokenization is performed using spaCy4 if the language is supported, otherwise, we utilize PyThaiNLP5 for Thai and Sacremoses6 for remaining languages. For each anchor entity, we construct 10 answerable MRC examples and 10 unanswerable MRC examples as described in Sec. 2.2. Anchor entities with low frequency (below 10 occurrences for English entities and 5 occurrences for entities in other languages) were excluded." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 301, + 291, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 301, + 291, + 570 + ], + "spans": [ + { + "bbox": [ + 69, + 301, + 291, + 570 + ], + "type": "text", + "content": "In mPMR, we use Huggingface's implementations of XLM-R (Wolf et al., 2020). During the pre-training stage, the query length " + }, + { + "bbox": [ + 69, + 301, + 291, + 570 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 69, + 301, + 291, + 570 + ], + "type": "text", + "content": " is set to 50 words, and the context length " + }, + { + "bbox": [ + 69, + 301, + 291, + 570 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 69, + 301, + 291, + 570 + ], + "type": "text", + "content": " is set to 200 words. Both are computed before the subword segmentation. We follow the default learning rate schedule and dropout settings used in XLM-R. We use AdamW (Loshchilov and Hutter, 2019) as our optimizer. We train both " + }, + { + "bbox": [ + 69, + 301, + 291, + 570 + ], + "type": "inline_equation", + "content": "\\mathrm{mPMR_{base}}" + }, + { + "bbox": [ + 69, + 301, + 291, + 570 + ], + "type": "text", + "content": " and mPMR on 4 A100 GPU. The learning rate is set to 1e-5, and the effective batch size for each step is set to 256 and 80 for " + }, + { + "bbox": [ + 69, + 301, + 291, + 570 + ], + "type": "inline_equation", + "content": "\\mathrm{mPMR_{base}}" + }, + { + "bbox": [ + 69, + 301, + 291, + 570 + ], + "type": "text", + "content": " and mPMR respectively in order to maximize the usage of the GPU memory. We use the average scores of XQuAD, CoNLL, and PAWS-X to select the best mPMR checkpoint. In fact, we continually pre-train " + }, + { + "bbox": [ + 69, + 301, + 291, + 570 + ], + "type": "inline_equation", + "content": "\\mathrm{mPMR_{base}}" + }, + { + "bbox": [ + 69, + 301, + 291, + 570 + ], + "type": "text", + "content": " and mPMR for 250,000 and 100,000 steps. The training speed is around 6250 steps per hour. The hyper-parameters of " + }, + { + "bbox": [ + 69, + 301, + 291, + 570 + ], + "type": "inline_equation", + "content": "\\mathrm{mPMR_{large}}" + }, + { + "bbox": [ + 69, + 301, + 291, + 570 + ], + "type": "text", + "content": " on downstream XLU tasks can be found in Table 5." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 581, + 212, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 581, + 212, + 592 + ], + "spans": [ + { + "bbox": [ + 68, + 581, + 212, + 592 + ], + "type": "text", + "content": "A.2 Downstream XLU Tasks" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 598, + 291, + 720 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 598, + 291, + 720 + ], + "spans": [ + { + "bbox": [ + 67, + 598, + 291, + 720 + ], + "type": "text", + "content": "We evaluate mPMR on XLU tasks including both span extraction (EQA, NER, and ABSA) and sequence classification (sentence pair classification). We follow (Xu et al., 2022) to convert all tasks into MRC formulation and tackle them accordingly. We show concrete examples for each task in Table 6. Specifically, we evaluate the performance of EQA on three benchmarks: XQuAD (Artetxe et al., 2020), MLQA (Lewis et al., 2020), and Ty" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 71, + 526, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 248 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 248 + ], + "type": "text", + "content": "DiQA (Clark et al., 2020) covering 11, 7, and 9 languages respectively. For NER evaluation, we use the WikiAnn dataset (Pan et al., 2017) restricted to the 40 languages from XTREME (Hu et al., 2020), as well as the CoNLL dataset with 4 languages (Tjong Kim Sang, 2002; Tjong Kim Sang and De Meulder, 2003); We also evaluate the XLU performance of SemEval16 ABSA on 6 languages (Pontiki et al., 2016), where we collect the data from Li et al. (2020b); Zhang et al. (2021). Regarding the sequence classification task, we evaluate XNLI (Conneau et al., 2018) and PAWS-X (Yang et al., 2019) with 15 and 7 languages respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 256, + 497, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 256, + 497, + 269 + ], + "spans": [ + { + "bbox": [ + 302, + 256, + 497, + 269 + ], + "type": "text", + "content": "A.3 mPMR Performance per Language" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 274, + 526, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 274, + 526, + 340 + ], + "spans": [ + { + "bbox": [ + 302, + 274, + 526, + 340 + ], + "type": "text", + "content": "We show the detailed results for each language in each task in Table 7 (XQuAD), Table 8 (MLQA), Table 9 (TyDiQA), Table 10 (WikiAnn), Table 11 (CoNLL), Table 12 (SemEval16), Table 13 (PAWS-X), and Table 14 (XNLI)." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 80, + 728, + 238, + 740 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 728, + 238, + 740 + ], + "spans": [ + { + "bbox": [ + 80, + 728, + 238, + 740 + ], + "type": "text", + "content": "3https://dumps.wikimedia.org/enwiki/latest" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 81, + 739, + 214, + 751 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 739, + 214, + 751 + ], + "spans": [ + { + "bbox": [ + 81, + 739, + 214, + 751 + ], + "type": "text", + "content": "4https://github.com/explosion/spaCy" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 81, + 750, + 233, + 761 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 750, + 233, + 761 + ], + "spans": [ + { + "bbox": [ + 81, + 750, + 233, + 761 + ], + "type": "text", + "content": "5https://github.com/PyThaiNLP/pythainlp" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 81, + 761, + 230, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 761, + 230, + 772 + ], + "spans": [ + { + "bbox": [ + 81, + 761, + 230, + 772 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 81, + 761, + 230, + 772 + ], + "type": "text", + "content": "https://github.com/ Alvations/sacremoses" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "1540" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 97, + 157, + 497, + 368 + ], + "blocks": [ + { + "bbox": [ + 97, + 157, + 497, + 368 + ], + "lines": [ + { + "bbox": [ + 97, + 157, + 497, + 368 + ], + "spans": [ + { + "bbox": [ + 97, + 157, + 497, + 368 + ], + "type": "table", + "html": "
Language# Entities# MRC examplesLanguage# Entities# MRC examples
ar118,2922,020,502ko94,6161,597,076
bn25,081410,634nl251,3234,185,913
de864,74614,795,826pl283,9254,765,015
el56,383946,114pt216,6953,648,603
en966,19719,303,940ru432,4377,342,472
es412,4767,044,972sv169,0302,808,214
fi113,1181,960,636sw4,85765,724
fr595,87910,164,216te11,005170,664
hi15,350242,078th31,676522,434
id70,9601,164,662tr71,2941,175,276
it376,4176,421,850vi68,6651,147,772
ja423,8847,338,308zh259,7854,438,004
Total5,934,091103,680,905
", + "image_path": "661dbfbb6e843fe62e3662662a55c377abc972b6a939cb660e1f8aac153b5086.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 106, + 585, + 487, + 661 + ], + "blocks": [ + { + "bbox": [ + 67, + 375, + 525, + 400 + ], + "lines": [ + { + "bbox": [ + 67, + 375, + 525, + 400 + ], + "spans": [ + { + "bbox": [ + 67, + 375, + 525, + 400 + ], + "type": "text", + "content": "Table 4: Data statistics of mPMR pre-training data. The statistics is computed after removing the low-frequency entities. The number of MRC examples includes both answerable and unanswerable examples." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 106, + 585, + 487, + 661 + ], + "lines": [ + { + "bbox": [ + 106, + 585, + 487, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 585, + 487, + 661 + ], + "type": "table", + "html": "
DatasetXQuADMLQATyDiQAWikiAnnCoNLLSemEval16PAWS-XXNLI
Query Length6464643232326464
Input Length384384384192192192192192
Batch Size8881616321632
Learning Rate3e-53e-52e-51e-51e-52e-55e-53e-5
Epoch3310101020103
", + "image_path": "e95d9666cd872126df72706db3a769093eb6d05345671c9cb46bb11694796c53.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 172, + 667, + 420, + 681 + ], + "lines": [ + { + "bbox": [ + 172, + 667, + 420, + 681 + ], + "spans": [ + { + "bbox": [ + 172, + 667, + 420, + 681 + ], + "type": "text", + "content": "Table 5: Hyper-parameters settings in fine-tuning XLU tasks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1541" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 72, + 74, + 522, + 644 + ], + "blocks": [ + { + "bbox": [ + 72, + 74, + 522, + 644 + ], + "lines": [ + { + "bbox": [ + 72, + 74, + 522, + 644 + ], + "spans": [ + { + "bbox": [ + 72, + 74, + 522, + 644 + ], + "type": "table", + "html": "
TaskExample InputExample Output
EQA(XSQuAD)Ori.Question: Who lost to the Broncos in the divisional round?Context: The Broncos defeated the Pittsburgh Steelers in the divi-sional round, 23–16, by scoring 11 points in the final three minutes of the game.Answer: "Pittsburgh Steelers"
PMR[CLS] Who lost to the Broncos in the divisional round ? [SEP] [SEP]The Broncos defeated the Pittsburgh Steelers in the divisional round, 23–16 , by scoring 11 points in the final three minutes of the game .[SEP](17,18) - "Pittsburgh Steelers"
NER(CoNLL)Ori.Two goals in the last six minutes gave holders Japan an uninspiring 2-1 Asian Cup victory over Syria on Friday.("Japan", LOC);("Syria", LOC);("Asian Cup", MISC)
PMR[CLS] "ORG". Organization entities are limited to named corporate,governmental, or other organizational entities. [SEP] [SEP] Two goals in the last six minutes gave holders Japan an uninspiring 2-1 Asian Cup victory over Syria on Friday . [SEP]∅
[CLS] "PER". Person entities are named persons or family . [SEP] [SEP] Two goals in the last six minutes gave holders Japan an uninspiring 2-1 Asian Cup victory over Syria on Friday . [SEP]∅
[CLS] "LOC". Location entities are the name of politically or geo-graphically defined locations such as cities , countries . [SEP] [SEP] Two goals in the last six minutes gave holders Japan an uninspiring 2-1 Asian Cup victory over Syria on Friday . [SEP](32,32) - "Japan";(40,40) - "Syria"
[CLS] "MISC". Examples of miscellaneous entities include events ,nationalities , products and works of art . [SEP] [SEP] Two goals in the last six minutes gave holders Japan an uninspiring 2-1 Asian Cup victory over Syria on Friday . [SEP](34,35) - "Asian Cup"
ABSA(SemEval16)Ori.Nice ambience, but highly overrated place.("ambience", POS);("place", NEG)
PMR[CLS] "POS". For aspect terms of positive sentiment . [SEP] [SEP] Nice ambience , but highly overrated place . [SEP](13,13) - "ambience"
[CLS] "NEG". For aspect terms of negative sentiment . [SEP] [SEP] Nice ambience , but highly overrated place . [SEP](18,18) - "place"
[CLS] "NEU". For aspect terms of neutral sentiment . [SEP] [SEP] Nice ambience , but highly overrated place . [SEP]∅
Sen. Pair Classification(PAWS-X)Ori.Hypothesis: The Tabaci River is a tributary of the River Leurda in Romania.Premise: The Leurda River is a tributary of the River Tabaci in Romania.Contradiction
PMR[CLS] Contradiction . The hypothesis is a sentence with a contradic-tory meaning to the premise . [SEP] [SEP] Hypothesis : The Tabaci River is a tributary of the River Leurda in Romania . Premise : The Leurda River is a tributary of the River Tabaci in Romania . [SEP](0,0) - "[CLS]"
[CLS] Entailment . The hypothesis is a sentence with a similar meaning as the premise . [SEP] [SEP] Hypothesis : The Tabaci River is a tributary of the River Leurda in Romania . Premise : The Leurda River is a tributary of the River Tabaci in Romania . [SEP]∅
", + "image_path": "529d9483dbe31ace10a0db411bc0a58a7b4d5ae548b131db3858aa13e8bbc42e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 71, + 694, + 525, + 743 + ], + "blocks": [ + { + "bbox": [ + 67, + 650, + 525, + 676 + ], + "lines": [ + { + "bbox": [ + 67, + 650, + 525, + 676 + ], + "spans": [ + { + "bbox": [ + 67, + 650, + 525, + 676 + ], + "type": "text", + "content": "Table 6: MRC examples of XLU tasks. We use English examples here for demonstration purposes. Ori. indicates the original data format of these tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 71, + 694, + 525, + 743 + ], + "lines": [ + { + "bbox": [ + 71, + 694, + 525, + 743 + ], + "spans": [ + { + "bbox": [ + 71, + 694, + 525, + 743 + ], + "type": "table", + "html": "
ModelenardeeleshiruthtrvizhAvg.
XLM-Rbase82.2 / 72.065.5 / 49.973.9 / 59.771.2 / 56.376.3 / 59.466.4 / 52.073.7 / 58.964.7 / 54.667.0 / 52.873.3 / 54.765.0 / 55.970.8 / 56.9
mPMRbase84.4 / 73.469.6 / 53.276.4 / 61.574.9 / 58.477.4 / 60.269.2 / 54.575.2 / 58.869.2 / 57.670.4 / 55.874.8 / 55.871.8 / 65.574.0 / 59.5
XLM-R86.5 / 75.672.4 / 54.879.3 / 63.079.2 / 61.682.0 / 62.976.1 / 59.179.0 / 62.972.2 / 59.875.4 / 60.879.7 / 60.868.2 / 58.277.3 / 61.7
mPMR87.6 / 76.575.9 / 60.081.5 / 65.080.8 / 63.982.8 / 65.176.5 / 60.380.9 / 65.375.5 / 65.576.7 / 61.381.5 / 62.271.5 / 63.479.2 / 64.4
", + "image_path": "5553417b6d76c3c93b8475680099ce0d91d4dbfe3af4241c388707bac41aee2e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 187, + 751, + 404, + 764 + ], + "lines": [ + { + "bbox": [ + 187, + 751, + 404, + 764 + ], + "spans": [ + { + "bbox": [ + 187, + 751, + 404, + 764 + ], + "type": "text", + "content": "Table 7: XQuAD results (F1 / EM) for each language." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "text", + "content": "1542" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 71, + 74, + 523, + 143 + ], + "blocks": [ + { + "bbox": [ + 71, + 74, + 523, + 143 + ], + "lines": [ + { + "bbox": [ + 71, + 74, + 523, + 143 + ], + "spans": [ + { + "bbox": [ + 71, + 74, + 523, + 143 + ], + "type": "table", + "html": "
ModelenardeeshivizhAvg.
XLM-Rbase79.3 / 67.255.4 / 38.162.0 / 49.166.8 / 50.259.4 / 44.866.1 / 46.761.8 / 39.564.4 / 47.9
mPMRbase81.1 / 68.958.5 / 41.063.6 / 50.568.5 / 52.160.3 / 46.468.3 / 49.256.6 / 32.965.3 / 48.7
XLM-R83.4 / 71.064.9 / 45.869.6 / 54.874.1 / 56.870.7 / 53.473.3 / 53.064.4 / 42.471.5 / 53.9
mPMR84.0 / 71.466.4 / 47.070.3 / 56.274.5 / 57.171.4 / 54.174.7 / 54.470.5 / 47.373.1 / 55.4
", + "image_path": "8e6f20f0077c1a8a39e3ffdb3cfe824932d9d769058ee6587eff3dec4c024ed8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 71, + 181, + 521, + 238 + ], + "blocks": [ + { + "bbox": [ + 189, + 152, + 403, + 164 + ], + "lines": [ + { + "bbox": [ + 189, + 152, + 403, + 164 + ], + "spans": [ + { + "bbox": [ + 189, + 152, + 403, + 164 + ], + "type": "text", + "content": "Table 8: MLQA results (F1 / EM) for each language." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 71, + 181, + 521, + 238 + ], + "lines": [ + { + "bbox": [ + 71, + 181, + 521, + 238 + ], + "spans": [ + { + "bbox": [ + 71, + 181, + 521, + 238 + ], + "type": "table", + "html": "
ModelenarbnfiidkoruswteAvg.
XLM-Rbase66.8 / 57.355.7 / 42.031.5 / 20.452.6 / 40.369.1 / 55.636.3 / 27.954.8 / 36.553.0 / 34.737.4 / 28.850.8 / 38.2
mPMRbase71.1 / 61.666.3 / 52.656.5 / 41.665.5 / 53.173.9 / 63.750.4 / 38.864.4 / 37.957.4 / 41.165.3 / 50.463.4 / 49.0
XLM-R71.3 / 60.769.3 / 52.366.2 / 53.164.3 / 51.376.5 / 62.558.3 / 46.764.7 / 43.468.6 / 53.167.3 / 41.167.4 / 51.6
mPMR76.4 / 65.276.0 / 58.072.3 / 55.874.4 / 56.584.1 / 71.362.2 / 50.772.5 / 43.276.5 / 63.177.7 / 60.874.7 / 58.3
", + "image_path": "1b5ba5fe807c4a6f675f2589e3a00658e6068ac4dc90166fbd3297fd790dac4b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 71, + 277, + 523, + 384 + ], + "blocks": [ + { + "bbox": [ + 172, + 248, + 420, + 259 + ], + "lines": [ + { + "bbox": [ + 172, + 248, + 420, + 259 + ], + "spans": [ + { + "bbox": [ + 172, + 248, + 420, + 259 + ], + "type": "text", + "content": "Table 9: TyDiQA-GoldP results (F1 / EM) for each language." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 71, + 277, + 523, + 384 + ], + "lines": [ + { + "bbox": [ + 71, + 277, + 523, + 384 + ], + "spans": [ + { + "bbox": [ + 71, + 277, + 523, + 384 + ], + "type": "table", + "html": "
Modelenafarbgbndeeleseteufafifrhehihuiditjajv
XLM-Rbase84.275.347.379.066.377.575.378.069.656.038.170.481.450.867.972.451.079.619.663.9
mPMRbase85.180.757.680.271.981.277.679.579.171.349.680.482.465.271.782.258.683.543.272.0
XLM-R85.481.153.984.073.882.382.880.468.854.864.275.981.459.372.976.459.384.613.271.2
mPMR86.081.756.185.979.682.382.375.582.769.675.284.182.066.575.984.059.986.149.172.4
kakkkomlmrmsmynlptruswtatethtltrurviyozh
XLM-Rbase58.740.634.350.846.063.840.681.580.065.476.143.046.44.271.968.745.770.91.523.0
mPMRbase72.245.152.962.459.468.157.483.781.571.877.350.557.43.074.280.355.775.231.649.9
XLM-R59.941.741.356.858.276.729.686.185.272.277.652.351.67.178.870.964.080.027.222.4
mPMR77.346.857.970.668.173.857.886.083.672.879.862.658.13.883.080.376.283.636.154.4
", + "image_path": "2b924d373700d88f3d63ddbff2aad515ae9917aa0e0817abe6eb8f1fde306eb0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 184, + 421, + 408, + 503 + ], + "blocks": [ + { + "bbox": [ + 182, + 392, + 410, + 404 + ], + "lines": [ + { + "bbox": [ + 182, + 392, + 410, + 404 + ], + "spans": [ + { + "bbox": [ + 182, + 392, + 410, + 404 + ], + "type": "text", + "content": "Table 10: WikiAnn results (F1 Score) for each language." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 184, + 421, + 408, + 503 + ], + "lines": [ + { + "bbox": [ + 184, + 421, + 408, + 503 + ], + "spans": [ + { + "bbox": [ + 184, + 421, + 408, + 503 + ], + "type": "table", + "html": "
ModelendeesnlAvg.
XLM-Rbase91.371.078.775.779.2
mPMRbase91.974.380.879.781.7
XLM-R92.873.781.677.781.4
mPMR93.575.085.083.184.1
", + "image_path": "e39a5e9ac46de813adf7f5f2534e44b0c25b28019994c38c02fafc74419fd2ab.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 153, + 541, + 440, + 624 + ], + "blocks": [ + { + "bbox": [ + 184, + 512, + 407, + 524 + ], + "lines": [ + { + "bbox": [ + 184, + 512, + 407, + 524 + ], + "spans": [ + { + "bbox": [ + 184, + 512, + 407, + 524 + ], + "type": "text", + "content": "Table 11: CoNLL results (F1 Score) for each language." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 153, + 541, + 440, + 624 + ], + "lines": [ + { + "bbox": [ + 153, + 541, + 440, + 624 + ], + "spans": [ + { + "bbox": [ + 153, + 541, + 440, + 624 + ], + "type": "table", + "html": "
ModelenesfrnlrutrAvg.
XLM-Rbase76.565.455.661.256.145.460.0
mPMRbase77.668.656.462.259.548.462.1
XLM-R82.471.360.367.461.249.166.1
mPMR82.871.964.767.466.955.768.2
", + "image_path": "ff06593b4aad1ba656695b4b3b8ff3d54ea268de4b5e57efb95b4afc91919e8f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 138, + 661, + 455, + 743 + ], + "blocks": [ + { + "bbox": [ + 177, + 632, + 415, + 645 + ], + "lines": [ + { + "bbox": [ + 177, + 632, + 415, + 645 + ], + "spans": [ + { + "bbox": [ + 177, + 632, + 415, + 645 + ], + "type": "text", + "content": "Table 12: SemEval16 results (F1 Score) for each language." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 661, + 455, + 743 + ], + "lines": [ + { + "bbox": [ + 138, + 661, + 455, + 743 + ], + "spans": [ + { + "bbox": [ + 138, + 661, + 455, + 743 + ], + "type": "table", + "html": "
ModelendeesfrjakozhAvg.
XLM-Rbase94.387.789.188.777.076.681.385.0
mPMRbase94.388.490.188.979.079.482.486.1
XLM-R95.289.391.090.979.679.982.586.9
mPMR95.290.690.391.381.282.984.688.0
", + "image_path": "715bade8a7ae6d5e3f69176701c3b2b8e53d8ee58fca914c606c1bdf513c16f8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 171, + 751, + 420, + 764 + ], + "lines": [ + { + "bbox": [ + 171, + 751, + 420, + 764 + ], + "spans": [ + { + "bbox": [ + 171, + 751, + 420, + 764 + ], + "type": "text", + "content": "Table 13: PAWS-X accuracy scores (Acc.) for each language." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "1543" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 71, + 375, + 524, + 442 + ], + "blocks": [ + { + "bbox": [ + 71, + 375, + 524, + 442 + ], + "lines": [ + { + "bbox": [ + 71, + 375, + 524, + 442 + ], + "spans": [ + { + "bbox": [ + 71, + 375, + 524, + 442 + ], + "type": "table", + "html": "
ModelenarbgdeelesfrhiruswthtrurvizhAvg.
XLM-Rbase84.671.076.875.674.977.976.968.974.164.471.172.465.273.273.073.3
mPMRbase84.271.577.275.575.578.676.969.574.762.571.471.665.574.374.073.6
XLM-R88.277.081.781.281.284.281.774.978.970.875.777.470.678.077.778.6
mPMR88.377.982.982.281.083.582.275.279.871.276.178.971.678.979.079.3
", + "image_path": "d55ac649e3c8d6574665651add12dec735a7988f8e06f35f210203d46bd94124.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 178, + 450, + 414, + 463 + ], + "lines": [ + { + "bbox": [ + 178, + 450, + 414, + 463 + ], + "spans": [ + { + "bbox": [ + 178, + 450, + 414, + 463 + ], + "type": "text", + "content": "Table 14: XNLI accuracy scores (Acc.) for each language." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "text", + "content": "1544" + } + ] + } + ], + "index": 2 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 90, + 192, + 103 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 90, + 192, + 103 + ], + "spans": [ + { + "bbox": [ + 68, + 90, + 192, + 103 + ], + "type": "text", + "content": "A For every submission:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 106, + 317, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 106, + 317, + 121 + ], + "spans": [ + { + "bbox": [ + 77, + 106, + 317, + 121 + ], + "type": "text", + "content": "A1. Did you describe the limitations of your work?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 122, + 177, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 122, + 177, + 132 + ], + "spans": [ + { + "bbox": [ + 89, + 122, + 177, + 132 + ], + "type": "text", + "content": "Section Limitations" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 76, + 143, + 329, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 143, + 329, + 157 + ], + "spans": [ + { + "bbox": [ + 76, + 143, + 329, + 157 + ], + "type": "text", + "content": "A2. Did you discuss any potential risks of your work?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 158, + 208, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 158, + 208, + 169 + ], + "spans": [ + { + "bbox": [ + 89, + 158, + 208, + 169 + ], + "type": "text", + "content": "Not applicable. Left blank." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 77, + 179, + 414, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 179, + 414, + 192 + ], + "spans": [ + { + "bbox": [ + 77, + 179, + 414, + 192 + ], + "type": "text", + "content": "A3. Do the abstract and introduction summarize the paper's main claims?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 89, + 194, + 174, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 194, + 174, + 205 + ], + "spans": [ + { + "bbox": [ + 89, + 194, + 174, + 205 + ], + "type": "text", + "content": "Abstract, Section 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 77, + 215, + 398, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 215, + 398, + 229 + ], + "spans": [ + { + "bbox": [ + 77, + 215, + 398, + 229 + ], + "type": "text", + "content": "A4. Have you used AI writing assistants when working on this paper?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 89, + 230, + 138, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 230, + 138, + 242 + ], + "spans": [ + { + "bbox": [ + 89, + 230, + 138, + 242 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 68, + 252, + 290, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 252, + 290, + 266 + ], + "spans": [ + { + "bbox": [ + 68, + 252, + 290, + 266 + ], + "type": "text", + "content": "B Did you use or create scientific artifacts?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 79, + 270, + 122, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 270, + 122, + 282 + ], + "spans": [ + { + "bbox": [ + 79, + 270, + 122, + 282 + ], + "type": "text", + "content": "Section 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 77, + 291, + 315, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 291, + 315, + 306 + ], + "spans": [ + { + "bbox": [ + 77, + 291, + 315, + 306 + ], + "type": "text", + "content": "B1. Did you cite the creators of artifacts you used?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 91, + 306, + 132, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 306, + 132, + 317 + ], + "spans": [ + { + "bbox": [ + 91, + 306, + 132, + 317 + ], + "type": "text", + "content": "Section 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 76, + 328, + 463, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 328, + 463, + 342 + ], + "spans": [ + { + "bbox": [ + 76, + 328, + 463, + 342 + ], + "type": "text", + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 89, + 343, + 208, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 343, + 208, + 355 + ], + "spans": [ + { + "bbox": [ + 89, + 343, + 208, + 355 + ], + "type": "text", + "content": "Not applicable. Left blank." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 77, + 364, + 524, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 364, + 524, + 417 + ], + "spans": [ + { + "bbox": [ + 77, + 364, + 524, + 417 + ], + "type": "text", + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 89, + 419, + 196, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 419, + 196, + 431 + ], + "spans": [ + { + "bbox": [ + 89, + 419, + 196, + 431 + ], + "type": "text", + "content": "Section 3, Appendix A.1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 76, + 441, + 524, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 441, + 524, + 481 + ], + "spans": [ + { + "bbox": [ + 76, + 441, + 524, + 481 + ], + "type": "text", + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 89, + 482, + 208, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 482, + 208, + 495 + ], + "spans": [ + { + "bbox": [ + 89, + 482, + 208, + 495 + ], + "type": "text", + "content": "Not applicable. Left blank." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 77, + 503, + 524, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 503, + 524, + 531 + ], + "spans": [ + { + "bbox": [ + 77, + 503, + 524, + 531 + ], + "type": "text", + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 89, + 533, + 214, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 533, + 214, + 544 + ], + "spans": [ + { + "bbox": [ + 89, + 533, + 214, + 544 + ], + "type": "text", + "content": "Appendix A.1, Appendix A.2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 77, + 553, + 524, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 553, + 524, + 622 + ], + "spans": [ + { + "bbox": [ + 77, + 553, + 524, + 622 + ], + "type": "text", + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 89, + 623, + 150, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 623, + 150, + 634 + ], + "spans": [ + { + "bbox": [ + 89, + 623, + 150, + 634 + ], + "type": "text", + "content": "Appendix A.1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "spans": [ + { + "bbox": [ + 68, + 643, + 293, + 657 + ], + "type": "text", + "content": "C Did you run computational experiments?" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 79, + 662, + 122, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 662, + 122, + 673 + ], + "spans": [ + { + "bbox": [ + 79, + 662, + 122, + 673 + ], + "type": "text", + "content": "Section 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 77, + 683, + 524, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 683, + 524, + 711 + ], + "spans": [ + { + "bbox": [ + 77, + 683, + 524, + 711 + ], + "type": "text", + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 89, + 712, + 196, + 724 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 712, + 196, + 724 + ], + "spans": [ + { + "bbox": [ + 89, + 712, + 196, + 724 + ], + "type": "text", + "content": "Section 3, Appendix A.1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "spans": [ + { + "bbox": [ + 67, + 730, + 522, + 751 + ], + "type": "text", + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "spans": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "text", + "content": "ACL 2023 Responsible NLP Checklist" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "1545" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 77, + 70, + 523, + 238 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 77, + 70, + 523, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 70, + 523, + 111 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 523, + 111 + ], + "type": "text", + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Appendix A.1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 77, + 120, + 523, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 120, + 523, + 174 + ], + "spans": [ + { + "bbox": [ + 77, + 120, + 523, + 174 + ], + "type": "text", + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Section 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 183, + 523, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 183, + 523, + 238 + ], + "spans": [ + { + "bbox": [ + 77, + 183, + 523, + 238 + ], + "type": "text", + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Section 3, Appendix A.1" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 67, + 247, + 522, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 247, + 522, + 278 + ], + "spans": [ + { + "bbox": [ + 67, + 247, + 522, + 278 + ], + "type": "text", + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 76, + 286, + 523, + 539 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 76, + 286, + 523, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 286, + 523, + 326 + ], + "spans": [ + { + "bbox": [ + 76, + 286, + 523, + 326 + ], + "type": "text", + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 76, + 336, + 523, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 336, + 523, + 390 + ], + "spans": [ + { + "bbox": [ + 76, + 336, + 523, + 390 + ], + "type": "text", + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 76, + 400, + 523, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 400, + 523, + 453 + ], + "spans": [ + { + "bbox": [ + 76, + 400, + 523, + 453 + ], + "type": "text", + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 76, + 462, + 519, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 462, + 519, + 489 + ], + "spans": [ + { + "bbox": [ + 76, + 462, + 519, + 489 + ], + "type": "text", + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 76, + 498, + 523, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 498, + 523, + 539 + ], + "spans": [ + { + "bbox": [ + 76, + 498, + 523, + 539 + ], + "type": "text", + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 310, + 791 + ], + "type": "text", + "content": "1546" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 13 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/xSIM++_ An Improved Proxy to Bitext Mining Performance for Low-Resource Languages/632d8eaf-c025-45c6-9916-28d9ca6a54ad_content_list.json b/2023/xSIM++_ An Improved Proxy to Bitext Mining Performance for Low-Resource Languages/632d8eaf-c025-45c6-9916-28d9ca6a54ad_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..83ce45b601ef4225de8dad94ad4baf84f9a2e55c --- /dev/null +++ b/2023/xSIM++_ An Improved Proxy to Bitext Mining Performance for Low-Resource Languages/632d8eaf-c025-45c6-9916-28d9ca6a54ad_content_list.json @@ -0,0 +1,1577 @@ +[ + { + "type": "text", + "text": "xSIM++: An Improved Proxy to Bitext Mining Performance for Low-Resource Languages", + "text_level": 1, + "bbox": [ + 168, + 89, + 831, + 130 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mingda Chen*, Kevin Heffernan*, Onur Celebi, Alex Mourachko, Holger Schwenk", + "bbox": [ + 144, + 154, + 857, + 172 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{mingdachen,kevinheffernan,celebio,alexmourachko,schwenk}@meta.com", + "bbox": [ + 166, + 173, + 835, + 187 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Meta AI Research", + "bbox": [ + 426, + 189, + 574, + 203 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 252, + 339, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce a new proxy score for evaluating bitext mining based on similarity in a multilingual embedding space: $xsim++$ . In comparison to $xsim$ , this improved proxy leverages rule-based approaches to extend English sentences in any evaluation set with synthetic, hard-to-distinguish examples which more closely mirror the scenarios we encounter during large-scale mining. We validate this proxy by running a significant number of bitext mining experiments for a set of low-resource languages, and subsequently train NMT systems on the mined data. In comparison to $xsim$ , we show that $xsim++$ is better correlated with the downstream BLEU scores of translation systems trained on mined bitexts, providing a reliable proxy of bitext mining performance without needing to run expensive bitext mining pipelines. $xsim++$ also reports performance for different error types, offering more fine-grained feedback for model development.", + "bbox": [ + 141, + 281, + 460, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 114, + 594, + 260, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "When training neural machine translation (NMT) systems, it has been shown in prior works that generally, the quality of such systems increases with the availability of high-quality training data (Koehn and Knowles, 2017). However, for many low-resource languages there are few public corpora available, posing many challenges. In order to address this sparsity, one approach is to supplement existing datasets with automatically created parallel corpora, and a technique which has shown to be successful for such issues is the task of bitext mining (Schwenk et al., 2021b).", + "bbox": [ + 112, + 621, + 489, + 813 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In bitext mining, the aim is to find pairs of sentences with the same sentence meaning across collections of monolingual corpora. In this work, we adopt a global mining approach (Schwenk et al., 2021a), which has shown recent success in provid", + "bbox": [ + 112, + 815, + 489, + 896 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing high-quality data for low-resourced languages (NLLB Team et al., 2022).", + "bbox": [ + 507, + 253, + 880, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In order to evaluate any bitext mining method, a natural approach is to train a NMT system on the automatically created alignments. However, this is extremely costly. As an alternative, the BUCC task (Zweigenbaum et al., 2018) offers a method for evaluating bitext mining algorithms by embedding known alignments within monolingual corpora, and then reporting on the number of correctly aligned pairs. However, this task currently only covers 5 high-resourced languages (English, French, Russian, German and Chinese), and so is not applicable to the low-resource domain. In order to address this, another approach to evaluate bitext mining is to align existing multilingual parallel test sets. Two such test sets are Tatoeba1 and FLORES200.2 However, as shown by Heffernan et al. (2022), the Tatoeba corpus is not very reliable given that for some sentence pairs there are only a few hundred sentences. Therefore, we opt to use FLORES200, which is also n-way parallel.", + "bbox": [ + 507, + 285, + 884, + 607 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One existing method for evaluating bitext mining on parallel test sets is xsim. This method reports the error rate of misaligned sentences, and follows a margin-based global mining approach (Artetxe and Schwenk, 2019a). However, although using xsim on test sets such as FLORES200 has been shown to be useful as a proxy metric for bitext mining (NLLB Team et al., 2022), it has the following limitations:", + "bbox": [ + 507, + 608, + 882, + 752 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Using FLORES200 alone has proven to not be difficult enough as for many language pairs, existing approaches quickly saturate at $0\\%$ error (NLLB Team et al., 2022).", + "bbox": [ + 524, + 766, + 884, + 831 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "", + "bbox": [ + 507, + 841, + 863, + 866 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{2}$ https://github.com/facebookresearch/ Flores/tree/main/fores200", + "bbox": [ + 507, + 866, + 863, + 892 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{3}$ https://github.com/facebookresearch/LASER/tree/main/tasks/xsim", + "bbox": [ + 507, + 892, + 857, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 142, + 904, + 268, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "101", + "bbox": [ + 485, + 927, + 514, + 940 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics", + "bbox": [ + 226, + 945, + 771, + 957 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Volume 2: Short Papers, pages 101-109", + "bbox": [ + 376, + 958, + 621, + 971 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "July 9-14, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 295, + 972, + 700, + 985 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/617ea5b529f0f604b41bab80537eb9b6a874a7fcad6d68b455641e223b43b092.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Transformation CategoryOriginal SentenceTransformed Sentence
Causality AlternationApart from the fever and a sore throat, I feel well and in good shape to carry out my work by telecommuting.Apart from the fever and a sore throat, I feel well and in bad shape to carry out my work by telecommuting
Entity ReplacementCharles was the first member of the British Royal Family to be awarded a degree.M. Smith was the first member of The University to be awarded a degree.
Number ReplacementNadal bagged 88% net points in the match winning 76 points in the first serve.Nadal bagged 98% net points in the match winning 71 points in the sixth serve.
", + "bbox": [ + 122, + 80, + 873, + 200 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1: Examples of the transformations applied to the English sentences from FLORES200 dev set. The red texts indicate the places of alternations.", + "bbox": [ + 112, + 209, + 882, + 237 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "2. As the dev and devtest sets are quite small (997/1012 respectively), this is arguably not a good approximation for performance when mining against billions of possible candidate sentences.", + "3. We have observed that there is not a significant overlap in the semantics between candidate sentences, meaning that it is not possible to test difficult scenarios that arise in bitext mining when choosing between multiple (similar) candidate pairs." + ], + "bbox": [ + 127, + 263, + 485, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In order to address these limitations, in this work we introduce $x \\sin^{++}$ . This is an improved proxy for bitext mining performance which expands the dev and devtest sets of FLORES200 to include both more data points, and also difficult to distinguish cases which provide far greater challenges to the models. Our contributions can be summarised as follows:", + "bbox": [ + 112, + 458, + 485, + 586 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We create a more semantically challenging and expanded English test set for FLORES200.", + "2. We validate this new test set by independently performing 110 bitext mining runs, training 110 NMT systems on the output mined bittexts, and then determining both the correlation and statistical significance between $x \\sin + +$ and the resulting BLEU scores.", + "3. We open-source the expanded FLORES200 dev and devtest sets, and also the xsim++ code to evaluate them4." + ], + "bbox": [ + 127, + 596, + 487, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Methodology", + "text_level": 1, + "bbox": [ + 112, + 819, + 262, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 Background: xsim", + "text_level": 1, + "bbox": [ + 112, + 844, + 305, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given two lists of sentences in different languages, xsim seeks to align each sentence in the source", + "bbox": [ + 112, + 865, + 487, + 896 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "language to a corresponding sentence in the target language based on a margin-based similarity (Artetxe and Schwenk, 2019a). In doing so, xsim leverages the mining approach described in Artetxe and Schwenk (2019b) to first encode sentences into embedding vectors, assign pairwise scores between sentences in the lists, and then take the sentence in the target language that achieves the maximum score as the final prediction. xsim relies on human-annotated parallel corpora and measures the performance of bitext mining using the fraction of misaligned source sentences, i.e., error rates.", + "bbox": [ + 507, + 263, + 884, + 456 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2 xsim++", + "text_level": 1, + "bbox": [ + 507, + 469, + 610, + 483 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As the effectiveness of xsim is limited by the availability of parallel corpora, we choose to create xsim++ by automatically expanding the English sentences, and evaluate the sentence encoders on into-English language directions, following prior work on low-resource bitext mining (Heffernan et al., 2022). Aside from the expanded candidate set, xsim++ follows the same procedure as xsim.", + "bbox": [ + 505, + 491, + 882, + 620 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "$\\times \\mathrm{sim} + +$ seeks to capture more subtle improvements in bitext mining by adding challenging negative examples. The examples are human-written sentences transformed by various operations. These operations intend to perturb semantics through minimal alternations in the surface text. In particular, we use the following categories of transformations: causality alternation, entity replacement, and number replacement. We focus on these three transformation types only as they easily allow us to create negative examples. Examples of the transformed sentences are shown in Table 1. For these transformations, we adapt the implementation in Dhole et al. (2021) $^6$ and describe the details", + "bbox": [ + 507, + 621, + 882, + 845 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "5In this work we report all results using the absolute margin", + "bbox": [ + 507, + 854, + 882, + 879 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "Although this library has additional transformation methods available, many would create positive examples in this use case (e.g. paraphrases).", + "bbox": [ + 507, + 881, + 882, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "4https://github.com/facebookresearch/LASER", + "bbox": [ + 134, + 903, + 455, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "102", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/9be1e755847667d793e10b653d85b30354728abbbbd94ee250cb9d99825544ba.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Total ## per orig.
Original997-
Causality18681.87
Entity3774537.86
Number34763.49
", + "bbox": [ + 186, + 80, + 415, + 158 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 2: Total numbers of original sentences and transformed sentences in different transformation categories. We also report the averaged numbers of transformations per original sentence for each category.", + "bbox": [ + 112, + 167, + 489, + 225 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of these transformations below.", + "bbox": [ + 112, + 252, + 347, + 266 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Causality Alternation. To alter causality in a sentence, we (1) replace adjectives with their antonyms; (2) negate the meaning of sentences by adding or removing negation function words (e.g. \"did not\" and \"was not\") to the sentences; or (3) leverage the negation strengthening approach (Tan et al., 2021), which changes the causal relationships through more assertive function words (e.g. replacing \"may\" with \"will\"). For example, as shown in Table 1 we replace \"good\" with the antonym \"bad\".", + "bbox": [ + 112, + 279, + 487, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Entity Replacement. We collect candidate entities from large amounts of monolingual data. Then we replace entities in sentences with the ones randomly sampled from the candidate set. For both stages, we use the named entity recognizer from NLTK (Bird et al., 2009).", + "bbox": [ + 112, + 468, + 489, + 565 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Number Replacement. We use spaCy (Honni-bal and Montani, 2017) to detect dates, ordinals, cardinals, times, numbers, and percentages and then randomly replace their values.", + "bbox": [ + 112, + 577, + 489, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the strategies above, for each sentence we create multiple transformations (i.e. negative examples) of that source sentence. For example, consider Table 1. In the \"Entity Replacement\" example we create a transformation by replacing two named entities. We can then continue this process by replacing these with other named entities until we have reached the desired number of total transformations7. Note that since the opportunity to change each category is dependent on the frequency of that category in the evaluation sets, some transformations occurred more than others (e.g. entities were more frequent than numbers). We summarize the data statistics for xsim++ on the FLORES200 dev", + "bbox": [ + 112, + 658, + 489, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "set in Table 2. Results for the devtest set are in appendix A.", + "bbox": [ + 507, + 84, + 882, + 116 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 Experiment", + "text_level": 1, + "bbox": [ + 509, + 130, + 648, + 147 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In order to establish $x$ sim++ as a proxy for bitext mining performance, we measure the correlation between both $x$ sim and $x$ sim++ error rates, and the BLEU scores resulting from NMT systems trained on mined bittexts. More specifically, for each language we choose a sentence encoder model, followed by bitext mining using each respective encoder, and then train and evaluate bilingual NMT systems on the resulting mined bittexts. We use the FLORES200 development sets when computing the BLEU scores.", + "bbox": [ + 507, + 158, + 884, + 332 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In order to validate $\\times$ sim++ against varied embedding spaces, we encode (and mine) using two different multilingual encoder methods: LASER (Artetxe and Schwenk, 2019b) and LaBSE (Feng et al., 2022). For LASER, we trained our own custom encoders (details below). For LaBSE, we used a publicly available model as the code and data for training LaBSE are not publicly available.", + "bbox": [ + 507, + 336, + 882, + 464 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We randomly choose 10 low-resource languages to perform both encoder training (if applicable) and bitext mining. The languages are: Faroese (fao), Kabuverdianu (kea), Tok Pisin (tpi), Kikuyu (kik), Friulian (fur), Igbo (ibo), Luxembourgish (ltz), Swahili (swh), Zulu (zul), Bemba (bem).", + "bbox": [ + 507, + 464, + 882, + 562 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Encoder Training. We trained LASER encoders using the teacher-student approach described in Heffernan et al. (2022). We choose a LASER model (Artetxe and Schwenk, 2019b) as our teacher, and then trained specialised students for each language. In order to train each student, we used both publicly available code9 and bitexts (e.g. OPUS10)", + "bbox": [ + 507, + 573, + 882, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Bitext Mining. For each chosen encoder model, we perform bitext mining against approximately 3.7 billion sentences of English. For low-resource languages, the sizes of monolingual data range from 140k to 124 million. Details are in the appendix. We make use of monolingual data available from both Commoncrawl and Paracrawl11, and operationalize the mining using the stopes library (An", + "bbox": [ + 507, + 714, + 884, + 843 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "$^{8}$ https://huggingface.co/sentence-transformers/LaBSE \n $^{9}$ https://github.com/facebookresearch/fairseq/tree/nllb/examples/nllb/laser_distillation \n $^{10}$ https://opus.nlpl.eu \n $^{11}$ https://paracrawl.eu", + "bbox": [ + 507, + 852, + 873, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "7We set a maximum threshold of 100 transformations per category per sentence.", + "bbox": [ + 112, + 891, + 487, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "103", + "bbox": [ + 487, + 928, + 515, + 940 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "drews et al., 2022).12 For LASER, we use 1.06 as the margin threshold following Heffernan et al. (2022) and for LaBSE, we use 1.16.13 Following mining, for each language we concatenate publicly available bittexts and the mined bitext as training data for NMT bilingual models using fairseq,14 translating from each foreign text into English. For all NMT systems, we keep the hyperparameters fixed (details in Appendix).", + "bbox": [ + 112, + 83, + 489, + 231 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Evaluation. Model selection involves two use cases: comparisons within a model and across different models. For the former comparison, given our custom encoders, we choose to compare 10 checkpoints from each model.[15] For cross model comparisons, we compare each chosen encoder checkpoint against another existing system. In this case, the LaBSE encoder. To quantitatively measure these two cases, we report pairwise ranking accuracy (Kocmi et al., 2021) for xsim and xsim++. Formally, the accuracy is computed as follows", + "bbox": [ + 112, + 240, + 489, + 418 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\left| \\mathrm {s} (\\text {p r o x y} \\Delta) = \\mathrm {s} (\\text {m i n i n g} \\Delta) \\text {f o r a l l s y s t e m p a i r s} \\right|}{\\left| \\text {a l l s y s t e m p a i r s} \\right|}\n$$\n", + "text_format": "latex", + "bbox": [ + 124, + 444, + 478, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where proxy $\\Delta$ is the difference of the $x \\sin$ or $x \\sin + +$ scores, mining $\\Delta$ is the difference of the BLEU scores, $s(\\cdot)$ is the sign function, and $|\\cdot|$ returns the cardinal number of the input.", + "bbox": [ + 112, + 492, + 487, + 557 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this work, we have 550 system pairs with 55 pairs per language direction (i.e. $\\binom{11}{2}$ pairs given 10 custom LASER encoder checkpoints + LaBSE). We always compare systems within a language direction as the scores for system pairs across different directions are not comparable.[16]", + "bbox": [ + 112, + 557, + 489, + 653 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Results", + "text_level": 1, + "bbox": [ + 112, + 667, + 218, + 682 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Table 3, xsim++ significantly outperforms xsim on the pairwise ranking accuracy. Additionally, when comparing the computational cost to mining, xsim++ costs over $99.9\\%$ less GPU hours and saves approximately 3 metric tons of carbon", + "bbox": [ + 112, + 689, + 489, + 771 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/862fc12c58765f7375b533c1cef2da04658a980252a3da11a5f7dfbbb1e3420f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MetricAccuracyGPU hours
xsim35.480.43
xsim++72.00*0.52
Mining BLEU(Oracle)10019569
", + "bbox": [ + 531, + 80, + 863, + 147 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/c13fcad28dad65c1a159f599bc6129c3144421c8ff6b94a4c07ed1d22aa17929.jpg", + "table_caption": [ + "Table 3: Pairwise ranking accuracy along with the total number of GPU hours. For all experiments, we used NVIDIA A100 GPUs. An * indicates that the result passes the significance test proposed by Koehn (2004) with $p$ -value $< 0.05$ when compared to xsim." + ], + "table_footnote": [], + "table_body": "
Accuracy
xsim++72.00
Causality63.09
Entity65.45
Number60.73
Misaligned40.73
Causality + Entity68.55
Causality + Entity + Misaligned70.55
Causality + Misaligned68.00
Causality + Number66.73
Causality + Number + Misaligned71.45
Entity + Misaligned70.55
Number + Entity67.45
Number + Entity + Misaligned71.09
Number + Misaligned64.36
", + "bbox": [ + 542, + 241, + 852, + 435 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 4: Pairwise ranking accuracy when using combinations of error categories. Causality=Causality Alternation, Entity=Entity Replacement, Number=Number Replacement.", + "bbox": [ + 507, + 444, + 884, + 502 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "emissions, but still manages to achieve a competitive accuracy. We observe similar trends for the within a model and across models use cases and report their separate accuracies in the appendix.", + "bbox": [ + 505, + 527, + 882, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To better understand the contributions of each transformation category (cf. subsection 2.1) in measuring the final mining performance, we report accuracies for different combinations of categories in Table 4. In cases where an incorrect bitext alignment do does not map to any of the augmented sentences of the true alignment, we denote these as \"misaligned\". We find that entity replacement helps most in improving the accuracy and combing all the transformations gives the best performance.", + "bbox": [ + 505, + 592, + 882, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4 Related Work", + "text_level": 1, + "bbox": [ + 507, + 764, + 665, + 778 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As $xsim++$ uses rule-based data augmentation, it is related to work in other areas that also employ similar data augmentation methods, such as part-of-speech tagging (Şahin and Steedman, 2018), contrastive learning (Tang et al., 2022), text classification (Kobayashi, 2018; Wei and Zou, 2019), dialogue generation (Niu and Bansal, 2018) and summarization (Chen and Yang, 2021).", + "bbox": [ + 505, + 790, + 884, + 919 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "$^{12}$ https://github.com/facebookresearch/stopes", + "bbox": [ + 132, + 780, + 463, + 795 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "13We did grid search on threshold values from 1.11 to 1.25 on three languages (swh, ltz, and zul), decided the optimal one based on the BLEU scores, and used the threshold for the rest of languages.", + "bbox": [ + 115, + 795, + 485, + 843 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "$^{14}$ https://github.com/facebookresearch/fairseq", + "bbox": [ + 132, + 844, + 470, + 856 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "15Evenly spaced between epochs 1 and 30.", + "bbox": [ + 132, + 856, + 391, + 869 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "16There are factors varied across language directions that are unrelated to the quality of sentence encoders but could affect mining performance, such as amounts of monolingual data available for mining.", + "bbox": [ + 115, + 870, + 485, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "104", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5 Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 114, + 84, + 401, + 98 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We proposed a proxy score $x \\sin + +$ for bitext mining performance using three kinds of data augmentation techniques: causality alternation, entity replacement, and number replacement. To validate its effectiveness, we conducted large-scale bitext mining experiments for 10 low-resource languages, and reported pairwise ranking accuracies. We found that $x \\sin + +$ significantly improves over $x \\sin$ , doubling the accuracies. Analysis reveals that entity replacement helps most in the improvement. In future work, we plan to extend $x \\sin + +$ to non-English language pairs.", + "bbox": [ + 112, + 111, + 492, + 305 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6 Limitations", + "text_level": 1, + "bbox": [ + 112, + 318, + 250, + 332 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We highlight three limitations of our work. The first is that $xsim++$ is automatically constructed. There could be noisy sentences leading to errors that are irrelevant to the quality of encoders. The second is that $xsim++$ applies transformations solely to English sentences. Generalizing it to non-English language pairs requires additional research. Finally, we have experimented with the two most popular multilingual encoders: LASER and LaBSE. There are other available approaches which would be interesting to also validate $xsim++$ against.", + "bbox": [ + 112, + 346, + 489, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 114, + 551, + 213, + 565 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Pierre Andrews, Guillaume Wenzek, Kevin Heffernan, Onur Celebi, Anna Sun, Ammar Kamran, Yingzhe Guo, Alexandre Mourachko, Holger Schwenk, and Angela Fan. 2022. stopes - modular machine translation pipelines. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing: System Demonstrations. Association for Computational Linguistics.", + "Mikel Artetxe and Holger Schwenk. 2019a. Margin-based parallel corpus mining with multilingual sentence embeddings. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3197-3203, Florence, Italy. Association for Computational Linguistics.", + "Mikel Artetxe and Holger Schwenk. 2019b. Massively multilingual sentence embeddings for zero-shot cross-lingual transfer and beyond. Transactions of the Association for Computational Linguistics, 7:597-610.", + "Steven Bird, Ewan Klein, and Edward Loper. 2009. Natural language processing with Python: analyzing text with the natural language toolkit. \"O'Reilly Media, Inc.\"" + ], + "bbox": [ + 115, + 576, + 489, + 917 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Jiaao Chen and Diyi Yang. 2021. Simple conversational data augmentation for semi-supervised abstractive dialogue summarization. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6605-6616, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.", + "Kaustubh D. Dhole, Varun Gangal, Sebastian Gehrmann, Aadesh Gupta, Zhenhao Li, Saad Mahamood, Abinaya Mahendiran, Simon Mille, Ashish Srivastava, Samson Tan, Tongshuang Wu, Jascha Sohl-Dickstein, Jinho D. Choi, Eduard Hovy, Ondrej Dusek, Sebastian Ruder, Sajant Anand, Nagender Aneja, Rabin Banjade, Lisa Barthe, Hanna Behnke, Ian Berlot-Attwell, Connor Boyle, Caroline Brun, Marco Antonio Sobrevilla Cabezudo, Samuel Cahyawijaya, Emile Chapuis, Wanxiang Che, Mukund Choudhary, Christian Clauss, Pierre Colombo, Filip Cornell, Gautier Dagan, Mayukh Das, Tanay Dixit, Thomas Dopierre, Paul-Alexis Dray, Suchitra Dubey, Tatiana Ekeinhor, Marco Di Giovanni, Rishabh Gupta, Rishabh Gupta, Louanes Hamla, Sang Han, Fabrice Harel-Canada, Antoine Honore, Ishan Jindal, Przemyslaw K. Joniak, Denis Kleyko, Venelin Kovatchev, Kalpesh Krishna, Ashutosh Kumar, Stefan Langer, Seungjae Ryan Lee, Corey James Levinson, Hualou Liang, Kaizhao Liang, Zhexiong Liu, Andrey Lukyanenko, Vukosi Marivate, Gerard de Melo, Simon Meoni, Maxime Meyer, Afnan Mir, Nafise Sadat Moosavi, Niklas Muennighoff, Timothy Sum Hon Mun, Kenton Murray, Marcin Namysl, Maria Obedkova, Priti Oli, Nivranshu Pasricha, Jan Pfister, Richard Plant, Vinay Prabhu, Vasile Pais, Libo Qin, Shahab Raji, Pawan Kumar Rajpoot, Vikas Raunak, Roy Rinberg, Nicolas Roberts, Juan Diego Rodriguez, Claude Roux, Vasconcellos P. H. S., Ananya B. Sai, Robin M. Schmidt, Thomas Scialom, Tshephisho Sefara, Saqib N. Shamsi, Xudong Shen, Haoyue Shi, Yiwen Shi, Anna Shvets, Nick Siegel, Damien Sileo, Jamie Simon, Chandan Singh, Roman Sitelew, Priyank Soni, Taylor Sorensen, William Soto, Aman Srivastava, KV Aditya Srivatsa, Tony Sun, Mukund Varma T, A Tabassum, Fiona Anting Tan, Ryan Teehan, Mo Tiwari, Marie Tolkiehn, Athena Wang, Zijian Wang, Gloria Wang, Zijie J. Wang, Fuxuan Wei, Bryan Wilie, Genta Indra Winata, Xinyi Wu, Witold Wydmanski, Tianbao Xie, Usama Yaseen, M. Yee, Jing Zhang and Yue Zhang. 2021. Nl-augmenter: A framework for task-sensitive natural language augmentation.", + "Fangxiaoyu Feng, Yinfei Yang, Daniel Cer, Naveen Arivazhagan, and Wei Wang. 2022. Language-agnostic BERT sentence embedding. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 878–891, Dublin, Ireland. Association for Computational Linguistics.", + "Kevin Heffernan, Onur Celebi, and Holger Schwenk. 2022. Bitext mining using distilled sentence representations for low-resource languages. Findings of EMNLP." + ], + "bbox": [ + 510, + 85, + 884, + 917 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "105", + "bbox": [ + 487, + 928, + 515, + 940 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Matthew Honnibal and Ines Montani. 2017. spaCy 2: Natural language understanding with Bloom embeddings, convolutional neural networks and incremental parsing. To appear.", + "Sosuke Kobayashi. 2018. Contextual augmentation: Data augmentation by words with paradigmatic relations. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 452-457, New Orleans, Louisiana. Association for Computational Linguistics.", + "Tom Kocmi, Christian Federmann, Roman Grundkiewicz, Marcin Junczys-Dowmunt, Hitokazu Matsushita, and Arul Menezes. 2021. To ship or not to ship: An extensive evaluation of automatic metrics for machine translation. In Proceedings of the Sixth Conference on Machine Translation, pages 478-494, Online. Association for Computational Linguistics.", + "Philipp Koehn. 2004. Statistical significance tests for machine translation evaluation. In Proceedings of the 2004 Conference on Empirical Methods in Natural Language Processing, pages 388-395, Barcelona, Spain. Association for Computational Linguistics.", + "Philipp Koehn and Rebecca Knowles. 2017. Six challenges for neural machine translation. In Proceedings of the First Workshop on Neural Machine Translation, pages 28-39, Vancouver. Association for Computational Linguistics.", + "Tong Niu and Mohit Bansal. 2018. Adversarial over-sensitivity and over-stability strategies for dialogue models. In Proceedings of the 22nd Conference on Computational Natural Language Learning, pages 486-496, Brussels, Belgium. Association for Computational Linguistics.", + "NLLB Team, Marta R Costa-jussa, James Cross, Onur Celebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, et al. 2022. No language left behind: Scaling human-centered machine translation. arXiv preprint arXiv:2207.04672.", + "Gözde Gül Şahin and Mark Steedman. 2018. Data augmentation via dependency tree morphing for low-resource languages. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 5004-5009, Brussels, Belgium. Association for Computational Linguistics.", + "Holger Schwenk, Vishrav Chaudhary, Shuo Sun, Hongyu Gong, and Francisco Guzmán. 2021a. WikiMatrix: Mining 135M parallel sentences in 1620 language pairs from Wikipedia. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 1351-1361, Online. Association for Computational Linguistics." + ], + "bbox": [ + 115, + 85, + 485, + 917 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Holger Schwenk, Guillaume Wenzek, Sergey Edunov, Edouard Grave, Armand Joulin, and Angela Fan. 2021b. CCMatrix: Mining billions of high-quality parallel sentences on the web. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 6490-6500, Online. Association for Computational Linguistics.", + "Fiona Anting Tan, Devamanyu Hazarika, See-Kiong Ng, Soujanya Poria, and Roger Zimmermann. 2021. Causal augmentation for causal sentence classification. In Proceedings of the First Workshop on Causal Inference and NLP, pages 1-20, Punta Cana, Dominican Republic. Association for Computational Linguistics.", + "Zilu Tang, Muhammed Yusuf Kocyigit, and Derry Tanti Wijaya. 2022. AugCSE: Contrastive sentence embedding with diverse augmentations. In Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 375-398, Online only. Association for Computational Linguistics.", + "Jason Wei and Kai Zou. 2019. EDA: Easy data augmentation techniques for boosting performance on text classification tasks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 6382-6388, Hong Kong, China. Association for Computational Linguistics.", + "Pierre Zweigenbaum, Serge Sharoff, and Reinhard Rapp. 2018. Overview of the third bucc shared task: Spotting parallel sentences in comparable corpora. In Proceedings of 11th workshop on building and using comparable corpora, pages 39-42." + ], + "bbox": [ + 510, + 85, + 880, + 624 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "106", + "bbox": [ + 487, + 928, + 515, + 939 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A Data Statistics for xsim++ with FLORES200 devtest set", + "text_level": 1, + "bbox": [ + 114, + 83, + 415, + 115 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f3fb07f619cd222ae753326f05cb302c8ef348d388a9ea4315b1e5e474540496.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Total ##per orig.
Original1012-
Causality19161.89
Entity3885538.39
Number32623.22
", + "bbox": [ + 186, + 135, + 415, + 212 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We report the data statistics for $xsim++$ with FLORES200 devtest set in Table 5.", + "bbox": [ + 112, + 299, + 487, + 331 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "B Sizes of Monolingual data for Low-Resource Languages", + "text_level": 1, + "bbox": [ + 114, + 344, + 406, + 378 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e8b253c25b9af2b88dfb476d911b269691af3e938a3248bfc6206ae4437d89e4.jpg", + "table_caption": [ + "Table 5: Total numbers of original sentences and transformed sentences in different transformation categories. We also report the averaged numbers of transformations per original sentence for each category." + ], + "table_footnote": [], + "table_body": "
LanguageSize
kik147,902
kea226,507
fur737,178
fao1,179,475
tpi1,661,743
bem2,302,805
ibo8,124,418
zul20,477,331
swh55,399,821
ltz123,944,670
", + "bbox": [ + 211, + 399, + 391, + 541 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We report the sizes of monolingual data for each language in Table 6.", + "bbox": [ + 112, + 596, + 485, + 627 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "C Hyperparameters for NMT systems", + "text_level": 1, + "bbox": [ + 114, + 642, + 460, + 659 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ee4a25d73373525143c8163b0a4b4c72be5b4905693a4615d6c675e9d2964888.jpg", + "table_caption": [ + "Table 6: Number of monolingual sentences for each language." + ], + "table_footnote": [], + "table_body": "
encoder layers6
encoder attention heads8
encoder embed dim512
encoder FFNN embed dim4096
decoder layers6
decoder attention heads8
decoder embed dim512
decoder FFNN embed dim4096
optimiserAdam
adam betas(0.9, 0.98)
learning rate0.001
dropout0.3
spm vocab size7000
", + "bbox": [ + 168, + 680, + 436, + 845 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We report hyperparameters for NMT evaluations in Table 7.", + "bbox": [ + 112, + 887, + 487, + 917 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "D Within and Across Model Accuracies", + "text_level": 1, + "bbox": [ + 509, + 83, + 868, + 99 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7c6e984dc1c9ff46d3a17a581244a0e636e5d0ee78f2311b33c2ca96c314a444.jpg", + "table_caption": [ + "Table 7: Hyperparameters for NMT systems." + ], + "table_footnote": [], + "table_body": "
MetricWithinAcross
xsim31.3354.04
xsim++69.77*82.00*
", + "bbox": [ + 598, + 117, + 794, + 164 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 8: Pairwise ranking accuracy for comparisons within a model and across different models. An * indicates that the result passes the significance test proposed by Koehn (2004) with $p$ -value $< 0.05$ when compared to xsim.", + "bbox": [ + 507, + 173, + 882, + 244 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We report accuracies for within a model (i.e., LASER) and across different models (i.e., the 10 LASER checkpoints vs LaBSE) in Table 8.", + "bbox": [ + 507, + 263, + 882, + 312 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "107", + "bbox": [ + 485, + 928, + 515, + 940 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A For every submission:", + "bbox": [ + 115, + 107, + 322, + 122 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A1. Did you describe the limitations of your work?", + "bbox": [ + 129, + 126, + 532, + 143 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 151, + 143, + 231, + 159 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A2. Did you discuss any potential risks of your work?", + "bbox": [ + 129, + 170, + 552, + 186 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Not applicable. Left blank.", + "bbox": [ + 151, + 187, + 349, + 200 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A3. Do the abstract and introduction summarize the paper's main claims?", + "bbox": [ + 129, + 212, + 695, + 229 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 152, + 230, + 231, + 244 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A4. Have you used AI writing assistants when working on this paper?", + "bbox": [ + 129, + 255, + 668, + 272 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 152, + 273, + 231, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "B Did you use or create scientific artifacts?", + "bbox": [ + 115, + 299, + 487, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Section 2", + "bbox": [ + 132, + 321, + 205, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "B1. Did you cite the creators of artifacts you used?", + "bbox": [ + 129, + 346, + 529, + 363 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Section 2", + "bbox": [ + 152, + 363, + 221, + 376 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?", + "bbox": [ + 129, + 390, + 778, + 406 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Not applicable. Left blank.", + "bbox": [ + 151, + 407, + 349, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?", + "bbox": [ + 129, + 432, + 880, + 495 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Not applicable. Left blank.", + "bbox": [ + 151, + 498, + 349, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?", + "bbox": [ + 129, + 524, + 880, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Not applicable. Left blank.", + "bbox": [ + 151, + 573, + 349, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?", + "bbox": [ + 129, + 598, + 880, + 631 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Section 2", + "bbox": [ + 152, + 632, + 221, + 646 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be.", + "bbox": [ + 129, + 657, + 880, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Section 2", + "bbox": [ + 152, + 740, + 221, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "C Did you run computational experiments?", + "bbox": [ + 115, + 764, + 492, + 781 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Section 3", + "bbox": [ + 132, + 787, + 205, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?", + "bbox": [ + 129, + 812, + 880, + 845 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Section 3", + "bbox": [ + 152, + 846, + 221, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.", + "bbox": [ + 112, + 866, + 877, + 889 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "ACL 2023 Responsible NLP Checklist", + "bbox": [ + 132, + 84, + 433, + 99 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "108", + "bbox": [ + 487, + 928, + 515, + 940 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Section 3 and Appendix", + "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Section 3 and Appendix", + "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Section 3" + ], + "bbox": [ + 129, + 83, + 878, + 280 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?", + "text_level": 1, + "bbox": [ + 112, + 293, + 877, + 310 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Left blank.", + "bbox": [ + 132, + 313, + 213, + 329 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response.", + "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response.", + "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response.", + "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response.", + "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response." + ], + "bbox": [ + 127, + 341, + 878, + 640 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "109", + "bbox": [ + 487, + 928, + 515, + 940 + ], + "page_idx": 8 + } +] \ No newline at end of file diff --git a/2023/xSIM++_ An Improved Proxy to Bitext Mining Performance for Low-Resource Languages/632d8eaf-c025-45c6-9916-28d9ca6a54ad_model.json b/2023/xSIM++_ An Improved Proxy to Bitext Mining Performance for Low-Resource Languages/632d8eaf-c025-45c6-9916-28d9ca6a54ad_model.json new file mode 100644 index 0000000000000000000000000000000000000000..96e4eea2f9a8a940093b2006ba210a9708414d60 --- /dev/null +++ b/2023/xSIM++_ An Improved Proxy to Bitext Mining Performance for Low-Resource Languages/632d8eaf-c025-45c6-9916-28d9ca6a54ad_model.json @@ -0,0 +1,1956 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.169, + 0.09, + 0.833, + 0.131 + ], + "angle": 0, + "content": "xSIM++: An Improved Proxy to Bitext Mining Performance for Low-Resource Languages" + }, + { + "type": "text", + "bbox": [ + 0.146, + 0.155, + 0.858, + 0.173 + ], + "angle": 0, + "content": "Mingda Chen*, Kevin Heffernan*, Onur Celebi, Alex Mourachko, Holger Schwenk" + }, + { + "type": "text", + "bbox": [ + 0.168, + 0.174, + 0.836, + 0.189 + ], + "angle": 0, + "content": "{mingdachen,kevinheffernan,celebio,alexmourachko,schwenk}@meta.com" + }, + { + "type": "text", + "bbox": [ + 0.427, + 0.19, + 0.576, + 0.204 + ], + "angle": 0, + "content": "Meta AI Research" + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.253, + 0.341, + 0.269 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.282, + 0.461, + 0.581 + ], + "angle": 0, + "content": "We introduce a new proxy score for evaluating bitext mining based on similarity in a multilingual embedding space: \\( xsim++ \\). In comparison to \\( xsim \\), this improved proxy leverages rule-based approaches to extend English sentences in any evaluation set with synthetic, hard-to-distinguish examples which more closely mirror the scenarios we encounter during large-scale mining. We validate this proxy by running a significant number of bitext mining experiments for a set of low-resource languages, and subsequently train NMT systems on the mined data. In comparison to \\( xsim \\), we show that \\( xsim++ \\) is better correlated with the downstream BLEU scores of translation systems trained on mined bitexts, providing a reliable proxy of bitext mining performance without needing to run expensive bitext mining pipelines. \\( xsim++ \\) also reports performance for different error types, offering more fine-grained feedback for model development." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.595, + 0.262, + 0.611 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.622, + 0.49, + 0.814 + ], + "angle": 0, + "content": "When training neural machine translation (NMT) systems, it has been shown in prior works that generally, the quality of such systems increases with the availability of high-quality training data (Koehn and Knowles, 2017). However, for many low-resource languages there are few public corpora available, posing many challenges. In order to address this sparsity, one approach is to supplement existing datasets with automatically created parallel corpora, and a technique which has shown to be successful for such issues is the task of bitext mining (Schwenk et al., 2021b)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.816, + 0.49, + 0.897 + ], + "angle": 0, + "content": "In bitext mining, the aim is to find pairs of sentences with the same sentence meaning across collections of monolingual corpora. In this work, we adopt a global mining approach (Schwenk et al., 2021a), which has shown recent success in provid" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.254, + 0.882, + 0.284 + ], + "angle": 0, + "content": "ing high-quality data for low-resourced languages (NLLB Team et al., 2022)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.286, + 0.885, + 0.608 + ], + "angle": 0, + "content": "In order to evaluate any bitext mining method, a natural approach is to train a NMT system on the automatically created alignments. However, this is extremely costly. As an alternative, the BUCC task (Zweigenbaum et al., 2018) offers a method for evaluating bitext mining algorithms by embedding known alignments within monolingual corpora, and then reporting on the number of correctly aligned pairs. However, this task currently only covers 5 high-resourced languages (English, French, Russian, German and Chinese), and so is not applicable to the low-resource domain. In order to address this, another approach to evaluate bitext mining is to align existing multilingual parallel test sets. Two such test sets are Tatoeba1 and FLORES200.2 However, as shown by Heffernan et al. (2022), the Tatoeba corpus is not very reliable given that for some sentence pairs there are only a few hundred sentences. Therefore, we opt to use FLORES200, which is also n-way parallel." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.609, + 0.884, + 0.753 + ], + "angle": 0, + "content": "One existing method for evaluating bitext mining on parallel test sets is xsim. This method reports the error rate of misaligned sentences, and follows a margin-based global mining approach (Artetxe and Schwenk, 2019a). However, although using xsim on test sets such as FLORES200 has been shown to be useful as a proxy metric for bitext mining (NLLB Team et al., 2022), it has the following limitations:" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.768, + 0.885, + 0.832 + ], + "angle": 0, + "content": "1. Using FLORES200 alone has proven to not be difficult enough as for many language pairs, existing approaches quickly saturate at \\(0\\%\\) error (NLLB Team et al., 2022)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.508, + 0.843, + 0.864, + 0.868 + ], + "angle": 0, + "content": "" + }, + { + "type": "page_footnote", + "bbox": [ + 0.508, + 0.868, + 0.865, + 0.893 + ], + "angle": 0, + "content": "\\(^{2}\\)https://github.com/facebookresearch/ Flores/tree/main/fores200" + }, + { + "type": "page_footnote", + "bbox": [ + 0.508, + 0.893, + 0.858, + 0.918 + ], + "angle": 0, + "content": "\\(^{3}\\)https://github.com/facebookresearch/LASER/tree/main/tasks/xsim" + }, + { + "type": "list", + "bbox": [ + 0.508, + 0.843, + 0.865, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.143, + 0.905, + 0.269, + 0.919 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.928, + 0.515, + 0.941 + ], + "angle": 0, + "content": "101" + }, + { + "type": "footer", + "bbox": [ + 0.227, + 0.946, + 0.772, + 0.958 + ], + "angle": 0, + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + }, + { + "type": "footer", + "bbox": [ + 0.377, + 0.959, + 0.623, + 0.972 + ], + "angle": 0, + "content": "Volume 2: Short Papers, pages 101-109" + }, + { + "type": "footer", + "bbox": [ + 0.297, + 0.973, + 0.702, + 0.986 + ], + "angle": 0, + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.124, + 0.081, + 0.875, + 0.201 + ], + "angle": 0, + "content": "
Transformation CategoryOriginal SentenceTransformed Sentence
Causality AlternationApart from the fever and a sore throat, I feel well and in good shape to carry out my work by telecommuting.Apart from the fever and a sore throat, I feel well and in bad shape to carry out my work by telecommuting
Entity ReplacementCharles was the first member of the British Royal Family to be awarded a degree.M. Smith was the first member of The University to be awarded a degree.
Number ReplacementNadal bagged 88% net points in the match winning 76 points in the first serve.Nadal bagged 98% net points in the match winning 71 points in the sixth serve.
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.21, + 0.883, + 0.239 + ], + "angle": 0, + "content": "Table 1: Examples of the transformations applied to the English sentences from FLORES200 dev set. The red texts indicate the places of alternations." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.265, + 0.487, + 0.343 + ], + "angle": 0, + "content": "2. As the dev and devtest sets are quite small (997/1012 respectively), this is arguably not a good approximation for performance when mining against billions of possible candidate sentences." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.355, + 0.487, + 0.451 + ], + "angle": 0, + "content": "3. We have observed that there is not a significant overlap in the semantics between candidate sentences, meaning that it is not possible to test difficult scenarios that arise in bitext mining when choosing between multiple (similar) candidate pairs." + }, + { + "type": "list", + "bbox": [ + 0.129, + 0.265, + 0.487, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.459, + 0.487, + 0.587 + ], + "angle": 0, + "content": "In order to address these limitations, in this work we introduce \\( x \\sin^{++} \\). This is an improved proxy for bitext mining performance which expands the dev and devtest sets of FLORES200 to include both more data points, and also difficult to distinguish cases which provide far greater challenges to the models. Our contributions can be summarised as follows:" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.597, + 0.489, + 0.643 + ], + "angle": 0, + "content": "1. We create a more semantically challenging and expanded English test set for FLORES200." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.655, + 0.487, + 0.75 + ], + "angle": 0, + "content": "2. We validate this new test set by independently performing 110 bitext mining runs, training 110 NMT systems on the output mined bittexts, and then determining both the correlation and statistical significance between \\( x \\sin + + \\) and the resulting BLEU scores." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.761, + 0.486, + 0.807 + ], + "angle": 0, + "content": "3. We open-source the expanded FLORES200 dev and devtest sets, and also the xsim++ code to evaluate them4." + }, + { + "type": "list", + "bbox": [ + 0.129, + 0.597, + 0.489, + 0.807 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.82, + 0.263, + 0.837 + ], + "angle": 0, + "content": "2 Methodology" + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.845, + 0.307, + 0.86 + ], + "angle": 0, + "content": "2.1 Background: xsim" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.866, + 0.488, + 0.897 + ], + "angle": 0, + "content": "Given two lists of sentences in different languages, xsim seeks to align each sentence in the source" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.265, + 0.885, + 0.457 + ], + "angle": 0, + "content": "language to a corresponding sentence in the target language based on a margin-based similarity (Artetxe and Schwenk, 2019a). In doing so, xsim leverages the mining approach described in Artetxe and Schwenk (2019b) to first encode sentences into embedding vectors, assign pairwise scores between sentences in the lists, and then take the sentence in the target language that achieves the maximum score as the final prediction. xsim relies on human-annotated parallel corpora and measures the performance of bitext mining using the fraction of misaligned source sentences, i.e., error rates." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.47, + 0.611, + 0.484 + ], + "angle": 0, + "content": "2.2 xsim++" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.492, + 0.883, + 0.621 + ], + "angle": 0, + "content": "As the effectiveness of xsim is limited by the availability of parallel corpora, we choose to create xsim++ by automatically expanding the English sentences, and evaluate the sentence encoders on into-English language directions, following prior work on low-resource bitext mining (Heffernan et al., 2022). Aside from the expanded candidate set, xsim++ follows the same procedure as xsim." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.622, + 0.884, + 0.846 + ], + "angle": 0, + "content": "\\(\\times \\mathrm{sim} + +\\) seeks to capture more subtle improvements in bitext mining by adding challenging negative examples. The examples are human-written sentences transformed by various operations. These operations intend to perturb semantics through minimal alternations in the surface text. In particular, we use the following categories of transformations: causality alternation, entity replacement, and number replacement. We focus on these three transformation types only as they easily allow us to create negative examples. Examples of the transformed sentences are shown in Table 1. For these transformations, we adapt the implementation in Dhole et al. (2021)\\(^6\\) and describe the details" + }, + { + "type": "page_footnote", + "bbox": [ + 0.508, + 0.856, + 0.884, + 0.881 + ], + "angle": 0, + "content": "5In this work we report all results using the absolute margin" + }, + { + "type": "page_footnote", + "bbox": [ + 0.508, + 0.882, + 0.884, + 0.919 + ], + "angle": 0, + "content": "Although this library has additional transformation methods available, many would create positive examples in this use case (e.g. paraphrases)." + }, + { + "type": "list", + "bbox": [ + 0.508, + 0.856, + 0.884, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.136, + 0.904, + 0.456, + 0.918 + ], + "angle": 0, + "content": "4https://github.com/facebookresearch/LASER" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.929, + 0.516, + 0.941 + ], + "angle": 0, + "content": "102" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.188, + 0.082, + 0.416, + 0.159 + ], + "angle": 0, + "content": "
Total ## per orig.
Original997-
Causality18681.87
Entity3774537.86
Number34763.49
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.168, + 0.49, + 0.227 + ], + "angle": 0, + "content": "Table 2: Total numbers of original sentences and transformed sentences in different transformation categories. We also report the averaged numbers of transformations per original sentence for each category." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.253, + 0.348, + 0.267 + ], + "angle": 0, + "content": "of these transformations below." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.28, + 0.489, + 0.456 + ], + "angle": 0, + "content": "Causality Alternation. To alter causality in a sentence, we (1) replace adjectives with their antonyms; (2) negate the meaning of sentences by adding or removing negation function words (e.g. \"did not\" and \"was not\") to the sentences; or (3) leverage the negation strengthening approach (Tan et al., 2021), which changes the causal relationships through more assertive function words (e.g. replacing \"may\" with \"will\"). For example, as shown in Table 1 we replace \"good\" with the antonym \"bad\"." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.469, + 0.49, + 0.566 + ], + "angle": 0, + "content": "Entity Replacement. We collect candidate entities from large amounts of monolingual data. Then we replace entities in sentences with the ones randomly sampled from the candidate set. For both stages, we use the named entity recognizer from NLTK (Bird et al., 2009)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.578, + 0.49, + 0.643 + ], + "angle": 0, + "content": "Number Replacement. We use spaCy (Honni-bal and Montani, 2017) to detect dates, ordinals, cardinals, times, numbers, and percentages and then randomly replace their values." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.659, + 0.49, + 0.884 + ], + "angle": 0, + "content": "Given the strategies above, for each sentence we create multiple transformations (i.e. negative examples) of that source sentence. For example, consider Table 1. In the \"Entity Replacement\" example we create a transformation by replacing two named entities. We can then continue this process by replacing these with other named entities until we have reached the desired number of total transformations7. Note that since the opportunity to change each category is dependent on the frequency of that category in the evaluation sets, some transformations occurred more than others (e.g. entities were more frequent than numbers). We summarize the data statistics for xsim++ on the FLORES200 dev" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.883, + 0.117 + ], + "angle": 0, + "content": "set in Table 2. Results for the devtest set are in appendix A." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.131, + 0.649, + 0.148 + ], + "angle": 0, + "content": "3 Experiment" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.159, + 0.885, + 0.334 + ], + "angle": 0, + "content": "In order to establish \\( x \\) sim++ as a proxy for bitext mining performance, we measure the correlation between both \\( x \\) sim and \\( x \\) sim++ error rates, and the BLEU scores resulting from NMT systems trained on mined bittexts. More specifically, for each language we choose a sentence encoder model, followed by bitext mining using each respective encoder, and then train and evaluate bilingual NMT systems on the resulting mined bittexts. We use the FLORES200 development sets when computing the BLEU scores." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.337, + 0.884, + 0.465 + ], + "angle": 0, + "content": "In order to validate \\( \\times \\)sim++ against varied embedding spaces, we encode (and mine) using two different multilingual encoder methods: LASER (Artetxe and Schwenk, 2019b) and LaBSE (Feng et al., 2022). For LASER, we trained our own custom encoders (details below). For LaBSE, we used a publicly available model as the code and data for training LaBSE are not publicly available." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.466, + 0.884, + 0.563 + ], + "angle": 0, + "content": "We randomly choose 10 low-resource languages to perform both encoder training (if applicable) and bitext mining. The languages are: Faroese (fao), Kabuverdianu (kea), Tok Pisin (tpi), Kikuyu (kik), Friulian (fur), Igbo (ibo), Luxembourgish (ltz), Swahili (swh), Zulu (zul), Bemba (bem)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.574, + 0.884, + 0.703 + ], + "angle": 0, + "content": "Encoder Training. We trained LASER encoders using the teacher-student approach described in Heffernan et al. (2022). We choose a LASER model (Artetxe and Schwenk, 2019b) as our teacher, and then trained specialised students for each language. In order to train each student, we used both publicly available code9 and bitexts (e.g. OPUS10)" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.715, + 0.885, + 0.844 + ], + "angle": 0, + "content": "Bitext Mining. For each chosen encoder model, we perform bitext mining against approximately 3.7 billion sentences of English. For low-resource languages, the sizes of monolingual data range from 140k to 124 million. Details are in the appendix. We make use of monolingual data available from both Commoncrawl and Paracrawl11, and operationalize the mining using the stopes library (An" + }, + { + "type": "page_footnote", + "bbox": [ + 0.508, + 0.853, + 0.874, + 0.918 + ], + "angle": 0, + "content": "\\(^{8}\\)https://huggingface.co/sentence-transformers/LaBSE \n\\(^{9}\\)https://github.com/facebookresearch/fairseq/tree/nllb/examples/nllb/laser_distillation \n\\(^{10}\\)https://opus.nlpl.eu \n\\(^{11}\\)https://paracrawl.eu" + }, + { + "type": "page_footnote", + "bbox": [ + 0.113, + 0.892, + 0.488, + 0.919 + ], + "angle": 0, + "content": "7We set a maximum threshold of 100 transformations per category per sentence." + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.929, + 0.516, + 0.941 + ], + "angle": 0, + "content": "103" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.084, + 0.49, + 0.232 + ], + "angle": 0, + "content": "drews et al., 2022).12 For LASER, we use 1.06 as the margin threshold following Heffernan et al. (2022) and for LaBSE, we use 1.16.13 Following mining, for each language we concatenate publicly available bittexts and the mined bitext as training data for NMT bilingual models using fairseq,14 translating from each foreign text into English. For all NMT systems, we keep the hyperparameters fixed (details in Appendix)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.241, + 0.49, + 0.419 + ], + "angle": 0, + "content": "Evaluation. Model selection involves two use cases: comparisons within a model and across different models. For the former comparison, given our custom encoders, we choose to compare 10 checkpoints from each model.[15] For cross model comparisons, we compare each chosen encoder checkpoint against another existing system. In this case, the LaBSE encoder. To quantitatively measure these two cases, we report pairwise ranking accuracy (Kocmi et al., 2021) for xsim and xsim++. Formally, the accuracy is computed as follows" + }, + { + "type": "equation", + "bbox": [ + 0.125, + 0.445, + 0.479, + 0.48 + ], + "angle": 0, + "content": "\\[\n\\frac {\\left| \\mathrm {s} (\\text {p r o x y} \\Delta) = \\mathrm {s} (\\text {m i n i n g} \\Delta) \\text {f o r a l l s y s t e m p a i r s} \\right|}{\\left| \\text {a l l s y s t e m p a i r s} \\right|}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.493, + 0.489, + 0.558 + ], + "angle": 0, + "content": "where proxy\\(\\Delta\\) is the difference of the \\(x \\sin\\) or \\(x \\sin + +\\) scores, mining\\(\\Delta\\) is the difference of the BLEU scores, \\(s(\\cdot)\\) is the sign function, and \\(|\\cdot|\\) returns the cardinal number of the input." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.558, + 0.49, + 0.655 + ], + "angle": 0, + "content": "In this work, we have 550 system pairs with 55 pairs per language direction (i.e. \\(\\binom{11}{2}\\) pairs given 10 custom LASER encoder checkpoints + LaBSE). We always compare systems within a language direction as the scores for system pairs across different directions are not comparable.[16]" + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.668, + 0.22, + 0.683 + ], + "angle": 0, + "content": "3.1 Results" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.69, + 0.49, + 0.772 + ], + "angle": 0, + "content": "As shown in Table 3, xsim++ significantly outperforms xsim on the pairwise ranking accuracy. Additionally, when comparing the computational cost to mining, xsim++ costs over \\(99.9\\%\\) less GPU hours and saves approximately 3 metric tons of carbon" + }, + { + "type": "table", + "bbox": [ + 0.532, + 0.081, + 0.864, + 0.148 + ], + "angle": 0, + "content": "
MetricAccuracyGPU hours
xsim35.480.43
xsim++72.00*0.52
Mining BLEU(Oracle)10019569
" + }, + { + "type": "table_caption", + "bbox": [ + 0.508, + 0.156, + 0.883, + 0.228 + ], + "angle": 0, + "content": "Table 3: Pairwise ranking accuracy along with the total number of GPU hours. For all experiments, we used NVIDIA A100 GPUs. An * indicates that the result passes the significance test proposed by Koehn (2004) with \\( p \\)-value \\( < 0.05 \\) when compared to xsim." + }, + { + "type": "table", + "bbox": [ + 0.543, + 0.242, + 0.853, + 0.436 + ], + "angle": 0, + "content": "
Accuracy
xsim++72.00
Causality63.09
Entity65.45
Number60.73
Misaligned40.73
Causality + Entity68.55
Causality + Entity + Misaligned70.55
Causality + Misaligned68.00
Causality + Number66.73
Causality + Number + Misaligned71.45
Entity + Misaligned70.55
Number + Entity67.45
Number + Entity + Misaligned71.09
Number + Misaligned64.36
" + }, + { + "type": "table_caption", + "bbox": [ + 0.508, + 0.445, + 0.885, + 0.503 + ], + "angle": 0, + "content": "Table 4: Pairwise ranking accuracy when using combinations of error categories. Causality=Causality Alternation, Entity=Entity Replacement, Number=Number Replacement." + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.528, + 0.884, + 0.592 + ], + "angle": 0, + "content": "emissions, but still manages to achieve a competitive accuracy. We observe similar trends for the within a model and across models use cases and report their separate accuracies in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.593, + 0.884, + 0.754 + ], + "angle": 0, + "content": "To better understand the contributions of each transformation category (cf. subsection 2.1) in measuring the final mining performance, we report accuracies for different combinations of categories in Table 4. In cases where an incorrect bitext alignment do does not map to any of the augmented sentences of the true alignment, we denote these as \"misaligned\". We find that entity replacement helps most in improving the accuracy and combing all the transformations gives the best performance." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.765, + 0.667, + 0.78 + ], + "angle": 0, + "content": "4 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.791, + 0.885, + 0.92 + ], + "angle": 0, + "content": "As \\( xsim++ \\) uses rule-based data augmentation, it is related to work in other areas that also employ similar data augmentation methods, such as part-of-speech tagging (Şahin and Steedman, 2018), contrastive learning (Tang et al., 2022), text classification (Kobayashi, 2018; Wei and Zou, 2019), dialogue generation (Niu and Bansal, 2018) and summarization (Chen and Yang, 2021)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.133, + 0.781, + 0.464, + 0.796 + ], + "angle": 0, + "content": "\\(^{12}\\)https://github.com/facebookresearch/stopes" + }, + { + "type": "page_footnote", + "bbox": [ + 0.116, + 0.796, + 0.486, + 0.844 + ], + "angle": 0, + "content": "13We did grid search on threshold values from 1.11 to 1.25 on three languages (swh, ltz, and zul), decided the optimal one based on the BLEU scores, and used the threshold for the rest of languages." + }, + { + "type": "page_footnote", + "bbox": [ + 0.134, + 0.845, + 0.471, + 0.857 + ], + "angle": 0, + "content": "\\(^{14}\\)https://github.com/facebookresearch/fairseq" + }, + { + "type": "page_footnote", + "bbox": [ + 0.134, + 0.857, + 0.393, + 0.87 + ], + "angle": 0, + "content": "15Evenly spaced between epochs 1 and 30." + }, + { + "type": "page_footnote", + "bbox": [ + 0.116, + 0.871, + 0.487, + 0.919 + ], + "angle": 0, + "content": "16There are factors varied across language directions that are unrelated to the quality of sentence encoders but could affect mining performance, such as amounts of monolingual data available for mining." + }, + { + "type": "list", + "bbox": [ + 0.116, + 0.781, + 0.487, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.929, + 0.517, + 0.941 + ], + "angle": 0, + "content": "104" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.115, + 0.085, + 0.402, + 0.099 + ], + "angle": 0, + "content": "5 Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.112, + 0.493, + 0.306 + ], + "angle": 0, + "content": "We proposed a proxy score \\( x \\sin + + \\) for bitext mining performance using three kinds of data augmentation techniques: causality alternation, entity replacement, and number replacement. To validate its effectiveness, we conducted large-scale bitext mining experiments for 10 low-resource languages, and reported pairwise ranking accuracies. We found that \\( x \\sin + + \\) significantly improves over \\( x \\sin \\), doubling the accuracies. Analysis reveals that entity replacement helps most in the improvement. In future work, we plan to extend \\( x \\sin + + \\) to non-English language pairs." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.319, + 0.251, + 0.334 + ], + "angle": 0, + "content": "6 Limitations" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.347, + 0.49, + 0.523 + ], + "angle": 0, + "content": "We highlight three limitations of our work. The first is that \\( xsim++ \\) is automatically constructed. There could be noisy sentences leading to errors that are irrelevant to the quality of encoders. The second is that \\( xsim++ \\) applies transformations solely to English sentences. Generalizing it to non-English language pairs requires additional research. Finally, we have experimented with the two most popular multilingual encoders: LASER and LaBSE. There are other available approaches which would be interesting to also validate \\( xsim++ \\) against." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.552, + 0.214, + 0.567 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.577, + 0.49, + 0.684 + ], + "angle": 0, + "content": "Pierre Andrews, Guillaume Wenzek, Kevin Heffernan, Onur Celebi, Anna Sun, Ammar Kamran, Yingzhe Guo, Alexandre Mourachko, Holger Schwenk, and Angela Fan. 2022. stopes - modular machine translation pipelines. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing: System Demonstrations. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.695, + 0.49, + 0.775 + ], + "angle": 0, + "content": "Mikel Artetxe and Holger Schwenk. 2019a. Margin-based parallel corpus mining with multilingual sentence embeddings. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3197-3203, Florence, Italy. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.786, + 0.49, + 0.853 + ], + "angle": 0, + "content": "Mikel Artetxe and Holger Schwenk. 2019b. Massively multilingual sentence embeddings for zero-shot cross-lingual transfer and beyond. Transactions of the Association for Computational Linguistics, 7:597-610." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.865, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Steven Bird, Ewan Klein, and Edward Loper. 2009. Natural language processing with Python: analyzing text with the natural language toolkit. \"O'Reilly Media, Inc.\"" + }, + { + "type": "list", + "bbox": [ + 0.116, + 0.577, + 0.49, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.086, + 0.885, + 0.179 + ], + "angle": 0, + "content": "Jiaao Chen and Diyi Yang. 2021. Simple conversational data augmentation for semi-supervised abstractive dialogue summarization. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6605-6616, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.189, + 0.885, + 0.751 + ], + "angle": 0, + "content": "Kaustubh D. Dhole, Varun Gangal, Sebastian Gehrmann, Aadesh Gupta, Zhenhao Li, Saad Mahamood, Abinaya Mahendiran, Simon Mille, Ashish Srivastava, Samson Tan, Tongshuang Wu, Jascha Sohl-Dickstein, Jinho D. Choi, Eduard Hovy, Ondrej Dusek, Sebastian Ruder, Sajant Anand, Nagender Aneja, Rabin Banjade, Lisa Barthe, Hanna Behnke, Ian Berlot-Attwell, Connor Boyle, Caroline Brun, Marco Antonio Sobrevilla Cabezudo, Samuel Cahyawijaya, Emile Chapuis, Wanxiang Che, Mukund Choudhary, Christian Clauss, Pierre Colombo, Filip Cornell, Gautier Dagan, Mayukh Das, Tanay Dixit, Thomas Dopierre, Paul-Alexis Dray, Suchitra Dubey, Tatiana Ekeinhor, Marco Di Giovanni, Rishabh Gupta, Rishabh Gupta, Louanes Hamla, Sang Han, Fabrice Harel-Canada, Antoine Honore, Ishan Jindal, Przemyslaw K. Joniak, Denis Kleyko, Venelin Kovatchev, Kalpesh Krishna, Ashutosh Kumar, Stefan Langer, Seungjae Ryan Lee, Corey James Levinson, Hualou Liang, Kaizhao Liang, Zhexiong Liu, Andrey Lukyanenko, Vukosi Marivate, Gerard de Melo, Simon Meoni, Maxime Meyer, Afnan Mir, Nafise Sadat Moosavi, Niklas Muennighoff, Timothy Sum Hon Mun, Kenton Murray, Marcin Namysl, Maria Obedkova, Priti Oli, Nivranshu Pasricha, Jan Pfister, Richard Plant, Vinay Prabhu, Vasile Pais, Libo Qin, Shahab Raji, Pawan Kumar Rajpoot, Vikas Raunak, Roy Rinberg, Nicolas Roberts, Juan Diego Rodriguez, Claude Roux, Vasconcellos P. H. S., Ananya B. Sai, Robin M. Schmidt, Thomas Scialom, Tshephisho Sefara, Saqib N. Shamsi, Xudong Shen, Haoyue Shi, Yiwen Shi, Anna Shvets, Nick Siegel, Damien Sileo, Jamie Simon, Chandan Singh, Roman Sitelew, Priyank Soni, Taylor Sorensen, William Soto, Aman Srivastava, KV Aditya Srivatsa, Tony Sun, Mukund Varma T, A Tabassum, Fiona Anting Tan, Ryan Teehan, Mo Tiwari, Marie Tolkiehn, Athena Wang, Zijian Wang, Gloria Wang, Zijie J. Wang, Fuxuan Wei, Bryan Wilie, Genta Indra Winata, Xinyi Wu, Witold Wydmanski, Tianbao Xie, Usama Yaseen, M. Yee, Jing Zhang and Yue Zhang. 2021. Nl-augmenter: A framework for task-sensitive natural language augmentation." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.762, + 0.885, + 0.854 + ], + "angle": 0, + "content": "Fangxiaoyu Feng, Yinfei Yang, Daniel Cer, Naveen Arivazhagan, and Wei Wang. 2022. Language-agnostic BERT sentence embedding. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 878–891, Dublin, Ireland. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.866, + 0.885, + 0.918 + ], + "angle": 0, + "content": "Kevin Heffernan, Onur Celebi, and Holger Schwenk. 2022. Bitext mining using distilled sentence representations for low-resource languages. Findings of EMNLP." + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.086, + 0.885, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.929, + 0.516, + 0.941 + ], + "angle": 0, + "content": "105" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.086, + 0.487, + 0.139 + ], + "angle": 0, + "content": "Matthew Honnibal and Ines Montani. 2017. spaCy 2: Natural language understanding with Bloom embeddings, convolutional neural networks and incremental parsing. To appear." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.153, + 0.487, + 0.257 + ], + "angle": 0, + "content": "Sosuke Kobayashi. 2018. Contextual augmentation: Data augmentation by words with paradigmatic relations. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 452-457, New Orleans, Louisiana. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.272, + 0.487, + 0.363 + ], + "angle": 0, + "content": "Tom Kocmi, Christian Federmann, Roman Grundkiewicz, Marcin Junczys-Dowmunt, Hitokazu Matsushita, and Arul Menezes. 2021. To ship or not to ship: An extensive evaluation of automatic metrics for machine translation. In Proceedings of the Sixth Conference on Machine Translation, pages 478-494, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.377, + 0.487, + 0.442 + ], + "angle": 0, + "content": "Philipp Koehn. 2004. Statistical significance tests for machine translation evaluation. In Proceedings of the 2004 Conference on Empirical Methods in Natural Language Processing, pages 388-395, Barcelona, Spain. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.457, + 0.487, + 0.522 + ], + "angle": 0, + "content": "Philipp Koehn and Rebecca Knowles. 2017. Six challenges for neural machine translation. In Proceedings of the First Workshop on Neural Machine Translation, pages 28-39, Vancouver. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.536, + 0.487, + 0.615 + ], + "angle": 0, + "content": "Tong Niu and Mohit Bansal. 2018. Adversarial over-sensitivity and over-stability strategies for dialogue models. In Proceedings of the 22nd Conference on Computational Natural Language Learning, pages 486-496, Brussels, Belgium. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.629, + 0.487, + 0.707 + ], + "angle": 0, + "content": "NLLB Team, Marta R Costa-jussa, James Cross, Onur Celebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, et al. 2022. No language left behind: Scaling human-centered machine translation. arXiv preprint arXiv:2207.04672." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.721, + 0.487, + 0.8 + ], + "angle": 0, + "content": "Gözde Gül Şahin and Mark Steedman. 2018. Data augmentation via dependency tree morphing for low-resource languages. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 5004-5009, Brussels, Belgium. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.814, + 0.487, + 0.918 + ], + "angle": 0, + "content": "Holger Schwenk, Vishrav Chaudhary, Shuo Sun, Hongyu Gong, and Francisco Guzmán. 2021a. WikiMatrix: Mining 135M parallel sentences in 1620 language pairs from Wikipedia. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 1351-1361, Online. Association for Computational Linguistics." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.487, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.086, + 0.882, + 0.203 + ], + "angle": 0, + "content": "Holger Schwenk, Guillaume Wenzek, Sergey Edunov, Edouard Grave, Armand Joulin, and Angela Fan. 2021b. CCMatrix: Mining billions of high-quality parallel sentences on the web. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 6490-6500, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.214, + 0.882, + 0.306 + ], + "angle": 0, + "content": "Fiona Anting Tan, Devamanyu Hazarika, See-Kiong Ng, Soujanya Poria, and Roger Zimmermann. 2021. Causal augmentation for causal sentence classification. In Proceedings of the First Workshop on Causal Inference and NLP, pages 1-20, Punta Cana, Dominican Republic. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.316, + 0.882, + 0.433 + ], + "angle": 0, + "content": "Zilu Tang, Muhammed Yusuf Kocyigit, and Derry Tanti Wijaya. 2022. AugCSE: Contrastive sentence embedding with diverse augmentations. In Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 375-398, Online only. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.444, + 0.882, + 0.548 + ], + "angle": 0, + "content": "Jason Wei and Kai Zou. 2019. EDA: Easy data augmentation techniques for boosting performance on text classification tasks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 6382-6388, Hong Kong, China. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.559, + 0.882, + 0.625 + ], + "angle": 0, + "content": "Pierre Zweigenbaum, Serge Sharoff, and Reinhard Rapp. 2018. Overview of the third bucc shared task: Spotting parallel sentences in comparable corpora. In Proceedings of 11th workshop on building and using comparable corpora, pages 39-42." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.882, + 0.625 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.489, + 0.929, + 0.516, + 0.94 + ], + "angle": 0, + "content": "106" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.115, + 0.084, + 0.416, + 0.116 + ], + "angle": 0, + "content": "A Data Statistics for xsim++ with FLORES200 devtest set" + }, + { + "type": "table", + "bbox": [ + 0.188, + 0.136, + 0.416, + 0.214 + ], + "angle": 0, + "content": "
Total ##per orig.
Original1012-
Causality19161.89
Entity3885538.39
Number32623.22
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.223, + 0.49, + 0.281 + ], + "angle": 0, + "content": "Table 5: Total numbers of original sentences and transformed sentences in different transformation categories. We also report the averaged numbers of transformations per original sentence for each category." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.3, + 0.489, + 0.332 + ], + "angle": 0, + "content": "We report the data statistics for \\( xsim++ \\) with FLORES200 devtest set in Table 5." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.345, + 0.407, + 0.379 + ], + "angle": 0, + "content": "B Sizes of Monolingual data for Low-Resource Languages" + }, + { + "type": "table", + "bbox": [ + 0.212, + 0.4, + 0.393, + 0.542 + ], + "angle": 0, + "content": "
LanguageSize
kik147,902
kea226,507
fur737,178
fao1,179,475
tpi1,661,743
bem2,302,805
ibo8,124,418
zul20,477,331
swh55,399,821
ltz123,944,670
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.551, + 0.489, + 0.581 + ], + "angle": 0, + "content": "Table 6: Number of monolingual sentences for each language." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.598, + 0.487, + 0.629 + ], + "angle": 0, + "content": "We report the sizes of monolingual data for each language in Table 6." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.643, + 0.461, + 0.66 + ], + "angle": 0, + "content": "C Hyperparameters for NMT systems" + }, + { + "type": "table", + "bbox": [ + 0.169, + 0.681, + 0.437, + 0.846 + ], + "angle": 0, + "content": "
encoder layers6
encoder attention heads8
encoder embed dim512
encoder FFNN embed dim4096
decoder layers6
decoder attention heads8
decoder embed dim512
decoder FFNN embed dim4096
optimiserAdam
adam betas(0.9, 0.98)
learning rate0.001
dropout0.3
spm vocab size7000
" + }, + { + "type": "table_caption", + "bbox": [ + 0.147, + 0.856, + 0.454, + 0.871 + ], + "angle": 0, + "content": "Table 7: Hyperparameters for NMT systems." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.888, + 0.489, + 0.918 + ], + "angle": 0, + "content": "We report hyperparameters for NMT evaluations in Table 7." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.084, + 0.87, + 0.1 + ], + "angle": 0, + "content": "D Within and Across Model Accuracies" + }, + { + "type": "table", + "bbox": [ + 0.599, + 0.118, + 0.795, + 0.165 + ], + "angle": 0, + "content": "
MetricWithinAcross
xsim31.3354.04
xsim++69.77*82.00*
" + }, + { + "type": "table_caption", + "bbox": [ + 0.508, + 0.174, + 0.884, + 0.245 + ], + "angle": 0, + "content": "Table 8: Pairwise ranking accuracy for comparisons within a model and across different models. An * indicates that the result passes the significance test proposed by Koehn (2004) with \\( p \\)-value \\( < 0.05 \\) when compared to xsim." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.265, + 0.884, + 0.313 + ], + "angle": 0, + "content": "We report accuracies for within a model (i.e., LASER) and across different models (i.e., the 10 LASER checkpoints vs LaBSE) in Table 8." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.929, + 0.516, + 0.941 + ], + "angle": 0, + "content": "107" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.134, + 0.085, + 0.435, + 0.1 + ], + "angle": 0, + "content": "ACL 2023 Responsible NLP Checklist" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.108, + 0.323, + 0.123 + ], + "angle": 0, + "content": "A For every submission:" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.127, + 0.533, + 0.144 + ], + "angle": 0, + "content": "A1. Did you describe the limitations of your work?" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.145, + 0.233, + 0.16 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.171, + 0.553, + 0.187 + ], + "angle": 0, + "content": "A2. Did you discuss any potential risks of your work?" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.188, + 0.351, + 0.202 + ], + "angle": 0, + "content": "Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.213, + 0.696, + 0.23 + ], + "angle": 0, + "content": "A3. Do the abstract and introduction summarize the paper's main claims?" + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.231, + 0.233, + 0.246 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.256, + 0.669, + 0.273 + ], + "angle": 0, + "content": "A4. Have you used AI writing assistants when working on this paper?" + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.274, + 0.233, + 0.288 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.3, + 0.489, + 0.317 + ], + "angle": 0, + "content": "B Did you use or create scientific artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.322, + 0.206, + 0.336 + ], + "angle": 0, + "content": "Section 2" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.347, + 0.53, + 0.364 + ], + "angle": 0, + "content": "B1. Did you cite the creators of artifacts you used?" + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.365, + 0.223, + 0.378 + ], + "angle": 0, + "content": "Section 2" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.391, + 0.779, + 0.407 + ], + "angle": 0, + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.408, + 0.351, + 0.423 + ], + "angle": 0, + "content": "Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.434, + 0.882, + 0.497 + ], + "angle": 0, + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.499, + 0.351, + 0.514 + ], + "angle": 0, + "content": "Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.525, + 0.882, + 0.573 + ], + "angle": 0, + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.574, + 0.351, + 0.589 + ], + "angle": 0, + "content": "Not applicable. Left blank." + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.599, + 0.882, + 0.632 + ], + "angle": 0, + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?" + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.633, + 0.223, + 0.647 + ], + "angle": 0, + "content": "Section 2" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.658, + 0.882, + 0.74 + ], + "angle": 0, + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be." + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.741, + 0.223, + 0.754 + ], + "angle": 0, + "content": "Section 2" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.765, + 0.494, + 0.782 + ], + "angle": 0, + "content": "C Did you run computational experiments?" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.788, + 0.206, + 0.802 + ], + "angle": 0, + "content": "Section 3" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.813, + 0.882, + 0.846 + ], + "angle": 0, + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?" + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.847, + 0.223, + 0.86 + ], + "angle": 0, + "content": "Section 3" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.867, + 0.878, + 0.89 + ], + "angle": 0, + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.929, + 0.516, + 0.941 + ], + "angle": 0, + "content": "108" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.13, + 0.084, + 0.88, + 0.133 + ], + "angle": 0, + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Section 3 and Appendix" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.143, + 0.88, + 0.208 + ], + "angle": 0, + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Section 3 and Appendix" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.218, + 0.88, + 0.281 + ], + "angle": 0, + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Section 3" + }, + { + "type": "list", + "bbox": [ + 0.13, + 0.084, + 0.88, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.294, + 0.878, + 0.311 + ], + "angle": 0, + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.315, + 0.215, + 0.33 + ], + "angle": 0, + "content": "Left blank." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.342, + 0.88, + 0.388 + ], + "angle": 0, + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.401, + 0.88, + 0.463 + ], + "angle": 0, + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.476, + 0.88, + 0.539 + ], + "angle": 0, + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.551, + 0.873, + 0.581 + ], + "angle": 0, + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.593, + 0.88, + 0.641 + ], + "angle": 0, + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response." + }, + { + "type": "list", + "bbox": [ + 0.129, + 0.342, + 0.88, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.929, + 0.516, + 0.941 + ], + "angle": 0, + "content": "109" + } + ] +] \ No newline at end of file diff --git a/2023/xSIM++_ An Improved Proxy to Bitext Mining Performance for Low-Resource Languages/632d8eaf-c025-45c6-9916-28d9ca6a54ad_origin.pdf b/2023/xSIM++_ An Improved Proxy to Bitext Mining Performance for Low-Resource Languages/632d8eaf-c025-45c6-9916-28d9ca6a54ad_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7501b642e161993bf970af7d2613052d58c84b86 --- /dev/null +++ b/2023/xSIM++_ An Improved Proxy to Bitext Mining Performance for Low-Resource Languages/632d8eaf-c025-45c6-9916-28d9ca6a54ad_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:efd47ded50746a2668df188f9386468d60bf6cb9d84ba459ac55e0e4a8634d99 +size 253326 diff --git a/2023/xSIM++_ An Improved Proxy to Bitext Mining Performance for Low-Resource Languages/full.md b/2023/xSIM++_ An Improved Proxy to Bitext Mining Performance for Low-Resource Languages/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ce8adcac5e13017363a87148225834803ac9aeef --- /dev/null +++ b/2023/xSIM++_ An Improved Proxy to Bitext Mining Performance for Low-Resource Languages/full.md @@ -0,0 +1,250 @@ +# xSIM++: An Improved Proxy to Bitext Mining Performance for Low-Resource Languages + +Mingda Chen*, Kevin Heffernan*, Onur Celebi, Alex Mourachko, Holger Schwenk + +{mingdachen,kevinheffernan,celebio,alexmourachko,schwenk}@meta.com + +Meta AI Research + +# Abstract + +We introduce a new proxy score for evaluating bitext mining based on similarity in a multilingual embedding space: $xsim++$ . In comparison to $xsim$ , this improved proxy leverages rule-based approaches to extend English sentences in any evaluation set with synthetic, hard-to-distinguish examples which more closely mirror the scenarios we encounter during large-scale mining. We validate this proxy by running a significant number of bitext mining experiments for a set of low-resource languages, and subsequently train NMT systems on the mined data. In comparison to $xsim$ , we show that $xsim++$ is better correlated with the downstream BLEU scores of translation systems trained on mined bitexts, providing a reliable proxy of bitext mining performance without needing to run expensive bitext mining pipelines. $xsim++$ also reports performance for different error types, offering more fine-grained feedback for model development. + +# 1 Introduction + +When training neural machine translation (NMT) systems, it has been shown in prior works that generally, the quality of such systems increases with the availability of high-quality training data (Koehn and Knowles, 2017). However, for many low-resource languages there are few public corpora available, posing many challenges. In order to address this sparsity, one approach is to supplement existing datasets with automatically created parallel corpora, and a technique which has shown to be successful for such issues is the task of bitext mining (Schwenk et al., 2021b). + +In bitext mining, the aim is to find pairs of sentences with the same sentence meaning across collections of monolingual corpora. In this work, we adopt a global mining approach (Schwenk et al., 2021a), which has shown recent success in provid + +ing high-quality data for low-resourced languages (NLLB Team et al., 2022). + +In order to evaluate any bitext mining method, a natural approach is to train a NMT system on the automatically created alignments. However, this is extremely costly. As an alternative, the BUCC task (Zweigenbaum et al., 2018) offers a method for evaluating bitext mining algorithms by embedding known alignments within monolingual corpora, and then reporting on the number of correctly aligned pairs. However, this task currently only covers 5 high-resourced languages (English, French, Russian, German and Chinese), and so is not applicable to the low-resource domain. In order to address this, another approach to evaluate bitext mining is to align existing multilingual parallel test sets. Two such test sets are Tatoeba1 and FLORES200.2 However, as shown by Heffernan et al. (2022), the Tatoeba corpus is not very reliable given that for some sentence pairs there are only a few hundred sentences. Therefore, we opt to use FLORES200, which is also n-way parallel. + +One existing method for evaluating bitext mining on parallel test sets is xsim. This method reports the error rate of misaligned sentences, and follows a margin-based global mining approach (Artetxe and Schwenk, 2019a). However, although using xsim on test sets such as FLORES200 has been shown to be useful as a proxy metric for bitext mining (NLLB Team et al., 2022), it has the following limitations: + +1. Using FLORES200 alone has proven to not be difficult enough as for many language pairs, existing approaches quickly saturate at $0\%$ error (NLLB Team et al., 2022). + +
Transformation CategoryOriginal SentenceTransformed Sentence
Causality AlternationApart from the fever and a sore throat, I feel well and in good shape to carry out my work by telecommuting.Apart from the fever and a sore throat, I feel well and in bad shape to carry out my work by telecommuting
Entity ReplacementCharles was the first member of the British Royal Family to be awarded a degree.M. Smith was the first member of The University to be awarded a degree.
Number ReplacementNadal bagged 88% net points in the match winning 76 points in the first serve.Nadal bagged 98% net points in the match winning 71 points in the sixth serve.
+ +Table 1: Examples of the transformations applied to the English sentences from FLORES200 dev set. The red texts indicate the places of alternations. + +2. As the dev and devtest sets are quite small (997/1012 respectively), this is arguably not a good approximation for performance when mining against billions of possible candidate sentences. +3. We have observed that there is not a significant overlap in the semantics between candidate sentences, meaning that it is not possible to test difficult scenarios that arise in bitext mining when choosing between multiple (similar) candidate pairs. + +In order to address these limitations, in this work we introduce $x \sin^{++}$ . This is an improved proxy for bitext mining performance which expands the dev and devtest sets of FLORES200 to include both more data points, and also difficult to distinguish cases which provide far greater challenges to the models. Our contributions can be summarised as follows: + +1. We create a more semantically challenging and expanded English test set for FLORES200. +2. We validate this new test set by independently performing 110 bitext mining runs, training 110 NMT systems on the output mined bittexts, and then determining both the correlation and statistical significance between $x \sin + +$ and the resulting BLEU scores. +3. We open-source the expanded FLORES200 dev and devtest sets, and also the xsim++ code to evaluate them4. + +# 2 Methodology + +# 2.1 Background: xsim + +Given two lists of sentences in different languages, xsim seeks to align each sentence in the source + +language to a corresponding sentence in the target language based on a margin-based similarity (Artetxe and Schwenk, 2019a). In doing so, xsim leverages the mining approach described in Artetxe and Schwenk (2019b) to first encode sentences into embedding vectors, assign pairwise scores between sentences in the lists, and then take the sentence in the target language that achieves the maximum score as the final prediction. xsim relies on human-annotated parallel corpora and measures the performance of bitext mining using the fraction of misaligned source sentences, i.e., error rates. + +# 2.2 xsim++ + +As the effectiveness of xsim is limited by the availability of parallel corpora, we choose to create xsim++ by automatically expanding the English sentences, and evaluate the sentence encoders on into-English language directions, following prior work on low-resource bitext mining (Heffernan et al., 2022). Aside from the expanded candidate set, xsim++ follows the same procedure as xsim. + +$\times \mathrm{sim} + +$ seeks to capture more subtle improvements in bitext mining by adding challenging negative examples. The examples are human-written sentences transformed by various operations. These operations intend to perturb semantics through minimal alternations in the surface text. In particular, we use the following categories of transformations: causality alternation, entity replacement, and number replacement. We focus on these three transformation types only as they easily allow us to create negative examples. Examples of the transformed sentences are shown in Table 1. For these transformations, we adapt the implementation in Dhole et al. (2021) $^6$ and describe the details + +
Total ## per orig.
Original997-
Causality18681.87
Entity3774537.86
Number34763.49
+ +Table 2: Total numbers of original sentences and transformed sentences in different transformation categories. We also report the averaged numbers of transformations per original sentence for each category. + +of these transformations below. + +Causality Alternation. To alter causality in a sentence, we (1) replace adjectives with their antonyms; (2) negate the meaning of sentences by adding or removing negation function words (e.g. "did not" and "was not") to the sentences; or (3) leverage the negation strengthening approach (Tan et al., 2021), which changes the causal relationships through more assertive function words (e.g. replacing "may" with "will"). For example, as shown in Table 1 we replace "good" with the antonym "bad". + +Entity Replacement. We collect candidate entities from large amounts of monolingual data. Then we replace entities in sentences with the ones randomly sampled from the candidate set. For both stages, we use the named entity recognizer from NLTK (Bird et al., 2009). + +Number Replacement. We use spaCy (Honni-bal and Montani, 2017) to detect dates, ordinals, cardinals, times, numbers, and percentages and then randomly replace their values. + +Given the strategies above, for each sentence we create multiple transformations (i.e. negative examples) of that source sentence. For example, consider Table 1. In the "Entity Replacement" example we create a transformation by replacing two named entities. We can then continue this process by replacing these with other named entities until we have reached the desired number of total transformations7. Note that since the opportunity to change each category is dependent on the frequency of that category in the evaluation sets, some transformations occurred more than others (e.g. entities were more frequent than numbers). We summarize the data statistics for xsim++ on the FLORES200 dev + +set in Table 2. Results for the devtest set are in appendix A. + +# 3 Experiment + +In order to establish $x$ sim++ as a proxy for bitext mining performance, we measure the correlation between both $x$ sim and $x$ sim++ error rates, and the BLEU scores resulting from NMT systems trained on mined bittexts. More specifically, for each language we choose a sentence encoder model, followed by bitext mining using each respective encoder, and then train and evaluate bilingual NMT systems on the resulting mined bittexts. We use the FLORES200 development sets when computing the BLEU scores. + +In order to validate $\times$ sim++ against varied embedding spaces, we encode (and mine) using two different multilingual encoder methods: LASER (Artetxe and Schwenk, 2019b) and LaBSE (Feng et al., 2022). For LASER, we trained our own custom encoders (details below). For LaBSE, we used a publicly available model as the code and data for training LaBSE are not publicly available. + +We randomly choose 10 low-resource languages to perform both encoder training (if applicable) and bitext mining. The languages are: Faroese (fao), Kabuverdianu (kea), Tok Pisin (tpi), Kikuyu (kik), Friulian (fur), Igbo (ibo), Luxembourgish (ltz), Swahili (swh), Zulu (zul), Bemba (bem). + +Encoder Training. We trained LASER encoders using the teacher-student approach described in Heffernan et al. (2022). We choose a LASER model (Artetxe and Schwenk, 2019b) as our teacher, and then trained specialised students for each language. In order to train each student, we used both publicly available code9 and bitexts (e.g. OPUS10) + +Bitext Mining. For each chosen encoder model, we perform bitext mining against approximately 3.7 billion sentences of English. For low-resource languages, the sizes of monolingual data range from 140k to 124 million. Details are in the appendix. We make use of monolingual data available from both Commoncrawl and Paracrawl11, and operationalize the mining using the stopes library (An + +drews et al., 2022).12 For LASER, we use 1.06 as the margin threshold following Heffernan et al. (2022) and for LaBSE, we use 1.16.13 Following mining, for each language we concatenate publicly available bittexts and the mined bitext as training data for NMT bilingual models using fairseq,14 translating from each foreign text into English. For all NMT systems, we keep the hyperparameters fixed (details in Appendix). + +Evaluation. Model selection involves two use cases: comparisons within a model and across different models. For the former comparison, given our custom encoders, we choose to compare 10 checkpoints from each model.[15] For cross model comparisons, we compare each chosen encoder checkpoint against another existing system. In this case, the LaBSE encoder. To quantitatively measure these two cases, we report pairwise ranking accuracy (Kocmi et al., 2021) for xsim and xsim++. Formally, the accuracy is computed as follows + +$$ +\frac {\left| \mathrm {s} (\text {p r o x y} \Delta) = \mathrm {s} (\text {m i n i n g} \Delta) \text {f o r a l l s y s t e m p a i r s} \right|}{\left| \text {a l l s y s t e m p a i r s} \right|} +$$ + +where proxy $\Delta$ is the difference of the $x \sin$ or $x \sin + +$ scores, mining $\Delta$ is the difference of the BLEU scores, $s(\cdot)$ is the sign function, and $|\cdot|$ returns the cardinal number of the input. + +In this work, we have 550 system pairs with 55 pairs per language direction (i.e. $\binom{11}{2}$ pairs given 10 custom LASER encoder checkpoints + LaBSE). We always compare systems within a language direction as the scores for system pairs across different directions are not comparable.[16] + +# 3.1 Results + +As shown in Table 3, xsim++ significantly outperforms xsim on the pairwise ranking accuracy. Additionally, when comparing the computational cost to mining, xsim++ costs over $99.9\%$ less GPU hours and saves approximately 3 metric tons of carbon + +
MetricAccuracyGPU hours
xsim35.480.43
xsim++72.00*0.52
Mining BLEU(Oracle)10019569
+ +Table 3: Pairwise ranking accuracy along with the total number of GPU hours. For all experiments, we used NVIDIA A100 GPUs. An * indicates that the result passes the significance test proposed by Koehn (2004) with $p$ -value $< 0.05$ when compared to xsim. + +
Accuracy
xsim++72.00
Causality63.09
Entity65.45
Number60.73
Misaligned40.73
Causality + Entity68.55
Causality + Entity + Misaligned70.55
Causality + Misaligned68.00
Causality + Number66.73
Causality + Number + Misaligned71.45
Entity + Misaligned70.55
Number + Entity67.45
Number + Entity + Misaligned71.09
Number + Misaligned64.36
+ +Table 4: Pairwise ranking accuracy when using combinations of error categories. Causality=Causality Alternation, Entity=Entity Replacement, Number=Number Replacement. + +emissions, but still manages to achieve a competitive accuracy. We observe similar trends for the within a model and across models use cases and report their separate accuracies in the appendix. + +To better understand the contributions of each transformation category (cf. subsection 2.1) in measuring the final mining performance, we report accuracies for different combinations of categories in Table 4. In cases where an incorrect bitext alignment do does not map to any of the augmented sentences of the true alignment, we denote these as "misaligned". We find that entity replacement helps most in improving the accuracy and combing all the transformations gives the best performance. + +# 4 Related Work + +As $xsim++$ uses rule-based data augmentation, it is related to work in other areas that also employ similar data augmentation methods, such as part-of-speech tagging (Şahin and Steedman, 2018), contrastive learning (Tang et al., 2022), text classification (Kobayashi, 2018; Wei and Zou, 2019), dialogue generation (Niu and Bansal, 2018) and summarization (Chen and Yang, 2021). + +# 5 Conclusion and Future Work + +We proposed a proxy score $x \sin + +$ for bitext mining performance using three kinds of data augmentation techniques: causality alternation, entity replacement, and number replacement. To validate its effectiveness, we conducted large-scale bitext mining experiments for 10 low-resource languages, and reported pairwise ranking accuracies. We found that $x \sin + +$ significantly improves over $x \sin$ , doubling the accuracies. Analysis reveals that entity replacement helps most in the improvement. In future work, we plan to extend $x \sin + +$ to non-English language pairs. + +# 6 Limitations + +We highlight three limitations of our work. The first is that $xsim++$ is automatically constructed. There could be noisy sentences leading to errors that are irrelevant to the quality of encoders. The second is that $xsim++$ applies transformations solely to English sentences. Generalizing it to non-English language pairs requires additional research. Finally, we have experimented with the two most popular multilingual encoders: LASER and LaBSE. There are other available approaches which would be interesting to also validate $xsim++$ against. + +# References + +Pierre Andrews, Guillaume Wenzek, Kevin Heffernan, Onur Celebi, Anna Sun, Ammar Kamran, Yingzhe Guo, Alexandre Mourachko, Holger Schwenk, and Angela Fan. 2022. stopes - modular machine translation pipelines. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing: System Demonstrations. Association for Computational Linguistics. +Mikel Artetxe and Holger Schwenk. 2019a. Margin-based parallel corpus mining with multilingual sentence embeddings. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3197-3203, Florence, Italy. Association for Computational Linguistics. +Mikel Artetxe and Holger Schwenk. 2019b. Massively multilingual sentence embeddings for zero-shot cross-lingual transfer and beyond. Transactions of the Association for Computational Linguistics, 7:597-610. +Steven Bird, Ewan Klein, and Edward Loper. 2009. Natural language processing with Python: analyzing text with the natural language toolkit. "O'Reilly Media, Inc." + +Jiaao Chen and Diyi Yang. 2021. Simple conversational data augmentation for semi-supervised abstractive dialogue summarization. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6605-6616, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics. +Kaustubh D. Dhole, Varun Gangal, Sebastian Gehrmann, Aadesh Gupta, Zhenhao Li, Saad Mahamood, Abinaya Mahendiran, Simon Mille, Ashish Srivastava, Samson Tan, Tongshuang Wu, Jascha Sohl-Dickstein, Jinho D. Choi, Eduard Hovy, Ondrej Dusek, Sebastian Ruder, Sajant Anand, Nagender Aneja, Rabin Banjade, Lisa Barthe, Hanna Behnke, Ian Berlot-Attwell, Connor Boyle, Caroline Brun, Marco Antonio Sobrevilla Cabezudo, Samuel Cahyawijaya, Emile Chapuis, Wanxiang Che, Mukund Choudhary, Christian Clauss, Pierre Colombo, Filip Cornell, Gautier Dagan, Mayukh Das, Tanay Dixit, Thomas Dopierre, Paul-Alexis Dray, Suchitra Dubey, Tatiana Ekeinhor, Marco Di Giovanni, Rishabh Gupta, Rishabh Gupta, Louanes Hamla, Sang Han, Fabrice Harel-Canada, Antoine Honore, Ishan Jindal, Przemyslaw K. Joniak, Denis Kleyko, Venelin Kovatchev, Kalpesh Krishna, Ashutosh Kumar, Stefan Langer, Seungjae Ryan Lee, Corey James Levinson, Hualou Liang, Kaizhao Liang, Zhexiong Liu, Andrey Lukyanenko, Vukosi Marivate, Gerard de Melo, Simon Meoni, Maxime Meyer, Afnan Mir, Nafise Sadat Moosavi, Niklas Muennighoff, Timothy Sum Hon Mun, Kenton Murray, Marcin Namysl, Maria Obedkova, Priti Oli, Nivranshu Pasricha, Jan Pfister, Richard Plant, Vinay Prabhu, Vasile Pais, Libo Qin, Shahab Raji, Pawan Kumar Rajpoot, Vikas Raunak, Roy Rinberg, Nicolas Roberts, Juan Diego Rodriguez, Claude Roux, Vasconcellos P. H. S., Ananya B. Sai, Robin M. Schmidt, Thomas Scialom, Tshephisho Sefara, Saqib N. Shamsi, Xudong Shen, Haoyue Shi, Yiwen Shi, Anna Shvets, Nick Siegel, Damien Sileo, Jamie Simon, Chandan Singh, Roman Sitelew, Priyank Soni, Taylor Sorensen, William Soto, Aman Srivastava, KV Aditya Srivatsa, Tony Sun, Mukund Varma T, A Tabassum, Fiona Anting Tan, Ryan Teehan, Mo Tiwari, Marie Tolkiehn, Athena Wang, Zijian Wang, Gloria Wang, Zijie J. Wang, Fuxuan Wei, Bryan Wilie, Genta Indra Winata, Xinyi Wu, Witold Wydmanski, Tianbao Xie, Usama Yaseen, M. Yee, Jing Zhang and Yue Zhang. 2021. Nl-augmenter: A framework for task-sensitive natural language augmentation. +Fangxiaoyu Feng, Yinfei Yang, Daniel Cer, Naveen Arivazhagan, and Wei Wang. 2022. Language-agnostic BERT sentence embedding. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 878–891, Dublin, Ireland. Association for Computational Linguistics. +Kevin Heffernan, Onur Celebi, and Holger Schwenk. 2022. Bitext mining using distilled sentence representations for low-resource languages. Findings of EMNLP. + +Matthew Honnibal and Ines Montani. 2017. spaCy 2: Natural language understanding with Bloom embeddings, convolutional neural networks and incremental parsing. To appear. +Sosuke Kobayashi. 2018. Contextual augmentation: Data augmentation by words with paradigmatic relations. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 452-457, New Orleans, Louisiana. Association for Computational Linguistics. +Tom Kocmi, Christian Federmann, Roman Grundkiewicz, Marcin Junczys-Dowmunt, Hitokazu Matsushita, and Arul Menezes. 2021. To ship or not to ship: An extensive evaluation of automatic metrics for machine translation. In Proceedings of the Sixth Conference on Machine Translation, pages 478-494, Online. Association for Computational Linguistics. +Philipp Koehn. 2004. Statistical significance tests for machine translation evaluation. In Proceedings of the 2004 Conference on Empirical Methods in Natural Language Processing, pages 388-395, Barcelona, Spain. Association for Computational Linguistics. +Philipp Koehn and Rebecca Knowles. 2017. Six challenges for neural machine translation. In Proceedings of the First Workshop on Neural Machine Translation, pages 28-39, Vancouver. Association for Computational Linguistics. +Tong Niu and Mohit Bansal. 2018. Adversarial over-sensitivity and over-stability strategies for dialogue models. In Proceedings of the 22nd Conference on Computational Natural Language Learning, pages 486-496, Brussels, Belgium. Association for Computational Linguistics. +NLLB Team, Marta R Costa-jussa, James Cross, Onur Celebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, et al. 2022. No language left behind: Scaling human-centered machine translation. arXiv preprint arXiv:2207.04672. +Gözde Gül Şahin and Mark Steedman. 2018. Data augmentation via dependency tree morphing for low-resource languages. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 5004-5009, Brussels, Belgium. Association for Computational Linguistics. +Holger Schwenk, Vishrav Chaudhary, Shuo Sun, Hongyu Gong, and Francisco Guzmán. 2021a. WikiMatrix: Mining 135M parallel sentences in 1620 language pairs from Wikipedia. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 1351-1361, Online. Association for Computational Linguistics. + +Holger Schwenk, Guillaume Wenzek, Sergey Edunov, Edouard Grave, Armand Joulin, and Angela Fan. 2021b. CCMatrix: Mining billions of high-quality parallel sentences on the web. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 6490-6500, Online. Association for Computational Linguistics. +Fiona Anting Tan, Devamanyu Hazarika, See-Kiong Ng, Soujanya Poria, and Roger Zimmermann. 2021. Causal augmentation for causal sentence classification. In Proceedings of the First Workshop on Causal Inference and NLP, pages 1-20, Punta Cana, Dominican Republic. Association for Computational Linguistics. +Zilu Tang, Muhammed Yusuf Kocyigit, and Derry Tanti Wijaya. 2022. AugCSE: Contrastive sentence embedding with diverse augmentations. In Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 375-398, Online only. Association for Computational Linguistics. +Jason Wei and Kai Zou. 2019. EDA: Easy data augmentation techniques for boosting performance on text classification tasks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 6382-6388, Hong Kong, China. Association for Computational Linguistics. +Pierre Zweigenbaum, Serge Sharoff, and Reinhard Rapp. 2018. Overview of the third bucc shared task: Spotting parallel sentences in comparable corpora. In Proceedings of 11th workshop on building and using comparable corpora, pages 39-42. + +# A Data Statistics for xsim++ with FLORES200 devtest set + +
Total ##per orig.
Original1012-
Causality19161.89
Entity3885538.39
Number32623.22
+ +We report the data statistics for $xsim++$ with FLORES200 devtest set in Table 5. + +# B Sizes of Monolingual data for Low-Resource Languages + +Table 5: Total numbers of original sentences and transformed sentences in different transformation categories. We also report the averaged numbers of transformations per original sentence for each category. + +
LanguageSize
kik147,902
kea226,507
fur737,178
fao1,179,475
tpi1,661,743
bem2,302,805
ibo8,124,418
zul20,477,331
swh55,399,821
ltz123,944,670
+ +We report the sizes of monolingual data for each language in Table 6. + +# C Hyperparameters for NMT systems + +Table 6: Number of monolingual sentences for each language. + +
encoder layers6
encoder attention heads8
encoder embed dim512
encoder FFNN embed dim4096
decoder layers6
decoder attention heads8
decoder embed dim512
decoder FFNN embed dim4096
optimiserAdam
adam betas(0.9, 0.98)
learning rate0.001
dropout0.3
spm vocab size7000
+ +We report hyperparameters for NMT evaluations in Table 7. + +# D Within and Across Model Accuracies + +Table 7: Hyperparameters for NMT systems. + +
MetricWithinAcross
xsim31.3354.04
xsim++69.77*82.00*
+ +Table 8: Pairwise ranking accuracy for comparisons within a model and across different models. An * indicates that the result passes the significance test proposed by Koehn (2004) with $p$ -value $< 0.05$ when compared to xsim. + +We report accuracies for within a model (i.e., LASER) and across different models (i.e., the 10 LASER checkpoints vs LaBSE) in Table 8. + +A For every submission: + +A1. Did you describe the limitations of your work? + +Left blank. + +A2. Did you discuss any potential risks of your work? + +Not applicable. Left blank. + +A3. Do the abstract and introduction summarize the paper's main claims? + +Left blank. + +A4. Have you used AI writing assistants when working on this paper? + +Left blank. + +B Did you use or create scientific artifacts? + +Section 2 + +B1. Did you cite the creators of artifacts you used? + +Section 2 + +B2. Did you discuss the license or terms for use and / or distribution of any artifacts? + +Not applicable. Left blank. + +B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? + +Not applicable. Left blank. + +B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? + +Not applicable. Left blank. + +B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? + +Section 2 + +B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. + +Section 2 + +C Did you run computational experiments? + +Section 3 + +C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? + +Section 3 + +The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance. + +C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Section 3 and Appendix +C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Section 3 and Appendix +C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Section 3 + +# D Did you use human annotators (e.g., crowdworkers) or research with human participants? + +Left blank. + +D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response. +D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response. +D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response. +D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response. +D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response. \ No newline at end of file diff --git a/2023/xSIM++_ An Improved Proxy to Bitext Mining Performance for Low-Resource Languages/images.zip b/2023/xSIM++_ An Improved Proxy to Bitext Mining Performance for Low-Resource Languages/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3d1e5a53876e7b7ed4178098953eae9ff93ab67c --- /dev/null +++ b/2023/xSIM++_ An Improved Proxy to Bitext Mining Performance for Low-Resource Languages/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4fade4bc87c68f61696678d508c786ed62e4ebb5e9f3b0dd742ef8fa010cc78a +size 199106 diff --git a/2023/xSIM++_ An Improved Proxy to Bitext Mining Performance for Low-Resource Languages/layout.json b/2023/xSIM++_ An Improved Proxy to Bitext Mining Performance for Low-Resource Languages/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8d30c810bac90742bdc81cbdb9a0ab0c9749d416 --- /dev/null +++ b/2023/xSIM++_ An Improved Proxy to Bitext Mining Performance for Low-Resource Languages/layout.json @@ -0,0 +1,6567 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 100, + 75, + 495, + 110 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 75, + 495, + 110 + ], + "spans": [ + { + "bbox": [ + 100, + 75, + 495, + 110 + ], + "type": "text", + "content": "xSIM++: An Improved Proxy to Bitext Mining Performance for Low-Resource Languages" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 86, + 130, + 510, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 130, + 510, + 145 + ], + "spans": [ + { + "bbox": [ + 86, + 130, + 510, + 145 + ], + "type": "text", + "content": "Mingda Chen*, Kevin Heffernan*, Onur Celebi, Alex Mourachko, Holger Schwenk" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 99, + 146, + 497, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 146, + 497, + 158 + ], + "spans": [ + { + "bbox": [ + 99, + 146, + 497, + 158 + ], + "type": "text", + "content": "{mingdachen,kevinheffernan,celebio,alexmourachko,schwenk}@meta.com" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 254, + 159, + 342, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 159, + 342, + 171 + ], + "spans": [ + { + "bbox": [ + 254, + 159, + 342, + 171 + ], + "type": "text", + "content": "Meta AI Research" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 155, + 212, + 202, + 226 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 212, + 202, + 226 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 202, + 226 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 84, + 237, + 274, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 237, + 274, + 488 + ], + "spans": [ + { + "bbox": [ + 84, + 237, + 274, + 488 + ], + "type": "text", + "content": "We introduce a new proxy score for evaluating bitext mining based on similarity in a multilingual embedding space: " + }, + { + "bbox": [ + 84, + 237, + 274, + 488 + ], + "type": "inline_equation", + "content": "xsim++" + }, + { + "bbox": [ + 84, + 237, + 274, + 488 + ], + "type": "text", + "content": ". In comparison to " + }, + { + "bbox": [ + 84, + 237, + 274, + 488 + ], + "type": "inline_equation", + "content": "xsim" + }, + { + "bbox": [ + 84, + 237, + 274, + 488 + ], + "type": "text", + "content": ", this improved proxy leverages rule-based approaches to extend English sentences in any evaluation set with synthetic, hard-to-distinguish examples which more closely mirror the scenarios we encounter during large-scale mining. We validate this proxy by running a significant number of bitext mining experiments for a set of low-resource languages, and subsequently train NMT systems on the mined data. In comparison to " + }, + { + "bbox": [ + 84, + 237, + 274, + 488 + ], + "type": "inline_equation", + "content": "xsim" + }, + { + "bbox": [ + 84, + 237, + 274, + 488 + ], + "type": "text", + "content": ", we show that " + }, + { + "bbox": [ + 84, + 237, + 274, + 488 + ], + "type": "inline_equation", + "content": "xsim++" + }, + { + "bbox": [ + 84, + 237, + 274, + 488 + ], + "type": "text", + "content": " is better correlated with the downstream BLEU scores of translation systems trained on mined bitexts, providing a reliable proxy of bitext mining performance without needing to run expensive bitext mining pipelines. " + }, + { + "bbox": [ + 84, + 237, + 274, + 488 + ], + "type": "inline_equation", + "content": "xsim++" + }, + { + "bbox": [ + 84, + 237, + 274, + 488 + ], + "type": "text", + "content": " also reports performance for different error types, offering more fine-grained feedback for model development." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 500, + 155, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 500, + 155, + 513 + ], + "spans": [ + { + "bbox": [ + 68, + 500, + 155, + 513 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 523, + 291, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 523, + 291, + 684 + ], + "spans": [ + { + "bbox": [ + 67, + 523, + 291, + 684 + ], + "type": "text", + "content": "When training neural machine translation (NMT) systems, it has been shown in prior works that generally, the quality of such systems increases with the availability of high-quality training data (Koehn and Knowles, 2017). However, for many low-resource languages there are few public corpora available, posing many challenges. In order to address this sparsity, one approach is to supplement existing datasets with automatically created parallel corpora, and a technique which has shown to be successful for such issues is the task of bitext mining (Schwenk et al., 2021b)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 686, + 291, + 754 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 686, + 291, + 754 + ], + "spans": [ + { + "bbox": [ + 67, + 686, + 291, + 754 + ], + "type": "text", + "content": "In bitext mining, the aim is to find pairs of sentences with the same sentence meaning across collections of monolingual corpora. In this work, we adopt a global mining approach (Schwenk et al., 2021a), which has shown recent success in provid" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 213, + 524, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 213, + 524, + 238 + ], + "spans": [ + { + "bbox": [ + 302, + 213, + 524, + 238 + ], + "type": "text", + "content": "ing high-quality data for low-resourced languages (NLLB Team et al., 2022)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 240, + 526, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 240, + 526, + 511 + ], + "spans": [ + { + "bbox": [ + 302, + 240, + 526, + 511 + ], + "type": "text", + "content": "In order to evaluate any bitext mining method, a natural approach is to train a NMT system on the automatically created alignments. However, this is extremely costly. As an alternative, the BUCC task (Zweigenbaum et al., 2018) offers a method for evaluating bitext mining algorithms by embedding known alignments within monolingual corpora, and then reporting on the number of correctly aligned pairs. However, this task currently only covers 5 high-resourced languages (English, French, Russian, German and Chinese), and so is not applicable to the low-resource domain. In order to address this, another approach to evaluate bitext mining is to align existing multilingual parallel test sets. Two such test sets are Tatoeba1 and FLORES200.2 However, as shown by Heffernan et al. (2022), the Tatoeba corpus is not very reliable given that for some sentence pairs there are only a few hundred sentences. Therefore, we opt to use FLORES200, which is also n-way parallel." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 512, + 525, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 512, + 525, + 633 + ], + "spans": [ + { + "bbox": [ + 302, + 512, + 525, + 633 + ], + "type": "text", + "content": "One existing method for evaluating bitext mining on parallel test sets is xsim. This method reports the error rate of misaligned sentences, and follows a margin-based global mining approach (Artetxe and Schwenk, 2019a). However, although using xsim on test sets such as FLORES200 has been shown to be useful as a proxy metric for bitext mining (NLLB Team et al., 2022), it has the following limitations:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 645, + 526, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 645, + 526, + 699 + ], + "spans": [ + { + "bbox": [ + 312, + 645, + 526, + 699 + ], + "type": "text", + "content": "1. Using FLORES200 alone has proven to not be difficult enough as for many language pairs, existing approaches quickly saturate at " + }, + { + "bbox": [ + 312, + 645, + 526, + 699 + ], + "type": "inline_equation", + "content": "0\\%" + }, + { + "bbox": [ + 312, + 645, + 526, + 699 + ], + "type": "text", + "content": " error (NLLB Team et al., 2022)." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 708, + 514, + 729 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 708, + 514, + 729 + ], + "spans": [ + { + "bbox": [ + 302, + 708, + 514, + 729 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 729, + 514, + 751 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 729, + 514, + 751 + ], + "spans": [ + { + "bbox": [ + 302, + 729, + 514, + 751 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 302, + 729, + 514, + 751 + ], + "type": "text", + "content": "https://github.com/facebookresearch/ Flores/tree/main/fores200" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 751, + 510, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 751, + 510, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 510, + 772 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 302, + 751, + 510, + 772 + ], + "type": "text", + "content": "https://github.com/facebookresearch/LASER/tree/main/tasks/xsim" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 85, + 761, + 160, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 761, + 160, + 772 + ], + "spans": [ + { + "bbox": [ + 85, + 761, + 160, + 772 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 289, + 780, + 306, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 780, + 306, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 780, + 306, + 791 + ], + "type": "text", + "content": "101" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 135, + 795, + 459, + 805 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 795, + 459, + 805 + ], + "spans": [ + { + "bbox": [ + 135, + 795, + 459, + 805 + ], + "type": "text", + "content": "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 224, + 806, + 370, + 817 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 806, + 370, + 817 + ], + "spans": [ + { + "bbox": [ + 224, + 806, + 370, + 817 + ], + "type": "text", + "content": "Volume 2: Short Papers, pages 101-109" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "spans": [ + { + "bbox": [ + 176, + 818, + 417, + 829 + ], + "type": "text", + "content": "July 9-14, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 68, + 520, + 169 + ], + "blocks": [ + { + "bbox": [ + 73, + 68, + 520, + 169 + ], + "lines": [ + { + "bbox": [ + 73, + 68, + 520, + 169 + ], + "spans": [ + { + "bbox": [ + 73, + 68, + 520, + 169 + ], + "type": "table", + "html": "
Transformation CategoryOriginal SentenceTransformed Sentence
Causality AlternationApart from the fever and a sore throat, I feel well and in good shape to carry out my work by telecommuting.Apart from the fever and a sore throat, I feel well and in bad shape to carry out my work by telecommuting
Entity ReplacementCharles was the first member of the British Royal Family to be awarded a degree.M. Smith was the first member of The University to be awarded a degree.
Number ReplacementNadal bagged 88% net points in the match winning 76 points in the first serve.Nadal bagged 98% net points in the match winning 71 points in the sixth serve.
", + "image_path": "617ea5b529f0f604b41bab80537eb9b6a874a7fcad6d68b455641e223b43b092.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 176, + 525, + 200 + ], + "lines": [ + { + "bbox": [ + 67, + 176, + 525, + 200 + ], + "spans": [ + { + "bbox": [ + 67, + 176, + 525, + 200 + ], + "type": "text", + "content": "Table 1: Examples of the transformations applied to the English sentences from FLORES200 dev set. The red texts indicate the places of alternations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 76, + 222, + 289, + 379 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 76, + 222, + 289, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 222, + 289, + 288 + ], + "spans": [ + { + "bbox": [ + 76, + 222, + 289, + 288 + ], + "type": "text", + "content": "2. As the dev and devtest sets are quite small (997/1012 respectively), this is arguably not a good approximation for performance when mining against billions of possible candidate sentences." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 76, + 298, + 289, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 298, + 289, + 379 + ], + "spans": [ + { + "bbox": [ + 76, + 298, + 289, + 379 + ], + "type": "text", + "content": "3. We have observed that there is not a significant overlap in the semantics between candidate sentences, meaning that it is not possible to test difficult scenarios that arise in bitext mining when choosing between multiple (similar) candidate pairs." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 67, + 386, + 289, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 386, + 289, + 493 + ], + "spans": [ + { + "bbox": [ + 67, + 386, + 289, + 493 + ], + "type": "text", + "content": "In order to address these limitations, in this work we introduce " + }, + { + "bbox": [ + 67, + 386, + 289, + 493 + ], + "type": "inline_equation", + "content": "x \\sin^{++}" + }, + { + "bbox": [ + 67, + 386, + 289, + 493 + ], + "type": "text", + "content": ". This is an improved proxy for bitext mining performance which expands the dev and devtest sets of FLORES200 to include both more data points, and also difficult to distinguish cases which provide far greater challenges to the models. Our contributions can be summarised as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 76, + 502, + 290, + 678 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 77, + 502, + 290, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 502, + 290, + 540 + ], + "spans": [ + { + "bbox": [ + 77, + 502, + 290, + 540 + ], + "type": "text", + "content": "1. We create a more semantically challenging and expanded English test set for FLORES200." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 76, + 550, + 289, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 550, + 289, + 630 + ], + "spans": [ + { + "bbox": [ + 76, + 550, + 289, + 630 + ], + "type": "text", + "content": "2. We validate this new test set by independently performing 110 bitext mining runs, training 110 NMT systems on the output mined bittexts, and then determining both the correlation and statistical significance between " + }, + { + "bbox": [ + 76, + 550, + 289, + 630 + ], + "type": "inline_equation", + "content": "x \\sin + +" + }, + { + "bbox": [ + 76, + 550, + 289, + 630 + ], + "type": "text", + "content": " and the resulting BLEU scores." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 76, + 640, + 289, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 640, + 289, + 678 + ], + "spans": [ + { + "bbox": [ + 76, + 640, + 289, + 678 + ], + "type": "text", + "content": "3. We open-source the expanded FLORES200 dev and devtest sets, and also the xsim++ code to evaluate them4." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 67, + 689, + 156, + 703 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 689, + 156, + 703 + ], + "spans": [ + { + "bbox": [ + 67, + 689, + 156, + 703 + ], + "type": "text", + "content": "2 Methodology" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 710, + 182, + 723 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 710, + 182, + 723 + ], + "spans": [ + { + "bbox": [ + 67, + 710, + 182, + 723 + ], + "type": "text", + "content": "2.1 Background: xsim" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 67, + 728, + 290, + 754 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 728, + 290, + 754 + ], + "spans": [ + { + "bbox": [ + 67, + 728, + 290, + 754 + ], + "type": "text", + "content": "Given two lists of sentences in different languages, xsim seeks to align each sentence in the source" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 222, + 526, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 222, + 526, + 384 + ], + "spans": [ + { + "bbox": [ + 302, + 222, + 526, + 384 + ], + "type": "text", + "content": "language to a corresponding sentence in the target language based on a margin-based similarity (Artetxe and Schwenk, 2019a). In doing so, xsim leverages the mining approach described in Artetxe and Schwenk (2019b) to first encode sentences into embedding vectors, assign pairwise scores between sentences in the lists, and then take the sentence in the target language that achieves the maximum score as the final prediction. xsim relies on human-annotated parallel corpora and measures the performance of bitext mining using the fraction of misaligned source sentences, i.e., error rates." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 395, + 363, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 395, + 363, + 407 + ], + "spans": [ + { + "bbox": [ + 302, + 395, + 363, + 407 + ], + "type": "text", + "content": "2.2 xsim++" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 301, + 413, + 525, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 413, + 525, + 522 + ], + "spans": [ + { + "bbox": [ + 301, + 413, + 525, + 522 + ], + "type": "text", + "content": "As the effectiveness of xsim is limited by the availability of parallel corpora, we choose to create xsim++ by automatically expanding the English sentences, and evaluate the sentence encoders on into-English language directions, following prior work on low-resource bitext mining (Heffernan et al., 2022). Aside from the expanded candidate set, xsim++ follows the same procedure as xsim." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 523, + 525, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 523, + 525, + 711 + ], + "spans": [ + { + "bbox": [ + 302, + 523, + 525, + 711 + ], + "type": "inline_equation", + "content": "\\times \\mathrm{sim} + +" + }, + { + "bbox": [ + 302, + 523, + 525, + 711 + ], + "type": "text", + "content": " seeks to capture more subtle improvements in bitext mining by adding challenging negative examples. The examples are human-written sentences transformed by various operations. These operations intend to perturb semantics through minimal alternations in the surface text. In particular, we use the following categories of transformations: causality alternation, entity replacement, and number replacement. We focus on these three transformation types only as they easily allow us to create negative examples. Examples of the transformed sentences are shown in Table 1. For these transformations, we adapt the implementation in Dhole et al. (2021)" + }, + { + "bbox": [ + 302, + 523, + 525, + 711 + ], + "type": "inline_equation", + "content": "^6" + }, + { + "bbox": [ + 302, + 523, + 525, + 711 + ], + "type": "text", + "content": " and describe the details" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 719, + 525, + 740 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 719, + 525, + 740 + ], + "spans": [ + { + "bbox": [ + 302, + 719, + 525, + 740 + ], + "type": "text", + "content": "5In this work we report all results using the absolute margin" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 302, + 741, + 525, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 741, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 525, + 772 + ], + "type": "text", + "content": "Although this library has additional transformation methods available, many would create positive examples in this use case (e.g. paraphrases)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 80, + 760, + 271, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 760, + 271, + 772 + ], + "spans": [ + { + "bbox": [ + 80, + 760, + 271, + 772 + ], + "type": "text", + "content": "4https://github.com/facebookresearch/LASER" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "102" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 111, + 68, + 247, + 133 + ], + "blocks": [ + { + "bbox": [ + 111, + 68, + 247, + 133 + ], + "lines": [ + { + "bbox": [ + 111, + 68, + 247, + 133 + ], + "spans": [ + { + "bbox": [ + 111, + 68, + 247, + 133 + ], + "type": "table", + "html": "
Total ## per orig.
Original997-
Causality18681.87
Entity3774537.86
Number34763.49
", + "image_path": "9be1e755847667d793e10b653d85b30354728abbbbd94ee250cb9d99825544ba.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 141, + 291, + 190 + ], + "lines": [ + { + "bbox": [ + 67, + 141, + 291, + 190 + ], + "spans": [ + { + "bbox": [ + 67, + 141, + 291, + 190 + ], + "type": "text", + "content": "Table 2: Total numbers of original sentences and transformed sentences in different transformation categories. We also report the averaged numbers of transformations per original sentence for each category." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 212, + 207, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 212, + 207, + 224 + ], + "spans": [ + { + "bbox": [ + 67, + 212, + 207, + 224 + ], + "type": "text", + "content": "of these transformations below." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 235, + 290, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 235, + 290, + 383 + ], + "spans": [ + { + "bbox": [ + 67, + 235, + 290, + 383 + ], + "type": "text", + "content": "Causality Alternation. To alter causality in a sentence, we (1) replace adjectives with their antonyms; (2) negate the meaning of sentences by adding or removing negation function words (e.g. \"did not\" and \"was not\") to the sentences; or (3) leverage the negation strengthening approach (Tan et al., 2021), which changes the causal relationships through more assertive function words (e.g. replacing \"may\" with \"will\"). For example, as shown in Table 1 we replace \"good\" with the antonym \"bad\"." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 394, + 291, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 394, + 291, + 476 + ], + "spans": [ + { + "bbox": [ + 67, + 394, + 291, + 476 + ], + "type": "text", + "content": "Entity Replacement. We collect candidate entities from large amounts of monolingual data. Then we replace entities in sentences with the ones randomly sampled from the candidate set. For both stages, we use the named entity recognizer from NLTK (Bird et al., 2009)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 486, + 291, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 486, + 291, + 540 + ], + "spans": [ + { + "bbox": [ + 67, + 486, + 291, + 540 + ], + "type": "text", + "content": "Number Replacement. We use spaCy (Honni-bal and Montani, 2017) to detect dates, ordinals, cardinals, times, numbers, and percentages and then randomly replace their values." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 554, + 291, + 743 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 554, + 291, + 743 + ], + "spans": [ + { + "bbox": [ + 67, + 554, + 291, + 743 + ], + "type": "text", + "content": "Given the strategies above, for each sentence we create multiple transformations (i.e. negative examples) of that source sentence. For example, consider Table 1. In the \"Entity Replacement\" example we create a transformation by replacing two named entities. We can then continue this process by replacing these with other named entities until we have reached the desired number of total transformations7. Note that since the opportunity to change each category is dependent on the frequency of that category in the evaluation sets, some transformations occurred more than others (e.g. entities were more frequent than numbers). We summarize the data statistics for xsim++ on the FLORES200 dev" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 71, + 525, + 98 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 525, + 98 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 525, + 98 + ], + "type": "text", + "content": "set in Table 2. Results for the devtest set are in appendix A." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 303, + 110, + 386, + 124 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 110, + 386, + 124 + ], + "spans": [ + { + "bbox": [ + 303, + 110, + 386, + 124 + ], + "type": "text", + "content": "3 Experiment" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 133, + 526, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 133, + 526, + 280 + ], + "spans": [ + { + "bbox": [ + 302, + 133, + 526, + 280 + ], + "type": "text", + "content": "In order to establish " + }, + { + "bbox": [ + 302, + 133, + 526, + 280 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 302, + 133, + 526, + 280 + ], + "type": "text", + "content": " sim++ as a proxy for bitext mining performance, we measure the correlation between both " + }, + { + "bbox": [ + 302, + 133, + 526, + 280 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 302, + 133, + 526, + 280 + ], + "type": "text", + "content": " sim and " + }, + { + "bbox": [ + 302, + 133, + 526, + 280 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 302, + 133, + 526, + 280 + ], + "type": "text", + "content": " sim++ error rates, and the BLEU scores resulting from NMT systems trained on mined bittexts. More specifically, for each language we choose a sentence encoder model, followed by bitext mining using each respective encoder, and then train and evaluate bilingual NMT systems on the resulting mined bittexts. We use the FLORES200 development sets when computing the BLEU scores." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 283, + 525, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 283, + 525, + 391 + ], + "spans": [ + { + "bbox": [ + 302, + 283, + 525, + 391 + ], + "type": "text", + "content": "In order to validate " + }, + { + "bbox": [ + 302, + 283, + 525, + 391 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 302, + 283, + 525, + 391 + ], + "type": "text", + "content": "sim++ against varied embedding spaces, we encode (and mine) using two different multilingual encoder methods: LASER (Artetxe and Schwenk, 2019b) and LaBSE (Feng et al., 2022). For LASER, we trained our own custom encoders (details below). For LaBSE, we used a publicly available model as the code and data for training LaBSE are not publicly available." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 391, + 525, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 391, + 525, + 473 + ], + "spans": [ + { + "bbox": [ + 302, + 391, + 525, + 473 + ], + "type": "text", + "content": "We randomly choose 10 low-resource languages to perform both encoder training (if applicable) and bitext mining. The languages are: Faroese (fao), Kabuverdianu (kea), Tok Pisin (tpi), Kikuyu (kik), Friulian (fur), Igbo (ibo), Luxembourgish (ltz), Swahili (swh), Zulu (zul), Bemba (bem)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 482, + 525, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 482, + 525, + 591 + ], + "spans": [ + { + "bbox": [ + 302, + 482, + 525, + 591 + ], + "type": "text", + "content": "Encoder Training. We trained LASER encoders using the teacher-student approach described in Heffernan et al. (2022). We choose a LASER model (Artetxe and Schwenk, 2019b) as our teacher, and then trained specialised students for each language. In order to train each student, we used both publicly available code9 and bitexts (e.g. OPUS10)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 601, + 526, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 601, + 526, + 709 + ], + "spans": [ + { + "bbox": [ + 302, + 601, + 526, + 709 + ], + "type": "text", + "content": "Bitext Mining. For each chosen encoder model, we perform bitext mining against approximately 3.7 billion sentences of English. For low-resource languages, the sizes of monolingual data range from 140k to 124 million. Details are in the appendix. We make use of monolingual data available from both Commoncrawl and Paracrawl11, and operationalize the mining using the stopes library (An" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 717, + 520, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 717, + 520, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 717, + 520, + 772 + ], + "type": "inline_equation", + "content": "^{8}" + }, + { + "bbox": [ + 302, + 717, + 520, + 772 + ], + "type": "text", + "content": "https://huggingface.co/sentence-transformers/LaBSE \n" + }, + { + "bbox": [ + 302, + 717, + 520, + 772 + ], + "type": "inline_equation", + "content": "^{9}" + }, + { + "bbox": [ + 302, + 717, + 520, + 772 + ], + "type": "text", + "content": "https://github.com/facebookresearch/fairseq/tree/nllb/examples/nllb/laser_distillation \n" + }, + { + "bbox": [ + 302, + 717, + 520, + 772 + ], + "type": "inline_equation", + "content": "^{10}" + }, + { + "bbox": [ + 302, + 717, + 520, + 772 + ], + "type": "text", + "content": "https://opus.nlpl.eu \n" + }, + { + "bbox": [ + 302, + 717, + 520, + 772 + ], + "type": "inline_equation", + "content": "^{11}" + }, + { + "bbox": [ + 302, + 717, + 520, + 772 + ], + "type": "text", + "content": "https://paracrawl.eu" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 67, + 750, + 290, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 750, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 750, + 290, + 772 + ], + "type": "text", + "content": "7We set a maximum threshold of 100 transformations per category per sentence." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 290, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 290, + 781, + 307, + 791 + ], + "type": "text", + "content": "103" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 70, + 291, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 70, + 291, + 195 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 291, + 195 + ], + "type": "text", + "content": "drews et al., 2022).12 For LASER, we use 1.06 as the margin threshold following Heffernan et al. (2022) and for LaBSE, we use 1.16.13 Following mining, for each language we concatenate publicly available bittexts and the mined bitext as training data for NMT bilingual models using fairseq,14 translating from each foreign text into English. For all NMT systems, we keep the hyperparameters fixed (details in Appendix)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 202, + 291, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 202, + 291, + 352 + ], + "spans": [ + { + "bbox": [ + 67, + 202, + 291, + 352 + ], + "type": "text", + "content": "Evaluation. Model selection involves two use cases: comparisons within a model and across different models. For the former comparison, given our custom encoders, we choose to compare 10 checkpoints from each model.[15] For cross model comparisons, we compare each chosen encoder checkpoint against another existing system. In this case, the LaBSE encoder. To quantitatively measure these two cases, we report pairwise ranking accuracy (Kocmi et al., 2021) for xsim and xsim++. Formally, the accuracy is computed as follows" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 74, + 374, + 285, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 374, + 285, + 403 + ], + "spans": [ + { + "bbox": [ + 74, + 374, + 285, + 403 + ], + "type": "interline_equation", + "content": "\\frac {\\left| \\mathrm {s} (\\text {p r o x y} \\Delta) = \\mathrm {s} (\\text {m i n i n g} \\Delta) \\text {f o r a l l s y s t e m p a i r s} \\right|}{\\left| \\text {a l l s y s t e m p a i r s} \\right|}", + "image_path": "ae912b2b7b749392d69b49bd5d162e6fc378f9d8a21745114d75f7cba426a522.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 414, + 290, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 414, + 290, + 469 + ], + "spans": [ + { + "bbox": [ + 67, + 414, + 290, + 469 + ], + "type": "text", + "content": "where proxy" + }, + { + "bbox": [ + 67, + 414, + 290, + 469 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 67, + 414, + 290, + 469 + ], + "type": "text", + "content": " is the difference of the " + }, + { + "bbox": [ + 67, + 414, + 290, + 469 + ], + "type": "inline_equation", + "content": "x \\sin" + }, + { + "bbox": [ + 67, + 414, + 290, + 469 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 67, + 414, + 290, + 469 + ], + "type": "inline_equation", + "content": "x \\sin + +" + }, + { + "bbox": [ + 67, + 414, + 290, + 469 + ], + "type": "text", + "content": " scores, mining" + }, + { + "bbox": [ + 67, + 414, + 290, + 469 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 67, + 414, + 290, + 469 + ], + "type": "text", + "content": " is the difference of the BLEU scores, " + }, + { + "bbox": [ + 67, + 414, + 290, + 469 + ], + "type": "inline_equation", + "content": "s(\\cdot)" + }, + { + "bbox": [ + 67, + 414, + 290, + 469 + ], + "type": "text", + "content": " is the sign function, and " + }, + { + "bbox": [ + 67, + 414, + 290, + 469 + ], + "type": "inline_equation", + "content": "|\\cdot|" + }, + { + "bbox": [ + 67, + 414, + 290, + 469 + ], + "type": "text", + "content": " returns the cardinal number of the input." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 469, + 291, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 469, + 291, + 550 + ], + "spans": [ + { + "bbox": [ + 67, + 469, + 291, + 550 + ], + "type": "text", + "content": "In this work, we have 550 system pairs with 55 pairs per language direction (i.e. " + }, + { + "bbox": [ + 67, + 469, + 291, + 550 + ], + "type": "inline_equation", + "content": "\\binom{11}{2}" + }, + { + "bbox": [ + 67, + 469, + 291, + 550 + ], + "type": "text", + "content": " pairs given 10 custom LASER encoder checkpoints + LaBSE). We always compare systems within a language direction as the scores for system pairs across different directions are not comparable.[16]" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 561, + 130, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 561, + 130, + 574 + ], + "spans": [ + { + "bbox": [ + 67, + 561, + 130, + 574 + ], + "type": "text", + "content": "3.1 Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 580, + 291, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 580, + 291, + 649 + ], + "spans": [ + { + "bbox": [ + 67, + 580, + 291, + 649 + ], + "type": "text", + "content": "As shown in Table 3, xsim++ significantly outperforms xsim on the pairwise ranking accuracy. Additionally, when comparing the computational cost to mining, xsim++ costs over " + }, + { + "bbox": [ + 67, + 580, + 291, + 649 + ], + "type": "inline_equation", + "content": "99.9\\%" + }, + { + "bbox": [ + 67, + 580, + 291, + 649 + ], + "type": "text", + "content": " less GPU hours and saves approximately 3 metric tons of carbon" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 316, + 68, + 514, + 124 + ], + "blocks": [ + { + "bbox": [ + 316, + 68, + 514, + 124 + ], + "lines": [ + { + "bbox": [ + 316, + 68, + 514, + 124 + ], + "spans": [ + { + "bbox": [ + 316, + 68, + 514, + 124 + ], + "type": "table", + "html": "
MetricAccuracyGPU hours
xsim35.480.43
xsim++72.00*0.52
Mining BLEU(Oracle)10019569
", + "image_path": "862fc12c58765f7375b533c1cef2da04658a980252a3da11a5f7dfbbb1e3420f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 323, + 203, + 507, + 366 + ], + "blocks": [ + { + "bbox": [ + 302, + 131, + 525, + 191 + ], + "lines": [ + { + "bbox": [ + 302, + 131, + 525, + 191 + ], + "spans": [ + { + "bbox": [ + 302, + 131, + 525, + 191 + ], + "type": "text", + "content": "Table 3: Pairwise ranking accuracy along with the total number of GPU hours. For all experiments, we used NVIDIA A100 GPUs. An * indicates that the result passes the significance test proposed by Koehn (2004) with " + }, + { + "bbox": [ + 302, + 131, + 525, + 191 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 302, + 131, + 525, + 191 + ], + "type": "text", + "content": "-value " + }, + { + "bbox": [ + 302, + 131, + 525, + 191 + ], + "type": "inline_equation", + "content": "< 0.05" + }, + { + "bbox": [ + 302, + 131, + 525, + 191 + ], + "type": "text", + "content": " when compared to xsim." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 203, + 507, + 366 + ], + "lines": [ + { + "bbox": [ + 323, + 203, + 507, + 366 + ], + "spans": [ + { + "bbox": [ + 323, + 203, + 507, + 366 + ], + "type": "table", + "html": "
Accuracy
xsim++72.00
Causality63.09
Entity65.45
Number60.73
Misaligned40.73
Causality + Entity68.55
Causality + Entity + Misaligned70.55
Causality + Misaligned68.00
Causality + Number66.73
Causality + Number + Misaligned71.45
Entity + Misaligned70.55
Number + Entity67.45
Number + Entity + Misaligned71.09
Number + Misaligned64.36
", + "image_path": "c13fcad28dad65c1a159f599bc6129c3144421c8ff6b94a4c07ed1d22aa17929.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 374, + 526, + 423 + ], + "lines": [ + { + "bbox": [ + 302, + 374, + 526, + 423 + ], + "spans": [ + { + "bbox": [ + 302, + 374, + 526, + 423 + ], + "type": "text", + "content": "Table 4: Pairwise ranking accuracy when using combinations of error categories. Causality=Causality Alternation, Entity=Entity Replacement, Number=Number Replacement." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 301, + 444, + 525, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 444, + 525, + 497 + ], + "spans": [ + { + "bbox": [ + 301, + 444, + 525, + 497 + ], + "type": "text", + "content": "emissions, but still manages to achieve a competitive accuracy. We observe similar trends for the within a model and across models use cases and report their separate accuracies in the appendix." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 301, + 498, + 525, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 498, + 525, + 634 + ], + "spans": [ + { + "bbox": [ + 301, + 498, + 525, + 634 + ], + "type": "text", + "content": "To better understand the contributions of each transformation category (cf. subsection 2.1) in measuring the final mining performance, we report accuracies for different combinations of categories in Table 4. In cases where an incorrect bitext alignment do does not map to any of the augmented sentences of the true alignment, we denote these as \"misaligned\". We find that entity replacement helps most in improving the accuracy and combing all the transformations gives the best performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 643, + 396, + 655 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 643, + 396, + 655 + ], + "spans": [ + { + "bbox": [ + 302, + 643, + 396, + 655 + ], + "type": "text", + "content": "4 Related Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 301, + 665, + 526, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 665, + 526, + 773 + ], + "spans": [ + { + "bbox": [ + 301, + 665, + 526, + 773 + ], + "type": "text", + "content": "As " + }, + { + "bbox": [ + 301, + 665, + 526, + 773 + ], + "type": "inline_equation", + "content": "xsim++" + }, + { + "bbox": [ + 301, + 665, + 526, + 773 + ], + "type": "text", + "content": " uses rule-based data augmentation, it is related to work in other areas that also employ similar data augmentation methods, such as part-of-speech tagging (Şahin and Steedman, 2018), contrastive learning (Tang et al., 2022), text classification (Kobayashi, 2018; Wei and Zou, 2019), dialogue generation (Niu and Bansal, 2018) and summarization (Chen and Yang, 2021)." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 79, + 656, + 276, + 669 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 656, + 276, + 669 + ], + "spans": [ + { + "bbox": [ + 79, + 656, + 276, + 669 + ], + "type": "inline_equation", + "content": "^{12}" + }, + { + "bbox": [ + 79, + 656, + 276, + 669 + ], + "type": "text", + "content": "https://github.com/facebookresearch/stopes" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 69, + 669, + 289, + 709 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 669, + 289, + 709 + ], + "spans": [ + { + "bbox": [ + 69, + 669, + 289, + 709 + ], + "type": "text", + "content": "13We did grid search on threshold values from 1.11 to 1.25 on three languages (swh, ltz, and zul), decided the optimal one based on the BLEU scores, and used the threshold for the rest of languages." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 79, + 710, + 280, + 720 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 710, + 280, + 720 + ], + "spans": [ + { + "bbox": [ + 79, + 710, + 280, + 720 + ], + "type": "inline_equation", + "content": "^{14}" + }, + { + "bbox": [ + 79, + 710, + 280, + 720 + ], + "type": "text", + "content": "https://github.com/facebookresearch/fairseq" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 79, + 720, + 233, + 731 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 720, + 233, + 731 + ], + "spans": [ + { + "bbox": [ + 79, + 720, + 233, + 731 + ], + "type": "text", + "content": "15Evenly spaced between epochs 1 and 30." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 69, + 732, + 289, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 732, + 289, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 732, + 289, + 772 + ], + "type": "text", + "content": "16There are factors varied across language directions that are unrelated to the quality of sentence encoders but could affect mining performance, such as amounts of monolingual data available for mining." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "104" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 71, + 239, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 71, + 239, + 83 + ], + "spans": [ + { + "bbox": [ + 68, + 71, + 239, + 83 + ], + "type": "text", + "content": "5 Conclusion and Future Work" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 94, + 293, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 94, + 293, + 257 + ], + "spans": [ + { + "bbox": [ + 67, + 94, + 293, + 257 + ], + "type": "text", + "content": "We proposed a proxy score " + }, + { + "bbox": [ + 67, + 94, + 293, + 257 + ], + "type": "inline_equation", + "content": "x \\sin + +" + }, + { + "bbox": [ + 67, + 94, + 293, + 257 + ], + "type": "text", + "content": " for bitext mining performance using three kinds of data augmentation techniques: causality alternation, entity replacement, and number replacement. To validate its effectiveness, we conducted large-scale bitext mining experiments for 10 low-resource languages, and reported pairwise ranking accuracies. We found that " + }, + { + "bbox": [ + 67, + 94, + 293, + 257 + ], + "type": "inline_equation", + "content": "x \\sin + +" + }, + { + "bbox": [ + 67, + 94, + 293, + 257 + ], + "type": "text", + "content": " significantly improves over " + }, + { + "bbox": [ + 67, + 94, + 293, + 257 + ], + "type": "inline_equation", + "content": "x \\sin" + }, + { + "bbox": [ + 67, + 94, + 293, + 257 + ], + "type": "text", + "content": ", doubling the accuracies. Analysis reveals that entity replacement helps most in the improvement. In future work, we plan to extend " + }, + { + "bbox": [ + 67, + 94, + 293, + 257 + ], + "type": "inline_equation", + "content": "x \\sin + +" + }, + { + "bbox": [ + 67, + 94, + 293, + 257 + ], + "type": "text", + "content": " to non-English language pairs." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 268, + 149, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 268, + 149, + 280 + ], + "spans": [ + { + "bbox": [ + 67, + 268, + 149, + 280 + ], + "type": "text", + "content": "6 Limitations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 291, + 291, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 291, + 291, + 439 + ], + "spans": [ + { + "bbox": [ + 67, + 291, + 291, + 439 + ], + "type": "text", + "content": "We highlight three limitations of our work. The first is that " + }, + { + "bbox": [ + 67, + 291, + 291, + 439 + ], + "type": "inline_equation", + "content": "xsim++" + }, + { + "bbox": [ + 67, + 291, + 291, + 439 + ], + "type": "text", + "content": " is automatically constructed. There could be noisy sentences leading to errors that are irrelevant to the quality of encoders. The second is that " + }, + { + "bbox": [ + 67, + 291, + 291, + 439 + ], + "type": "inline_equation", + "content": "xsim++" + }, + { + "bbox": [ + 67, + 291, + 291, + 439 + ], + "type": "text", + "content": " applies transformations solely to English sentences. Generalizing it to non-English language pairs requires additional research. Finally, we have experimented with the two most popular multilingual encoders: LASER and LaBSE. There are other available approaches which would be interesting to also validate " + }, + { + "bbox": [ + 67, + 291, + 291, + 439 + ], + "type": "inline_equation", + "content": "xsim++" + }, + { + "bbox": [ + 67, + 291, + 291, + 439 + ], + "type": "text", + "content": " against." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 464, + 127, + 476 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 464, + 127, + 476 + ], + "spans": [ + { + "bbox": [ + 68, + 464, + 127, + 476 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 485, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 69, + 485, + 291, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 485, + 291, + 575 + ], + "spans": [ + { + "bbox": [ + 69, + 485, + 291, + 575 + ], + "type": "text", + "content": "Pierre Andrews, Guillaume Wenzek, Kevin Heffernan, Onur Celebi, Anna Sun, Ammar Kamran, Yingzhe Guo, Alexandre Mourachko, Holger Schwenk, and Angela Fan. 2022. stopes - modular machine translation pipelines. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing: System Demonstrations. Association for Computational Linguistics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 584, + 291, + 651 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 584, + 291, + 651 + ], + "spans": [ + { + "bbox": [ + 69, + 584, + 291, + 651 + ], + "type": "text", + "content": "Mikel Artetxe and Holger Schwenk. 2019a. Margin-based parallel corpus mining with multilingual sentence embeddings. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3197-3203, Florence, Italy. Association for Computational Linguistics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 661, + 291, + 717 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 661, + 291, + 717 + ], + "spans": [ + { + "bbox": [ + 69, + 661, + 291, + 717 + ], + "type": "text", + "content": "Mikel Artetxe and Holger Schwenk. 2019b. Massively multilingual sentence embeddings for zero-shot cross-lingual transfer and beyond. Transactions of the Association for Computational Linguistics, 7:597-610." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 727, + 291, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 727, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 727, + 291, + 772 + ], + "type": "text", + "content": "Steven Bird, Ewan Klein, and Edward Loper. 2009. Natural language processing with Python: analyzing text with the natural language toolkit. \"O'Reilly Media, Inc.\"" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 526, + 772 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 304, + 72, + 526, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 526, + 150 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 526, + 150 + ], + "type": "text", + "content": "Jiaao Chen and Diyi Yang. 2021. Simple conversational data augmentation for semi-supervised abstractive dialogue summarization. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6605-6616, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 158, + 526, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 158, + 526, + 631 + ], + "spans": [ + { + "bbox": [ + 304, + 158, + 526, + 631 + ], + "type": "text", + "content": "Kaustubh D. Dhole, Varun Gangal, Sebastian Gehrmann, Aadesh Gupta, Zhenhao Li, Saad Mahamood, Abinaya Mahendiran, Simon Mille, Ashish Srivastava, Samson Tan, Tongshuang Wu, Jascha Sohl-Dickstein, Jinho D. Choi, Eduard Hovy, Ondrej Dusek, Sebastian Ruder, Sajant Anand, Nagender Aneja, Rabin Banjade, Lisa Barthe, Hanna Behnke, Ian Berlot-Attwell, Connor Boyle, Caroline Brun, Marco Antonio Sobrevilla Cabezudo, Samuel Cahyawijaya, Emile Chapuis, Wanxiang Che, Mukund Choudhary, Christian Clauss, Pierre Colombo, Filip Cornell, Gautier Dagan, Mayukh Das, Tanay Dixit, Thomas Dopierre, Paul-Alexis Dray, Suchitra Dubey, Tatiana Ekeinhor, Marco Di Giovanni, Rishabh Gupta, Rishabh Gupta, Louanes Hamla, Sang Han, Fabrice Harel-Canada, Antoine Honore, Ishan Jindal, Przemyslaw K. Joniak, Denis Kleyko, Venelin Kovatchev, Kalpesh Krishna, Ashutosh Kumar, Stefan Langer, Seungjae Ryan Lee, Corey James Levinson, Hualou Liang, Kaizhao Liang, Zhexiong Liu, Andrey Lukyanenko, Vukosi Marivate, Gerard de Melo, Simon Meoni, Maxime Meyer, Afnan Mir, Nafise Sadat Moosavi, Niklas Muennighoff, Timothy Sum Hon Mun, Kenton Murray, Marcin Namysl, Maria Obedkova, Priti Oli, Nivranshu Pasricha, Jan Pfister, Richard Plant, Vinay Prabhu, Vasile Pais, Libo Qin, Shahab Raji, Pawan Kumar Rajpoot, Vikas Raunak, Roy Rinberg, Nicolas Roberts, Juan Diego Rodriguez, Claude Roux, Vasconcellos P. H. S., Ananya B. Sai, Robin M. Schmidt, Thomas Scialom, Tshephisho Sefara, Saqib N. Shamsi, Xudong Shen, Haoyue Shi, Yiwen Shi, Anna Shvets, Nick Siegel, Damien Sileo, Jamie Simon, Chandan Singh, Roman Sitelew, Priyank Soni, Taylor Sorensen, William Soto, Aman Srivastava, KV Aditya Srivatsa, Tony Sun, Mukund Varma T, A Tabassum, Fiona Anting Tan, Ryan Teehan, Mo Tiwari, Marie Tolkiehn, Athena Wang, Zijian Wang, Gloria Wang, Zijie J. Wang, Fuxuan Wei, Bryan Wilie, Genta Indra Winata, Xinyi Wu, Witold Wydmanski, Tianbao Xie, Usama Yaseen, M. Yee, Jing Zhang and Yue Zhang. 2021. Nl-augmenter: A framework for task-sensitive natural language augmentation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 640, + 526, + 718 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 640, + 526, + 718 + ], + "spans": [ + { + "bbox": [ + 304, + 640, + 526, + 718 + ], + "type": "text", + "content": "Fangxiaoyu Feng, Yinfei Yang, Daniel Cer, Naveen Arivazhagan, and Wei Wang. 2022. Language-agnostic BERT sentence embedding. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 878–891, Dublin, Ireland. Association for Computational Linguistics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 728, + 526, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 728, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 728, + 526, + 772 + ], + "type": "text", + "content": "Kevin Heffernan, Onur Celebi, and Holger Schwenk. 2022. Bitext mining using distilled sentence representations for low-resource languages. Findings of EMNLP." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 290, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 290, + 781, + 307, + 791 + ], + "type": "text", + "content": "105" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 289, + 772 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 69, + 72, + 289, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 289, + 116 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 289, + 116 + ], + "type": "text", + "content": "Matthew Honnibal and Ines Montani. 2017. spaCy 2: Natural language understanding with Bloom embeddings, convolutional neural networks and incremental parsing. To appear." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 128, + 289, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 128, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 69, + 128, + 289, + 216 + ], + "type": "text", + "content": "Sosuke Kobayashi. 2018. Contextual augmentation: Data augmentation by words with paradigmatic relations. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 452-457, New Orleans, Louisiana. Association for Computational Linguistics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 228, + 289, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 228, + 289, + 305 + ], + "spans": [ + { + "bbox": [ + 69, + 228, + 289, + 305 + ], + "type": "text", + "content": "Tom Kocmi, Christian Federmann, Roman Grundkiewicz, Marcin Junczys-Dowmunt, Hitokazu Matsushita, and Arul Menezes. 2021. To ship or not to ship: An extensive evaluation of automatic metrics for machine translation. In Proceedings of the Sixth Conference on Machine Translation, pages 478-494, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 317, + 289, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 317, + 289, + 371 + ], + "spans": [ + { + "bbox": [ + 69, + 317, + 289, + 371 + ], + "type": "text", + "content": "Philipp Koehn. 2004. Statistical significance tests for machine translation evaluation. In Proceedings of the 2004 Conference on Empirical Methods in Natural Language Processing, pages 388-395, Barcelona, Spain. Association for Computational Linguistics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 384, + 289, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 384, + 289, + 439 + ], + "spans": [ + { + "bbox": [ + 69, + 384, + 289, + 439 + ], + "type": "text", + "content": "Philipp Koehn and Rebecca Knowles. 2017. Six challenges for neural machine translation. In Proceedings of the First Workshop on Neural Machine Translation, pages 28-39, Vancouver. Association for Computational Linguistics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 450, + 289, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 450, + 289, + 517 + ], + "spans": [ + { + "bbox": [ + 69, + 450, + 289, + 517 + ], + "type": "text", + "content": "Tong Niu and Mohit Bansal. 2018. Adversarial over-sensitivity and over-stability strategies for dialogue models. In Proceedings of the 22nd Conference on Computational Natural Language Learning, pages 486-496, Brussels, Belgium. Association for Computational Linguistics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 528, + 289, + 594 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 528, + 289, + 594 + ], + "spans": [ + { + "bbox": [ + 69, + 528, + 289, + 594 + ], + "type": "text", + "content": "NLLB Team, Marta R Costa-jussa, James Cross, Onur Celebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, et al. 2022. No language left behind: Scaling human-centered machine translation. arXiv preprint arXiv:2207.04672." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 606, + 289, + 672 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 606, + 289, + 672 + ], + "spans": [ + { + "bbox": [ + 69, + 606, + 289, + 672 + ], + "type": "text", + "content": "Gözde Gül Şahin and Mark Steedman. 2018. Data augmentation via dependency tree morphing for low-resource languages. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 5004-5009, Brussels, Belgium. Association for Computational Linguistics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 684, + 289, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 684, + 289, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 684, + 289, + 772 + ], + "type": "text", + "content": "Holger Schwenk, Vishrav Chaudhary, Shuo Sun, Hongyu Gong, and Francisco Guzmán. 2021a. WikiMatrix: Mining 135M parallel sentences in 1620 language pairs from Wikipedia. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 1351-1361, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 524, + 525 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 304, + 72, + 524, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 524, + 170 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 524, + 170 + ], + "type": "text", + "content": "Holger Schwenk, Guillaume Wenzek, Sergey Edunov, Edouard Grave, Armand Joulin, and Angela Fan. 2021b. CCMatrix: Mining billions of high-quality parallel sentences on the web. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 6490-6500, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 179, + 524, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 179, + 524, + 257 + ], + "spans": [ + { + "bbox": [ + 304, + 179, + 524, + 257 + ], + "type": "text", + "content": "Fiona Anting Tan, Devamanyu Hazarika, See-Kiong Ng, Soujanya Poria, and Roger Zimmermann. 2021. Causal augmentation for causal sentence classification. In Proceedings of the First Workshop on Causal Inference and NLP, pages 1-20, Punta Cana, Dominican Republic. Association for Computational Linguistics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 265, + 524, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 265, + 524, + 364 + ], + "spans": [ + { + "bbox": [ + 304, + 265, + 524, + 364 + ], + "type": "text", + "content": "Zilu Tang, Muhammed Yusuf Kocyigit, and Derry Tanti Wijaya. 2022. AugCSE: Contrastive sentence embedding with diverse augmentations. In Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 375-398, Online only. Association for Computational Linguistics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 373, + 524, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 524, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 524, + 460 + ], + "type": "text", + "content": "Jason Wei and Kai Zou. 2019. EDA: Easy data augmentation techniques for boosting performance on text classification tasks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 6382-6388, Hong Kong, China. Association for Computational Linguistics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 470, + 524, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 470, + 524, + 525 + ], + "spans": [ + { + "bbox": [ + 304, + 470, + 524, + 525 + ], + "type": "text", + "content": "Pierre Zweigenbaum, Serge Sharoff, and Reinhard Rapp. 2018. Overview of the third bucc shared task: Spotting parallel sentences in comparable corpora. In Proceedings of 11th workshop on building and using comparable corpora, pages 39-42." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 290, + 781, + 307, + 790 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 781, + 307, + 790 + ], + "spans": [ + { + "bbox": [ + 290, + 781, + 307, + 790 + ], + "type": "text", + "content": "106" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 70, + 247, + 97 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 70, + 247, + 97 + ], + "spans": [ + { + "bbox": [ + 68, + 70, + 247, + 97 + ], + "type": "text", + "content": "A Data Statistics for xsim++ with FLORES200 devtest set" + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 111, + 114, + 247, + 179 + ], + "blocks": [ + { + "bbox": [ + 111, + 114, + 247, + 179 + ], + "lines": [ + { + "bbox": [ + 111, + 114, + 247, + 179 + ], + "spans": [ + { + "bbox": [ + 111, + 114, + 247, + 179 + ], + "type": "table", + "html": "
Total ##per orig.
Original1012-
Causality19161.89
Entity3885538.39
Number32623.22
", + "image_path": "f3fb07f619cd222ae753326f05cb302c8ef348d388a9ea4315b1e5e474540496.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 252, + 290, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 252, + 290, + 279 + ], + "spans": [ + { + "bbox": [ + 67, + 252, + 290, + 279 + ], + "type": "text", + "content": "We report the data statistics for " + }, + { + "bbox": [ + 67, + 252, + 290, + 279 + ], + "type": "inline_equation", + "content": "xsim++" + }, + { + "bbox": [ + 67, + 252, + 290, + 279 + ], + "type": "text", + "content": " with FLORES200 devtest set in Table 5." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 290, + 242, + 318 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 290, + 242, + 318 + ], + "spans": [ + { + "bbox": [ + 68, + 290, + 242, + 318 + ], + "type": "text", + "content": "B Sizes of Monolingual data for Low-Resource Languages" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 126, + 336, + 233, + 455 + ], + "blocks": [ + { + "bbox": [ + 67, + 187, + 291, + 236 + ], + "lines": [ + { + "bbox": [ + 67, + 187, + 291, + 236 + ], + "spans": [ + { + "bbox": [ + 67, + 187, + 291, + 236 + ], + "type": "text", + "content": "Table 5: Total numbers of original sentences and transformed sentences in different transformation categories. We also report the averaged numbers of transformations per original sentence for each category." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 126, + 336, + 233, + 455 + ], + "lines": [ + { + "bbox": [ + 126, + 336, + 233, + 455 + ], + "spans": [ + { + "bbox": [ + 126, + 336, + 233, + 455 + ], + "type": "table", + "html": "
LanguageSize
kik147,902
kea226,507
fur737,178
fao1,179,475
tpi1,661,743
bem2,302,805
ibo8,124,418
zul20,477,331
swh55,399,821
ltz123,944,670
", + "image_path": "e8b253c25b9af2b88dfb476d911b269691af3e938a3248bfc6206ae4437d89e4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 502, + 289, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 502, + 289, + 528 + ], + "spans": [ + { + "bbox": [ + 67, + 502, + 289, + 528 + ], + "type": "text", + "content": "We report the sizes of monolingual data for each language in Table 6." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 68, + 540, + 274, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 540, + 274, + 555 + ], + "spans": [ + { + "bbox": [ + 68, + 540, + 274, + 555 + ], + "type": "text", + "content": "C Hyperparameters for NMT systems" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 100, + 572, + 260, + 711 + ], + "blocks": [ + { + "bbox": [ + 67, + 463, + 290, + 488 + ], + "lines": [ + { + "bbox": [ + 67, + 463, + 290, + 488 + ], + "spans": [ + { + "bbox": [ + 67, + 463, + 290, + 488 + ], + "type": "text", + "content": "Table 6: Number of monolingual sentences for each language." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 100, + 572, + 260, + 711 + ], + "lines": [ + { + "bbox": [ + 100, + 572, + 260, + 711 + ], + "spans": [ + { + "bbox": [ + 100, + 572, + 260, + 711 + ], + "type": "table", + "html": "
encoder layers6
encoder attention heads8
encoder embed dim512
encoder FFNN embed dim4096
decoder layers6
decoder attention heads8
decoder embed dim512
decoder FFNN embed dim4096
optimiserAdam
adam betas(0.9, 0.98)
learning rate0.001
dropout0.3
spm vocab size7000
", + "image_path": "ee4a25d73373525143c8163b0a4b4c72be5b4905693a4615d6c675e9d2964888.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "type": "text", + "content": "We report hyperparameters for NMT evaluations in Table 7." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 303, + 70, + 517, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 70, + 517, + 84 + ], + "spans": [ + { + "bbox": [ + 303, + 70, + 517, + 84 + ], + "type": "text", + "content": "D Within and Across Model Accuracies" + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 356, + 99, + 473, + 138 + ], + "blocks": [ + { + "bbox": [ + 87, + 719, + 270, + 732 + ], + "lines": [ + { + "bbox": [ + 87, + 719, + 270, + 732 + ], + "spans": [ + { + "bbox": [ + 87, + 719, + 270, + 732 + ], + "type": "text", + "content": "Table 7: Hyperparameters for NMT systems." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 356, + 99, + 473, + 138 + ], + "lines": [ + { + "bbox": [ + 356, + 99, + 473, + 138 + ], + "spans": [ + { + "bbox": [ + 356, + 99, + 473, + 138 + ], + "type": "table", + "html": "
MetricWithinAcross
xsim31.3354.04
xsim++69.77*82.00*
", + "image_path": "7c6e984dc1c9ff46d3a17a581244a0e636e5d0ee78f2311b33c2ca96c314a444.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 146, + 525, + 206 + ], + "lines": [ + { + "bbox": [ + 302, + 146, + 525, + 206 + ], + "spans": [ + { + "bbox": [ + 302, + 146, + 525, + 206 + ], + "type": "text", + "content": "Table 8: Pairwise ranking accuracy for comparisons within a model and across different models. An * indicates that the result passes the significance test proposed by Koehn (2004) with " + }, + { + "bbox": [ + 302, + 146, + 525, + 206 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 302, + 146, + 525, + 206 + ], + "type": "text", + "content": "-value " + }, + { + "bbox": [ + 302, + 146, + 525, + 206 + ], + "type": "inline_equation", + "content": "< 0.05" + }, + { + "bbox": [ + 302, + 146, + 525, + 206 + ], + "type": "text", + "content": " when compared to xsim." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 302, + 222, + 525, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 222, + 525, + 263 + ], + "spans": [ + { + "bbox": [ + 302, + 222, + 525, + 263 + ], + "type": "text", + "content": "We report accuracies for within a model (i.e., LASER) and across different models (i.e., the 10 LASER checkpoints vs LaBSE) in Table 8." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 289, + 781, + 307, + 791 + ], + "type": "text", + "content": "107" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "spans": [ + { + "bbox": [ + 69, + 90, + 192, + 103 + ], + "type": "text", + "content": "A For every submission:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 106, + 317, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 106, + 317, + 121 + ], + "spans": [ + { + "bbox": [ + 77, + 106, + 317, + 121 + ], + "type": "text", + "content": "A1. Did you describe the limitations of your work?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 90, + 121, + 138, + 134 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 121, + 138, + 134 + ], + "spans": [ + { + "bbox": [ + 90, + 121, + 138, + 134 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 143, + 329, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 143, + 329, + 157 + ], + "spans": [ + { + "bbox": [ + 77, + 143, + 329, + 157 + ], + "type": "text", + "content": "A2. Did you discuss any potential risks of your work?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 90, + 158, + 208, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 158, + 208, + 169 + ], + "spans": [ + { + "bbox": [ + 90, + 158, + 208, + 169 + ], + "type": "text", + "content": "Not applicable. Left blank." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 77, + 179, + 414, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 179, + 414, + 193 + ], + "spans": [ + { + "bbox": [ + 77, + 179, + 414, + 193 + ], + "type": "text", + "content": "A3. Do the abstract and introduction summarize the paper's main claims?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 91, + 194, + 138, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 194, + 138, + 206 + ], + "spans": [ + { + "bbox": [ + 91, + 194, + 138, + 206 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 77, + 215, + 398, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 215, + 398, + 229 + ], + "spans": [ + { + "bbox": [ + 77, + 215, + 398, + 229 + ], + "type": "text", + "content": "A4. Have you used AI writing assistants when working on this paper?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 91, + 230, + 138, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 230, + 138, + 242 + ], + "spans": [ + { + "bbox": [ + 91, + 230, + 138, + 242 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 252, + 290, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 252, + 290, + 266 + ], + "spans": [ + { + "bbox": [ + 69, + 252, + 290, + 266 + ], + "type": "text", + "content": "B Did you use or create scientific artifacts?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 79, + 270, + 122, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 270, + 122, + 282 + ], + "spans": [ + { + "bbox": [ + 79, + 270, + 122, + 282 + ], + "type": "text", + "content": "Section 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 77, + 291, + 315, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 291, + 315, + 306 + ], + "spans": [ + { + "bbox": [ + 77, + 291, + 315, + 306 + ], + "type": "text", + "content": "B1. Did you cite the creators of artifacts you used?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 91, + 306, + 132, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 306, + 132, + 317 + ], + "spans": [ + { + "bbox": [ + 91, + 306, + 132, + 317 + ], + "type": "text", + "content": "Section 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 77, + 328, + 463, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 328, + 463, + 342 + ], + "spans": [ + { + "bbox": [ + 77, + 328, + 463, + 342 + ], + "type": "text", + "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 90, + 343, + 208, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 343, + 208, + 355 + ], + "spans": [ + { + "bbox": [ + 90, + 343, + 208, + 355 + ], + "type": "text", + "content": "Not applicable. Left blank." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 77, + 364, + 524, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 364, + 524, + 417 + ], + "spans": [ + { + "bbox": [ + 77, + 364, + 524, + 417 + ], + "type": "text", + "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 90, + 419, + 208, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 419, + 208, + 432 + ], + "spans": [ + { + "bbox": [ + 90, + 419, + 208, + 432 + ], + "type": "text", + "content": "Not applicable. Left blank." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 77, + 441, + 524, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 441, + 524, + 481 + ], + "spans": [ + { + "bbox": [ + 77, + 441, + 524, + 481 + ], + "type": "text", + "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 90, + 482, + 208, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 482, + 208, + 495 + ], + "spans": [ + { + "bbox": [ + 90, + 482, + 208, + 495 + ], + "type": "text", + "content": "Not applicable. Left blank." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 77, + 503, + 524, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 503, + 524, + 531 + ], + "spans": [ + { + "bbox": [ + 77, + 503, + 524, + 531 + ], + "type": "text", + "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 91, + 532, + 132, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 532, + 132, + 544 + ], + "spans": [ + { + "bbox": [ + 91, + 532, + 132, + 544 + ], + "type": "text", + "content": "Section 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 77, + 553, + 524, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 553, + 524, + 622 + ], + "spans": [ + { + "bbox": [ + 77, + 553, + 524, + 622 + ], + "type": "text", + "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 91, + 623, + 132, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 623, + 132, + 634 + ], + "spans": [ + { + "bbox": [ + 91, + 623, + 132, + 634 + ], + "type": "text", + "content": "Section 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 69, + 643, + 293, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 643, + 293, + 657 + ], + "spans": [ + { + "bbox": [ + 69, + 643, + 293, + 657 + ], + "type": "text", + "content": "C Did you run computational experiments?" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 79, + 662, + 122, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 662, + 122, + 674 + ], + "spans": [ + { + "bbox": [ + 79, + 662, + 122, + 674 + ], + "type": "text", + "content": "Section 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 77, + 683, + 524, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 683, + 524, + 711 + ], + "spans": [ + { + "bbox": [ + 77, + 683, + 524, + 711 + ], + "type": "text", + "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 91, + 712, + 132, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 712, + 132, + 723 + ], + "spans": [ + { + "bbox": [ + 91, + 712, + 132, + 723 + ], + "type": "text", + "content": "Section 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 67, + 729, + 522, + 748 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 729, + 522, + 748 + ], + "spans": [ + { + "bbox": [ + 67, + 729, + 522, + 748 + ], + "type": "text", + "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "spans": [ + { + "bbox": [ + 79, + 71, + 258, + 84 + ], + "type": "text", + "content": "ACL 2023 Responsible NLP Checklist" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 290, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 290, + 781, + 307, + 791 + ], + "type": "text", + "content": "108" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 77, + 70, + 523, + 236 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 77, + 70, + 523, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 70, + 523, + 111 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 523, + 111 + ], + "type": "text", + "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Section 3 and Appendix" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 77, + 120, + 523, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 120, + 523, + 174 + ], + "spans": [ + { + "bbox": [ + 77, + 120, + 523, + 174 + ], + "type": "text", + "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Section 3 and Appendix" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 183, + 523, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 183, + 523, + 236 + ], + "spans": [ + { + "bbox": [ + 77, + 183, + 523, + 236 + ], + "type": "text", + "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Section 3" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 67, + 247, + 522, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 247, + 522, + 261 + ], + "spans": [ + { + "bbox": [ + 67, + 247, + 522, + 261 + ], + "type": "text", + "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 79, + 264, + 127, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 264, + 127, + 277 + ], + "spans": [ + { + "bbox": [ + 79, + 264, + 127, + 277 + ], + "type": "text", + "content": "Left blank." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 76, + 287, + 523, + 539 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 76, + 287, + 523, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 287, + 523, + 326 + ], + "spans": [ + { + "bbox": [ + 76, + 287, + 523, + 326 + ], + "type": "text", + "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 76, + 337, + 523, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 337, + 523, + 389 + ], + "spans": [ + { + "bbox": [ + 76, + 337, + 523, + 389 + ], + "type": "text", + "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 76, + 400, + 523, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 400, + 523, + 453 + ], + "spans": [ + { + "bbox": [ + 76, + 400, + 523, + 453 + ], + "type": "text", + "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 76, + 463, + 519, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 463, + 519, + 488 + ], + "spans": [ + { + "bbox": [ + 76, + 463, + 519, + 488 + ], + "type": "text", + "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 76, + 498, + 523, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 498, + 523, + 539 + ], + "spans": [ + { + "bbox": [ + 76, + 498, + 523, + 539 + ], + "type": "text", + "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 290, + 781, + 307, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 781, + 307, + 791 + ], + "spans": [ + { + "bbox": [ + 290, + 781, + 307, + 791 + ], + "type": "text", + "content": "109" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file